Express package safety management system and method based on artificial intelligence technology

Abstract

The application relates to an express package safety management system and method based on artificial intelligence technology, the system comprising a data acquisition layer, an intelligent analysis layer, a decision execution layer and a counter-feeding optimization layer; the data acquisition layer collects multi-source heterogeneous data of a to-be-identified package and binds the data into associated data; the intelligent analysis layer comprises at least two risk research and judgment modules arranged in parallel, each module taking different dimension data as input and output to correspond to a risk research and judgment result; the decision execution layer executes comprehensive decision on the risk research and judgment results and issues a hierarchical disposal instruction; the counter-feeding optimization layer returns the disposal result and the identification process data to a sample library to iteratively update the risk research and judgment module. The application has the advantages that multi-dimensional risk identification is mutually verified, disposal decision is hierarchical and unified, and the model can evolve with new risks.

Description

Technical Field

[0001] This application relates to the field of express delivery sorting, and in particular to an express parcel security management system and related devices based on artificial intelligence technology. Background Technology

[0002] With the continuous growth of the express delivery industry's business volume, the security risks faced during the parcel circulation process are becoming more diversified and dynamically evolving, placing higher demands on the express parcel security management system, moving from single-point identification to full-chain control. The security risks of express parcels are not singular; they simultaneously involve compliance risks of the parcel's contents, consistency risks between the declared parcel information and its actual physical attributes, and risks related to the identity of the parcel's delivery entity and the entities associated with the circulation chain. These risk dimensions differ in their identification criteria, judgment methods, and required analytical techniques, and there is a possibility of cross-verification between them.

[0003] Existing parcel security management systems are typically composed of several isolated identification devices. For example, a device for detecting the compliance of parcel contents operates independently, only recognizing image data; a device for verifying weight consistency operates independently, only collecting and sampling weight data for comparison; and the verification process for the delivery entity relies on manual spot checks or post-event traceability. This fragmented architecture first and foremost artificially fragments the data foundation for the identification process, meaning each identification method can only rely on data from a single dimension for judgment, unable to cross-verify with data from other dimensions. The inherent blind spots of single-dimensional identification are difficult to compensate for through internal mechanisms within the solution.

[0004] The fragmentation of the identification process further exacerbates the lack of focus in the decision-making process. When multiple identification methods output results independently, the system lacks a mechanism to aggregate these results into a unified handling basis. Each identification device can only issue handling instructions independently based on its own output. This leads to repeated interception or release of the same package, and because any missed detection in any dimension can cause the package to be misjudged as a safe item in that dimension, the overall robustness of control is constrained by the weakest dimension. The decision-making process cannot reflect the collaborative assessment between multiple dimensions, nor can it apply differentiated handling intensity to packages according to the overall severity of the risk.

[0005] The separation between decision-making and identification further hinders the reverse correction of the identification model based on the handling results. The actual handling results of packages and the intermediate data left during the identification process are crucial for identifying identification deviations and updating the judgment model. However, in an architecture where each device operates in isolation, the handling results remain at the execution end, and the identification process data remains at the identification end, unable to be uniformly collected and used to update training parameters. When new package disguise methods or new delivery behavior patterns emerge, the system's identification capabilities cannot evolve accordingly. Over time, identification accuracy gradually declines, the investment in manual verification continues to increase, and the gap between control efficiency and the speed of industry development widens.

[0006] Therefore, the disconnect between the aforementioned identification, decision-making, and feedback stages stems from the same systemic problem: the lack of an architecture that can unify multi-dimensional risk assessment, comprehensive decision-making, and feedback through a closed-loop iterative process. Thus, an AI-based express parcel security management system is needed to perform parallel multi-dimensional risk assessment, comprehensive decision-making based on the assessment results, and feedback updates to the assessment model using the handling results, all within a single architecture. This ensures that the identification capabilities of each dimension complement each other, that the assessment results from each dimension jointly support tiered handling, and that the system possesses the ability to continuously evolve in response to emerging risks. Summary of the Invention

[0007] To address the problems in existing express parcel security management systems where the identification, decision-making, and feedback processes are fragmented, resulting in the inability to corroborate multidimensional risks, the lack of unified classification of handling decisions, and the inability of models to evolve autonomously with new risks, this application provides an express parcel security management system and related devices based on artificial intelligence technology.

[0008] Firstly, this application provides a parcel security management system based on artificial intelligence technology, which adopts the following technical solution: An artificial intelligence-based express parcel security management system includes a data acquisition layer, an intelligent analysis layer, a decision execution layer, and a feedback optimization layer. The data acquisition layer is configured to collect multi-source heterogeneous data of the parcels to be identified and perform association binding on the multi-source heterogeneous data to obtain associated data corresponding to the parcels to be identified. The intelligent analysis layer includes at least two parallel risk assessment modules, each configured to select at least one type of data from the associated data as input and output the corresponding risk assessment result. The input data types or assessment dimensions of different risk assessment modules are different. The decision execution layer is configured to perform a comprehensive decision based on the risk assessment results output by the at least two risk assessment modules, determine the handling level of the parcels to be identified, and issue corresponding handling instructions according to the handling level. The feedback optimization layer is configured to feed back the handling results from the decision execution layer and the identification process data from the intelligent analysis layer to a sample database, and perform iterative updates on at least one of the at least two risk assessment modules based on the sample database.

[0009] By adopting the above technical solution, multiple risk identification methods that originally operated independently on their respective devices are unified into an intelligent analysis layer with the same architecture. This allows each risk assessment module to derive assessment results from different dimensions based on common related data. The decision execution layer then makes a comprehensive decision based on the multi-dimensional assessment results and issues a unified tiered handling instruction, avoiding duplicate handling caused by separate interception by multiple dimensions or overall failure due to missed detection in any dimension. At the same time, the handling results of the decision execution layer and the identification process data of the intelligent analysis layer are fed back to the sample library through the optimization layer and drive the iterative update of the risk assessment modules. This enables the system to transform handling feedback into the continuous evolution of assessment capabilities, thus establishing a closed-loop control architecture that links identification, decision-making, and feedback.

[0010] Optionally, at least two risk assessment modules include a prohibited items identification module, a weight deviation control module, and a fraud risk assessment module. The prohibited items identification module is configured to perform multimodal fusion on image data and waybill text data in the associated data and output the prohibited items identification result. The weight deviation control module is configured to calculate the predicted weight range of the package to be identified based on the image data in the associated data and determine the weight risk level based on the deviation between the predicted weight range and the measured weight. The fraud risk assessment module is configured to determine the comprehensive fraud risk score based on sender information, order information, and branch information in the associated data. The prohibited items identification result, weight risk level, and comprehensive fraud risk score are all used as inputs for comprehensive decision-making.

[0011] By adopting the above technical solution, the three risk dimensions are implemented as independent yet unified judgment modules for the same comprehensive decision-making process. This allows the three heterogeneous judgment results of prohibited items compliance, weight consistency, and delivery entity risk to be considered by the same decision-making mechanism, forming a triple complementary judgment capability covering package contents, physical attributes, and delivery entity.

[0012] Optionally, the contraband identification module includes a visual feature extraction submodule, a text feature extraction submodule, and a cross-modal fusion submodule. The visual feature extraction submodule includes a first visual branch and a second visual branch set in parallel. The first visual branch is configured to extract global structural features of the image data, and the second visual branch is configured to extract local texture features of the image data. The text feature extraction submodule is configured to perform bidirectional sequence encoding and intra-sequence association weighting on the waybill text data to output text semantic features. The cross-modal fusion submodule is configured to generate text-guided visual features and visually guided text features based on the similarity between global structural features, local texture features, and text semantic features, and to perform fusion on the text-guided visual features and visually guided text features to output the contraband identification result.

[0013] By adopting the above technical solution, the visual branch extracts features at both global and local granularities in parallel, making up for the problem of the uneven recognition ability of a single visual network for large and small contraband items; the cross-modal fusion submodule aligns visual features and textual semantic features in both directions, so that the image and the waybill text can corroborate each other, improving the robustness of recognizing disguised packages.

[0014] Optionally, the weight deviation control module includes an appearance feature extraction submodule, a weight prediction submodule, and a weight risk determination submodule. The appearance feature extraction submodule is configured to extract the size features, shape category features, and texture category features of the package to be identified from the image data. The weight prediction submodule is configured to output a predicted weight range based on the size features, shape category features, and texture category features through a cascaded structure of a visual neural network and a regression model. The weight risk determination submodule is configured to classify the weight risk level into three levels—high risk, medium risk, and low risk—that are sequentially adjacent and do not overlap, based on the relative deviation between the measured weight and the predicted weight range.

[0015] By adopting the above technical solution, the reasonable weight range of the package can be calculated based on the visual characteristics of the package, so that the weight verification is transformed from a post-event sampling inspection to an online verification of each package. The three-level deviation range supports graded handling, so that the response intensity to the weight deviation can match the severity of the deviation.

[0016] Optionally, the fraud risk assessment module includes a user risk scoring submodule, an order risk scoring submodule, a branch risk scoring submodule, and a comprehensive scoring submodule. The user risk scoring submodule is configured to output a user risk score based on sender information; the order risk scoring submodule is configured to output an order risk score based on order information; the branch risk scoring submodule is configured to output a branch risk score based on branch information; and the comprehensive scoring submodule is configured to output a comprehensive fraud risk score based on at least the user risk score, order risk score, and branch risk score.

[0017] By adopting the above technical solution, the risk of fraud is broken down from a single subject dimension into three independent scoring lines: user, order, and outlet, and then incorporated into the same comprehensive score. This allows the subject profile, single delivery behavior, and the operational status of the outlets involved in the same delivery behavior to be evaluated simultaneously, avoiding overall omissions caused by omissions in a single dimension.

[0018] Optionally, the fraud risk assessment module also includes at least one of a risk cluster mining submodule and a multimodal association submodule. The risk cluster mining submodule is configured to calculate the inter-sample similarity between samples to be clustered using mixed distance, and perform clustering based on sample distribution density based on the inter-sample similarity to mine at least one of high-risk customer clusters and high-risk branch clusters, and feed the mining results back to the comprehensive scoring submodule. The multimodal association submodule is configured to aggregate unstructured data and association graph data in the association data into supplementary features, and use the supplementary features as additional input features of the comprehensive scoring submodule, which together with the user risk score, order risk score and branch risk score participate in the calculation of the comprehensive fraud risk score.

[0019] By adopting the above technical solutions, risk cluster mining aggregates isolated high-risk individuals into group profiles, enabling the system to identify gang-related fraudulent activities; the multimodal association submodule introduces supplementary features from image and graph dimensions, so that the comprehensive score is not limited to the linear superposition of structured scores, but can integrate implicit risk signals from heterogeneous modalities.

[0020] Optionally, the decision execution layer includes a hierarchical decision submodule and a disposal execution submodule. The hierarchical decision submodule is configured to match the risk assessment results output by each risk assessment module with preset multi-level risk thresholds. In response to any one of the highest-level risk thresholds being hit, the disposal level is determined to be automatic interception level and an automatic interception instruction is generated. In response to at least one intermediate-level risk threshold being hit and none of the highest-level risk thresholds being hit, the disposal level is determined to be manual review level and a manual review task is generated. In response to all of the lowest-level risk thresholds being hit, the disposal level is determined to be normal flow level and a normal flow instruction is generated. The disposal execution submodule is configured to send the automatic interception instruction, the manual review task, or the normal flow instruction to the corresponding execution terminal, respectively.

[0021] By adopting the above technical solutions, the parallel decision-making logic of "intercepting any package that hits the highest level" ensures the priority handling of the most critical packages, while the hierarchical mechanism of "reviewing packages that partially hit the intermediate level" accurately allocates human resources to packages that are truly questionable, avoiding the one-size-fits-all approach caused by single-dimensional thresholds.

[0022] Optionally, the feedback optimization layer includes at least one of a sample management submodule, a model iteration submodule, and a collaborative output submodule; the sample management submodule is configured to archive missed cases, misjudged cases, and detected cases into a sample library according to a preset classification; the model iteration submodule is configured to extract samples from the sample library and divide them into training sets, validation sets, and test sets, retrain at least one of at least two risk assessment modules, and deploy the retrained risk assessment module to the intelligent analysis layer in a canary release manner; the collaborative output submodule is configured to synchronize the de-identified identification results and related data to an external system through a standardized interface.

[0023] By adopting the above technical solution, the three types of boundary samples—missed detection, false judgment, and detected—are continuously archived as training inputs for the judgment module, enabling the model to evolve with new risks. The gray-scale release mechanism ensures that the risk judgment module after retraining has a rollback-friendly stability verification space before being deployed to the intelligent analysis layer.

[0024] Optionally, the data acquisition layer includes a set of sensing devices and an association binding submodule. The set of sensing devices includes at least two of the following: image acquisition devices, weighing devices, barcode scanning devices, and business system interfaces. These devices are configured to collect at least two of the following types of data from the package to be identified: image data, weight data, identification data, and business data. The identification data includes at least waybill text data, and the business data includes at least sender information, order information, and branch information. The image data, weight data, identification data, and business data together constitute multi-source heterogeneous data. The association binding submodule is configured to perform verification on the multi-source heterogeneous data and bind the verified multi-source heterogeneous data to the unique identifier of the package to be identified, thereby obtaining the associated data.

[0025] By adopting the above technical solution, heterogeneous data scattered across different sensing devices and business systems are bound together with the unique identifier of the package to be identified as the anchor point, so that all subsequent risk assessment modules can run on the same data base, avoiding assessment bias caused by data misalignment.

[0026] Optionally, the image data includes X-ray image data, which includes top view image data and side view image data of the package to be identified. The visual feature extraction submodule constructs a three-dimensional structural representation of the package to be identified based on the top view image data and side view image data, and then extracts global structural features and local texture features.

[0027] By adopting the above technical solution, dual-view image fusion is used to obtain a three-dimensional structural representation of the package, which alleviates the loss of visual features caused by the stacking and occlusion of items inside the package.

[0028] Optionally, the bidirectional sequence encoding in the text feature extraction submodule is implemented by a bidirectional gated recurrent unit or a bidirectional long short-term memory network, the intra-sequence association weighting is implemented by a self-attention mechanism, and the text feature extraction submodule loads a special dictionary for express delivery waybill text as prior knowledge.

[0029] By adopting the above technical solutions, sequence encoding and associated weighting are implemented by mature structures in the field, and the parsing accuracy of waybill abbreviations, misspellings and industry terms is improved by combining industry-specific dictionaries.

[0030] Optionally, the first visual branch includes a network based on a visual Transformer structure, and the second visual branch includes a convolutional neural network based on a residual structure.

[0031] By adopting the above technical solution, two visual networks with different inductive biases run in parallel, respectively undertaking large-scale shape recognition and small-scale texture recognition, so that the recognition advantages of the two branches complement each other.

[0032] Optionally, the fusion strategy of the cross-modal fusion submodule includes two levels: feature-level fusion and decision-level fusion. Feature-level fusion concatenates the text-guided visual features and the visual-guided text features to obtain multimodal fusion features. Decision-level fusion weights and synthesizes the first classification result generated by the multimodal fusion features with the auxiliary classification results generated by each single modal feature to obtain the contraband identification result.

[0033] By adopting the above technical solution, a two-level fusion of feature level and decision level is introduced, so that multimodal collaboration is reflected not only in the alignment at the feature level, but also in the voting at the output level, thereby enhancing the robustness of the recognition results to abnormal modalities.

[0034] Optionally, the shape category feature includes at least one of cuboid, spherical, sheet-like, and irregular shapes, and the texture category feature includes at least one of cardboard box, eco-friendly bag, and other categories.

[0035] By adopting the above technical solution, the visual appearance of packages can be classified into a limited number of categories, enabling the regression model to model in segments according to categories and reducing the prediction error caused by cross-category sample mixing training.

[0036] Optionally, the weight risk determination submodule maps the relative deviation to a weight deviation ratio, and classifies the weight deviation ratio into high risk, medium risk and low risk respectively when it is located in the first preset interval, the second preset interval and the third preset interval. The first preset interval, the second preset interval and the third preset interval are adjacent to each other and do not overlap.

[0037] By adopting the above technical solution, the relative deviation is normalized in the form of a ratio, so that packages with different absolute weights can be judged on the same scale, and the connection relationship of the three intervals avoids judgment conflicts caused by the overlap of adjacent intervals.

[0038] Optionally, the weight prediction submodule trains an independent prediction model for each shape category. When performing weight prediction on the package to be identified, the corresponding prediction model is first selected based on the shape category features, and then the selected prediction model outputs the predicted weight range based on the size features and texture category features.

[0039] By adopting the above technical solution, the prediction model trained by shape makes the feature space of each shape category independent, which alleviates the phenomenon that the parameters of categories with small sample sizes, such as spherical, sheet-like, and irregular shapes, are squeezed out by the mainstream categories.

[0040] Optionally, the input indicators for user risk scoring may include at least the sender's identity authenticity indicator, delivery behavior stability indicator, contact information validity indicator, and historical compliance record indicator.

[0041] By adopting the above technical solutions, a comprehensive profile of the sender is created from four dimensions: identity, behavior, contact information, and historical compliance, avoiding scoring distortion caused by the falsification or circumvention of a single indicator.

[0042] Optionally, the input indicators for order risk scoring may include at least the bulk delivery characteristics of the order, the risk characteristics of the goods, and the delivery method characteristics.

[0043] By adopting the above technical solutions, abnormal signals in terms of quantity, product attributes, and delivery mode of a single delivery behavior can be captured, thus compensating for the risks of a single behavior that cannot be captured from the user perspective.

[0044] Optionally, the input indicators for the risk score of the outlet should include at least the indicators for implementing real-name registration for mail and parcels, the pass rate of mail and parcel inspection, and the indicators for reporting suspicious clues.

[0045] By adopting the above technical solutions, the compliance performance of the network layer is incorporated into the assessment, extending the scoring chain from both ends of "sender-order" to the "delivery location" side, covering the entire risk source.

[0046] Optionally, the mixed distance is a weighted sum of the Euclidean distance component, the Jaccard distance component, and the cosine distance component. The Jaccard distance is used to measure the difference in categorical indicators among the samples to be clustered.

[0047] By adopting the above technical solution, numerical, categorical and directional distance components are introduced simultaneously in the measurement of similarity between samples, so that clustering can reflect the comprehensive dissimilarity under the combination of heterogeneous indicators.

[0048] Optionally, the supplementary features of the multimodal association submodule include image feature components extracted from image data via a convolutional neural network, and graph feature components aggregated from the association graph composed of sender information, recipient information, and branch information via a graph neural network.

[0049] By adopting the above technical solution, the two types of heterogeneous data, unstructured images and associated maps, are aggregated through their respective adapted networks and uniformly converted into supplementary features in vector form, which facilitates unified processing by the comprehensive scoring submodule.

[0050] Optionally, the sample management submodule also constructs a fraud-related feature knowledge base, which includes a basic database containing fraud-related features clearly defined by laws and regulations, and an incremental database supplemented by actual cases at preset intervals. The sample management submodule merges the incremental database into the basic database at preset intervals.

[0051] By adopting the above technical solution, static compliance features and dynamic incremental features are organized in a two-layer structure, which ensures both the stability and timeliness of the knowledge base and avoids the long-term absence of new fraud-related patterns from the training process.

[0052] Optionally, the model iteration submodule divides the training set, validation set, and test set according to a preset ratio of 7:2:1.

[0053] By adopting the above technical solution, a set of commonly used sample division ratios are given as a reference implementation method, which facilitates the direct implementation of the model iteration process.

[0054] Optionally, the canary release of the model iteration sub-module further includes: first, deploying the retrained risk assessment module to some execution nodes; in response to the fact that the operating indicators of some execution nodes meet the preset stability conditions within the preset observation period, deploying the retrained risk assessment module to all execution nodes; and in response to the fact that the operating indicators of some execution nodes do not meet the preset stability conditions within the preset observation period, reverting some execution nodes to the risk assessment module version before retraining.

[0055] By adopting the above technical solution, the gray release includes both forward promotion conditions and backward rollback conditions, which allows abnormal training versions to be withdrawn within the observation window, thus preventing defective models from affecting the entire network.

[0056] Optionally, the collaborative output submodule has a one-click synchronization mechanism and a three-level linkage mechanism. The one-click synchronization mechanism enables the collaborative output submodule to synchronize the associated data with the comprehensive risk score of fraud exceeding the preset threshold to the external system in a single operation. The three-level linkage mechanism enables the synchronized data to reach the headquarters level, local level and grassroots level three docking points in the external system respectively.

[0057] By adopting the above technical solution, the transfer of fraud-related clues is integrated from multiple independent operations into a single multi-terminal synchronous operation, improving the efficiency of collaboration with external systems.

[0058] Optionally, the collaborative output submodule is also configured to synchronize the de-identified recognition results and associated data to the industry collaboration platform through the industry sharing interface for use by other similar systems.

[0059] By adopting the above technical solutions, cross-entity industry-level collaboration is introduced in addition to internal enterprise handling, enabling the risk intelligence accumulated by a single enterprise to support joint defense across the entire industry.

[0060] Optionally, the feedback optimization layer is also configured to perform reverse diffusion tracing on the delivery chain associated with the seized case based on the logistics trajectory and association graph in the seized case, and output the tracing results including the upstream associated entities of the delivery chain.

[0061] By adopting the above technical solutions, starting from the cases uncovered, we can trace the related entities upstream and provide link-level source tracing analysis support for external systems such as public security.

[0062] Optionally, the feedback optimization layer is also configured to construct pseudo samples based on the sample library. The pseudo samples include pseudo image data generated from text semantic features through cross-modal generation and pseudo text data generated from global structural features and local texture features of image data through cross-modal generation. The pseudo samples are used to supplement the training samples of the target category in the sample library where the number of samples is lower than a preset sample threshold.

[0063] By adopting the above technical solution, the problem of difficulty in collecting real samples for rare contraband or rare fraud patterns is addressed by expanding the training samples of the target category through cross-modal generation, thereby alleviating the insufficient recognition ability of long-tail categories.

[0064] Optionally, the edge analysis sublayer of the system is deployed on the sensing device side. The edge analysis sublayer carries a lightweight version of the risk assessment module. The lightweight version is constructed using at least one of model compression, depthwise separable convolution, and knowledge distillation. Its inference output is used to perform preliminary screening of the packages to be identified on the edge side. Only the associated data of the packages to be identified that are suspected in the preliminary screening results are uploaded to the cloud analysis sublayer. The cloud analysis sublayer carries the full version of the risk assessment module.

[0065] By adopting the above technical solution, the vast majority of low-risk packages are initially screened and eliminated at the edge before suspicious data is sent to the cloud, which greatly reduces the cloud computing pressure and link bandwidth usage, while taking into account both real-time performance and in-depth analysis capabilities.

[0066] Optionally, the system is also configured to perform robust processing for at least one of the following abnormal scenarios: missing waybill information, poor image data quality, or irregular package shape. This includes supplementing missing data through cross-modal generation and enhancing the risk assessment module's adaptability to low-quality data through data augmentation.

[0067] By adopting the above technical solutions, the ability of the judgment module to handle boundary data is improved through two paths: cross-modal completion and data augmentation, targeting the unavoidable abnormal samples in actual collection scenarios.

[0068] Optionally, the system can be configured with encryption for data transmission links, encryption for data storage layers, role-based access control for data access interfaces, and desensitization processing for personal information in associated data.

[0069] By adopting the above technical solutions, security measures are set up simultaneously in the three stages of transmission, storage, and access, so that the system meets the compliance requirements for personal information protection.

[0070] Optionally, the system is configured with a multi-node disaster recovery architecture, including at least two sets of computing nodes and data nodes that serve as backups for each other. In response to a failure of the primary node, business traffic is switched to the backup node.

[0071] By adopting the above technical solutions, the impact of system failures on online services can be constrained within a controllable range through node redundancy, thereby improving overall availability.

[0072] Secondly, this application provides a method for safe management of express parcels based on artificial intelligence technology, which adopts the following technical solution: A method for secure management of express parcels based on artificial intelligence technology includes the following steps: S1. Collect multi-source heterogeneous data of the package to be identified, and perform association binding on the multi-source heterogeneous data to obtain the associated data corresponding to the package to be identified; S2. Perform at least two parallel risk assessments on the associated data. Each risk assessment selects at least one type of data from the associated data as input and outputs the corresponding risk assessment result. The input data types or assessment dimensions of different risk assessments are different. S3. Make a comprehensive decision based on the risk assessment results of at least two risk assessments, determine the disposal level of the package to be identified, and perform corresponding disposal operations on the package to be identified according to the disposal level to obtain the disposal result; S4. Feedback the disposal results and process data of at least two risk assessments back to the sample library, and perform iterative updates on at least one of the corresponding models of the at least two risk assessments based on the sample library.

[0073] By adopting the above technical solution, the data processing flow relationship during system operation is presented in a step-by-step manner, which enables implementers to deploy it either as a device or incrementally in the existing logistics business system. The data flow from S1 to S4 constitutes a complete risk identification-disposal-feedback closed loop.

[0074] In summary, this application includes at least one of the following beneficial technical effects: 1. By integrating multi-dimensional risk assessment, comprehensive decision-making, and feedback iteration into the same architecture, the identification of prohibited items, weight deviation prevention, and fraud risk assessment, which were originally distributed on various independent devices, can share the same related data base. After the multi-dimensional assessment results are mutually verified, the same decision-making mechanism will uniformly issue graded disposal instructions, avoiding the overall control failure caused by single-dimensional missed detection or repeated interception in various dimensions.

[0075] 2. Through advanced mechanisms such as cross-modal feature alignment, visual weight estimation, three-dimensional risk scoring, and cluster mining, each risk assessment module has the ability to complement multiple sources and fuse heterogeneous indicators when running independently. With the addition of hierarchical threshold decision-making and desensitized collaborative output, the system is simultaneously enhanced in terms of assessment accuracy, handling granularity, and cross-subject collaboration.

[0076] 3. Through a model iteration mechanism that combines continuous archiving of bad samples, cross-modal generation to expand rare samples, and gray-scale release with reverse rollback, the system can continuously evolve itself when faced with new methods of disguising contraband and new fraudulent behavior patterns. It also has rollback space when abnormal training versions appear, overcoming the problems of lagging model updates and difficulty in rolling back once released in traditional solutions. Attached Figure Description

[0077] Figure 1 This paper illustrates the overall architecture of an express parcel security management system based on artificial intelligence technology in one embodiment of this application. Detailed Implementation

[0078] The present application will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the application and are not intended to limit the scope of the application.

[0079] This application discloses a parcel security management system and related devices based on artificial intelligence technology.

[0080] To facilitate understanding, several core terms are explained first. A package to be identified refers to a physical package that enters the express delivery network and is identified, decided upon, and handled by this system. In this application, its unique business identifier serves as the anchor point at the data level. Multi-source heterogeneous data refers to the data set collected by different types of sensing devices and business systems during the circulation of the same package to be identified. The data in this set differ in data type, data format, collection frequency, and semantic space. Examples include penetrating content imaging data, appearance morphology data, quality data, waybill identification data, and delivery entity data from business systems. A risk assessment module refers to several algorithm modules that can be deployed in parallel within the intelligent analysis layer in this application embodiment. Each risk assessment module selects one or more types of data from the associated data as input and outputs assessment results around a specific risk dimension, with the dimensions being independent of each other.

[0081] To support the subsequent numerical examples, this application establishes a baseline scenario throughout. In this baseline scenario, the package to be identified is a paper package with an external shape approximately rectangular, measuring 40cm × 30cm × 20cm, with a measured weight of 2.5kg. The waybill describes the goods as "daily necessities," and the receiving and dispatching point is a branch in a provincial capital city, with a daily business volume at a medium level for the province. This system is deployed on a typical express sorting center, with the sorting line carrying the package to be identified through the sensing device area at a speed of 2.3m / s. The business system interface connects to the logistics business system of the express company. In subsequent embodiments of this application, the measured weight, waybill text, external dimensions, and delivery entity information, unless otherwise stated, are all based on values ​​from this baseline scenario.

[0082] This system unifies various risk identification methods, previously distributed across different devices and business processes, into a single architecture. The data acquisition layer provides a unified, interconnected data foundation for the entire system; the intelligent analysis layer runs multiple risk assessment modules in parallel; the decision execution layer performs comprehensive decisions based on the multi-dimensional assessment results; and the feedback optimization layer transforms the handling feedback into the continuous evolution of the assessment model. Based on this architecture, a closed loop is formed between the identification, decision-making, and feedback stages, avoiding the blind spots in identification, decision-making defocus, and feedback interruptions that often occur with traditional systems composed of separate devices.

[0083] Reference Figure 1 The following section will first provide a detailed explanation of the data acquisition layer, which serves as the starting point for system data, and then proceed layer by layer in the order of identification, decision-making, and feedback.

[0084] The data acquisition layer is deployed at business nodes such as express pickup points, sorting centers, and transit hubs, and is the data starting point for the entire system. The data acquisition layer includes a collection of sensing devices and associated binding sub-modules.

[0085] The sensing device set includes at least two of the following: image acquisition devices, weighing devices, barcode scanning devices, and business system interfaces. Image acquisition devices are used to acquire image data of the package to be identified. For example, the image acquisition device may include an X-ray scanner, which acquires a penetrating image of the contents through the outer packaging of the package; the image acquisition device may also include a visible light camera, which acquires visible light image data of the exterior of the package. Weighing devices are used to acquire weight data of the package to be identified, which is the measured weight of the package. For example, the weighing device may include a dynamic weighing system, which weighs the package as it moves along the sorting line, avoiding sorting interruptions. Barcode scanning devices are used to acquire identification data of the package to be identified, which includes at least waybill text data. For example, the barcode scanning device may include a six-sided barcode scanner, which scans the waybill from six directions of the package to address situations where the waybill's orientation is uncertain; the obtained waybill text data includes text fields such as the waybill number, goods description, and the sender's and recipient's names and contact information. The business system interface is used to obtain business data for packages to be identified. This business data includes at least sender information, order information, and branch information. For example, the business system interface connects to the logistics business system of express delivery companies, the order system of e-commerce platforms, and the branch management system to retrieve sender real-name registration information, order completion information, and branch operation information.

[0086] In the baseline scenario, the image acquisition device captures X-ray image data of the package to be identified; the weighing device outputs the measured weight of 2.5kg; the barcode scanning device outputs waybill text data, with the goods description field set to "daily necessities"; the business system interface outputs sender information, order information, and branch information. Sender information includes the sender's ID number, contact information, and historical delivery record number; order information includes the order number, order completion time, and the total number of orders placed by the same sender in the last 24 hours; and branch information includes the branch code, the acceptance and inspection pass rate for the last three months, and the number of suspicious leads reported. These various types of data are collected and output by different sensing devices and business systems, forming multi-source heterogeneous data.

[0087] Image data, weight data, identification data, and business data together constitute multi-source heterogeneous data. The outputs of image acquisition devices, weighing devices, and barcode scanning devices correspond to each other according to the timestamp at the time of acquisition; the business data returned by the business system interface corresponds to the waybill text data output by the barcode scanning device, indexed by the waybill number. Multi-source heterogeneous data does not directly enter the intelligent analysis layer, but is first processed by the association and binding submodule to ensure a one-to-one correspondence between various types of data and the packages to be identified.

[0088] The association and binding submodule performs verification on multi-source heterogeneous data, binding the verified multi-source heterogeneous data to the unique identifier of the package to be identified, thus obtaining associated data. The general verification methods include cross-referencing the timestamps of the data output by each sensing device, comparing the waybill number output by the barcode scanner with the waybill number returned by the business system interface, and comparing the number of packages to be identified entering the sensing device area within the same time window with the number of outputs from each sensing device. When all the above comparisons are consistent, the association and binding submodule considers the verification successful; when any comparison is inconsistent, the association and binding submodule considers the verification unsuccessful and marks the unsuccessful multi-source heterogeneous data as "data to be reviewed," excluding it from subsequent risk assessment. In the baseline scenario, the image acquisition device... Output X-ray image data, the barcode scanning device at any time Output waybill text data, weighing equipment at time Output weight data; the business system interface at any time Return the business data corresponding to the waybill number; the timestamp difference of the four types of data are all within the time window preset by the association binding submodule, the waybill number field is consistent, the number of packages entering the sensing device area per unit time is 1, the association binding submodule determines that the verification is passed, and binds the four types of data to the unique identifier of the package to be identified, thus obtaining the associated data.

[0089] In some embodiments, the image data includes X-ray image data, which includes top-view and side-view image data of the package to be identified. The image acquisition device acquires top-view and side-view image data along two different directions of the package to be identified, so that the stacking and obstruction of items inside the package can be distinguished from at least one viewpoint. This dual-view image data format is a lower-level implementation of image data.

[0090] As another subordinate implementation of image data, in some embodiments, the image acquisition device can be a millimeter-wave imaging device or a terahertz imaging device instead of an X-ray scanner. Millimeter-wave and terahertz imaging devices can also penetrate the outer packaging of the package to be identified and obtain penetrating images of the contents, making them semantically equivalent to X-ray image data and also examples of image data in the embodiments of this application. While the above alternatives differ from X-ray scanners in image resolution and penetration capability to specific materials, they all fall within the scope of image data.

[0091] This system comprises a data acquisition layer, an intelligent analysis layer, a decision execution layer, and a feedback optimization layer. The data acquisition layer provides relevant data to the intelligent analysis layer; the intelligent analysis layer provides risk assessment results from various dimensions to the decision execution layer; the decision execution layer provides the response results to the feedback optimization layer; and the feedback optimization layer sends the retrained risk assessment module back to the intelligent analysis layer. These four layers, starting with relevant data and ending with the iteratively updated risk assessment module, form a closed data loop that runs through identification, decision-making, and feedback.

[0092] The intelligent analysis layer comprises at least two risk assessment modules configured in parallel. "Parallelism" means that different risk assessment modules execute independently of each other, without any temporal dependency where a later module must wait for the output of a previous one. In engineering implementation, parallel configuration can be achieved by allocating independent computing processes to different risk assessment modules within the same computing node, or by deploying different risk assessment modules on different computing nodes and having a scheduler concurrently issue inference requests. Each risk assessment module selects at least one type of data from the associated data as input. Different risk assessment modules select different data types or assessment dimensions. One module focuses on the compliance of the package contents, another on the consistency between the package declaration and its physical attributes, and yet another on the risk of the delivery entity and delivery behavior. It is precisely because the assessment dimensions do not overlap that the outputs of each risk assessment module have complementary value in the decision-making process.

[0093] In this embodiment, at least two risk assessment modules include a prohibited items identification module, a weight deviation control module, and a fraud risk assessment module. The prohibited items identification module takes image data and waybill text data from the associated data as input, performs multimodal fusion, and outputs the prohibited items identification result. The weight deviation control module takes image data from the associated data as input, calculates the predicted weight range of the package to be identified, and combines it with the measured weight carried by the weight data in the associated data to output a weight risk level. The fraud risk assessment module takes sender information, order information, and branch information from the associated data as input and outputs a comprehensive fraud risk score. The prohibited items identification result, weight risk level, and comprehensive fraud risk score are generated in parallel and used together as input for the comprehensive decision-making of the decision execution layer.

[0094] In the baseline scenario, three risk assessment modules operate in parallel. The contraband identification module receives X-ray image data and waybill text data, completes feature extraction and cross-modal fusion within milliseconds, and outputs the contraband identification result. The weight deviation control module receives X-ray image data and calculates and predicts the weight range based on appearance features, combining the measured weight of 2.5kg to output the weight risk level. The fraud risk assessment module receives sender information, order information, and branch information, calculates three scoring lines for users, orders, and branches respectively, and then merges them into a comprehensive fraud risk score. The three assessment results are fed into the decision execution layer within the same decision-making cycle.

[0095] It should be noted that the three risk assessment modules are a typical combination of at least two risk assessment modules in the embodiments of this application. In some embodiments, the at least two risk assessment modules may also include only a prohibited items identification module and a weight deviation control module to address local business scenarios where anti-fraud responsibilities are not assumed; in other embodiments, the at least two risk assessment modules may further introduce additional assessment modules for risk dimensions such as cold chain temperature compliance and dangerous goods temperature and humidity compliance to adapt to delivery scenarios in specific industries. The above combinations all belong to the implementation methods of at least two risk assessment modules.

[0096] The specific internal structure of each risk assessment module will be described in detail in subsequent embodiments.

[0097] The contraband identification module includes a visual feature extraction submodule, a text feature extraction submodule, and a cross-modal fusion submodule. The visual feature extraction submodule and the text feature extraction submodule execute in parallel, and their outputs are jointly fed into the cross-modal fusion submodule.

[0098] The visual feature extraction submodule extracts visual features at two granularities from the image data in the associated data, including a first visual branch and a second visual branch set up in parallel. The first visual branch extracts global structural features from the image data, which are used to characterize the outlines and overall shape distribution of large-scale objects in the image, such as the elongated outer contour of a knife, the near-cylindrical outer contour of a bottle or jar, and the square stacked outer contour of a battery pack. The second visual branch extracts local texture features from the image data, which are used to characterize the edges, spots, and details of small-sized areas in the image, such as the high-density metal edges of a knife blade, the array of spots on an electric shock electrode, and the fine texture of a lighter's flint wheel.

[0099] For large or well-defined contraband, which appears as a large outline in the image, the global structural features of the first visual branch can be reliably captured. For small or concealed contraband, the discrimination information in the image is concentrated in local areas, and the local texture features of the second visual branch can be specifically captured. The two granularities are extracted in parallel and then combined in the subsequent cross-modal fusion submodule. Compared to a single-granularity visual feature extraction scheme, this approach can better address the recognition capabilities of contraband at different scales.

[0100] In some embodiments, the first visual branch employs a network based on a visual Transformer structure. The visual Transformer structure divides image data into several image patches, treating each patch as an element of a sequence. It models the global dependencies between image patches through a self-attention mechanism within the sequence, making it suitable for capturing cross-regional structural information in image data. The second visual branch employs a convolutional neural network based on a residual structure. The residual structure allows the network to train stably at deeper layers, with convolutional layers progressively extracting texture features from the local receptive field of the image data. The two networks have different focuses in inductive bias, respectively undertaking the tasks of extracting global structural features and local texture features.

[0101] As another subordinate implementation of the first visual branch, in some embodiments, the first visual branch may also employ other neural networks with global receptive fields, such as networks based on the Swin Transformer structure, or convolutional neural networks that incorporate global attention mechanisms. These networks can also model cross-regional dependencies in image data to output global structural features. As another subordinate implementation of the second visual branch, in some embodiments, the second visual branch may also employ other convolutional neural networks with local receptive fields, such as convolutional neural networks based on densely connected structures, or convolutional neural networks that incorporate depthwise separable convolutional structures. These networks can also extract texture features within local regions of image data.

[0102] In some embodiments, when the image data includes top-view image data and side-view image data, the visual feature extraction submodule first constructs a 3D structural representation of the object to be identified based on the top-view image data and side-view image data, and then extracts global structural features and local texture features from the 3D structural representation. Specifically, the visual feature extraction submodule performs geometric registration on the top-view image data and side-view image data according to the geometric parameters known to the sensing device, mapping the pixels of the same object in the two views to obtain a 3D structural representation with depth information. Compared to single-view image data, the 3D structural representation can distinguish objects that would originally overlap in a single view. When the associated data does not contain side-view image data, the visual feature extraction submodule directly extracts global structural features and local texture features using the top-view image data as input, without constructing a 3D structural representation.

[0103] The text feature extraction submodule performs bidirectional sequence encoding and intra-sequence association weighting on the waybill text data in the associated data to output text semantic features. Bidirectional sequence encoding refers to simultaneously reading characters or word units from the beginning and end of the text sequence and fusing contextual information from both directions at each position. Intra-sequence association weighting calculates weights for the correlation between any two positions in the sequence based on the bidirectional sequence encoding, and then weights and summarizes the sequence features according to these weights, ensuring that important fields occupy a larger proportion of the text semantic features. For the waybill text data "daily necessities", the literal semantics of a single character are weak. Bidirectional sequence encoding can enrich the representation of the character by combining the sender information, recipient information, and product description context that appear before and after it. Intra-sequence association weighting can highlight the weight of the "daily necessities" product description field in the entire waybill text, so that subsequent cross-modal fusion can receive text semantic features focused on the product description.

[0104] In some embodiments, bidirectional sequence encoding is implemented by a bidirectional gated recurrent unit (GRU) or a bidirectional long short-term memory (LSTM) network, while intra-sequence association weighting is implemented by a self-attention mechanism. Both GRUs and LSTMs are recurrent neural network structures capable of bidirectional sequence modeling, each with its own emphasis when modeling long sequences. The self-attention mechanism directly establishes weighted connections between any two positions within the input sequence. The text feature extraction submodule further loads a dedicated dictionary for express delivery waybill text as prior knowledge. This dedicated dictionary includes frequently occurring abbreviations, industry jargon, and easily confused fields in the express delivery industry. When the waybill text data contains entries from the dedicated dictionary, the text feature extraction submodule assigns a higher initial weight to that entry, enabling bidirectional sequence encoding and intra-sequence association weighting to prioritize these industry-specific words.

[0105] The cross-modal fusion submodule generates text-guided visual features and visual-guided text features based on the similarity between global structural features, local texture features, and text semantic features. It then performs fusion on the text-guided visual features and visual-guided text features to output the contraband identification result.

[0106] The cross-modal fusion submodule first calculates the similarity matrix between textual semantic features and global structural features, as well as the similarity matrix between textual semantic features and local texture features. Each element in the similarity matrix measures the correlation strength between a location in the textual semantic features and a location in the visual features. After the similarity matrices are generated, the cross-modal fusion submodule uses the similarity matrices as weights to perform a weighted summation of the visual features, obtaining text-guided visual features; simultaneously, it uses the similarity matrix as weights to perform a weighted summation of the textual semantic features, obtaining visual-guided text features. Text-guided visual features are relevant parts selected from visual features based on textual semantics, while visual-guided text features are relevant parts selected from textual semantic features based on visual information. Together, they constitute a bidirectional aligned representation between visual and textual information.

[0107] Text-guided visual features and visual-guided text features are not two final classification results, but rather intermediate features fed into the subsequent fusion stage. In the baseline scenario, the goods description field of the waybill text data is "daily necessities." The representation of "daily necessities" in the text semantic features drives the cross-modal fusion submodule to search for visual evidence consistent with "daily necessities" in the visual features, such as the appearance of bottles, paper products, and clothing. Conversely, if the X-ray image data contains local textures with metal blade features, the visual-guided text features will amplify the weight of keywords such as "knife" and "tool" in the waybill text. After bidirectional alignment, if the visual evidence and text evidence are semantically mutually corroborative, the two types of features point to the same category. If there is a significant misalignment between the visual evidence and text evidence, for example, the visual features point to a long metal blade while the waybill text points to "daily necessities," the bidirectional aligned features will reflect this misalignment, providing significant discriminative basis for the subsequent classification layer.

[0108] In some embodiments, the fusion strategy of the cross-modal fusion submodule includes two levels: feature-level fusion and decision-level fusion. Feature-level fusion concatenates text-guided visual features and visual-guided text features to obtain multimodal fusion features; these multimodal fusion features are input into a classification layer, outputting a first classification result. Decision-level fusion, on the other hand, retains three single-modal classification paths in parallel, inputting global structural features, local texture features, and text semantic features into independent classification layers respectively, outputting three auxiliary classification results. The cross-modal fusion submodule performs weighted synthesis on the first classification result and the three auxiliary classification results to obtain the contraband identification result. The first classification result reflects the joint judgment under multimodal collaboration, while the three auxiliary classification results reflect the independent judgment of each single modality. Weighted synthesis ensures that the classification results mutually reinforce each other when the multimodalities are consistent, and mutually constrain each other when the single modality is abnormal.

[0109] As another subordinate implementation of the cross-modal fusion submodule, in some embodiments, the cross-modal fusion submodule may only employ decision-level fusion, that is, without the first classification result of multimodal fusion feature concatenation, it directly synthesizes the contraband identification result by combining the three single-modal auxiliary classification results according to preset weights; in other embodiments, the cross-modal fusion submodule may only employ feature-level fusion, that is, without generating auxiliary classification results, it directly uses the first classification result generated by multimodal fusion features as the contraband identification result. All of the above different fusion strategies can achieve collaborative recognition of visual and textual modalities, but they differ in their ability to combat single-modal noise and the complexity of the classification layer.

[0110] For rare target categories in the prohibited items category where the number of samples is lower than a preset sample threshold, in some embodiments, the feedback optimization layer constructs pseudo-samples based on the sample library. Pseudo-samples include pseudo-image data generated from text semantic features through cross-modal processing, and pseudo-text data generated from the global structural features and local texture features of image data through cross-modal processing. The former uses a text-to-image model to generate corresponding pseudo-X-ray image data from the "rare prohibited item name" in the waybill text, while the latter uses a graph-to-text model to reverse-generate corresponding pseudo-waybill text data from the X-ray image features of rare prohibited items. These pseudo-samples serve as expanded training samples for the target categories in the sample library, participating in the iterative updates of the prohibited items identification module to alleviate the insufficient identification ability of rare categories due to the difficulty in accumulating real samples.

[0111] The weight deviation control module includes an appearance feature extraction submodule, a weight prediction submodule, and a weight risk assessment submodule. The three submodules are executed in sequence in the order of "extracting appearance features → predicting weight range → determining risk level".

[0112] The appearance feature extraction submodule extracts the size, shape category, and texture category features of the package to be identified from the image data. Size features include the length, width, and height of the package. The appearance feature extraction submodule determines these dimensions by performing edge detection and outer contour fitting on the image data. Shape category features classify the overall geometry of the package. The appearance feature extraction submodule compares the outer contour with several preset standard shape templates and selects the template with the highest fit as the shape category. Texture category features classify the material appearance of the package's outer packaging. The appearance feature extraction submodule extracts the texture pattern of the package surface from the image data and compares it with several preset material texture prototypes.

[0113] A predicted weight range refers to a range of weight values ​​with upper and lower boundaries, rather than a single predicted weight point. The purpose of introducing ranges is to acknowledge that the mapping between appearance features and weight is not strictly deterministic; that is, even with the same length, width, height, and shape, the density of the contents can still fluctuate. Therefore, appearance features can only constrain the weight to a reasonable range. The range-based output not only covers this inherent uncertainty but also provides a clear criterion for subsequent risk assessment: whether the weight falls within the range.

[0114] In some embodiments, the shape category features include at least one of cuboid, spheroid, sheet, and irregular shapes. The cuboid category corresponds to packages with distinct edges in all three directions (length, width, and height), such as paper packages in the baseline scenario; the spheroid category corresponds to packages with similar dimensions in all three directions and no distinct edges, such as vacuum-sealed clothing packages; the sheet category corresponds to packages whose size in one direction is much smaller than in the other two directions, such as document bags; and the irregular shape category corresponds to other shaped packages that cannot be classified into the above three categories.

[0115] In some embodiments, the texture category features include at least one of cardboard boxes, eco-friendly bags, and other categories. Cardboard boxes correspond to packages with outer packaging made of rigid corrugated cardboard or paperboard; eco-friendly bags correspond to packages with outer packaging made of biodegradable plastic woven bags or non-woven bags; other categories include all outer packaging materials other than the first two categories, such as bubble wrap, waterproof plastic bags, etc.

[0116] In the baseline scenario, the size features extracted from the X-ray image data by the appearance feature extraction submodule are 40cm long, 30cm wide, and 20cm high. The shape category feature matches the cuboid category, and the texture category feature matches the cardboard box category.

[0117] The weight prediction submodule outputs a predicted weight range based on size, shape, and texture features, using a cascaded structure of a visual neural network and a regression model. The visual neural network first performs further deep encoding on the aforementioned appearance features to extract high-dimensional visual representations. The regression model takes these visual representations as input and outputs the lower and upper bounds of the predicted weight range. The width between the lower and upper bounds is adaptively determined by the regression model based on the prediction bias statistics in the training set; the more typical the appearance features, the narrower the range; the closer the appearance features are to the category boundary, the wider the range.

[0118] The mapping relationship between appearance features and actual weight varies significantly across different shape categories. The combination of cuboid and cardboard box shapes corresponds to regular paper packaging containers, where the contents are relatively concentrated in terms of distribution and density. Spherical shapes correspond to compressed clothing, where density is greatly affected by the compression ratio. Sheet-shaped shapes correspond to paper or documents, where density mainly depends on the number of paper layers. Irregularly shaped shapes correspond to special packaging that cannot be categorized into the above three types, and the mapping from appearance to weight lacks a unified pattern. If samples from all the above shape categories are trained in a unified prediction model, the number of samples from the mainstream categories will dominate the model parameters, and the prediction bias of the marginal categories will be averaged out. Therefore, in some embodiments, the weight prediction submodule trains an independent prediction model for each shape category. When performing weight prediction on the package to be identified, the corresponding prediction model is first selected based on the shape category features, and then the selected prediction model outputs a predicted weight range based on size and texture category features.

[0119] In the baseline scenario, the shape category of the package to be identified is cuboid, and the weight prediction submodule selects the prediction model corresponding to the cuboid category. This prediction model is based on the size feature 40cm×30cm×20cm and the texture category feature cardboard box, and outputs a predicted weight range of [5.5kg, 6.1kg]. The center value of this predicted weight range, 5.8kg, is called the predicted weight and is used for subsequent risk assessment.

[0120] The weight risk assessment submodule classifies the weight risk level into three levels—high risk, medium risk, and low risk—based on the relative deviation between the measured weight and the predicted weight range. These levels are sequentially adjacent and do not overlap.

[0121] In some embodiments, the weight risk assessment submodule maps relative deviations to a weight deviation ratio. The formula for calculating the weight deviation ratio is:

[0122] in, For actual measured weight, To predict weight, WDR is the weight deviation ratio. When When WDR=0; when When WDR < 0, the larger its absolute value, the more significantly the measured weight is lower than the predicted weight, and the greater the suspicion that the declared weight is inflated or the contents have been replaced; when When WDR > 0, it usually indicates that the package contains additional items beyond the declared contents.

[0123] The weight risk assessment submodule classifies weight deviation ratios into high-risk, medium-risk, and low-risk categories based on whether they fall within a first, second, or third preset interval. These three preset intervals are sequentially adjacent and do not overlap. The first preset interval is defined as a high-risk interval if the absolute value of the weight deviation ratio falls within a larger range (e.g., a ratio within (−1.0, −0.5)). The second preset interval is defined as a medium-risk interval if the absolute value of the weight deviation ratio falls within a medium range (e.g., a ratio within (−0.5, −0.2)). The third preset interval is defined as a low-risk interval if the absolute value of the weight deviation ratio falls within a smaller range (e.g., a ratio within (−0.2, 0)). The sequential adjacency of these intervals ensures that each weight deviation ratio uniquely falls into a single interval, preventing ambiguity caused by overlapping intervals and avoiding missing assessments due to uncovered intervals.

[0124] In the benchmark scenario, the measured weight Predicted weight Substituting the values ​​into the formula, the weight deviation ratio WDR is calculated as WDR = (2.5 − 5.8) / 5.8 ≈ −0.57 (rounded to two decimal places). This weight deviation ratio falls within the first preset range [−1.0, −0.5]. The weight risk determination submodule determines that the weight risk level of the package to be identified is high risk, indicating that the actual contents of the package to be identified deviate significantly from the declared goods description "daily necessities" in terms of weight, and it needs to be intercepted in subsequent decision-making.

[0125] As another subordinate implementation, in some embodiments, the weight prediction submodule can also adopt a cascaded structure based on appearance feature engineering and gradient boosting trees instead of the cascaded structure of visual neural network and regression model. In this implementation, appearance features are not deep encoded by visual neural network, but are directly fed into the gradient boosting tree model as feature vectors using the numerical values ​​of size features, the category codes of shape category, and the category codes of texture category. The gradient boosting tree model then outputs the predicted weight range. The gradient boosting tree model has low training and inference overhead in scenarios with a small number of features and a moderate sample size. This alternative approach can also map the predicted weight range of the package to be identified from appearance features and is one of the implementation methods of the weight prediction submodule.

[0126] The fraud risk assessment module includes a user risk scoring submodule, an order risk scoring submodule, a branch risk scoring submodule, and a comprehensive scoring submodule. The first three submodules execute in parallel, each outputting a scoring line; the comprehensive scoring submodule aggregates the three scoring lines to obtain a comprehensive fraud risk score.

[0127] User risk scoring is a quantitative characterization of an individual sender's profile. The user risk scoring submodule outputs a user risk score based on sender information from associated data, with each field in the sender information being converted into several risk indicators. In some embodiments, the input indicators for user risk scoring include at least the sender's identity authenticity indicator, delivery behavior stability indicator, contact information validity indicator, and historical compliance record indicator. The identity authenticity indicator measures the consistency between the sender's real-name information (such as ID card) and the public identity database; the delivery behavior stability indicator measures the fluctuation in the sender's recent delivery frequency and destination distribution; the contact information validity indicator measures the sender's contact connection rate and SMS delivery rate; and the historical compliance record indicator measures the frequency of violations and seizures among the sender's historical packages.

[0128] Order risk scoring is a quantitative characterization of a single delivery behavior. The order risk scoring submodule outputs an order risk score based on order information in the associated data. In some embodiments, the input indicators for order risk scoring include at least the order's batch delivery characteristics, commodity risk characteristics, and delivery method characteristics. Batch delivery characteristics measure the total number of orders sent by the same sender within a short time window; commodity risk characteristics measure the frequency of the declared commodities in this order appearing in past fraud cases; delivery method characteristics measure the delivery channel, payment method, and pickup mode of this order.

[0129] The branch risk score is a quantitative characterization of the compliance performance of the branches through which a package passes. The branch risk score submodule outputs the branch risk score based on branch information from the associated data. In some embodiments, the input indicators for the branch risk score include at least the branch's real-name registration implementation indicator, the acceptance and inspection pass rate indicator, and the suspicious clue reporting indicator. The real-name registration implementation indicator measures the branch's compliance rate with sender real-name registration; the acceptance and inspection pass rate indicator measures the branch's pass rate for package opening and inspection; and the suspicious clue reporting indicator measures the number of suspicious packages proactively reported by the branch historically.

[0130] The three scoring sub-modules independently output user risk scores, order risk scores, and branch risk scores. The reason for using three independent scoring lines instead of a single composite score is that fraud risk originates from multiple dimensions. For example, a known compliant sender may experience batch anomalies in a single order, and a long-established, well-regulated branch may temporarily malfunction. Therefore, a single-dimensional score cannot cover fraud risk in all situations. The parallel output of the three scoring lines allows the comprehensive fraud risk score to reflect independent signals from each dimension.

[0131] The comprehensive scoring submodule outputs a comprehensive fraud risk score based on at least the user risk score, order risk score, and branch risk score. In some embodiments, the comprehensive scoring submodule performs a weighted synthesis of the three scoring lines according to preset weights, for example:

[0132] in, For comprehensive risk assessment related to fraud, These are user risk score, order risk score, and branch risk score. These are the weights of the three scoring lines, which are fitted by the model based on fraud-related labels from the historical training set.

[0133] User risk scores range from 0 to 100, with lower scores indicating higher risk. For new users with missing historical records, historical compliance record indicators are filled in using preset default values ​​before being included in the score calculation. In the baseline scenario, the sender is a new user who is making their first delivery at this branch within the past month and has a low contact rate. This sender's risk score is high due to their identity verification meeting the requirements but insufficient stability in their delivery behavior and low contact rate. This order is the sender's first order of the day, the product description is "daily necessities," and the characteristics of bulk delivery are relatively weak. The order risk score is... The branch's acceptance and inspection pass rate for mail and parcels over the past three months has been 85%, indicating good implementation of real-name registration for mail and parcels. The branch's risk score is [not specified]. =70. Weighting Substituting into the above formula yields the comprehensive risk score for fraud. .

[0134] As another subordinate implementation of the comprehensive scoring submodule, in some embodiments, the comprehensive scoring submodule can also adopt a segmented synthesis method based on rule logic instead of linear weighting. The segmented synthesis method refers to configuring synthesis rules separately for different combinations of values ​​of the three scoring lines. For example, when any scoring line is lower than a preset threshold, the comprehensive scoring submodule directly takes the score of that scoring line as the basis for the comprehensive risk score related to fraud, and then makes corrections based on the remaining two scoring lines. The segmented synthesis method can avoid omissions caused by dilution from other normal scores in linear weighting when a certain scoring line has an extremely low score.

[0135] The fraud risk assessment module can also include a risk cluster mining submodule. The aforementioned three-dimensional scoring submodule outputs independent scores for individual senders, orders, and branches. However, fraudulent activities often occur in organized, clustered forms in practice. Implicit similarities exist between a group of interconnected sender accounts, a group of similar order patterns, and a group of branches controlled by the same group. Individual scoring considers samples independently and cannot uncover these cross-sample group patterns. The risk cluster mining submodule serves to supplement this dimension.

[0136] The risk cluster mining submodule uses hybrid distance to calculate the similarity between samples to be clustered, and performs clustering based on sample distribution density based on the similarity between samples to mine at least one of high-risk customer clusters and high-risk branch clusters. The mining results are then fed back to the comprehensive scoring submodule. The samples to be clustered can be sender accounts, branch accounts, or both; the feature vector of the sample is composed of various indicators from sender information, order information, and branch information.

[0137] In some embodiments, the mixed distance is a weighted sum of the Euclidean distance component, the Jaccard distance component, and the cosine distance component. The formulas for calculating the above three distance components are as follows:

[0138] Where x and y are the feature vectors of two samples to be clustered. For Euclidean distance components, For the Jaccard distance component, For the cosine distance component, For mixed distance, The weights are for the three components. The Euclidean distance component measures the difference between two samples in numerical indicators, and is suitable for continuous numerical indicators such as delivery frequency and contact connection rate; the Jaccard distance component measures the difference between two samples in a set of categorical indicators, and is suitable for indicators composed of discrete sets such as the set of delivery addresses and the set of product categories; the cosine distance component measures the difference in direction between the feature vectors of two samples, is not sensitive to the magnitude of the numerical value, and is suitable for reflecting pattern similarity when the feature dimensions of the samples are inconsistent.

[0139] For example, let Calculate the distance between sender sample x in the baseline scenario and a candidate cluster member y, and let... Substituting into the formula, we get The smaller this value, the more similar sample x is to candidate cluster member y in terms of value, category, and orientation.

[0140] Clustering based on sample distribution density uses the aforementioned mixture distance as a measure of similarity between samples, examining the neighborhood density of each sample in the feature space. The clustering algorithm sets two parameters: neighborhood radius. and minimum number of contained points For any sample to be clustered, when the sample is the center and the neighborhood radius is... The number of samples within a region of radius is greater than or equal to the minimum number of contained points. When a sample is identified as a core sample, it is assigned to the same cluster as all samples in its neighborhood. Core samples expand into complete clusters through density reachability relationships. Isolated samples in sparsely distributed areas are not assigned to any cluster. This mechanism makes clustering naturally robust to noisy samples, unlike partitioning clustering which uses distance as the sole criterion.

[0141] In some embodiments, neighborhood radius and minimum number of contained points Use preset values. For example, set the neighborhood radius. =0.3, minimum number of contained points =5 means that a sender is considered to be at the core of a high-risk cluster only when there are at least 5 sender samples within a neighborhood centered on a particular sender sample and with a mixing distance of no more than 0.3. Parameter selection directly affects the granularity of the cluster: the larger the neighborhood radius, the more lenient the cluster, and more samples are included in the same cluster; the larger the minimum number of inclusion points, the more stringent the cluster, and only denser groups of samples are identified as a cluster.

[0142] The data mining results are fed back to the comprehensive scoring submodule. When the sender of the package to be identified falls into an already identified high-risk customer cluster, or the outlets it traversed fall into an already identified high-risk outlet cluster, the comprehensive scoring submodule introduces a cluster correction term into the calculation of the fraud risk score, adjusting the fraud risk score towards higher risk. When neither the sender nor the outlets it traversed fall into any identified high-risk cluster, no cluster correction term is applied. Through this feedback mechanism, the collective signal from risk cluster mining is injected into the single-sample scoring logic of the comprehensive scoring submodule.

[0143] As another subordinate implementation, in some embodiments, the risk cluster mining submodule can also replace one of the Euclidean distance component, Jaccard distance component, and cosine distance component with a Mahalanobis distance component or a Hamming distance component to adapt to data with different indicator distributions. The Mahalanobis distance component is sensitive to covariance and is suitable for scenarios where there is a significant correlation between indicators; the Hamming distance component is suitable for scenarios consisting entirely of discrete Boolean indicators. The aforementioned alternative distance components can also be weighted with the other two components to synthesize a mixed distance, which plays a role in measuring similarity between samples.

[0144] The fraud risk assessment module can also include a multimodal correlation submodule. User risk scores, order risk scores, and branch risk scores are all numerical representations of structured indicators, carrying fraud signals in the form of discrete score values. However, fraud clues may also be hidden in data beyond two types of structured scores: one is unstructured data in the correlation data, such as images of package appearance, handwritten images on waybills, and surveillance images of pickup and delivery sites, which carry information in the form of pixels or raw signals; the other is graph data composed of the relationships between senders, recipients, and branches, such as the same mobile phone number serving multiple sender accounts, the same delivery address associated with multiple different senders, and the same branch providing pickup and delivery services to multiple suspicious accounts in a short period of time, which carries correlation patterns in the form of graph nodes and graph edges. The role of the multimodal correlation submodule is to transform the above two types of unstructured or correlational data into supplementary features to participate in the calculation of the comprehensive fraud risk score.

[0145] The multimodal association submodule aggregates unstructured data and association graph data in the association data into supplementary features. These supplementary features are used as additional input features of the comprehensive scoring submodule, and together with user risk scores, order risk scores, and branch risk scores, they participate in the calculation of the comprehensive risk score for fraud.

[0146] In some embodiments, the supplementary features of the multimodal association submodule include image feature components extracted from image data via a convolutional neural network, and graph feature components aggregated from an association graph composed of sender information, recipient information, and branch information via a graph neural network. The image feature component extraction path involves the multimodal association submodule feeding image data from the association data into a convolutional neural network, which outputs a fixed-dimensional image feature vector. The graph feature component extraction path involves the multimodal association submodule using sender, recipient, and branch as graph nodes and their relationships as graph edges to construct an association graph; the graph neural network performs multiple rounds of neighbor aggregation on the graph, aggregating the neighbor information of each node onto its own representation to obtain a fixed-dimensional graph feature vector. Graph edges in the association graph include, but are not limited to, associations based on the same contact method, the same recipient address, and the same branch service. Through multiple rounds of aggregation, the graph neural network can aggregate information from nodes that are geographically distant but have linked connections to the target node, thereby amplifying the risk signal of nodes with "shared identity elements."

[0147] After the image feature components and the spectral feature components are concatenated, they constitute the supplementary features output by the multimodal correlation submodule. The supplementary features, along with the user risk score, order risk score, and branch risk score, participate in the calculation of the comprehensive fraud risk score through the comprehensive scoring submodule. For example, the comprehensive scoring submodule adds a supplementary scoring term obtained by nonlinear mapping of the supplementary features to the calculation formula of the comprehensive fraud risk score.

[0148] As another subordinate implementation, in some embodiments, the multimodal association submodule can also use other feature aggregation methods instead of convolutional neural networks and graph neural networks. For example, the extraction of image feature components can use a self-supervised pre-trained visual encoder, which has better generalization ability for small sample image data; the extraction of graph feature components can use a graph embedding method based on random walk sampling, encoding each graph node as a vector representation. The above alternative methods can also aggregate unstructured data and associated graph data into supplementary features, and belong to one of the implementation methods of the multimodal association submodule.

[0149] In some embodiments, the sample management submodule of the feedback optimization layer also constructs a fraud-related feature knowledge base. The fraud-related feature knowledge base consists of two parts: a basic library and an incremental library.

[0150] The basic database includes fraud-related characteristics explicitly defined by regulations. These characteristics are derived from lists of prohibited items related to fraud, descriptions of fraudulent behaviors, and suspicious items explicitly listed in regulations on postal and express delivery security issued by regulatory authorities at all levels. The basic database was imported by administrators once before the system went live and has remained relatively stable thereafter, only undergoing centralized revisions when regulations are updated. The stability of the basic database ensures that the fraud risk assessment module can consistently cover the static characteristics established at the compliance level during the training process.

[0151] The incremental database is supplemented by real-world cases at preset intervals. As new types of fraudulent activities continue to emerge, the incremental database includes fraudulent characteristic terms extracted from handling results, seized cases, and user reports. These include newly emerging disguised product names, newly identified delivery route patterns, and newly emerging suspicious contact information formats. Updates to the incremental database originate from the sample database archived in the optimization layer. The sample management submodule periodically extracts new characteristic entries from seized cases in the sample database, reviews them, and then writes them into the incremental database.

[0152] The sample management submodule merges the incremental library into the basic library according to a preset cycle. The two-layer structure of the basic library and the incremental library corresponds to the static compliance and dynamic timeliness of fraud-related features, respectively. This allows the fraud risk assessment module to acquire both stable features at the regulatory level and the latest fraud-related pattern features during each iteration of training, alleviating the contradiction between "stability" and "timeliness" in a single-layer knowledge base.

[0153] The role of the decision-making and execution layer is to transform the multi-path risk assessment results output by the intelligent analysis layer into actual actions to handle the packages to be identified. The decision-making and execution layer includes a hierarchical decision-making submodule and a handling and execution submodule.

[0154] The hierarchical decision-making submodule matches the risk assessment results output by each risk assessment module with preset multi-level risk thresholds. The multi-level risk thresholds classify each risk assessment result into several levels from low to high, with each level corresponding to a threshold. The "highest level risk threshold" corresponds to the most severe level of risk assessment, the "lowest level risk threshold" corresponds to the least severe level, and intermediate levels fall between the two. In the baseline scenario with the three modules combined, the multi-level risk thresholds for prohibited item identification results divide the results from "clearly not detected" to "clearly detected" into several levels; the multi-level risk thresholds for weight risk levels directly correspond to high, medium, and low risk levels; and the multi-level risk thresholds for the comprehensive fraud risk score divide the numerical score into several segments. For example, if the highest threshold range for the comprehensive risk score related to fraud is 0 to 30, the intermediate threshold range is 30 to 60, and the lowest threshold range is 60 to 100, when the comprehensive risk score related to fraud falls into a certain range, the hierarchical decision-making submodule determines the risk level of the comprehensive risk score related to fraud for that range.

[0155] The decision-making logic of the tiered decision-making submodule adopts a parallel approach: If any of the three risk assessment results hit the highest-level risk threshold, the handling level is determined to be automatic interception, and an automatic interception instruction is generated; if at least one intermediate-level risk threshold is hit and none hit the highest-level risk threshold, the handling level is determined to be manual review, and a manual review task is generated; if all three risk thresholds are hit, the handling level is determined to be normal flow, and a normal flow instruction is generated. The logical implication of this parallel approach is that if any one of the three risk assessment results shows the highest-level signal, it is sufficient to trigger interception; only when all three results show the lowest-level signal will the package be released. This approach ensures the system's fallback interception capability in case of failure in any dimension, while avoiding over-handling caused by all intermediate-level risks triggering interception through intermediate-level aggregation.

[0156] In the baseline scenario, the prohibited item identification result is output as "suspected knife type" by the prohibited item identification module, which is intermediate level; the weight risk level is output as high risk by the weight deviation control module, which is the highest level; and the comprehensive risk score for fraud is... The risk assessment module outputs data that falls within the intermediate threshold range, classifying it as intermediate. Following the parallel decision-making logic of the tiered decision-making submodule, the weight risk level hits the highest-level risk threshold. Therefore, the tiered decision-making submodule determines the handling level as automatic interception and generates an automatic interception command.

[0157] The execution submodule distributes automatic interception instructions, manual review tasks, or normal flow instructions to the corresponding execution terminals. In this embodiment, the execution terminals include at least one of an automatic sorting device, a cloud security inspection platform, a mobile terminal APP, and an on-site control terminal. The automatic sorting device executes automatic interception instructions—in response to receiving an automatic interception instruction, the automatic sorting device diverts the package to be identified to the suspicious item isolation area and removes it from the main sorting line. The cloud security inspection platform receives manual review tasks—in response to receiving a manual review task, the cloud security inspection platform pushes the task to security inspectors located in different sorting centers, who remotely access the associated data of the package and provide a review conclusion. The mobile terminal APP pushes early warning information and review results to frontline personnel such as delivery personnel and branch managers. The on-site control terminal displays the interception status, review progress, and handling records locally at the sorting center for real-time monitoring by on-site management personnel.

[0158] As another subordinate implementation of the hierarchical decision-making submodule, in some embodiments, the judgment logic of the hierarchical decision-making submodule can adopt a weighted synthesis approach instead of a parallel approach. The weighted synthesis approach involves first converting the three risk assessment results into scores under a unified dimension, synthesizing a comprehensive handling score according to preset weights, and then matching the handling level based on the comprehensive handling score. The weighted synthesis approach can provide a finer level classification than the parallel approach when the overall risk assessment results of multiple sources are biased towards the intermediate level; however, when the highest-level signal appears in a single dimension, it may be diluted by the lower scores of other dimensions. Both judgment logics are implementation methods of the hierarchical decision-making submodule, and the specific choice can be flexibly configured according to the business's preference for the strictness of the safety net and the granularity of the handling.

[0159] The feedback optimization layer is the final stage of the identification-decision-feedback closed loop. Its function is to transform the handling results and identification process data into a driving force for the continuous optimization of the risk assessment module. The feedback optimization layer includes at least one of the following sub-modules: sample management, model iteration, and collaborative output.

[0160] The sample management submodule archives missed detection cases, misjudged cases, and detected cases into the sample library according to preset categories. Missed detection cases refer to packages that the risk assessment module outputs as low-risk or medium-risk but are subsequently confirmed as high-risk by manual review or other means; misjudged cases refer to packages that the risk assessment module outputs as high-risk but are confirmed as not high-risk after manual review; detected cases refer to packages that are intercepted by the decision-making execution layer and confirmed as illegal or fraudulent after review. These three types of cases together constitute the "boundary samples" in the sample library, that is, samples that fall near or outside the current decision boundary of the risk assessment module, and they have special value for the subsequent training of the assessment model.

[0161] The model iteration submodule extracts samples from the sample library and divides them into training, validation, and test sets. It retrains at least one of the at least two risk assessment modules and deploys the retrained risk assessment module to the intelligent analysis layer using a canary release method. In some embodiments, the model iteration submodule divides the training, validation, and test sets in a 7:2:1 ratio. For example, when the sample library accumulates 10,000 new samples in a certain iteration cycle, the model iteration submodule uses 7,000 samples as the training set, 2,000 samples as the validation set, and 1,000 samples as the test set. The training set is used to drive model parameter updates, the validation set is used to evaluate the model's performance on unseen data during training, and the test set is used to provide the final model evaluation metrics after training is completed.

[0162] In some embodiments, canary deployment further includes: first deploying the retrained risk assessment module to a portion of the execution nodes; in response to the fact that the operating indicators of the portion of the execution nodes meet the preset stability conditions within a preset observation period, deploying the retrained risk assessment module to all execution nodes; in response to the fact that the operating indicators of the portion of the execution nodes do not meet the preset stability conditions within a preset observation period, reverting the portion of the execution nodes to the risk assessment module version before retraining. For example, the specific execution method is as follows: the model iteration submodule first deploys the retrained risk assessment module to 10% of all execution nodes, with an observation period of 3 days; if the recognition accuracy of these nodes within 3 days is not lower than the baseline before retraining, and the false positive rate is not higher than the baseline before retraining, then the retrained risk assessment module is determined to meet the preset stability conditions, and the model iteration submodule pushes it to the remaining 90% of the execution nodes; if any indicator of these nodes within 3 days does not reach the baseline, then the stability conditions are not met, and the model iteration submodule reverts these 10% of execution nodes to the risk assessment module version before retraining, preventing the defective retrained version from affecting the entire network. The rollback mechanism makes canary releases controllable in both directions, reducing iteration risks.

[0163] The collaborative output submodule synchronizes the anonymized identification results and associated data to external systems via a standardized interface. The standardized interface refers to the data export that adheres to a unified data format and communication protocol when connecting to external systems. External systems refer to data consumers other than the courier company itself, such as regulatory agencies or industry collaboration platforms. Anonymization refers to replacing, masking, or hashing personal information fields such as the sender's name, contact information, and ID number before data export, ensuring that the data synchronized to external systems retains the necessary information for risk assessment without revealing personal identity.

[0164] In some embodiments, the collaborative output submodule has a one-click synchronization mechanism and a three-level linkage mechanism. The one-click synchronization mechanism enables the collaborative output submodule to synchronize associated data with a comprehensive fraud risk score higher than a preset threshold to an external system in a single operation. Specifically, the collaborative output submodule maintains a synchronization task queue, and associated data with a comprehensive fraud risk score higher than the preset threshold is automatically added to the queue; the collaborative output submodule performs batch synchronization on the queue at a preset period, and each synchronization completes the encapsulation, encryption, and transmission of all data to be synchronized in a single remote procedure call. One-click synchronization integrates what were originally multiple independent synchronization operations into a single multi-data synchronization, reducing the connection establishment overhead and operation confirmation overhead during the synchronization process.

[0165] The three-tiered linkage mechanism ensures that synchronized data reaches the headquarters, local, and grassroots connection points in the external system. Specifically, the collaborative output submodule sends synchronized data simultaneously to the headquarters, local, and grassroots connection points in a single synchronization. Each connection point handles the data according to its jurisdiction and response procedures. For example, the headquarters connection point is responsible for nationwide case coordination and cross-regional clue aggregation; the local connection point is responsible for localized handling and collaborative response within its province / city; and the grassroots connection point is responsible for on-site investigation of specific cases. Through this three-tiered linkage, synchronized data reaches the three connection points at different execution levels immediately after a single transmission, eliminating the need to wait for top-down hierarchical transmission and significantly reducing the time window between synchronization and handling of clues.

[0166] In some embodiments, the collaborative output submodule also synchronizes the anonymized identification results and related data to the industry collaboration platform via an industry-shared interface, allowing other similar systems to access them. The industry collaboration platform is operated by an industry association or a trusted third-party organization. Each express delivery company joins as a member, uploading anonymized information on newly discovered characteristics of prohibited items and new fraudulent patterns to the platform, and also retrieving similar intelligence uploaded by other companies. The industry-shared interface differs from the aforementioned three-tiered (headquarters, local, and grassroots) linkage in its synchronization targets: the former targets other companies in the same industry, while the latter targets regulatory or law enforcement agencies. Through the industry-shared interface, the risk intelligence accumulated by a single express delivery company can be expanded into industry-wide joint defense, enabling the entire industry to identify new fraudulent patterns as soon as they first appear.

[0167] In some embodiments, the feedback optimization layer also performs reverse diffusion tracing on the delivery chain associated with the seized case based on the logistics trajectory and association graph of the seized case, outputting tracking results containing upstream related entities in the delivery chain. The specific execution method of reverse diffusion tracing is as follows: starting with a seized fraudulent package, tracing upstream along its logistics trajectory to identify each collection and delivery point the package passes through, the sender, and their associated accounts; simultaneously, using the sender and associated accounts as the starting node of the association graph, performing multi-level neighbor diffusion on the association graph to discover other nodes that share contact information, delivery address, or pickup personnel with the starting node. The diffusion results constitute a profile of the upstream related entities in the fraudulent chain, providing link-level tracing clues for external systems. For example, if the sender corresponding to a seized case shares contact information with multiple other sender accounts in the association graph, the tracking results output by reverse diffusion tracing will include these sender accounts sharing contact information, allowing external systems to conduct further collaborative investigations.

[0168] The method embodiments corresponding to the above system embodiments are as follows.

[0169] In some embodiments, a method for managing the security of express parcels based on artificial intelligence technology is provided, including steps S1-S4.

[0170] S1. Collect multi-source heterogeneous data of the package to be identified, and perform association binding on the multi-source heterogeneous data to obtain the associated data corresponding to the package to be identified.

[0171] S2. Perform at least two parallel risk assessments on the associated data. Each risk assessment selects at least one type of data from the associated data as input and outputs the corresponding risk assessment result. The input data types or assessment dimensions of different risk assessments are different.

[0172] S3. Execute a comprehensive decision based on the risk assessment results of at least two risk assessments, determine the disposal level of the package to be identified, and perform corresponding disposal operations on the package to be identified according to the disposal level to obtain the disposal result.

[0173] S4. Feedback the disposal results and process data of at least two risk assessments back to the sample library, and perform iterative updates on at least one of the corresponding models of the at least two risk assessments based on the sample library.

[0174] S1 corresponds one-to-one with the execution process of the data acquisition layer in the aforementioned system embodiments. For the specific execution method of S1, please refer to the relevant embodiments of the data acquisition layer. S2 corresponds one-to-one with the execution process of the intelligent analysis layer in the aforementioned system embodiments. At least two risk assessments in S2 may include one or more of contraband identification, weight deviation control, and fraud risk assessment. For the specific execution methods of various risk assessments, please refer to the relevant embodiments of the corresponding risk assessment modules. S3 corresponds one-to-one with the execution process of the decision execution layer in the aforementioned system embodiments. For the specific execution method of S3, please refer to the relevant embodiments of the decision execution layer. S4 corresponds one-to-one with the execution process of the feedback optimization layer in the aforementioned system embodiments. For the specific execution method of S4, please refer to the relevant embodiments of the feedback optimization layer. The method embodiments and system embodiments correspond to each other in technical features. The implementer of the method embodiments may be an independently deployed software module, a functional component embedded in an existing logistics business system, or a hybrid deployment combining the above two forms.

[0175] In some embodiments, the intelligent analytics layer further includes an edge analytics sublayer and a cloud analytics sublayer deployed at different locations. The edge analytics sublayer is deployed on the sensing device side, i.e., on server nodes at business sites such as sorting centers, transit points, or pickup points; the cloud analytics sublayer is deployed on a remote computing platform, i.e., in the express delivery company's data center or its subscribed cloud computing service platform. The two layers are connected via a dedicated business network, the bandwidth and latency of which are determined by the express delivery company's on-site conditions.

[0176] The motivation for splitting the intelligent analysis layer into edge and cloud layers is that if all risk assessments were handled by the cloud, the associated data for every package would have to be uploaded to the cloud, placing unbearable pressure on both bandwidth and cloud computing power. Conversely, if all risk assessments were handled by the edge, the edge servers' computing power would be insufficient to support the complete multimodal model, thus limiting recognition accuracy. This division of labor between the edge and cloud layers allows each to undertake tasks appropriate to its computing power level: the edge performs rapid initial screening, while the cloud performs in-depth analysis.

[0177] The edge analysis sublayer carries a lightweight version of at least two risk assessment modules. This lightweight version is constructed using at least one of the following methods: model compression, depthwise separable convolution, and knowledge distillation. Model compression involves performing weight pruning, quantization, or low-rank decomposition on the trained full-version risk assessment module, significantly reducing the model size while maintaining basic recognition capabilities. Depthwise separable convolution breaks down standard convolution into two steps: channel-wise convolution and pointwise convolution, significantly reducing the number of convolution operations and parameters compared to standard convolution. Knowledge distillation involves using a full-version teacher model with a larger number of parameters to guide the training of a student model with a smaller number of parameters, enabling the student model to output inference results close to those of the teacher model, even with a much smaller number of parameters. These three methods can be used individually or in combination.

[0178] The lightweight version performs a preliminary screening of the associated data of the packages to be identified, uploading only the associated data of packages whose preliminary screening results are suspicious to the cloud analysis sublayer. The "suspicious" criterion is determined by the edge analysis sublayer based on a preset preliminary screening threshold: when the confidence level of the risk assessment result output by the lightweight version falls into the uncertainty range, or when the output is not the lowest risk level, the package to be identified is determined to be suspicious. The full version, which carries the risk assessment module in the cloud analysis sublayer, performs in-depth analysis on the uploaded associated data and outputs the risk assessment results. The full version uses cloud computing power to support models with a larger number of parameters, and performs secondary confirmation on samples initially identified as suspicious by the edge screening, avoiding false positives and false negatives from the edge screening that directly affect the final disposal decision.

[0179] After initial edge screening, only a few related data of packages to be identified are sent to the cloud, significantly reducing link bandwidth usage. Cloud computing power is concentrated on samples that truly require in-depth analysis, and the overall processing latency is kept within an acceptable range for the business.

[0180] In some embodiments, the association binding submodule of the data acquisition layer performs binding based on cross-validation of multi-source heterogeneous data, enabling the sensing devices to maintain the stability of the association between multi-source heterogeneous data and the unique identifier of the package to be identified even in high-speed sorting scenarios. In the baseline scenario, the sorting line carries the packages to be identified through the sensing device area at a speed of 2.3 m / s. The time interval between adjacent packages may be shortened to the order of hundreds of milliseconds, and the order of data output by the sensing devices may be misaligned with the order of package arrival. The association binding submodule performs cross-validation on the waybill number output by the barcode scanner, the weight trigger time output by the weighing device, the image timestamp output by the image acquisition device, and the waybill field returned by the business system interface. Binding is considered successful only if the four types of data are consistent within a preset time window. The cross-validation method can tolerate occasional frame loss or time drift of a single sensing device. As long as the outputs of other sensing devices can corroborate each other, the binding process can still be executed stably.

[0181] In some embodiments, this system performs robust processing for at least one of the following abnormal scenarios: missing waybill information, poor image data quality, or irregular package shape. Missing waybill information refers to situations where some key fields in the waybill text data cannot be scanned or cannot be recognized after scanning, such as an empty or garbled goods description field. Poor image data quality refers to situations where local areas of the image data are blurred or overexposed due to noise from the sensing device, reflections on the surface of the package to be identified, etc. Irregular package shape refers to situations where the external shape of the package to be identified does not fall into any typical category of shape category features, and the appearance feature extraction submodule cannot reliably classify it.

[0182] Robustness handling includes filling in missing data through cross-modal generation and improving the risk assessment module's adaptability to low-quality data through data augmentation. Cross-modal generation and completion are implemented by generating missing data from another modality using a cross-modal generation model when data from one modality is missing. For example, if the cargo description field is missing in waybill text data, X-ray image data is used to generate corresponding cargo description completion data using a graph-to-text model for subsequent multimodal fusion; if the overall quality of X-ray image data is so poor that effective visual features cannot be extracted, waybill text data is used to generate corresponding pseudo-image completion data using a graph-to-text model for cross-modal alignment. Data augmentation is implemented by performing geometric transformations such as rotation, scaling, cropping, and occlusion simulation, as well as noise injection, on the image data in the training samples during risk assessment module training. This improves the trained model's adaptability to low-quality data of similar quality. When cross-modal generation and data augmentation fail to restore the abnormal sample to a level that can be assessed, this system directly marks the sample as a sample to be manually reviewed, without forcibly assigning a risk assessment result, thus avoiding erroneous assessments that could contaminate subsequent decisions.

[0183] In some embodiments, the system configures data transmission links with transmission encryption, data storage layers with storage encryption, and data access interfaces with role-based access control. Personal information in associated data is also anonymized. Transmission encryption specifically employs a transport layer security protocol to encrypt all data transmissions between the data acquisition layer, intelligent analysis layer, decision execution layer, and feedback optimization layer, as well as data synchronization with external systems. Storage encryption uses key encryption for the sample database, fraud-related feature knowledge base, and other storage media carrying associated data. Role-based access control assigns roles to access accounts within the system; each role corresponds to a set of accessible interfaces and executable operations. Accounts without the corresponding permissions cannot call interfaces or initiate operations. Anonymization is performed on personal information fields such as sender's name, contact information, and ID number when associated data enters the sample database or is collaboratively output.

[0184] As another subordinate implementation of access control, in some embodiments, the system's data access interface can also use attribute-based access control instead of role-based access control. Attribute-based access control dynamically determines access permissions based on a combination of multiple attributes of the access subject and multiple attributes of the access object. Compared with role-based access control, it has stronger rule expression capabilities and is suitable for use in scenarios with finer-grained permission division. Both access control methods can constrain the scope of access to system interfaces and operations, and both are implementation methods of access control.

[0185] In some embodiments, this system is configured with a multi-node disaster recovery architecture, which includes at least two sets of compute nodes and data nodes that serve as backups for each other. In response to a failure of the primary node, the system switches business traffic to the backup node. Specifically, during normal operation, the multi-node disaster recovery architecture replicates the data from the primary node to the backup node in a primary-backup synchronization manner, ensuring data consistency between the primary and backup nodes. System health monitoring continuously performs heartbeat checks on the primary node. When the heartbeat times out or the primary node returns an abnormal health status, health monitoring triggers a business traffic switch, forwarding business requests from the data acquisition layer, intelligent analysis layer, decision execution layer, and feedback optimization layer from the primary node to the backup node. After the primary node recovers, the system reverse-synchronizes the data generated during the backup node period back to the primary node, completing the primary-backup recovery.

[0186] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0187] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A parcel security management system based on artificial intelligence technology, characterized in that, It includes a data acquisition layer, an intelligent analysis layer, a decision execution layer, and a feedback optimization layer; The data acquisition layer is configured to acquire multi-source heterogeneous data of the package to be identified, and perform association binding on the multi-source heterogeneous data to obtain associated data corresponding to the package to be identified; The intelligent analysis layer includes at least two risk assessment modules set in parallel. Each risk assessment module is configured to select at least one type of data from the associated data as input and output the corresponding risk assessment result. The input data types or assessment dimensions of different risk assessment modules are different. The decision execution layer is configured to perform a comprehensive decision on the risk assessment results output by the at least two risk assessment modules, determine the handling level of the package to be identified, and issue a corresponding handling instruction according to the handling level; The feedback optimization layer is configured to feed back the processing results of the decision execution layer and the identification process data of the intelligent analysis layer to the sample library, and perform iterative updates on at least one of the at least two risk assessment modules based on the sample library.

2. The express parcel security management system based on artificial intelligence technology according to claim 1, characterized in that, The at least two risk assessment modules include a prohibited items identification module, a weight deviation control module, and a fraud risk assessment module. The prohibited items identification module is configured to perform multimodal fusion on the image data and waybill text data in the associated data and output the prohibited items identification result. The weight deviation control module is configured to calculate the predicted weight range of the package to be identified based on the image data in the associated data and determine the weight risk level based on the deviation between the predicted weight range and the measured weight. The fraud risk assessment module is configured to determine a comprehensive fraud risk score based on the sender information, order information, and branch information in the associated data. The results of the identification of prohibited items, the weight risk level, and the comprehensive risk score related to fraud are all used as inputs to the comprehensive decision.

3. The express parcel security management system based on artificial intelligence technology according to claim 2, characterized in that, The contraband identification module includes a visual feature extraction submodule, a text feature extraction submodule, and a cross-modal fusion submodule. The visual feature extraction submodule includes a first visual branch and a second visual branch set in parallel. The first visual branch is configured to extract global structural features of the image data, and the second visual branch is configured to extract local texture features of the image data. The text feature extraction submodule is configured to perform bidirectional sequence encoding and intra-sequence association weighting on the waybill text data to output text semantic features. The cross-modal fusion submodule is configured to generate text-guided visual features and visually guided text features based on the similarity between the global structural features, the local texture features, and the text semantic features, and to perform fusion on the text-guided visual features and the visually guided text features to output the contraband identification result.

4. The express parcel security management system based on artificial intelligence technology according to claim 2, characterized in that, The weight deviation control module includes an appearance feature extraction submodule, a weight prediction submodule, and a weight risk determination submodule. The appearance feature extraction submodule is configured to extract the size features, shape category features, and texture category features of the package to be identified from the image data. The weight prediction submodule is configured to output the predicted weight range based on the size features, shape category features, and texture category features through a cascaded structure of a visual neural network and a regression model. The weight risk determination submodule is configured to classify the weight risk level into three levels—high risk, medium risk, and low risk—that are sequentially adjacent and do not overlap, based on the relative deviation between the measured weight and the predicted weight range.

5. The express parcel security management system based on artificial intelligence technology according to claim 2, characterized in that, The fraud risk assessment module includes a user risk scoring submodule, an order risk scoring submodule, a branch risk scoring submodule, and a comprehensive scoring submodule; the user risk scoring submodule is configured to output a user risk score based on the sender information; The order risk scoring submodule is configured to output an order risk score based on the order information; The branch risk scoring submodule is configured to output a branch risk score based on the branch information; The comprehensive scoring submodule is configured to output the fraud-related comprehensive risk score based at least on the user risk score, the order risk score, and the branch risk score.

6. The express parcel security management system based on artificial intelligence technology according to claim 5, characterized in that, The fraud risk assessment module further includes at least one of a risk cluster mining submodule and a multimodal association submodule; the risk cluster mining submodule is configured to use mixed distance to calculate the inter-sample similarity between samples to be clustered, and perform clustering based on sample distribution density based on the inter-sample similarity to mine at least one of high-risk customer clusters and high-risk branch clusters, and feed the mining results back to the comprehensive scoring submodule. The multimodal association submodule is configured to aggregate unstructured data and association graph data in the association data into supplementary features, and use the supplementary features as additional input features of the comprehensive scoring submodule, which together with the user risk score, the order risk score and the branch risk score participate in the calculation of the fraud-related comprehensive risk score.

7. The express parcel security management system based on artificial intelligence technology according to claim 1, characterized in that, The decision execution layer includes a hierarchical decision submodule and a disposal execution submodule. The hierarchical decision submodule is configured to match the risk assessment results output by each risk assessment module with preset multi-level risk thresholds. In response to any one of the highest-level risk thresholds being hit, the disposal level is determined to be the automatic interception level and an automatic interception instruction is generated. In response to at least one intermediate-level risk threshold being hit and none of the highest-level risk thresholds being hit, the disposal level is determined to be the manual review level and a manual review task is generated. In response to all hits of the lowest risk threshold, the handling level is determined to be normal flow level and a normal flow instruction is generated; the handling execution submodule is configured to send the automatic interception instruction, the manual review task, or the normal flow instruction to the corresponding execution terminal respectively.

8. The express parcel security management system based on artificial intelligence technology according to claim 1, characterized in that, The feedback optimization layer includes at least one of a sample management submodule, a model iteration submodule, and a collaborative output submodule; the sample management submodule is configured to archive missed cases, misjudged cases, and detected cases into the sample library according to a preset classification; the model iteration submodule is configured to extract samples from the sample library and divide them into training sets, validation sets, and test sets, retrain at least one of the at least two risk assessment modules, and deploy the retrained risk assessment module to the intelligent analysis layer in a canary release manner; The collaborative output submodule is configured to synchronize the de-identified recognition results and associated data to an external system through a standardized interface.

9. The express parcel security management system based on artificial intelligence technology according to claim 1, characterized in that, The data acquisition layer includes a set of sensing devices and an associated binding submodule. The set of sensing devices includes at least two of the following: image acquisition devices, weighing devices, barcode scanning devices, and business system interfaces. These devices are configured to acquire at least two of the following: image data, weight data, identification data, and business data of the package to be identified. The identification data includes at least waybill text data, and the business data includes at least sender information, order information, and branch information. The image data, weight data, identification data, and business data together constitute the multi-source heterogeneous data. The association binding submodule is configured to perform verification on the multi-source heterogeneous data, bind the verified multi-source heterogeneous data to the unique identifier of the package to be identified, and obtain the association data.

10. A method for secure management of express parcels based on artificial intelligence technology, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous data of the package to be identified, and perform association binding on the multi-source heterogeneous data to obtain associated data corresponding to the package to be identified; S2. Perform at least two parallel risk assessments on the associated data. Each risk assessment selects at least one type of data from the associated data as input and outputs the corresponding risk assessment result. The input data types or assessment dimensions of different risk assessments are different. S3. Perform a comprehensive decision based on the risk assessment results of the at least two risk assessments to determine the handling level of the package to be identified, and perform corresponding handling operations on the package to be identified according to the handling level to obtain the handling result; S4. The disposal results and the process data of the at least two risk assessments are fed back to the sample library, and the model corresponding to at least one of the at least two risk assessments is iteratively updated based on the sample library.

Images