An artificial intelligence feature fusion-based case-related property valuation method

Abstract

The application discloses a kind of based on artificial intelligence feature fusion's case-involved property valuation method, comprising the following steps: step one: the corresponding multi-source original data of the case-involved property to be valued is collected, and unified processing is carried out;Step two: according to the case-involved property valuation element set, establish the standard identification image of property;Step three: calculate real-time market quotation benchmark price and construct real-time market quotation benchmark matrix;Step four: calculate initial value decay ratio, and depreciation correction is carried out to real-time market quotation benchmark matrix, obtain depreciation correction price;Step five: the standard identification image of property, real-time market quotation benchmark matrix and corresponding depreciation correction price are input to the improved TabNet model and carry out case-involved property valuation, obtain initial valuation interval;Step six: according to depreciation correction price, calibrate initial valuation interval and generate traceable valuation interval.The application realizes case-involved property intelligent valuation and traceability by improved TabNet model.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for valuing seized property based on artificial intelligence feature fusion. Background Technology

[0002] In judicial and law enforcement practice, the valuation of seized assets is crucial for case handling, determining the amount involved, and disposing of the assets. Traditional asset valuation methods often rely on manual experience, single market price references, or fixed depreciation schedules, making it difficult to comprehensively quantify the actual condition, health, and market value of the assets from multiple sources. This approach is not only time-consuming and highly subjective, but also fails to accurately reflect the true market value of the assets, leading to discrepancies between the valuation results and the actual disposable value. This affects the accuracy of determining the amount involved and the fairness of judicial proceedings.

[0003] In existing technologies, even those employing partially data-driven methods, most only consider market prices from a single source or simple depreciation calculations, lacking comprehensive analysis of multi-dimensional characteristics such as brand, model, specifications, service life, depreciation status, and condition. Furthermore, traditional methods struggle to handle raw data from different sources, formats, or time periods, lacking a unified processing mechanism and failing to achieve traceability and verification of valuation results. Therefore, when faced with multi-source information such as judicial disposal prices, secondhand platform prices, industry quotations, e-commerce new product prices, and auction prices, existing technologies cannot effectively integrate price data from different sources, nor can they dynamically adjust based on the property's usage status and depreciation information, resulting in insufficient accuracy and interpretability of valuation results.

[0004] Therefore, how to provide a valuation method for involved assets based on artificial intelligence feature fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a valuation method for seized assets based on artificial intelligence feature fusion. This invention unifies the processing of multi-source data on seized assets, constructs a standard identification profile of the assets and a real-time market benchmark matrix, combines the years of use, usage intensity, maintenance and repair conditions to perform depreciation correction, and utilizes an improved TabNet model to fuse identity, market conditions and depreciation features to achieve intelligent generation, dynamic calibration and result traceability of valuation ranges.

[0006] According to an embodiment of the present invention, a method for valuing seized property based on artificial intelligence feature fusion includes the following steps: Step 1: Collect multi-source raw data corresponding to the assets to be valued, and unify the data from different sources in the multi-source raw data to form a set of valuation elements for the assets to be valued. Step Two: Establish a standard identification profile of the assets based on the set of valuation elements of the assets involved in the case; Step 3: Calculate the real-time market benchmark price based on the financial standard identification profile, and construct the real-time market benchmark matrix; Step 4: Calculate the annual consumption ratio, usage intensity ratio, maintenance attenuation, and maintenance deficiency attenuation according to the set of valuation elements of the involved property to obtain the initial value attenuation ratio, and then adjust the depreciation based on the real-time market benchmark matrix to obtain the depreciation adjusted price. Step 5: Input the standard identification profile of the property, the real-time market benchmark matrix, and the corresponding depreciation adjustment price into the improved TabNet model to value the property involved in the case and obtain an initial valuation range. The improved TabNet model includes a feature input embedding layer, a feature transformer, and an attention transformer. Step 6: Calibrate the initial valuation range based on the depreciation adjustment price and generate a traceable valuation range.

[0007] Optionally, the multi-source raw data includes property identity data, market information data, usage and depreciation data, condition data, circulation and transaction data, and source traceability data.

[0008] Optionally, the process of unifying the data from different sources in the multi-source raw data to form a set of valuation elements for the involved property specifically includes: Generate a unique data number for each piece of data in the multi-source raw data, convert synonymous fields from different sources into unified field names, unify prices to the same currency unit, unify time to the same date format, and unify usage to the corresponding property category's unit of measurement. The system then performs outlier removal, duplicate data removal, and missing value imputation to form a set of valuation elements for the assets involved in the case.

[0009] Optionally, step two specifically includes: Retrieve brand name, model name, specifications, nameplate content, and purchase receipt content from the set of valuation elements of the involved property; Perform character cleansing on the brand name to convert it into a standard brand name; The model name is split into letter segment, number segment and version segment, and the splitting results are compared with the standard model codes in the preset model library item by item to filter candidate models with the same model body; The specifications are arranged by parameter name and parameter value. The standard specifications of each candidate model are compared with the actual specifications of the property involved in the case, and candidate models with more than the preset value of parameter conflicts are deleted. For the multiple candidate models that are retained, cross-verification is carried out in the following order: nameplate field first, voucher field second, and appearance image field third, to determine the unique standard model. An identification code is generated based on the standard brand, standard model, specification level, production batch, and asset category, and the identification code is used as the primary index. The circulation status is determined based on the current availability and transaction status; The asset standard identification profile is composed of the main index, brand value level, and circulation status.

[0010] Optionally, step three specifically includes: Extract market records corresponding to standard brand, standard model and specification level from the standard identification profile of the property, and classify the records with the same model into the first market layer; Records from the same series with specifications that are within the allowable range are classified into the second market level; Records of the same category and with the same main functions are assigned to the third market data layer; The judicial disposal transaction records, second-hand platform transaction records, industry quotation records, e-commerce new machine price records, and auction transaction records in each market level are respectively formed into corresponding price sub-matrices; Calculate the median transaction price for each row in the price submatrix; The median transaction prices of each source are summed and divided by the number of sources with valid data to obtain the real-time market benchmark price; and the real-time market benchmark prices of each source are used to form the real-time market benchmark matrix.

[0011] Optionally, step four specifically involves: The actual service life is calculated based on the interval between the valuation benchmark date and the commissioning date. The standard economic life is read from the economic life table corresponding to the property category. The actual service life is divided by the standard economic life to obtain the service life consumption ratio. The usage intensity ratio is calculated based on the ratio of cumulative usage to the current model's rated usage. The maintenance activities in the maintenance records are divided into routine maintenance, general maintenance and major maintenance. Routine maintenance does not increase depreciation. Each general maintenance increases the preset maintenance attenuation amount. Major maintenance increases the maintenance attenuation amount based on the degree of damage to the core components. Compare the replaced parts in the parts replacement record with the critical parts catalog; if a new original part is replaced, add a preset maintenance attenuation amount; if a non-original part or a used part is replaced, leave it unchanged. The difference between the actual maintenance interval and the specified maintenance interval in the maintenance record is calculated as the maintenance loss attenuation amount; The initial value decay ratio is obtained by adding the proportion of annual consumption, the proportion of usage intensity, the amount of maintenance decay, and the amount of maintenance deficiency decay according to preset weights. The depreciation correction factor is obtained by subtracting the initial value decay ratio from 1, and the depreciation correction price is generated by multiplying the real-time market benchmark prices in the real-time market benchmark matrix by the depreciation correction factor.

[0012] Optionally, the improved TabNet model is specifically as follows: The identity coding features in the standard identification profile of assets, the market features in the real-time market benchmark matrix, and the corresponding depreciation features are concatenated in a fixed column order to form the original fused feature vector; The original fused feature vector is input into the feature input embedding layer, and each discrete code is converted into a dense feature vector of fixed length. The dense discrete features are input into the first feature transformer of the improved TabNet model. The first feature transformer performs linear transformation through a fully connected layer, stabilizes the feature distribution through a batch normalization layer, and divides the input features into content branches and gated branches through a gated linear unit. The content branch undergoes a nonlinear transformation, and the gated branch is activated by Sigmoid and then multiplied position by position with the content branch to obtain the first-stage transformation features. The first stage transformation features are input into the attention transformer, and the selection scores of each input feature are generated through the fully connected layer. The features used in the previous decision step are combined with the feature usage record to remove features that have been repeatedly selected. A sparse activation function is used to compress the selection scores that are less than a preset score threshold to 0, forming a sparse feature mask for the current decision step. The sparse feature mask is multiplied element-wise with the original fused feature vector and then input into the feature transformer again. The estimated decision features for the current decision step are generated through fully connected transformation, batch normalization and gated linear units. The remaining features are then passed to the next decision step to continue mask selection and feature transformation. The valuation decision features generated from multiple decision steps are transformed into the same dimension through a linear layer and then added position by position to form a multi-stage fusion feature; The multi-stage fusion features are input into the fully connected layer respectively, and the linear combination of the fusion features is calculated to obtain the predicted center price; Subtracting the preset initial deviation value from the prediction center price yields the initial valuation lower limit, and adding the preset initial deviation value to the prediction center price yields the initial valuation upper limit. The initial valuation range is formed by using the lower limit and the upper limit of the initial valuation.

[0013] Optionally, step six specifically includes: Compare the initial valuation range with the depreciation adjustment prices. If the initial valuation range covers all depreciation adjustment prices, retain the initial valuation range. When the lower limit of the initial valuation range is higher than all depreciation adjustment prices, the lower limit of the initial valuation range will be adjusted to the average of all depreciation adjustment prices. When the upper limit of the initial valuation range is lower than all depreciation adjustment prices, the upper limit of the initial valuation range will be adjusted to the average of all depreciation adjustment prices and the forecast center price. After adjusting the interval boundaries, the target valuation lower limit, target valuation median, and target valuation upper limit are generated, forming a traceable valuation interval.

[0014] The beneficial effects of this invention are: This invention collects data on the identity, market conditions, depreciation, condition, circulation, and provenance of the assets to be valued. It then standardizes field names, currency units, date formats, and units of measurement for usage across different data sources. This reduces the impact of format differences, semantic conflicts, and data redundancy between multiple data sources on the valuation results, providing a unified and complete set of valuation elements for the assets. By comparing and cross-checking brand names, model names, specifications, nameplate content, and purchase receipts at each level, a standard identification profile of the assets is established, primarily indexed by identification codes. This accurately distinguishes assets of similar brands, models, and specifications, reducing market matching errors caused by inaccurate asset identification.

[0015] By dividing the market data into different layers and processing judicial disposal transaction records, second-hand platform transaction records, industry quotation records, e-commerce new product price records, and auction transaction records separately, and calculating the median of each type of price and the real-time market benchmark price, the interference of single-source price fluctuations and abnormal transaction data on the valuation results can be reduced, allowing the real-time market benchmark matrix to more comprehensively reflect the market price level of the involved property. By calculating the annual consumption ratio, usage intensity ratio, repair attenuation, and maintenance deficiency attenuation to obtain the initial value depreciation ratio, and using this to correct the real-time market benchmark matrix for depreciation, the valuation results can correspond to the actual use and maintenance status of the involved property, avoiding value deviations caused by depreciating only according to a fixed depreciation period. Through the improved TabNet model, sparse selection, multi-decision-step transformation, and multi-stage fusion of identity coding features, market features, and depreciation features, key features that significantly impact the value of the property can be automatically extracted, reducing interference from irrelevant and duplicate features, and generating an initial valuation range that better reflects the actual situation of the property. The initial valuation range is calibrated based on the depreciation adjustment price, resulting in a target valuation lower limit, target valuation median, and target valuation upper limit. This ensures that the final valuation result takes into account both the artificial intelligence prediction result and the actual depreciation adjustment price, providing a reliable reference for the determination of the amount involved and the disposal of assets. Attached Figure Description

[0016] 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: Figure 1 This is an overall flowchart of a method for valuing seized assets based on artificial intelligence feature fusion proposed in this invention; Figure 2 This is a schematic diagram illustrating the steps of constructing a real-time market benchmark matrix for a valuation method of involved assets based on artificial intelligence feature fusion proposed in this invention. Detailed Implementation

[0017] 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.

[0018] refer to Figures 1-2 A method for valuing seized assets based on artificial intelligence feature fusion includes the following steps: Step 1: Collect multi-source raw data corresponding to the assets to be valued, and unify the data from different sources in the multi-source raw data to form a set of valuation elements for the assets to be valued. Step Two: Establish a standard identification profile of the assets based on the set of valuation elements of the assets involved in the case; Step 3: Calculate the real-time market benchmark price based on the financial standard identification profile, and construct the real-time market benchmark matrix; Step 4: Calculate the annual consumption ratio, usage intensity ratio, maintenance attenuation, and maintenance deficiency attenuation according to the set of valuation elements of the involved property to obtain the initial value attenuation ratio, and then adjust the depreciation based on the real-time market benchmark matrix to obtain the depreciation adjusted price. Step 5: Input the standard identification profile of the property, the real-time market benchmark matrix, and the corresponding depreciation adjustment price into the improved TabNet model to value the property involved in the case and obtain an initial valuation range. The improved TabNet model includes a feature input embedding layer, a feature transformer, and an attention transformer. Step 6: Calibrate the initial valuation range based on the depreciation adjustment price and generate a traceable valuation range.

[0019] In this embodiment, step one specifically includes: Collect multi-source raw data corresponding to the assets involved in the case to be valued. The multi-source raw data includes asset identity data, market information data, usage and depreciation data, condition data, circulation and transaction data, and source tracing data. Among them, the property identification data includes property category, brand name, model name, specifications, manufacturing date, purchase date, original purchase price, and purchase voucher; The market data includes judicial disposal transaction prices, second-hand platform transaction prices, industry quotations, e-commerce new machine prices, auction transaction prices, and the corresponding transaction time, transaction area, and transaction status. The depreciation data includes actual usage time, cumulative usage, repair records, component replacement records, and maintenance records; The condition data includes appearance inspection results, structural inspection results, functional inspection results, and accessory inventory results; The circulation and transaction data includes the number of similar assets issued, the number of transactions, the transaction cycle, and the regional circulation range. The source traceability data includes the source platform, collection time, source record number, and data provider; Generate a unique data number for each data entry, convert synonymous fields from different sources into uniform field names, unify prices to the same currency unit, unify time to the same date format, and unify usage to the corresponding property category's unit of measurement; Outlier removal, duplicate data deletion, and missing value imputation are performed to form a set of valuation elements for the assets involved in the case. The missing value imputation process involves taking the average of the neighboring values ​​and imputing them.

[0020] In the specific implementation process, the first step is to establish a data collection task around the assets to be valued, and assign corresponding case numbers and temporary asset numbers to the current assets involved in the case. These case numbers and temporary asset numbers serve as the association identifiers for multi-source raw data. Asset identity data is collected through asset registration records, purchase vouchers, nameplate content, and manually entered information; market data is collected through judicial disposal records, second-hand transaction records, industry quotation records, e-commerce sales records, and auction transaction records; depreciation data is collected through asset usage ledgers, repair receipts, parts replacement vouchers, and maintenance records; condition data is collected through on-site inspection records, functional test records, and accessory inventory records; and circulation and transaction data is collected through the release and transaction records of similar assets. Simultaneously, the source platform, collection time, source record number, and data provider for each data point are recorded, forming multi-source raw data associated with the current assets involved in the case.

[0021] After completing the collection of multi-source raw data, a unique data number is generated for each data entry using a combination of "case number, temporary asset number, data type number, and sequence number," ensuring that data from different sources, data types, and collection times are independently identified. A unified field name lookup table is established to convert fields with the same meaning but different names from different sources into unified field names. For example, "transaction amount," "total transaction price," and "transaction price" are uniformly converted to the "transaction price" field, and "production date," "manufacturing time," and "manufacturing date" are uniformly converted to the "manufacturing time" field. For price data, it is converted to a preset currency unit based on the currency type at the time of collection, while retaining both the original price and the converted price. For time data, it is uniformly converted to a date format of year, month, and day. For cumulative usage, it is converted to the corresponding unified unit of measurement based on the asset category, enabling comparisons of usage for similar assets under the same measurement standard.

[0022] Outlier detection is performed on data with unified fields and units. Records with prices less than 0, transaction times earlier than production times, cumulative usage less than 0, or field values ​​exceeding the reasonable range of their respective data are identified as outliers and removed from the current data processing. A unique data ID and removal flag are retained for each outlier. Subsequently, duplicate data is checked. If two data sets have the same source record ID, data provider, collection content, and collection time, they are considered completely duplicate data, and only one is retained. If the data content is identical but the source record IDs are different, the data from different sources is retained to avoid deleting records with independent source traceability relationships.

[0023] For data with missing values, neighboring data is first searched within the same asset category, data field, and under similar collection conditions. Centering on the data to be filled, a preset number of neighboring values ​​are selected based on the similarity of brand name, model name, specifications, or transaction time. These selected neighboring values ​​are summed and divided by the number of neighboring values ​​to obtain the neighboring average. This average is then written into the corresponding missing field, while simultaneously recording the neighboring data IDs and missing value filling markers used in the average calculation. If no matching neighboring values ​​exist, the missing status is retained, and data from other categories or unrelated data is not used for filling. After unique number generation, field unification, unit conversion, outlier removal, duplicate data deletion, and missing value averaging, the retained data is aggregated according to the asset's temporary ID and data type to form a set of valuation elements for the involved assets. This provides a unified, complete, and traceable data foundation for the subsequent establishment of a standardized asset identification profile.

[0024] In this embodiment, step two specifically includes: Retrieve brand name, model name, specifications, nameplate content, and purchase receipt content from the set of valuation elements of the involved property; Perform character cleaning on the brand name, which involves converting the brand abbreviation, foreign language name, and historical name into the standard brand name; The model name is split into letter segment, number segment and version segment, and the splitting results are compared with the standard model codes in the preset model library item by item to filter candidate models with the same model body; The specifications are arranged by parameter name and parameter value. The standard specifications of each candidate model are compared with the actual specifications of the property involved in the case, and candidate models with more than the preset value of parameter conflicts are deleted. For multiple candidate models that are retained, the model characters, factory number and production batch are read from the nameplate image, and the product name and configuration description are read from the purchase voucher. Cross-verification is performed in the order of nameplate fields first, voucher fields second, and appearance image fields third to determine the unique standard model. An identification code is generated based on the standard brand, standard model, specification level, production batch, and asset category, and the identification code is used as the primary index. The specification level is determined based on historical transaction records to identify the brand value level corresponding to this model. The circulation status is determined based on the current availability and transaction status; The asset standard identification profile is composed of the main index, brand value level, and circulation status.

[0025] In the implementation process, a brand name lookup table, a model code library, and a specification parameter lookup table are pre-established. The brand name lookup table records the mapping relationship between standard brand names and their corresponding abbreviations, foreign language names, and historical names. The model code library is stored hierarchically according to brand, model body, version identifier, and specification parameters. The specification parameter lookup table unifies the names and units of measurement for the same parameters from different sources. During character cleaning, spaces, connectors, parentheses, and enterprise nature descriptions are first removed from the brand field, and then the standard name is replaced using the brand name lookup table. When the same brand field can correspond to multiple standard brand names, a second verification is performed by combining the model name and nameplate content.

[0026] Model segmentation uses character type changes and preset separators as boundaries, writing consecutive letters, consecutive numbers, and characters indicating generation or configuration differences into corresponding fields. During candidate model screening, complete consistency between the letter segment and the main number segment is used as a model body consistency condition; the version segment is only used for subsequent subdivision and not as an initial exclusion criterion. When comparing specifications, parameter names and units of measurement are first standardized, then the number of conflicting parameters between candidate models and actual assets is calculated. The preset value can be set based on the total number of parameters; when the number of conflicting parameters exceeds 20% of the total number of parameters, the corresponding candidate model is deleted.

[0027] During cross-verification, the nameplate field has the highest confidence level, the purchase voucher field has the second highest confidence level, and the appearance image field is used for supplementary identification. If the nameplate model characters are completely identical to a candidate model, that candidate model is directly retained; if the nameplate characters are incomplete, the model range is narrowed down based on the factory number and production batch; the product name and configuration description in the purchase voucher are used to further eliminate candidate models with inconsistent specifications or versions; the appearance image is used to compare differences in interface layout, structural dimensions, or component arrangement. After elimination at each level, only one candidate model is retained and determined as the standard model.

[0028] The identification code can be concatenated in the order of property category code, standard brand code, standard model code, specification grade code, and production batch code, with fixed separators between fields to ensure a unique master index for each involved property. The brand value level is determined based on the price distribution in historical transaction records of the same model. The median historical transaction price can be compared with the price range of similar properties to classify them into corresponding levels. The circulation status is determined based on the number of items for sale and the number of transactions within a preset statistical period. A normal circulation status is defined when both the number of items for sale and the number of transactions reach a preset number; a low transaction status is defined when only the number of items for sale exists but the number of transactions is lacking; and a low circulation status is defined when both the number of items for sale and the number of transactions are insufficient. Finally, the master index, brand value level, and circulation status are written into the same profile record and linked to a unique data number in the set of valuation elements for the involved property, forming a standard identification profile for the property that can be directly accessed for subsequent market data matching and valuation calculations.

[0029] In this embodiment, step three specifically includes: Extract market data records that match the standard brand, standard model, and specification level from the standard identification profile of the property, and classify the records with completely identical models into the first market data layer; Records from the same series with specifications that are within the allowable range are classified into the second market level; Records of the same category and with the same main functions are assigned to the third market data layer; The judicial disposal transaction records, second-hand platform transaction records, industry quotation records, e-commerce new machine price records and auction transaction records in each market layer are respectively formed into corresponding price sub-matrices. Each row of the price sub-matrices corresponds to one market record, and each column is written with the price value in sequence. Calculate the median transaction price for each row in the price submatrix; Add the median of each transaction price and divide by the number of source categories with valid data to obtain the real-time benchmark price. If a source does not have valid data, then the current source will not be included in the divisor. If the number of valid sources is less than the preset number, the subsequent valuation range will be expanded and marked as insufficient market data samples. The real-time market benchmark matrix is ​​composed of real-time market benchmark prices from various sources.

[0030] In the specific implementation process, the allowable range for specification differences is set to ±10% of the corresponding standard specification parameters. Identical main functions are determined through preset function category codes. Each market data record is written into a price sub-matrix according to its source category and market data layer, and records with empty prices, invalid transaction status, or exceeding the preset time range are deleted. When calculating the median transaction price, valid prices within the same price sub-matrix are arranged in ascending order; if the number of valid prices is odd, the middle price is taken; if the number of valid prices is even, the average of the two middle prices is taken. The preset number of valid sources is 3 types. When there are fewer than 3 valid sources, an insufficient market data sample marker is written, and the upper and lower boundaries of the subsequent initial valuation interval are extended outwards by 10%. The real-time market data benchmark matrix is ​​arranged in the order of the first, second, and third market data layers, with each row corresponding to one market data layer and each column corresponding to a real-time market data benchmark price for one source category. Missing source positions are marked with a null value.

[0031] In this embodiment, step four specifically includes: The actual service life is calculated based on the interval between the valuation benchmark date and the commissioning date. The standard economic life is read from the economic life table corresponding to the property category. The actual service life is divided by the standard economic life to obtain the service life consumption ratio. The usage intensity ratio is calculated based on the ratio of cumulative usage to the current model's rated usage. When there is no cumulative usage, the cumulative usage intensity is calculated based on the average daily usage time, activation time, and usage frequency. Read the maintenance records and classify maintenance activities into routine maintenance, general maintenance and major maintenance. Routine maintenance does not increase depreciation. Each general maintenance increases the pre-set maintenance attenuation amount. Major maintenance increases the maintenance attenuation amount based on the degree of damage to the core components. Read the component replacement record and compare the replaced components with the critical component catalog; if a new original component is replaced, add a preset maintenance attenuation amount. If non-original or second-hand parts are replaced, the repair attenuation will be retained. Read the maintenance records and calculate the difference between the actual maintenance interval and the specified maintenance interval as the maintenance loss attenuation amount; The initial value decay ratio is obtained by adding the proportion of annual consumption, the proportion of usage intensity, the amount of maintenance decay, and the amount of maintenance deficiency decay according to preset weights.

[0032] The depreciation correction factor is obtained by subtracting the initial value decay ratio from 1, and the depreciation correction price is generated by multiplying the real-time market benchmark prices in the real-time market benchmark matrix by the depreciation correction factor.

[0033] In practice, the actual service life is accurate to the month and converted to years; when the service life consumption ratio and the usage intensity ratio are greater than 1, they are both counted as 1. When the cumulative usage is missing, the average daily usage time, the number of days used per year, and the service life are multiplied to obtain the cumulative usage time, which is then multiplied by the usage frequency coefficient to calculate the cumulative usage intensity. The preset maintenance attenuation amount for general maintenance is set to 0.02; for major maintenance, the amount is increased by 0.04, 0.07, and 0.10 for minor, moderate, and severe damage to core components, respectively; replacing original key components increases the amount by 0.01; non-original or used components do not change the existing maintenance attenuation amount. When the actual maintenance interval does not exceed the specified interval, the maintenance attenuation amount is 0; when it exceeds the specified interval, the difference in intervals is divided by the specified maintenance interval to obtain the maintenance attenuation amount. The preset weights for the service life consumption ratio, usage intensity ratio, maintenance attenuation amount, and maintenance attenuation amount are set to 0.40, 0.25, 0.20, and 0.15, respectively. When the initial value decay rate exceeds 0.90, it is calculated as 0.90, so that the depreciation adjustment factor for property is not less than 0.10.

[0034] In this embodiment, the improved TabNet model is specifically as follows: The identity coding features in the standard identification profile of assets, the market features in the real-time market benchmark matrix, and the corresponding depreciation features are concatenated in a fixed column order to form the original fused feature vector; The original fused feature vector is input into the feature embedding layer, and each discrete code is converted into a dense feature vector of fixed length. Dense discrete features are input into the first feature transformer of TabNet. The first feature transformer performs linear transformation through fully connected layers, stabilizes the feature distribution through batch normalization layers, and then divides the input features into content branches and gated branches through gated linear units. The content branches undergo nonlinear transformation, and the gated branches are multiplied with the content branches position by position after being activated by Sigmoid to obtain the first stage transformed features. The first-stage transformation features are input into the attention transformer, and the selection scores of each input feature are generated through a fully connected layer. The features used in the previous decision step are combined with the record to suppress features that have been repeatedly selected. A sparse activation function is used to compress low selection scores to 0, forming a sparse feature mask for the current decision step. The sparse feature mask is multiplied element-wise with the original fused feature vector, retaining only the market conditions, depreciation, condition, and identity features selected in the current decision step.

[0035] The filtered features are input into the feature transformer again, which generates the estimated decision features for the current decision step through fully connected transformation, batch normalization and gated linear units, and the remaining features are passed to the next decision step to continue to perform mask selection and feature transformation. The valuation decision features generated from multiple decision steps are transformed into the same dimension through a linear layer, and then added position by position to form a multi-stage fusion feature. The multi-stage fusion features are input into the fully connected layer, and the linear combination of the fusion features is calculated to obtain the predicted center price. The initial valuation lower limit is obtained by subtracting the preset initial deviation value from the forecast center price, and the initial valuation upper limit is obtained by adding the preset initial deviation value to the forecast center price, thus forming the initial valuation range.

[0036] In the specific implementation, the embedding dimension of the discrete encoding is set to 16, the output dimension of the fully connected layer of the first feature transformer is set to 64, and the gated linear unit divides the output equally into content branches and gated branches. The attention transformer sets four decision steps, uses a sparse activation function to generate a feature selection mask, and sets features with selection scores below 0.05 to 0; features selected in the previous decision step have their subsequent selection scores reduced by using records to reduce repeated extraction. Each decision step outputs 32-dimensional valuation decision features, which are uniformly converted to 32 dimensions by the linear layer and then added position by position. During training, historical valuation results of involved assets are used as labels, and the parameters of the embedding layer, feature transformer, and attention transformer are updated through backpropagation. The preset initial bias value is the median of the absolute values ​​of the prediction errors of the training samples, which is used to determine the boundaries of the initial valuation interval.

[0037] In this embodiment, step six specifically includes: Read the initial valuation range and compare it with each depreciation adjustment price. If the initial valuation range covers all depreciation adjustment prices, retain the initial valuation range. When the lower limit of the initial valuation range is higher than all depreciation adjustment prices, the lower limit of the initial valuation range will be adjusted to the average of all depreciation adjustment prices. When the upper limit of the initial valuation range is lower than all depreciation adjustment prices, the upper limit of the initial valuation range will be adjusted to the average of all depreciation adjustment prices and the forecast center price. When there is insufficient market data sample, the distance between the lower and upper valuation limits is increased according to the preset expansion ratio. After adjusting the interval boundaries, the lower limit of the target valuation, the median of the target valuation, and the upper limit of the target valuation are generated. Extract the source record number and status record number of the data involved in the depreciation adjustment price calculation, generate a valuation traceability chain according to the data reading and processing order, and bind the target valuation range to the valuation traceability chain to form a traceable valuation range.

[0038] In the specific implementation process, all depreciation adjustment prices are first sorted in ascending order, and their average is calculated. The initial valuation range covers all depreciation adjustment prices, meaning that the lower limit of the initial valuation is not higher than the minimum depreciation adjustment price and the upper limit of the initial valuation is not lower than the maximum depreciation adjustment price. When adjusting the upper limit of the valuation, the average of all depreciation adjustment prices is first calculated, and then averaged with the predicted center price. The target valuation median is the average of the adjusted target valuation lower limit and the target valuation upper limit. When the market data sample is insufficient, the upper and lower boundary distances are increased by 10% with the current interval median as the center. The valuation traceability chain adopts the following sequence: "Source record number - Market data layer - Real-time market benchmark price - Depreciation adjustment coefficient - Depreciation adjustment price - Interval adjustment result", and is synchronously associated with the records number of annual consumption, usage intensity, and maintenance status, so that each target interval boundary can be traced back to the corresponding original data and processing node.

[0039] Example 1: To verify the feasibility of this invention in practice, it was applied to a case-related property management center for centralized valuation of case-related property. The center received case-related property from different case-handling departments, including industrial equipment, electronic equipment, photographic equipment, office equipment, and professional testing instruments. Due to inconsistent data formats submitted by different departments, the brand name, model name, and specifications of the same property were sometimes abbreviated, with foreign names and historical names mixed together. Market prices were also scattered across different sources such as judicial disposal, secondhand transactions, industry quotations, e-commerce sales, and auctions. Traditional valuation methods typically involve staff manually searching for several reference prices and then applying depreciation based on the years of use. This method is easily affected by the search scope, manual judgment, and a single price source, especially when the property has undergone repairs, component replacements, or lack of maintenance; fixed-term depreciation cannot reflect the actual value of the property. To verify the application effect of this invention, 36 case-related properties with complete provenance records and subsequent verification results were selected as test objects at the case-related property management center. The entire testing process lasted 20 working days, with an industrial-grade 3D scanning device selected as the specific example.

[0040] The equipment to be valued is registered as a professional testing equipment. The brand name is recorded in different documents as a brand abbreviation, English name, and company historical name. The model name is recorded as "X7PRO," "X7-Pro," and "X7 Pro Industrial Edition." The model name recorded on the nameplate is "X7-PRO." The batch number is B3. The product name recorded on the purchase voucher is "Industrial 3D Scanning System," and the original purchase price is 168,000 yuan. The system collected original data from multiple sources, including the case-related property registration form, nameplate content, purchase vouchers, judicial disposal records, second-hand transaction records, industry quotation records, e-commerce sales records, auction transaction records, equipment usage ledgers, maintenance documents, parts replacement records, and maintenance records, totaling 214 original records. After field unification, "Transaction Amount," "Total Transaction Price," and "Transaction Price" were unified into a single "Transaction Price" field; different forms of model names were unified into a single "Model Name" field; all prices were unified to RMB; the time field was unified to the same date format; and the cumulative equipment usage was unified to hours. During the processing, 18 duplicate records, 7 logically abnormal records, and 11 missing records were found. Duplicate records were deleted. Records with a transaction price less than 0, a transaction time earlier than the equipment's manufacturing date, or a negative cumulative usage were removed as outliers. Missing records were filled by averaging the values ​​of neighboring records under similar transaction conditions for the same brand, model, and specifications. After processing, a set of valuation elements for the involved property containing 189 valid data entries was formed. Each data entry is bound to a unique data number, source platform, collection time, source record number, and data provider.

[0041] The system retrieves brand name, model name, specifications, nameplate content, and purchase voucher content from the set of valuation elements for the involved property. It then uses a pre-set brand name lookup table to convert the brand abbreviation, English name, and historical name into a standard brand name. The model name is split into letter, number, and version segments and compared item by item with the standard model codes in the pre-set model library, resulting in four candidate models with the same basic structure. Next, specifications such as scanning accuracy, single scan range, working distance, interface type, and supporting software version are arranged by parameter name and value, and compared item by item with the standard specifications of the four candidate models. Two candidate models were deleted because the number of parameter conflicts exceeded the pre-set value. For the remaining two candidate models, the system continues to retrieve the model characters, factory number, and production batch from the nameplate, and the product name and configuration description from the purchase voucher. Cross-verification is performed in the order of nameplate fields first, voucher fields second, and appearance image fields third, ultimately determining the unique standard model as the X7-Pro Industrial Edition. The system generates an identification code based on the standard brand, standard model, specification level, production batch, and asset category, and determines the brand value level by combining historical transaction records. It also determines the circulation status based on the current sales records and transaction records, thus forming a corresponding asset standard identification profile.

[0042] During market data processing, the system categorizes market records with identical standard brands, models, and specifications into the first market layer; records from the same series with specification differences within permissible limits into the second market layer; and records of the same category with identical main functions into the third market layer. The first market layer yielded 48 valid market records, the second layer 37, and the third layer 29. Each market layer is further divided into price sub-matrices based on judicial disposal transaction records, secondhand platform transaction records, industry quotation records, e-commerce new machine price records, and auction transaction records. After sorting prices from each source in ascending order, the median is calculated. The median price for judicial disposal transactions in the first market layer is 95,800 yuan; the median price for secondhand platform transactions is 103,200 yuan; the median price for industry quotations is 110,000 yuan; the median price for new machines on e-commerce platforms is 166,800 yuan; and the median price for auction transactions is 94,500 yuan. Adding the medians from the five valid sources and dividing by 5 yields the real-time benchmark price for the first market layer: 114,060 yuan. Using the same method, the real-time benchmark price for the second market layer is 108,780 yuan, and the real-time benchmark price for the third market layer is 101,800 yuan. The real-time benchmark matrix is ​​composed of the above three real-time benchmark prices.

[0043] The equipment usage log shows that the actual service life of the equipment is 3.5 years, while the standard economic life corresponding to the asset category is 8 years, therefore the service life ratio is 0.4375. The cumulative usage of the equipment is 7860 hours, and the current model's rated usage is 20000 hours, resulting in a usage intensity ratio of 0.393. The maintenance record includes 2 general maintenance and 1 major maintenance. The general maintenance resulted in a maintenance attenuation of 0.04, while the major maintenance involved moderate damage to core components, resulting in a maintenance attenuation of 0.07. The component replacement record shows that one original key component was replaced, increasing the preset maintenance attenuation by 0.01, thus the total maintenance attenuation is 0.12. The specified maintenance interval is 6 months, while the actual average maintenance interval is 7.2 months. The difference between the actual and specified maintenance intervals yields a maintenance deficiency attenuation of 0.20. The initial value depreciation ratio is calculated based on preset weights of 0.40, 0.25, 0.20, and 0.15 for the proportion of use over years, the proportion of use intensity, the amount of depreciation due to maintenance, and the amount of depreciation due to lack of maintenance. The depreciation correction factor is 0.67275. Multiplying the real-time market benchmark matrix of 114,060 yuan, 108,780 yuan, and 101,800 yuan by 0.67275 respectively yields three depreciation correction prices of 76,733.87 yuan, 73,181.75 yuan, and 68,485.95 yuan.

[0044] The system concatenates the identity coding features from the asset standard identification profile, the market data features from the real-time market data benchmark matrix, and the depreciation features in a fixed column order to form an original fused feature vector. Discrete codes are converted into a 16-dimensional dense feature vector via a feature input embedding layer. Continuous market data features and depreciation features are batch normalized and then concatenated with the dense features. The first feature transformer completes the first stage of feature transformation through a fully connected layer, a batch normalization layer, and a gated linear unit. The attention transformer generates sparse feature masks sequentially through four decision steps, compressing features with selection scores below 0.05 to 0 and reducing the subsequent selection scores of features already used in the previous decision step. The 32-dimensional valuation decision features generated in each decision step are transformed by a linear layer and added position by position to form a multi-stage fused feature. After the multi-stage fusion feature is input into the fully connected layer, the predicted center price is 73,500 yuan, and the median absolute value of the prediction error of the training samples is 6,200 yuan. Therefore, the lower limit of the initial valuation is 67,300 yuan, and the upper limit of the initial valuation is 79,700 yuan, forming an initial valuation range of 67,300 yuan to 79,700 yuan.

[0045] The initial valuation range is compared with three depreciation adjustment prices. The lower limit of the initial valuation is lower than the minimum depreciation adjustment price of 68,485.95 yuan, and the upper limit is higher than the maximum depreciation adjustment price of 76,733.87 yuan. Therefore, the initial valuation range can cover all depreciation adjustment prices, and this initial valuation range is retained. Since the number of valid sources is 5 categories, there is no insufficient market data sample flag, and there is no need to expand the range boundary. The lower limit of the target valuation is set at 67,300 yuan, the upper limit of the target valuation is set at 79,700 yuan, and the median of the target valuation is the average of the upper and lower limits, which is 73,500 yuan. The system establishes a valuation traceability chain according to the processing order of source record number, market data layer, real-time market data benchmark price, depreciation adjustment coefficient, depreciation adjustment price, and range adjustment result, and associates it with the status record number corresponding to the actual years of use, cumulative usage, maintenance records, and repair records, ultimately forming a traceable valuation range. The device was subsequently independently reviewed by three individuals with relevant valuation experience. The average value of the reviewed price was 74,800 yuan. The deviation between the median value of the target valuation output by this invention and the reviewed price was 1,300 yuan, with a relative deviation of 1.74%. The valuation result obtained by using the traditional method of manually searching for prices on a single second-hand platform and depreciating them over a fixed period was 62,800 yuan, with a deviation of 12,000 yuan from the reviewed price and a relative deviation of 16.04%.

[0046] To further verify the applicability of this invention to various types of involved assets, a batch valuation was conducted on 36 test assets, including 8 industrial equipment items, 10 electronic equipment items, 6 photographic equipment items, 7 office equipment items, and 5 professional testing instruments. Traditional methods involve staff manually searching for reference prices and calculating depreciation using a fixed depreciation table. The method of this invention processes the valuation according to the set of valuation elements defined in the claims, the standard identification profile of the assets, the real-time market benchmark matrix, the depreciation-adjusted price, the improved TabNet model, and the interval calibration process. The average of the results independently reviewed by three valuation personnel was used as a reference to form the following comparative data.

[0047] Table 1. Comparison of the effects of different valuation methods on involved assets. As shown in Table 1, traditional manual valuation methods use an average of only 1.8 data sources, and staff typically prioritize easily searchable prices from secondhand platforms or e-commerce sites, making it difficult to simultaneously cover information from judicial disposals, industry quotations, and auction transactions. This invention, by unifying the processing of multi-source raw data, increases the average number of valid data records per item from 23 to 167, and improves the brand and model accuracy recognition rate from 86.11% to 97.22%, reducing market matching errors caused by model abbreviations, configuration differences, and confusion between similar models. Through three-layer market segmentation and processing of price sub-matrices from different sources, the abnormal market data recognition rate reaches 93.85%, preventing a single abnormal price from directly affecting the valuation result. The average relative deviation of 36 items decreased from 13.82% to 3.46%, and the number of items with a relative valuation deviation of no more than 5% increased from 11 to 31, indicating that this invention can make the valuation results closer to the verification results. The proportion of the valuation range covering the reviewed price increased from 58.33% to 94.44%, indicating that by using the improved TabNet model for multi-stage feature fusion and combining it with depreciation-adjusted prices to calibrate the initial valuation range, it is possible to take into account differences in market conditions and the actual use status of the property, thereby improving the accuracy, stability, and convenience of the valuation of the property involved in the case, and providing a more reliable valuation reference for the determination of the amount involved in the case and the disposal of the property.

[0048] 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 valuing seized assets based on artificial intelligence feature fusion, characterized in that, Includes the following steps: Step 1: Collect multi-source raw data corresponding to the assets to be valued, and unify the data from different sources in the multi-source raw data to form a set of valuation elements for the assets to be valued. Step Two: Establish a standard identification profile of the assets based on the set of valuation elements of the assets involved in the case; Step 3: Calculate the real-time market benchmark price based on the financial standard identification profile, and construct the real-time market benchmark matrix; Step 4: Calculate the annual consumption ratio, usage intensity ratio, maintenance attenuation, and maintenance deficiency attenuation according to the set of valuation elements of the involved property to obtain the initial value attenuation ratio, and then adjust the depreciation based on the real-time market benchmark matrix to obtain the depreciation adjusted price. Step 5: Input the standard identification profile of the property, the real-time market benchmark matrix, and the corresponding depreciation adjustment price into the improved TabNet model to value the property involved in the case and obtain an initial valuation range. The improved TabNet model includes a feature input embedding layer, a feature transformer, and an attention transformer. Step 6: Calibrate the initial valuation range based on the depreciation adjustment price and generate a traceable valuation range.

2. The method for valuing seized property based on artificial intelligence feature fusion according to claim 1, characterized in that, The multi-source raw data includes property identity data, market information data, usage and depreciation data, condition data, circulation and transaction data, and source traceability data.

3. The method for valuing seized property based on artificial intelligence feature fusion according to claim 1, characterized in that, The process of unifying data from different sources within the multi-source raw data to form a set of valuation elements for the involved assets specifically involves: Generate a unique data number for each piece of data in the multi-source raw data, convert synonymous fields from different sources into unified field names, unify prices to the same currency unit, unify time to the same date format, and unify usage to the corresponding property category's unit of measurement. The system then performs outlier removal, duplicate data removal, and missing value imputation to form a set of valuation elements for the assets involved in the case.

4. The method for valuing seized property based on artificial intelligence feature fusion according to claim 1, characterized in that, Step two specifically involves: Retrieve brand name, model name, specifications, nameplate content, and purchase receipt content from the set of valuation elements of the involved property; Perform character cleansing on the brand name to convert it into a standard brand name; The model name is split into letter segment, number segment and version segment, and the splitting results are compared with the standard model codes in the preset model library item by item to filter candidate models with the same model body; The specifications are arranged by parameter name and parameter value. The standard specifications of each candidate model are compared with the actual specifications of the property involved in the case, and candidate models with more than the preset value of parameter conflicts are deleted. For the multiple candidate models that are retained, cross-verification is carried out in the following order: nameplate field first, voucher field second, and appearance image field third, to determine the unique standard model. An identification code is generated based on the standard brand, standard model, specification level, production batch, and asset category, and the identification code is used as the primary index. The circulation status is determined based on the current availability and transaction status; The standard identification profile of assets is composed of the main index, brand value level, and circulation status.

5. The method for valuing seized property based on artificial intelligence feature fusion according to claim 1, characterized in that, Step three specifically involves: Extract market records corresponding to standard brand, standard model and specification level from the standard identification profile of the property, and classify the records with the same model into the first market layer; Records from the same series with specifications that are within the allowable range are classified into the second market level; Records of the same category and with the same main functions will be placed in the third market data layer; The judicial disposal transaction records, second-hand platform transaction records, industry quotation records, e-commerce new machine price records, and auction transaction records in each market level are respectively formed into corresponding price sub-matrices; Calculate the median transaction price for each row in the price submatrix; Add the median of each transaction price and divide by the number of source categories with valid data to obtain the real-time benchmark price. The real-time market benchmark matrix is ​​composed of real-time market benchmark prices from various sources.

6. The method for valuing seized property based on artificial intelligence feature fusion according to claim 1, characterized in that, Step four specifically involves: The actual service life is calculated based on the interval between the valuation benchmark date and the commissioning date. The standard economic life is read from the economic life table corresponding to the property category. The actual service life is divided by the standard economic life to obtain the service life consumption ratio. The usage intensity ratio is calculated based on the ratio of cumulative usage to the current model's rated usage. The maintenance activities in the maintenance records are divided into routine maintenance, general maintenance and major maintenance. Routine maintenance does not increase depreciation. Each general maintenance increases the preset maintenance attenuation amount. Major maintenance increases the maintenance attenuation amount based on the degree of damage to the core components. Compare the replaced parts in the parts replacement record with the critical parts catalog; if a new original part is replaced, add a preset maintenance attenuation amount; if a non-original part or a used part is replaced, leave it unchanged. The difference between the actual maintenance interval and the specified maintenance interval in the maintenance record is calculated as the maintenance deficiency attenuation amount; The initial value decay ratio is obtained by adding the proportion of annual consumption, the proportion of usage intensity, the amount of maintenance decay, and the amount of maintenance deficiency decay according to preset weights. The depreciation correction factor is obtained by subtracting the initial value decay ratio from 1, and the depreciation correction price is generated by multiplying the real-time market benchmark prices in the real-time market benchmark matrix by the depreciation correction factor.

7. The method for valuing seized property based on artificial intelligence feature fusion according to claim 1, characterized in that, The improved TabNet model is specifically as follows: The identity coding features in the standard identification profile of assets, the market features in the real-time market benchmark matrix, and the corresponding depreciation features are concatenated in a fixed column order to form the original fused feature vector; The original fused feature vector is input into the feature input embedding layer, and each discrete code is converted into a dense feature vector of fixed length. The dense discrete features are input into the first feature transformer of the improved TabNet model. The first feature transformer performs linear transformation through a fully connected layer, stabilizes the feature distribution through a batch normalization layer, and divides the input features into content branches and gated branches through a gated linear unit. The content branch undergoes a nonlinear transformation, and the gated branch is activated by Sigmoid and then multiplied position by position with the content branch to obtain the first-stage transformation features. The first stage transformation features are input into the attention transformer, and the selection scores of each input feature are generated through the fully connected layer. The features used in the previous decision step are combined with the feature usage record to remove features that have been repeatedly selected. A sparse activation function is used to compress the selection scores that are less than a preset score threshold to 0, forming a sparse feature mask for the current decision step. The sparse feature mask is multiplied element-wise with the original fused feature vector and then input into the feature transformer again. The estimated decision features for the current decision step are generated through fully connected transformation, batch normalization and gated linear units. The remaining features are then passed to the next decision step to continue mask selection and feature transformation. The valuation decision features generated from multiple decision steps are transformed into the same dimension through a linear layer and then added position by position to form a multi-stage fusion feature; The multi-stage fusion features are input into the fully connected layer respectively, and the linear combination of the fusion features is calculated to obtain the predicted center price; Subtracting the preset initial deviation value from the prediction center price yields the initial valuation lower limit, and adding the preset initial deviation value to the prediction center price yields the initial valuation upper limit. The initial valuation range is formed by using the lower limit and the upper limit of the initial valuation.

8. The method for valuing seized property based on artificial intelligence feature fusion according to claim 1, characterized in that, Step six specifically involves: Compare the initial valuation range with the depreciation adjustment prices. If the initial valuation range covers all depreciation adjustment prices, retain the initial valuation range. When the lower limit of the initial valuation range is higher than all depreciation adjustment prices, the lower limit of the initial valuation range will be adjusted to the average of all depreciation adjustment prices. When the upper limit of the initial valuation range is lower than all depreciation adjustment prices, the upper limit of the initial valuation range will be adjusted to the average of all depreciation adjustment prices and the forecast center price. After adjusting the interval boundaries, the target valuation lower limit, target valuation median, and target valuation upper limit are generated, forming a traceable valuation interval.

Images