A port non-inductive customs clearance verification method and system based on multi-modal data fusion

Abstract

This invention relates to the field of intelligent immigration management, specifically disclosing a port-based contactless clearance verification method and system based on multimodal data fusion. The method includes: collecting and processing multimodal data from passengers to generate a standardized set of evidence elements containing identity, behavior, and historical risks; inputting the initial set of evidence elements into a risk fusion decision model for judgment, automatically generating a structured dynamic decision evidence chain based on the model's decision logic; the system executing tiered clearance instructions according to risk levels and simultaneously pushing the dynamic decision evidence chain to the corresponding processing terminal; the system collecting manual processing results and the corresponding evidence chain, continuously optimizing the risk decision model through online learning; the system includes four corresponding modules. This invention realizes the transformation of verification decision-making from a black box to a white box and constructs a closed-loop self-optimization mechanism, significantly improving clearance efficiency, processing effectiveness, and system adaptability while ensuring security.

Description

Technical Field

[0001] This invention relates to the field of intelligent immigration management technology, and in particular to a port-based contactless customs clearance verification method and system based on multimodal data fusion. Background Technology

[0002] In recent years, to improve port clearance efficiency, contactless clearance systems based on biometrics and radio frequency identification have been widely used. These existing technologies typically employ multimodal data fusion schemes, such as simultaneously collecting passengers' facial and iris biometric features along with their electronic document chip information, achieving "person-document matching" verification through rapid comparison. If the comparison is successful and the passenger is not on a blacklist, passage is automatically granted; otherwise, an alarm is triggered and the passenger is blocked.

[0003] However, existing technical solutions of this kind suffer from two prominent and interconnected technical problems. First, their verification logic is inherently static and rigid, relying primarily on preset comparison thresholds and risk lists for binary "black and white" decisions. This fails to provide refined risk assessment and tiered handling for passengers who pass identity verification but exhibit abnormal behavior or possess complex risk profiles. Second, and more critically, the decision-making process is completely opaque. When the system issues an instruction to intercept or guide a review, the specific decision-making basis behind it (e.g., whether it's insufficient confidence in biometric matching, a specific behavioral pattern triggering a warning, or historical risk tags) cannot be presented to on-site staff in a timely and intuitive manner. This forces customs officers to often begin manual questioning and investigation from the beginning when faced with alerts, resulting in low efficiency and an inability to provide clear explanations to passengers, impacting the customs clearance experience and public trust in management.

[0004] Therefore, there is an urgent need for a new, seamless customs clearance verification method that can not only achieve accurate risk assessment, but also make the decision-making process traceable and explainable, and can continuously optimize itself using business feedback. Summary of the Invention

[0005] The purpose of this invention is to solve the core problems of low processing efficiency, insufficient public trust, and inability of the model to self-optimize caused by the lack of transparency in the decision-making process of existing contactless customs clearance systems. In response, this invention proposes a port contactless customs clearance verification method and system based on multimodal data fusion.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A port-based contactless customs clearance verification method based on multimodal data fusion includes: S1. Collect passenger data and process it to generate an initial set of evidence elements; S2. Input the initial set of evidence elements into the risk model and simultaneously output the risk level and dynamic decision-making evidence chain; S3. Execute the instruction corresponding to the risk level and synchronize the dynamic decision evidence chain to the corresponding terminal; S4. Recover the dynamic decision evidence chain labeled with the final disposal result to optimize the risk model.

[0007] As a further technical solution of the present invention, S1 specifically includes: S11. Real-time collection of biometric data, electronic document data, and behavioral posture data of passengers passing through the preset contactless channel; S12. Compare the biometric data and electronic certificate data to generate identity verification result elements; analyze the behavioral posture data to generate behavioral compliance assessment elements; and simultaneously query the passenger's risk profile to generate historical risk profile elements. S13. The identity verification result elements, behavioral compliance assessment elements, and historical risk profile elements together constitute the initial evidence element set.

[0008] As a further technical solution of the present invention, S12 specifically includes: S121. Perform multimodal fusion comparison between the facial image, iris image, and voiceprint data and the corresponding biometric templates in the electronic certificate data, and output identity verification result elements including similarity score and liveness detection result; S122. Input the behavioral posture data into the posture analysis model trained by abnormal behavior samples, extract the gait cycle abnormality, gaze deflection frequency, and dwell time in specific areas, and output quantitative behavioral compliance assessment elements. S123. Using the document number in the electronic document data as an index, perform parallel queries to the port risk database, international travel record database, and related case database, return risk labels, historical passage frequency, and related attention, and generate structured historical risk profile elements.

[0009] As a further technical solution of the present invention, S2 specifically includes: S21. The initial set of evidence elements is simultaneously input into the risk calculation engine and the evidence chain construction engine of the risk fusion decision model, and the risk calculation engine generates the final risk level determination result. S22. The evidence chain construction engine receives intermediate data generated by the risk calculation engine in real time during the calculation process. The intermediate data includes at least the weighted scores of each sub-element and its contribution to the total risk score. Based on the internal logic and weight relationship, it initiates a multi-stage processing flow and automatically encapsulates it into a structured dynamic decision evidence chain.

[0010] As a further technical solution of the present invention, in step S21, the risk calculation engine generates the final risk level determination result, specifically including: The risk calculation engine performs weighted fusion calculations on the similarity scores in the identity verification result elements, the liveness detection results, the quantitative values ​​of each feature in the behavior compliance assessment elements, and the risk labels in the historical risk profile elements, based on pre-configured element weights. Combined with preset multi-level thresholds, it outputs the final risk level determination result.

[0011] As a further technical solution of the present invention, the multi-stage processing flow in S22 specifically includes: The evidence chain building engine first parses the intermediate data packets transmitted by the risk calculation engine and extracts the standardized contribution value of each sub-element; Based on the contribution level, the sub-elements in the initial evidence element set are sorted and associated. After sorting, the engine establishes the logical relationship between the elements. Based on the element association, the engine encapsulates the logical association relationship and the final risk level determination result into a dynamic decision evidence chain in JSON-LD or Protocol Buffers format; The dynamic decision evidence chain explicitly records key elements, the contribution weight of each element, and the decision logic path. The decision logic path is recorded in the form of a directed acyclic graph, where nodes represent calculation steps and edges represent data flow. After encapsulation, the evidence chain building engine pushes the dynamic decision evidence chain to the terminal that matches the customs clearance control command. When the passage is released without human intervention, it is archived to the background log. When the passage is guided for review or manual intervention, it is pushed to the reviewer's terminal or the mobile terminal of the on-site law enforcement personnel.

[0012] As a further technical solution of the present invention, S3 specifically includes: S31. Generate corresponding clearance control instructions based on the final risk level determination result and send them to the channel control unit for execution; wherein, the clearance control instructions include: seamless release instructions, guided review instructions, and manual intervention instructions; S32. While executing the customs clearance control command, adapt the dynamic decision evidence chain according to the command type and push it synchronously: For the seamless passage instruction, the dynamic decision evidence chain is bound to the passage record, compressed, and archived to the background log database; For the guided review instruction, the visualization template engine is invoked to convert the dynamic decision evidence chain into an HTML5 format visualization report that highlights key risk elements, and then pushes it to the designated back-end reviewer's terminal via a message queue; For the manual intervention command, the dynamic decision-making evidence chain and real-time video stream segments are packaged and pushed to the mobile terminal of the on-site law enforcement personnel through a low-latency communication protocol, and a visual report is displayed on the terminal interface in the form of an overlay layer.

[0013] As a further technical solution of the present invention, S4 specifically includes: S41. For cases where the guided review instruction or manual intervention instruction is executed, the system receives and records the final handling result confirmed after manual verification at the back-end reviewer's terminal or the mobile terminal of the on-site law enforcement personnel; the final handling result is used as an authoritative label and uniquely bound to the dynamic decision-making evidence chain corresponding to the case generated in step S2 to form a labeled training sample pair. S42. The accumulated training sample pairs are input into the online learning module of the risk fusion decision model; the online learning module calculates the prediction bias of the model on each sub-element of the initial evidence element set and the contribution of each sub-element to the judgment bias; based on the prediction bias, the weight coefficients of each sub-element in the risk fusion decision model and the decision thresholds between different risk levels are dynamically adjusted; by iterating the above process, the risk level judgment results output by the model for the subsequently input initial evidence element set are continuously aligned with the manual handling standards, thereby realizing the model's self-iteration and optimization.

[0014] As a further technical solution of the present invention, in step S42, the online learning module calculates the prediction bias of the model on each sub-element of the initial evidence element set and the contribution of each sub-element to the judgment bias, specifically including: Prediction bias: ,in: The risk level assessment result output by the model. The label represents the final processing result confirmed by human verification. The deviation penalty index; Gradient backpropagation and influence analysis were used to calculate each sub-element. Contribution to judgment bias ,in: For the first The values ​​of each sub-element, The influence factor of this element: In the formula: For the first Data confidence level of each sub-factor For the number of sub-elements, For the first The weight of each sub-element.

[0015] A port-based contactless customs clearance verification system based on multimodal data fusion, used to implement a port-based contactless customs clearance verification method based on multimodal data fusion, comprising: The evidence element collection and generation module is used to collect and process multimodal passenger data to generate an initial evidence element set that includes identity verification results, behavioral compliance assessment, and historical risk profiles. The integrated decision-making and evidence chain construction module is used to determine risks based on the initial set of evidence elements and generate a dynamic decision-making evidence chain with risk level and structure. The instruction execution and feedback module is used to execute clearance instructions corresponding to the risk level and synchronously feed back the dynamic decision evidence chain to the corresponding terminal. The model self-optimization module is used to optimize the risk assessment model online based on the results of manual handling and the dynamic decision-making evidence chain.

[0016] The beneficial effects of this invention are as follows: 1. By generating a structured, dynamic chain of decision-making evidence in parallel, the black-box decision-making process of the traditional seamless customs clearance system is transformed into a white-box operation where every step is verifiable and traceable. This fundamentally solves the obscure problems of opaque verification decisions and the inability to immediately reproduce the evidence, greatly enhancing the system's credibility and providing clear guidance for rapid on-site handling.

[0017] 2. By intelligently adapting and synchronously pushing dynamic decision-making evidence chains with handling instructions of different risk levels, a leap from binary interception to tiered and precise response has been achieved. The system can automatically convert key evidence into visual reports and accurately push them to the corresponding terminals, enabling personnel to instantly locate the core risk, thereby significantly improving the overall throughput efficiency and handling effectiveness of high-traffic channels while ensuring safety.

[0018] 3. By introducing an online incremental learning loop based on evidence chain feedback, the system can automatically fine-tune the risk fusion decision model in real time using the final results confirmed by manual review. This allows the model to continuously adapt to new risk patterns and correct misjudgments of long-tail anomaly scenarios. This enables the system to continuously self-optimize from business practices, effectively overcoming the shortcomings of traditional static models that gradually become rigid and degrade in performance after deployment. Attached Figure Description

[0019] Figure 1 This is a flowchart of a port-based contactless customs clearance verification method based on multimodal data fusion proposed in Embodiment 1 of the present invention; Figure 2 This is an architecture diagram of the risk fusion decision-making model in Embodiment 1 of the present invention; Figure 3 This is a block diagram of a port-based contactless customs clearance verification system based on multimodal data fusion, as proposed in Embodiment 2 of the present invention. Figure 4 This is a comparison diagram of the effects of the present invention and the prior art in Example 3. Detailed Implementation

[0020] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0021] Example 1 Please see the appendix Figure 1 - Appendix Figure 2 A port-based contactless customs clearance verification method based on multimodal data fusion includes: S1. Collect passenger data and process it to generate an initial set of evidence elements; specifically including: S11. Deploy a multi-source heterogeneous sensing system within the pre-designed contactless passage to collect non-contact, multi-dimensional data from passing passengers: S111. Biometric data collection: Multispectral cameras covering the visible and near-infrared bands are deployed within the passageway to simultaneously capture facial and iris images of passengers. A microphone array is installed in a ring layout to collect voice signals generated during natural movement, forming voiceprint data. All biometric data includes time-synchronized markers.

[0022] S112. Electronic document data acquisition: A parallel deployment of an ultra-high frequency RFID reader and an NFC near-field communication module enables dual-channel redundant reading of electronic passport or pass chips. The RFID module reads document chip information via far-field electromagnetic coupling, while the NFC module completes data interaction via near-field magnetic induction; both serve as backups for each other. The read content includes structured text data, digital signatures, and biometric template data stored within the document chip.

[0023] S113. Behavioral posture data acquisition: Millimeter-wave radar arrays are deployed on the side walls and top of the passageway, transmitting broadband frequency-modulated signals and receiving human body reflection echoes to extract motion micro-Doppler features and generate passenger body point cloud data. A 3D depth camera simultaneously acquires depth image sequences, and a skeleton extraction algorithm obtains the 3D coordinates of human body joints, forming a skeleton joint sequence. The millimeter-wave radar and depth camera data are precisely aligned in time, together forming a behavioral posture data stream.

[0024] S12. Multimodal fusion comparison, behavior analysis, and risk profile generation, specifically: S121. Generation of Identity Verification Result Elements: a. Multimodal biometric comparison: Modal feature extraction and similarity calculation are performed on the collected biometric data and electronic document templates. Facial similarity comparison Iris similarity Voiceprint similarity in: , , These are the similarity calculation functions for the face, iris, and voiceprint, respectively. These represent the corresponding deep feature extraction functions; These represent the facial image sequence, iris image, and speech signal acquired in step S111, respectively. , , These are facial, iris, and voiceprint biometric template data stored in the electronic certificate chip.

[0025] b. Multimodal fusion similarity: An adaptive weight fusion strategy is adopted, with weights dynamically adjusted based on the liveness detection confidence. in: For multimodal similarity fusion, weights satisfy and For the first Modal liveness detection confidence. This is the modality-specific scaling factor.

[0026] c. Liveness detection results: Comprehensive score for liveness detection Calculations were performed using multispectral consistency verification and micro-Doppler motion analysis: in: The score for protection against photo / video attacks is calculated based on multispectral reflectance variation. This is a live score based on the spectral analysis of facial micro-movements and heartbeat micro-vibrations.

[0027] d. Identity verification result elements: Generate structured element vectors ,in: For identity verification timestamps This is the transpose of the matrix.

[0028] S122. Generation of behavioral compliance assessment elements, which involves generating behavioral posture data. Input is fed into a spatiotemporal graph convolutional neural network (ST-GCN) pose analysis model trained on abnormal behavior samples. : a. Gait cycle anomaly: Extracting gait cycle sequences Calculation and normal gait template deviation : in: The standard deviation of the normal gait cycle. b. Gait cycle number extracted. Angle with the central axis of the channel Statistical anomaly bias ( )frequency : in: For indicator functions, For the preset threshold, This represents the number of frames within the observation time window.

[0029] c. Duration of stay in specific areas: Calculate the cumulative time passengers spend in sensitive areas (such as corners of passageways or blind spots of equipment). : in: The sampling interval is... A set of 3D spatial coordinates for the sensitive area. Represents the three-dimensional coordinates of the trunk joints at time t. This is an indicator function.

[0030] d. Behavioral compliance assessment elements: generating quantitative assessment vectors .

[0031] S123. Generation of historical risk profile elements, using electronic document numbers. For indexing, perform parallel queries on multi-source heterogeneous databases: a. Port Risk Database Query: Retrieve Risk Tag Set ,in Indicates whether the first exists. Class of risks, calculate risk density ,in The total number of risk label categories preset by the system; b. International Travel Records Database Search: Obtain recent travel records Passage record Calculate the traffic frequency index ,in: For the current time, The first time the passage was permitted is recorded in months.

[0032] c. Related Case Database Query: Retrieve the number of related individuals Weighted by the severity of related cases ,in: For a collection of related cases, For the case The severity level, This represents the weight of the correlation.

[0033] d. Historical Risk Profile Elements: Generate Structured Profile Vectors .

[0034] S13. Spatiotemporally correlate and standardize the identity verification result elements, behavioral compliance assessment elements, and historical risk profile elements to form a unified evidence data structure: S131. Timestamp alignment mechanism: Using the timestamp of the identity verification step as the baseline clock, the timestamps of behavioral compliance assessment elements and historical risk profile elements are checked for deviation. The system presets a time tolerance threshold. If the absolute difference between the timestamps of the other two types of elements and the baseline timestamp is within the tolerance range, it is determined to be synchronized and valid; if it exceeds the threshold, a data integrity alarm is triggered, and the time reliability attribute of the element is marked.

[0035] S132. Data format encapsulation: Three types of element vectors are vertically concatenated in a fixed order to form an initial evidence element set. This set is stored in matrix form, with each row corresponding to one type of element and each column corresponding to the numerical components of the element. An additional metadata data package is also included, containing the electronic ID number, the actual collected biometric modal set (such as a combination of facial, iris, and voiceprint identifiers), and a baseline timestamp. The encapsulated data structure possesses machine readability and serializability, supporting efficient transmission and storage.

[0036] S133. Integrity Verification: During the element integration process, the system verifies the validity flag of each type of element. If any core element is missing or its value exceeds a reasonable range, a data quality marker is added to the initial evidence element set for adaptive weight adjustment in the subsequent risk fusion decision model. The encapsulated initial evidence element set will serve as the sole input to the risk fusion decision model in step S2 for unified processing.

[0037] S2. Input the initial set of evidence elements into the risk model, and simultaneously output the risk level and dynamic decision-making evidence chain; specifically including: S21. Parallel Processing and Decision-Making: S211. Dual-engine parallel processing architecture: Initial evidence element set The risk calculation engine is a parallel processing unit that is simultaneously injected into the risk fusion decision model. Execute numerical risk quantification and level determination, evidence chain construction engine Perform interpretable data capture and traceability encapsulation; The two engines share the underlying computation graph. The message bus enables real-time sharing of intermediate data, ensuring that risk calculation and evidence tracing are strictly synchronized in terms of timing and logic.

[0038] S212. Weighted fusion calculation of the risk calculation engine: The engine performs weighted fusion calculation on the initial set of evidence elements. A hierarchical adaptive weighted fusion strategy is implemented, with calculations divided into two levels: sub-feature level and main element level. a. First layer: Sub-element level weighting, dynamically allocating weights to sub-elements within the three main categories of elements: a1. Identity Verification Result Elements Internal integration The effective sub-elements are calculated as fused similarity scores. Compared with liveness detection results .

[0039] The sub-feature weights are defined to be dynamically modulated by the liveness confidence score: ,in: The liveness sensitivity coefficient controls the rate of change of the weights with liveness confidence. and The fusion similarity scores are respectively Compared with liveness detection results The weight.

[0040] The weighted score for the identity verification elements is as follows: , in: The nonlinear amplification index for in vivo results, with typical values... This is used to enhance the risk mitigation effect of liveness detection.

[0041] a2. Elements of Behavioral Compliance Assessment Internal integration The effective sub-factor is the gait period anomaly degree. Frequency of eye movement Duration of stay in a specific area .

[0042] A weight allocation mechanism based on anomaly sensitivity is adopted. First, the anomaly activation value of each sub-element is calculated: ,in: , , These are the gait cycle anomalies. Frequency of eye movement Duration of stay in a specific area Abnormal activation values; For each sub-element, the anomaly sensitivity coefficient is... This serves as the normalized benchmark for the maximum permissible stay duration.

[0043] The weights of each sub-element are obtained through Softmax normalization: ,in: The contribution amplification factor for each sub-element.

[0044] The weighted score of the behavioral compliance element. ,in: Set saturation thresholds for gait and gaze to prevent extreme outliers from dominating the results; This is a normalized parameter for dwell time.

[0045] a3. Historical Risk Profile Elements Internal integration The effective sub-element is risk density. Traffic frequency index Weight of related cases .

[0046] A strategy combining static expert experience weights and dynamic data-driven weights is adopted: , in: Weighted fusion score for historical risk profile elements; Weighting based on expert experience to meet [the following criteria] ; This is the frequency exponential sensitivity parameter; This is the baseline value for normal passenger traffic frequency; This represents the maximum normalized value for the weight of related cases.

[0047] b. Second layer: Principal element level fusion, which involves weighting the scores of the three principal elements. Final fusion: b1. Define the principal element weight vector Its value is determined by the adaptive context modulation factor. Dynamically generated: in: Priority coefficient for contribution of main elements.

[0048] b2. Context modulation factor Calculated in real time based on the signal-to-noise ratio and data completeness of each key element: ,in: For the normalized scaling parameters of behavioral elements, This is an estimate of the variance of historical risk data.

[0049] b3. Final Comprehensive Risk Score Calculation formula: in: The linear-nonlinear mixing coefficient of the main elements is used to adjust the shape of the risk response curve of each main element. The symbol is a general one. When i traverses the set {id, bhv, risk}, it represents the weighted scores of the three main elements in turn.

[0050] c. Multi-level threshold risk assessment, which divides the comprehensive risk... A stepwise comparison is performed with a preset threshold system, where This represents the theoretical maximum value of the comprehensive risk score. Define three levels of judgment thresholds: low-risk threshold This corresponds to seamless passage; medium-risk threshold Corresponding guidance for review; high-risk threshold This corresponds to both manual intervention and on-site interception; and meets the following requirements: .

[0051] Risk level determination rules: threshold During system initialization, the system is calibrated based on statistical analysis of historical port cases and expert experience, and dynamic optimization is achieved through incremental learning in the S4 stage.

[0052] d. Evidence Chain Construction Engine Intermediate data generated by the risk calculation engine is captured in real time through a computational graph hook mechanism. The captured data includes at least the following: Weighted scores for each sub-element: ,in For the first The original values ​​of each sub-element For its dynamic weight, The total number of sub-elements ; Contribution of each sub-element: (The rest of the text appears to be a list of sub-elements and their corresponding values Sub-factors affect the overall risk score. Contribution Defined as: ; Major element fusion path: Record the weighted scores of the three major elements. , , and its dynamic weights , ; Captured data with timestamps By associating the computing node ID with the final risk level determination result, a complete determination trajectory data packet is formed, ensuring a strict temporal and logical correspondence between the data packet and the final risk level determination result. .

[0053] S22. The evidence chain construction engine receives intermediate data generated by the risk calculation engine in real time during the calculation process. This intermediate data includes at least the weighted scores of each sub-element and its contribution to the total risk score. Based on internal logic and weight relationships, it initiates a multi-stage processing flow and automatically encapsulates this data into a structured, dynamic decision-making evidence chain. Specifically: S221. The engine first parses the intermediate data packets transmitted by the risk calculation engine and extracts the standardized contribution value of each sub-element. The contribution reflects the actual influence of the sub-element in the final risk score, and is determined by the proportion of the weighted score of the element to the overall risk score, while also considering the weakening effect of the time decay factor on historical data. The engine sorts the contribution values ​​from high to low to identify the core sub-element clusters that play a decisive role in the current risk assessment.

[0054] S222. After sorting, the engine establishes logical relationships between elements. For top-ranking sub-elements, the engine traces their original source in the initial evidence element set, marks their primary element category, and constructs hierarchical relationships between parent and child elements. Simultaneously, the engine analyzes the coupling between high-contribution elements; if multiple elements have dependency or reinforcing relationships in the calculation logic, association edges are established between them. For example, a low confidence level in liveness detection leads to a reduction in the overall weight of identity verification result elements; this weight adjustment is recorded as an association rule.

[0055] S223. Based on element association, the engine performs structured encapsulation. The encapsulation format uses two extensible data structures: JSON-LD or Protocol Buffers. The JSON-LD format is suitable for scenarios requiring both human readability and machine parsing, enabling self-explanatory evidence chain data through semantic tags and contextual embedding. The Protocol Buffers format is suitable for high-throughput, low-latency terminal push scenarios, achieving efficient serialization and deserialization through binary encoding and schema definition. The encapsulated data object contains five core parts: a unique identifier for the evidence chain, a risk level determination result, a list of elements sorted by contribution, a graph of relationships between elements, and a complete traceability record of the determination logic path.

[0056] S224. The decision logic path is recorded in the form of a directed acyclic graph, where nodes represent calculation steps and edges represent data flow. The path starts with the raw data of the initial evidence element set, goes through key calculation steps such as sub-element weighting, main element fusion, and threshold comparison, and ends with the final risk level determination result. Each node records the input data summary, the weight coefficients used, and the intermediate output value for that step. The path data enables human reviewers to fully reproduce the model's decision-making process, verify whether the contributions of key elements conform to safety logic, and provide interpretable deviation analysis basis for model optimization in stage S4.

[0057] S225. After encapsulation, the evidence chain construction engine submits the structured data packet to the instruction execution module. The data flow is determined according to the instruction type in step S3: if the judgment result is seamless release, the evidence chain data packet is compressed and archived to the background log database, and the storage period is set according to the port audit requirements; if the judgment result is to guide review or manual intervention, the evidence chain data packet is pushed to the corresponding terminal interface through a secure communication channel. It is presented in the form of a visual report on the reviewer's terminal and in the form of an element summary card on the mobile terminal of the on-site law enforcement personnel to support manual handling decisions.

[0058] S3. Execute the instruction corresponding to the risk level and synchronize the dynamic decision-making evidence chain to the corresponding terminal; specifically including: S31. The instruction generation logic unit receives the final risk level determination result output by the risk calculation engine, generates the corresponding clearance control instruction according to the preset instruction mapping rules, and sends it to the channel control unit for execution, specifically: S311. Seamless release command in low-risk assessment scenarios: When the risk level assessment result falls into the low-risk range, the instruction generation logic unit activates the contactless access instruction generation module. This instruction is the highest priority execution instruction and includes the gate opening authorization code and the passage validity token.

[0059] After receiving the instruction, the channel control unit immediately sends an opening signal to the gate controller. The gate wings complete the opening action within the preset response time, and at the same time trigger the green passage indicator light in the channel and the subtle guidance sound effect.

[0060] The system automatically associates the passenger's travel record in the background, marking the verification process as closed-loop completed, without any additional manual intervention or review triggering.

[0061] S312. Guided review instructions in medium-risk assessment scenarios: When the risk level assessment result falls into the medium-risk range, the instruction generation logic unit activates the guidance review instruction generation module. This instruction is a medium-priority execution instruction, which includes directional guidance parameters and gate locking control codes.

[0062] Upon receiving the instruction, the channel control unit first triggers the directional audio-visual guidance device, sending a directional voice prompt to the passenger through a narrow-beam speaker. Simultaneously, it activates the channel floor projection system, projecting a bright directional arrow or text sign onto the ground in front of the passenger, clearly indicating their path to the designated verification area. At the same time, the channel control unit sends a locking command to the turnstiles ahead of the passenger's path, keeping the turnstile doors closed and displaying a red warning sign to prevent the passenger from continuing to pass.

[0063] The system simultaneously pushes reminders of tasks to be reviewed to the back-end review terminal, including basic passenger information and a summary of risk factors.

[0064] S313. Manual intervention instructions in high-risk assessment scenarios: When the risk level assessment result falls into the high-risk range, the instruction generation logic unit activates the manual intervention instruction generation module. This instruction is the highest security level execution instruction, containing a full-channel lockout control code and a field alarm activation signal.

[0065] Upon receiving the instruction, the channel control unit immediately sends an emergency lock command to all turnstiles within the channel. All turnstile doors are forcibly locked in their current positions and flash red warning lights. Simultaneously, on-site alarm devices are activated, including high-decibel sirens, high-brightness strobe lights, and an emergency interception notification sent to handheld terminals of on-site law enforcement personnel.

[0066] The system locks the passage of all other passengers within the channel to prevent the spread of risk, and automatically calls the location information of on-site law enforcement personnel from nearest to push detailed handling instructions and risk factor details.

[0067] S32. While the channel control unit executes the clearance control command, the evidence chain distribution engine performs differentiated processing and precise push of the dynamic decision-making evidence chain according to the command type: S321. Evidence chain archiving in seamless release scenarios: For seamless release instructions, the evidence chain distribution engine initiates a lightweight archiving process; First, the dynamic decision-making evidence chain is bound to the passenger's travel record identifier to form a complete verification-decision-release data package.

[0068] Then, the data compression module is invoked, and an efficient compression algorithm is used to perform lossless compression of the evidence chain, reducing storage space usage.

[0069] The compressed evidence chain is archived in the long-term storage partition of the backend log database, with the storage period set according to port audit requirements and data compliance policies.

[0070] The archived records include data integrity check codes and timestamps to ensure data reliability for subsequent audits and traceability. This archiving process is executed asynchronously and does not affect the efficiency of the channel.

[0071] S322. Visualized Push of Evidence Chain in Guided Review Scenarios: For guided review instructions, the evidence chain distribution engine initiates a visual conversion process; First, the visualization template engine is invoked to load a preset HTML5 report template for review. The template engine parses the structured data in the dynamic decision-making evidence chain, highlights the high-risk sub-elements with the highest contribution ranking, uses different color codes to distinguish the three main elements of identity, behavior, and history, and displays the judgment logic path in the form of a timeline.

[0072] The generated visualization report includes risk factor values, contribution percentages, a comparison chart of judgment thresholds, and key video screenshots. The report is pushed to the designated backend reviewer's terminal via a highly reliable message queue system, with a task priority indicator and response time limit attached to the push. Upon receiving the report, the reviewer's terminal automatically displays a review interface, showing the visualization report and allowing the reviewer to retrieve the original data for in-depth verification.

[0073] S323. Real-time push of evidence chain in scenarios with manual intervention: For manual intervention commands, the evidence chain distribution engine initiates a real-time push process; First, video stream segments from the channel monitoring system are extracted within a preset time window before and after the high-risk determination. The video data is then timestamped and packaged with the dynamic decision-making evidence chain. The packaged data packet is then directly pushed to the mobile terminals of law enforcement personnel on site via a low-latency communication protocol, ensuring real-time reception before or during law enforcement operations.

[0074] After receiving the data packet, the mobile terminal automatically activates the evidence chain display interface, showing a visual report as a semi-transparent overlay layer on the real-time video screen. Arrows mark the spatiotemporal locations of key risky behaviors, and numerical labels display the quantitative assessment values ​​of each risk element. Law enforcement officers can expand or collapse detailed information via touch operations and can directly record the on-site handling results on the terminal. The handling results will be sent back to the system as authoritative labels for model optimization.

[0075] S4. Retrieve the dynamic decision-making evidence chain labeled with the final disposal result to optimize the risk model; specifically including: S41. Evidence Chain Recovery and Tag Binding: S411. For cases involving guided review instructions or manual intervention instructions, the system establishes a two-way data transmission channel between the reviewer's terminal in the background and the mobile terminal of the on-site law enforcement officer. Once the reviewer completes in-depth verification and confirms the handling conclusion, or the on-site law enforcement officer completes the on-site inspection and records the final handling result, the terminal application transmits the manually confirmed final handling result back to the central data server via an encrypted communication link. The final handling result includes the handling conclusion code, the text of the handling basis, the handling timestamp, and the identity authentication information of the executing personnel, constituting a legally valid authoritative judgment label.

[0076] S412. Upon receiving the authoritative tag, the system immediately activates the data binding engine. The binding engine retrieves the original record of the dynamic decision evidence chain generated in stage S2 based on the unique identifier of the case, and ensures the integrity of the evidence chain has not been tampered with through hash verification.

[0077] Subsequently, authoritative labels are bound one-to-one with the dynamic decision-making evidence chain, forming a labeled training sample pair data structure. This structure contains four core parts: the complete numerical vector of the initial evidence element set, the original risk level judgment result output by the risk fusion decision model, intermediate data snapshots of the model's internal calculation path, and the final disposal result label confirmed by humans. The binding process uses digital signature technology to ensure the authenticity and non-repudiation of the training sample pairs.

[0078] S413. The completed training sample pairs are stored in the incremental learning sample pool and prioritized according to the severity of the bias based on the risk level. Samples with larger biases will receive higher sampling weights and be used more frequently in subsequent online learning processes to quickly improve the model's ability to distinguish difficult cases.

[0079] S42. Incremental Learning and Optimization: S421. Let the accumulated set of labeled training sample pairs be... for: ,in: To accumulate the number of samples, For the first The original data of the dynamic decision-making evidence chain for each sample. This is the initial set vector of evidence elements for this sample. The risk level determination result (discrete value) output by the model (corresponding to four levels) The label for the final disposal result confirmed by humans (as a truth value); Accumulated training sample pairs Input into the online learning module of the risk fusion decision model.

[0080] S422. The online learning module calculates the prediction bias of the model on each sub-element of the initial evidence element set by comparing the risk level judgment results initially output by the model with the final handling results confirmed by humans, based on the differences between the training samples and the results. The prediction bias is as follows: in This is the bias penalty index. The penalty increases non-linearly as the difference between the model's judgment and the manual handling becomes larger. A typical value is... .

[0081] Gradient backpropagation and influence analysis were used to calculate each sub-element. Contribution to judgment bias ,in For the first The values ​​of each sub-element, The influence factor of this element: In the formula: For the first The data confidence scores of each sub-feature (calculated from sensor signal-to-noise ratio, data integrity, etc.) are used to reduce the attribution weight of low-quality data; The number of sub-elements; For the first The weight of each sub-element.

[0082] S423. The weights of sub-elements in the risk fusion decision model are dynamically adjusted using the truncated stochastic gradient descent (T-SGD) algorithm: in: The updated weights; The online learning rate controls the step size of a single update; typical values ​​are... ; The gradient of the bias with respect to the weights is obtained through automatic differentiation calculation; This is a priori regularization coefficient used by experts to prevent the model from deviating too far from expert experience; typical values ​​are provided. ; The weighted prior values ​​preset by the experts.

[0083] S424. Adaptive adjustment of decision threshold: a. Threshold deviation calculation, for each threshold Calculate its adjustment offset, where This serves as the dividing line between low and medium risk. This serves as the dividing line between medium and high risk. This serves as the dividing threshold between high-risk and interception levels. If the model is determined to be Level- The manual assessment is Level- , ( < (i.e., missed judgment) If the model is determined to be Level- The manual assessment is Level- , ( > (i.e., missed judgment) in: and These represent the upward and downward adjustment amounts for the threshold, respectively. For the set of samples that were missed, This is the set of misjudged samples; These represent the number of samples that were missed or incorrectly judged, respectively. The cost coefficient for false positives and false negatives is usually... ; The comprehensive risk score for the nth sample is calculated in step S212. This sets the threshold boundary corresponding to the current sample model's classification level. For example, if the model classifies it as Level-1 (medium risk), then... = If it is determined to be Level-2 (high risk), then = .

[0084] b. Threshold updates employ a conservative update strategy, smoothly adjusting the threshold: in: and These are the thresholds before and after the adjustment, respectively; The threshold learning rate is a relatively small value (usually). To ensure system stability.

[0085] S425. Model Iteration Convergence and Performance Monitoring a. Implementation of the Forgetting Mechanism: To prevent model drift caused by over-reliance on outdated historical samples during online learning, the system introduces an exponentially weighted moving average mechanism to control the influence weight of outdated samples. This mechanism assigns each training sample a weight that decays exponentially over time: the most recently collected samples receive the highest weight, while the influence of earlier samples gradually decreases over time. The decay rate is regulated by a forgetting factor, which ranges from 0 to 1. When the forgetting factor is close to 1, the model maintains a longer memory period for historical samples, resulting in a smoother and more stable parameter update process; when the forgetting factor is small, the model focuses more on recent sample data and can quickly adapt to changes in new risk patterns. The system dynamically adjusts the forgetting factor value based on the stability of the port risk environment: a larger value is used during periods of stable risk patterns to maintain model stability, while a smaller value is used during periods of rapid changes in risk patterns to enhance the model's environmental adaptability.

[0086] b. Performance Monitoring and Rollback Protection: To ensure the stability and security of the online learning process, the system establishes a periodic performance monitoring mechanism. After processing a fixed number of training samples, the model's performance is evaluated on an independent validation dataset. The validation dataset consists of recently collected samples that did not participate in online training, used to objectively reflect the model's actual discriminative ability. The performance evaluation metrics not only include the model's classification accuracy on the validation set, but also introduce a stability regularization term to constrain the deviation of model parameters from the initial state, preventing over-adjustment during online learning that could cause the model to lose its basic discriminative ability.

[0087] If monitoring detects a significant drop in model performance compared to the previous evaluation, exceeding a preset tolerance level, the system immediately triggers a rollback protection mechanism. This mechanism restores all model parameters to their stable state at the time of the last evaluation and automatically halves the online learning rate to reduce subsequent update steps and prevent drastic fluctuations. This mechanism ensures that model performance does not continue to deteriorate and quickly recovers to a reliable state in the event of learning anomalies, guaranteeing the continuity and security of port clearance and verification operations.

[0088] Example 2 Please see Figure 3 A port-based contactless customs clearance verification system based on multimodal data fusion, used to implement a port-based contactless customs clearance verification method based on multimodal data fusion, including: The evidence element collection and generation module is used to collect biometric data, electronic document data, and behavioral posture data of passengers passing through a preset contactless channel in real time; compare the biometric data and electronic document data to generate identity verification result elements; analyze the behavioral posture data to generate behavioral compliance assessment elements; and simultaneously query the passenger's risk profile to generate historical risk profile elements. The identity verification result elements, behavioral compliance assessment elements, and historical risk profile elements together constitute the initial evidence element set, specifically including: The data acquisition unit is equipped with a multispectral camera, microphone array, UHF RFID reader, NFC near-field communication module, millimeter-wave radar array and three-dimensional depth camera deployed in the non-sensory channel, for collecting the biometric data, electronic certificate data and behavioral posture data respectively. The element generation unit is used to process the data output by the data acquisition unit: perform multimodal biometric comparison to generate the identity verification result element, run the posture analysis model to generate the behavior compliance assessment element, and concurrently query multiple risk databases to generate historical risk profile elements. The element integration unit is used to encapsulate the elements output by the element generation unit in a preset format and align them with timestamps to generate an initial set of evidence elements.

[0089] The fusion decision-making and evidence chain construction module, connected to the evidence element collection and generation module, is used to input the initial set of evidence elements into the risk fusion decision-making model to generate the final risk level determination result. Simultaneously, based on the internal logic and weight relationships of the risk fusion decision-making model, the initial set of evidence elements is associated and sorted according to their contribution to the final risk level determination result, automatically encapsulating it into a structured dynamic decision-making evidence chain. This dynamic decision-making evidence chain records the reasoning path and key evidence from the initial evidence elements to the final risk level determination result in a machine-readable data structure; specifically, it includes: The risk calculation engine receives an initial set of evidence elements, performs weighted fusion calculations based on pre-configured element weights, and outputs the final risk level determination result by combining multiple thresholds. The evidence chain building engine works in parallel with the risk calculation engine and shares data. It is used to acquire intermediate data in real time during the calculation process, sort and associate each sub-element according to its contribution to the risk score, and encapsulate the association results and the risk level determination results into a structured data package to form a dynamic decision evidence chain.

[0090] The instruction execution and feedback module, connected to the integrated decision-making and evidence chain construction module, is used to execute corresponding customs clearance control instructions based on the risk level assessment results. Simultaneously, it pushes the dynamic decision-making evidence chain to the terminal matching the customs clearance control instruction: if it's a seamless release instruction, the evidence chain is archived in the background log; if it's a guided review or manual intervention instruction, the evidence chain is converted into a visual report and pushed to the corresponding reviewer's terminal or the mobile terminal of on-site law enforcement personnel as the basis for handling. Specifically, this includes: The command and control unit is used to generate instructions for seamless release, guidance review, or manual intervention based on the risk level determination results, and send them to the access gate, sound and light guidance device, and alarm device for execution; The evidence chain adaptation and distribution unit is used to differentiate the evidence chain for dynamic decision-making based on the type of instruction generated by the instruction control unit: compressing and archiving the non-intrusive release instruction, calling the template engine to generate an HTML5 visual report and pushing it to the backend terminal for the guided review instruction, and packaging the evidence chain and video stream for the manual intervention instruction and pushing it to the mobile terminal through a low-latency protocol.

[0091] The model self-optimization module, connected to the instruction execution and feedback module and the fusion decision and evidence chain construction module, is used to collect the final human-confirmed handling results for cases that have executed guided review or manual intervention instructions. This final handling result is used as an authoritative label and bound to the dynamic decision evidence chain from the fusion decision and evidence chain construction module, forming a labeled training sample pair. The accumulated training sample pairs are periodically or in real-time input into the online learning unit of the risk fusion decision model. By comparing the model's initial risk level judgment with the final human handling result, the prediction bias of the model on each evidence element is calculated. Based on this bias, an incremental learning algorithm is used to dynamically adjust the weight coefficients and decision thresholds of each evidence element in the risk fusion decision model, achieving self-iteration and optimization of the model. Specifically, this includes: The sample collection unit is connected to the back-end terminal and mobile terminal of the instruction execution and feedback module. It is used to receive and record the final disposal results confirmed by the human, and bind them with the corresponding dynamic decision evidence chain as training sample pairs. The online learning unit, built into the risk fusion decision model of the fusion decision and evidence chain construction module, is used to receive training sample pairs, calculate prediction bias and apply incremental learning algorithms to dynamically adjust the weight coefficients and decision thresholds of the risk fusion decision model.

[0092] Example 3 To verify the technical effectiveness of the present invention, a highly realistic digital twin test platform for seamless customs clearance at ports was constructed, and the method of the present invention was compared and verified with typical existing technical methods.

[0093] 1. Simulation Environment and Data Settings The hardware environment simulates the deployment of a standard seamless channel, integrating sensors such as multispectral cameras, millimeter-wave radar, and RFID / NFC card readers.

[0094] The software environment deploys the core software stacks of the present invention and existing solutions on the same hardware and basic algorithm library. The existing solution adopts the classic "feature comparison - threshold judgment - binary processing" pipeline.

[0095] The test dataset is constructed using anonymized real historical customs clearance data, containing 10,000 customs clearance samples. The samples are pre-labeled with real risk tags (9,500 low-risk, 400 medium-risk, and 100 high-risk), and cover typical scenarios such as document mismatch, abnormal behavior, and high-risk individuals.

[0096] Key parameters: Simulation duration: 30 consecutive days of operation; throughput: set to a peak of 600 people per hour.

[0097] 2. Validation Indicators and Methods For comprehensive evaluation, the following core performance metrics were defined: Overall customs clearance efficiency refers to the average time it takes for a passenger to complete the process (release or handover) from entering the channel.

[0098] Risk identification accuracy refers to the degree of consistency between the risk level determined by the system and the actual label of the sample.

[0099] The false alarm rate of the system, which is the proportion of low-risk passengers who are mistakenly identified as medium- or high-risk, directly affects passenger experience and resource utilization.

[0100] The efficiency of manual processing refers to the average time it takes for staff to complete confirmation or processing for cases requiring review or intervention.

[0101] The model adaptive gain improves the system's accuracy in identifying "long-tail" rare risks (specific abnormal patterns that account for less than 1% of the test set) after running for a period of time.

[0102] 3. Verification process and results analysis Both schemes were run under the same test set and load. The existing scheme only outputs a "allow" or "block" command; the scheme of this invention fully executes steps S1 to S4, generates a dynamic decision evidence chain, and enables online learning.

[0103] After multiple rounds of simulation tests, the proposed solution demonstrates significant advantages in all indicators. Specific data comparisons are shown in Table 1 below. Figure 4 As shown: Table 1: Comparison of Results 4. Summary of Implementation Examples Simulation results show that, compared with existing mainstream technologies, the present invention achieves a leap from "high false alarms, low interpretability, and rigid response" to "high accuracy, interpretability, and flexible optimization".

[0104] In terms of core safety features, the risk identification accuracy rate is close to 98%, and the false alarm rate is less than 1%, which greatly reduces interference to normal passengers while improving safety margins.

[0105] In terms of operational efficiency, through tiered processing and evidence-driven approaches, overall customs clearance efficiency has increased by over 30%, and manual processing time has been reduced by nearly 60%, effectively resolving the contradiction between large passenger flows and refined management.

[0106] In terms of system evolution, the unique self-optimization capability enables the system to continuously learn and adapt to new threats, which is a long-term advantage that static systems cannot match.

[0107] This embodiment fully verifies the completeness of the method and the significant progress in the technical effect of the present invention, which meets the requirements of the patent law on inventiveness and utility, and provides a clear and efficient technical path for the next generation of contactless customs clearance system for smart ports.

[0108] As can be seen from the above description, the embodiments of the present invention achieve the following technical effects: This invention achieves white-box and credible verification decisions, solving the fundamental problem of opaque decision-making processes. It generates a dynamic chain of decision-making evidence in parallel, automatically providing a structured, machine-readable criterion description for each risk assessment (especially guided review and human intervention). This transforms decision-making from unknowable code logic into auditable and traceable data records for system administrators, greatly enhancing the system's credibility and auditability.

[0109] For on-site staff, they no longer need to guess the reasons for interception based on experience, thus upgrading the handling mode from time-consuming and labor-intensive manual blind inspection to efficient evidence-guided handling. The 59% increase in manual handling efficiency in simulation verification is proof of this. For passengers, the system can provide brief compliance prompts through the terminal (such as "Please look directly at the camera"), alleviating the confusion and resistance caused by being inexplicably blocked, and improving the customs clearance experience and the image of public services.

[0110] This invention constructs an evidence-driven, differentiated, and precise handling process, achieving an optimal balance between safety and efficiency. By intelligently adapting and synchronously pushing dynamic decision-making evidence chains with instructions at different risk levels, this invention innovates a human-machine collaborative handling model. Its effectiveness lies in achieving precise matching of handling resources and strategies: for low-risk passengers (the vast majority), the system performs seamless release and silently archives the evidence chain, ensuring an extremely smooth passage experience (a core contribution to a 34% improvement in overall clearance efficiency in simulations). For medium-risk passengers, the system guides them to the review channel while simultaneously converting key evidence into a visual report and pushing it to the reviewer's screen. This allows reviewers to complete targeted confirmation within seconds, avoiding a complete initial investigation and minimizing the impact on efficiency while ensuring safety. For high-risk passengers, the system locks the channel while simultaneously pushing an alarm package containing a complete evidence chain and real-time footage to the law enforcement officer's mobile terminal, ensuring they have full risk information before approaching the target, achieving "preemptive awareness and precise control," greatly improving the success rate and safety of the handling process.

[0111] A closed-loop self-evolutionary mechanism of perception-decision-learning has been established, enabling the system to continuously optimize and grow. This invention introduces an online incremental learning module based on evidence chains and human feedback. This mechanism breaks the fate of traditional static models gradually becoming rigid after deployment, bringing the following profound effects: Continuously improving accuracy, the system can automatically use the standard answers (final handling results) of each manual review to correct its own biases, achieving increasing accuracy with use. In simulations, the model's recognition rate for long-tail rare risks increased from 70% to 92% within 30 days, which is a direct manifestation of this effect. Adapting to business changes, port risk patterns are not static. This self-optimization mechanism enables the system to dynamically track and learn new risk patterns, automatically adjusting model weights and thresholds without frequent, high-cost manual model retraining and version upgrades. Reducing long-term operation and maintenance costs, the system's self-iterative capability reduces the reliance on continuous manual tuning by the algorithm team, shifting the operation and maintenance mode from manual maintenance to autonomous maintenance, significantly reducing the system's total lifecycle management costs.

[0112] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

[0113] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this specification. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A port-based contactless customs clearance verification method based on multimodal data fusion, characterized in that, include: Collect passenger data and process it to generate an initial set of evidence elements; Input the initial set of evidence elements into the risk model, and simultaneously output the risk level and dynamic decision-making evidence chain; Execute instructions corresponding to the risk level and synchronize the dynamic decision evidence chain to the corresponding terminal; The dynamic decision-making evidence chain, labeled with the final disposal result, is recovered to optimize the risk model.

2. The port-based contactless customs clearance verification method based on multimodal data fusion according to claim 1, characterized in that, The process of collecting passenger data and processing it to generate an initial set of evidence elements specifically includes: Real-time collection of biometric data, electronic document data, and behavioral posture data of passengers passing through the preset contactless channel; The biometric data and electronic certificate data are compared to generate identity verification result elements; the behavioral posture data is analyzed to generate behavioral compliance assessment elements; and the passenger's risk profile is queried simultaneously to generate historical risk profile elements. The identity verification result elements, behavioral compliance assessment elements, and historical risk profile elements together constitute the initial evidence element set.

3. The port-based contactless customs clearance verification method based on multimodal data fusion according to claim 2, characterized in that, The biometric data and electronic certificate data are compared to generate identity verification result elements; the behavioral posture data is analyzed to generate behavioral compliance assessment elements. The risk profile of the passenger is queried simultaneously to generate historical risk profile elements, including: The facial image, iris image, and voiceprint data are compared with the corresponding biometric templates in the electronic certificate data through multimodal fusion, and the output includes identity verification result elements including similarity score and liveness detection result; Behavioral posture data is input into a posture analysis model trained with abnormal behavior samples to extract features such as gait cycle abnormality, gaze deflection frequency, and dwell time in specific areas, and output quantitative behavioral compliance assessment elements. Using the document number in the electronic document data as an index, the system queries the port risk database, international travel record database, and related case database in parallel, returning risk labels, historical passage frequency, and related attention levels to generate structured historical risk profile elements.

4. The port-based contactless customs clearance verification method based on multimodal data fusion according to claim 1, characterized in that, The initial set of evidence elements is input into the risk model, which simultaneously outputs the risk level and the dynamic decision-making evidence chain, specifically including: The initial set of evidence elements is simultaneously input into the risk calculation engine and the evidence chain construction engine of the risk fusion decision model, and the risk calculation engine generates the final risk level determination result. The evidence chain building engine receives intermediate data generated by the risk calculation engine in real time during the calculation process. Based on internal logic and weight relationships, it initiates a multi-stage processing flow and automatically encapsulates it into a structured dynamic decision evidence chain.

5. The port-based contactless customs clearance verification method based on multimodal data fusion according to claim 4, characterized in that, The risk calculation engine generates the final risk level determination result, specifically including: The risk calculation engine performs weighted fusion calculations on the similarity scores in the identity verification result elements, the liveness detection results, the quantitative values ​​of each feature in the behavior compliance assessment elements, and the risk labels in the historical risk profile elements, based on pre-configured element weights. Combined with preset multi-level thresholds, it outputs the final risk level determination result.

6. The port-based contactless customs clearance verification method based on multimodal data fusion according to claim 5, characterized in that, The multi-stage processing flow specifically includes: The evidence chain building engine first parses the intermediate data packets transmitted by the risk calculation engine and extracts the standardized contribution value of each sub-element; Based on the contribution level, the sub-elements in the initial evidence element set are sorted and associated. After sorting, the engine establishes the logical relationship between the elements. Based on the element association, the engine encapsulates the logical association relationship and the final risk level determination result into a dynamic decision evidence chain in JSON-LD or Protocol Buffers format; The dynamic decision evidence chain explicitly records key elements, the contribution weight of each element, and the decision logic path. The decision logic path is recorded in the form of a directed acyclic graph, where nodes represent calculation steps and edges represent data flow. After encapsulation, the evidence chain building engine pushes the dynamic decision evidence chain to the terminal that matches the customs clearance control command. When the passage is released without human intervention, it is archived to the background log. When the passage is guided for review or manual intervention, it is pushed to the reviewer's terminal or the mobile terminal of the on-site law enforcement personnel.

7. The port-based contactless customs clearance verification method based on multimodal data fusion according to claim 4, characterized in that, Execute instructions corresponding to the risk level and synchronize the dynamic decision-making evidence chain to the corresponding terminal, specifically including: Based on the final risk level determination result, a corresponding clearance control instruction is generated and sent to the channel control unit for execution; wherein, the clearance control instruction includes: seamless release instruction, guided review instruction, and manual intervention instruction; While executing the customs clearance control command, the dynamic decision-making evidence chain is adapted and pushed synchronously according to the command type: For the seamless passage instruction, the dynamic decision evidence chain is bound to the passage record, compressed, and archived to the background log database; For the guided review instruction, the visualization template engine is invoked to convert the dynamic decision evidence chain into an HTML5 format visualization report that highlights key risk elements, and then pushes it to the designated back-end reviewer's terminal via a message queue; For the manual intervention command, the dynamic decision-making evidence chain and real-time video stream segments are packaged and pushed to the mobile terminal of the on-site law enforcement personnel through a low-latency communication protocol, and a visual report is displayed on the terminal interface in the form of an overlay layer.

8. The port-based contactless customs clearance verification method based on multimodal data fusion according to claim 7, characterized in that, The dynamic decision-making evidence chain, labeled with the final disposal result, is used to optimize the risk model and specifically includes: For cases where the guided review instruction or manual intervention instruction is executed, the system receives and records the final handling result confirmed after manual verification at the back-end reviewer's terminal or the mobile terminal of the on-site law enforcement personnel; the final handling result is used as an authoritative label and uniquely bound to the dynamic decision-making evidence chain corresponding to the case to form a labeled training sample pair. The accumulated training samples are input into the online learning module of the risk fusion decision model; the online learning module calculates the prediction bias of the model on each sub-element of the initial evidence element set and the contribution of each sub-element to the judgment bias; based on the prediction bias, the weight coefficients of each sub-element in the risk fusion decision model and the decision thresholds between different risk levels are dynamically adjusted; through iterating the above process, the risk level judgment results output by the model for the subsequently input initial evidence element set are continuously aligned with the manual handling standards.

9. A port-based contactless customs clearance verification method based on multimodal data fusion according to claim 8, characterized in that, The online learning module calculates the prediction bias of the model on each sub-element of the initial evidence element set and the contribution of each sub-element to the judgment bias, specifically including: Prediction bias: ,in: The risk level assessment result output by the model. The label represents the final processing result confirmed by human verification. The deviation penalty index; Gradient backpropagation and influence analysis were used to calculate each sub-element. Contribution to judgment bias ,in: For the first The values ​​of each sub-element, The influence factor of this element: In the formula: For the first Data confidence level of each sub-factor For the number of sub-elements, For the first The weight of each sub-element.

10. A port-based contactless customs clearance verification system based on multimodal data fusion, characterized in that, To implement the port-based contactless customs clearance verification method based on multimodal data fusion as described in any one of claims 1-9, the method includes: The evidence element collection and generation module is used to collect and process multimodal passenger data to generate an initial evidence element set that includes identity verification results, behavioral compliance assessment, and historical risk profiles. The integrated decision-making and evidence chain construction module is used to determine risks based on the initial set of evidence elements and generate a dynamic decision-making evidence chain with risk level and structure. The instruction execution and feedback module is used to execute clearance instructions corresponding to the risk level and synchronously feed back the dynamic decision evidence chain to the corresponding terminal. The model self-optimization module is used to optimize the risk assessment model online based on the results of manual handling and the dynamic decision-making evidence chain.

Images