Large model-based waste and old material treatment plan examination method and system, terminal and medium

Abstract

The invention relates to the field of industrial material management, in particular to a waste and old material treatment plan review method and system based on a large model, a terminal and a medium, and the method comprises the steps: calling a pre-trained prediction model to generate a reference treatment plan; a declaration document text, material image features and structured business data are obtained through OCR, computer vision and a data interface; executing multi-modal comparison review, inputting the data and the reference plan into a field large model together for semantic analysis to obtain a first review result, and inputting a quantitative index into a rule engine for logic and deviation verification to obtain a second review result; and generating an examination conclusion according to the two results, and calling a robot process automation service to execute corresponding examination and approval operation. According to the method, the review dimension and conclusion reliability is improved, and then the accuracy and efficiency of review are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial materials management, specifically to a method, system, terminal, and medium for reviewing waste material disposal plans based on a large model. Background Technology

[0002] In the power industry's asset management system, the disposal of obsolete materials generated from decommissioned projects is a crucial step in realizing the full life-cycle value recovery of assets and preventing asset loss. As the starting point of the disposal process, the review of disposal plans suffers from poor adaptability and low accuracy when facing complex and unstructured business scenarios.

[0003] The existing technologies are based on rudimentary automated systems using simple rule engines and keyword matching. While these systems automate some review points through pre-defined rigid business rules and keyword scanning of documents, they cannot understand the contextual semantics of the application materials. For aspects requiring cognitive judgment, such as assessing the sufficiency of the justifications, manual intervention is still necessary. For text recognized by OCR, only simple pattern matching is possible, without information association, affecting the accuracy of the review results. Furthermore, any changes to the business rules require rewriting and redeploying the code, resulting in high maintenance costs. Related solutions use classification or regression models to learn from historical data to predict specific review points or quantitative indicators; however, these models struggle to achieve cross-modal and cross-data source information fusion and cross-validation, similarly impacting the accuracy of the review results. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a method, system, terminal, and medium for reviewing waste material disposal plans based on a large model, thereby enhancing the review dimensions and the reliability of the conclusions, and ultimately improving the accuracy and efficiency of the review process.

[0005] The technical solution of this invention provides a method for reviewing waste material disposal plans based on a large model, comprising the following steps: The pre-trained prediction model is invoked, and a baseline disposal plan is output based on the input project decommissioning data and asset ledger data. The text content of the declaration document is obtained through the optical character recognition interface, the feature data of the image of the scrapped material is obtained through the computer vision model interface, and the relevant structured business data is retrieved from the business database through the data interface. Performing multimodal comparative review includes: inputting the text content, feature data, and structured business data, along with the benchmark handling plan, into a pre-trained domain large model to obtain a first semantic review result; inputting the quantitative indicators in the text content, the quantitative indicators in the structured business data, and the estimated indicators in the benchmark handling plan into a rule engine for logical operations and deviation comparison to obtain a second rule review result. Based on the first semantic review result and the second rule review result, a review conclusion is generated; based on the review conclusion, the robot process automation service is invoked to execute the corresponding approval process operation.

[0006] As can be seen from the above technical solutions, this application has the following advantages: First, by using a domain-wide model trained with industry knowledge, the system possesses deep semantic understanding and logical reasoning capabilities, enabling it to handle cognitive tasks such as rationality and consistency, thus solving the problem that related automated systems can only handle rigid rules and shallow information; Second, by collaboratively analyzing OCR text, visual images, structured business data, and the baseline handling plan generated by the system, it achieves mutual verification of images and text, plan comparison, and the fusion processing of multi-source heterogeneous information, improving the comprehensiveness of the review dimensions and enhancing the reliability of conclusions; Third, through the parallel operation and intelligent agent scheduling of the domain-wide model and the rule engine, the domain-wide model achieves flexible cognition, while the rule engine achieves rigid calculation, realizing the unification of review rules in both flexible understanding and rigid execution, thereby improving the accuracy and efficiency of the review. Attached Figure Description

[0007] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a schematic diagram of a method for reviewing waste material disposal plans based on a large model, provided as an embodiment of the present invention.

[0009] Figure 2 This is a schematic block diagram of a waste material disposal plan review system based on a large model, provided as an embodiment of the present invention.

[0010] Figure 3 This is a schematic diagram of the structure of a terminal provided in an embodiment of the present invention. Detailed Implementation

[0011] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0013] Figure 1 This is a schematic flowchart illustrating a method for reviewing waste material disposal plans based on a large model, provided as an embodiment of the present invention. Figure 1 The executing entity can be a large-scale model-based waste material disposal plan review system. The large-scale model-based waste material disposal plan review method provided in this embodiment of the invention is executed by computer equipment; correspondingly, the large-scale model-based waste material disposal plan review system runs on the computer equipment. Depending on different needs, the order of steps in this flowchart can be changed, and some steps can be omitted.

[0014] like Figure 1 As shown, the method includes the following steps.

[0015] S1 calls a pre-trained prediction model and outputs a baseline disposal plan based on the input project decommissioning data and asset ledger data.

[0016] S2 obtains the text content of the declaration document through the optical character recognition interface, obtains the feature data of the waste disposal material image through the computer vision model interface, and retrieves the relevant structured business data from the business database through the data interface.

[0017] S3, Perform multimodal comparison review, including: inputting the text content, feature data, and structured business data, together with the benchmark handling plan, into a pre-trained domain large model to obtain a first semantic review result; inputting the quantitative indicators in the text content, the quantitative indicators in the structured business data, and the estimated indicators in the benchmark handling plan into a rule engine for logical operation and deviation comparison to obtain a second rule review result.

[0018] S4. Generate a review conclusion based on the first semantic review result and the second rule review result; based on the review conclusion, call the robot process automation service to execute the corresponding approval process operation.

[0019] As a refinement and extension of the specific implementation of the above embodiments, in order to fully illustrate the specific implementation process of this embodiment, another embodiment of the waste material disposal plan review method based on a large model is provided, which includes the following steps.

[0020] S101, Generate a baseline processing plan.

[0021] This step calls a pre-trained prediction model, which outputs a baseline disposal plan based on the input project decommissioning data and asset ledger data. Specifically, it includes the following sub-steps.

[0022] S101.1 Extract structured project decommissioning data and asset ledger data from the asset management system and project management system through a preset data interface; clean and standardize the extracted data to generate standardized material feature vectors and project time sequence features.

[0023] Asset ledger data is extracted from the asset management system and project decommissioning data is extracted from the project management system via a pre-defined RESTful API or direct database connection. An incremental extraction strategy is used during the extraction process, retrieving only data added or changed since the last extraction, reducing data transmission volume and processing time.

[0024] The asset ledger data includes fields such as asset code, material category, original purchase cost, service life, depreciation rate, and current status. Project decommissioning data includes fields such as decommissioning project number, decommissioning time, decommissioning material list, reason for decommissioning, and on-site inspection records.

[0025] Outliers, missing values, and duplicate values ​​in the extracted data are handled as follows: The first step is outlier handling, using the IQR (Interquartile Range) method to identify outliers in quantified fields. For example, if the "service life" of an asset exceeds three times the average service life of similar materials, it is considered an outlier. Corrections are made by querying the original survey records; if no correction basis exists, it is marked as "pending manual confirmation" and temporarily stored in the outlier data pool. The second step is missing value handling. For missing values ​​in critical fields, such as asset code and material category, they are filled by referring to business system logs or calling historical data completion interfaces. For missing values ​​in non-critical fields, the default value of "no supplementary explanation" is used. The third step is duplicate value handling. Based on the composite primary key of "asset code + project number," duplicate data is identified and deleted, retaining only the most recent data record.

[0026] Next, data standardization was performed, converting non-uniformly formatted fields to preset standards. For example, expressions such as "cable," "power cable," and "high-voltage cable" in the "material category" field were uniformly mapped to the "020101-power cable" code in the material classification standard; the "retirement time" field was uniformly converted to the time format "YYYY-MM-DDHH:MM:SS". The standardized asset ledger data was converted into material feature vectors, such as [asset code hash value, material category code, normalized value of service life, normalized value of depreciation rate, normalized value of original procurement cost], and project retirement data was converted into project time-series features, such as [project retirement timestamp, quantity of retired materials, retirement batch number, time interval since the last retirement of a similar project]. The normalization process used the Min-Max standardization formula, mapping the quantitative indicators to the [0,1] interval.

[0027] S101.2, the standardized material feature vector and project time series features are input into a pre-trained prediction model. The prediction model is trained by machine learning, with the standardized material feature vector and project time series features corresponding to historical project data as input and the material weight and disposal method in the approved historical disposal plan as the supervision target.

[0028] The standardized material feature vector generated by S101.1 is concatenated with the project time series features to form a joint input vector of length n+m, where n is the dimension of the material feature vector and m is the dimension of the project time series features.

[0029] The pre-trained prediction model used in this step is a hybrid model architecture of Multilayer Perceptron (MLP) + Gradient Boosting Tree (XGBoost): The bottom MLP module contains three hidden layers. The number of neurons in the first layer is twice the dimension of the input vector, the second layer is half the dimension of the first layer, and the third layer is half the dimension of the second layer. The activation function is ReLU, which is used to learn non-linear features in the data, such as the implicit correlation between material category and disposal method. The top XGBoost module integrates the non-linear features output by the MLP with linear features such as service life and depreciation rate for ensemble learning, focusing on optimizing the prediction accuracy of quantitative indicators, such as optimizing the prediction accuracy of estimated weight.

[0030] After receiving the joint input vector, the model performs forward propagation computation. First, the MLP module performs a nonlinear transformation on the input vector, outputting a 20-dimensional intermediate feature vector. This intermediate feature vector is then input into the XGBoost module, which outputs three core prediction results through a voting mechanism using multiple decision trees. Material Classification List: Based on the category code in the material feature vector and historical disposal data, output the sub-classification corresponding to the decommissioned material, such as "02010101-10kV power cable"; Estimated weight: Combining material category, service life, original dimensions obtained from the asset ledger's associated fields, and historical data on the decommissioning weight of similar materials, the estimated weight is output with an error range controlled within ±5%. Recommended disposal methods: Based on the status of the materials (including "repairable", "disassembled and recyclable", "non-recyclable and must be destroyed"), environmental protection requirements, and historical disposal revenue data, the recommended disposal methods are output, including "public auction", "targeted recycling" and "environmentally friendly destruction".

[0031] S101.3 Receive the prediction results output by the pre-trained prediction model, the prediction results including a material classification list, estimated weight and recommended disposal method; combine the prediction results with the basic project information and fill them into a standard document template to generate the baseline disposal plan.

[0032] The material classification list, estimated weight, and recommended disposal method output by S101.2 are associated and combined with the basic project information, which may include the project name, construction unit, and decommissioning approval number, to form a complete set of basic data for the disposal plan, as shown in Table 1.

[0033] Table 1: Basic Data Set for the Disposal Plan

[0034] The system calls a preset standard template for waste material disposal plans and automatically fills the corresponding positions in the template with the aforementioned basic data set through the template engine.

[0035] The training process of the prediction model is based on historical business data and is carried out in four stages: data preparation, model building, training optimization, and evaluation and verification, as detailed below.

[0036] (1) Training data preparation Data collected from waste material disposal projects completed in recent years can cover 2,000 decommissioning projects and 50,000 asset ledger records. Each data entry contains two types of information: input characteristics and monitoring targets.

[0037] The input features are consistent with the material feature vector, project time series features, and structure in S101.1, with a total of 9 dimensions.

[0038] The monitoring targets are based on three types of key data from historical handling plans approved by the business departments: Material classification label: adopts the 6-level coding of the material classification standard, such as "02010101-10kV power cable"; Actual weight: Weighing data during the disposal of decommissioned materials (unit: kg); Disposal method labels: The actual disposal method determined by approval, such as "public auction", "targeted recycling" and "environmentally friendly destruction", totaling 5 types of labels.

[0039] The dataset was divided into training, validation, and test sets in a 7:2:1 ratio. Stratified sampling was used during the division process to ensure that the distribution ratio of each material category and disposal method was consistent across the three datasets, thus avoiding data skew that could affect the model's generalization ability.

[0040] For material categories with a small number of samples in the training set, SMOTE (Synthetic Minority Oversampling) is used to generate virtual samples: based on the K nearest neighbors of minority class samples, new sample features are generated through linear interpolation, while keeping the supervision target consistent with the original samples, so that the difference in sample size between each category is controlled within 10 times.

[0041] (2) Model architecture construction The prediction model adopts a multi-task learning architecture, which simultaneously optimizes three tasks: material classification, weight prediction, and disposal method recommendation. The architecture is divided into a shared layer and a task-specific layer.

[0042] The shared layer, also known as the underlying module of the MLP mentioned in S101.2, contains three hidden layers with ReLU activation function. It is used to learn the common features of the three types of tasks, and the parameters of the shared layer are updated together during the training of the three types of tasks.

[0043] The task-specific layers include a material classification task layer, a weight prediction task layer, and a disposal method recommendation task layer. The material classification task layer, after the output of the shared layer, is connected to a fully connected layer and a Softmax layer, outputting the probability distribution of various material classification labels. The weight prediction task layer, after the output of the shared layer, is connected to a fully connected layer and a linear output layer, outputting continuous values ​​of the estimated weight. The disposal method recommendation task layer, after the output of the shared layer, is connected to a fully connected layer and a Softmax layer, outputting the probability distribution of various disposal methods.

[0044] (3) Loss function design Since the model employs a multi-task learning architecture, a weighted summation joint loss function is used. The calculation formula is as follows:

[0045] Wherein, α, β, and γ are task weights, and the weight allocation is based on the degree of impact of the task on the baseline plan.

[0046] The loss function for the material classification task is the cross-entropy loss function, calculated using the following formula:

[0047] Where N is the number of samples, C1 is the number of material classification categories, is the true label of sample i belonging to category c, and is the probability predicted by the model that sample i belongs to category c.

[0048] For the weight prediction task, the loss function is the root mean square error (RMSE) loss function, calculated as follows:

[0049] in, The actual weight of sample i. This is the weight predicted by the model.

[0050] The recommended loss function for the disposal method is the weighted cross-entropy loss function, calculated as follows:

[0051] Where D represents the number of disposal method categories, The weight of category d, For sample i, the true label belongs to treatment method d. Predict probabilities for the model.

[0052] (4) Model training and optimization The Adam optimizer is used with an initial learning rate of 1e-3 and a learning rate decay strategy of "decreasing to 0.8 times the previous learning rate every 5 epochs"; the batch size is 32; the number of iterations is 50, and an early stopping strategy is adopted, whereby the learning rate decreases after 5 epochs. Training is stopped if the performance does not decrease for three consecutive epochs to avoid overfitting. L2 regularization is added to both the shared layer and the task-specific layer, and Dropout layers are added to the MLP hidden layers to reduce overfitting.

[0053] The training process includes initializing model parameters, inputting the training set into the model in batches, calculating the loss of each task and the joint loss, calculating the gradient through backpropagation (BP algorithm), and updating the model parameters using the Adam optimizer. After each epoch, the metrics of each task are calculated on the validation set, and the optimal model parameters are recorded. After training, the optimal parameters of the validation set are loaded to obtain the final pre-trained prediction model.

[0054] (5) Model evaluation and validation The performance of the pre-trained model was evaluated using a test set. The evaluation metric for the material classification task was accuracy, the evaluation metric for the weight prediction task was RMSE, the evaluation metric for the disposal method recommendation task was accuracy, and the evaluation metric for the joint task was joint loss.

[0055] S102, Project Data Acquisition.

[0056] This step obtains the text content of the declaration document through the optical character recognition interface, acquires the feature data of the waste disposal material image through the computer vision model interface, and retrieves the relevant structured business data from the business database through the data interface.

[0057] S102.1, Obtaining the text content of the application document.

[0058] This step involves extracting text information from unstructured application documents such as transfer lists, assessment reports, and disposal applications using an optical character recognition (OCR) interface to ensure that no key content is omitted and that the recognition accuracy meets the standards.

[0059] After receiving the application documents uploaded by users, the system converts non-PDF documents into standard PDF format through a document conversion interface to avoid recognition errors caused by format differences. For the scanned PDF documents, the system calls the image preprocessing module to perform noise reduction, tilt correction, and contrast enhancement operations to provide high-quality image input for OCR recognition.

[0060] The PDF document is partitioned according to business logic. An OCR interface is used to individually recognize each area, outputting the corresponding text content and its coordinates within the document. For example, the transfer list is divided into an asset basic information area, a material details area, and a signature confirmation area; the appraisal report is divided into an appraisal target area, a value estimation area, and an appraisal conclusion area. Tables, formulas, and stamped information in the document are automatically marked as "special content blocks" by the OCR interface. Table content is extracted as structured table data, formulas are temporarily stored as image blocks, and stamped information is extracted by extracting the text content of the stamp and recording its position.

[0061] Post-processing is then performed, with semantic verification of the OCR recognition results based on a power industry terminology database. Redundant information such as whitespace characters, repeated line breaks, headers, and footers are removed, while core text relevant to the review of the disposal plan is retained. The processed text content is then stored in an unstructured database according to a hierarchical structure of "document type-region name-text content-coordinate location," and a unique text identifier is generated.

[0062] S102.2 Acquisition of image feature data of scrapped materials.

[0063] This step involves taking images of the waste disposal site and extracting visual feature data through a computer vision model interface to establish a basis for image-text verification.

[0064] Employing a dedicated computer vision model, which can be trained on a ResNet-50 backbone network, the training dataset contains over 100,000 images of waste electrical materials. The model supports three core tasks: material classification, defect detection, and nameplate parameter recognition, and the interface outputs feature data for the corresponding tasks.

[0065] The model performs category identification on materials in images, outputting category labels and confidence scores, and extracting 2048-dimensional visual feature vectors. The model automatically detects material defect types, outputting defect region coordinates, defect category labels, and confidence scores, and generating local feature vectors for the defect regions. The model performs character recognition and structured parsing on nameplate images, extracting parameters such as material model, rated voltage, rated current, and production date, and outputting parameter key-value pairs and corresponding visual feature vectors.

[0066] S102.3, Retrieval of Structured Business Data.

[0067] This step retrieves structured data associated with the disposal plan from the State Grid business database through a standardized data interface, thus establishing the basis for verifying consistency between accounts and actual data.

[0068] Based on the "project number" and "asset code" extracted from the declaration documents as related keywords, a precise data query is initiated to the business system to ensure that the retrieved data corresponds one-to-one with the current disposal plan and to avoid data confusion.

[0069] Retrieve key structured fields related to the review from various business systems, as follows: Asset Management System: Asset code, material name, specifications, original purchase cost, service life, depreciation amount, current asset status (e.g., "pending retirement" or "retired"), and the unit to which the asset belongs; Project Management System: Project Number, Project Name, Decommissioning Application Time, Decommissioning Approval Status, Site Survey Record Number, Survey Personnel Information; Financial system: asset book value, accumulated depreciation, estimated residual value, disposal cost budget, and historical disposal revenue records.

[0070] S103, Perform multimodal comparison review.

[0071] First, a first semantic review is performed using a domain-wide large model, followed by a second rule review using a rule engine.

[0072] S103.1, The text content, feature data, and structured business data, together with the baseline processing plan, are input into the pre-trained domain large model to obtain the first semantic review result.

[0073] S103.11, the text content of the application document and the text content of the benchmark processing plan are concatenated to form a joint text sequence. The joint text sequence is then processed by the text encoder of the domain large model to obtain a text feature vector sequence.

[0074] This step involves text encoding, represented as follows:

[0075] in, The text content of the application document, The text content of the baseline disposal plan, For text encoders of large domain models, To obtain the text feature vector sequence.

[0076] S103.12, the visual feature data extracted from the image of discarded materials is mapped to the same semantic space as the text features through a projection layer to obtain a sequence of visual feature vectors.

[0077] For example, vectors representing the degree of damage and model characteristics. Through a projection layer Mapping it to the same semantic space as the text features, it is represented as:

[0078] in, The visual feature vector sequence is obtained.

[0079] S103.13 After converting structured business data into natural language description fragments, it is encoded through an independent embedding layer to obtain a sequence of structured data feature vectors.

[0080] Independent Embedding Layer For structured business data The encoding representation is as follows:

[0081] in, To obtain the sequence of feature vectors for structured data.

[0082] S103.14, fuse the text feature vector sequence, visual feature vector sequence, and structured data feature vector sequence to generate a fused context representation.

[0083] For example, using text features As a query, visual and structured features serve as keys and values, generating the final fusion context representation C, denoted as:

[0084] S103.15, Construct prompts that include review instructions and relevant context.

[0085] For example: "Based on the following baseline plan: [\(T_{base}\)] and the user-submitted plan: [\(T_{submit}\)], and the material image description: [\(H_{vision}\)], please determine: 1. Is the handling of the user's plan reasonable? 2. Is the declared material description consistent with the image evidence?" S103.16, The fused context representation and the constructed prompt words are input together into the Transformer-based decoder of the domain large model, and a natural language inference chain and preliminary conclusion for each review point are generated through autoregression.

[0086] S103.17, the preliminary conclusions are structured and analyzed to form the first semantic review result.

[0087] S103.17a, the preliminary conclusion is semantically labeled using a sequence labeling model to extract predefined review elements, which include at least review items, judgment results, and evidence entities.

[0088] The system provides a pre-defined structured review element schemaS, which defines the fields to be extracted, for example:

[0089] A lightweight sequence labeling model M is adopted. tag (e.g., a BERT-based token classifier) ​​Preliminary conclusions regarding the generated data. Each token in the model is labeled. The model assigns each token to a predefined set of labels. = In the middle. Let the input sequence be... Model output labeled sequence ,in .

[0090] The probability of the model assigning a label to each token is:

[0091] in, It’s Token In model M tag The context vector representation in the text.

[0092] Based on the labeled sequence Y, key-value pairs and entity relationships conforming to schemaS are extracted using rule templates or models. For example, triples (consistency of treatment method, judgment result, inconsistency) and (consistency of treatment method, evidence entity, assessment report page X) can be extracted.

[0093] S103.17b maps the extracted text-type judgment results to a state in a predefined set of standardized states using a normalization function.

[0094] Define a normalization function This maps text-based "judgment results" to a finite, predefined set of states. In, for example:

[0095] S103.17c, Based on the confidence level extracted from the preliminary conclusion and the predicted probability of the sequence labeling model, the final confidence level is generated by weighted calculation and fusion.

[0096] The final confidence level for each review item It is a combination of the model generation probability and the extraction probability. Let's assume that from... The numerical value corresponding to the original confidence score text extracted from it is The sequence labeling model has an average prediction probability of 100% for key decision words. The final confidence level can then be calculated using a weighted geometric mean:

[0097] This is an adjustable hyperparameter used to balance the weights of confidence levels from different sources.

[0098] S103.17d, the review items, standardized judgment results, final confidence level and evidence entities are assembled into a structured data object to form the first semantic review result.

[0099] The processed information is then assembled into the final first semantic review result. Specifically, the normalized review items, judgment results, confidence levels, evidence, etc., are assembled into a structured data object.

[0100] Through the above process, multi-source heterogeneous data is compared, reasoned, and judged at a deep semantic level with the benchmark plan, and finally a structured and quantifiable first semantic review result is output.

[0101] This embodiment inputs the baseline disposal plan into the domain-wide model, providing a reliable and compliant reference benchmark. This transforms the review task from an absolute judgment to a difference analysis. The baseline disposal plan is an ideal template generated by the system based on historical data and rules. It represents a compliant, reasonable, and optimized disposal solution under existing knowledge and rules, providing a basis for judgment for the large model. For example, the baseline plan suggests "auctioning the transformer as a whole," while the user plan is "disassembling and then auctioning." The large model, combining both, can determine whether this change is due to "the on-site assessment found that the transformer is severely damaged and does not have value for overall auction" (reasonable) or "low-value disposal without justifiable reason" (questionable). Moreover, the generated review conclusion can provide specific and verifiable reasons based on the differences from the benchmark. For example, the first semantic review result could be: "Not approved. Reason: The submitted estimated weight (5 tons) deviates from the baseline plan's predicted value (8 tons) by more than the threshold (±15%), and the submitter has not provided a reasonable explanation in the description." The baseline plan is generated for specific projects, making the review standards dynamic and personalized. For projects of different sizes and types, the system automatically adjusts its review focus. Furthermore, the model does not need to have seen all possible correct and incorrect cases in the training data. It only needs to learn how to compare and analyze. As long as the baseline plan is relatively reliable, the model can discover anomalies through comparison, which reduces the dependence on massive, precisely labeled error case data and improves the system's usability and buildability.

[0102] S103.2, input the quantitative indicators in the text content, the quantitative indicators in the structured business data, and the estimated indicators in the benchmark handling plan into the rule engine for logical operation and deviation comparison to obtain the second rule review result.

[0103] S103.21, Extract the set of quantitative indicators submitted by the user from the text content and structured business data.

[0104] S103.22 Input the set of quantitative indicators submitted by the user and the set of predicted indicators in the benchmark treatment plan into the rule engine.

[0105] S103.23, The rule engine loads a preset business rule set and performs logical operations on the input quantitative indicators. The logical operations include at least: absolute rule verification based on a preset fixed threshold, and deviation rule verification that calculates the relative deviation between the user-submitted indicator and the benchmark estimated indicator and compares the relative deviation with a preset deviation threshold.

[0106] S103.24, summarize the execution status of all rules and generate a structured second rule review result, which includes the pass status, actual value, expected value and deviation value of each rule.

[0107] Specifically, the first step is to accurately extract the quantitative indicators that require logical verification from multi-source data and convert them into a unified numerical format for calculation. Then, using pre-configured parsing rules, the target quantitative indicators are retrieved from different sources. Text content recognized by OCR In the process, the set of quantitative indicators declared by users is extracted through regular expressions or keyword matching. For example, declaring weight and declaring residual value; From structured business data In the middle, directly read the relevant set of quantitative indicators. For example, the original value of the asset and its book value; Obtain the estimated set of quantitative indicators from the baseline treatment plan. For example, estimating weight and estimating residual value.

[0108] All extracted indicator values ​​are standardized to the same units and dimensions, and null or outlier values ​​are handled. For indicators that cannot be directly compared, conversion is performed using predefined conversion formulas.

[0109] The rules engine loads the pre-defined business rule set. The extracted indicators are subjected to Boolean logic operations and arithmetic deviation comparisons, including absolute rule verification and relative deviation verification.

[0110] Absolute rule validation performs compliance checks based on fixed thresholds. For example, it checks whether the residual value rate is within a specified range.

[0111] in, This is the maximum residual value rate threshold stipulated by the system. The rule output is a Boolean value. .

[0112] Relative deviation verification will check the metrics submitted by the user. Projected targets compared to the baseline plan Perform a difference comparison. First, calculate the absolute deviation. and relative deviation :

[0113] Compare the calculated deviation with the dynamic or static threshold. For comparison, the threshold can be fixed or dynamically adjusted based on the importance of the indicators:

[0114] The output of this rule is also a Boolean value, along with the specific deviation value. .

[0115] Combining multiple basic rules into complex business rules using logical operators:

[0116] This is used to express complex business logic such as "when the weight deviation exceeds the threshold, unless there is a special approved instruction, it will be judged as unacceptable".

[0117] The rule engine aggregates the execution results of all rules and generates a structured second rule review result. , It is a collection containing all rule validation details and their conclusions. For each rule... Output a result tuple:

[0118] PassStatus is a boolean value indicating whether the rule is passed.

[0119] The overall conclusion of the second rule review is determined by the pass status of all rules, using either a "one-vote veto" logic (i.e., all rules must pass) or a weighted voting logic: This is represented as:

[0120] in, This represents a logical AND operation, where N is the total number of rules.

[0121] S104, Generate review conclusions and execute them.

[0122] S104.1, Based on the results of the first semantic review and the results of the second rule review, generate a review conclusion.

[0123] S104.11, extract and vectorize the features of the first semantic review result and the second rule review result, and concatenate them to form a comprehensive decision tensor. .

[0124] First, the key features of the two input results are converted into a unified numerical representation that can be used for decision-making.

[0125] from Features are extracted from the data, including one-hot encoding of the state of semantic review items to form vectors. Extract the confidence scores for each item. This constitutes the confidence vector. Ultimately, the comprehensive feature vector of the semantic review result is: .

[0126] from Extracting features, including each rule The state is binary encoded (1 / 0) to form a regular vector. Extract the deviation values ​​of key rules , forming the deviation vector Ultimately, the comprehensive feature vector of the rule review results is: .

[0127] The two eigenvectors are merged into a single comprehensive decision tensor. , represented as:

[0128] in, This indicates a vector concatenation operation.

[0129] S104.12, the comprehensive decision tensor Input to a predefined decision function Calculate the fusion decision score ; where, decision function It is a linear weighted function, and its form is: In the formula, W is the weight vector and b is the bias term.

[0130] The final review conclusion is calculated by a decision function, the input of which is the decision tensor. ,Right now .

[0131] The first m components of the weight vector W (corresponding to) The passing item and the passing item in The sum of the coefficients of the coupling is α, and the subsequent n components (corresponding to) The sum of the coefficients of ) is β.

[0132] S104.13 will integrate decision scores. The results are compared with a preset decision threshold, and a final review conclusion is generated based on the comparison results. The review conclusion includes one of three states: "passed", "failed", and "manual review".

[0133] Integrate decision scores Compared with the preset decision threshold and By comparing the results to determine the final conclusion, it can be represented as follows:

[0134] The final output of the review conclusion is a structured object containing the decision itself and the reasons for its generation.

[0135] S104.2 Based on the review conclusion, invoke the Robotic Process Automation service to execute the corresponding approval process operation.

[0136] S104.21, based on the ruling status in the review conclusion, map it to a predefined robotic process automation operation instruction; the instruction includes at least the operation type, target system, and parameters.

[0137] S104.22 If the decision status is "passed", the Robotic Process Automation Service performs an automatic approval operation and transfers the disposal plan to the next business node.

[0138] Specifically, the RPA robot automatically logs into the business system, clicks the "Approve" button on the approval interface, pushes the process node to the handling department, and attaches a preset explanation in the approval opinion.

[0139] S104.23 If the ruling status is "not approved", the Robotic Process Automation Service performs a rejection operation, generates a rejection reason based on the rule violation list in the review conclusion, and sends a notification to the applicant.

[0140] Specifically, the RPA robot performs the rejection operation and automatically triggers the email system to send the rejection notice and a list of specific reasons to the applicant.

[0141] S104.24 If the adjudication status is "manual review", the Robotic Process Automation Service creates a manual review task in the task management system and attaches the semantic review summary and rule violation list in the review conclusion as task context data to the task.

[0142] Specifically, the RPA robot automatically creates a review work order in the task management system and fills in the work order description with key semantic review summaries and a list of violated rule IDs as attachments, providing decision support for manual review. The task priority is dynamically set based on whether rule violations exist.

[0143] The foregoing has described in detail an embodiment of a method for reviewing waste material disposal plans based on a large model. Based on the above embodiment of the method for reviewing waste material disposal plans based on a large model, this invention also provides a corresponding system for reviewing waste material disposal plans based on a large model.

[0144] Figure 2 This is a schematic block diagram of a waste material disposal plan review system based on a large model, provided as an embodiment of the present invention. In this embodiment, the waste material disposal plan review system 200 based on the large model can be divided into multiple functional modules according to its functions. A module, as referred to in this invention, is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and is stored in memory.

[0145] The baseline disposal plan generation module 210 is used to call a pre-trained prediction model and output a baseline disposal plan based on the input project decommissioning data and asset ledger data.

[0146] The data extraction module 220 is used to obtain the text content of the declaration document through the optical character recognition interface, obtain the feature data of the waste disposal material image through the computer vision model interface, and retrieve relevant structured business data from the business database through the data interface.

[0147] The review result generation module 230 is used to perform multimodal comparison review, including: inputting the text content, feature data, and structured business data together with the benchmark handling plan into a pre-trained domain large model to obtain a first semantic review result; inputting the quantitative indicators in the text content, the quantitative indicators in the structured business data, and the estimated indicators in the benchmark handling plan into the rule engine for logical operation and deviation comparison to obtain a second rule review result.

[0148] The review conclusion generation module 240 is used to generate a review conclusion based on the first semantic review result and the second rule review result; based on the review conclusion, it calls the robot process automation service to execute the corresponding approval process operation.

[0149] The large-model-based waste material disposal plan review system of this embodiment is used to implement the aforementioned large-model-based waste material disposal plan review method. Therefore, the specific implementation of this system can be found in the embodiment section of the large-model-based waste material disposal plan review method above. Thus, its specific implementation can be referred to the description of the corresponding embodiments, and will not be elaborated here.

[0150] Furthermore, since the large-model-based waste material disposal plan review system in this embodiment is used to implement the aforementioned large-model-based waste material disposal plan review method, its function corresponds to the function of the above method, and will not be repeated here.

[0151] Figure 3 This is a schematic diagram of the structure of a terminal 300 provided in an embodiment of the present invention, including: a processor 310, a memory 320, and a communication unit 330. The processor 310 is used to implement the process steps of the above-described embodiment of the waste material disposal plan review method based on a large model when implementing the waste material disposal plan review procedure based on a large model stored in the memory 320.

[0152] This invention also provides a computer storage medium, which may be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. The computer storage medium stores a large-scale model-based waste disposal plan review program. When executed by a processor, the large-scale model-based waste disposal plan review program implements the process steps of the above-described large-scale model-based waste disposal plan review method embodiment.

[0153] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for reviewing waste material disposal plans based on a large model, characterized in that, Includes the following steps: The pre-trained prediction model is invoked, and a baseline disposal plan is output based on the input project decommissioning data and asset ledger data. The text content of the declaration document is obtained through the optical character recognition interface, the feature data of the image of the scrapped material is obtained through the computer vision model interface, and the relevant structured business data is retrieved from the business database through the data interface. Performing multimodal comparative review includes: inputting the text content, feature data, and structured business data, along with the benchmark handling plan, into a pre-trained domain large model to obtain a first semantic review result; inputting the quantitative indicators in the text content, the quantitative indicators in the structured business data, and the estimated indicators in the benchmark handling plan into a rule engine for logical operations and deviation comparison to obtain a second rule review result. Based on the first semantic review result and the second rule review result, a review conclusion is generated; based on the review conclusion, the robot process automation service is invoked to execute the corresponding approval process operation.

2. The method for reviewing waste material disposal plans based on a large model according to claim 1, characterized in that, The pre-trained prediction model is invoked, and based on the input project decommissioning data and asset ledger data, a baseline disposal plan is output, specifically including: Structured project decommissioning data and asset ledger data are extracted from the asset management system and project management system through preset data interfaces; the extracted data is cleaned and standardized to generate standardized material feature vectors and project time-series features. Standardized material feature vectors and project time-series features are input into a pre-trained prediction model. This prediction model is trained through machine learning, with standardized material feature vectors and project time-series features corresponding to historical project data as inputs and the weight and disposal method of materials in the approved historical disposal plan as the supervision target. Receive the prediction results output by the pre-trained prediction model, which include a list of material categories, estimated weight, and recommended disposal methods; combine the prediction results with the basic project information and populate them into a standard document template to generate the baseline disposal plan.

3. The method for reviewing waste material disposal plans based on a large model according to claim 2, characterized in that, The text content, feature data, and structured business data, along with the baseline processing plan, are input into a pre-trained domain-specific large model to obtain the first semantic review result, which specifically includes: The text content of the application document and the text content of the benchmark processing plan are concatenated to form a joint text sequence. The joint text sequence is then processed by the text encoder of the domain large model to obtain a text feature vector sequence. Visual feature data extracted from images of discarded materials are mapped to the same semantic space as text features through a projection layer to obtain a sequence of visual feature vectors. After converting structured business data into natural language description fragments, it is encoded through an independent embedding layer to obtain a sequence of structured data feature vectors. The text feature vector sequence, visual feature vector sequence, and structured data feature vector sequence are fused to generate a fused context representation; Construct prompts that include review instructions and relevant context; The fused context representation and the constructed prompt words are input together into the Transformer-based decoder of the domain large model, and a natural language inference chain and preliminary conclusion for each review point are generated through autoregression. The preliminary conclusions were structured and analyzed to form the first semantic review result.

4. The method for reviewing waste material disposal plans based on a large model according to claim 3, characterized in that, The preliminary conclusions were structured and analyzed to form the first semantic review results, which specifically include: Semantic role labeling is performed on the preliminary conclusions using a sequence labeling model to extract predefined review elements, which include at least review items, judgment results, and evidence entities. The extracted text-based judgment results are mapped to a certain state in a predefined set of standardized states using a normalization function; Based on the confidence level extracted from the preliminary conclusions and the predicted probability of the sequence labeling model, the final confidence level is generated by weighted calculation and fusion. The review items, standardized judgment results, final confidence levels, and evidence entities are assembled into a structured data object to form the first semantic review result.

5. The method for reviewing waste material disposal plans based on a large model according to claim 4, characterized in that, The quantitative indicators in the text content, the quantitative indicators in the structured business data, and the estimated indicators in the benchmark handling plan are input into the rule engine for logical operations and deviation comparisons to obtain the second rule review result, which specifically includes: From the text content and structured business data, a set of quantitative indicators submitted by the user is parsed and extracted; Input the set of quantitative indicators submitted by the user and the set of predicted indicators in the benchmark treatment plan into the rule engine; The rule engine loads a pre-set set of business rules and performs logical operations on the input quantitative indicators. The logical operations include at least: absolute rule verification based on a preset fixed threshold, and deviation rule verification that calculates the relative deviation between the user-submitted indicator and the benchmark estimated indicator and compares the relative deviation with a preset deviation threshold. The execution status of all rules is summarized to generate a structured second rule review result, which includes the pass status, actual value, expected value and deviation value of each rule.

6. The method for reviewing waste material disposal plans based on a large model according to claim 5, characterized in that, Based on the results of the first semantic review and the second rule review, a review conclusion is generated, specifically including: The first semantic review result and the second rule review result are subjected to feature extraction and vectorization, and then concatenated to form a comprehensive decision tensor. ; Comprehensive decision tensor Input to a predefined decision function Calculate the fusion decision score ; where, decision function It is a linear weighted function, and its form is: In the formula, W is the weight vector and b is the bias term; Integrate decision scores The system compares the results with a preset decision threshold and generates a final review conclusion based on the comparison results. The review conclusion includes one of three states: "passed", "failed", or "manual review".

7. The method for reviewing waste material disposal plans based on a large model according to claim 6, characterized in that, Based on this review conclusion, the Robotic Process Automation (RPA) service is invoked to execute the corresponding approval process operations, specifically including: Based on the ruling status in the review conclusion, it is mapped to a predefined robotic process automation (RoLA) operation instruction; the instruction includes at least the operation type, target system, and parameters. If the ruling status is "passed", the Robotic Process Automation Service will perform an automatic approval operation and transfer the disposal plan to the next business node; If the ruling status is "not approved", the Robotic Process Automation Service will perform a rejection operation, generate a rejection reason based on the rule violation list in the review conclusion, and send a notification to the applicant. If the adjudication status is "manual review", the Robotic Process Automation Service creates a manual review task in the task management system and attaches the semantic review summary and rule violation list from the review conclusion as task context data to the task.

8. A waste material disposal plan review system based on a large model, characterized in that, include: The baseline disposal plan generation module is used to call a pre-trained prediction model and output a baseline disposal plan based on the input project decommissioning data and asset ledger data. The data extraction module is used to obtain the text content of the declaration document through the optical character recognition interface, obtain the feature data of the waste disposal material image through the computer vision model interface, and retrieve relevant structured business data from the business database through the data interface. The review result generation module is used to perform multimodal comparative review, including: inputting the text content, feature data, and structured business data together with the benchmark handling plan into a pre-trained domain large model to obtain a first semantic review result; inputting the quantitative indicators in the text content, the quantitative indicators in the structured business data, and the estimated indicators in the benchmark handling plan into the rule engine for logical operation and deviation comparison to obtain a second rule review result; The review conclusion generation module is used to generate a review conclusion based on the first semantic review result and the second rule review result; based on the review conclusion, the robot process automation service is invoked to execute the corresponding approval process operation.

9. A terminal, characterized in that, include: Memory is used to store the review process for waste disposal plans based on large models; A processor is configured to implement the steps of the large-model-based waste disposal plan review method as described in any one of claims 1 to 7 when executing the large-model-based waste disposal plan review procedure.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a large-model-based waste disposal plan review program, which, when executed by a processor, implements the steps of the large-model-based waste disposal plan review method as described in any one of claims 1 to 7.

Images