ESG environmental automatic rating system and method based on multi-source data fusion

Abstract

This invention discloses an automatic ESG environment rating system and method based on multi-source data fusion. For contactless rating scenarios, it introduces flux intensity indicators and inter-source difference entropy values ​​as de-obfuscation criteria, constructing a hysteresis state machine to form a nonlinear locking logic that is sensitive to triggering but stringent to unlocking. After entering the hysteresis-locked state, asymmetric strict constraints are applied to subjective and objective data, and the reset effect of feature fallback on the system's confidence state is masked. Finally, the state unlocking condition is strictly bound to the total accumulated value of objective evidence across time windows; a stable state is restored and a rating is output only when the total accumulated value of objective evidence exceeds the unlocking threshold. This invention effectively suppresses the instantaneous impact of misleading data through nonlinear state locking and existing evidence verification mechanisms, ensuring that the rating results have objective confidence support.

Description

Technical Field

[0001] This invention relates to the field of multi-source heterogeneous data processing and automatic rating technology, and more specifically, to an ESG environment automatic rating system and method based on multi-source data fusion. Background Technology

[0002] Multi-source heterogeneous data processing and automatic rating technology specifically involves the access perception, feature extraction, and fusion decision-making of non-contact multimodal data streams. Existing systems typically access enterprise disclosure reports and environmental sensor or regulatory records, and use time window linear weighting models or real-time threshold gating logic to calculate a comprehensive score on the standardized data to reflect the real-time status of the rated object.

[0003] The existing technology has the following shortcomings: Existing technologies generally suffer from two main problems: First, subjective disclosure data, such as corporate disclosure reports, are published frequently and at low cost. Rated entities often employ a strategy of high-frequency publication coupled with intermittent pauses to disrupt the model, but existing models lack the ability to non-linearly lock onto historical adversarial states. As a result, once the high-frequency subjective data flow pauses, the system incorrectly judges the risk as resolved based on real-time window statistics, causing rating results to fluctuate with data publication behavior. Second, for scarce and low-frequency objective monitoring data, such as environmental sensor data or regulatory records, there is a lack of rigid verification mechanisms based on available data. As a result, the system is prone to prematurely lifting risk controls during data gaps simply due to the passage of time, leading to erroneous output of low-confidence rating results. To address these problems, this invention proposes a solution. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an automatic ESG environment rating system and method based on multi-source data fusion to address the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Automatic ESG environment rating methods based on multi-source data fusion include: S101, acquire multi-source heterogeneous data streams of the object to be rated, and distinguish between subjective disclosure and objective monitoring data sources; based on the feature indicators of the data streams and the system confidence state inherited from the previous time window, update the current fusion scenario status identifier; wherein, the system confidence state includes at least a stationary state and a hysteresis-locked state; the feature indicators include at least a flux intensity indicator and an inter-source difference entropy value. S102, when the fusion scenario status identifier enters the hysteresis-locked state, an asymmetric strict constraint configuration is applied to the subjective disclosure data and the objective monitoring data; when the feature index falls back but the system confidence state is still in the hysteresis-locked state, the asymmetric strict constraint configuration is forcibly maintained, and the objective evidence sufficiency verification logic is initiated; in the hysteresis-locked state, the direct reset effect of the decline in flux intensity index and inter-source difference entropy value on the system confidence state is masked, and the objective evidence sufficiency verification logic determines whether to allow the transition from the hysteresis-locked state to the steady state based on the total accumulated value of objective evidence across time windows and the unlocking threshold; S103, in the hysteresis-locked state, determine whether the total accumulated value of the objective evidence exceeds the unlocking threshold; if it exceeds, unlock the hysteresis-locked state and output the ESG rating result; if it does not exceed, determine that the confidence level is non-convergent and block the output of the ESG rating result.

[0006] In a preferred embodiment, distinguishing between subjective disclosure data sources and objective monitoring data sources specifically includes: parsing the metadata tags or source protocol headers of the access data stream; marking subjective disclosure data sources originating from enterprise disclosure channels or third-party dissemination channels as subjective disclosure data sources; and marking objective monitoring data sources originating from sensor observation data, remote sensing observation data, or regulatory record data as objective monitoring data sources.

[0007] In a preferred embodiment, the update of the fusion scenario status identifier based on the feature index of the data flow includes real-time calculation of the throughput intensity index. The specific steps are as follows: setting a historical sliding time window, calculating the moving average throughput of the data flow within the historical sliding time window as a historical baseline; statistically analyzing the data access throughput within the current time window, and calculating one of its multiplier factor or deviation relative to the historical baseline; and using one of the multiplier factor or deviation as the throughput intensity index to reflect the sudden characteristics of the data flow relative to its historical level.

[0008] In a preferred embodiment, the feature index further includes an inter-source difference entropy value, which is calculated by: mapping multi-source heterogeneous data to a unified feature vector space, wherein text feature vectors are extracted for text data and normalization mapping is performed for numerical data; calculating the feature distance between different information sources in the unified feature vector space; and calculating the information entropy representing the degree of conflict between information sources based on the feature distance, which is used as the inter-source difference entropy value.

[0009] In a preferred embodiment, updating the current fusion scenario state identifier specifically includes: weighted coupling of the flux intensity index and the inter-source difference entropy value to generate an adversarial feature index; determining whether the adversarial feature index exceeds a preset anomaly warning threshold; when the anomaly warning threshold is exceeded and the system confidence state of the previous time window is in a stationary state, triggering a state transition, updating the current fusion scenario state identifier to a hysteresis-locked state, and updating the system confidence state to a hysteresis-locked state.

[0010] In a preferred embodiment, the asymmetric strict constraint configuration applied to subjective disclosure data and objective monitoring data specifically includes: performing a shrinking operation on the consistency verification space for subjective disclosure data, switching the tolerance range for determining the validity of data from a stationary state to a strict mode of hysteresis locking, and only allowing data whose feature distance in multi-source comparison is not greater than a preset verification threshold to pass the verification; and performing a shielding protection operation on the fusion weight for objective monitoring data, cutting off the positive correlation between its weight and data throughput intensity, and assigning fixed weights or compensation weights based on data scarcity to prevent objective monitoring data from being diluted under the impact of high-throughput subjective disclosure data.

[0011] In a preferred embodiment, when the feature index falls back but the system confidence state remains in a hysteresis-locked state, the asymmetric strict constraint configuration is forcibly maintained. Specifically, this includes: in the hysteresis-locked state, shielding the direct reset effect of the real-time feature index fallback on the system confidence state; even if the throughput intensity index and inter-source difference entropy value of the real-time data stream are both detected to have recovered to historical baseline levels, maintaining the shrinkage operation of the consistency verification space and the shielding protection operation of the fusion weights; and limiting the migration permission of the system confidence state to be triggered only by the objective evidence sufficiency verification logic.

[0012] In a preferred embodiment, the initiation of the objective evidence sufficiency verification logic specifically includes: employing a confidence accumulation mechanism across time windows; within each time window of the hysteresis-locked state, selecting objective monitoring data that satisfy multi-source logic constraints, and converting their effective information content into the current period's confidence contribution value, wherein the effective information content is one of the number of effective observation items, the number of coverage indicators, or the source coverage; performing linear accumulation or nonlinear superposition on the confidence contribution values ​​of multiple consecutive time windows to update the system's total accumulated objective evidence value; and generating a release command for the system's confidence state only when the total accumulated objective evidence value exceeds a preset unlocking threshold, allowing the system to recover to a stable state.

[0013] In a preferred embodiment, determining that the confidence level is not convergent and blocking the output of the ESG rating result specifically includes: suspending the current ESG rating task and suppressing the release of the current ESG rating result; generating and outputting a data supplementation collection instruction containing the type and magnitude of the objective monitoring data gap; pushing the task into the evidence retention queue, and retaining the accumulated total value of objective evidence in the evidence retention queue without resetting it, until the objective monitoring data input in the subsequent time window responds to the data supplementation collection instruction and fills the gap, or the preset retention period expires.

[0014] The ESG environment automatic rating system based on multi-source data fusion includes a hysteresis state machine maintenance unit, an asymmetric locking and holding unit, and an objective evidence sufficiency verification and rating output unit. The hysteresis state machine maintenance unit executes S101, reads multi-source heterogeneous data streams and distinguishes between subjective disclosure and objective monitoring, and updates the fusion scenario state identifier and system confidence state based on the flux intensity index, the inter-source difference entropy value, and the system confidence state inherited from the previous time window. The asymmetric locking and holding unit executes S102, applies asymmetric strict constraint configuration when the fusion scenario state identifier enters the hysteresis locking state, and maintains the constraint configuration and updates the total accumulated value of objective evidence when the feature index falls back and the system confidence state is still in the hysteresis locking state. The objective evidence sufficiency verification and rating output unit executes S103, arbitrates the release or blocking based on the total accumulated value of objective evidence and the unlocking threshold, outputting the ESG rating result when the release is made and the supplementary evidence control result when the blocking is made.

[0015] The advantages and benefits of the ESG environment automatic rating system and method based on multi-source data fusion in this invention are as follows: This invention introduces flux intensity and inter-source difference entropy as criteria for de-obfuscation and constructs a hysteresis state machine to form a trigger-sensitive but release-rigid nonlinear locking logic. This allows the system to shield itself from the reset effect of real-time feature fallback on the system's confidence state when high-frequency and highly differentiated adversarial features are detected, thus enabling the rating system to possess state memory capabilities to resist instantaneous data fluctuations. In the hysteresis-locked state, this invention uses locking constraints and the evidence accumulation result set R102 to enforce the contraction verification of subjective data and the weight protection of objective data. The state unlocking condition is strictly bound to the cumulative value of the effective information content of objective monitoring data across a time window, rather than relying on the natural passage of time. Based on the above state memory and stock verification mechanism, it can suppress repeated oscillations in rating results caused by impulsive interference from subjective data, prevent the risk of misjudgment during the window period of missing objective evidence, and thus ensure that the final output ESG rating result has quantifiable objective confidence. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0017] 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.

[0018] This invention provides an automatic ESG environment rating system and method based on multi-source data fusion, applicable to scenarios lacking real-time on-site verification methods and relying on non-contact rating based on multi-source heterogeneous data streams. In such scenarios, the rating system can only acquire the release frequency and content characteristics of the accessed data. Furthermore, companies often use frequently released public relations press releases or online trolls to create high-frequency subjective disclosure data streams, attempting to cover up low-frequency and scarce objective monitoring data such as sensor observations or regulatory records. This high-frequency release combined with intermittent cessation of the adversarial strategy makes it difficult for traditional fusion models based on linear weighting or real-time threshold judgments to distinguish between genuine hotspots and misleading data interference. It is highly susceptible to misjudging the risk as resolved when adversarial characteristics temporarily subside, leading to unexpected fluctuations in rating results and failure of confidence level determination.

[0019] The technical problem this invention aims to solve is how to construct a nonlinear control mechanism with state memory capabilities to identify and defend against subjective data shocks and prevent rating results from being interfered with by instantaneous data fluctuations, under the conditions of subjective data interfering with objective data and logical judgment difficulties. To this end, this method introduces flux intensity indicators and inter-source difference entropy values ​​as the criteria for de-obfuscation, and constructs a hysteresis state machine to form a nonlinear locking logic that is sensitive to triggering but stringent to unlocking. This method uses hysteresis features and a state result set as a unified state record carrier. When high-frequency and highly differentiated adversarial features are detected, a hysteresis locking state is triggered, and an asymmetric strict constraint configuration is maintained until the total accumulated value of objective evidence in the locking constraint and evidence accumulation result set meets a preset unlocking threshold, thereby effectively blocking abnormally disclosed data when objective evidence is insufficient.

[0020] Based on the above design, this invention constructs an automatic ESG environment rating method flow based on multi-source data fusion, consisting of steps S101 to S103 sequentially. (Refer to...) Figure 1 , Figure 1 This is a schematic diagram of the method flow of the present invention, which includes: Step S101, hysteresis state machine maintenance, is used during the multi-source heterogeneous data access and sensing phase to perform feature calculations and state machine updates on the data stream, completing adversarial feature quantification and system confidence state updates. This step reads the multi-source heterogeneous data stream X101 and the hysteresis feature and state result set R101 from the previous time window, analyzes the source attributes to distinguish between subjective disclosure and objective monitoring data, calculates the throughput intensity index and inter-source difference entropy value in parallel, generates an adversarial feature index through entropy gating coupling, and updates the fusion scenario state identifier and system confidence state based on the adversarial feature index. Finally, the feature index field and state field are written into R101 for subsequent step S102 to determine whether to enable asymmetric constraints. Specifically, X101 provides raw data access and protocol header information; entropy gating logic is used to determine whether the entropy value of the inter-source difference reaches the gating threshold. When it does not reach the threshold, the adversarial feature index is suppressed or set to zero to reduce false triggering caused by short-term noise and occasional fluctuations; the fusion scenario state identifier is used to indicate whether the data stream triggers the adversarial feature discrimination result; the system confidence state is used to indicate whether the output of ESG rating results is allowed and to limit the state transition to be triggered only by the objective evidence sufficiency verification logic; R101 serves as a memory carrier, carrying the perception result at the current moment and supporting cross-time window inheritance. This step only allows the system confidence state to transition from a stationary state to a hysteresis-locked state when the triggering condition is met, and does not perform the operation of exiting the hysteresis-locked state based on feature fallback.

[0021] Step S102, Asymmetric Locking and Holding, is used in the risk control phase after the perception state is established to perform asymmetric constraint configuration and forced maintenance operations on heterogeneous data sources, completing noise reduction and evidence accumulation for subjective data noise. This step reads the hysteresis feature and state result set R101 and the objective monitoring data subset from the multi-source heterogeneous data stream X101, generates a locking constraint and evidence accumulation result set R102 by masking feature fallback and accumulating effective information, for subsequent step S103 to read and perform blocking determination. R102 records the currently effective asymmetric strict constraint configuration and the total value of accumulated objective evidence used to determine whether the lock can be released.

[0022] Step S103, the blocking fusion rating, is used during the evidence assessment and delivery phase to perform threshold comparison and confidence level arbitration on accumulated objective evidence, completing the rating blocking or release output for abnormally disclosed data. This step reads the locking constraints and evidence accumulation result set R102, determines unlocking or blocking based on the comparison result between the total accumulated value of objective evidence and the unlocking threshold, and generates a blocking output and supplementary evidence control result set R103, which is used to issue ratings or provide feedback supplementary evidence instructions. The blocking output and supplementary evidence control result set R103 is the final delivery carrier, encapsulating ESG rating results that meet confidence requirements, a supplementary evidence maintenance queue when confidence levels have not converged, and data supplementation collection instructions for gaps.

[0023] Through the above steps, this method utilizes hysteresis features and the state result set R101 as a cross-time window carrier for maintaining the system's confidence state. When the fused scene state identifier enters the hysteresis-locked state, logical masking cuts off the reset effect of the decline in flux intensity indicators and inter-source difference entropy values ​​on the state. This method strictly binds the release condition of the hysteresis-locked state to the locking constraints and the evidence accumulation result set R102, forcing that the total value of accumulated objective evidence must exceed the unlocking threshold before an unlocking command can be generated. Thus, the system transforms the defense against misleading data interference into the quantitative verification of the effective stock of objective monitoring data, ensuring that the final output ESG rating result is based on objective evidence that meets the threshold requirements, rather than depending on the duration of adversarial feature disappearance.

[0024] The implementation process and operational effects of the method of the present invention will be described in detail below with reference to specific embodiments. It should be understood that the embodiments are only used to illustrate the technical solution of the present invention, and not to limit it. The relevant steps, parameters and module divisions can be appropriately adjusted without changing the essence of the invention.

[0025] For ease of understanding, this embodiment is described under a unified system architecture, which can be modified equivalently according to actual needs. The system includes a hysteresis state machine maintenance unit, an asymmetric locking and holding unit, and an objective evidence sufficiency verification and rating output unit. (Refer to...) Figure 2 , Figure 2 This is a schematic diagram of the system structure of the present invention. Specifically, the hysteresis state machine maintenance unit is used to execute step S101, which reads the multi-source heterogeneous data stream X101 and the hysteresis characteristics and state result set R101 of the previous time window during the access sensing stage of multi-source heterogeneous data, analyzes the source attributes to distinguish between subjective disclosure type and objective monitoring type data, calculates the throughput intensity index and the inter-source difference entropy value, generates the adversarial feature index, and updates the fusion scene state identifier and system confidence state accordingly to obtain the updated R101. The asymmetric locking and holding unit is used to execute step S102, which applies asymmetric strict constraint configuration to the subjective disclosure type data and objective monitoring type data when the data stream feature index falls but the system confidence state is still in the hysteresis locking state, and forcibly maintains the strict constraint configuration, while updating the total value of objective evidence accumulation and writing it into the locking constraint and evidence accumulation result set R102. The objective evidence sufficiency verification and rating output unit is used to perform step S103, which verifies the evidence coverage and consistency of objective monitoring data under the strict constraint configuration, reads the cumulative total value of objective evidence and the unlocking threshold in R102 and arbitrates release or block accordingly, generates the blocking output and supplementary evidence control result set R103 and outputs it.

[0026] In an optional embodiment, step S101 is executed by the hysteresis state machine maintenance unit. This unit generates adversarial discrimination features at the data access layer and maintains the system confidence state. It maps conflict features of multi-source data to hysteresis-locked states or stationary states and writes them into R101, thereby providing a unified state triggering caliber for the subsequent step S102. This step is only responsible for feature calculation and state identifier updating; it does not directly perform constraint operations. Its implementation process includes source attribute parsing and classification, feature index calculation, adversarial feature gating coupling, state machine transition, and write-back.

[0027] In source attribute parsing and classification, metadata tags or source protocol headers are extracted from X101. Based on metadata such as protocol header fields, channel type, source authentication mark, collection link identifier, and regulatory record type mark, data from enterprise disclosure channels or third-party dissemination channels are marked as subjective disclosure, and data from sensor observation, remote sensing observation, or regulatory records are marked as objective monitoring. The classification marks are then appended to the data stream metadata for subsequent channel-specific calculations.

[0028] In the calculation of feature indicators, a historical sliding time window is set for the throughput dimension. The moving average of the data access rate within the window is calculated as the historical baseline. The real-time access throughput of the current time window is statistically analyzed, and its multiplier factor or deviation from the historical baseline is calculated to obtain the throughput intensity index. For the difference dimension, multi-source heterogeneous data is mapped to a unified feature vector space. The feature distances of different information sources in this space are calculated, and the information entropy representing the degree of conflict between information sources is calculated based on the distance distribution to obtain the inter-source difference entropy value.

[0029] In the adversarial feature gating coupling, an adversarial feature index is generated using entropy gating logic based on the flux intensity index and the inter-source difference entropy value. When the inter-source difference entropy value is lower than a preset entropy gating threshold, the adversarial feature index is forced to zero to exercise a veto power and avoid false triggering. Otherwise, the flux intensity index and the inter-source difference entropy value are weighted and coupled to generate a non-zero adversarial feature index. The larger the value, the higher the probability that it reflects high-heat and high-conflict anomaly data characteristics.

[0030] During state machine transitions and write-back, the system confidence state and fusion scenario state identifier from the previous time window R101 are read. When there is no previous time window R101, the system confidence state is initialized to a stationary state, and the feature index is initialized to the baseline value or zero value. When the adversarial feature index exceeds the anomaly warning threshold and the previous state is a stationary state, a state transition is triggered, updating the fusion scenario state identifier to a hysteresis-locked state, and updating the system confidence state to a hysteresis-locked state. When the previous state is a hysteresis-locked state, the direct reset effect of the decline in flux intensity index and inter-source difference entropy value on the system confidence state is masked; only the current feature index is recorded, and the hysteresis-locked state is forcibly maintained unchanged. The system confidence state transition permission is limited to being triggered only by the objective evidence sufficiency verification logic. Finally, the flux intensity index, inter-source difference entropy value, adversarial feature index, system confidence state, and fusion scenario state identifier are written into R101.

[0031] To facilitate implementation and standardize the calculation of adversarial features, this embodiment presents an optional calculation scheme. Let... This is the entropy gate threshold used to determine the entropy value of inter-source differences. Whether the consensus-breaking level has been reached or not, if it is below this threshold, the adversarial index is directly rejected to avoid false triggering. Let... Let be the flux intensity index, representing the degree of deviation of the current time window's data access flux from the historical baseline flux, and take a non-negative real number value. Let be the inter-source difference entropy value, representing the degree of conflict between different information sources in a unified feature vector space, and take the value of a non-negative real number. and These are coupling weights, used to balance the contributions of the flux term and the conflict term, respectively, satisfying... and and All are non-negative. Let... To counteract the characteristic index, used to quantify the impact of subjective disclosure and drive the triggering of a hysteresis state machine, a non-negative real number is taken. For example, a logarithmic term is introduced into the flux intensity index to smooth extreme flow shocks, thus forming the counteracting characteristic index. The calculation method is as follows, where The natural logarithm function: The threshold and weight examples and reasons are as follows.

[0032] Entropy Gating Threshold Normal operating period can be taken as an example The lower quantile, such as the 20th percentile, is chosen because this threshold acts as a veto gate, aiming to reduce false triggers rather than increase sensitivity. The coupling weight can be exemplified by taking... , As the default balanced configuration, if the business is more concerned about abnormal heat levels, then increase the [configuration / adjustment]. If more attention is paid to cross-source conflicts, it will improve The reason for this is that the two correspond to the sensitivity adjustment knobs for abnormal heat and abnormal conflict, respectively. It is recommended that the state machine trigger threshold be uniformly named the "Abnormal Warning Threshold". For example, the normal operating period can be taken. The high quantile, such as the 80th percentile with an added safety margin, is chosen because the trigger threshold needs to prioritize controlling false alarms while ensuring a rapid transition to a hysteresis-locked state when an abnormal shock occurs. Using this approach, S101 quantifies the risk characteristics and status indicators within R101, allowing subsequent step S102 to directly execute asymmetric constraints based on the R101 status field without repeatedly tracing back the original data stream.

[0033] In an optional embodiment, step S102 is executed by the asymmetric locking and holding unit, which performs a differentiated defense strategy in the hysteresis-locked state and establishes an unlocking verification mechanism based on the total accumulated value of objective evidence. This step forces the maintenance of the asymmetric strict constraint configuration in the hysteresis-locked state and updates the total accumulated value of objective evidence to ensure that the defense posture is maintained even when objective evidence is insufficient. Its implementation process includes state reading and policy routing, asymmetric constraint configuration generation, state reset and shielding protection, objective evidence screening and cross-window accumulation, and result writing to R102.

[0034] In state reading and policy routing, the fusion scenario state identifier is read from R101. If it is in a stable state, the normal verification tolerance and dynamic weight configuration are maintained. If it is in a hysteresis-locked state, the asymmetric defense logic is triggered to proceed to subsequent processing.

[0035] In the generation of asymmetric constraint configurations, asymmetric strict constraint configurations are generated for subjective disclosure data and objective monitoring data in a hysteresis-locked state. For subjective disclosure data, a consistency verification space contraction is performed, switching the tolerance range for valid data to a stringent mode, allowing only data with a feature distance not greater than the consistency verification threshold to pass verification. Specifically, the consistency verification space contraction is used when the system confidence state is in a hysteresis-locked state, shrinking the judgment tolerance boundary of the consistency verification index from the basic tolerance boundary to the lock tolerance boundary, making the unlocking determination require more sufficient objective evidence. For objective monitoring data, a fusion weight masking protection is performed, severing the positive correlation between its weight and data throughput intensity, assigning fixed weights or compensating weights based on data scarcity, preventing the dilution of objective evidence under the impact of high-throughput subjective disclosure data. The consistency verification threshold can be, for example, taken as the high quantile of the normal consistency distance distribution, such as the 90th percentile, and can be tightened to the 80th percentile in the stringent mode. This is because the stringent mode needs to reduce the risk of false positives and use a quantifiable error distribution as the basis for setting the threshold.

[0036] In the state reset shielding protection, when the system confidence state is in a hysteresis-locked state, the direct reset effect of the system confidence state by the decline of the flux intensity index and the inter-source difference entropy value is shielded. Even if the flux intensity index and the inter-source difference entropy value are detected to have recovered to the historical baseline level, the asymmetric strict constraint configuration is still forcibly maintained and not revoked, and the migration permission of the system confidence state is limited to being triggered only by the objective evidence sufficiency verification logic established in this step.

[0037] In the objective evidence screening and cross-window accumulation process, a subset of objective monitoring data is screened from X101, and the historical accumulated values ​​inherited from the previous time window R102 are read. Outlier noise is eliminated based on multi-source logical constraints. The effective information content of the current valid data is calculated and converted into the current confidence contribution value. Then, cross-time window accumulation is performed, and the current confidence contribution value is superimposed with the historical accumulated value to update the total accumulated value of objective evidence. When there is no previous time window R102, the total accumulated value of objective evidence is initialized to zero or the baseline value. The multi-source logical constraints include at least time window consistency constraints and source consistency constraints. Time window consistency constraints are used to eliminate objective monitoring items that are not within the current time window, and source consistency constraints are used to eliminate objective monitoring items with missing source labels or unverifiable acquisition links.

[0038] Finally, the asymmetric strict constraint configuration, the updated cumulative total of objective evidence, and the unlocking threshold are written into R102. The unlocking threshold is a system-preset strategy configuration parameter used to characterize the minimum cumulative total of objective evidence required to unlock the hysteresis lock state. The cumulative total of objective evidence quantification system in R102 uses the progress of unlocking the current risk state based on the existing evidence as the core quantitative basis for the subsequent blocking determination in step S103.

[0039] To ensure the reproducibility of the evidence accumulation process, this embodiment provides an optional integration iteration approach: Let... Let the time window number be... The first in the current time window One valid objective monitoring data item. (Set) Let the total accumulated objective evidence at the end of the current time window be denoted as . Let this be the cumulative total of objective evidence inherited from the previous time window R102. This represents the number of valid objective monitoring data entries within the current time window, and is a non-negative integer. Let... For the first The effective coverage or effective information content of objective data is a non-negative value. For example, it can be the ratio of the number of covered indicators to the total number of predefined key indicators, ensuring it falls within the range of 0 to 1. This is because this caliber is comparable across sources and has stable dimensions, facilitating cross-window accumulation and threshold setting. Let... For the first The source weight of each objective data point is a non-negative value, used to reflect the difference between the reliability of the source and the credibility of the data collection process. Therefore: Examples of weights and thresholds, and the reasons for them, are as follows.

[0040] Source weight For example, verifiability can be graded and assigned values, such as 1.2 for regulatory records, 1.0 for sensor observations, and 0.8 for remote sensing observations. This is because different sources have different anti-counterfeiting capabilities and traceability, and the weights are used to explicitly reflect the differences in reliability in the amount of evidence. It is recommended that the unlocking threshold be uniformly named "Unlocking Threshold". , and With the same dimensions, it can be exemplified by setting the minimum objective coverage target backwards. For instance, if the requirement is to cover K categories of key indicators with each category appearing at least once, then the typical scenario for meeting this coverage target would be... Horizontal as The reason is that the unlocking conditions are tied to the total value of accumulated objective evidence, rather than to the passage of time.

[0041] In an optional embodiment, step S103 is executed by the objective evidence sufficiency verification and rating output unit, establishing a threshold-based mandatory verification mechanism based on evidence reserves. This means that a rating result can only be output after the effective accumulation of objective monitoring data forms credibility support. This step, acting as the exit gate of the hysteresis state machine, is responsible for deciding whether the current rating task has sufficient objective evidence to proceed, or whether it must be suspended due to the unresolved inconsistency data characteristics, thereby preventing low-confidence rating results from being released.

[0042] The implementation process of this step includes determining the sufficiency threshold of evidence, arbitrating the confidence convergence state, calculating and publishing the rating results, generating blocking and supplementary evidence instructions, and clearing timeouts and preventing deadlocks.

[0043] In determining the sufficiency of evidence threshold, the total accumulated value of objective evidence and the unlocking threshold are read from the locking constraint and evidence accumulation result set R102. The total accumulated value of objective evidence is then compared with the unlocking threshold according to numerical comparison rules to determine whether the evidence is sufficiently sufficient. This determination process relies solely on the current amount of evidence and does not consider the passage of time.

[0044] In the confidence convergence arbitration, the system processing branch is determined based on the above judgment results. If the judgment evidence is sufficient, the system confidence level is considered to have converged to the credible interval, an unlocking command is generated, and the system confidence state is updated to a stationary state. The multi-source heterogeneous data stream X101 is read, and after filtering or weighting according to the asymmetric strict constraints in R102 to form valid data for rating, the rating model is called to calculate the ESG rating result based on the current valid data and output it externally. If the judgment evidence is insufficient, the system is considered to still be in a confidence non-convergence state with a high incidence of abnormal disclosure data, triggering a blocking branch, prohibiting the output of the current rating result, and entering the blocking remediation process. Time-based judgments are only used for suspended task cleanup and resource protection and do not participate in the unlocking judgment.

[0045] In the blocking suspension and supplementary evidence generation process, the pending evidence retention queue in the previous time window blocking output and supplementary evidence control result set R103 is first read as historical inventory. For non-convergent confidence states, the current rating task is encapsulated and pushed into the pending evidence retention queue, and the release of the current ESG rating result is suppressed. The system analyzes the gap composition of evidence accumulation in the locking constraints and evidence accumulation result set R102, generates a data supplementation collection instruction that includes the type and magnitude of objective monitoring data gaps, and outputs the instruction to the data acquisition end to guide targeted supplementary evidence collection.

[0046] In the timeout cleanup and deadlock prevention process, the suspension duration of tasks remaining in the pending evidence queue is scanned periodically. If the suspension duration of a task exceeds the preset retention duration, a cleanup operation is performed, removing the task from the queue and marking it as an evaluation failure. Simultaneously, a summary of the blocking reason is recorded as insufficient evidence and timeout, to prevent deadlocked tasks from indefinitely consuming system resources. This cleanup process does not change the unlocking threshold or the criteria for determining the amount of evidence remaining. The preset retention duration is a system-preset duration-based policy configuration parameter used to limit the maximum suspension time of tasks in the pending evidence retention queue.

[0047] Finally, the calculated ESG rating results, the updated pending certification queue, the generated data supplementation collection instructions, and the summary of the blocking cause are written into the blocking output and certification control result set R103. This result set interfaces with external publishing interfaces and data source feedback interfaces to achieve closed-loop control of the rating task.

[0048] To clarify the logical boundaries of the unlocking determination, this embodiment provides the following status arbitration criteria. The output status code is denoted as... ,in This indicates that the passage is permitted and the rating result is output. This indicates that the process has been blocked and is pending further documentation. This indicates that the process was blocked and failed due to timeout. (The item being rated is...) In the current time window The cumulative total of objective evidence at the end is denoted as The unlock threshold is denoted as The task suspension time is recorded as follows: For example, the count is accumulated according to the time window, and the preset duration is recorded as... Accordingly, the arbitration output status code is determined according to the following rules. .when season .when and season .when and season The threshold example and reason are as follows. Preset hold duration. For example, 30 time windows can be used, or the number of time windows can be calculated based on the longest suspension time allowed by the business. The reason is that this parameter is responsible for resource protection and closed-loop convergence, preventing the queue from growing indefinitely and allowing reasonable time for supplementary certification. The status is used for resource recycling and task closure, and does not participate in unlocking judgment. It ensures that only when the total value of accumulated objective evidence reaches the unlocking threshold can the output rating result be released.

[0049] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0050] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0051] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0052] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0053] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic ESG environment rating method based on multi-source data fusion, characterized in that, include: S101, Obtain multi-source heterogeneous data streams of the object to be rated, and distinguish between subjective disclosure data sources and objective monitoring data sources; Based on the feature indicators of the data stream and the system confidence state inherited from the previous time window, update the current fusion scenario status identifier; wherein, the system confidence state includes at least the stationary state and the hysteresis-locked state; the feature indicators include at least the flux intensity index and the inter-source difference entropy value. S102, when the fusion scenario status identifier enters the hysteresis-locked state, an asymmetric strict constraint configuration is applied to the subjective disclosure data and the objective monitoring data; when the feature index falls back but the system confidence state is still in the hysteresis-locked state, the asymmetric strict constraint configuration is forcibly maintained, and the objective evidence sufficiency verification logic is initiated; in the hysteresis-locked state, the direct reset effect of the decline in flux intensity index and inter-source difference entropy value on the system confidence state is masked, and the objective evidence sufficiency verification logic determines whether to allow the transition from the hysteresis-locked state to the steady state based on the total accumulated value of objective evidence across time windows and the unlocking threshold; S103, in the hysteresis-locked state, determine whether the total accumulated value of the objective evidence exceeds the unlocking threshold; if it exceeds, unlock the hysteresis-locked state and output the ESG rating result; if it does not exceed, determine that the confidence level is non-convergent and block the output of the ESG rating result.

2. The automatic ESG environment rating method based on multi-source data fusion according to claim 1, characterized in that, The distinction between subjective disclosure and objective monitoring data sources specifically includes: parsing the metadata tags or source protocol headers of the access data stream; marking subjective disclosure data sources originating from enterprise disclosure channels or third-party dissemination channels as subjective disclosure data sources; and marking objective monitoring data sources originating from sensor observation data, remote sensing observation data, or regulatory record data as objective monitoring data sources.

3. The automatic ESG environment rating method based on multi-source data fusion according to claim 1, characterized in that, The method of updating the fusion scenario status identifier based on data flow feature indicators includes real-time calculation of throughput intensity indicators. The specific steps are as follows: setting a historical sliding time window, calculating the moving average throughput of the data flow within the historical sliding time window as a historical baseline; statistically analyzing the data access throughput within the current time window, and calculating one of its multiplier factor or deviation relative to the historical baseline; and using one of the multiplier factor or deviation as the throughput intensity indicator to reflect the sudden characteristics of the data flow relative to its own historical level.

4. The automatic ESG environment rating method based on multi-source data fusion according to claim 3, characterized in that, The feature index also includes the inter-source difference entropy value, which is calculated as follows: mapping multi-source heterogeneous data to a unified feature vector space, wherein text feature vectors are extracted for text data and normalization mapping is performed for numerical data; calculating the feature distance between different information sources in the unified feature vector space; and calculating the information entropy representing the degree of conflict between information sources based on the feature distance, which is used as the inter-source difference entropy value.

5. The automatic ESG environment rating method based on multi-source data fusion according to claim 4, characterized in that, The step of updating the current fusion scenario status identifier specifically includes: weighted coupling of the flux intensity index and the inter-source difference entropy value to generate an adversarial feature index; determining whether the adversarial feature index exceeds a preset anomaly warning threshold; when the anomaly warning threshold is exceeded and the system confidence state of the previous time window is in a stationary state, triggering a state transition, updating the current fusion scenario status identifier to a hysteresis-locked state, and updating the system confidence state to a hysteresis-locked state.

6. The automatic ESG environment rating method based on multi-source data fusion according to claim 1, characterized in that, The asymmetric and strict constraint configuration applied to subjective disclosure data and objective monitoring data specifically includes: performing a shrinking operation on the consistency verification space for subjective disclosure data, switching the tolerance range for determining the validity of data from a stationary state to a strict mode of hysteresis locking, and only allowing data whose feature distance in multi-source comparison is not greater than a preset verification threshold to pass the verification; and performing a shielding protection operation on the fusion weight for objective monitoring data, cutting off the positive correlation between its weight and data throughput intensity, and assigning fixed weights or compensation weights based on data scarcity to prevent objective monitoring data from being diluted under the impact of high-throughput subjective disclosure data.

7. The automatic ESG environment rating method based on multi-source data fusion according to claim 6, characterized in that, When the feature index falls back but the system confidence state remains in a hysteresis-locked state, the asymmetric strict constraint configuration is forcibly maintained. Specifically, this includes: in the hysteresis-locked state, shielding the direct reset effect of the real-time feature index fallback on the system confidence state; even if the throughput intensity index and the inter-source difference entropy value of the real-time data stream are both detected to have recovered to the historical baseline level, the shrinkage operation of the consistency verification space and the shielding protection operation of the fusion weight are still not revoked; and limiting the migration permission of the system confidence state to be triggered only by the objective evidence sufficiency verification logic.

8. The automatic ESG environment rating method based on multi-source data fusion according to claim 7, characterized in that, The logic for initiating the objective evidence sufficiency verification specifically includes: employing a confidence accumulation mechanism across time windows; within each time window of the hysteresis-locked state, selecting objective monitoring data that meet multi-source logical constraints, and converting their effective information content into the current period's confidence contribution value, wherein the effective information content is one of the number of effective observation items, the number of coverage indicators, or the source coverage; performing linear accumulation or non-linear superposition on the confidence contribution values ​​of multiple consecutive time windows to update the system's total objective evidence accumulation value; and generating a release command for the system's confidence state only when the total objective evidence accumulation value exceeds a preset unlocking threshold, allowing the system to recover to a stable state.

9. The automatic ESG environment rating method based on multi-source data fusion according to claim 8, characterized in that, The determination that the confidence level is not convergent and the output of the ESG rating result is blocked specifically includes: suspending the current ESG rating task and suppressing the release of the current ESG rating result; generating and outputting a data supplement collection instruction containing the type and magnitude of the objective monitoring data gap; pushing the task into the evidence retention queue, and retaining the accumulated total value of objective evidence in the evidence retention queue without resetting it, until the objective monitoring data input in the subsequent time window responds to the data supplement collection instruction and fills the gap, or the preset retention period expires.

10. An automatic ESG environmental rating system based on multi-source data fusion, used to implement the automatic ESG environmental rating method based on multi-source data fusion as described in any one of claims 1-9, characterized in that, The system includes a hysteresis state machine maintenance unit, an asymmetric locking and holding unit, and an objective evidence sufficiency verification and rating output unit. The hysteresis state machine maintenance unit executes S101, reads multi-source heterogeneous data streams and distinguishes between subjective disclosure and objective monitoring types, and updates the fusion scenario state identifier and system confidence state based on the flux intensity index, the inter-source difference entropy value, and the system confidence state inherited from the previous time window. The asymmetric locking and holding unit executes S102, applies an asymmetric strict constraint configuration when the fusion scenario state identifier enters the hysteresis locking state, and maintains the constraint configuration and updates the total accumulated value of objective evidence when the feature index falls back and the system confidence state is still in the hysteresis locking state. The objective evidence sufficiency verification and rating output unit executes S103, arbitrates the release or blocking based on the total accumulated value of objective evidence and the unlocking threshold, outputting the ESG rating result when releasing and the supplementary evidence control result when blocking.

Images