A classification system for surveying and mapping results incorporating time-dynamic weights and its application methods

Abstract

This invention discloses a surveying and mapping results classification system and its usage method that incorporates time-dynamic weights. The classification system includes a sample weight calculation module, a multi-sub-model training module, and an upper-level fusion decision module. The training method involves the sample weight calculation module assigning a time-dynamic weight to each initial sample based on its temporal attributes to obtain training samples, and then dividing the training samples into multiple independent sub-sample sets according to time periods. The multi-sub-model training module independently configures a classifier algorithm for each sub-sample set and performs parallel training to generate a sub-model corresponding to the time dimension. The upper-level fusion decision module outputs the final classification result. This invention effectively improves the accuracy and timeliness of surveying and mapping results classification: through the design of time-dynamic weights, it strengthens recent surveying and mapping results.

Description

Technical Field

[0001] This invention relates to the field of surveying and mapping data processing technology, specifically to a surveying and mapping results classification system and its usage method that incorporates time-dynamic weights. Background Technology

[0002] The inherent characteristics of surveying and mapping work and the diversity of application needs lead to a complex array of data types. Traditional classification of surveying and mapping results relies on manual annotation, which is inefficient and easily influenced by subjective factors. Existing automatic classification technologies suffer from the following problems: they fail to consider the correlation between surveying and mapping results and projects, contracts, and operating departments, resulting in a lack of business logic support for classification; business connections are missing: existing systems only support fuzzy searches by result name and cannot trace results back through contract numbers or project leader business attributes, leading to data silos; they do not consider the temporal evolution characteristics of historical data (such as differences in surveying and mapping techniques and task objectives over time); and the classification model is outdated and cannot adapt to new data characteristics.

[0003] The prior art CN117688183A discloses a text classification method based on TF-IDF weighting, belonging to the field of semantic network technology. It includes the following steps: A) Calculating the TF-IDF of documents in each category of the dataset; B) Ranking the TF-IDF of the documents; C) Remapping the values ​​based on the TF-IDF ranking information of each word in the document; D) Calculating the similarity score between the current sentence and the category documents; E) Evaluating the classification effect. By calculating word-level category similarity scores to represent the category features of documents, it solves to some extent the problem that TF-IDF features cannot well represent the category information of documents due to differences in data volume or data feature distribution between categories.

[0004] The current classification system has not yet established spatiotemporal coupling coding rules. When facing the actual operation scenarios of multi-period remote sensing monitoring data fusion and BIM+GIS cross-stage modeling, there is both a time sequence gap between historical archived data and newly measured data, and a lack of metadata specifications for multi-source spatiotemporal data fusion. Summary of the Invention

[0005] The purpose of this invention is to provide a classification system and method for surveying and mapping results that incorporates time-dynamic weights, in order to address the aforementioned shortcomings in the prior art.

[0006] This invention relates to the intersection of surveying and mapping geographic information processing and machine learning technologies, specifically providing a surveying and mapping results classification system that integrates multi-source heterogeneous data correlation analysis, dynamic weighting of time-sensitive features, and hierarchical label semantic mapping. By incorporating time weights, a dynamic weighted TF-IDF feature engineering framework is constructed to address the problems of missing business logic, insufficient timeliness of historical data, and weak model generalization ability in traditional surveying and mapping results classification.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A mapping result classification system incorporating time-dynamic weights includes a sample weight calculation module, a multi-sub-model training module, and a higher-level fusion decision module. The training method for this mapping result classification system includes...

[0009] The sample weight calculation module assigns a time dynamic weight to each initial sample based on the time attribute of the initial sample to obtain training samples, and divides the training samples into multiple independent sub-sample sets according to the time period. The initial sample contains a hybrid feature vector and a corresponding multi-level classification label. The hybrid feature vector is generated by weighted fusion of structured feature vectors, semantic feature vectors and metadata feature vectors obtained by processing multi-source mapping data through an attention mechanism.

[0010] The multi-sub-model training module independently configures classifier algorithms for each sub-sample set and performs parallel training to generate sub-models for the corresponding time dimension. During training, the classification hierarchy and business entity association relationships of the surveying and mapping results in the rule graph are called in real time.

[0011] The upper-level fusion decision module uses a stacking fusion strategy to construct an upper-level linear regression model. It receives the prediction results of all sub-models, dynamically allocates weights based on the information gain of each sub-sample set, and outputs the final classification result after weighted fusion.

[0012] As one of the preferred options, the classifier algorithm is LightGBM or GBDT.

[0013] As one preferred solution, the surveying and mapping results text information, project contract structured data, and operation metadata are extracted from the OA system; the structured data is converted into structured feature vectors, the surveying and mapping results text information is converted into semantic feature vectors after word segmentation and entity recognition, and the operation metadata is converted into metadata feature vectors.

[0014] As one of the preferred solutions, the bidirectional TF-IDF algorithm is used to calculate the word frequency matrix during the semantic feature vector construction process, and the semantic feature vector is optimized through a dynamic weight adjustment mechanism. At the same time, entity recognition is used to map the project code and equipment model in the text to the project contract structured data, thereby enhancing the association between semantic features and structured features.

[0015] As one preferred approach, the exponential decay model for calculating the time dynamic weights is: , γ is the time dynamic weight, γ is the time decay factor, γ∈[0.1,0.3], where C is the current year and t is the data year.

[0016] A method for using the aforementioned time-dynamic weighted mapping result classification system includes the following steps:

[0017] 1) New data preprocessing: Obtain multi-source data of surveying and mapping results to be classified, including text information of surveying and mapping results, structured business data of project contracts and operation metadata, and convert them into structured feature vectors, semantic feature vectors and metadata feature vectors respectively. Generate hybrid feature vectors by weighted fusion through attention mechanism. The semantic feature vector is optimized by dynamic weighted TF-IDF feature engineering, and the project code and equipment model in the text are mapped to the structured data of project contracts through entity recognition.

[0018] 2) Parallel prediction by sub-models: The mixed feature vector is input into multiple trained time-dimensional sub-models, and each sub-model outputs the predicted probability of the corresponding classification label at each level.

[0019] 3) Upper-level fusion decision: The upper-level linear regression model receives the predicted probabilities of all sub-models, dynamically allocates weights based on the information gain of the sub-sample sets corresponding to each sub-model, and outputs preliminary three-level classification results after weighted fusion. The rule graph is called to perform hierarchical relationship verification and business entity association verification on the preliminary three-level classification results, and outputs the final classification results that meet the verification rules.

[0020] As one preferred approach, a model optimization step is also included, which comprises...

[0021] 31) When the cumulative amount of new surveying and mapping data reaches a preset threshold, incremental training is triggered, and historical model weights are loaded using a warm-start method;

[0022] 32) Manually review samples with classification confidence scores below a set threshold, add the reviewed samples to the training set, and update the rule graph.

[0023] 33) Based on the newly added samples, fine-tune the model parameters to complete the model optimization.

[0024] In the above technical solution, the beneficial effects of the present invention are as follows:

[0025] 1. Effectively improve the accuracy and timeliness of surveying and mapping results classification: Through time-dynamic weight design, the influence weight of recent surveying and mapping results, such as UAV surveying and laser point cloud and other new results, is strengthened, while the effective experience of historical data is retained to avoid the feature dilution problem when new and old data are mixed. This significantly improves the classification accuracy of new surveying and mapping results and adapts to the industry's technology iteration needs.

[0026] 2. Fully ensure that the classification results comply with industry business standards. Relying on the built-in surveying and mapping rule map, combined with entity recognition and structured data association technology, the classification process realizes the hierarchical relationship, such as the third-level category belonging to the corresponding second-level category, and the automatic verification of business qualification matching, such as the matching of the qualification of the operation department with the surveying and mapping type. This eliminates classification errors that do not conform to industry logic and reduces the cost of manual compliance review.

[0027] 3. Through the synergy of the three-level label tree and the Stacking integration strategy, cascading classification from the first-level major category to the third-level minor category can be completed without manual intervention. This avoids the limitations of traditional classification, which can only cover major categories and requires manual supplementation of classification details, thereby improving the efficiency of classification work and meeting the actual needs of refined classification in the management of surveying and mapping results.

[0028] 4. By integrating structured data, semantic features, and metadata through an attention mechanism, and combining dynamic weighted TF-IDF to optimize text feature vectors, the scattered results information, contract data, and metadata in the OA system can be effectively integrated to solve the problems of inconsistent formats of multi-source data and low feature utilization, thereby improving the model's adaptability to processing complex surveying and mapping data. Detailed Implementation

[0029] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below.

[0030] The mapping result classification system of the present invention, which incorporates time-dynamic weights, includes a sample weight calculation module, a multi-sub-model training module, and an upper-level fusion decision module. The training method of the mapping result classification system includes the following steps.

[0031] 1. The sample weight calculation module assigns dynamic weights to each initial sample based on its time attribute to obtain training samples. The training samples are then divided into multiple independent subsets according to time periods. Each initial sample contains a hybrid feature vector and corresponding multi-level classification labels. The hybrid feature vector is generated by weighted fusion of structured feature vectors, semantic feature vectors, and metadata feature vectors obtained from multi-source surveying data processing using an attention mechanism. For example, a composite feature vector is generated by weighted concatenation of structured features (35% weight), text features (55% weight), and metadata feature vectors (10%) using an attention mechanism. Specifically, the system extracts surveying result text information containing the result name and technical description from the OA system, including contract fields such as number, signing date, amount, project field, name, type, and work department code, as well as work department metadata including qualification level, work scope, coordinate system, and accuracy level. The surveying result text information includes the result name and technical description extracted from the OA system; the project contract structure data includes the basic contract field (contract number). The system includes: contract signing date, contract amount, project-related fields (project name, project type, project fields, etc.), and related identifier fields (such as those used to associate work metadata and work departments). Work metadata includes qualification level, work scope, coordinate system, and accuracy level. Structured data is converted into structured feature vectors. For example, contract numbers are standardized using regular expressions, and the "C-2023-045" format is uniformly converted to the "CONTRACT_2023045" encoding system. Work departments are converted into structured feature vectors using One-Hot encoding. The surveying and mapping results text information is converted into semantic feature vectors after word segmentation and entity recognition. In the process of constructing semantic feature vectors, a professional dictionary and stop word library in the surveying and mapping field are introduced. A bidirectional TF-IDF algorithm is used to calculate the word frequency matrix, and a dynamic weight adjustment mechanism is used to optimize the semantic feature vectors. At the same time, the project code and equipment model in the text are optimized through feature engineering and mapped to the project contract structured data through entity recognition. Feature engineering optimization refers to filling in some missing values, which can be done by using the mean, median, or other methods. Outliers can be identified and cleaned using statistical methods. Mapping involves storing data in a database, which refers to storing structured data in the database.

[0032] 1. For example, entity recognition technology can be used to automatically map "Project Code ZJ-2024" from unstructured text to the project name in structured data, enhancing the association between semantic features and structured project contract data. Each hybrid feature vector is labeled with a corresponding multi-level classification label. This labeling can be done manually by domain experts based on the technical description of the outcome, contract fields, and metadata, referring to a three-level label tree; or by using rule graphs to label samples; or a combination of both. The purpose is to add this hybrid feature vector-label pair to the training set for training a classification model. In the future, when encountering similar outcomes, the model can automatically output classification results. For instance, a three-level label tree can be constructed based on the GB / T35648-2023 standard, with first-level categories such as "Engineering Surveying," second-level categories such as "Underground Space Surveying," and third-level categories such as "Underground Pipeline Mapping."

[0033] The exponential decay model for calculating the time dynamic weights is as follows: , The time-based dynamic weights are defined by γ, which is a time decay factor ∈ [0.1, 0.3], where C is the current year and t is the data year. For example, training subsets are created with a 5-year cycle for each decade, training is performed separately for the 2010-2015 and 2016-2020 subsets, and finally, the prediction results of each sub-model are fused using a Stacking ensemble algorithm to improve the learning ability of time-sensitive features. The sample weights of the subset after 2020 are increased by 30%-50% to strengthen the learning of new technology features.

[0034] The multi-sub-model training module independently configures classifier algorithms for each sub-sample set and trains them in parallel to generate sub-models for the corresponding time dimension. During training, the classification hierarchy and business entity relationships of the surveying results in the rule graph are invoked in real time. The classifier algorithms are LightGBM or GBDT. LightGBM / XGBoost has strong capabilities in handling high-dimensional mixed features, is easy to embed hierarchical constraints of the rule graph, such as the subordinate relationship between parent and child labels, has fast training speed, is robust to noisy data, and can be adapted to sub-sample sets of different time periods through parameter adjustment. Random Forest / Gradient Boosting Tree (GBDT) has the advantage of outputting feature importance, making it easy to trace "which surveying features play a key role in classification", has good anti-overfitting ability, and is suitable for sub-model training of small to medium-sized sample sets.

[0035] The upper-level fusion decision module employs a stacking fusion strategy to construct the upper-level linear regression model. It receives predictions from all sub-models, dynamically assigns weights based on the information gain of each sub-sample set, and outputs the final classification result after weighted fusion. The core of information gain is to measure how much "certainty" an attribute (in this case, the model's prediction) can bring to the classification result. A larger gain indicates a greater contribution of the model's prediction to correct classification, and should be assigned a higher weight. Specifically, the weights of each model can be dynamically calculated based on relevant attribute information such as results, projects, contracts, and departments, especially operational equipment and project classifications.

[0036] To achieve the aforementioned dynamic allocation of information gain weights, two types of dynamic adjustment rules are preset based on surveying and mapping industry knowledge and historical classification data:

[0037] Terminology importance rules: Assign basic weight coefficients to specialized terms, such as setting the basic weight of core terms like "laser point cloud," "oblique photography," and "CGCS2000 coordinate system" to 1.5-2.0, and setting the basic weight of general terms like "report" and "data" to 0.5-0.8;

[0038] Contextual association rules: Weight enhancement is triggered when a specific combination of terms appears. For example, when "underground pipeline" and "detection" appear at the same time, the weight of both is increased by 20%; when "drone" and "topographic mapping" appear together, the weight is increased by 30%.

[0039] Specifically, in the process of constructing semantic feature vectors, a dynamic weighting factor is embedded in the TF-IDF calculation to perform a secondary adjustment on the traditional TF-IDF value, as shown in the formula:

[0040] in: For vocabulary The original term frequency - inverse document frequency value; These are the basic weighting coefficients in the terminology importance rules; As a dynamic enhancement factor in context association rules, based on vocabulary With context The co-occurrence relationship is calculated, ranging from 1.0 to 1.5. Iterative optimization of weights based on sample classification error is combined with the classification error during model training to dynamically correct the weight rules: The feature words of "misclassified samples" in the statistical validation set are analyzed. For example, if "underground space mapping" is repeatedly misclassified as "ground mapping," then the basic weight of "underground space" is adjusted. Automatically increase by 10%; for frequently occurring term combinations in "samples with high classification accuracy," such as "real estate" + "clave," the association enhancement factor is increased. This improves performance by 5%-10%, forming a closed loop of "data feedback - weight optimization." The optimized semantic feature vector is generated by dynamically adjusting the word weights and arranging them in the order of the vocabulary to form the final semantic feature vector, ensuring that the weights of surveying and mapping terms in the vector match their actual importance in the classification task.

[0041] Furthermore, the rule graph of this invention is a structured industry knowledge system built to adapt to the classification scenarios of surveying and mapping results. It transforms the business rules and classification standards of the surveying and mapping field into constraints that can be called by the model, ensuring that the classification results conform to the data patterns. The rule graph stores key information in a "node-relationship" graph structure, including classification hierarchy rules, clearly defining the subordinate logic of each level of classification label, forming a tree-like subordinate relationship of "level one → level two → level three", and preventing errors such as "level three categories belonging to different parent levels" (e.g., classifying "underground pipeline surveying" as "real estate surveying").

[0042] The rule graph also includes business entity association rules that associate core entities in surveying and mapping operations, such as the matching logic between work departments, equipment types, project attributes, and classification tags, reflecting the qualification and technical suitability requirements in actual work. For example, the work department "XX Surveying and Mapping Institute (Qualification Scope: Real Estate Surveying)" can only be associated with tags at all levels under the "Real Estate Surveying" category, and cannot be matched with tags in the "Marine Surveying" or "Geodesy" categories; the equipment type "Laser Point Cloud Scanner" is preferentially associated with third-level tags such as "3D Topographic Surveying" and "Underground Space Surveying," and is weakly associated or not associated with tags such as "Plane Control Surveying" that do not require 3D data. By establishing a network of associations between tags and business entities through the rule graph, a mapping relationship database is established between classification tags and projects, contracts, and work departments, enabling multi-hop queries of "Surveying Results → Second-level Category → Third-level Category → Corresponding Contract."

[0043] Using rule graphs alone can easily lead to rule bloat, requiring manual writing of each rule. However, using a TF-IDF feature vector library as the source of classification rules effectively solves this problem. Furthermore, since the feature vector library is dynamically updated, the rule graph can also be updated dynamically. Traditional rule graphs require experts to manually define rules, such as "if the text contains 'underground pipeline' and the project type is 'infrastructure,' then classify it as 'underground pipeline mapping.'" This method is accurate but has high maintenance costs, struggles to cover all situations, and cannot adapt to newly emerging terms or concepts. Utilizing a TF-IDF feature vector library automatically and continuously discovers and quantifies these judgment criteria, thereby dynamically generating rules. For the category "control measurement," the system automatically selects the N words with the highest TF-IDF weights (e.g., words with a weight > 0.8) as the key feature words for this category. Each category generates a suitable set of data-driven discrimination rules. The automatically generated rules need to be filtered and integrated before being used. The automatically generated rules are then injected into the rule graph. Rule graphs excel at understanding and processing long texts, particularly suitable for projects or achievements with numerous keywords. However, they struggle to distinguish between core and peripheral labels and lack weighting. This makes them a complementary combination to TF-IDF. Specifically, rule graphs use the TF-IDF feature vector library as their rule source, automatically generating age-sensitive classification rules. For example, they prioritize matching "electronic ranging" class weights to data from the 1990s and activate "real-world 3D modeling" class rules to data from the 2020s.

[0044] This invention integrates structured data, semantic features, and metadata through an attention mechanism, and combines dynamically weighted TF-IDF to optimize text feature vectors. This effectively integrates scattered results information, contract data, and metadata from OA systems, addressing the issues of inconsistent data formats and low feature utilization across multiple sources, and improving the model's adaptability to complex surveying data. Through a time-dynamic weighting design, it strengthens the influence weight of recent surveying results, such as UAV surveying and laser point cloud features, while retaining valuable experience from historical data. This avoids feature dilution when mixing new and old data, significantly improving the classification accuracy of new surveying results and adapting to the industry's technological iteration needs.

[0045] The method of using the time-dynamic weighted classification system for surveying and mapping results described in this invention includes the following steps:

[0046] 1) New data preprocessing: Obtain multi-source data of the surveying and mapping results to be classified, including text information, structured business data and metadata, and convert them into structured feature vectors, semantic feature vectors and metadata feature vectors respectively. Generate hybrid feature vectors by weighted fusion through attention mechanism. The semantic feature vectors are optimized by dynamic weighted TF-IDF feature engineering, and the project code and equipment model in the text are mapped to the project contract structured data through entity recognition.

[0047] 2) Parallel prediction of sub-models: The mixed feature vector is input into multiple trained time-dimensional sub-models, and each sub-model outputs the predicted probability of the corresponding classification label at each level;

[0048] 3) Upper-level fusion decision: The upper-level linear regression model receives the predicted probabilities of all sub-models, dynamically allocates weights based on the information gain of the sub-sample sets corresponding to each sub-model, outputs preliminary three-level classification results after weighted fusion, calls the rule graph to perform hierarchical relationship verification and business entity association verification on the preliminary three-level classification results, and outputs the final classification results that meet the verification rules.

[0049] Furthermore, the present invention also includes a model optimization step, which includes,

[0050] 31) When the cumulative amount of new surveying and mapping data reaches the preset threshold, incremental training is triggered, and historical model weights are loaded in a warm-start manner; when new results are added to the database, model fine-tuning is triggered, and XGBoost with warm-start is used to update parameters, while retaining historical model weights to avoid cold start; if a minimum update sample size (e.g., 100) is set, performance fluctuations caused by frequent training can be prevented.

[0051] 32) Samples with classification confidence scores below a set threshold are manually reviewed, and the reviewed samples are added to the training set and the rule graph is updated; the classification confidence score is calculated based on the prediction probability entropy value. , When the prediction probability entropy value Samples with a confidence level >0.8 are considered low-confidence samples, automatically pushed to the manual review interface, and the data entry process is suspended. The entropy value represents the prediction probability.

[0052] To measure the degree of uncertainty in the predicted probability distribution, i is the index variable for summation, used to iterate through the categories from 1 to k, where k represents the total number of categories, i.e., the number of categories to be predicted. This represents the model's predicted probability for the i-th category.

[0053] 33) Based on the newly added samples, fine-tune the model parameters to complete the model optimization. That is, use the newly added data as sample data to retrain the model using the training strategy of this invention to adjust the optimization parameters.

[0054] Simultaneously, after classifying the results based on the rule graph, feedback can be provided to TF-IDF to update thresholds and weights, forming a complete feedback loop. When TF-IDF classification results conflict with the rule engine, manual intervention is initiated, and the correction results are written into the knowledge base. The TF-IDF weight coefficients of the corresponding labels are adjusted synchronously, increasing the weight of new terms like "UAV cadastral mapping". ≥0.2.

[0055] In the feature fusion stage, an adversarial text enhancement mechanism is introduced. The semantic valence perturbation text is generated through the BERT-CRF model, and cross-era adversarial samples containing "coastal topographic map, marine mapping / integrated natural resource survey" are constructed to improve the robustness of the model.

[0056] This invention achieves multi-dimensional feature fusion, combining business-related data (contracts, projects) with text content to improve the business rationality of classification. Data subsets are divided by era, the contribution of each era to the current task is calculated, and training sample weights are dynamically allocated. Dynamic time weights are also introduced to address the problem of historical data distribution shifts caused by the evolution of surveying and mapping technology. Furthermore, a closed-loop adaptive system is designed: incremental learning enables model self-optimization, reducing manual intervention.

[0057] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above description is illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A mapping product classification system that fuses temporal dynamic weights, characterized by, The surveying and mapping achievement classification system comprises a sample weight calculation module, a multi-sub-model training module and an upper layer fusion decision module, and the training method of the surveying and mapping achievement classification system comprises, The sample weight calculation module assigns a time dynamic weight to each initial sample according to the time attribute of the initial sample to obtain a training sample, and divides the training sample into a plurality of independent sub-sample sets according to a time period, wherein the initial sample comprises a mixed feature vector and a corresponding multi-level classification label, the mixed feature vector is generated by weighting and fusing a structured feature vector, a semantic feature vector and a metadata feature vector obtained by processing surveying and mapping multi-source data through an attention mechanism, wherein surveying and mapping achievement text information, project contract structured data and job metadata are extracted from an OA system; the structured data is converted into a structured feature vector, the surveying and mapping achievement text information is converted into a semantic feature vector after being segmented and recognized, and the job metadata is converted into a metadata feature vector; The multi-sub-model training module independently configures a classifier algorithm for each sub-sample set to perform parallel training to generate a sub-model corresponding to a time dimension, and the classification hierarchical relationship and the business entity association relationship of the surveying and mapping achievement in the rule graph are called in real time during the training; The upper layer fusion decision module adopts a Stacking integrated strategy to construct an upper layer linear regression model, receives the prediction results of all the sub-models, dynamically allocates weights based on the information gain of each sub-sample set, and outputs a final output classification result after weighted fusion.

2. The fused temporal dynamic weight mapping product classification system of claim 1, wherein, The classifier algorithm is LightGBM or GBDT.

3. The fusion time dynamic weight mapping product classification system of claim 1, wherein, In the semantic feature vector construction process, a bidirectional TF-IDF algorithm is used to calculate a word frequency matrix, and a dynamic weight adjustment mechanism is used to optimize the semantic feature vector; meanwhile, the project code and equipment model in the text are mapped to the project contract structured data through entity recognition, so as to realize the association enhancement of semantic features and structured features.

4. The fused temporal dynamic weight mapping product classification system of claim 1, wherein, The exponential decay model for calculating the time dynamic weight is , is the time dynamic weight, γ is the time decay factor, γ ∈ [0.1, 0.3], where C is the current year, and t is the data year.

5. A method of using a fusion time dynamic weight mapping product classification system as claimed in any one of claims 1 to 4, characterized by, The method comprises the following steps, 1) new data preprocessing, obtaining surveying and mapping achievement multi-source data to be classified, including surveying and mapping achievement text information, project contract structured business data and job metadata, converting them into structured feature vectors, semantic feature vectors and metadata feature vectors respectively, and generating a mixed feature vector by weighting and fusing through an attention mechanism, wherein the semantic feature vector is optimized through dynamic weighting TF-IDF feature engineering, and the project code and equipment model in the text are mapped to the project contract structured data through entity recognition; 2) sub-model parallel prediction, inputting the mixed feature vector into the trained plurality of time dimension sub-models, and each sub-model outputs a prediction probability corresponding to each level of classification label; 3) upper layer fusion decision, the upper layer linear regression model receives the prediction probability of all the sub-models, dynamically allocates weights based on the information gain of the sub-sample set corresponding to each sub-model, and outputs a preliminary three-level classification result after weighted fusion, calls the rule graph to perform hierarchical relationship verification and business entity association verification on the preliminary three-level classification result, and outputs a final classification result meeting the verification rules.

6. The method of using a fusion time dynamic weight mapping product classification system of claim 5, wherein, The method further comprises a model optimization step, which comprises, 31) when the cumulative amount of new surveying and mapping achievement data reaches a preset threshold, triggering incremental training, and loading the historical model weight in a warm-start manner. 32) Manually review samples with classification confidence below a set threshold, add the reviewed samples to the training set and update the rule graph, 33) Retrain the model parameters based on the newly added samples, complete the model optimization.