A customs inspection and release trigger dispatching closed-loop method based on counterfactual causal learning

Abstract

The application discloses a customs inspection and release trigger dispatching closed-loop method based on counterfactual causal learning, comprising the following steps: establishing a supervision object main index to form a global data platform for the current inspection and release trigger problem; constructing a layer-by-layer identification mechanism to form a trigger causal portrait for the current business object; performing controlled replacement on the decision variable, performing parallel deduction on the potential results of each counterfactual twin scene in different dimensions, and forming an alternative strategy effect spectrum; constructing a resonance balance relationship among risk and return, time efficiency and resource cost; performing causal backtracking and effect calibration on previous decisions; the scheme has the characteristics of improving the accuracy, stability and interpretability of the inspection and release trigger decision, improving the scientificity and rationality of the decision, improving the rationality of the inspection resource allocation, reducing the phenomena of on-site congestion, repeated circulation and invalid dispatching, and improving the adaptive optimization capability and the precision and intelligent level of customs supervision.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and smart ports, and in particular to a closed-loop method for customs clearance triggering dispatch based on counterfactual causal learning. Background Technology

[0002] With the continuous advancement of smart customs, smart ports, and the digitalization of international trade, the customs supervision model is gradually evolving from traditional manual document review, experience-based control, and post-event processing to automated triggering, intelligent document dispatching, and closed-loop supervision throughout the entire process.

[0003] The mainstream methods for triggering and dispatching customs clearance currently fall into the following categories: The first category is triggering methods based on fixed rules or expert experience. These methods typically set trigger conditions based on information such as commodity category, enterprise credit, declaration elements, control parameters, and historical anomaly records. When a preset threshold is met or a rule is hit, the process transitions to clearance processing or inspection dispatch. The second category is predictive methods based on statistical learning or risk scoring. These methods utilize historical customs declaration, control, inspection, and release data to construct classification or scoring models, ranking the current business objects by risk, and then determining whether to trigger clearance or proceed to manual review, machine inspection, or physical inspection based on the risk level. The third category is dispatching methods based on on-site resource status. After triggering, the system combines factors such as inspection bay occupancy, customs officer workload, equipment availability, and time constraints, using rule matching, priority ranking, or simple optimization algorithms to dispatch tasks to the corresponding on-site units. The fourth category is post-event rule correction methods based on business results. That is, after the inspection results, review results, or abnormal handling results are generated, the rules are manually adjusted, the parameters are updated, or the model is retrained to improve the triggering effect in the next stage.

[0004] While the aforementioned methods have improved the automation level of clearance to some extent, they still have significant shortcomings. First, most existing methods are based on correlation judgments rather than causal judgments. Whether it's fixed rules, scorecards, or general supervised learning models, they essentially predict risk based on "which features and abnormal results occur simultaneously" in historical samples, but struggle to identify "which key factors actually lead to the need to trigger clearance, transfer to inspection, or implement special dispatch." Second, existing solutions generally lack counterfactual analysis capabilities. For the same business object, the system can usually only provide a single judgment of "whether it is currently triggered," but cannot further answer questions such as: would canceling this trigger still produce abnormal consequences? Due to the lack of ability to evaluate "alternative paths that did not occur but may have occurred," existing technologies cannot provide deeper evidence for optimizing clearance triggering and dispatch strategies. Third, in existing systems, "trigger decision" and "dispatch execution" are often separate two-stage processes. The former only focuses on whether a trigger occurs, while the latter deals with how to dispatch, lacking a unified modeling and collaborative optimization mechanism between the two. Fourth, most existing feedback mechanisms are coarse-grained, slow-to-time, and weakly attributable post-hoc corrections. While the system can record inspection results, release results, or anomaly handling results, these results are usually simply used as new training samples or as a basis for manual revision, without further analysis. Due to the lack of refined attribution and causal backpropagation, the model and rules struggle to form a truly self-optimizing closed loop. Fifth, traditional methods based on static rules or purely statistical learning are insufficient to stably characterize the real-world mechanisms by which multiple factors affect triggering results and dispatch effectiveness in dynamic environments, thus making it difficult to balance regulatory accuracy and customs clearance efficiency.

[0005] Therefore, there is an urgent need for a new technical solution that can identify key causal factors from multi-source business data for customs clearance triggering and dispatch scenarios, conduct joint analysis on "whether it is triggered, when it is triggered, where it is dispatched, and what disposal path is adopted", and build a closed-loop optimization mechanism by combining inspection results, release results and subsequent feedback, thereby improving the interpretability of triggering decisions, the rationality of dispatch execution and the effectiveness of system strategy iteration. Summary of the Invention

[0006] To achieve the above objectives, the inventors provide a closed-loop method for customs clearance-triggered dispatch based on counterfactual causal learning, comprising the following steps:

[0007] S1. Establish a master index for the regulated objects, perform cross-source alignment, temporal correlation and status splicing to form a global data base for the current release triggering issues;

[0008] S2, construct a layered identification mechanism for triggering causal candidate set, causal contribution screening, and triggering main cause focus, and decouple the disturbance factors in layers to form a triggering causal profile for the current business object;

[0009] S3, based on the current real business state, while maintaining consistency in subject attributes, environmental constraints and key contexts, performs controlled replacement of decision variables to generate several counterfactual twin scenarios, and performs parallel extrapolation of the potential results of each counterfactual twin scenario in different dimensions to form an alternative strategy effect spectrum;

[0010] S4, based on the causal focus identification results and the counterfactual substitution strategy deduction results, incorporates the trigger necessity, dispatch object, dispatch timing, priority order and disposal path into the same decision space to construct a resonance balance relationship between risk and benefit, timeliness benefit and resource cost;

[0011] S5 receives the results of the inspection hit, release, transfer to manual processing, time consumption, abnormal handling, and resource usage. Based on the correspondence between the execution results, the triggering cause, and the dispatch path, it performs causal backtracking and effect calibration on the previous decisions.

[0012] In a preferred embodiment of the present invention, step S1 includes the following steps:

[0013] S101, for each record, extract its identifier field, time field, business context field, and spatial attribution field to form a record-level candidate description. Then, using the candidate business object as the merging unit, comprehensively consider identifier consistency, time proximity, business context consistency, and spatial attribution consistency to perform merging judgment on cross-source records. Generate a unique regulatory object master index for each single shipment, single vehicle trip, single container trip, or single batch of business, obtaining the master index set and its corresponding cross-source record attribution relationship, expressed as:

[0014] ;

[0015] ;

[0016] in, Indicates the first The primary index generated by each regulated object Indicates the primary index candidate space. Indicates the total number of entities under supervision. This represents the number of original records associated with the u-th supervised object. This represents the identifier consistency score of the nth record in the uth supervised object. This represents the temporal proximity score of the nth record within the uth monitored object. This represents the business context consistency score of the nth record in the uth supervised object. This represents the spatial attribution consistency score of the nth record within the uth supervised object. , , and Let represent the weight coefficients for identifier consistency, temporal proximity, business context consistency, and spatial affiliation consistency, respectively, satisfying: , ;

[0017] S102, around the master index of each regulated object, performs field extraction and attribute renaming on data from different sources, maps synonymous fields to a unified attribute space, and unifies discrete states, classification codes, timestamps, numerical units, and missing values. Based on the master index of the regulated object and the event timeline, aligns the attributes from each source to a unified temporal reference framework, and outputs the alignment attribute matrix of each regulated object in the unified attribute space and unified temporal framework. The expression is:

[0018] ;

[0019] in, This represents the alignment attribute matrix of the u-th supervised object. This represents the number of time slices or business phase nodes for the u-th regulated object within the unified time-series framework. This represents the total number of dimensions in the uniform attribute space. Indicates the total number of data sources. This represents the original attribute block of the u-th supervised object in the s-th data source. This represents the source-level normalization function of the s-th data source. This represents the set of attribute mapping parameters for the s-th data source. This represents the time alignment parameter set of the s-th data source. This represents the cross-source projection function for the s-th data source. Indicates the feature concatenation operator;

[0020] S103 divides the attributes into several causal factor clusters, constructs an event evolution chain in the time dimension, and weaves the causal factor clusters, event evolution chain, and context constraints together into a multi-dimensional field-of-view representation oriented towards triggering decisions. The triggering causal field-of-view tensor for each monitored object is output, with the expression:

[0021] ;

[0022] in, This represents the triggering causal field tensor of the u-th monitored object. This represents the length of the u-th monitored object in the causal attribute expansion dimension. This represents the length of the u-th monitored object in the event evolution dimension. Let G represent the length of the u-th regulated object in the context constraint dimension, and let G represent the number of causal factor clusters. Indicates the first The aggregate weights of each regulated object on the g-th causal factor cluster satisfy: , This indicates that the u-th regulated object is in the th... Attribute representation vectors on a cluster of causal factors This represents the event evolution vector of the u-th regulated object on the g-th causal factor cluster. Indicates the first The context constraint vector of a regulated object on the g-th causal factor cluster This represents the tensor outer product operator.

[0023] In a preferred embodiment of the present invention, in step 103, the plurality of causal factor clusters include subject attribute clusters, historical behavior clusters, risk transmission clusters, on-site constraint clusters, and feedback prior clusters. The construction of an event evolution chain in the time dimension is used to characterize the stage relationship between the reporting, deployment, arrival, inspection, release, and feedback links.

[0024] In a preferred embodiment of the present invention, step S2 includes the following steps:

[0025] S201, analyze the various local structures in the triggering causal field of view, and jointly extract the local attribute fragments, local event chain fragments, and local constraint context fragments into a complete set of candidate causal units. Based on the temporal precedence, disturbance sensitivity, and structural salience between each candidate causal unit and the triggering result, perform a candidateliness measure on the candidate causal units, and aggregate the candidate causal units that meet the candidate determination criteria into the triggering causal candidate set of the current business object, expressed as:

[0026] ;

[0027] ;

[0028] ;

[0029] in, This represents the triggering causal field tensor of the u-th monitored object. The complete set of candidate causal units obtained by analysis Indicates the first The first among the regulatory targets One candidate causal unit, This represents the total number of candidate causal units resolved in the triggering causal field of the u-th monitored object. This represents the candidate set of triggering causes for the u-th regulated object. This represents the temporal precedence measure of the j-th candidate causal unit. This represents the perturbation sensitivity measure of the j-th candidate causal unit. This represents the structural saliency measure of the j-th candidate causal unit. , and Let the weights represent the fusion weights for temporal precedence, perturbation sensitivity, and structural saliency, respectively, satisfying: , This represents the threshold for determining the candidate set of the u-th regulatory object;

[0030] S202, constructing a local intervention perspective around each candidate causal unit, examining the magnitude of change in the triggered result when the factor undergoes controlled perturbation, and combining the factor's propagation effect on other factor chains, measuring its direct and indirect contributions, jointly estimating the direct driving strength, conduction amplification strength, and net contribution after background stripping of each candidate factor, and selecting a subset of causal factors that substantially contribute to the triggered result, expressed as:

[0031] ;

[0032] ;

[0033] in, This represents the overall causal contribution strength of the j-th candidate causal unit among the u-th regulatory objects. This indicates the direct driving strength of the candidate causal unit. This indicates the conduction amplification intensity of the candidate causal unit. This indicates the net contribution strength of the candidate causal unit. , and The fusion weights for direct driving intensity, conduction amplification intensity, and net contribution intensity, respectively, satisfy: , , This represents the causal contribution screening threshold for the u-th regulated object. This represents the set of causal contribution factors retained after contribution screening for the u-th regulated object;

[0034] S203: Cluster the causal contribution factors according to business semantics and triggering link location to form a set of principal cause clusters. Estimate the dominance and synergistic effect among factors within each principal cause cluster, extract the core principal causes and auxiliary factors that represent the triggering effect of the principal cause cluster, sort and profile the focusing results of each principal cause cluster, generate a triggering causal profile for the current business object, and output the set of triggering principal causes, the set of auxiliary factors, and their intensity ranking results. The expression is:

[0035] ;

[0036] ;

[0037] ;

[0038] in, This represents the set of causal contribution factors for the u-th regulated object. The set of principal clusters obtained by clustering according to business semantics and trigger link location This represents the b-th principal factor cluster in the u-th regulated object. This represents the number of principal clusters formed by the u-th regulated object. This represents the trigger causal profile vector of the u-th monitored object. This represents the focus weight of the b-th principal factor cluster. This represents the core principal factor encoding vector of the b-th principal factor cluster. This represents the auxiliary factor encoding vector of the b-th principal factor cluster. This represents the sorting structure encoding vector of the b-th principal cluster. This represents the set of triggering causes for the u-th monitored object. Indicates selection based on focus weight. Each principal factor cluster corresponds to the operation of the core principal factor. This represents the number of main factors output by the u-th monitored object. This represents the set of focus weights for all principal factor clusters of the u-th regulatory object.

[0039] In a preferred embodiment of the present invention, step S3 includes the following steps:

[0040] S301, for the u-th regulated object, its subject attributes, environmental constraints, key contexts, and current decision state are uniformly encoded to form a real-world base state. Decision variables that allow controlled substitution are extracted from the real-world base state, while keeping the subject attributes, environmental constraints, and key contexts unchanged. The output is the real-world base state vector used for twin generation and its corresponding set of substitutable decision variables, expressed as:

[0041] ;

[0042] ;

[0043] in, This represents the real-world base state vector of the u-th monitored object. This represents the base state encoding function for the u-th monitored object. This represents the main attribute state vector of the u-th supervised object. This represents the environmental constraint state vector of the u-th regulated object. This represents the actual decision-making state vector of the u-th regulated object. Let u represent the set of substitutable decision variables for the u-th regulatory object. Let r be the r-th alternative decision variable in the u-th regulated object. This represents the number of substitutable decision variables for the u-th regulatory object;

[0044] S302, decompose the real-world base state into an invariant part and a permissible substituent part. Around the decision variables, perform several controlled substitutions on the real-world decision state, constructing several virtual decision clones that run parallel to the real-world path. Recombine each virtual decision clone after controlled substitution with the invariant subject attribute state and environmental constraint state to form a family of counterfactual twin scenarios, expressed as:

[0045] ;

[0046] ;

[0047] ;

[0048] in, This represents the counterfactual twin scenario family generated by the u-th regulated object. Indicates the first The wth counterfactual twin scenario among the regulated entities This represents the number of counterfactual twin scenarios generated for the u-th regulatory object. This represents the virtual decision state vector corresponding to the w-th counterfactual twin scenario. Let w represent the replacement rule vector for the w-th counterfactual twin scenario. This represents the controlled replacement function for the u-th supervised object;

[0049] S303, for each counterfactual twin scenario, estimates its potential outcomes across different dimensions, performs parallel aggregation and normalization comparison of all scenario potential outcomes to form an effect vector for each scenario, and organizes all scenario effect vectors into an alternative strategy effect spectrum to characterize the comprehensive advantages and disadvantages of each alternative relative to the real path, expressed as:

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] in, This represents the effect vector of the w-th counterfactual twin scenario in the u-th regulated object. This represents the scenario effect deduction function for the u-th monitored object. This represents the potential outcome of the w-th counterfactual twin scenario in terms of regulatory hit rate. This represents the potential outcome of the w-th counterfactual twin scenario in terms of release timeliness. This represents the potential outcome of the w-th counterfactual twin scenario in terms of resource consumption. This represents the potential outcome of the w-th counterfactual twin scenario in terms of congestion risk. This represents the potential outcome of the w-th counterfactual twin scenario in terms of the probability of repeated flow. , , , and These represent the weighting coefficients for five dimensions: regulatory hit rate, release timeliness, resource consumption, congestion risk, and probability of duplicate transfer, respectively, satisfying the following: , , This represents the spectrum of alternative strategies for the u-th regulated object. This represents the number of counterfactual twin scenarios generated for the u-th regulatory object.

[0055] In a preferred embodiment of the present invention, in step S303, the decision variables for controlled replacement include whether it is triggered, the triggering time, the dispatch target, and the inspection path; in step S303, the different dimensions include regulatory hit rate, release timeliness, resource consumption, congestion risk, and probability of repeated transfer.

[0056] In a preferred embodiment of the present invention, step S4 includes the following steps:

[0057] S401, focusing on the same regulatory object, abstracts the necessity of triggering, the object to be dispatched, the timing of dispatching, the priority, and the disposal path into a unified decision vector. It maps the intensity of the triggering cause, the effect of alternative scenarios, on-site resource constraints, and business time window constraints into a unified decision space, forming a joint decision space that includes decision candidates, constraint boundaries, and benefit-cost structures. The expression is:

[0058] ;

[0059] ;

[0060] ;

[0061] in, This represents the joint decision space for triggering order dispatch for the u-th monitored object. Let z be the joint decision candidate vector among the u-th regulatory objects. This represents the number of joint decision-making candidates for the u-th regulatory object. This represents the trigger decision sub-vector in the z-th candidate vector. This represents the dispatch object subvector in the z-th candidate vector. This represents the time vector for dispatching orders in the z-th candidate vector. Let represent the priority sub-vector among the z-th candidate vectors. This represents the sub-vector of the disposal path in the z-th candidate vector. Let the joint decision boundary vector of the u-th regulated object be denoted as . Represents the decision space mapping function for the u-th regulatory object;

[0062] S402, for each joint decision candidate vector, estimate its risk-reward, time-efficiency benefit, and resource cost. By constructing a resonance balance function among the three types of benefits and costs, measure the overall synergy between each candidate vector in terms of regulatory effectiveness, customs clearance efficiency, and resource burden. Select the joint decision candidate vector with the resonance balance value as the triggering dispatch scheme for the current regulatory object. The expression is:

[0063] ;

[0064] ;

[0065] in, This represents the resonance equilibrium value of the z-th joint decision candidate vector among the u-th regulatory objects. Let z represent the risk-reward term of the z-th joint decision candidate vector. Let z represent the time-efficiency benefit term of the z-th joint decision candidate vector. Let z represent the resource cost term of the z-th joint decision candidate vector. , and Let these represent the balanced weights of risk-return, time-efficiency benefit, and resource cost, respectively, satisfying the following: , , This represents the joint decision vector obtained after resonant collaborative solution for the u-th regulatory object. This represents the number of joint decision-making candidates for the u-th regulatory object;

[0066] S403 decodes the joint decision vector, restoring the trigger status, dispatch target, time parameters, priority parameters, and path parameters into executable business fields. Based on the business system interface specification, the decoding results are organized into structured joint instructions, outputting joint decision instructions that can be directly sent to the inspection and scheduling system, the field operation system, or the subsequent feedback system. The expression is:

[0067] ;

[0068] ;

[0069] in, This represents the joint decision-making instruction vector for the u-th regulated object. This represents the instruction decoding function for the u-th monitored object. Represents the vector of trigger status fields. This represents a vector of target fields for dispatch orders. This represents a vector of fields indicating the order dispatch timing. Represents a priority field vector. Represents a vector of disposal path fields. This represents the structured joint decision-making instruction for the u-th regulated object. This represents the metadata vector associated with the joint decision-making instruction.

[0070] In a preferred embodiment of the present invention, step S5 includes the following steps:

[0071] S501, focusing on the feedback received regarding the inspection results, release results, transfer to manual processing results, time consumption results, anomaly handling results, and resource usage for the same regulated object, aligns various execution feedbacks with the original decision-making actions based on the regulated object identifier, trigger status field, dispatch target field, and processing path field in the joint decision-making instruction. This constructs a three-way correspondence between execution results, triggering causes, and dispatch paths, forming an execution evidence mapping that can be used for subsequent causal backtracking. The expression is:

[0072] ;

[0073] ;

[0074] ;

[0075] in, This represents the set of execution feedback records corresponding to the u-th supervised object. This represents the nth execution feedback record in the uth supervised object. This represents the number of execution feedback records received by the u-th monitored object. This represents the set of execution evidence mappings for the u-th regulated object. This represents the evidence mapping unit constructed from the nth execution feedback record. This represents the evidence mapping function for the u-th supervised object;

[0076] S502, based on the execution evidence mapping set, analyzes the causal correspondence between the verification hit results, release results, manual transfer results, time consumption results, anomaly handling results, and resource usage feedback and the triggering main cause and dispatch path. It estimates the verification strength of each triggering main cause, the execution deviation strength of each dispatch path, and the environmental mismatch strength of each strategy item. The verification strength, execution deviation strength, and environmental mismatch strength are jointly summarized into a calibration signal, used to characterize the direction of effectiveness correction of the original decision after actual execution. The expression is:

[0077] ;

[0078] ;

[0079] in, This represents the calibration strength of the h-th decision element within the u-th regulated object. This represents the total number of decision-making factors that require calibration assessment for the u-th regulated object. This represents the verification strength of the h-th decision element. This represents the intensity of the execution deviation of the h-th decision element. This represents the intensity of environmental mismatch for the h-th decision element. , and The fusion weights for validation strength, execution bias strength, and environment mismatch strength, respectively, satisfy: , , Represents the set of calibration strengths for the u-th monitored object;

[0080] S503 decomposes the calibration intensity set into three levels: triggering cause, dispatch object, and handling path, forming the cause weight correction, dispatch weight correction, and path preference correction respectively. These corrections are written into the corresponding parameter sets to update the original trigger weight, dispatch weight, and path selection preference, outputting a new closed-loop evolution parameter set for direct use by subsequent monitored objects. The expression is:

[0081] ;

[0082] ;

[0083] ;

[0084] in, This represents the triggering cause weight vector of the u-th monitored object before the update. This represents the updated trigger cause weight vector for the u-th monitored object. This represents the dispatch weight vector of the u-th monitored object before the update. This represents the updated dispatch weight vector for the u-th monitored object. This represents the path selection preference vector of the u-th monitored object before the update. This represents the updated path selection preference vector for the u-th monitored object. This represents the modified mapping function applied to the triggering factor weight vector. This represents the correction mapping function applied to the dispatch weight vector. This represents the correction mapping function acting on the path selection preference vector. This represents the closed-loop evolution update parameter set for the u-th monitored object.

[0085] Unlike existing technologies, the above technical solution achieves the following beneficial effects:

[0086] (1) This method enables focused identification and analysis of key causal factors affecting the release triggering results, thereby improving the accuracy, stability and interpretability of the release triggering decision;

[0087] (2) This method realizes counterfactual comparison and effect evaluation of different trigger schemes, dispatch object schemes, dispatch timing schemes and disposal path schemes, which improves the scientificity and rationality of decision-making;

[0088] (3) This method realizes the integrated joint decision-making of inspection and release triggering and dispatch decision, improves the rationality of inspection resource allocation, and reduces on-site congestion, duplicate transfer and invalid dispatch.

[0089] (4) This method realizes evidence reinjection, causal attribution and closed-loop update of execution results, and constructs a closed-loop mechanism of triggering, dispatching, execution, feedback and optimization, which improves the adaptive optimization capability and the accuracy and intelligence of customs supervision. Attached Figure Description

[0090] Figure 1 The flowchart is for a specific implementation method. Detailed Implementation

[0091] To explain in detail the technical content, structural features, objectives, and effects of the technical solution, the following description is provided in conjunction with specific embodiments and accompanying drawings.

[0092] like Figure 1As shown, this embodiment provides a closed-loop method for customs clearance trigger dispatch based on counterfactual causal learning, including the following steps: S1, establish a master index for the supervised object, perform cross-source alignment, temporal correlation, and state splicing to form a global data base for the current clearance trigger problem; S2, construct a layer-by-layer identification mechanism for trigger causal candidate set, causal contribution screening, and trigger main cause focusing, and decouple the disturbing factors in layers to form a trigger causal profile for the current business object; S3, based on the current real business state, under the premise of maintaining consistency of subject attributes, environmental constraints, and key contexts, perform controlled replacement of decision variables to generate several counterfactual causal... In the real-fact twin scenario, the potential outcomes of each counterfactual twin scenario in different dimensions are extrapolated in parallel to form an alternative strategy effect spectrum; S4, based on the causal focus identification results and the counterfactual alternative strategy extrapolation results, the trigger necessity, dispatch object, dispatch timing, priority order and handling path are jointly incorporated into the same decision space to construct a resonance balance relationship between risk and benefit, time efficiency benefit and resource cost; S5, the system receives the inspection hit results, release results, transfer to manual results, time consumption results, abnormal handling results and resource consumption feedback, and performs causal backtracking and effect calibration on the previous decisions based on the correspondence between the execution results, the triggering cause and the dispatch path.

[0093] In the specific implementation of step S1 of the above embodiment, the following steps are included:

[0094] S101, the input for this step is the set of original regulatory records from the reporting end, risk end, on-site end, and feedback end. For each record, extract its identifier field, time field, business context field, and spatial attribution field to form a record-level candidate description; then, using the candidate business object as the merging unit, comprehensively consider identifier consistency, time proximity, business context consistency, and spatial attribution consistency to perform merging judgment on cross-source records; finally, generate a unique regulatory object master index for each single shipment, single vehicle trip, single container trip, or single batch of business, with the expression:

[0095] ;

[0096] ;

[0097] in, Indicates the first The primary index generated by each regulated object Indicates the primary index candidate space. Indicates the total number of entities under supervision. This represents the number of original records associated with the u-th supervised object. This represents the identifier consistency score of the nth record within the uth supervised object, used to measure the degree of matching between key identifiers such as customs declaration number, waybill number, container number, train number, and task number. This represents the time proximity score of the nth record in the uth regulatory object, used to measure the degree of time consistency between the declaration time, arrival time, inspection time, and feedback time. This represents the business context consistency score of the nth record in the uth regulatory object, used to measure the consistency of business semantics such as product attributes, enterprise entity, transportation mode, and regulatory method; This represents the spatial attribution consistency score of the nth record in the uth regulatory object, used to measure the matching relationship between spatial locations such as ports, checkpoints, storage yards, and inspection areas; , , and These represent the weight coefficients for identifier consistency, temporal proximity, business context consistency, and spatial affiliation consistency, respectively, satisfying... , .

[0098] S102, around the master index of each regulated object, performs field extraction and attribute renaming on data from different sources, mapping synonymous and heteronymous fields to a unified attribute space; uniformly standardizes discrete states, classification codes, timestamps, numerical units, and missing values; aligns attributes from each source to a unified temporal reference framework based on the master index of the regulated object and the event timeline; finally, outputs the alignment attribute matrix of each regulated object in the unified attribute space and unified temporal framework, the expression of which is:

[0099] ;

[0100] in, This represents the alignment attribute matrix of the u-th supervised object; This represents the number of time slices or business phase nodes for the u-th monitored object within the unified time-series framework; M represents the total number of dimensions in the unified attribute space; and S represents the total number of data sources. This represents the original attribute block of the u-th supervised object in the s-th data source; This represents the source normalization function for the s-th data source, used to complete field renaming, unit unification, encoding mapping, missing data normalization, and time format standardization; This represents the set of attribute mapping parameters for the s-th data source, which describes the mapping rules from the attributes of that data source to the unified attribute space. This represents the time alignment parameter set of the s-th data source, used to describe the alignment rules of this data source on a unified time axis or a unified business phase axis; This represents the cross-source projection function for the s-th data source, used to map the normalized in-source properties to a unified property space and align them to a unified temporal framework; The feature concatenation operator is used to combine the projection results of multiple data sources into a unified alignment attribute matrix for the same regulatory object according to a unified attribute dimension.

[0101] S103 divides attributes into multiple causal factor clusters, including main attribute cluster, historical behavior cluster, risk transmission cluster, on-site constraint cluster, and feedback prior cluster; it constructs an event evolution chain in the time dimension to represent the stage relationships between reporting, deployment, arrival, inspection, release, and feedback; it weaves the causal factor clusters, event evolution chain, and contextual constraints together into a multi-dimensional field-of-view representation oriented towards trigger decision-making; finally, it outputs the trigger causal field-of-view tensor for each regulated object, with the expression:

[0102] ;

[0103] in, This represents the triggering causal field tensor of the u-th monitored object; This represents the length of the u-th supervised object in the causal attribute expansion dimension; This represents the length of the u-th monitored object in the event evolution dimension; G represents the length of the u-th regulatory object in the context constraint dimension; G represents the number of causal factor clusters. Represents the aggregate weight of the u-th regulated object on the g-th causal factor cluster, satisfying , This represents the attribute representation vector of the u-th regulatory object on the g-th causal factor cluster, used to represent the static or semi-static business attributes in that factor cluster; This represents the event evolution vector of the u-th regulatory object on the g-th causal factor cluster, used to characterize the business stage evolution relationship related to this factor cluster; This represents the context constraint vector of the u-th regulated object on the g-th causal factor cluster, used to characterize background factors such as port status, resource load, rule boundaries, and time window constraints; This represents the tensor outer product operator, used to couple attribute representation vectors, event evolution vectors, and context constraint vectors into a three-dimensional field of view unit.

[0104] In the specific implementation of step S2 of the above embodiment, the following steps are included:

[0105] S201, analyze the various local structures in the triggering causal field of view, and jointly extract the local attribute fragments, local event chain fragments, and local constraint context fragments into a complete set of candidate causal units; then, based on the temporal precedence, disturbance sensitivity, and structural saliency of each candidate causal unit and the triggering result, perform a candidateliness measure on the candidate causal units; aggregate the candidate causal units that meet the candidate determination criteria into the triggering causal candidate set of the current business object, expressed as:

[0106] ;

[0107] ;

[0108] ;

[0109] in, This represents the triggering causal field tensor of the u-th monitored object. The complete set of candidate causal units obtained through analysis; This represents the j-th candidate causal unit in the u-th supervised object, which is a combined representation consisting of local attribute fragments, local event evolution fragments, and local constraint context fragments. This represents the total number of candidate causal units resolved in the triggering causal field of the u-th monitored object; This represents the candidate set of triggering causes for the u-th monitored object; This represents the temporal precedence measure of the j-th candidate causal unit, used to measure whether the factor appears before the triggering result in time and has precedence; This represents the perturbation sensitivity measure of the j-th candidate causal unit, used to measure how sensitive the factor is to the triggering result when local changes occur; represents the structural saliency measure of the j-th candidate causal unit, used to measure the degree of structural prominence of the factor in the triggering causal field of view; , and Let the weights represent the fusion weights for temporal precedence, perturbation sensitivity, and structural saliency, respectively, satisfying: , This represents the threshold for determining the candidate set of the u-th monitored object, used to control the lower bound of causal units entering the candidate set of triggering causality.

[0110] S202, construct a local intervention perspective around each candidate causal unit to examine the magnitude of change in the triggered result when the factor undergoes controlled perturbation; combine the factor's propagation effect on other factor chains to measure its direct and indirect contributions; jointly estimate the direct driving strength, conduction amplification strength, and net contribution after background stripping of each candidate factor; and select a subset of causal factors that substantially contribute to the triggered result, expressed as:

[0111] ;

[0112] ;

[0113] in, This represents the overall causal contribution strength of the j-th candidate causal unit among the u-th regulatory objects; This indicates the direct driving strength of the candidate causal unit, used to characterize the degree of direct influence of the controlled change of the factor on the triggering result; This indicates the conduction amplification intensity of the candidate causal unit, used to characterize the propagation effect of the factor indirectly influencing the triggering result through other causal links; This represents the net contribution strength of the candidate causal unit, used to characterize the pure driving effect retained by the factor after removing background covariants; , and The fusion weights for direct driving intensity, conduction amplification intensity, and net contribution intensity, respectively, satisfy: , ; This represents the causal contribution screening threshold for the u-th regulated object; This represents the set of causal contribution factors retained after contribution screening for the u-th regulated object.

[0114] S203, although the causal contribution screening has retained the truly explanatory causal factors, these factors may still be scattered across multiple levels in terms of business semantics, such as abnormal product attributes, abnormal corporate historical behavior, regulatory rule conflicts, risk transmission links, and on-site disturbances, making it difficult to directly form an actionable trigger basis. This step further performs principal cause focusing processing on the causal contribution factor set: the causal contribution factors are clustered according to business semantics and trigger link positions to form a principal cause cluster set; within each principal cause cluster, the dominance and synergistic effect between factors are estimated, and the core principal cause and auxiliary factors that best represent the triggering effect of the principal cause cluster are extracted; the focusing results of each principal cause cluster are sorted and profiled to generate a trigger causal profile for the current business object; finally, the trigger principal cause set, auxiliary factor set, and their intensity ranking results are output. Through this step, the scattered causal contribution information can be transformed into a clearly structured, interpretable, and directly serviceable trigger principal cause representation for subsequent joint decision-making, expressed as:

[0115] ;

[0116] ;

[0117] ;

[0118] in, This represents the set of causal contribution factors for the u-th regulated object. The set of principal clusters obtained by clustering according to business semantics and trigger link location; This represents the b-th principal factor cluster within the u-th regulated object; This represents the number of principal clusters formed by the u-th regulatory object; This represents the trigger causal profile vector of the u-th regulatory object, which is used to comprehensively characterize the object's main cause structure, auxiliary factor structure, and intensity ranking structure. This represents the focus weight of the b-th primary factor cluster, used to measure the overall dominance of this primary factor cluster in the current triggering result; This represents the core principal factor encoding vector of the b-th principal factor cluster, used to characterize the most representative dominant triggering factor in this cluster; This represents the auxiliary factor encoding vector of the b-th main factor cluster, used to characterize the auxiliary driving factors that work together with the core main factor; The sorting structure encoding vector of the b-th principal factor cluster is used to characterize the importance order and synergistic relationship of the factors within the cluster; The feature concatenation operator is used to combine the core principal factor encoding, auxiliary factor encoding, and sorting structure encoding into a single principal factor cluster representation. This represents the set of triggering factors for the u-th monitored object; Indicates selection based on focus weight. Each primary factor cluster corresponds to the operation of the core primary factor; This represents the number of main factors ultimately output by the u-th monitored object; This represents the set of focus weights for all principal factor clusters of the u-th regulatory object.

[0119] In the specific implementation of step S3 of the above embodiment, the following steps are included:

[0120] S301. Since counterfactual twin generation cannot be arbitrarily constructed detached from the current real business state, but should be based on the real state of the same regulated object, this step first uniformly encodes the subject attributes, environmental constraints, key contexts, and current decision state of the u-th regulated object to form a real base state. Then, it extracts controllable substitution decision variables from the real base state, including whether it is triggered, the triggering timing, the dispatch target, and the inspection path, while keeping the subject attributes, environmental constraints, and key contexts unchanged. Finally, it outputs the real base state vector used for twin generation and its corresponding set of substitutable decision variables, expressed as:

[0121] ;

[0122] ;

[0123] in, This represents the real-world base state vector of the u-th monitored object; The base state encoding function for the u-th monitored object is used to jointly map the trigger causal field tensor, the trigger causal profile vector, and the set of triggering main causes into a unified state representation. This represents the subject attribute state vector of the u-th regulated object, used to characterize relatively stable business subject information such as cargo attributes, enterprise entity, and transportation attributes; The environmental constraint state vector of the u-th regulated object is used to characterize external constraint information such as on-site resource load, rule boundaries, time window limits, and port status. This represents the current decision-making state vector of the u-th regulated object, used to characterize the triggering, dispatching, and inspection and handling status under the current real path; Let represent the set of alternative decision variables for the u-th regulatory object; Let r represent the r-th alternative decision variable in the u-th regulatory object; This represents the number of alternative decision variables for the u-th regulatory object.

[0124] S302, since the goal of counterfactual analysis is to replace only controlled decision variables while maintaining consistency in subject attributes, environmental constraints, and key contexts, this step first decomposes the real-world base state into "unchanging parts" and "replaceable parts." Around replaceable decision variables such as whether a trigger occurs, the trigger timing, the order assignment target, and the verification path, various controlled substitutions are performed on the real-world decision state to construct multiple virtual decision clones running parallel to the real-world path. Each virtual decision clone after controlled substitution is then combined with the unchanging subject attribute state and environmental constraint state to form a counterfactual twin scenario family, expressed as:

[0125] ;

[0126] ;

[0127] ;

[0128] in, This represents the counterfactual twin scenario family generated by the u-th regulatory object; This represents the w-th counterfactual twin scenario within the u-th regulated object; This represents the number of counterfactual twin scenarios generated for the u-th regulatory object; This represents the virtual decision state vector corresponding to the w-th counterfactual twin scenario; The vector representing the replacement rule for the w-th counterfactual twin scenario indicates which decision variables are replaced and what alternative values ​​are used to replace them in this scenario. This represents the controlled substitution function for the u-th supervised object, used to perform controlled substitution of specified variables in the actual decision state while keeping the subject attribute state and environmental constraint state unchanged.

[0129] S303. Since each counterfactual twin scenario differs from the real path only in the configuration of decision variables, it can be considered as a test vehicle for different alternative strategies for the same regulatory object. This step first estimates the potential outcomes of each counterfactual twin scenario across multiple dimensions, including regulatory hit rate, release timeliness, resource consumption, congestion risk, and probability of repeated transfers. Then, the multidimensional potential outcomes of all scenarios are aggregated and normalized in parallel for comparison, forming an effect vector for each scenario. Finally, the effect vectors of all scenarios are organized into an alternative strategy effect spectrum to characterize the comprehensive advantages and disadvantages of each alternative scheme relative to the real path, expressed as:

[0130] ;

[0131] ;

[0132] ;

[0133] ;

[0134] in, This represents the effect vector of the w-th counterfactual twin scenario in the u-th regulated object; Represents the scenario effect inference function for the u-th regulatory object, used to estimate the potential outcomes of a single counterfactual twin scenario across multiple effect metrics; This represents the potential outcome of the w-th counterfactual twin scenario in terms of regulatory hit rate. This represents the potential outcome of the w-th counterfactual twin scenario in terms of release timeliness; This represents the potential outcome of the w-th counterfactual twin scenario in terms of resource consumption. This represents the potential outcome of the w-th counterfactual twin scenario in terms of congestion risk. This represents the potential outcome of the w-th counterfactual twin scenario in the dimension of repeated flow probability; , , , and These represent the weighting coefficients for five dimensions: regulatory hit rate, release timeliness, resource consumption, congestion risk, and probability of duplicate transfer, respectively, satisfying the following: , , Let represent the alternative strategy effect spectrum of the u-th regulatory object, which consists of the effect vectors of all counterfactual twin scenarios; This represents the number of counterfactual twin scenarios generated for the u-th regulatory object.

[0135] In the specific implementation of step S4 of the above embodiment, the following steps are included:

[0136] S401. Because "whether to trigger" and "how to dispatch" are usually separated into two independent processing stages in traditional solutions, the trigger judgment in the previous stage cannot explicitly perceive the resource constraints and path benefits in the subsequent stage. Therefore, this step first focuses on the same regulatory object and abstracts the trigger necessity, dispatch object, dispatch timing, priority, and disposal path into a unified decision vector; it then maps the trigger cause strength, alternative scenario effects, on-site resource constraints, and business time window constraints into a unified decision space; forming a joint decision space that includes decision candidates, constraint boundaries, and benefit-cost structures, expressed as:

[0137] ;

[0138] ;

[0139] ;

[0140] in, This represents the joint decision space for triggering dispatch for the u-th monitored object; This represents the z-th joint decision candidate vector among the u-th regulatory objects; This represents the number of joint decision-making candidates for the u-th regulatory object; This represents the trigger determination subvector in the z-th candidate vector, used to describe whether a trigger has occurred and the trigger level; This represents the dispatch object subvector in the z-th candidate vector, used to describe the inspection unit, post, equipment, or disposal entity; This represents the time vector for dispatching orders in the z-th candidate vector, used to describe when the dispatching action is issued; This represents the priority subvector in the z-th candidate vector, used to describe the priority level of tasks; This represents the processing path sub-vector in the z-th candidate vector, used to describe the path selection such as machine inspection, manual review, and physical inspection; Let represent the joint decision boundary vector of the u-th regulatory object, which is used to characterize the constraint boundary and benefit-cost structure jointly determined by the triggering causal profile, the triggering main cause, the effect of alternative strategies, and the counterfactual scenario; Let represent the decision space mapping function for the u-th regulatory object.

[0141] S402. Because different joint decision-making candidates may have mutual constraints or even conflicts in dimensions such as risk-reward, timeliness benefit, and resource cost, they cannot be selected independently based on a single indicator. This step first estimates the risk-reward, timeliness benefit, and resource cost of each joint decision-making candidate vector; then, by constructing a resonance balance function among the three types of benefits and costs, the overall synergy of each candidate in "regulatory effectiveness - customs clearance efficiency - resource burden" is measured; finally, the joint decision-making candidate vector with the optimal resonance balance value is selected as the optimal triggering dispatch scheme for the current regulatory object, expressed as:

[0142] ;

[0143] ;

[0144] in, This represents the resonance equilibrium value of the z-th joint decision candidate vector among the u-th regulatory objects; The risk-benefit term for the z-th joint decision candidate vector is used to measure the benefits of the candidate solution in terms of verification hit rate, risk interception, and anomaly detection. The time efficiency benefit term represents the z-th joint decision candidate vector, which measures the benefits of the candidate solution in shortening waiting time, reducing release delay, and improving customs clearance efficiency. The resource cost term represents the z-th joint decision candidate vector, which measures the cost of the candidate solution in terms of equipment usage, job load, inspection capacity consumption, and path burden. , and Let these represent the balanced weights of risk-return, time-efficiency benefit, and resource cost, respectively, satisfying the following: , , This represents the optimal joint decision vector obtained after resonant collaborative solution for the u-th regulatory object; This represents the constructed joint decision space for triggering order dispatch; This represents the number of joint decision-making candidates for the u-th regulatory object.

[0145] S403, although the optimal joint decision vector provides the optimal combination of trigger judgment, dispatch object, dispatch timing, priority order, and disposal path for the current monitored object, it is essentially still an internal representation oriented towards decision-making and cannot be directly used as an execution instruction in the business system. This step first decodes the fields of the optimal joint decision vector, restoring the trigger status, dispatch target, time parameter, priority parameter, and path parameter into executable business fields; according to the business system interface specification, the decoding result is organized into a structured joint instruction; the output is a joint decision instruction that can be directly issued to the inspection and scheduling system, the field operation system, or the subsequent feedback system. Through this step, the optimal decision result obtained can be converted into a business instruction with actual execution significance, expressed as:

[0146] ;

[0147] ;

[0148] in, This represents the joint decision-making instruction vector for the u-th supervised object; This represents the instruction decoding function for the u-th supervised object, used to convert the optimal joint decision vector into business-executable fields; This represents a vector of trigger status fields, used to describe whether a trigger has occurred, the trigger level, and the trigger identifier. This represents a vector of dispatch target fields, used to describe the inspection unit, post, or equipment object to which the order is dispatched; This represents a vector of dispatch timing fields, used to describe the dispatch execution time, window, or delay parameters; This represents a priority field vector used to describe task priority and sorting position; This represents a vector of processing path fields, used to describe the selected inspection path and subsequent processing methods; This represents the structured joint decision-making instruction for the u-th supervised object; This represents a metadata vector associated with the joint decision-making instruction, used to record auxiliary information such as the regulatory object identifier, generation time, decision version, and constraint summary.

[0149] In the specific implementation of step S5 of the above embodiment, the following steps are included:

[0150] S501, The input for this step is the structured joint decision instruction. This includes the on-site execution feedback data corresponding to the instruction. Since on-site execution feedback includes not only results such as inspection hit, release, transfer to manual processing, time consumption, and anomaly handling, but also resource feedback information such as job occupancy, equipment occupancy, time window offset, and congestion spread, it is necessary to first organize the scattered execution feedback into standardized evidence representations that can participate in decision correction. Firstly, focusing on the same regulated object, feedback on inspection hit results, release results, transfer to manual processing results, time consumption results, anomaly handling results, and resource consumption is received. Based on the regulated object identifier, trigger status field, dispatch target field, and handling path field in the joint decision instruction, various execution feedbacks are aligned with the original decision actions. A three-element correspondence of "execution result - triggering cause - dispatch path" is constructed to form an execution evidence mapping that can be used for subsequent causal backtracking. The expression is:

[0151] ;

[0152] ;

[0153] ;

[0154] in, This represents the set of execution feedback records corresponding to the u-th supervised object; This represents the nth execution feedback record in the uth supervised object; This represents the number of execution feedback records received by the u-th supervised object; Represents the set of execution evidence mappings for the u-th regulated object; This represents the evidence mapping unit constructed from the nth execution feedback record, used to characterize the correspondence between the feedback record and the original triggering cause, dispatch path, and decision instruction; The evidence mapping function for the u-th regulatory object is used to associate and align execution feedback records with joint decision instructions, triggering cause sets, and joint decision instruction vectors.

[0155] S502. Since the execution feedback contains both positive evidence that verifies the correctness of the original triggering cause and dispatch path, and negative evidence that indicates the original decision-making bias, strategy mismatch, and path inefficiency, further causal backtracking and effect calibration of the previous decision are required. First, based on the execution evidence mapping set, analyze the causal correspondence between the verification hit results, release results, manual transfer results, time consumption results, anomaly handling results, and resource usage feedback and the triggering cause and dispatch path; estimate the verification strength of each triggering cause, the execution bias strength of each dispatch path, and the environmental mismatch strength of each strategy item; combine the verification strength, execution bias strength, and environmental mismatch strength into a calibration signal to characterize the direction of effectiveness correction of the original decision after actual execution, expressed as: .

[0156] ;

[0157] ;

[0158] in, This represents the calibration strength of the h-th decision element in the u-th regulatory object; This represents the total number of decision-making factors that require calibration assessment for the u-th regulated object; This represents the verification strength of the h-th decision element, used to characterize the degree to which the decision element is supported by positive evidence after actual implementation; This represents the intensity of the execution deviation of the h-th decision element, used to characterize the degree to which this decision element causes additional time consumption, resource waste, congestion spread, or redundant flow during the execution process; This represents the environmental mismatch intensity of the h-th decision element, used to characterize the degree to which this decision element becomes ineffective, weakened, or no longer applicable under changes in the current business environment; , and The fusion weights for validation strength, execution bias strength, and environment mismatch strength, respectively, satisfy: , , Let represent the set of calibration strengths for the u-th regulated object, which consists of the calibration strengths of all decision elements.

[0159] S503, since the ultimate goal of evidence reinjection is not merely to provide a one-time calibration result, but to continuously feed the calibration result back into the subsequent decision-making mechanism, this step further converts the calibration intensity into an update quantity that can be applied to the triggering strategy, dispatching strategy, and path preference. First, the calibration intensity set is decomposed into three levels: "triggering cause - dispatching object - handling path," forming the cause weight correction quantity, dispatching weight correction quantity, and path preference correction quantity, respectively. The correction quantities are written into the corresponding parameter groups to update the original triggering weight, dispatching weight, and path selection preference. A new closed-loop evolution parameter group is output for direct use when processing subsequent monitored objects; the expression is as follows.

[0160] ;

[0161] ;

[0162] ;

[0163] in, This represents the triggering cause weight vector of the u-th monitored object before the update; This represents the updated trigger cause weight vector for the u-th monitored object; This represents the dispatch weight vector of the u-th supervised object before the update; This represents the updated dispatch weight vector for the u-th monitored object; This represents the path selection preference vector of the u-th monitored object before the update; This represents the updated path selection preference vector for the u-th monitored object; This represents the correction mapping function applied to the triggering cause weight vector, used to convert the calibration intensity set into the triggering cause weight correction amount; This represents the correction mapping function applied to the dispatch weight vector, used to convert the calibration intensity set into the dispatch weight correction amount; This represents the correction mapping function acting on the path selection preference vector, used to convert the set of calibration intensity into path preference correction values; This represents the closed-loop evolution update parameter set for the u-th monitored object, which consists of the updated trigger cause weight vector, the updated dispatch weight vector, and the updated path selection preference vector.

[0164] To demonstrate the effectiveness of this embodiment, an internal verification sample set was constructed based on historical business data from real customs business scenarios. This historical business data included declaration data, risk supervision data, on-site operation status data, and subsequent feedback data, which were then anonymized, cleaned, standardized, and correlated to form the business samples used for testing. This method was compared with fixed-rule clearance methods, correlation risk scoring methods, and methods that separate triggering and dispatching. The experimental results are shown in Table 1. Overall, this method performs best in all core indicators, demonstrating that the proposed method can simultaneously improve the accuracy of clearance triggering, the effectiveness of inspection hits, and the efficiency of dispatching, while effectively reducing false triggers, missed triggers, and invalid dispatches.

[0165] From the perspective of release triggering effectiveness, this method, by introducing causal focusing identification and counterfactual reasoning mechanisms, can more accurately identify the key factors that truly affect release triggering, avoiding judgment biases caused by relying solely on static rules or superficial correlations. Regarding false triggering and missed triggering rates, this method not only reduces unnecessary release triggering and wastes regulatory resources, but also reduces situations where triggers should have been triggered but were not, thereby improving the overall accuracy and reliability of supervision. In terms of inspection hit rate, the triggering decisions output by this method are more targeted, enabling limited inspection resources to be allocated more towards truly high-risk or high-value business objects, thus improving the effectiveness of inspection operations. From the perspective of clearance efficiency, this method does not simply pursue higher inspection intensity, but rather, while ensuring regulatory effectiveness, reduces waiting, congestion, and redundant processing through integrated and coordinated optimization of triggering and dispatching, thereby improving overall clearance efficiency. Regarding dispatching resource utilization and invalid dispatching rate, this method, while making triggering judgments, comprehensively considers the on-site resource status and disposal path selection, making dispatching results more reasonable and reducing resource idleness and inefficient scheduling.

[0166] Table 1 Comparison of experimental results

[0167] method Release trigger accuracy rate / % False trigger rate / % Missed trigger rate / % Check hit rate / % Average clearance time / m Order dispatch resource utilization rate / % Invalid order rate / % Fixed rule inspection and release 84.6 18.7 16.4 41.3 128 71.5 14.2 Correlation risk score 88.9 13.8 11.9 48.6 112 76.3 10.8 Separation of triggering and dispatching 90.7 11.6 9.8 53.9 98 81.4 8.7 This method 95.8 6.1 5.4 65.7 76 89.8 4.1

[0168] It should be noted that although the above embodiments have been described herein, this does not limit the scope of patent protection of the present invention. Therefore, any changes and modifications made to the embodiments described herein based on the innovative concept of the present invention, or equivalent structural or procedural transformations made using the content of the present invention's specification and drawings, directly or indirectly applying the above technical solutions to other related technical fields, are all included within the scope of patent protection of the present invention.

Claims

1. A closed-loop method for customs clearance-triggered dispatching based on counterfactual causal learning, characterized in that, Includes the following steps: S1. Establish a master index for the regulated objects, perform cross-source alignment, temporal correlation and status splicing to form a global data base for the current release triggering issues; S2, construct a layered identification mechanism for triggering causal candidate set, causal contribution screening, and triggering main cause focus, and decouple the disturbance factors in layers to form a triggering causal profile for the current business object; S3, based on the current real business state, while maintaining consistency in subject attributes, environmental constraints and key contexts, performs controlled replacement of decision variables to generate several counterfactual twin scenarios, and performs parallel extrapolation of the potential results of each counterfactual twin scenario in different dimensions to form an alternative strategy effect spectrum; S4, based on the causal focus identification results and the counterfactual substitution strategy deduction results, incorporates the trigger necessity, dispatch object, dispatch timing, priority order and disposal path into the same decision space to construct a resonance balance relationship between risk and benefit, timeliness benefit and resource cost; S5 receives the results of the inspection hit, release, transfer to manual processing, time consumption, abnormal handling, and resource usage. Based on the correspondence between the execution results, the triggering cause, and the dispatch path, it performs causal backtracking and effect calibration on the previous decisions. Step S2 includes the following steps: S201, analyze the various local structures in the triggering causal field of view, and jointly extract the local attribute fragments, local event chain fragments and local constraint context fragments into a complete set of candidate causal units. Based on the temporal precedence, disturbance sensitivity and structural salience between each candidate causal unit and the triggering result, measure the candidateness of the candidate causal units, and aggregate the candidate causal units that meet the candidate determination conditions into the triggering causal candidate set of the current business object. S202: Construct a local intervention perspective around each candidate causal unit, examine the magnitude of change in the triggered result when the factor is subject to controlled perturbation, combine the propagation effect of the factor on other factor links, measure its direct and indirect contributions, jointly estimate the direct driving strength, transmission amplification strength and net contribution after background stripping of each candidate factor, and screen out a subset of causal factors that have substantial contributions to the triggered result. S203: The causal contribution factors are clustered according to business semantics and triggering link position to form a set of main cause clusters. Within each main cause cluster, the dominance and synergistic effect between factors are estimated. The core main cause and auxiliary factors that can represent the triggering effect of the main cause cluster are extracted. The focusing results of each main cause cluster are sorted and profiled to generate a triggering causal profile for the current business object. The set of triggering main causes, the set of auxiliary factors and their intensity ranking results are output.

2. The customs clearance-triggered order dispatch closed-loop method based on counterfactual causal learning according to claim 1, characterized in that, Step S1 includes the following steps: S101, for each record, extract its identifier field, time field, business context field, and spatial attribution field to form a record-level candidate description. Then, using the candidate business object as the merging unit, comprehensively consider identifier consistency, time proximity, business context consistency, and spatial attribution consistency to perform merging judgment on cross-source records. Generate a unique regulatory object master index for each single shipment, single vehicle trip, single container trip, or single batch of business, obtaining the master index set and its corresponding cross-source record attribution relationship, expressed as: , , in, Indicates the first The primary index generated by each regulated object Indicates the primary index candidate space. Indicates the total number of entities under supervision. Indicates the first The number of original records associated with each regulated entity. This represents the identifier consistency score of the nth record in the uth supervised object. This represents the temporal proximity score of the nth record within the uth monitored object. This represents the business context consistency score of the nth record in the uth supervised object. This represents the spatial attribution consistency score of the nth record within the uth supervised object. , , and Let represent the weight coefficients for identifier consistency, temporal proximity, business context consistency, and spatial affiliation consistency, respectively, satisfying: , ; S102, around the master index of each regulated object, performs field extraction and attribute renaming on data from different sources, maps synonymous fields to a unified attribute space, and unifies discrete states, classification codes, timestamps, numerical units, and missing values. Based on the master index of the regulated object and the event timeline, aligns the attributes from each source to a unified temporal reference framework, and outputs the alignment attribute matrix of each regulated object in the unified attribute space and unified temporal framework. The expression is: , in, This represents the alignment attribute matrix of the u-th supervised object. This represents the number of time slices or business phase nodes for the u-th regulated object within the unified time-series framework. This represents the total number of dimensions in the uniform attribute space. Indicates the total number of data sources. This represents the original attribute block of the u-th supervised object in the s-th data source. This represents the source-level normalization function of the s-th data source. This represents the set of attribute mapping parameters for the s-th data source. This represents the time alignment parameter set of the s-th data source. This represents the cross-source projection function for the s-th data source. Indicates the feature concatenation operator; S103 divides the attributes into several causal factor clusters, constructs an event evolution chain in the time dimension, and weaves the causal factor clusters, event evolution chain, and context constraints together into a multi-dimensional field-of-view representation oriented towards triggering decisions. The triggering causal field-of-view tensor for each monitored object is output, with the expression: , in, This represents the triggering causal field tensor of the u-th monitored object. This represents the length of the u-th monitored object in the causal attribute expansion dimension. This represents the length of the u-th monitored object in the event evolution dimension. Let G represent the length of the u-th regulated object in the context constraint dimension, and let G represent the number of causal factor clusters. Indicates the first The aggregate weights of each regulated object on the g-th causal factor cluster satisfy: , This indicates that the u-th regulated object is in the th... Attribute representation vectors on a cluster of causal factors This represents the event evolution vector of the u-th regulated object on the g-th causal factor cluster. Indicates the first The context constraint vector of a regulated object on the g-th causal factor cluster This represents the tensor outer product operator.

3. The customs clearance-triggered order dispatch closed-loop method based on counterfactual causal learning according to claim 2, characterized in that: In step 103, the plurality of causal factor clusters include subject attribute clusters, historical behavior clusters, risk transmission clusters, on-site constraint clusters, and feedback prior clusters. The event evolution chain constructed in the time dimension is used to characterize the stage relationship between the reporting, deployment, arrival, inspection, release, and feedback links.

4. The customs clearance trigger dispatch closed-loop method based on counterfactual causal learning according to claim 2, characterized in that: The expression for S201 is: , , , in, This represents the triggering causal field tensor of the u-th monitored object. The complete set of candidate causal units obtained by analysis Indicates the first The first among the regulatory targets One candidate causal unit, This represents the total number of candidate causal units resolved in the triggering causal field of the u-th monitored object. This represents the candidate set of triggering causes for the u-th regulated object. This represents the temporal precedence measure of the j-th candidate causal unit. This represents the perturbation sensitivity measure of the j-th candidate causal unit. This represents the structural saliency measure of the j-th candidate causal unit. , and Let the weights represent the fusion weights for temporal precedence, perturbation sensitivity, and structural saliency, respectively, satisfying: , This represents the threshold for determining the candidate set of the u-th regulatory object; The expression for S202 is: , , in, This represents the overall causal contribution strength of the j-th candidate causal unit among the u-th regulatory objects. This indicates the direct driving strength of the candidate causal unit. This indicates the conduction amplification intensity of the candidate causal unit. This indicates the net contribution strength of the candidate causal unit. , and The fusion weights for direct driving intensity, conduction amplification intensity, and net contribution intensity, respectively, satisfy: , , This represents the causal contribution screening threshold for the u-th regulated object. This represents the set of causal contribution factors retained after contribution screening for the u-th regulated object; The expression for S203 is: , , , in, This represents the set of causal contribution factors for the u-th regulated object. The set of principal clusters obtained by clustering according to business semantics and trigger link location This represents the b-th principal factor cluster in the u-th regulated object. This represents the number of principal clusters formed by the u-th regulated object. This represents the trigger causal profile vector of the u-th monitored object. This represents the focus weight of the b-th principal factor cluster. This represents the core principal factor encoding vector of the b-th principal factor cluster. This represents the auxiliary factor encoding vector of the b-th principal factor cluster. This represents the sorting structure encoding vector of the b-th principal cluster. This represents the set of triggering causes for the u-th monitored object. Indicates selection based on focus weight. Each principal factor cluster corresponds to the operation of the core principal factor. This represents the number of main factors output by the u-th monitored object. This represents the set of focus weights for all principal factor clusters of the u-th regulatory object.

5. The customs clearance trigger dispatch closed-loop method based on counterfactual causal learning according to claim 4, characterized in that, Step S3 includes the following steps: S301, for the u-th regulated object, its subject attributes, environmental constraints, key contexts, and current decision state are uniformly encoded to form a real-world base state. Decision variables that allow controlled substitution are extracted from the real-world base state, while keeping the subject attributes, environmental constraints, and key contexts unchanged. The output is the real-world base state vector used for twin generation and its corresponding set of substitutable decision variables, expressed as: , , in, This represents the real-world base state vector of the u-th monitored object. This represents the base state encoding function for the u-th monitored object. This represents the main attribute state vector of the u-th supervised object. This represents the environmental constraint state vector of the u-th regulated object. This represents the actual decision-making state vector of the u-th regulated object. Let u represent the set of substitutable decision variables for the u-th regulatory object. Let r be the r-th alternative decision variable in the u-th regulated object. This represents the number of substitutable decision variables for the u-th regulatory object; S302, decompose the real-world base state into an invariant part and a permissible substituent part. Around the decision variables, perform several controlled substitutions on the real-world decision state, constructing several virtual decision clones that run parallel to the real-world path. Recombine each virtual decision clone after controlled substitution with the invariant subject attribute state and environmental constraint state to form a family of counterfactual twin scenarios, expressed as: , , , in, This represents the counterfactual twin scenario family generated by the u-th regulated object. Indicates the first The wth counterfactual twin scenario among the regulated entities This represents the number of counterfactual twin scenarios generated for the u-th regulatory object. This represents the virtual decision state vector corresponding to the w-th counterfactual twin scenario. Let w represent the replacement rule vector for the w-th counterfactual twin scenario. This represents the controlled replacement function for the u-th supervised object; S303, for each counterfactual twin scenario, estimates its potential outcomes across different dimensions, performs parallel aggregation and normalization comparison of all scenario potential outcomes to form an effect vector for each scenario, and organizes all scenario effect vectors into an alternative strategy effect spectrum to characterize the comprehensive advantages and disadvantages of each alternative relative to the real path, expressed as: , , , , in, This represents the effect vector of the w-th counterfactual twin scenario in the u-th regulated object. This represents the scenario effect deduction function for the u-th monitored object. This represents the potential outcome of the w-th counterfactual twin scenario in terms of regulatory hit rate. This represents the potential outcome of the w-th counterfactual twin scenario in terms of release timeliness. This represents the potential outcome of the w-th counterfactual twin scenario in terms of resource consumption. This represents the potential outcome of the w-th counterfactual twin scenario in terms of congestion risk. This represents the potential outcome of the w-th counterfactual twin scenario in terms of the probability of repeated flow. , , , and These represent the weighting coefficients for five dimensions: regulatory hit rate, release timeliness, resource consumption, congestion risk, and probability of duplicate transfer, respectively, satisfying the following: , , This represents the spectrum of alternative strategies for the u-th regulated object. This represents the number of counterfactual twin scenarios generated for the u-th regulatory object.

6. The customs clearance-triggered order dispatch closed-loop method based on counterfactual causal learning according to claim 5, characterized in that: In step S303, the decision variables for controlled replacement include whether it is triggered, the triggering time, the dispatch target, and the inspection path; in step S303, the different dimensions include regulatory hit rate, release timeliness, resource consumption, congestion risk, and probability of repeated transfer.

7. The customs clearance trigger dispatch closed-loop method based on counterfactual causal learning according to claim 5, characterized in that, Step S4 includes the following steps: S401, focusing on the same regulatory object, abstracts the necessity of triggering, the object to be dispatched, the timing of dispatching, the priority, and the disposal path into a unified decision vector. It maps the intensity of the triggering cause, the effect of alternative scenarios, on-site resource constraints, and business time window constraints into a unified decision space, forming a joint decision space that includes decision candidates, constraint boundaries, and benefit-cost structures. The expression is: , , , in, This represents the joint decision space for triggering order dispatch for the u-th monitored object. Let z be the joint decision candidate vector among the u-th regulatory objects. This represents the number of joint decision-making candidates for the u-th regulatory object. This represents the trigger decision sub-vector in the z-th candidate vector. This represents the dispatch object subvector in the z-th candidate vector. This represents the time vector for dispatching orders in the z-th candidate vector. Let represent the priority subvector in the z-th candidate vector. This represents the sub-vector of the disposal path in the z-th candidate vector. Let the joint decision boundary vector of the u-th regulated object be denoted as . Represents the decision space mapping function for the u-th regulated object; S402, for each joint decision candidate vector, estimate its risk-reward, time-efficiency benefit, and resource cost. By constructing a resonance balance function among the three types of benefits and costs, measure the overall synergy between each candidate vector in terms of regulatory effectiveness, customs clearance efficiency, and resource burden. Select the joint decision candidate vector with the resonance balance value as the triggering dispatch scheme for the current regulatory object. The expression is: , , in, This represents the resonance equilibrium value of the z-th joint decision candidate vector among the u-th regulatory objects. Let z represent the risk-reward term of the z-th joint decision candidate vector. Let z represent the time-efficiency benefit term of the z-th joint decision candidate vector. Let z represent the resource cost term of the z-th joint decision candidate vector. , and Let these represent the balanced weights of risk-return, time-efficiency benefit, and resource cost, respectively, satisfying the following: , , This represents the joint decision vector obtained after resonant collaborative solution for the u-th regulatory object. This represents the number of joint decision-making candidates for the u-th regulatory object; S403 decodes the joint decision vector, restoring the trigger status, dispatch target, time parameters, priority parameters, and path parameters into executable business fields. Based on the business system interface specification, the decoding results are organized into structured joint instructions, outputting joint decision instructions that can be directly sent to the inspection and scheduling system, the field operation system, or the subsequent feedback system. The expression is: , , in, This represents the joint decision-making instruction vector for the u-th regulated object. This represents the instruction decoding function for the u-th monitored object. Represents the vector of trigger status fields. This represents a vector of target fields for dispatch orders. This represents a vector of fields indicating the order dispatch timing. Represents a priority field vector. Represents a vector of disposal path fields. This represents the structured joint decision-making instruction for the u-th regulated object. This represents the metadata vector associated with the joint decision-making instruction.

8. The customs clearance-triggered order dispatch closed-loop method based on counterfactual causal learning according to claim 7, characterized in that, Step S5 includes the following steps: S501, focusing on the feedback received regarding the inspection results, release results, transfer to manual processing results, time consumption results, anomaly handling results, and resource usage for the same regulated object, aligns various execution feedbacks with the original decision-making actions based on the regulated object identifier, trigger status field, dispatch target field, and processing path field in the joint decision-making instruction. This constructs a three-way correspondence between execution results, triggering causes, and dispatch paths, forming an execution evidence mapping that can be used for subsequent causal backtracking. The expression is: , , , in, This represents the set of execution feedback records corresponding to the u-th supervised object. This represents the nth execution feedback record in the uth supervised object. This represents the number of execution feedback records received by the u-th monitored object. This represents the set of execution evidence mappings for the u-th regulated object. This represents the evidence mapping unit constructed from the nth execution feedback record. This represents the evidence mapping function for the u-th supervised object; S502, based on the execution evidence mapping set, analyzes the causal correspondence between the verification hit results, release results, manual transfer results, time consumption results, anomaly handling results, and resource usage feedback and the triggering main cause and dispatch path. It estimates the verification strength of each triggering main cause, the execution deviation strength of each dispatch path, and the environmental mismatch strength of each strategy item. The verification strength, execution deviation strength, and environmental mismatch strength are jointly summarized into a calibration signal, used to characterize the direction of effectiveness correction of the original decision after actual execution. The expression is: , , in, This represents the calibration strength of the h-th decision element within the u-th regulated object. This represents the total number of decision-making factors that require calibration assessment for the u-th regulated object. This represents the verification strength of the h-th decision element. This represents the intensity of the execution deviation of the h-th decision element. This represents the intensity of environmental mismatch for the h-th decision element. , and The fusion weights for validation strength, execution bias strength, and environment mismatch strength, respectively, satisfy: , , Represents the set of calibration strengths for the u-th monitored object; S503 decomposes the calibration intensity set into three levels: triggering cause, dispatch object, and handling path, forming the cause weight correction, dispatch weight correction, and path preference correction respectively. These corrections are written into the corresponding parameter sets to update the original trigger weight, dispatch weight, and path selection preference, outputting a new closed-loop evolution parameter set for direct use by subsequent monitored objects. The expression is: , , , in, This represents the triggering cause weight vector of the u-th monitored object before the update. This represents the updated trigger cause weight vector for the u-th monitored object. This represents the dispatch weight vector of the u-th monitored object before the update. This represents the updated dispatch weight vector for the u-th monitored object. This represents the path selection preference vector of the u-th monitored object before the update. This represents the updated path selection preference vector for the u-th monitored object. This represents the modified mapping function applied to the triggering factor weight vector. This represents the correction mapping function applied to the dispatch weight vector. This represents the correction mapping function acting on the path selection preference vector. This represents the closed-loop evolution update parameter set for the u-th monitored object.

Images