An asset ledger intelligent checking method based on a learnable fuzzy measure
By employing a learnable fuzzy measure method based on Choquet similarity, the problem of data inconsistency in asset ledgers was solved, enabling efficient and intelligent identification and matching of asset anomalies, and improving the accuracy and efficiency of asset inventory.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG YOUCHANG ELECTRIC POWER TECH CO LTD
- Filing Date
- 2026-02-11
- Publication Date
- 2026-05-22
Smart Images
Figure CN121684340B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent auditing technology, and in particular to an intelligent asset ledger verification method based on learnable fuzzy measures. Background Technology
[0002] Asset ledger records are typically distributed across multiple business systems, such as the financial system, the physical inventory management system, and departmental self-built ledger systems. Due to inconsistent data entry standards, differences in field naming, time delays, and human error, inconsistencies, duplicate entries, or assets being incorrectly attributed often occur between different systems. These inconsistencies not only affect the accuracy of asset records but can also lead to distorted financial statements and difficulties in asset inventory checks.
[0003] Traditional asset verification methods primarily rely on rule-based matching and manual comparison, such as matching using fixed field names, amount thresholds, or fuzzy search algorithms. This approach often fails for records with minor field differences or inconsistent formats, especially when fields like name, model, and responsible person have spelling variations or semantic similarities, making it difficult for rule-based matching to identify potential relationships. Furthermore, manual verification is time-consuming and labor-intensive, making it unsuitable for dynamic, massive, and multi-source ledger environments.
[0004] Some studies have attempted to utilize machine learning methods for asset matching, such as similarity calculations based on vectorized representations or supervised classification models. However, these methods generally rely on pre-defined similarity metrics (such as Euclidean distance or cosine similarity), which cannot effectively describe the non-linear interactions between multiple fields, resulting in insufficient generalization performance in complex matching scenarios. Furthermore, existing models lack targeted learning mechanisms for "difficult-to-distinguish" similar assets, making it difficult to achieve intelligent, fine-grained asset anomaly identification.
[0005] Therefore, how to provide an intelligent asset ledger verification method based on learnable fuzzy measures is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0006] One objective of this invention is to propose an intelligent asset inventory method based on learnable fuzzy measures. This invention proposes an intelligent optimization method for asset inventory based on Choquet similarity comparison learning. By constructing learnable fuzzy measures and combining them with Lovarz extensions to achieve differentiability of Choquet integrals, and using Choquet similarity as the sole contrast measure, the method completes soft matching and anomaly identification of asset inventory fields, thereby significantly improving the accuracy and intelligence level of asset inventory.
[0007] An intelligent asset ledger verification method based on learnable fuzzy measure according to an embodiment of the present invention includes the following steps:
[0008] Collect asset ledger field information, perform standardization processing and vectorization encoding on the fields, and generate a unified-dimensional ledger field vector representation;
[0009] Construct learnable fuzzy measures and define corresponding fuzzy measure capacity parameters for any subset of the ledger field vectors;
[0010] Based on the fuzzy measure, the differentiability of the Choquet integral is achieved using the Lovatz extension, and the Choquet similarity of the candidate asset pairs is obtained.
[0011] Using the Choquet similarity as the sole contrast measure, a contrastive learning objective function is constructed, which pairs positive samples from multiple sources of the same asset and pairs negative samples from records from different assets but with similar field features.
[0012] In the contrastive learning process, monotonic projection constraints are applied to the fuzzy measure, and interactive sparse regularization is applied to obtain the Choquet similarity matrix obtained through training optimization.
[0013] Based on the Choquet similarity matrix obtained through training and optimization, a Top-k matching and threshold filtering strategy is adopted to output the soft matching results between different ledger records or the soft matching results between ledger records and physical assets, along with the corresponding confidence levels.
[0014] Based on the soft matching results, determine whether there are discrepancies in the asset accounting, such as duplicate entries, incorrect account settings, abnormal amounts, or inconsistent locations, and generate processing suggestions.
[0015] Furthermore, the asset ledger field information includes asset number, name, model and specifications, original value, net value, department, location, responsible person, accounting subject and date information;
[0016] The standardized and vectorized representation includes:
[0017] Numeric fields are standardized using a range scaling method;
[0018] Learnable embedding vectors are introduced for categorical fields to map discrete values of department, location, responsible person, and accounting subject to a continuous vector space;
[0019] Text fields are encoded through word decomposition;
[0020] Date fields are standardized using timestamp differentiation.
[0021] Generate a vector representation of the ledger fields.
[0022] Furthermore, the construction of the learnable fuzzy measure includes:
[0023] Define a fuzzy measure capacity parameter for all possible subsets of the ledger field vector;
[0024] An end-to-end training method based on gradient descent is used to update the fuzzy measure capacity parameter;
[0025] A set of fuzzy measure capacity parameters is obtained.
[0026] Furthermore, the calculation of the Choquet integral using the Lovarz extension includes:
[0027] Sort the ledger field vectors in descending order of their values;
[0028] Based on the arrangement result, the differences between adjacent field vectors are calculated sequentially, and the differences between adjacent field vectors are weighted and summed with the corresponding fuzzy measure capacity.
[0029] The capacity function is made continuous and differentiable using the Lovarz extension function.
[0030] Obtain the Choquet similarity of candidate asset pairs.
[0031] Furthermore, the construction of the contrastive learning objective function includes:
[0032] Multiple ledger records from the same asset are grouped into positive sample pairs, and the positive sample loss term is calculated based on Choquet similarity.
[0033] Records from different assets but with similar field features are grouped into hard negative sample pairs, and the negative sample loss term is calculated based on Choquet similarity.
[0034] The positive sample loss term and the negative sample loss term are weighted and combined to form a contrastive learning objective function with Choquet similarity as the sole similarity metric.
[0035] During training, the contrastive learning objective function is minimized to update the embedding model parameters.
[0036] Furthermore, the steps for generating the Choquet similarity matrix obtained through training and optimization include:
[0037] Choquet similarity of candidate asset pairs is obtained by performing Choquet integral calculation based on the vector representation of ledger fields and the fuzzy measure capacity parameter.
[0038] After each update, a monotonic projection operation is performed on the capacity parameter to generate a capacity parameter that satisfies the monotonicity constraint under the subset inclusion relation.
[0039] An interactive sparse regularization term is introduced into the training objective of the capacity function to impose sparsity constraints on the capacity parameters of higher-order field subsets.
[0040] The capacity parameter, after being processed by monotonic projection and interactive sparse regularization, is applied to the subsequent iterative calculation of Choquet similarity, and finally the trained and optimized similarity matrix is obtained.
[0041] Furthermore, the Top-k filtering based on the Choquet similarity matrix and threshold selection includes:
[0042] For each asset ledger record, calculate the Choquet similarity with all candidate records to obtain a similarity matrix;
[0043] From the similarity matrix, the top k candidate records are selected as preliminary matching results in order of their similarity values.
[0044] A similarity threshold is further set for the preliminary matching results, and candidate records below the threshold are removed to obtain the final soft matching results;
[0045] While outputting the soft matching results, a confidence score based on Choquet similarity is attached to each pair of matching records.
[0046] Furthermore, the setting of the similarity lower limit threshold includes:
[0047] In the training samples, the Choquet similarity distribution of positive sample matching pairs is statistically analyzed, and the lower limit threshold of similarity is determined according to the preset quantile of the distribution.
[0048] The calculation of the confidence score includes:
[0049] The Choquet similarity of each pair of matching records is mapped to the [0,1] interval using a normalization function to obtain the corresponding confidence score.
[0050] Furthermore, the step of determining asset discrepancies and generating processing suggestions based on soft matching results includes:
[0051] When multiple highly similar records of the same asset appear in different ledgers, it is determined to be a duplicate entry, and processing suggestions for merging or writing off are generated;
[0052] When there is an inconsistency in the accounting subject field of the matching records, it is determined that the subject is incorrectly set, and a processing suggestion for adjusting the subject is generated;
[0053] When the difference in the amount field of the matched records exceeds the preset tolerance threshold, it is determined to be an amount abnormal, and processing suggestions for supplementation or correction are generated;
[0054] When the matching records do not match in the location or responsible person field, it is determined that the location is inconsistent, and processing suggestions for transfer or responsibility tracing are generated.
[0055] The beneficial effects of this invention are:
[0056] This invention constructs a learnable fuzzy measure and utilizes Lovarz extensions to achieve differentiability of the Choquet integral, using Choquet similarity as the sole contrast measure. This accurately characterizes the nonlinear interactions between asset ledger fields. Compared to traditional inventory methods based on Euclidean distance or cosine similarity, this invention effectively identifies complex asset entries that require multiple field combinations for significant differentiation, thereby significantly improving matching accuracy and anomaly detection rate during asset inventory.
[0057] This invention applies monotonic projection and interactive sparse regularization constraints to the fuzzy measure during the contrastive learning training process. This ensures that the capacity function satisfies mathematical monotonicity while highlighting key field combinations and mitigating interference from invalid or redundant fields. This design enhances the model's generalization ability across different business scenarios and improves the interpretability of the matching results. It allows managers to intuitively understand which field combinations play a crucial role in asset inventory, thereby increasing the credibility of the inventory conclusions.
[0058] This invention utilizes a Top-k similarity matrix and threshold filtering mechanism to output soft-match results with confidence scores. Based on these results, it identifies discrepancies such as duplicate entries, incorrect account entries, abnormal amounts, and location discrepancies, and generates corresponding processing suggestions. Compared to traditional methods relying on manual verification, this invention enables an efficient and intelligent inventory clearance process in large-scale asset data environments, significantly reducing manual workload, minimizing errors and omissions, and improving the overall efficiency and application value of asset management and auditing. Attached Figure Description
[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0060] Figure 1 This is a flowchart of an intelligent asset ledger inventory method based on learnable fuzzy measure proposed in this invention;
[0061] Figure 2 This is a data collection and vectorized representation diagram for an intelligent asset ledger inventory method based on learnable fuzzy measure proposed in this invention;
[0062] Figure 3 This is a comparative learning target construction and training graph for an intelligent asset ledger inventory method based on learnable fuzzy measures proposed in this invention. Detailed Implementation
[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0064] refer to Figure 1 - Figure 3 A method for intelligent asset ledger verification based on learnable fuzzy measures includes the following steps:
[0065] Collect asset ledger field information, perform standardization processing and vectorization encoding on the fields, and generate a unified-dimensional ledger field vector representation;
[0066] Construct learnable fuzzy measures and define corresponding fuzzy measure capacity parameters for any subset of the ledger field vectors;
[0067] Based on the fuzzy measure, the differentiability of the Choquet integral is achieved using the Lovatz extension, and the Choquet similarity of the candidate asset pairs is obtained.
[0068] Using the Choquet similarity as the sole contrast measure, a contrastive learning objective function is constructed, which pairs positive samples from multiple sources of the same asset and pairs negative samples from records from different assets but with similar field features.
[0069] In the contrastive learning process, monotonic projection constraints are applied to the fuzzy measure, and interactive sparse regularization is applied to obtain the Choquet similarity matrix obtained through training optimization.
[0070] Based on the Choquet similarity matrix obtained through training and optimization, a Top-k matching and threshold filtering strategy is adopted to output the soft matching results between different ledger records or the soft matching results between ledger records and physical assets, along with the corresponding confidence levels.
[0071] Based on the soft matching results, determine whether there are discrepancies in the asset accounting, such as duplicate entries, incorrect account settings, abnormal amounts, or inconsistent locations, and generate processing suggestions.
[0072] In this embodiment, the asset ledger fields include asset number, name, model / specification, original value, net value, department, location, responsible person, accounting subject, and date information. To ensure that all fields can be processed within a unified representation space, this embodiment first performs standardization and vectorization representation, specifically including:
[0073] Range scaling is applied to numerical fields such as original value and net value, linearly mapping all values to a fixed range according to preset upper and lower limits to eliminate the interference of different units on subsequent similarity calculations; when values exceed the range, they can be truncated to the range boundary before scaling is performed, thereby ensuring that the mapping result is stable within the specified range.
[0074] For discrete value fields such as department, location, responsible person, and accounting subject, an index table is created and a learnable embedding vector is assigned. Each category is mapped to a continuous vector. This embedding is updated along with the model parameters during training to capture potential associations between categories.
[0075] For free text fields such as asset names and model specifications, the original string is divided into sub-word units by word decomposition, then encoded into a fixed-dimensional vector through an embedding table, and finally the overall representation of the field is formed by average pooling. This approach can mitigate the impact of differences in spelling or naming.
[0076] For time fields such as purchase date and transfer date, they are converted into timestamps and subtracted from the current reference time to reflect the asset's time characteristics in numerical form. These are then scaled to obtain a standardized numerical representation. All fields are converted into numerical vector fragments and concatenated in a predetermined order to form a complete ledger field vector representation. This representation can serve as a unified input for subsequent fuzzy measure construction and Choquet integral calculation, thereby enabling the fusion modeling of different types of fields during asset inventory.
[0077] In this embodiment, fuzzy measure capacity parameters are defined for all possible subsets of the ledger field vector.
[0078] The capacity parameter is stored in numerical form and is used to describe the weighting benchmark of different field subsets in subsequent Choquet integral calculations.
[0079] In order to adapt the capacity parameter to different business data, the capacity parameter and the embedded model parameter are treated as variables to be optimized during the model training process, and iteratively updated through an end-to-end training method based on gradient descent.
[0080] After each round of parameter updates, a monotonicity constraint is applied to the capacity parameter to ensure that the capacity value corresponding to any subset of fields is not less than the capacity value of its proper subset.
[0081] After the above processing, a set of fuzzy measure capacity parameters that meet the monotonicity constraint are obtained, which are used as inputs for Choquet integral similarity calculation.
[0082] In this embodiment, the Choquet similarity calculation uses the ledger field vector and the fuzzy measure capacity parameter as input. For any pair of ledger field vectors, an ordered sequence is formed based on the corresponding numerical values or similarity scores of each field, with the larger scores placed first.
[0083] A subset of cumulative fields is constructed sequentially, and the corresponding capacity value is retrieved from a predefined fuzzy metric capacity parameter. Then, following the Lovaz extension's calculation rules, the differences between adjacent sorting fields are weighted and summed, with the weights provided by the capacity parameter. This weighted summation process yields the Choquet similarity, reflecting the overall matching degree of the pair of ledger records.
[0084] In the above process, the implementation of the Choquet integral is placed within a differentiable framework, enabling the integral output to be backpropagated to the fuzzy measure capacity parameter and the embedding model parameter.
[0085] During training, the capacity parameter is recalculated after each round of updates, thereby gradually forming a similarity matrix that matches the actual asset matching relationship over multiple iteration cycles.
[0086] Any two ledger records can be used to obtain a unique Choquet similarity value through this calculation process, which is then used as the input to the contrastive learning objective function.
[0087] In this embodiment, the contrastive learning objective function is constructed and trained using Choquet similarity as the sole similarity metric. In the training dataset, multiple ledger records from the same asset are paired to form positive sample pairs. Then, from ledger records of different assets, several records that are similar to the positive samples in terms of field features are selected and paired to form hard negative sample pairs. For each positive sample pair, the Choquet similarity obtained in the third step is used to calculate the positive sample loss term; for each hard negative sample pair, the Choquet similarity is used to calculate the negative sample loss term.
[0088] The positive and negative sample loss terms are combined using preset weighting coefficients to obtain the complete contrastive learning objective function. During training, this objective function is the sole optimization target. By minimizing the value of the objective function, the embedding parameters of the ledger field vectors and the fuzzy measure capacity parameter are iteratively updated. Since Choquet similarity is used as the sole metric throughout the training process, it avoids mixing with traditional metrics such as Euclidean distance and cosine similarity, thus ensuring the consistency of the model's optimization direction.
[0089] Through the above construction and training steps, a contrastive learning model that can distinguish between positive sample pairs and hard negative sample pairs is obtained. Its embedding parameters and capacity parameters are updated and converged during the iteration process.
[0090] In this embodiment, the capacity parameters are constrained during training through monotonic projection and interactive sparse regularization. After each round of gradient update, monotonic projection is performed sequentially on all fuzzy metric capacity parameters. The monotonic projection operation iterates according to the inclusion relationship of field subsets. When it is detected that the capacity value of a subset is less than the capacity value of any of its proper subsets, the capacity value is adjusted to a value that is not less than the capacity values of all proper subsets, thereby generating a set of capacity parameters that satisfy the monotonicity constraint.
[0091] Building upon this, an interactive sparse regularization term is introduced into the training objective function to impose sparsity constraints on all high-order subset capacity parameters containing two or more fields. This sparsity constraint controls the update intensity with an additional regularization coefficient, causing some capacity parameters that lack practical distinguishing value to gradually approach zero. After joint processing by monotonic projection and interactive sparse regularization, the updated set of capacity parameters is obtained and used in subsequent Choquet integral calculations.
[0092] In this embodiment, the steps for generating the Choquet similarity matrix obtained through training and optimization include: calculating the Choquet similarity for each pair of all ledger records based on the updated fuzzy measure capacity parameter to form a complete similarity matrix. For each ledger record, selecting the first k candidate records from the corresponding row of the similarity matrix in descending order of similarity value to obtain preliminary matching results.
[0093] Based on the statistical results of the similarity distribution of positive sample pairs in the training samples, a lower similarity threshold is set, and candidate records below the threshold are removed from the preliminary matching results to obtain the final soft matching results.
[0094] After the above filtering is completed, for each pair of matching records that remain, their Choquet similarity is mapped to the [0,1] interval according to a preset normalization function to generate a corresponding confidence score. The confidence score is output together with the soft matching result to support subsequent difference determination and result analysis.
[0095] In this embodiment, the setting of the similarity lower limit threshold includes:
[0096] In the training data, the Choquet similarity distribution of positive sample matching pairs is statistically analyzed, and the value closest to the upper quantile is selected as the lower threshold. For example, the 95th percentile of the positive sample similarity distribution is taken as a fixed threshold. Therefore, candidate pairs below this threshold are directly eliminated during the screening process.
[0097] The calculation of the confidence score includes:
[0098] For each record's candidate matches, its Choquet similarity is proportionally normalized using Softmax within the candidate set for that record, ensuring all confidence scores fall within the [0,1] range. The closer the normalized score is to 1, the higher the relative confidence of the match.
[0099] In this embodiment, the difference determination and result output are based on the soft matching result and its corresponding confidence score.
[0100] When multiple soft-match results with high similarity appear for the same asset in different ledgers, it is determined to be a duplicate entry, and the record number to be merged or written off is output in the results.
[0101] When there is a discrepancy in the accounting subject field of the matching pair, it is determined that the subject is incorrectly registered, and the correspondence between the original subject and the matching subject is output in the result;
[0102] When the difference between the matching pairs in the amount field exceeds the preset ratio threshold, it is judged as an amount abnormality, and the difference amount and the corresponding confidence score are output in the result;
[0103] If there is a discrepancy in the location or responsible person field of the matching pair, it is determined that the location is inconsistent, and the location information and confidence score of the two records are output in the result.
[0104] The aforementioned difference types and their corresponding matching pairs are compiled into a structured results file, along with a confidence score for each difference determination. This file can be used by finance personnel during asset verification to generate processing suggestions or for further review.
[0105] Example 1:
[0106] To verify the feasibility of this invention in practice, it was applied to a fixed asset inventory task of a large enterprise. This organization manages approximately 12,000 assets distributed across multiple departments, including administration, finance, logistics, and research, with different accounting systems used concurrently. Due to historical reasons, some assets suffer from issues such as duplicate registrations, incorrect accounting classifications, inconsistent amount entries, and location discrepancies across different ledgers. Traditional manual inventory methods are insufficient for large-scale comparisons in a short time. Therefore, typical sample data was selected for automated inventory verification using the method of this invention.
[0107] During the validation phase, 10 representative asset records were selected as samples from different accounting systems. Each record included fields such as asset number, name, model and specifications, original value, net value, department, location, responsible person, accounting subject, and purchase date. Table 1 shows some of the original data from the experimental sample.
[0108] Table 1 Sample Data of Asset Ledger
[0109]
[0110] First, standardization and vectorization are performed on the various fields in Table 1:
[0111] Numerical fields such as original value and net value are range-scaled, mapping values to the [0,1] range according to preset upper and lower limits; discrete fields such as department, location, responsible person, and accounting subject are mapped into continuous vectors using learnable embedding; textual fields such as asset name and model specification are generated into fixed-length vectors using sub-word decomposition encoding and average pooling; and date fields are represented by timestamp differences to reflect the temporal characteristics of asset acquisition. This results in a consistent ledger field vector, providing a unified input for fuzzy measurement and Schulke integral calculation.
[0112] A fuzzy measure capacity parameter is defined for all field subsets and updated end-to-end using gradient descent. After each training round, a monotonic projection operation is performed to ensure that the capacity of any field subset is not less than the capacity of its proper subset, thus maintaining monotonicity. At the same time, interactive sparse regularization is added to the training objective function to impose sparse constraints on the capacity parameter of higher-order field combinations, so that the weights of irrelevant field combinations gradually decay.
[0113] The updated capacity parameters are used, and the Shukai integral is made differentiable through Lovaz extension to obtain the Shukai similarity between any two asset records. The system uses this similarity as the sole metric to construct a contrastive learning objective function: records from multiple sources for the same asset are considered positive sample pairs, while records from different assets with similar field features are considered hard negative sample pairs. By minimizing the objective function, the model achieves co-optimization of the capacity and embedding parameters through multiple iterations, ultimately forming a convergent similarity matrix.
[0114] Based on the trained and optimized similarity matrix, a Top-k screening is performed on each asset record, selecting the highest similarity records as candidates. The system calculates the 95th percentile in the positive sample similarity distribution as a lower similarity threshold, and candidate pairs below the threshold are eliminated to generate the final soft matching result. For each pair of retained matching records, the system normalizes the Shukai similarity proportionally to obtain a confidence score in the [0,1] interval.
[0115] During the difference determination phase, the system compares each key field item by item based on the soft matching results:
[0116] When the same asset number appears in different ledgers and the similarity between the two assets exceeds 0.9, it is marked as a duplicate entry.
[0117] When the accounting subject fields are different but the names and models are the same and the similarity exceeds 0.85, it is marked as an incorrect subject;
[0118] When the difference in amount exceeds 3%, it is marked as an abnormal amount;
[0119] If the location field is inconsistent but other fields are highly similar, it is marked as a location mismatch.
[0120] Finally, a list of discrepancies is generated and confidence scores are output.
[0121] Table 2. Results of Difference Determination
[0122]
[0123] As shown in Table 2, duplicate entries (such as A001-A002) were successfully identified and given a high confidence level of 0.95; for inconsistent accounting subjects (such as A003-A004, A005-A006), it can be determined that the subject was incorrectly recorded; for location field mismatch (such as A007-A008), the location difference is accurately output; for significant differences in amount (such as A009-A010), it is determined that the amount is abnormal; all differences are output with corresponding confidence levels, which makes it convenient for financial personnel to sort and process according to confidence level.
[0124] Compared to traditional asset matching methods based on Euclidean distance or cosine similarity, this invention achieves nonlinear interaction modeling between fields through fuzzy measures and Schukian integrals, enabling accurate capture of complex similarity relationships under multi-field combined features. On the same sample set, the accuracy rate of traditional cosine similarity matching is 83%, while the accuracy rate of the method of this invention reaches 95%, reducing the false positive rate by approximately 40%.
[0125] This embodiment demonstrates that the method of the present invention can be directly applied to actual accounting data, enabling automatic identification of duplicate entries, incorrect accounts, abnormal amounts, and location discrepancies, thus possessing significant practical value.
[0126] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for intelligent asset ledger verification based on learnable fuzzy measures, characterized in that, Includes the following steps: Collect asset ledger field information, perform standardization processing and vectorization encoding on the fields, and generate a unified-dimensional ledger field vector representation; Construct learnable fuzzy measures and define corresponding fuzzy measure capacity parameters for any subset of the ledger field vectors; Based on the fuzzy measure, the differentiability of the Choquet integral is achieved using the Lovatz extension, and the Choquet similarity of the candidate asset pairs is obtained. Using the Choquet similarity as the sole contrast measure, a contrastive learning objective function is constructed, which pairs positive samples from multiple sources of the same asset and pairs negative samples from records from different assets but with similar field features. In the contrastive learning process, monotonic projection constraints are applied to the fuzzy measure, and interactive sparse regularization is applied to obtain the Choquet similarity matrix obtained through training optimization. Based on the Choquet similarity matrix obtained through training and optimization, a Top-k matching and threshold filtering strategy is adopted to output the soft matching results between different ledger records or the soft matching results between ledger records and physical assets, along with the corresponding confidence levels. Based on the soft matching results, determine whether there are discrepancies in the asset accounting, such as duplicate entries, incorrect account settings, abnormal amounts, or inconsistent locations, and generate processing suggestions; The calculation of Choquet integrals using the Lovarz extension includes: Sort the ledger field vectors in descending order of their values; Based on the arrangement result, the differences between adjacent field vectors are calculated sequentially, and the differences between adjacent field vectors are weighted and summed with the corresponding fuzzy measure capacity. The capacity function is made continuous and differentiable using the Lovarz extension function. Obtain the Choquet similarity of candidate asset pairs.
2. The intelligent asset ledger verification method based on learnable fuzzy measure according to claim 1, characterized in that, The asset ledger fields include asset number, name, model and specifications, original value, net value, department, location, responsible person, accounting subject, and date information; The standardized and vectorized representation includes: Numeric fields are standardized using a range scaling method; Learnable embedding vectors are introduced for categorical fields to map discrete values of department, location, responsible person, and accounting subject to a continuous vector space; Text fields are encoded through word decomposition; Date fields are standardized using timestamp differentiation. Generate a vector representation of the ledger fields.
3. The intelligent asset ledger verification method based on learnable fuzzy measure according to claim 1, characterized in that, The construction of learnable fuzzy measures includes: Define a fuzzy measure capacity parameter for all possible subsets of the ledger field vector; An end-to-end training method based on gradient descent is used to update the fuzzy measure capacity parameter; A set of fuzzy measure capacity parameters is obtained.
4. The intelligent asset ledger verification method based on learnable fuzzy measure according to claim 1, characterized in that, The construction of the contrastive learning objective function includes: Multiple ledger records from the same asset are grouped into positive sample pairs, and the positive sample loss term is calculated based on Choquet similarity. Records from different assets but with similar field features are grouped into hard negative sample pairs, and the negative sample loss term is calculated based on Choquet similarity. The positive sample loss term and the negative sample loss term are weighted and combined to form a contrastive learning objective function with Choquet similarity as the sole similarity metric. During training, the contrastive learning objective function is minimized to update the embedding model parameters.
5. The intelligent asset ledger verification method based on learnable fuzzy measure according to claim 1, characterized in that, The steps for generating the Choquet similarity matrix obtained through training and optimization include: Choquet similarity of candidate asset pairs is obtained by performing Choquet integral calculation based on the vector representation of ledger fields and the fuzzy measure capacity parameter. After each update, a monotonic projection operation is performed on the capacity parameter to generate a capacity parameter that satisfies the monotonicity constraint under the subset inclusion relation. An interactive sparse regularization term is introduced into the training objective of the capacity function to impose sparsity constraints on the capacity parameters of higher-order field subsets. The capacity parameter, after being processed by monotonic projection and interactive sparse regularization, is applied to the subsequent iterative calculation of Choquet similarity, and finally the trained and optimized similarity matrix is obtained.
6. The intelligent asset ledger verification method based on learnable fuzzy measure according to claim 1, characterized in that, The Top-k and threshold filtering based on the Choquet similarity matrix includes: For each asset ledger record, calculate the Choquet similarity with all candidate records to obtain a similarity matrix; From the similarity matrix, the top k candidate records are selected as preliminary matching results in order of their similarity values. A similarity threshold is further set for the preliminary matching results, and candidate records below the threshold are removed to obtain the final soft matching results; While outputting the soft matching results, a confidence score based on Choquet similarity is attached to each pair of matching records.
7. The intelligent asset ledger verification method based on learnable fuzzy measure according to claim 6, characterized in that, The setting of the similarity lower limit threshold includes: In the training samples, the Choquet similarity distribution of positive sample matching pairs is statistically analyzed, and the lower limit threshold of similarity is determined according to the preset quantile of the distribution. The calculation of the confidence score includes: The Choquet similarity of each pair of matching records is mapped to the [0,1] interval using a normalization function to obtain the corresponding confidence score.
8. The intelligent asset ledger verification method based on learnable fuzzy measure according to claim 1, characterized in that, The process of determining asset discrepancies and generating processing suggestions based on soft matching results includes: When multiple highly similar records of the same asset appear in different ledgers, it is determined to be a duplicate entry, and processing suggestions for merging or write-off are generated. When there is an inconsistency in the accounting subject field of the matching records, it is determined that the subject is incorrectly set, and a processing suggestion for adjusting the subject is generated; When the difference in the amount field of the matched records exceeds the preset tolerance threshold, it is determined to be an amount abnormal, and processing suggestions for supplementation or correction are generated; When the matching records do not match in the location or responsible person field, it is determined that the location is inconsistent, and processing suggestions for transfer or responsibility tracing are generated.