An intelligent compliance verification system for dangerous goods packing operation

Abstract

The present application relates to the technical field of dangerous goods management, in particular to a dangerous goods packing operation intelligent compliance verification system, which comprises a data acquisition preprocessing unit, a regulation graph construction iteration unit, a multi-agent collaborative verification unit, a risk early warning report generation unit and a man-machine collaborative closed loop unit; the regulation graph construction iteration unit extracts a triple of "regulation clause number-verification dimension-constraint condition" of regulations in the field of dangerous goods packing and indexes according to the level of effectiveness, the main body of release and the scope of application, and introduces an improved three-dimensional dynamic weight conflict resolution algorithm; the multi-agent collaborative verification unit; the risk early warning report generation unit; the man-machine collaborative closed loop unit. The present application solves the problems of regulation update lag and clause conflict by adopting the triple of "regulation clause number-verification dimension-constraint condition" extraction, the three-dimensional dynamic weight conflict resolution algorithm based on regulation effectiveness-scene matching degree-release time, and the technical means of periodically optimizing weight parameters through reinforcement learning model.

Description

Technical Field

[0001] This invention relates to the field of dangerous goods management technology, and more specifically, to an intelligent compliance verification system for dangerous goods packing operations. Background Technology

[0002] Before the transport of dangerous goods, the packing process must strictly comply with a series of constantly updated international and domestic regulations and standards. Traditional verification methods rely heavily on the manual memory and experience of operators, supplemented by offline verification of paper documents. This method is not only inefficient but also highly susceptible to misjudgments and omissions due to misunderstandings of regulations, information omissions, or operator fatigue, posing significant safety hazards and compliance risks. Therefore, achieving efficient, accurate, and automated compliance verification has become a key technical challenge for ensuring the safety of dangerous goods transport and improving the operational efficiency of ports and logistics hubs.

[0003] In the existing technology, relevant patents have proposed solutions for the supervision and safety inspection of dangerous goods in ports. For example, Chinese patent CN202510681319.3 discloses "A method for safety supervision of dangerous goods in ports." This solution relates to the field of dangerous goods management technology. It collects internal status and location information by setting status acquisition devices and radio frequency positioning tags inside dangerous goods containers, and generates a safety supervision patrol path to control the patrol vehicle for patrol. Based on the collected information, safety supervision is carried out, aiming to improve the quality of safety supervision of dangerous goods in ports. Another example is Chinese patent CN202211515936.9, which discloses "A safety traceability management system for dangerous goods in ports." This solution stores safety inspection items and standards for each operational stage by constructing a safety inspection standard library, and generates safety inspection content and results accordingly. It aims to standardize inspection processes, reduce reliance on personnel, and improve the efficiency of safety management of goods during their stay in port.

[0004] Despite the design advantages of the aforementioned technical solutions, they also suffer from the following technical shortcomings: First, the problems of outdated regulations and conflicting clauses are prominent, and the existing system lacks adaptability: The "safety inspection standard library" adopted by CN202211515936.9 is essentially a static database, while regulations in the field of dangerous goods transportation are frequently updated, and the regulatory requirements of different countries and different modes of transportation differ or even conflict. The existing system cannot automatically identify regulatory updates, and it lacks an intelligent mechanism for resolving conflicting clauses. This leads to situations where the system often relies on outdated regulations or cannot determine the applicable clauses in actual use, requiring frequent manual intervention and maintenance, which seriously affects the accuracy and reliability of verification. Second, fragmented multi-source evidence makes it difficult to form effective corroboration, and the existing technology lacks deep semantic fusion capabilities: CN202510681319.3 mainly relies on sensors to collect physical state data, while the key evidence for the compliance verification of dangerous goods is scattered in multimodal data such as dangerous goods packaging certificates, UN labels, and stacking images. Existing technologies lack the ability to deeply understand the semantics of these heterogeneous pieces of evidence, failing to establish the intrinsic connections between document information, visual features, and regulatory provisions. This results in systems that can only perform simple rule matching, making it difficult to detect contradictions between pieces of evidence or accurately locate and interpret complex violations. Therefore, we propose an intelligent compliance verification system for dangerous goods packing operations. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent compliance verification system for dangerous goods packing operations to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, the present invention aims to provide an intelligent compliance verification system for dangerous goods packing operations, comprising:

[0007] The data acquisition and preprocessing unit uses optical character recognition and image preprocessing technology to acquire scanned copies of dangerous goods packaging certificates, UN number information, stacking images, real-shot images of labels, and basic packing data, and performs standardized conversion, storing the structured verification data source in the database;

[0008] The regulatory graph construction iterative unit extracts the "regulatory clause number-verification dimension-constraint condition" triples from regulations in the field of dangerous goods packaging and indexes them according to the level of effectiveness, issuing entity, and scope of application. An improved three-dimensional dynamic weight conflict resolution algorithm is introduced to construct a dynamic weight matrix based on regulatory effectiveness, scenario matching degree, and publication time. The parameters of the dynamic weight matrix are periodically optimized by an offline-trained reinforcement learning model, which uses the consistency between historical verification decisions and manual review results as a reward signal. By monitoring regulatory update announcements, based on preset version lifecycle rules and time windows, the unit automatically switches the confidence level of new and old clauses in the knowledge graph and reconstructs their associations. The text, image, and historical evidence from the data acquisition and preprocessing unit are constructed into a heterogeneous graph, and evidence associations are learned through a graph attention mechanism to output interpretable difference localization results.

[0009] The multi-agent collaborative verification unit, based on a structured verification data source and a scenario-based dynamic regulatory knowledge graph, completes single-dimensional compliance judgment and multi-dimensional comprehensive result summarization through collaborative reasoning of intelligent agent modules for dangerous goods packaging certificate verification, UN number matching, stacking compliance judgment, and label identification verification.

[0010] The risk warning report generation unit, based on the multi-agent collaborative verification results, triggers warning signals and automatically generates a regulatory matching report with explanatory details through a hierarchical risk triggering mechanism and standardized report templates.

[0011] The human-machine collaborative closed-loop unit establishes an interactive interface between AI judgment results and manual review, and feeds back the manual review results to the regulatory map construction iteration unit.

[0012] As a further improvement to this technical solution, the data acquisition and preprocessing unit includes a multi-source data acquisition module and a data standardization and conversion module, wherein:

[0013] The multi-source data acquisition module uses optical character recognition technology to acquire text data of scanned dangerous goods packaging certificates and UN number information, and uses image preprocessing technology to acquire image data of stacking images and actual label photos, while simultaneously acquiring basic container data.

[0014] The data standardization and conversion module performs format normalization on the collected text data, extracts features and performs structure conversion on the image data, and stores the processed data in the database to form a structured verification data source.

[0015] As a further improvement to this technical solution, the regulatory map construction iterative unit includes a triplet extraction and indexing module, a dynamic weight conflict resolution module, a map optimization and reconstruction module, and a multimodal evidence fusion module, wherein:

[0016] The triple extraction indexing module is used to extract the triple "legal clause number-verification dimension-constraint condition" and to perform structured indexing of validity level, issuing entity, and scope of application;

[0017] The dynamic weight conflict resolution module constructs a dynamic weight matrix based on the legal effect, scenario matching degree, and release time, and quantifies and resolves conflicting clauses.

[0018] The graph optimization and reconstruction module periodically optimizes weight parameters through a reinforcement learning model, while monitoring regulatory updates and completing the dynamic reconstruction of the knowledge graph.

[0019] The multimodal evidence fusion module is used to construct a heterogeneous graph from multi-source evidence, learn evidence associations through a graph attention mechanism, and output differential localization results.

[0020] As a further improvement to this technical solution, the triplet extraction and structured indexing of the triplet extraction and indexing module includes the following steps:

[0021] S21.1. A hybrid extraction method combining regular rules and a fine-tuned BERT model is adopted. After matching the regulatory clause number and filtering redundant text, the target text containing verification dimensions and constraints is input into the BERT model.

[0022] S21.2 Calculate the class probabilities of the verification dimensions and constraints using the Softmax function. ;

[0023] S21.3. Associate and combine constraints according to the format of "legal clause number-verification dimension-constraint condition", and split the constraints into three types of structured data: numerical, logical, and range.

[0024] S21.4 Mapping effectiveness level quantification coefficients according to preset rules ,in Hierarchical binary identifier With corresponding level preset weight The association is obtained; a three-dimensional index vector is constructed according to "goods type - packaging category - transportation method". As a further improvement to this technical solution, the weight matrix construction and conflict resolution of the dynamic weight conflict resolution module include the following steps:

[0025] S22.1, Weight of Legal Validity The effectiveness level quantification coefficients output from S21.4 are used. ; Calculate scene matching weights using cosine similarity ,in Vector of Applicable Regulations Vector of the current packing operation scenario ; Calculate the release time weight using a linear decay function ,in Current time of associated system Date of promulgation of regulations With respect to the pre-set validity period of regulations ;

[0026] S22.2, to Normalization process is performed to obtain ,based on Constructing a dynamic weight matrix ;

[0027] S22.3 Calculate the overall score (Score) for each conflicting clause under the same verification dimension, where the Score is associated with the dynamic weight matrix. Quantitative values ​​corresponding to the dimensions of each clause The clause with the highest overall score is selected as the adaptation result. When the score difference meets the preset threshold, the multimodal evidence cross-validation process is triggered.

[0028] As a further improvement to this technical solution, the parameter optimization and map reconstruction of the map optimization and reconstruction module includes the following steps:

[0029] S23.1, Define the agent as the dynamic weight matrix constructed in S22.2. parameter set Define the state space Define the action space for the feature vectors of the historical verification scenario. For parameters Adjustment set ;

[0030] S23.2 Collect historical verification data from the past 3 years and divide it into training set and verification set according to a preset ratio. Each data point includes AI decision results and manual review results.

[0031] S23.3, Define the reward function ,in Correlation weight coefficient Decision accuracy With F1 score The DDPG algorithm is used, and samples are stored in an empirical replay buffer. Update parameters using gradient descent ;

[0032] S23.4. Set an optimization cycle. At the end of each cycle, use the validation set to evaluate the model performance. If the F1 score is lower than a preset threshold, trigger a new round of offline training and update the dynamic weight matrix. Parameters;

[0033] S23.5. Use web crawling technology to retrieve regulations and announcements from official websites, and use the TF-IDF algorithm to extract keywords, including TF-IDF-related keywords. ,document Total number of documents With keywords Number of documents When the keyword matching degree meets the preset threshold, it is determined to be a relevant regulatory update, and the updated clauses and effective date are extracted. ;

[0034] S23.6. Set a transition time window and use the Sigmoid function to adjust the confidence levels of the old and new clauses. Among them, the confidence level correlation smoothing coefficient Current time With the effective date of the regulations ; Calculate the similarity of constraints between the old and new clauses. , When the preset similarity threshold is met, the old clause's associated edge is retained and points to the new clause; If the preset similarity threshold is not met, delete the associated edges of the old clauses and construct associated edges with the verification dimension nodes for the new clauses.

[0035] As a further improvement to this technical solution, the evidence fusion and difference localization of the multimodal evidence fusion module includes the following steps:

[0036] S24.1 Map the textual evidence, image evidence, and historical record evidence output by the data acquisition and preprocessing unit to text nodes, image nodes, and historical record nodes in a heterogeneous graph, respectively; generate feature vectors for the textual evidence using the BERT model. A CNN model is used to generate feature vectors for video evidence. LSTM model is used to generate feature vectors for historical evidence. ;

[0037] S24.2. Based on the original node feature vectors in S24.1, the initial correlation between nodes is calculated using the cosine similarity algorithm, and then transformed into initial attention coefficients. The initial edge weights of the heterogeneous graph are obtained by Softmax normalization. And based on the initial edge weights of the heterogeneous graph Construct the initial connection structure of the multimodal evidence heterogeneity graph;

[0038] S24.3, through the dynamic weight matrix A linear transformation is performed on the original node feature vectors to obtain enhanced node features. The initial graph structure and enhanced node features constructed based on S24.2 Combined with attention weight vector Using the LeakyReLU activation function, calculate the fine attention coefficients. ;

[0039] S24.4, to At the node The set of neighboring nodes Perform Softmax normalization within the range to obtain fine-grained edge weights. ;by The weighting coefficients are used to evaluate the features of neighboring nodes. After weighted summation, the activation function is applied. Transformation to generate node embedding vectors that incorporate multimodal information. ;

[0040] S24.5, Calculate the node embedding vector Node features corresponding to the regulatory map The Euclidean distance is used as the deviation value ;like If the threshold is exceeded, the output includes the deviation node identifier, the associated evidence ID, the deviation value, and the difference location result of the regulatory inconsistency.

[0041] As a further improvement to this technical solution, the multi-agent collaborative verification unit includes a dangerous goods package certificate verification agent module, an UN number matching agent module, a stacking compliance judgment agent module, and a label identification verification agent module, wherein:

[0042] The dangerous goods packaging certificate verification intelligent agent module verifies the compliance of the dangerous goods packaging certificate validity period, issuing agency qualifications, and packaging container model based on the electronic data of dangerous goods packaging certificates in the structured verification data source and the compliance rules of dangerous goods packaging certificates in the scenario-based dynamic legal knowledge graph, and generates a compliance judgment result for the dangerous goods packaging certificate dimension.

[0043] The UN number matching intelligent agent module compares the actual UN number of the goods with the regulatory requirements based on the UN number information of the goods in the structured verification data source and the UN number association rules in the scenario-based dynamic regulatory knowledge graph, and generates a compliance judgment result for the UN number dimension. The stacking compliance judgment intelligent agent module analyzes the compliance of the number of stacking layers and stacking method of the goods based on the stacking data of the goods in the structured verification data source and the stacking specifications in the scenario-based dynamic regulatory knowledge graph, and generates a compliance judgment result for the stacking dimension.

[0044] The label identification verification intelligent agent module verifies the compliance of the content, format, and placement of goods labels based on label identification data in a structured verification data source and label identification rules in a scenario-based dynamic legal knowledge graph, and generates a label identification dimension compliance judgment result.

[0045] As a further improvement to this technical solution, the dangerous goods packaging certificate verification intelligent agent module includes a data parsing submodule, a rule comparison submodule, and an anomaly tracing submodule, wherein:

[0046] The data parsing submodule extracts core fields directly related to the compliance verification of dangerous goods packaging from the electronic data of dangerous goods packaging certificates based on OCR recognition technology and structured data parsing technology.

[0047] The rule comparison submodule calls the dangerous goods packaging certificate compliance rule library in the scenario-based dynamic legal knowledge graph, and compares the parsed fields with the "validity period, list of issuing agency qualifications, and list of compliant packaging container models" in the dangerous goods packaging certificate compliance rule library one by one;

[0048] When the anomaly tracing submodule determines an anomaly through rule comparison, it automatically associates real-time data collected from cargo packaging with historical verification records. Through multi-dimensional data correlation analysis, it locates the root cause of the anomaly and outputs a tracing report containing "anomaly type, legal basis, and rectification suggestions." This report is then synchronized to the UN number matching intelligent agent module, the stacking compliance determination intelligent agent module, and the label identification verification intelligent agent module, triggering a multi-agent joint verification process.

[0049] As a further improvement to this technical solution, the risk warning report generation unit includes a verification result parsing module, a risk classification determination module, and a standardized report generation module, wherein:

[0050] The verification result parsing module receives compliance judgment results from various dimensions output by the multi-agent collaborative verification unit, extracts key information directly related to the compliance judgment, including violation type, violation-related data, and regulatory matching basis, and completes the structured integration of the results;

[0051] The risk assessment module is based on the priority of compliance clauses in the scenario-based dynamic legal knowledge graph, combined with the degree of impact of violations on packing safety, and triggers corresponding level warning signals through preset risk thresholds.

[0052] The standardized report generation module calls a preset compliance report template, integrates structured verification results, graded risk levels, specific regulatory clauses corresponding to violations, interpretable compliance judgment explanations, and targeted rectification guidelines, and automatically generates a complete regulatory matching risk warning report.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] 1. This invention effectively solves the problems of regulatory update lag and clause conflict by using a triplet extraction of "regulatory clause number-verification dimension-constraint condition", a three-dimensional dynamic weight conflict resolution algorithm based on regulatory validity-scenario matching degree-release time, and a reinforcement learning model to periodically optimize weight parameters, thereby realizing the dynamic construction and adaptive optimization of the regulatory knowledge graph.

[0055] 2. This invention solves the problem of fragmented multi-source evidence by mapping multimodal evidence to heterogeneous graph nodes, constructing a graph structure based on initial correlation, and then performing fine feature fusion through a graph attention mechanism. This enables deep semantic association and interpretable differential localization between evidence.

[0056] 3. This invention deploys four dedicated intelligent agents for dangerous goods packaging certificate verification, UN number matching, stacking compliance judgment, and label identification verification. Based on structured verification data sources and dynamic regulatory knowledge graphs, collaborative reasoning is performed to form a professional and automated verification capability covering all key dimensions of dangerous goods packaging.

[0057] 4. By constructing a closed-loop architecture that integrates data collection, intelligent verification, and human-machine collaborative feedback, this invention feeds back the results of manual review to the regulatory map construction iteration unit, enabling continuous optimization of the system based on actual usage data and improving the long-term usability and reliability of the system. Attached Figure Description

[0058] Figure 1 This is a system structure diagram of the present invention;

[0059] The meanings of the labels in the diagram are as follows:

[0060] 1. Data acquisition and preprocessing unit; 11. Multi-source data acquisition module; 12. Data standardization and conversion module;

[0061] 2. Iterative unit for constructing the regulatory graph; 21. Triple extraction and indexing module; 22. Dynamic weight conflict resolution module; 23. Graph optimization and reconstruction module; 24. Multimodal evidence fusion module;

[0062] 3. Multi-agent collaborative verification unit; 31. Dangerous goods packaging certificate verification agent module; 32. UN number matching agent module; 33. Stacking compliance judgment agent module; 34. Label identification verification agent module;

[0063] 4. Risk warning report generation unit; 41. Verification result analysis module; 42. Tiered risk assessment module; 43. Standardized report generation module;

[0064] 5. Human-machine collaborative closed-loop unit. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0066] like Figure 1 As shown, this embodiment provides an intelligent compliance verification system for dangerous goods packing operations, including:

[0067] Data acquisition and preprocessing unit 1 uses optical character recognition and image preprocessing technology to acquire scanned copies of dangerous goods packaging certificates, UN number information, stacking images, real photos of labels, and basic packing data, and performs standardized conversion, storing the structured verification data source in the database;

[0068] In this embodiment, the data acquisition and preprocessing unit 1 includes a multi-source data acquisition module 11 and a data standardization and transformation module 12, wherein:

[0069] The multi-source data acquisition module 11 uses optical character recognition technology to acquire text data of dangerous goods packaging certificate scans and UN number information, and uses image preprocessing technology to acquire image data of stacking images and actual label photos, while simultaneously acquiring container basic data.

[0070] Specifically, the multi-source data acquisition module 11 completes the acquisition of multiple types of data by adapting to hardware devices and preprocessing technologies:

[0071] Industrial-grade high-definition scanning equipment is used to scan dangerous goods packaging certificates to generate electronic scans. Portable handheld scanning terminals are used to collect UN numbers. Before collection, the electronic images are preprocessed by Gaussian filtering for noise reduction, Hough transform for correction, and edge cropping. Then, an OCR recognition engine that supports mixed Chinese, English, and numbers is called to extract core text fields such as the certificate number, issuance date, and validity period, as well as the UN number (and initially verify it according to the "UN + 4 digits" format; if it does not match, it is marked for further verification).

[0072] High-definition industrial cameras at fixed workstations (installed directly above the packing station, with adjustable height) are used to capture stacking images (3-5 images from different angles are taken after each stacking is completed, with supplementary lighting to avoid glare). Handheld smart terminal cameras are used to capture real-life images of the labels (shot around the four key sides of the packaging to ensure that the label information is clear and associated with the workstation number and timestamp).

[0073] Meanwhile, basic packing information (such as cargo name, packaging quantity, etc., including format verification) can be manually entered through the system's interactive interface, or the system can be connected to external workstation management and cargo management systems through standard API interfaces to synchronize data such as order numbers and container identification numbers in real time, so as to achieve complete collection of dangerous goods packaging certificates, UN numbers, stacking images, real photos of labels and basic packing data.

[0074] The data standardization and conversion module 12 performs format normalization on the collected text data, extracts features and performs structure conversion on the image data, and stores the processed data in the database to form a structured verification data source.

[0075] Specifically, the data standardization and conversion module 12 performs unified format processing and structured storage on the collected data:

[0076] Text data is normalized according to rules. Date fields are uniformly formatted as “YYYY-MM-DD” and their logical rationality is verified (e.g., the validity period start date is not later than the issuance date). Number fields are removed of redundant spaces and non-inherent special symbols and verified according to encoding rules. Text description fields use full-width Chinese characters and redundant spaces are removed. Goods names are standardized and mapped based on the dangerous goods name standard library.

[0077] For image data, a convolutional neural network (CNN) model is called to identify individual goods from stacked images, count the number of stacking layers, analyze the stacking method, and detect the stability of stacking (marking the coordinates of abnormal areas). By combining OCR with a pre-trained image classification model, UN number, dangerous goods category, warning symbol code, and label affixing location are extracted from the actual image of the label. The extracted features are converted into structured fields and associated with the original image storage path.

[0078] Ultimately, all structured data is stored in a relational database, and tables such as "Dangerous Goods Packaging Certificate Information Table," "UN Number Information Table," "Stacking Information Table," "Label Information Table," and "Basic Packing Information Table" are designed. Foreign key relationships are established through fields such as container identification number and UN number. At the same time, a daily scheduled full backup mechanism is set up to form a structured verification data source that can be used for subsequent verification.

[0079] The regulatory graph construction iteration unit 2 extracts the "regulatory clause number-verification dimension-constraint condition" triples from regulations in the field of dangerous goods packaging and indexes them according to their validity level, issuing entity, and scope of application. It introduces an improved three-dimensional dynamic weight conflict resolution algorithm and constructs a dynamic weight matrix based on regulatory validity, scenario matching degree, and publication time. The parameters of the dynamic weight matrix are periodically optimized by an offline-trained reinforcement learning model, which uses the consistency between historical verification decisions and manual review results as a reward signal. By monitoring regulatory update announcements, based on preset version lifecycle rules and time windows, it automatically switches the confidence level of new and old clauses in the knowledge graph and reconstructs their associations. It also constructs a heterogeneous graph from the text, image, and historical evidence in data collection and preprocessing unit 1, learns evidence associations through a graph attention mechanism, and outputs interpretable difference localization results. The regulatory graph construction iteration unit 2 includes a triple extraction and indexing module 21, a dynamic weight conflict resolution module 22, a graph optimization and reconstruction module 23, and a multimodal evidence fusion module 24, wherein:

[0080] In this embodiment, the triple extraction indexing module 21 is used to extract the triple "legal clause number-verification dimension-constraint condition" and perform structured indexing of validity level, issuing entity, and scope of application; the triple extraction and structured indexing of the triple extraction indexing module 21 includes the following steps:

[0081] S21.1. A hybrid extraction method combining regular rules and a fine-tuned BERT model is adopted. After matching the regulatory clause number and filtering redundant text, the target text containing verification dimensions and constraints is input into the BERT model.

[0082] Specifically, for regulations at different levels of effectiveness in the field of dangerous goods packing, such as international regulations, domestic standards, and industry specifications, dedicated clause number matching rules are designed: For international regulations (such as the IMO International Maritime Dangerous Goods Code), the number begins with a capital letter, may contain letter segments separated by periods, followed by numbers and optional numbers in parentheses; for domestic standards (such as the GB30000 series), the number begins with a capital letter, followed by one or two groups of numbers (with decimal point separators) and optional numbers with hyphens; for industry specifications (such as JT / T standards), the number begins with a capital letter followed by " / T", followed by numbers and optional numbers with hyphens. These rules accurately match clause numbers at each level of regulation, filtering redundant text before and after the clause number, laying the foundation for subsequent ternary extraction.

[0083] S21.2 Calculate the class probabilities of the verification dimensions and constraints using the Softmax function. ;

[0084] Specifically, the probability of the target text belonging to each verification dimension and constraint category is calculated using the Softmax function, with the following formula:

[0085] ;

[0086] in, Represents input text Category (e.g., the probability of "validity period of dangerous goods packaging certificate" or "stack layer ≤ 3 layers" constraint) It is a model for text Category The output score, It is the total number of categories of verification dimensions or constraints in the current scenario (e.g., there are 8 categories of verification dimensions, including dangerous goods packaging certificate, UN number, stacking, label, etc.).

[0087] Furthermore, a probability threshold is set (determined based on historical data sampling to ensure that the false negative rate is below the industry acceptable range). If the probability of a certain category is higher than the threshold, the text is directly determined to belong to that category. If the probability of all categories is lower than the threshold, it is marked as "awaiting manual annotation". After supplementary confirmation by the regulatory specialist, the text is added to the training set as a new sample to iteratively optimize the BERT model.

[0088] S21.3. Associate and combine constraints according to the format of "legal clause number-verification dimension-constraint condition", and split the constraints into three types of structured data: numerical, logical, and range.

[0089] Specifically, following the fixed format of "legal clause number - verification dimension - constraint", the extracted clause number is associated with the judgment dimension and constraint to form a complete triple (such as "GB30000.1-2013-stack layer number-≤3 layers").

[0090] Constraints are split according to data type: numerical constraints are directly extracted by numbers and units (e.g., "packaging thickness ≥ 5mm" is split into "5mm"); logical constraints are converted into "yes / no" binary values ​​(e.g., "label must be marked with UN number" is split into "yes"); range constraints are extracted by interval boundaries and units (e.g., "dangerous goods packaging certificate validity period 1-2 years" is split into "1 year - 2 years"), ensuring that constraints can be directly used for subsequent numerical comparisons and logical judgments.

[0091] S21.4 Mapping effectiveness level quantification coefficients according to preset rules ,in Hierarchical binary identifier With corresponding level preset weight The association is obtained; a three-dimensional index vector is constructed according to "goods type - packaging category - transportation method". .

[0092] Specifically, the effectiveness level quantification coefficient is calculated according to preset rules. The formula is:

[0093] ;

[0094] in, It is a hierarchical binary identifier ( =1 corresponds to international regulations. =2 corresponds to national regulations. =3 corresponds to industry standards. =4 corresponds to a certain level of corporate regulations or rules. =1, otherwise 0); These are the preset weights for the corresponding levels (international regulations 0.4, national regulations 0.3, industry standards 0.2, corporate regulations 0.1, which are set based on the priority of the effectiveness of regulations and conform to the general principle of "international over domestic, national over industry" in the field of dangerous goods management).

[0095] At the same time, a three-dimensional index vector is constructed according to "goods type - packaging category - transportation method". :

[0096] According to the "Classification and Numbering of Dangerous Goods" (GB6944) codes (e.g., flammable liquids 3, corrosive substances 8);

[0097] According to UN packaging standard codes (Class I 1, Class II 2, Class III 3);

[0098] The transportation mode is coded (sea transport 1, land transport 2, air transport 3, multimodal transport 4) to provide a structured basis for subsequent scenario matching.

[0099] In this embodiment, the dynamic weight conflict resolution module 22 constructs a dynamic weight matrix based on regulatory validity, scenario matching degree, and release time, and quantifies and resolves conflicting clauses. The dynamic weight conflict resolution module 22 achieves quantitative processing of conflicting regulatory clauses under the same verification dimension through a three-step process of "weight calculation - matrix construction - conflict resolution." The weight matrix construction and conflict resolution of the dynamic weight conflict resolution module 22 include the following steps:

[0100] S22.1, Weight of Legal Validity The effectiveness level quantification coefficients output from S21.4 are used. ; Calculate scene matching weights using cosine similarity ,in Vector of Applicable Regulations Vector of the current packing operation scenario ; Calculate the release time weight using a linear decay function ,in Current time of associated system Date of promulgation of regulations With respect to the pre-set validity period of regulations ;

[0101] Specifically, the calculation logic for the three-dimensional weights is as follows:

[0102] Weight of legal effect : Directly reuse the effectiveness level quantization coefficients output by the triplet extraction indexing module 21 ,Right now This ensures that the weighting is strongly correlated with the level of effectiveness of the regulations themselves, and avoids giving priority to regulations with lower effectiveness.

[0103] Scene matching weight The matching degree between the applicable scope of regulations and the current work scenario is calculated using cosine similarity. The formula is as follows:

[0104] ;

[0105] in, It is the three-dimensional index vector corresponding to the regulations (from the triplet indexing results). It is the three-dimensional vector of the current packing operation (based on actual operation parameters, such as [3,2,1] corresponding to "flammable liquids-Class II packaging-sea freight"). It is the dot product of two vectors. It is the L2 norm of two vectors. The value ranges from 0 to 1, with values ​​closer to 1 indicating a higher degree of scene matching.

[0106] Release time weight The timeliness of regulations is reflected by a linear decay function, the formula of which is:

[0107] ;

[0108] in, It is the current system time (format "YYYY-MM-DD", which can be converted to a numeric value such as 20240520). This refers to the date the regulation was issued (same format). It is the pre-set validity period of the regulations (set according to the explicit value in the text of the regulations; if there is no explicit value, it defaults to 5 years, which is converted into a numerical form such as 1825 days).

[0109] If the regulations have expired. Directly exclude its application;

[0110] If the deadline has not passed, the weight decreases linearly over time, and newly released regulations are given priority.

[0111] S22.2, to Normalization process is performed to obtain ,based on Constructing a dynamic weight matrix ;

[0112] Specifically, first... Min-Max normalization is performed to eliminate dimensional differences. The formula is as follows:

[0113] ;

[0114] in, These are the normalized weights (values ​​from 0 to 1). , These are the minimum and maximum values ​​of the three-dimensional weights, respectively.

[0115] Based on the normalized weights, a 3×3 dynamic weight matrix is ​​constructed. The matrix structure is as follows:

[0116] ;

[0117] The diagonal elements of the matrix are the normalized three-dimensional weights, while the off-diagonal elements are set to 0 to ensure that each weight dimension acts independently and avoids cross-interference.

[0118] S22.3 Calculate the overall score (Score) for each conflicting clause under the same verification dimension, where the Score is associated with the dynamic weight matrix. Quantitative values ​​corresponding to the dimensions of each clause The clause with the highest overall score is selected as the adaptation result. When the score difference meets the preset threshold, the multimodal evidence cross-validation process is triggered.

[0119] Specifically, for conflicting clauses under the same verification dimension (such as Clause A "Number of stacking layers ≤ 3 layers" and Clause B "Number of stacking layers ≤ 2 layers"), first calculate the comprehensive score for each clause, using the following formula:

[0120] ;

[0121] in, It is the three-dimensional quantification value of the conflict clause ( This is a quantification of the hierarchical effectiveness of the clauses. To match quantified values ​​to the scene, (Quantized value after release time decay) yes The transpose of the matrix, A higher value indicates a stronger compatibility with the terms.

[0122] Furthermore, the clause with the highest score is selected as the initial adaptation result, while a score difference threshold is set (based on historical conflict resolution case statistics to ensure that the misjudgment rate is controllable): if the difference between the highest score and the second highest score is less than the threshold, multimodal evidence cross-validation is triggered - the heterogeneous graph association result of the multimodal evidence fusion module 24 is called to compare the matching degree of text evidence (such as the stacking requirements marked on the dangerous goods package certificate), image evidence (such as the image recognition result of the actual stacking layer) with the two conflicting clauses, and finally the clause with higher evidence support is selected as the adaptation result to avoid the bias of single weight judgment.

[0123] In this embodiment, the graph optimization and reconstruction module 23 periodically optimizes the weight parameters through a reinforcement learning model, while simultaneously monitoring regulatory updates and completing the dynamic reconstruction of the knowledge graph. The parameter optimization and graph reconstruction of the graph optimization and reconstruction module 23 include the following steps:

[0124] S23.1, Define the agent as the dynamic weight matrix constructed in S22.2. parameter set Define the state space Define the action space for the feature vectors of the historical verification scenario. For parameters Adjustment set ;

[0125] Specifically, the core elements of reinforcement learning are defined as follows:

[0126] Intelligent agent: defined as a dynamic weight matrix parameter set That is, the normalized weights of regulatory effectiveness, scenario matching, and release time; the optimization goal is to make... It can adapt to the regulatory judgment requirements in different scenarios;

[0127] state space Construct a 10-dimensional feature vector, with specific dimensions including cargo type code, packaging category code, transportation mode code, legal validity level, time since publication (days), number of historical conflicts, number of manual review and corrections, decision accuracy rate in the past 30 days, F1 score in the past 30 days, and scenario complexity (quantified by the number of regulations involved in the operation; the more regulations involved, the higher the complexity), to comprehensively reflect the characteristics of the current verification scenario.

[0128] Action space Defined as a parameter Adjustment set The adjustment range is set to [-0.1, 0.1], with an adjustment step of 0.01 each time to ensure that the parameter adjustment is stable and to avoid drastic fluctuations that could lead to sudden changes in the judgment result.

[0129] S23.2 Collect historical verification data from the past 3 years and divide it into training set and verification set according to a preset ratio. Each data point includes AI decision results and manual review results.

[0130] Specifically, historical verification data from the past three years of system operation was collected, with the following selection criteria: complete AI decision results, manual review results, scene feature vectors, and weight parameters. Data with missing key fields and duplicate or invalid data was removed. The selected data was then divided into a training set (for model training) and a validation set (for performance evaluation) in a 7:3 ratio. Each data point was organized into a "status" column. -action -award -Next state The sample format is "" to ensure sample integrity.

[0131] S23.3, Define the reward function ,in Correlation weight coefficient Decision accuracy With F1 score The DDPG algorithm is used, and samples are stored in an empirical replay buffer. Update parameters using gradient descent ;

[0132] Specifically, define the reward function. The performance index of the association model is given by the following formula:

[0133] ;

[0134] in, (Decision accuracy weight) (F1 score weighting) To improve the consistency rate between AI decisions and human review results, The F1 score (combined precision and recall) is the result of AI judgment. The higher the value, the better the parameter. The better the compatibility.

[0135] The model is trained using the DDPG (Deep Deterministic Policy Gradient) algorithm: the experience replay buffer size is set to 10000, and 32 samples are randomly sampled from the buffer for training each time; the Adam optimizer is selected with a learning rate of 1e-4, and the loss value of the reward function is minimized through gradient descent, and the parameters are continuously updated. This improves model adaptability.

[0136] S23.4. Set an optimization cycle. At the end of each cycle, use the validation set to evaluate the model performance. If the F1 score is lower than a preset threshold, trigger a new round of offline training and update the dynamic weight matrix. Parameters;

[0137] Specifically, the optimization cycle is set to one month, and the model performance is evaluated using the validation set at the end of each cycle: the Acc and F1 scores on the validation set are calculated, and if the F1 score is lower than a preset threshold (set based on historical best performance to ensure the accuracy of the system's judgment), a new round of offline training is immediately triggered—the training set data is reread, and the parameters are updated according to the DDPG algorithm process described above. And will update Substitute into the dynamic weight matrix Complete the iterative optimization of the weight parameters; if the F1 score meets the target, maintain the current value. If unchanged, proceed to the next optimization cycle.

[0138] S23.5. Use web crawling technology to retrieve regulations and announcements from official websites, and use the TF-IDF algorithm to extract keywords, including TF-IDF-related keywords. ,document Total number of documents With keywords Number of documents When the keyword matching degree meets the preset threshold, it is determined to be a relevant regulatory update, and the updated clauses and effective date are extracted. ;

[0139] Specifically, a regulatory monitoring crawler is built using a Python web scraping framework (such as Scrapy) to regularly crawl regulatory announcements from designated official platforms, including: the "Transportation of Dangerous Goods" section of the International Maritime Organization (IMO) website, the "Policies and Regulations" section of the Ministry of Transport of China's website, and the "National Standards Announcement" section of the State Administration for Standardization's website, ensuring coverage of major regulatory release channels.

[0140] Furthermore, the TF-IDF algorithm is used to extract keywords (such as "dangerous goods", "stacking specifications", "UN number") from the captured announcement text, and the TF-IDF calculation logic associates these keywords. ,document Total number of documents With keywords Number of documents The system uses keyword matching to determine whether the announcement is related to the compliance of dangerous goods packing. If the matching degree meets a preset threshold (based on statistical analysis of regulatory text features), it is determined to be a relevant regulatory update, and the content, clause number, and effective date of the updated clauses are further extracted. (Format "YYYY-MM-DD").

[0141] Furthermore, to accurately determine the relevance of regulatory updates, this embodiment uses the TF-IDF algorithm to quantify the semantic correlation between keywords and regulatory documents. The specific calculation logic is as follows:

[0142] Term Frequency (TF) calculation: Used to measure the frequency of keywords in a single regulatory document. The formula is:

[0143] ;

[0144] in, Keywords In the document The number of times it appears in For document The total number of occurrences of all words in the text; Keywords In a single regulatory document The word frequency in the document is used to quantify the frequency of the keyword in the document. The frequency of local occurrences within a document reflects its importance within that single document.

[0145] Inverse Document Frequency (IDF) calculation: Used to measure the scarcity of keywords in the entire set of regulatory documents, the formula is:

[0146] ;

[0147] in, It is a collection of regulatory documents. Total number of documents It contains keywords The number of documents can be incremented by 1 to avoid the denominator being 0; Keywords In the collection of regulatory documents Inverse document frequency in the data.

[0148] TF-IDF composite score: The final relevance score is obtained by weighting term frequency and inverse document frequency, using the following formula:

[0149] ;

[0150] in Keywords In the document and document collection The semantic relevance score is as follows.

[0151] Understandably, regarding the "keyword matching threshold," considering the semantic features of regulatory texts in the field of dangerous goods packing, the actual needs of relevance determination, and the conventional threshold setting logic in this field, this embodiment initially sets the threshold to 0.6. The core considerations for setting this threshold are: to ensure the effective identification of regulatory update-related documents while avoiding weakly related documents from mistakenly triggering subsequent processes, thus balancing identification coverage and judgment efficiency;

[0152] The specific judgment rule is as follows: when the combined TF-IDF score of the keyword and the document to be detected is ≥0.6, it is judged as "strongly relevant to the regulatory update", triggering the subsequent difference location and map optimization process; when the score is <0.6, it is judged as weakly relevant, and no update mark is triggered.

[0153] It should be noted that the above threshold is the initial configuration adapted to this embodiment. Those skilled in the art can adjust the threshold value according to factors such as the type of regulatory document, update frequency, and semantic complexity in the actual application scenario, through conventional verification methods (such as small-scale testing based on actual business data) to adapt to specific recognition accuracy and efficiency requirements.

[0154] S23.6. Set a transition time window and use the Sigmoid function to adjust the confidence levels of the old and new clauses. Among them, the confidence level correlation smoothing coefficient Current time With the effective date of the regulations ; Calculate the similarity of constraints between the old and new clauses. , When the preset similarity threshold is met, the old clause's associated edge is retained and points to the new clause; If the preset similarity threshold is not met, delete the associated edges of the old clauses and construct associated edges with the verification dimension nodes for the new clauses. Specifically, set a transition time window (e.g., 15 days before the regulation takes effect to 15 days after its effect), and use the Sigmoid function to adjust the confidence levels of the old and new clauses. The formula logic is as follows:

[0155] ;

[0156] in, The old terms are in effect at the current time. Confidence level, It's about the confidence level of the new terms. It is a smoothing coefficient (based on statistical settings of the regulatory transition period to ensure a smooth switch in confidence levels); within the transition window, Follow near Gradually decrease, Gradually increase the level to avoid abrupt changes in the judgment before and after the effect takes effect.

[0157] At the same time, the similarity of the constraints between the old and new clauses is calculated. (Using text similarity algorithms, such as cosine similarity, to compare the core content of the constraints):

[0158] like If the preset threshold is met (indicating that the new terms are an update of the old terms), the relevant edges of the old terms are retained and point to the new terms, ensuring that historical data is traceable;

[0159] If Sim does not meet the threshold (indicating that the new clause is an addition or replacement clause), delete the associated edges of the old clause, construct associated edges with the corresponding verification dimension nodes for the new clause, and complete the structural reconstruction of the knowledge graph.

[0160] In this embodiment, the multimodal evidence fusion module 24 is used to construct a heterogeneous graph from multi-source evidence, learn evidence associations through a graph attention mechanism, and output the difference localization result. The evidence fusion and difference localization of the multimodal evidence fusion module 24 includes the following steps:

[0161] S24.1 Map the text evidence, image evidence, and historical record evidence output by data acquisition and preprocessing unit 1 to text nodes, image nodes, and historical record nodes in a heterogeneous graph, respectively; generate feature vectors for the text evidence using the BERT model. A CNN model is used to generate feature vectors for video evidence. LSTM model is used to generate feature vectors for historical evidence. ;

[0162] Specifically, the three types of evidence output by the data acquisition and preprocessing unit 1 are mapped to nodes in the heterogeneous graph: textual evidence (such as dangerous goods packaging certificate text and legal clause text) is mapped to “text nodes”, image evidence (such as stacking images and actual label photos) is mapped to “image nodes”, and historical evidence (such as historical verification results and manual review records) is mapped to “historical record nodes”.

[0163] Furthermore, feature extraction is performed separately for different types of nodes:

[0164] The text nodes use the BERT-base-chinese model to generate 768-dimensional feature vectors. (The model parameters reuse the fine-tuning results of the triplet extraction indexing module 21 to ensure consistency in terminology recognition.)

[0165] The image nodes use a CNN model (such as ResNet-50, with pre-trained weights based on ImageNet, and then fine-tuned using dangerous goods packing image data) to generate 2048-dimensional feature vectors. ;

[0166] Historical data nodes use an LSTM model (256 hidden layer dimensions, training data consists of the last 3 years of historical data) to generate 256-dimensional feature vectors. This ensures that the feature dimensions of each node are consistent and semantically representative.

[0167] S24.2. Based on the original node feature vectors in S24.1, the initial correlation between nodes is calculated using the cosine similarity algorithm, and then transformed into initial attention coefficients. The initial edge weights of the heterogeneous graph are obtained by Softmax normalization. And based on the initial edge weights of the heterogeneous graph Construct the initial connection structure of the multimodal evidence heterogeneity graph;

[0168] Specifically, based on the original feature vectors of each node, the initial correlation between nodes is calculated using the cosine similarity algorithm (e.g., the correlation between text nodes and image nodes, reflecting the degree of matching between text descriptions and image content), and the correlation is directly converted into initial attention coefficients. ( For the target node, (Neighboring nodes).

[0169] Specifically, for Softmax normalization is performed to obtain the initial edge weights of the heterogeneous graph. The formula is:

[0170] ;

[0171] in, It is a node The set of neighboring nodes, Values ​​range from 0 to 1, representing nodes. and The initial correlation strength.

[0172] based on Constructing the initial connection structure of a multimodal evidence heterogeneity graph: retaining only Edges exceeding a preset threshold (based on evidence-related case statistics) are removed due to weak correlation, thus reducing redundant calculations.

[0173] Furthermore, in the multimodal feature fusion stage, to address the dimensionality differences between text features (768 dimensions, extracted by the BERT model), image features (2048 dimensions, extracted by the ResNet-50 model), and historical record features (256 dimensions, composed of TF-IDF and statistical features), the following fully connected layer is used to achieve dimensionality unification:

[0174] Fully connected text feature mapping layer: input dimension 768, output dimension 256, using ReLU activation function, and configured with a Dropout layer (dropout rate 0.2) to prevent overfitting;

[0175] Image feature mapping fully connected layer: input dimension 2048, output dimension 256, using ReLU activation function, and also configured with Dropout layer (dropout rate 0.2).

[0176] Historical data features remain unchanged at 256 dimensions, requiring no additional mapping. Through the dimensional mapping of the fully connected layer described above, it is ensured that the dimensions of text, images, and historical data are uniformly 256 before entering the graph attention mechanism, satisfying the computational requirement of the graph attention mechanism for the consistency of neighborhood feature dimensions.

[0177] S24.3, through the dynamic weight matrix A linear transformation is performed on the original node feature vectors to obtain enhanced node features. The initial graph structure and enhanced node features constructed based on S24.2 Combined with attention weight vector Using the LeakyReLU activation function, calculate the fine attention coefficients. ;

[0178] Specifically, through the dynamic weight matrix (The dimension is consistent with the node feature dimension, such as text nodes) A linear transformation is performed on the original node feature vectors (256×256 dimensions) to obtain enhanced node features. The formula is:

[0179] ;

[0180] in, It is a node The original feature vector, Key information in the features was enhanced (such as the "validity period" field in the text and the "stack layer number" feature in the image).

[0181] Furthermore, based on the initial graph structure and Combined with attention weight vector (dimensions and) Consistent (through training optimization) and the LeakyReLU activation function, the fine attention coefficients are calculated. The formula is:

[0182] ;

[0183] in, yes and The concatenated vector, yes The transpose of LeakyReLU ensures effective gradient propagation and avoids gradient vanishing.

[0184] S24.4, to At the node The set of neighboring nodes Perform Softmax normalization within the range to obtain fine-grained edge weights. ;by The weighting coefficients are used to evaluate the features of neighboring nodes. After weighted summation, the activation function is applied. Transformation to generate node embedding vectors that incorporate multimodal information. ;

[0185] Specifically, for exist Perform Softmax normalization within the range to obtain fine-grained edge weights. , the formula is the same as The normalization logic; and then with The weighting coefficients are used to evaluate the features of neighboring nodes. Weighted summation followed by Sigmoid activation function Transformation to generate node embedding vectors that incorporate multimodal information. The formula is:

[0186] ;

[0187] By integrating the correlation information of evidence from different modalities, it can more comprehensively reflect the compliance characteristics of the evidence.

[0188] S24.5, Calculate the node embedding vector Node features corresponding to the regulatory map The Euclidean distance is used as the deviation value ;like If the threshold is exceeded, the output includes the deviation node identifier, the associated evidence ID, the deviation value, and the difference location result of the regulatory inconsistency.

[0189] Specifically, compute the node embedding vector. Node features corresponding to the regulatory map The Euclidean distance is used as the deviation value. The formula is:

[0190] ;

[0191] in, It is the dimension of the feature vector. yes The 3D eigenvalues It is the first node feature corresponding to the regulatory map 3D eigenvalues The larger the value, the more significant the difference between the evidence and the regulatory requirements.

[0192] Furthermore, if If the deviation exceeds a preset threshold (set based on compliance judgment accuracy requirements), the difference location result will be output, including: deviation node identifier (e.g., "stacked image node-001"), associated evidence ID (e.g., "image evidence-20240520-003"), and deviation value (i.e., ...). Specific values ​​and discrepancies with regulations (such as "the actual number of stacking layers is 4 layers, while the regulation requires ≤3 layers") provide a clear basis for subsequent rectification.

[0193] Multi-agent collaborative verification unit 3, based on structured verification data sources and scenario-based dynamic regulatory knowledge graphs, completes single-dimensional compliance judgment and multi-dimensional comprehensive result summarization through collaborative reasoning of intelligent agent modules such as dangerous goods package certificate verification, UN number matching, stacking compliance judgment, and label identification verification.

[0194] In this embodiment, the multi-agent collaborative verification unit 3 includes a dangerous goods package certificate verification agent module 31, an UN number matching agent module 32, a stacking compliance judgment agent module 33, and a label identification verification agent module 34, wherein:

[0195] The dangerous goods packaging certificate verification intelligent agent module 31 verifies the compliance of the certificate's validity period, issuing agency qualifications, and packaging container model based on the electronic data of dangerous goods packaging certificates in the structured verification data source and the compliance rules of dangerous goods packaging certificates in the scenario-based dynamic regulatory knowledge graph, and generates compliance judgment results for the dangerous goods packaging certificate dimension. The dangerous goods packaging certificate verification intelligent agent module 31 includes a data parsing submodule, a rule comparison submodule, and an anomaly tracing submodule, among which:

[0196] The data parsing submodule uses OCR recognition technology and structured data parsing technology to extract the core fields directly related to the compliance verification of dangerous goods packaging certificates from the electronic data of dangerous goods packaging certificates;

[0197] Specifically, based on OCR recognition technology and structured data parsing technology, core fields directly related to compliance verification are extracted from the electronic data of dangerous goods packaging certificates, such as certificate number, issuance date, validity period (start date + end date), issuing agency name, issuing agency qualification level, packaging container model, packaging container material, and applicable goods category (corresponding to GB6944 code). Date fields are uniformly converted to the "YYYY-MM-DD" format to be consistent with the standardized format of data collection and preprocessing unit 1. Redundant spaces and special symbols are removed from text fields such as agency name and container model to ensure that the parsed data format is uniform and can be directly used for subsequent comparison. If a core field is missing, it is marked as "field missing" to provide a basis for subsequent anomaly tracing.

[0198] The rule comparison submodule calls the dangerous goods packaging certificate compliance rule library in the scenario-based dynamic legal knowledge graph and compares the parsed fields with the "validity period, list of issuing agency qualifications, and list of compliant packaging container models" in the dangerous goods packaging certificate compliance rule library one by one;

[0199] Specifically, the system uses a standardized data interface to access the "Dangerous Goods Packaging Certificate Compliance Rule Library" within a scenario-based dynamic regulatory knowledge graph. This rule library is generated by matching a three-dimensional index of "goods category - packaging category - mode of transport" and includes three core rule categories: validity period rules, issuing agency qualification rules, and packaging container model compliance rules. The system verifies and parses each field in the order of "validity period - issuing agency qualification - packaging container model." The validity period rule requires that "the current system time must be within the validity period of the dangerous goods packaging certificate and not exceed the maximum validity period stipulated by regulations." The issuing agency qualification rule requires that "the parsed issuing agency must be in the list of legally qualified agencies and its qualification level must match the hazard level of the goods." The packaging container model compliance rule requires that "the parsed container model must be in the compliance list of the corresponding goods category and its material must meet regulatory requirements." The comparison results are categorized into three types: "compliant," "non-compliant," and "pending review." "Pending review" only applies to rule boundary cases such as the validity period ending at the current date.

[0200] When the anomaly tracing submodule determines an anomaly through rule comparison, it automatically associates real-time data collected from cargo packaging with historical verification records. Through multi-dimensional data correlation analysis, it locates the root cause of the anomaly and outputs a tracing report containing "anomaly type, legal basis, and rectification suggestions." This report is then synchronized to the UN number matching intelligent agent module 32, the stacking compliance determination intelligent agent module 33, and the label identification verification intelligent agent module 34, triggering a multi-agent joint verification process.

[0201] Specifically, when the rule comparison determines that the violation is "in violation" or "pending review," the traceability process is automatically initiated: Real-time data collected on cargo packaging (including original scanned copies of the dangerous goods packaging certificate, actual photos of the packaging containers, and operator notes) and historical verification records of similar goods from the same issuing agency over the past three years are retrieved through associated fields such as container identification number and cargo order ID; the root cause of the anomaly is located through multi-dimensional data correlation analysis, and the root cause category is subdivided according to the anomaly type (e.g., validity period violations are categorized as "expired," "not in effect," and "exceeding the regulatory limit"); a standardized traceability report containing "anomaly type, associated regulatory clause number, anomaly data details, root cause analysis, and rectification suggestions" is generated; the report is synchronized to the UN number matching intelligent agent module 32, the stacking compliance judgment intelligent agent module 33, and the label identification verification intelligent agent module 34 through the system's internal data interaction interface, triggering a joint verification process. For example, when the packaging is made of fragile materials, the stacking intelligent agent needs to additionally verify the crushing risk, and the label intelligent agent needs to verify the "fragile" warning symbol.

[0202] The UN number matching intelligent agent module 32 compares the actual UN number of the goods with the regulatory requirements based on the UN number information of the goods in the structured verification data source and the UN number association rules in the scenario-based dynamic regulatory knowledge graph, and generates a compliance judgment result in the UN number dimension.

[0203] Specifically, the UN number matching intelligent agent module 32 extracts the actual UN number of the goods (including OCR-recognized text and the result of extracting the label real-shot image) and basic information of the goods (goods category code, goods name, and description of hazardous characteristics) from the structured verification data source. It performs format standardization processing on the actual UN number, uniformly converting it into the "UN + 4-digit number" format and removing non-standard redundant characters. It calls the "UN number association rule library" in the scenario-based dynamic regulatory knowledge graph and obtains the set of compliant UN numbers and number association constraints by matching the three-dimensional index of "goods category code - packaging category - mode of transport". It adopts a two-level comparison logic of "complete matching + association matching" to first verify whether the actual UN number is completely consistent with the compliant set. If they are inconsistent, it further verifies whether they belong to the range of associated numbers and whether the hazardous characteristics match. It generates a compliance judgment result of the UN number dimension, clearly marking the compliance, the type of abnormality (such as number mismatch, format error, missing) and the corresponding regulatory basis.

[0204] The intelligent agent module 33 for stacking compliance judgment analyzes the compliance of the number of stacking layers and stacking method of goods based on the cargo stacking data in the structured verification data source and the stacking specifications in the scenario-based dynamic regulatory knowledge graph, and generates the stacking dimension compliance judgment result.

[0205] Specifically, the intelligent agent module 33 for judging stacking compliance extracts core data such as the number of stacking layers, stacking method, stacking stability status, cargo packaging size, packaging weight, and packaging category (I / II / III) from the structured verification data source;

[0206] The system invokes the "stacking compliance rule library" in the scenario-based dynamic regulatory knowledge graph, matching the three-dimensional index of "UN number-packaging category-transportation method" to obtain constraint rules such as the maximum number of stacking layers, allowed stacking methods, stacking spacing requirements, and weight limits. Compliance analysis is completed in three steps: verifying whether the number of stacking layers exceeds the upper limit and the load-bearing capacity of the lower packaging is within the rated range; verifying whether the stacking method is compliant and the stability status is "stable"; and verifying whether the stacking spacing meets the requirements. A compliance judgment result for the stacking dimension is generated, marking anomalies (such as exceeding the upper limit or improper stacking method), compliant stacking parameter suggestions, and regulatory clause numbers.

[0207] The label identification verification intelligent agent module 34 verifies the compliance of the content, format, and placement of goods labels based on label identification data in the structured verification data source and label identification rules in the scenario-based dynamic legal knowledge graph, and generates a label identification dimension compliance judgment result.

[0208] Specifically, the label identification verification intelligent agent module 34 extracts core data such as label content (UN number, hazard category, warning symbol name, etc.), format features (font size, color, size, material), posting location information, and number of labels from the structured verification data source; it calls the "Label Identification Compliance Rule Library" in the scenario-based dynamic regulatory knowledge graph, and obtains the constraint rules such as content requirements, format requirements, and posting location requirements by matching the "UN number-hazard category" index; it verifies each rule in the order of "content-format-location": confirming that the label contains all mandatory fields and the UN number is consistent with the actual label, the format meets the requirements such as font size ≥5mm and yellow background with black image, and the posting location meets the requirements of "centered on two opposite sides, unobstructed, and ≥2cm from the edge"; it generates a label identification dimension compliance judgment result, marks abnormal points (such as missing content, non-compliant format, and position offset), specific rectification guidelines, and regulatory basis.

[0209] It should be added that after the dangerous goods packaging certificate verification agent module 31, the UN number matching agent module 32, the stacking compliance judgment agent module 33, and the label identification verification agent module 34 complete their single-dimensional compliance judgments, the results are summarized and integrated through the system's internal collaborative interface: if all dimensions comply with the rules, the overall judgment is "packing operation compliant"; if any dimension violates the rules, the overall judgment is "packing operation non-compliant", and the abnormal items are sorted by priority according to "dangerous goods packaging certificate-UN number-stacking-label"; if a certain agent triggers the joint verification process, other agents adjust the verification logic based on the abnormal information; finally, a "multi-dimensional compliance verification summary report" is generated, which includes the judgment results of each dimension, a list of abnormal items, legal basis, and priority of rectification suggestions, providing a complete basis for risk warning and rectification.

[0210] Risk warning report generation unit 4, based on the multi-agent collaborative verification results, triggers warning signals and automatically generates a regulatory matching report with explanatory explanations through a hierarchical risk triggering mechanism and standardized report template;

[0211] In this embodiment, the risk warning report generation unit 4 includes a verification result parsing module 41, a risk classification determination module 42, and a standardized report generation module 43, wherein:

[0212] The verification result parsing module 41 receives the compliance judgment results of various dimensions output by the multi-agent collaborative verification unit 3, extracts key information directly related to the compliance judgment, including violation type, violation-related data, and regulatory matching basis, and completes the structured integration of the results;

[0213] Specifically, the verification result parsing module 41 receives, through the system's internal data interface, single-dimensional compliance judgment results output by the dangerous goods packaging certificate verification agent module 31, the UN number matching agent module 32, the stacking compliance judgment agent module 33, and the label identification verification agent module 34, as well as the comprehensive judgment conclusion summarized by multiple agents. During the parsing process, it follows the logic of "dimensional classification - key information extraction - structured integration," as follows:

[0214] First, the results are analyzed and determined according to four dimensions: “dangerous goods packaging certificate - UN number - stacking - label”, to ensure that the information in each dimension is independent and traceable.

[0215] Then, extract three types of core information for each dimension: violation type (such as "expired" dangerous goods packaging certificate, "exceeding the upper limit" stacking), violation-related data (including specific abnormal values, related evidence IDs, cargo order numbers, container identification numbers, etc., such as "actual stacking layers are 4, the compliant upper limit is 3 layers" and "related image evidence ID: IMG-20240520-008"), and regulatory matching basis (from the clause number and core content of the clause in the scenario-based dynamic regulatory knowledge graph, such as "GB30000.15-2013 Article 5.3: The stacking layers of Class I packaging shall not exceed 3 layers").

[0216] Finally, through data association technology, multi-dimensional violation information of the same goods is bound by "order ID-container identification number" to form a structured integration result, which provides complete data support for subsequent risk classification. The integration result is stored in a standardized JSON format for easy cross-module access.

[0217] The risk assessment module 42 is based on the priority of compliance clauses in the scenario-based dynamic legal knowledge graph, combined with the degree of impact of violations on packing safety, and triggers corresponding level warning signals through preset risk assessment thresholds.

[0218] Specifically, the risk assessment module 42 conducts risk classification based on two core criteria: first, the "compliance clause priority" pre-stored in the scenario-based dynamic legal knowledge graph (determined by the legal validity level and safety impact weight, such as clauses involving personal safety having higher priority than compliance filing clauses); and second, the degree of impact of violations on packing safety (quantified according to four dimensions: "personal safety risk - environmental hazard risk - compliance penalty risk - operational efficiency impact," with each dimension having a scoring standard of 0-10 points, such as "violations may lead to explosion" corresponding to 10 points for personal safety risk, and "non-standard label format" corresponding to 3 points for operational efficiency impact).

[0219] Meanwhile, three risk thresholds are preset: high risk (overall score ≥ 8 points), medium risk (4 points ≤ overall score < 8 points), and low risk (overall score < 4 points). The overall score is calculated as: priority weight of compliance clauses × total score of impact (priority weights are set according to the level of legal effectiveness: international regulations 0.4, national regulations 0.3, industry standards 0.2, and corporate regulations 0.1).

[0220] In addition, when assessing risks, the overall score for each violation is calculated first, and then the overall risk level is determined according to the principle of "highest risk priority" (e.g., if a certain commodity has both high-risk and medium-risk violations, it is judged as high-risk overall). Subsequently, the corresponding level of warning signal is triggered: high risk triggers "audio-visual alarm + system pop-up + management personnel SMS notification", medium risk triggers "system pop-up + department head in-system message", and low risk only triggers "system log recording + operator reminder", ensuring that different levels of risk receive differentiated responses.

[0221] The standardized report generation module 43 calls the preset compliance report template, integrates the structured verification results, graded risk levels, specific regulatory clauses corresponding to violations, interpretable compliance judgment explanations, and targeted rectification guidelines, and automatically generates a complete regulatory matching risk warning report.

[0222] Specifically, the standardized report generation module 43 calls the preset compliance report template, automatically integrates the structured verification data from the verification result analysis module 41, the risk level and warning signals from the graded risk assessment module 42, the specific regulatory clauses corresponding to each violation, the interpretable compliance assessment explanation (including AI assessment logic), and the targeted rectification guidelines sorted by priority, to generate a complete regulatory matching risk warning report. The report includes core content such as basic cargo information, comprehensive assessment results, multi-dimensional violation details, risk descriptions, and rectification requirements. It automatically links to original evidence and supports exporting and printing in PDF format. At the same time, it is pushed to relevant personnel according to preset permissions to meet archiving and regulatory inspection needs.

[0223] Human-machine collaboration closed-loop unit 5: Human-machine collaboration closed-loop unit 5 establishes an interactive interface between AI judgment results and manual review, and feeds back the manual review results to the regulatory map construction iteration unit 2.

[0224] Specifically, the human-machine collaboration closed-loop unit 5 focuses on the core logic of "AI judgment - human review - data feedback," building a lightweight and efficient interaction and optimization closed loop that ensures the necessity of human intervention while avoiding process redundancy, as detailed below:

[0225] An interactive interface for manual review, integrated into the system's operating platform, is built. The core of the interface displays a "List of Tasks to be Reviewed" and a "Review Operation Area." Tasks to be reviewed are automatically filtered into three scenarios: high-risk violations (mandatory manual review), AI-judged boundary cases (such as ambiguous rules or questionable data), and operator appeals. The list clearly indicates task priority, order number, AI judgment result, and quick access to original evidence. The review operation area allows viewing the complete AI judgment logic, violation-related data, and legal basis. It provides two core operations: "Agree to AI Judgment" and "Correct AI Judgment." "Correct AI Judgment" requires a brief reason for correction (such as "evidence recognition deviation" or "incorrect application of regulations") and the correct judgment result, eliminating the need for complex data entry.

[0226] The manual review process is simplified and executed according to the "task push - verification - judgment - submission" method: The system pushes the task to the designated reviewer (safety management specialist or department head) through the in-system message. The reviewer can quickly retrieve relevant information and original evidence (scanned copy of dangerous goods packaging certificate, image, etc.) through the interface, make a judgment based on experience, and submit the result.

[0227] The review results are fed back to the regulatory graph construction iteration unit 2 in real time: cases that are "agreed upon by AI" are automatically used as positive samples and added to the training set of the reinforcement learning model; cases that are "corrected by AI" are used as negative samples, and the reasons for the deviation are labeled simultaneously to optimize the dynamic weight matrix parameters and the rule association of the regulatory knowledge graph; the entire feedback process does not require any additional manual operation, and the system automatically completes the data flow, realizing a lightweight closed loop of "review once, optimize once", and continuously improving the accuracy and adaptability of AI judgment.

[0228] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0229] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent compliance verification system for dangerous goods packing operations, characterized in that, include: The data acquisition and preprocessing unit (1) uses optical character recognition and image preprocessing technology to acquire dangerous goods packaging certificate scans, UN number information, stacking images, label real photos and packing basic data and performs standardized conversion, and stores the structured verification data source in the database; The regulatory map construction iteration unit (2) extracts the "regulatory clause number-verification dimension-constraint condition" triple of the regulations in the field of dangerous goods packing and indexes them according to the level of effectiveness, issuing entity, and scope of application. It introduces an improved three-dimensional dynamic weight conflict resolution algorithm, constructs a dynamic weight matrix based on the legal effectiveness, scenario matching degree, and release time, and calculates the comprehensive score of each conflict clause under the same verification dimension. The comprehensive score is associated with the dynamic weight matrix and the quantitative value of the corresponding dimension of each clause. The clause with the highest comprehensive score is selected as the adaptation result. When the score difference meets the preset threshold, the multimodal evidence cross-verification process is triggered. The parameters of the dynamic weight matrix are periodically optimized by the offline trained reinforcement learning model, and the reinforcement learning model uses the consistency between the historical verification decision and the manual review result as the reward signal. By monitoring the announcements of regulatory updates, and based on the preset version lifecycle rules and time windows, the confidence level of the new and old clauses in the knowledge graph is automatically switched and the association is reconstructed; and the text, image and historical evidence of the data collection and preprocessing unit (1) are constructed into a heterogeneous graph, and the evidence association is learned through the graph attention mechanism to output interpretable difference localization results. Multi-agent collaborative verification unit (3), the multi-agent collaborative verification unit (3) is based on structured verification data source and scenario-based dynamic regulatory knowledge graph, and completes single-dimensional compliance judgment and multi-dimensional comprehensive result summary through collaborative reasoning of intelligent agent modules such as dangerous goods package certificate verification, UN number matching, stacking compliance judgment and label identification verification; Risk warning report generation unit (4), the risk warning report generation unit (4) triggers warning signals and automatically generates a regulatory matching report with explanatory explanations based on the multi-agent collaborative verification results, through a hierarchical risk triggering mechanism and standardized report template; Human-machine collaborative closed-loop unit (5) builds an interactive interface between AI judgment results and manual review, and feeds back the manual review results to the regulatory map construction iteration unit (2). 2.The dangerous goods packing operation intelligent compliance verification system according to claim 1, characterized in that, The data acquisition and preprocessing unit (1) includes a multi-source data acquisition module (11) and a data standardization and transformation module (12), wherein: The multi-source data acquisition module (11) uses optical character recognition technology to acquire text data of dangerous goods packaging certificate scans and UN number information, uses image preprocessing technology to acquire image data of stacking images and actual label photos, and simultaneously acquires container basic data. The data standardization and conversion module (12) performs format normalization on the collected text data, extracts features and performs structure conversion on the image data, and stores the processed data in the database to form a structured verification data source. 3.The dangerous goods packing operation intelligent compliance verification system according to claim 2, characterized in that, The regulatory map construction iteration unit (2) includes a triplet extraction and indexing module (21), a dynamic weight conflict resolution module (22), a map optimization and reconstruction module (23), and a multimodal evidence fusion module (24), wherein: The triple extraction indexing module (21) is used to extract the triple "legal clause number-verification dimension-constraint condition" and to perform structured indexing of validity level, issuing entity, and scope of application; The dynamic weight conflict resolution module (22) constructs a dynamic weight matrix based on the legal effect, scenario matching degree, and release time, and quantifies and resolves conflicting clauses. The graph optimization and reconstruction module (23) periodically optimizes the weight parameters through a reinforcement learning model, while monitoring regulatory updates and completing the dynamic reconstruction of the knowledge graph. The multimodal evidence fusion module (24) is used to construct a heterogeneous graph from multi-source evidence, learn evidence association through graph attention mechanism and output differential localization results. 4.The dangerous goods packing operation intelligent compliance verification system according to claim 3, characterized in that, The triplet extraction and structured indexing of the triplet extraction and indexing module (21) includes the following steps: S21.

1. A hybrid extraction method combining regular rules and a fine-tuned BERT model is adopted. After matching the regulatory clause number and filtering redundant text, the target text containing verification dimensions and constraints is input into the BERT model. S21.2, calculate the class probability of the verification dimension and the constraint condition by the Softmax function ; S21.

3. Associate and combine constraints according to the format of "legal clause number-verification dimension-constraint condition", and split the constraints into three types of structured data: numerical, logical, and range. S21.4 Mapping effectiveness level quantification coefficients according to preset rules ,in Hierarchical binary identifier With corresponding level preset weight The association is obtained; a three-dimensional index vector is constructed according to "goods type - packaging category - transportation method". . 5.The dangerous goods packing operation intelligent compliance verification system according to claim 4, characterized in that, The construction of the weight matrix of the dynamic weight conflict resolution module (22) includes the following steps: S22.1, Weight of Legal Validity The effectiveness level quantification coefficients output from S21.4 are used. ; Calculate scene matching weights using cosine similarity ,in Vector of Applicable Regulations Vector of the current packing operation scenario ; Calculate the release time weight using a linear decay function ,in Current time of associated system Date of promulgation of regulations With respect to the pre-set validity period of regulations ; S22.2, to Normalization processing is performed to obtain , based on A dynamic weight matrix is constructed . 6.The dangerous goods packing operation intelligent compliance verification system according to claim 5, characterized in that, The parameter optimization and map reconstruction of the map optimization and reconstruction module (23) includes the following steps: S23.1, Define the agent as the dynamic weight matrix constructed in S22.

2. parameter set Define the state space Define the action space for the feature vectors of the historical verification scenario. For parameters Adjustment set Define the next state To perform the action The resulting new historical verification scene feature vector; S23.2 Collect historical verification data from the past 3 years and divide it into training set and verification set according to a preset ratio. Each data point includes AI decision results and manual review results. S23.3, Define the reward function ,in Correlation weight coefficient Decision accuracy With F1 score The DDPG algorithm is used, and samples are stored in an empirical replay buffer. Update parameters using gradient descent ; S23.

4. Set an optimization cycle. At the end of each cycle, use the validation set to evaluate the model performance. If the F1 score is lower than a preset threshold, trigger a new round of offline training and update the dynamic weight matrix. Parameters; S23.

5. Use web crawling technology to retrieve regulations and announcements from official websites, and use the TF-IDF algorithm to extract keywords, including TF-IDF-related keywords. ,document Total number of documents With keywords Number of documents When the keyword matching degree meets the preset threshold, it is determined to be a relevant regulatory update, and the updated clauses and effective date are extracted. ; S23.

6. Set a transition time window and use the Sigmoid function to adjust the confidence levels of the old and new clauses. Among them, the confidence level correlation smoothing coefficient Current time With the effective date of the regulations ; Calculate the similarity of constraints between the old and new clauses. , When the preset similarity threshold is met, the old clause's associated edge is retained and points to the new clause; If the preset similarity threshold is not met, delete the associated edges of the old clauses and construct associated edges with the verification dimension nodes for the new clauses.

7. The dangerous goods packing operation intelligent compliance verification system according to claim 6, characterized in that, The evidence fusion and difference localization of the multimodal evidence fusion module (24) includes the following steps: S24.1 Map the text evidence, image evidence, and historical record evidence output by the data acquisition and preprocessing unit (1) to text nodes, image nodes, and historical record nodes of the heterogeneous graph, respectively; and use the BERT model to generate feature vectors for the text evidence. A CNN model is used to generate feature vectors for video evidence. LSTM model is used to generate feature vectors for historical evidence. ; S24.2, based on the original node feature vector of S24.1, the initial correlation degree between nodes is calculated by cosine similarity algorithm, which is converted into initial attention coefficient , the initial edge weight of the heterogeneous graph is obtained by Softmax normalization , and the initial connection structure of the multi-modal evidence heterogeneous graph is constructed based on the initial edge weight of the heterogeneous graph . S24.3, through the dynamic weight matrix Linear transformation is performed on the original node feature vector to obtain an enhanced node feature ; based on the initial graph structure and the enhanced node feature constructed in S24.2 , combined with the attention weight vector and the LeakyReLU activation function, the fine attention coefficient is calculated ; S24.4、to In the node neighbor node set The Softmax normalization is performed within the range of the neighbor node set, and the fine edge weight is obtained ; The neighbor node features are weighted and summed with the weighted coefficient , and then converted through the activation function to generate the node embedding vector fused with multi-modal information ; S24.5, Calculate the node embedding vector Node features corresponding to the regulatory map The Euclidean distance is used as the deviation value ;like If the threshold is exceeded, the output includes the deviation node identifier, the associated evidence ID, the deviation value, and the difference location result of the regulatory inconsistency. 8.The dangerous goods packing operation intelligent compliance verification system according to claim 1, characterized in that, The multi-agent collaborative verification unit (3) includes a dangerous goods certificate verification agent module (31), an UN number matching agent module (32), a stacking compliance judgment agent module (33), and a label identification verification agent module (34), wherein: The dangerous goods packaging certificate verification intelligent agent module (31) verifies the compliance of the dangerous goods packaging certificate validity period, the issuing agency qualification, and the packaging container model based on the electronic data of the dangerous goods packaging certificate in the structured verification data source and the dangerous goods packaging certificate compliance rules in the scenario-based dynamic legal knowledge graph, and generates a dangerous goods packaging certificate dimension compliance judgment result. The UN number matching intelligent agent module (32) compares the consistency between the actual UN number of the goods and the regulatory requirements based on the UN number information of the goods in the structured verification data source and the UN number association rules in the scenario-based dynamic regulatory knowledge graph, and generates a compliance judgment result for the UN number dimension. The stacking compliance judgment intelligent agent module (33) analyzes the compliance of the number of stacking layers and stacking method of goods based on the cargo stacking data in the structured verification data source and the stacking specifications in the scenario-based dynamic regulatory knowledge graph, and generates the stacking dimension compliance judgment result. The label identification verification intelligent agent module (34) verifies the compliance of the content, format, and posting location of the goods label identification based on the label identification data in the structured verification data source and the label identification rules in the scenario-based dynamic legal knowledge graph, and generates a label identification dimension compliance judgment result.

9. The intelligent compliance verification system for dangerous goods packing operations according to claim 8, characterized in that, The dangerous goods packaging certificate verification intelligent agent module (31) includes a data parsing submodule, a rule comparison submodule, and an anomaly tracing submodule, wherein: The data parsing submodule extracts core fields directly related to the compliance verification of dangerous goods packaging from the electronic data of dangerous goods packaging certificates based on OCR recognition technology and structured data parsing technology. The rule comparison submodule calls the dangerous goods packaging certificate compliance rule library in the scenario-based dynamic legal knowledge graph, and compares the parsed fields with the "validity period, list of issuing agency qualifications, and list of packaging container model compliance" in the dangerous goods packaging certificate compliance rule library one by one; When the anomaly tracing submodule determines an anomaly through rule comparison, it automatically associates real-time data collected from cargo packaging with historical verification records. Through multi-dimensional data association analysis, it locates the root cause of the anomaly and outputs a tracing report containing "anomaly type, legal basis, and rectification suggestions". It also synchronizes the report to the UN number matching agent module (32), the stacking compliance determination agent module (33), and the label identification verification agent module (34), triggering a multi-agent joint verification process.

10. The intelligent compliance verification system for dangerous goods packing operations according to claim 9, characterized in that, The risk warning report generation unit (4) includes a verification result parsing module (41), a risk classification judgment module (42), and a standardized report generation module (43), wherein: The verification result parsing module (41) receives the compliance judgment results of each dimension output by the multi-agent collaborative verification unit (3), extracts key information directly related to the compliance judgment, including violation type, violation-related data, and regulatory matching basis, and completes the structured integration of the results; The risk assessment module (42) triggers a corresponding level warning signal based on the priority of compliance clauses in the scenario-based dynamic legal knowledge graph and the degree of impact of violations on packing safety by setting a pre-defined risk threshold. The standardized report generation module (43) calls the preset compliance report template, integrates the structured verification results, graded risk levels, specific regulatory clauses corresponding to violations, interpretable compliance judgment explanations and targeted rectification guidelines, and automatically generates a complete regulatory matching risk warning report.

Images