Material disposal condition automatic summarization and asset estimation verification method

Abstract

The invention relates to the technical field of asset valuation, and discloses a material disposal condition automatic summarization and asset valuation verification method, which comprises the following steps of: obtaining attribute feature vectors of new warehousing materials, screening a reference category set from a mature material category library, and fusing a multiple regression model parameter set of a reference category to generate prior parameter distribution; limited historical records of new warehousing materials are extracted to generate a multivariate sample data set, posterior regression parameter distribution is generated based on Bayesian fusion, a prediction residual value rate and prediction uncertainty are calculated, a prediction estimated value is generated and compared with a declaration estimated value, a verification result is output, and the problem that multi-factor residual value prediction is insufficient in accuracy under the small sample condition is solved.

Description

Technical Field

[0001] This invention relates to the field of asset valuation technology, and more specifically, to a method for automatically summarizing the disposal status of materials and verifying asset valuation. Background Technology

[0002] In corporate asset disposal scenarios, the residual value decay of lead-acid and lithium batteries is affected by multiple factors, including usage time, charge-discharge cycle count, maintenance condition, and ambient temperature. Existing technologies employ two residual value prediction methods: The first method finds a reference category through battery attribute feature similarity matching, transferring its fluctuation characteristics as prior knowledge and fusing it with limited historical data using Bayesian methods. This method can solve the small sample estimation problem, but it can only transfer a single fluctuation characteristic and cannot reflect the differentiated impact of multiple factors on residual value. The second method uses a multivariate nonlinear regression model to predict the residual value rate using multiple factors such as charge-discharge cycle count, capacity decay rate, and environmental factors as input variables. This method can capture the interactive effects of multiple factors, but requires sufficient historical samples to train multiple parameters of the multivariate nonlinear regression model. Newly acquired battery materials face two challenges: insufficient historical data to reliably train multiple parameters of the multivariate nonlinear regression model, and the inability of simply transferring a single fluctuation characteristic of similar materials to reflect the differentiated impact of multiple factors on residual value. Consequently, neither method, used alone, can achieve accurate multi-factor residual value prediction under small sample conditions. Summary of the Invention

[0003] This invention provides a method for automatically summarizing the disposal status of materials and verifying asset valuation, which solves the technical problems in related technologies, such as low residual value prediction accuracy due to insufficient historical data for newly received materials, inability to fully utilize prior knowledge of mature categories, and lack of a mechanism for quantifying prediction uncertainty.

[0004] This invention provides a method for automatically summarizing material disposal status and verifying asset valuation, including:

[0005] Obtain the attribute feature vector of newly received materials. The attribute feature vector includes the material category, technical parameters, applicable scenarios, and a list of influencing factors.

[0006] Obtain the attribute feature vectors of each category in the mature material category library and the parameter set of the trained multivariate regression model. The parameter set of the multivariate regression model includes the regression coefficients of each factor and the interaction term parameters.

[0007] Calculate the similarity between the attribute feature vector of newly added materials and the attribute feature vector of each category in the mature material category library, and select the top K categories by similarity ranking as the reference category set;

[0008] The parameter sets of the multiple regression models for each category in the reference category set are weighted and fused to generate the prior parameter distribution of the multiple regression model for newly entered materials. The fusion weight is positively correlated with the similarity.

[0009] Obtain historical disposal records of newly received materials, extract multi-factor feature values ​​and actual residual value rates from each historical disposal record, and generate a multivariate sample dataset.

[0010] A likelihood function is constructed based on a multivariate sample dataset. The likelihood function is then fused with the prior parameter distribution using Bayesian methods to generate the posterior regression parameter distribution.

[0011] Obtain the multi-factor feature vector of the material to be valued, substitute the multi-factor feature vector into the posterior mean parameter of the posterior regression parameter distribution, and calculate the predicted residual value rate.

[0012] A predicted valuation is generated based on the predicted residual value rate and the purchase price of the materials to be valued. The declared valuation is then compared with the predicted valuation, and the verification result is output.

[0013] Furthermore, the attribute feature vector is constructed as follows: one-hot encoding is performed on categorical attributes, and mean normalization based on range is applied to numerical attributes. The encoded attribute values ​​are then concatenated to form a vector of fixed dimensions. The one-hot encoding is implemented as follows: for a categorical attribute with M possible values, a binary vector of length M is constructed, and a value of 1 is assigned to the position corresponding to the actual value of the categorical attribute, while a value of 0 is assigned to the other positions.

[0014] Furthermore, the residual rate output of the multiple regression model consists of the intercept term, the sum of the products of the main effect terms of each influencing factor and their corresponding regression coefficients, the sum of the products of the interaction terms between any two influencing factors and their corresponding interaction term parameters, and the random error term; the mature material category refers to the material category whose number of historical disposal records reaches the preset sample size threshold, which is 5 to 10 times the total number of parameters in the parameter set of the multiple regression model.

[0015] Furthermore, cosine similarity is used to calculate the similarity. When calculating the similarity, each dimension of the attribute feature vector is weighted. The weights are obtained by averaging the absolute values ​​of the regression coefficients of the multiple regression models of each category in the mature material category library and the absolute values ​​of the regression coefficients corresponding to the same attribute dimension. The number of reference categories K is an integer between 3 and 10, which is determined according to the size of the mature material category library.

[0016] Furthermore, the prior parameter distribution adopts a multivariate normal distribution. The prior mean is obtained by weighting the parameter sets of the multivariate regression models for each category in the reference category set according to the fusion weights. The fusion weights are obtained by normalizing the similarity between each reference category and the newly entered materials. The prior covariance matrix is ​​obtained by weighting the sum of the outer products of the differences between the parameter sets of the multivariate regression models for each reference category and the prior mean according to the fusion weights, and then adding a regularization term. The regularization term is the product of the regularization coefficient and the identity matrix.

[0017] Furthermore, the multi-factor feature values ​​include service life, usage intensity index, maintenance score and environmental factors; the multi-factor feature vector is constructed as follows: the normalized values ​​of each influencing factor are concatenated in a preset order as the main effect term, the product between any two influencing factors is calculated as the interaction term and concatenated, and a constant term is added to the first and second parts of the multi-factor feature vector to correspond to the intercept parameter.

[0018] Furthermore, the posterior regression parameters are distributed as a multivariate normal distribution. The posterior covariance matrix is ​​obtained by adding the inverse of the prior covariance matrix and the sum of the outer products of the multivariate sample dataset's eigenvectors, divided by the residual variance, and then taking the inverse of the matrix. The posterior mean is obtained by multiplying the posterior covariance matrix by the inverse of the prior covariance matrix and the prior mean, and then adding the product of the posterior covariance matrix and the sum of the multiplied eigenvectors of the corresponding actual residual rates, divided by the residual variance.

[0019] Furthermore, the prediction uncertainty is calculated based on the posterior covariance matrix of the posterior regression parameter distribution. The prediction uncertainty is the square root of the product of the multifactor eigenvector of the material to be valued, the posterior covariance matrix, and the transpose of the multifactor eigenvector. The uncertainty range is determined by the product of the prediction uncertainty and the purchase price. If the declared valuation falls within the uncertainty range, the verification result is passed; if the declared valuation exceeds the uncertainty range, the verification result is pending review.

[0020] Furthermore, when outputting the verification results, the contribution of each influencing factor to the predicted residual value rate is output simultaneously. For the j-th influencing factor, the main effect contribution is the product of the regression coefficient corresponding to the j-th influencing factor in the posterior mean vector and the normalized value of the j-th influencing factor of the material to be valued. The interaction term contribution is the sum of the interaction term parameters in which the j-th influencing factor participates and the product of the normalized values ​​of the corresponding two influencing factors. The total contribution is the sum of the main effect contribution and the interaction term contribution.

[0021] This invention provides an automatic system for summarizing material disposal status and verifying asset valuation, comprising:

[0022] The attribute feature acquisition module is used to acquire the attribute feature vectors of newly received materials.

[0023] The mature category library management module is used to obtain the attribute feature vectors of each category in the mature material category library and the parameter set of the trained multivariate regression model;

[0024] The similarity calculation and filtering module is used to calculate the similarity between the attribute feature vector of newly added materials and the attribute feature vector of each category in the mature material category library, and to filter the reference category set.

[0025] The prior parameter fusion module is used to perform weighted fusion of the parameter sets of multiple regression models for each category in the reference category set to generate the prior parameter distribution.

[0026] The sample dataset generation module is used to obtain historical disposal records of newly received materials and generate a multivariate sample dataset.

[0027] The Bayesian fusion module is used to construct a likelihood function based on a multivariate sample dataset, and then perform Bayesian fusion of the likelihood function with the prior parameter distribution to generate the posterior regression parameter distribution.

[0028] The residual value prediction module is used to substitute the multi-factor feature vector of the material to be valued into the posterior mean parameter of the posterior regression parameter distribution to calculate the predicted residual value rate.

[0029] The valuation verification module is used to generate a predicted valuation based on the predicted residual value rate and the purchase price, compare the declared valuation with the predicted valuation, and output the verification result.

[0030] The beneficial effects of this invention are as follows:

[0031] This invention uses the parameter set of a multiple regression model for similar materials as prior knowledge and performs Bayesian fusion with the limited historical data of newly entered materials. This solves the technical problem of needing to reflect the interaction of multiple factors under small sample conditions while facing unreliable parameter estimation. It achieves the technical effect of adaptively balancing the contribution of prior knowledge and actual data under different sample sizes, realizing accurate multi-factor residual value prediction and quantifying prediction uncertainty. Attached Figure Description

[0032] Figure 1 This is a flowchart of the method for automatically summarizing material disposal status and verifying asset valuation according to the present invention;

[0033] Figure 2 This is a diagram showing the construction of the attribute feature vector of newly received materials according to the present invention;

[0034] Figure 3 This is a mature material category similarity calculation and ranking diagram of the present invention;

[0035] Figure 4 This is the reference category fusion weight allocation diagram of the present invention;

[0036] Figure 5 This is a multi-factor feature distribution map of the historical disposal records of the present invention;

[0037] Figure 6 This is a comparison chart of parameters before and after Bayesian fusion according to the present invention;

[0038] Figure 7 This is a graph showing the contribution of the influencing factors of this invention to the residual value rate.

[0039] Figure 8 This is a verification chart of the residual value prediction results and the declared valuation of the present invention. Detailed Implementation

[0040] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0041] At least one embodiment of the present invention discloses a method for automatically summarizing material disposal status and verifying asset valuation, such as... Figure 1 As shown, it includes the following steps:

[0042] Step 1: Obtain the attribute feature vector of the newly received materials.

[0043] Obtain basic attribute information of newly received materials and generate attribute feature vectors. The attribute feature vector includes the material category, technical parameters, applicable scenarios, and a list of influencing factor types.

[0044] It should be noted that the material category includes the specific model classification of lead-acid batteries or lithium batteries; the technical parameters include specifications such as rated capacity, nominal voltage, and design cycle life; the applicable scenarios include application types such as energy storage, power, and backup power; the list of influencing factors includes the type identifiers of factors that affect the residual value decay of this type of material, such as enumeration identifiers of factors such as the number of charge and discharge cycles, ambient temperature, and maintenance frequency.

[0045] It should be noted that attribute feature vectors The construction method is as follows: one-hot encoding is performed on categorical attributes, mean normalization based on range is applied to numerical attributes, and the encoded attribute values ​​are concatenated to form a vector of fixed dimensions.

[0046] Furthermore, the implementation of one-hot encoding is as follows: for those with A categorical attribute with possible values, with a construction length of . A binary vector, where To determine the number of possible values ​​for a categorical attribute, a value of 1 is assigned to the position corresponding to the actual value of the categorical attribute, and a value of 0 is assigned to the other positions, thereby converting the categorical information into a numerical vector representation.

[0047] Step 2: Obtain the attribute feature vectors and multivariate regression model parameter set of the mature material category library.

[0048] Obtain a mature material category library with sufficient historical data from the system, and extract attribute feature vectors for each category from the mature material category library. and the parameter set of the trained multivariate regression model ,in The quantity of mature material categories, For indexing mature material categories, the parameter set of the multiple regression model It includes the regression coefficients and interaction term parameters for each factor.

[0049] It should be noted that a mature material category refers to a material category whose number of historical disposal records has reached a preset sample size threshold. The preset sample size threshold is determined based on the number of parameters in the parameter set of the multiple regression model to ensure the reliability of parameter estimation in the parameter set of the multiple regression model.

[0050] Furthermore, the preset sample size threshold is determined by: counting the total number of parameters in the parameter set of the multivariate regression model, multiplying the total number by a coefficient of 5 to 10 to obtain the preset sample size threshold, ensuring that each parameter corresponds to a sufficient number of sample observations on average.

[0051] It should be noted that the form of the multiple regression model is as follows:

[0052]

[0053] in, Residual value rate For the first The values ​​of each influencing factor, For the first The values ​​of each influencing factor, For the intercept term, For the first The regression coefficients of the influencing factors, For the first The first influencing factor and the second Interaction term parameters of each influencing factor, The number of influencing factors, For the index of influencing factors, For the index of influencing factors, This represents the random error term. (Parameter set of the multiple regression model) It includes all the parameters to be estimated mentioned above.

[0054] Furthermore, the multiple regression model is trained using the least squares method, with the input being multi-factor feature vectors. The multi-factor feature vector includes the main effect terms and interaction terms of each influencing factor; the output is the predicted residual rate. The training process determines the optimal parameter set by minimizing the sum of squared errors between the predicted residual rate and the actual residual rate.

[0055] Furthermore, the interaction term in the multiple regression model The physical meaning is: to capture the first The first influencing factor and the second The combined effect of several influencing factors on residual value rate, such as the combined effect of high temperature environment and high frequency charge-discharge cycle, may lead to accelerated residual value decay. This accelerated decay effect exceeds the simple summation of the individual effects of the two factors, as indicated by the interaction term parameters. Perform quantitative modeling.

[0056] Step 3: Calculate similarity and filter the reference category set.

[0057] Calculate the feature vector of newly received materials attributes Compared with the attribute feature vectors of each category in the mature material category library similarity Before filtering by similarity ranking The category of the bit is used as a reference category set. ,in For reference, the number of categories, The reference category ranked first in similarity. The reference category ranked second in similarity. Ranked by similarity The reference category of the bit.

[0058] It should be noted that similarity The calculation uses cosine similarity.

[0059] It should be noted that the number of reference categories The value ranges from 3 to 10, and is determined by the number of categories. The specific value is determined based on the size of the mature material category library.

[0060] Furthermore, refer to the number of categories. The method for determining the number of categories is as follows: when the size of the mature material category library is less than 20 categories, the number of reference categories is taken. The value is 3; when the size of the mature material category library is between 20 and 50 categories, the reference category number is used. The value is 5; when the size of the mature material category library exceeds 50 categories, the number of reference categories is used. It is 10.

[0061] In this embodiment of the application, in order to improve the accuracy of reference category screening, the attributes feature vectors are weighted when calculating similarity. The weights are determined according to the degree of influence of each attribute dimension on residual value decay, and the attribute dimensions with greater influence are given higher weights.

[0062] Furthermore, the method for determining the weight of the attribute dimension is as follows: the absolute values ​​of the regression coefficients of the multiple regression models of each category in the mature material category library are statistically analyzed, the absolute values ​​of the regression coefficients corresponding to the same attribute dimension are averaged, and the average value is used as the weight of the same attribute dimension, so that the attributes that have a greater impact on residual value decay will receive higher importance in similarity calculation.

[0063] Step 4: Integrate reference category parameters to generate prior parameter distribution.

[0064] For reference category set The parameter sets of various categories of multiple regression models are weighted and fused to generate the prior parameter distribution of the multiple regression model for newly received materials. The fusion weight is positively correlated with similarity.

[0065] It should be noted that the prior parameter distribution Adopting a multivariate normal distribution:

[0066]

[0067] in, For the parameter set of the multiple regression model, Represents a multivariate normal distribution. The prior mean, This is the prior covariance matrix. The prior mean is... Calculated by weighted average:

[0068]

[0069] in, For reference category index, For the first The fusion weight of each reference category, For reference category The parameter set of the multiple regression model. (Fusing weights) Obtained by similarity normalization:

[0070]

[0071] in, For reference category Similarity to newly received goods, For reference category Similarity to newly received goods, For summation index.

[0072] It should be noted that the prior covariance matrix Calculated using weighted covariance:

[0073]

[0074] in, This represents the matrix transpose operation. The regularization coefficient is . The identity matrix is ​​used, and the regularization term is used to ensure the positive definiteness of the prior covariance matrix.

[0075] Furthermore, the regularization coefficient The value range is from 0.001 to 0.1, and the regularization coefficient is... The specific value is determined based on the degree of dispersion of the parameter set of the multiple regression model in the reference category set. A larger value is taken when the difference between the parameter sets of the multiple regression model is large, and a smaller value is taken when the difference between the parameter sets of the multiple regression model is small.

[0076] Furthermore, the measure of the dispersion of the parameter set of a multiple regression model is as follows: calculate the parameter set of each multiple regression model in the reference category set relative to the prior mean. The average Euclidean distance is used to determine the degree of dispersion. If the average Euclidean distance exceeds a preset threshold, it is considered to have a large degree of dispersion; otherwise, it is considered to have a small degree of dispersion. The preset threshold ranges from 0.5 to 2.0.

[0077] Furthermore, the prior covariance matrix The purpose is to quantify the uncertainty of prior parameter estimation. The diagonal elements of the prior covariance matrix reflect the variance of each parameter's own estimation, while the off-diagonal elements of the prior covariance matrix reflect the correlation between the estimates of different parameters. The uncertainty information is used in subsequent Bayesian fusion to balance the contribution weights of prior knowledge and actual data.

[0078] Step 5: Extract historical records to generate a multivariate sample dataset.

[0079] Obtain limited historical disposal records of newly received materials, extract the service life, usage intensity index, maintenance score, environmental factors, and actual residual value rate of each historical disposal record, and generate a multivariate sample dataset. ,in For the first Multi-factor feature vector of historical disposal records For the first The actual residual value rate corresponding to each historical disposal record. The number of historical disposal records, An index for historical disposal records.

[0080] It should be noted that the usage intensity index includes quantitative indicators reflecting the battery's usage level, such as the cumulative number of charge-discharge cycles and the average depth of discharge; the maintenance score is a numerical score that is comprehensively evaluated based on dimensions such as the completeness of maintenance records, maintenance frequency, and maintenance quality; and environmental factors include quantitative values ​​of environmental conditions such as average storage temperature and humidity.

[0081] It should be noted that before constructing the multi-factor feature vector, data preprocessing is performed on each influencing factor: mean normalization based on range is used for numerical influencing factors such as service life, cumulative charge and discharge cycles, average discharge depth, maintenance score, average storage temperature, and humidity to eliminate the dimensional differences between different influencing factors and to enable regression modeling of each influencing factor on the same scale.

[0082] It should be noted that multi-factor feature vectors The construction method is as follows: Normalized values ​​of various influencing factors, such as service life, usage intensity index, maintenance score, and environmental factors, are concatenated in a preset order. The interaction terms between each pair of influencing factors are then calculated and concatenated to form a complete feature vector containing both main effect terms and interaction terms. Specifically, if there are... If there are multiple influencing factors, then the multi-factor eigenvector... Its structure is as follows: First, it contains One main effect term ,in For the first The normalized value of the first influencing factor in each historical disposal record. For the first The normalized value of the second influencing factor in the historical disposal record. For the first Article number of historical disposal records Normalized values ​​of each influencing factor; then including Interactive items (in ), For the first Article number of historical disposal records Normalized values ​​of each influencing factor. For the first Article number of historical disposal records The normalized values ​​of each influencing factor were then used; finally, a constant term of 1 was added to the first position of the multi-factor eigenvector to correspond to the intercept parameter. .

[0083] Furthermore, the maintenance score is calculated as follows: scores are given based on three dimensions: the completeness of maintenance operations in historical maintenance records, whether the maintenance interval meets the specifications, and the degree of recovery of battery performance indicators after maintenance. The scores of the three dimensions are weighted and summed to obtain the comprehensive maintenance score, which ranges from 0 to 100. The higher the comprehensive maintenance score, the better the maintenance condition.

[0084] Step 6: Bayesian fusion generates the posterior regression parameter distribution.

[0085] Based on multivariate sample datasets Constructing the likelihood function The likelihood function is compared with the prior parameter distribution. Perform Bayesian fusion to generate the posterior regression parameter distribution. And the confidence interval of the parameters.

[0086] It should be noted that the likelihood function is constructed based on the assumption that the residuals of the multiple regression model follow a normal distribution:

[0087]

[0088] in, This represents the matrix transpose operation. For residual variance, For the multiple regression model for the first The predicted residual value rate of each historical disposal record.

[0089] Furthermore, residual variance The estimation method is as follows: the sum of squared sample residuals is divided by the degrees of freedom, where the degrees of freedom is the number of historical disposal records. Subtract the total number of parameters in the parameter set of the multiple regression model.

[0090] It should be noted that the posterior regression parameter distribution is calculated based on Bayes' theorem:

[0091]

[0092] Since both the prior parameter distribution and the likelihood function are in normal form, the posterior regression parameter distribution... It is still a multivariate normal distribution ,in The posterior mean is... Let be the posterior covariance matrix and posterior mean. and posterior covariance The posterior covariance matrix is ​​obtained through the analytical formula of the conjugate prior. Specifically, the formula for calculating the posterior covariance matrix is:

[0093]

[0094] in, Let be the inverse of the prior covariance matrix. This represents the matrix transpose operation. The formula for calculating the posterior mean is:

[0095]

[0096] Furthermore, the adaptive mechanism of Bayesian fusion is manifested in: when the number of historical processing records... When smaller, item In the calculation of the posterior covariance matrix, the posterior regression parameter distribution is more closely similar to the prior parameter distribution, reflecting a dependence on prior knowledge; when the number of historical disposal records... When it increases, the term Gradually taking the lead, the posterior regression parameter distribution reflects more information from the actual data, thereby achieving an automatic balance between the contribution weights of prior knowledge and actual data.

[0097] It should be noted that the parameter confidence interval is calculated based on the posterior mean and posterior covariance of the posterior regression parameter distribution. For the parameter vector... The Each component , No. Each component The confidence level is The confidence interval of the parameter is ,in The posterior mean vector is the first... One portion, The square root of the diagonal elements of the posterior covariance matrix. The quantiles of the standard normal distribution The significance level is indicated by .

[0098] Furthermore, confidence level The value is 0.95, which corresponds to the standard normal distribution quantile. It is 1.96.

[0099] Step 7: Calculate the predicted residual value rate and the prediction uncertainty.

[0100] Obtain the multi-factor feature values ​​of the material to be valued, and apply the same normalization process as in step 5 to each influencing factor to construct the multi-factor feature vector of the material to be valued. The multi-factor feature vector of the material to be valued is substituted into the posterior mean parameter of the posterior regression parameter distribution. Calculate the predicted residual value rate Calculate prediction uncertainty based on parameter confidence intervals .

[0101] It should be noted that the formula for calculating the predicted residual value rate is:

[0102]

[0103] in, This represents the matrix transpose operation. The calculation of prediction uncertainty is based on the covariance propagation of the posterior regression parameter distribution:

[0104]

[0105] in, Let be the posterior covariance matrix of the posterior regression parameter distribution.

[0106] Furthermore, the multi-factor characteristic values ​​of the materials to be valued include measured or estimated values ​​of influencing factors such as current service life, cumulative charge-discharge cycles, average depth of discharge, maintenance score, average storage temperature, and humidity. The range and mean parameters used in the normalization process are consistent with the normalization parameters of the training samples in step 5 to ensure the uniformity of the feature scale.

[0107] Furthermore, prediction uncertainty The physical meaning is: to quantify the standard error of the predicted residual value rate caused by the uncertainty of parameter estimation. The prediction uncertainty comprehensively reflects the uncertainty of prior knowledge, the uncertainty of finite historical samples, and the influence of the position of the multi-factor feature vector of the material to be valued in the feature space, and provides a risk assessment basis for asset disposal decisions.

[0108] Step 8: Generate the predicted estimate and verification results.

[0109] Obtain the purchase price of the materials to be valued. The purchase price and the projected residual value rate Multiply to generate a predicted estimate and uncertainty interval ; Obtain the valuation for the application The submitted valuation is compared with the predicted valuation and the uncertainty range, and the verification result is output.

[0110] It should be noted that the judgment rule for the verification result is: if the declared valuation If the value falls within the uncertainty range, the verification result is considered passed; if the declared valuation... If the value exceeds the uncertainty range, the verification result will be "to be reviewed" and the deviation rate between the declared valuation and the predicted valuation will be output.

[0111] Furthermore, the formula for calculating the deviation rate is: The deviation rate is used to quantify the relative difference between the declared valuation and the predicted valuation.

[0112] In this embodiment of the application, in order to improve the interpretability of the verification results, the contribution of each influencing factor to the predicted residual value rate is output simultaneously when the verification results are output. The contribution is calculated based on the product of the posterior mean and the value of the corresponding influencing factor, which makes it easier for reviewers to identify the main factors affecting the residual value prediction.

[0113] Furthermore, the calculation method for the contribution of influencing factors is as follows: For the first... The influencing factor, the first The main effect contribution of each influencing factor is: ,in The th element in the posterior mean vector The regression coefficients corresponding to each influencing factor For the materials to be valued The normalized values ​​of the first influencing factor; the second The contribution of the interaction terms involving each influencing factor is ,in For interactive item parameters, For the materials to be valued Normalized values ​​of each influencing factor. For the summation index; the first... The total contribution of each influencing factor is the sum of the contribution of the main effect and the contribution of the interaction term.

[0114] This implementation addresses the problem of unreliable multi-parameter estimation under small sample conditions by transferring the parameter set of a multiple regression model of similar materials as prior knowledge. Since the prior parameter distribution integrates the regression coefficients and interaction term parameters of multiple reference categories, the multiple regression model of newly added materials can inherit the modeling ability of mature material categories for the interaction of multiple factors, rather than simply transferring a single fluctuation feature.

[0115] This implementation incorporates information from limited historical data on newly arrived materials into the posterior regression parameter distribution using Bayesian fusion, enabling the multiple regression model to adapt to the unique decay patterns of newly arrived materials based on prior knowledge. Because Bayesian fusion assigns greater weight to the prior parameter distribution when the sample size is small, and gradually transitions to a likelihood function-dominated approach as the sample size increases, this implementation can adaptively balance the contributions of prior knowledge and actual data under different sample size conditions.

[0116] This implementation method calculates the prediction uncertainty based on the posterior covariance of the posterior regression parameter distribution, which can quantify the impact of the uncertainty of parameter estimation on the prediction results under small sample conditions, and provide a risk reference for asset disposal decisions.

[0117] A power company Year The month received a batch of retired lithium iron phosphate power batteries (model: This batch of batteries originated from a battery pack replacement project for electric buses. These batteries are classified as newly received materials and are only listed in the company's asset disposal system. The historical disposal records have a sample size far smaller than that required for training a multiple regression model. Record threshold. The basic attributes of a battery are: rated capacity. nominal voltage Design cycle life The next application scenario is for power batteries. Mature lead-acid backup batteries, ternary lithium energy storage batteries, and lithium iron phosphate forklift batteries are already available in the system. There are several mature material categories, all of which have sufficient historical data and trained multivariate regression models.

[0118] Now we need to process one of the groups numbered as follows The residual value of the battery was predicted and the declared valuation was verified. The declared valuation of the battery was... Yuan, purchase price was Yuan.

[0119] Example of implementing core step 1: Constructing the attribute feature vector of newly received materials.

[0120] The system first extracts the basic attribute information of this batch of lithium iron phosphate power batteries, including the material category, technical parameters, applicable scenarios, and a list of influencing factors. The material category is "lithium iron phosphate power batteries," and the technical parameters include rated capacity. Nominal voltage Design cycle life The applicable scenario is "power application", and the influencing factors include four factors: charge and discharge cycle count, ambient temperature, maintenance frequency, and depth of discharge.

[0121] Table 1. Input data for step 1 (original attribute data of newly received materials):

[0122]

[0123] One-hot encoding is used for the categorical attribute "material category": There are a total of 100 battery categories in the system. Types (lead-acid backup, ternary lithium energy storage, lithium iron phosphate power, lithium iron phosphate forklift, nickel-metal hydride power), with a construction length of [missing information]. The vector, the lithium iron phosphate power corresponding to the first Bits, encoded as One-hot encoding is used for "applicable scenarios": There are a total of scenario types. Types (energy storage, power, backup), power application corresponds to the first Bits, encoded as Normalize numerical attributes:

[0124]

[0125]

[0126]

[0127] Table 2 Output data of step 1 (feature vector of newly received material attributes):

[0128]

[0129] The final generated attribute feature vector:

[0130] The vector dimension is .

[0131] Figure 2 The construction of the attribute feature vector of newly received materials is demonstrated.

[0132] Example of implementing core step 2: Extracting parameter information of mature material categories

[0133] The system extracts from the mature material category library The attribute feature vectors of each category and the parameter set of the multivariate regression model. (The rest is omitted as it is not part of the main text.) The data structure is illustrated using a typical category as an example.

[0134] Table 3. Input data for step 2 (example of feature vector for mature material category attributes):

[0135]

[0136] These mature categories of multiple regression models all include One main effect parameter ( to (corresponding to the number of cycles, ambient temperature, maintenance frequency, and depth of discharge) and Interaction item parameters ( to ), plus the intercept term Each model has a total of Several parameters. The system's preset sample size threshold is... All of the above categories meet the requirements.

[0137] Table 4 Output data for step 2 (Example of parameter set for multiple regression model of mature material categories):

[0138]

[0139] Note: Only some parameters are shown in the table; the complete parameter set includes all parameters. One parameter.

[0140] Example of implementing core step 3: Calculate similarity and filter reference categories.

[0141] The system calculates newly received materials and The cosine similarity of each mature category. Before calculation, the attribute dimensions are weighted according to the statistical results of the regression coefficients of the mature categories: the weight of the material category code is... Applicable scenario encoding weight is The rated capacity weight is The nominal voltage weight is The design cycle life weight is .

[0142] For category The weighted similarity is calculated as follows:

[0143]

[0144] in For the first Dimensional weights, For newly received materials 3D eigenvalues For category No. 1 eigenvalue. Substituting the numerical values, we obtain... .

[0145] Table 5 Output data of step 3 (similarity calculation results and reference category filtering):

[0146]

[0147] Because the size of the mature material category warehouse is One (in) to (between), take the number of reference categories Before filtering by similarity ranking The categories of bits constitute a reference category set. .

[0148] Figure 3 It demonstrates the similarity of various categories of mature materials.

[0149] Example of implementing core step 4: fusing reference category parameters to generate a prior distribution.

[0150] The system The parameter sets of the multiple regression models for each reference category are weighted and fused. First, the fusion weights are calculated, and the similarity is normalized:

[0151]

[0152] Similarly, the calculation yields , , , .

[0153] Table 6 Output data for step 4 (calculated with reference to category fusion weights and prior parameters):

[0154]

[0155] The components of the prior mean vector are obtained by weighted summation: (Intercept term) (Loop count coefficient), the rest The individual parameter components are calculated similarly.

[0156] When calculating the prior covariance matrix, the deviations of each reference class parameter from the prior mean are first calculated, for example, the class... The intercept term deviation is .for The covariance matrix, with diagonal elements calculated as follows:

[0157]

[0158] Where regularization coefficient (Referring to the category parameters, the dispersion is moderate, and the average Euclidean distance is...) (Take the middle value).

[0159] Figure 4 The specific fusion weights for each reference category are displayed.

[0160] Implementation example of core step 5: Extracting historical records to construct a multivariate sample dataset.

[0161] The system extracts the lithium iron phosphate power batteries from this batch. Each historical disposal record includes the service life, cumulative number of cycles, average storage temperature, maintenance score, average depth of discharge, and actual residual value rate.

[0162] Table 7 Input data for step 5 (raw data from historical processing records):

[0163]

[0164] Each influencing factor was normalized, and the normalization parameters were determined based on the range of historical data: service life range. Loop count range Temperature range Maintain the scoring range Discharge depth range To record For example, the normalized calculation is as follows:

[0165]

[0166]

[0167]

[0168]

[0169]

[0170] When constructing multi-factor feature vectors, in addition to the main effect terms, interaction terms also need to be calculated, for example... .

[0171] Table 8 shows the output data of step 5 (example of a normalized multivariate sample dataset):

[0172]

[0173] Complete feature vector Include One constant term, One main effect term and Each interactive item (take) ),common dimension.

[0174] Figure 5 The normalized values ​​of multi-factor characteristics of historical disposal records are displayed.

[0175] Implementation example of core step 6: Bayesian fusion to generate posterior parameter distribution.

[0176] based on The likelihood function is constructed from the historical records, and the residual variance is estimated as follows: (Calculated from sample residuals). In the calculation of the posterior covariance matrix, the prior accuracy matrix... With data information matrix Add and then take the inverse:

[0177]

[0178] The posterior mean is calculated by fusing prior information and sample information. Taking the intercept term as an example:

[0179]

[0180] Table 9 Output data of step 6 (posterior parameter distribution and confidence interval):

[0181]

[0182] The posterior mean of the iteration number coefficient is Compared to the prior mean Slight adjustments were made to reflect the corrective effect of the actual data. Diagonal elements of the posterior covariance matrix. Less than the prior covariance This indicates that uncertainty is reduced after data fusion.

[0183] Figure 6 This demonstrates a comparison of the prior and posterior mean parameters before and after Bayesian fusion.

[0184] Example of implementing core step 7: Calculate the predicted residual rate and uncertainty.

[0185] Batteries awaiting valuation The measured multi-factor characteristic values ​​are: service life Year, cumulative number of cycles Average storage temperature ℃, maintenance rating Average depth of discharge The same normalization parameters as historical samples are used for processing:

[0186]

[0187]

[0188]

[0189]

[0190]

[0191] Construct the complete feature vector:

[0192] It includes all main effect items and interaction items. dimension.

[0193] Table 10 Input data for step 7 (multi-factor characteristic values ​​of the battery to be valued):

[0194]

[0195] Predicted residual value rate calculation:

[0196]

[0197] Prediction uncertainty calculation:

[0198]

[0199] Table 11 Output data of step 7 (predicted residual rate and uncertainty):

[0200]

[0201] Figure 7 It shows the contribution of influencing factors to the residual value rate.

[0202] Implementation example of core step 8: generating predicted estimates and verification results.

[0203] Battery purchase price Yuan, with a projected valuation of:

[0204]

[0205] The uncertainty interval is:

[0206]

[0207] Declaration of Valuation The element falls within the uncertainty interval The verification result is passed. Deviation rate calculation:

[0208]

[0209] Table 12 Output data of step 8 (prediction estimates and verification results):

[0210]

[0211] The system synchronously outputs the contribution of each influencing factor:

[0212] Main effect contribution of cycle number Temperature main effect contribution Maintaining the contribution of the main effect The main effect contribution of depth of discharge Circular-deep interaction item contribution Analysis shows that the number of cycles and the depth of discharge are the main factors affecting the residual value of the battery.

[0213] Figure 8 It displays the residual value prediction results, the declared estimated value, and the verification results.

[0214] The data flow throughout the entire implementation process reflects a complete link from attribute characteristics to residual value prediction: First, the original attribute data of newly added lithium iron phosphate power batteries is encoded and normalized into... 3D attribute feature vector This vector corresponds to the mature material category library. Calculate the cosine similarity of the attribute feature vectors of each category, and filter out... The most similar reference categories; the parameter sets of the multiple regression models for the reference categories are weighted and fused according to similarity to generate a result containing... The prior mean vector of each parameter component and The prior covariance matrix This prior distribution transfers the ability to model multiple factors for similar materials; at the same time, The original influencing factor values ​​of each historical disposal record were normalized and expanded using interaction terms, and then transformed into... indivual The multivariate sample dataset is composed of multivariate feature vectors and their corresponding actual residual rates; the prior distribution and sample data are combined using a Bayesian fusion formula to generate the posterior mean. and posterior covariance The fusion process adaptively balances the weights of prior knowledge (dominant factor) and limited sample information (auxiliary correction); the measured influencing factor values ​​of the battery to be valued are converted to the same normalization process. 3D feature vector The predicted residual rate is obtained by multiplying the residual rate by the inner product of the posterior mean and the mean. The prediction uncertainty is calculated through covariance propagation. The final predicted residual value rate is multiplied by the purchase price to generate the predicted valuation. element and uncertainty interval Yuan, declared valuation The verification is completed by comparing the element with the interval.

[0215] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for automatically summarizing material disposal status and verifying asset valuation, characterized in that, Includes the following steps: Obtain the attribute feature vector of newly received materials. The attribute feature vector includes the material category, technical parameters, applicable scenarios, and a list of influencing factors. Obtain the attribute feature vectors of each category in the mature material category library and the parameter set of the trained multivariate regression model. The parameter set of the multivariate regression model includes the regression coefficients of each factor and the interaction term parameters. Calculate the similarity between the attribute feature vector of newly added materials and the attribute feature vector of each category in the mature material category library, and select the top K categories by similarity ranking as the reference category set; The parameter sets of the multiple regression models for each category in the reference category set are weighted and fused to generate the prior parameter distribution of the multiple regression model for newly entered materials. The fusion weight is positively correlated with the similarity. Obtain historical disposal records of newly received materials, extract multi-factor feature values ​​and actual residual value rates from each historical disposal record, and generate a multivariate sample dataset. A likelihood function is constructed based on a multivariate sample dataset. The likelihood function is then fused with the prior parameter distribution using Bayesian methods to generate the posterior regression parameter distribution. Obtain the multi-factor feature vector of the material to be valued, substitute the multi-factor feature vector into the posterior mean parameter of the posterior regression parameter distribution, and calculate the predicted residual value rate. A predicted valuation is generated based on the predicted residual value rate and the purchase price of the materials to be valued. The declared valuation is then compared with the predicted valuation, and the verification result is output.

2. The method for automatically summarizing material disposal status and verifying asset valuation according to claim 1, characterized in that, The attribute feature vector is constructed as follows: one-hot encoding is performed on categorical attributes, and mean normalization based on range is applied to numerical attributes. The encoded attribute values ​​are then concatenated to form a vector of fixed dimensions. The one-hot encoding is implemented as follows: for a categorical attribute with M possible values, a binary vector of length M is constructed, and a value of 1 is assigned to the position corresponding to the actual value of the categorical attribute, while a value of 0 is assigned to the other positions.

3. The method for automatically summarizing material disposal status and verifying asset valuation according to claim 1, characterized in that, The residual rate output of the multiple regression model consists of the intercept term, the sum of the products of the main effect terms of each influencing factor and their corresponding regression coefficients, the sum of the products of the interaction terms between any two influencing factors and their corresponding interaction term parameters, and the random error term. Mature material categories refer to material categories whose number of historical disposal records reaches the preset sample size threshold, which is 5 to 10 times the total number of parameters in the parameter set of the multiple regression model.

4. The method for automatically summarizing material disposal status and verifying asset valuation according to claim 1, characterized in that, Similarity is calculated using cosine similarity. When calculating similarity, each dimension of the attribute feature vector is weighted. The weights are obtained by averaging the absolute values ​​of the regression coefficients of the multiple regression models for each category in the mature material category library and the absolute values ​​of the regression coefficients corresponding to the same attribute dimension. The number of reference categories K is an integer between 3 and 10, determined according to the size of the mature material category library.

5. The method for automatically summarizing material disposal status and verifying asset valuation according to claim 1, characterized in that, The prior parameter distribution adopts a multivariate normal distribution. The prior mean is obtained by weighting the parameter sets of the multivariate regression models of each category in the reference category set according to the fusion weight. The fusion weight is obtained by normalizing the similarity between each reference category and the newly entered materials. The prior covariance matrix is ​​obtained by weighting the sum of the outer products of the differences between the parameter sets of the multivariate regression models of each reference category and the prior mean according to the fusion weight, and then adding a regularization term. The regularization term is the product of the regularization coefficient and the identity matrix.

6. The method for automatically summarizing material disposal status and verifying asset valuation according to claim 1, characterized in that, The multi-factor feature values ​​include service life, usage intensity index, maintenance score and environmental factors. The multi-factor feature vector is constructed as follows: the normalized values ​​of each influencing factor are concatenated in a preset order as the main effect term, the product between any two influencing factors is calculated as the interaction term and concatenated, and a constant term is added to the first and second parts of the multi-factor feature vector to correspond to the intercept parameter.

7. The method for automatically summarizing material disposal status and verifying asset valuation according to claim 1, characterized in that, The posterior regression parameters are distributed as a multivariate normal distribution. The posterior covariance matrix is ​​obtained by adding the inverse of the prior covariance matrix and the sum of the outer products of the multivariate sample dataset's eigenvectors, divided by the residual variance, and then taking the inverse of the result. The posterior mean is obtained by multiplying the posterior covariance matrix by the inverse of the prior covariance matrix and the prior mean, and then adding the product of the posterior covariance matrix and the sum of the multiplications of the eigenvectors of each multivariate factor by the corresponding actual residual rates, divided by the residual variance.

8. The method for automatically summarizing material disposal status and verifying asset valuation according to claim 1, characterized in that, The prediction uncertainty is calculated based on the posterior covariance matrix of the posterior regression parameter distribution. The prediction uncertainty is the square root of the product of the multifactor eigenvector of the material to be valued, the posterior covariance matrix, and the transpose of the multifactor eigenvector. The uncertainty range is determined by the product of the prediction uncertainty and the purchase price. If the declared valuation falls within the uncertainty range, the verification result is passed. If the declared valuation exceeds the uncertainty range, the verification result is pending review.

9. The method for automatically summarizing material disposal status and verifying asset valuation according to claim 1, characterized in that, When outputting the verification results, the contribution of each influencing factor to the predicted residual value rate is output simultaneously. For the j-th influencing factor, the main effect contribution is the product of the regression coefficient corresponding to the j-th influencing factor in the posterior mean vector and the normalized value of the j-th influencing factor of the material to be valued. The interaction term contribution is the sum of the interaction term parameters in which the j-th influencing factor participates and the product of the normalized values ​​of the corresponding two influencing factors. The total contribution is the sum of the main effect contribution and the interaction term contribution.

10. A system for automatically summarizing material disposal status and verifying asset valuation, used to execute the method for automatically summarizing material disposal status and verifying asset valuation as described in any one of claims 1 to 9, characterized in that, include: The attribute feature acquisition module is used to acquire the attribute feature vectors of newly received materials. The mature category library management module is used to obtain the attribute feature vectors of each category in the mature material category library and the parameter set of the trained multivariate regression model; The similarity calculation and filtering module is used to calculate the similarity between the attribute feature vector of newly added materials and the attribute feature vector of each category in the mature material category library, and to filter the reference category set. The prior parameter fusion module is used to perform weighted fusion of the parameter sets of multiple regression models for each category in the reference category set to generate the prior parameter distribution. The sample dataset generation module is used to obtain historical disposal records of newly received materials and generate a multivariate sample dataset. The Bayesian fusion module is used to construct a likelihood function based on a multivariate sample dataset, and then perform Bayesian fusion of the likelihood function with the prior parameter distribution to generate the posterior regression parameter distribution. The residual value prediction module is used to substitute the multi-factor feature vector of the material to be valued into the posterior mean parameter of the posterior regression parameter distribution to calculate the predicted residual value rate. The valuation verification module is used to generate a predicted valuation based on the predicted residual value rate and the purchase price, compare the declared valuation with the predicted valuation, and output the verification result.

Images