Logistics whole-process visualization method and system based on multi-source data fusion

Abstract

The application relates to the technical field of intelligent logistics, and discloses a logistics whole-process visualization method and system based on multi-source data fusion, which comprises the following steps: acquiring first logistics data containing time series data, business system data, external environment data and discrete event data, performing classification cleaning and format standardization processing to obtain second logistics data; taking a unified waybill number as a primary key, using a sliding time window to complete space-time alignment to obtain third logistics data, performing feature extraction and attention mechanism weighted fusion to generate a first fusion feature vector and construct a logistics digital twin model; dynamically calculating a cargo prediction state and a corresponding confidence interval through a multi-model fusion prediction algorithm, comparing an actual logistics state to identify abnormal events, types and risk levels, and generating a personalized visualization view based on user roles, permissions and behavior characteristics. The application solves the problems of incomplete multi-modal transport coverage, poor real-time performance and poor role adaptability of the logistics visualization technology.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent logistics, in particular to a logistics whole-process visualization method and system based on multi-source data fusion. BACKGROUND

[0002] With the deepening development of modern logistics network to multimodal transport coordination mode, the linking efficiency of multiple transportation modes such as highway, railway, sea transport and aviation directly determines the overall efficiency of the supply chain. The information transparency and traceability of the whole logistics process have become the core demand of shippers and operators. The current logistics visualization system has been widely used in single transportation scenarios, but there are still significant technical shortcomings in complex logistics scenarios across modes and whole links.

[0003] The logistics visualization data is scattered in transportation management systems, warehouse management systems, customs declaration systems, port information systems and other heterogeneous platforms. The data format, reporting frequency and waybill coding rules are not unified, there is a lack of unified data access standard and space-time alignment mechanism based on event timestamp, it is impossible to splice a continuous and complete logistics digital track, the data credibility is low, and there are data blind spots of several hours to several days in the transfer link of multimodal transport. At the same time, the existing system is a descriptive historical state backtracking display, which does not have the dynamic prediction ability of fusing real-time traffic, weather, port congestion and other multidimensional information, and cannot output accurate estimated arrival time of goods. The operators can only deal with it passively after the delay occurs, and lose the window period of early intervention.

[0004] In addition, the visualization display and business execution link are separated from each other, the abnormal events are only marked on the interface, and the response efficiency is low. The system does not build a logistics digital twin model, which cannot support scheduling scheme simulation deduction. A unified view is used for all user roles, the information matching degree is poor, and batch data synchronization mechanism is mostly used, so the state update has obvious delay, which is difficult to adapt to the fast-paced modern logistics operation demand.

[0005] In summary, the related logistics visualization technology has the defects of insufficient data fusion, lack of prediction ability, fragmented abnormal response, incomplete multimodal transport coverage, poor real-time performance and role adaptability, which cannot meet the actual needs of modern logistics whole-process intelligent management and control. SUMMARY

[0006] Therefore, the present application provides a logistics whole-process visualization method and system based on multi-source data fusion, to solve the problems of insufficient data fusion, lack of prediction ability, fragmented abnormal response, incomplete multimodal transport coverage, poor real-time performance and role adaptability of related logistics visualization technology, which cannot meet the actual needs of modern logistics whole-process intelligent management and control.

[0007] In a first aspect, the present application provides a logistics whole-process visualization method based on multi-source data fusion, the method comprising: acquiring first logistics data, the first logistics data comprising time series data, business system data, external environment data and discrete event data; performing classification cleaning and format standardization processing on the first logistics data to obtain second logistics data; using a unified waybill number as a primary key, performing spatio-temporal alignment on the second logistics data using a sliding time window to obtain third logistics data; performing feature extraction and attention mechanism weighted fusion on the third logistics data to generate a first fusion feature vector; constructing a logistics digital twin model based on the first fusion feature vector; based on the logistics digital twin model, dynamically calculating a cargo predicted state and a corresponding confidence interval using a multi-model fusion prediction algorithm; collecting an actual logistics state, and based on the actual logistics state, the cargo predicted state and the corresponding confidence interval, identifying an abnormal event, an abnormal type and a risk level; acquiring role data, permission data and behavior characteristic data of a current user, and generating a personalized visualization view based on the role data, the permission data, the behavior characteristic data, the actual logistics state, the abnormal event, the abnormal type and the risk level.

[0008] The logistics whole-process visualization method based on multi-source data fusion has the following beneficial effects: first, the first logistics data covering time series data, business system data, external environment data and discrete event data are collected in multiple dimensions, which completely covers the equipment perception dimension, business process dimension, external environment dimension and node event dimension of the whole logistics link, builds a multi-dimensional data base for subsequent whole-link visualization, and avoids the problem of incomplete information coverage caused by a single data source. The four types of data correspond to the equipment running state, business flow node, external influencing factor and discrete event marker in the logistics process, which can completely map the logistics running process in the physical world through multi-dimensional data, and ensure that the state changes of each link of the logistics have corresponding data carriers. Second, the first logistics data is classified, cleaned and standardized, the abnormal data is removed and the missing values are completed according to the characteristics of different source data, the coding rules and format specifications of various data are unified, and the data format difference and coding barrier between different business systems and equipment are eliminated. The classification cleaning can specifically process the typical abnormal problems of different types of data, avoid the interference of invalid data on the subsequent processing process, and the format standardization can make the multi-source heterogeneous data have a unified analysis basis, improve the accuracy and consistency of the original data. Then, the second logistics data is time and space aligned by using a sliding time window with a unified waybill number as the primary key, and the data of different sources and different reporting frequencies are integrated in order according to the event time dimension, realizing the accurate corresponding matching of multi-source data in the time dimension, solving the problem of different system data reporting time difference and data out-of-order arrival. Taking the unified waybill number as the association primary key can make the data scattered in various systems aggregated around the same logistics object, and the sliding time window alignment method can adapt to the frequency difference of data reporting, so that data of different update frequencies can form a coherent time series on the unified time axis. In addition, feature extraction and attention mechanism weighted fusion are performed on the third logistics data, which can mine deep representation information of logistics data from multiple dimensions such as time series features, statistical features, cross features and event sequence features, and dynamically allocate fusion weights of different feature dimensions based on the current logistics scene context, so that the features with higher relevance to the current scene occupy a higher proportion in the fusion process. Multi-dimensional feature extraction can fully capture the time sequence rules, statistical characteristics, cross influences and event sequence logic of logistics operation, and the weighted method of attention mechanism can improve the adaptability of feature vectors to different logistics scenes, generating a first fusion feature vector that can accurately reflect the actual running state of logistics. Then, the logistics digital twin model is constructed based on the first fusion feature vector, which can digitally abstract various entities in the whole logistics link, clearly define the attribute fields, state sets and standardized state transition paths of various entities, build the association relationship network between various entities, and realize real-time synchronization of entity state and establishment of whole-link state mirror.The digital twin model can fully replicate the operation process and entity relationship of the physical logistics in the digital space, so that the state change, flow process and interaction logic of the logistics entity can be synchronized and mapped in the digital space. Further, based on the logistics digital twin model, the predicted state of the goods and the corresponding confidence interval are dynamically calculated by using a multi-model fusion prediction algorithm, the time efficiency is calculated from multiple dimensions such as basic time efficiency, real-time deviation and external influence, the final prediction result is output by fusing the calculation advantages of different models, and the corresponding confidence interval is generated to reflect the uncertainty of the prediction result. The multi-model fusion method can take into account the stability of the baseline time efficiency, the sensitivity of the real-time state and the relevance of the external factors, so that the prediction result not only has the support of historical statistical rules, but also can respond to the dynamic changes of the real-time operation state and the external environment, and the output of the confidence interval quantifies the reliability of the prediction result. In addition, by collecting the actual logistics state, identifying abnormal events, abnormal types and risk levels based on the actual logistics state, the predicted state of the goods and the corresponding confidence interval, the physical logistics state collected in real time can be quantitatively compared with the predicted state output by the model, the confidence interval is used as the reference boundary for abnormal judgment, the abnormal events deviating from the normal operation interval are accurately identified, and the corresponding abnormal types and risk levels are matched. The comparison of real-time state and predicted state can make the abnormal judgment have a dynamic reference standard instead of relying on a fixed threshold, and the division of abnormal types and risk levels can realize the accurate positioning and hierarchical control of abnormal events. Finally, by obtaining the role data, permission data and behavior characteristic data of the current user, generating a personalized visual view based on the role data, permission data, behavior characteristic data, actual logistics state, abnormal events, abnormal types and risk levels, combining the identity permission and operation habit of different users, adapting the data display range and information granularity, and fusing the real-time logistics state and abnormal information to generate multi-form visual content. The adaptation of roles and permissions can ensure that different users only obtain logistics information within the permission range, the introduction of behavior characteristics can make the view layout and function priority conform to the operation habit of the user, and the multi-form visual content can adapt to the information viewing needs in different scenarios, so that different users can obtain logistics information matching their own needs, avoiding the problems of information overload or information deficiency, and improving the adaptability and practicality of the visualization system. By implementing the scheme, the problems of insufficient data fusion, lack of prediction ability, fragmented abnormal response, incomplete multi-modal transport coverage, poor real-time performance and role adaptability in related logistics visualization technology are solved, and the actual needs of modern logistics intelligent control are met.

[0009] In an optional implementation, the first logistics data is acquired, including: collecting time series data by using an Internet of Things device access protocol, the time series data including GPS positioning data and vehicle-mounted sensor data, supporting an offline caching and breakpoint retransmission mechanism; connecting business system data by using an application program interface and a message queue, the business system data including transportation, warehousing, order, and customs clearance link business status data; collecting external environment data by using a standard service interface, the external environment data including meteorological data, road real-time operation data, port real-time operation data, and railway real-time operation data; collecting discrete event data in a domain event format, the discrete event data including port node state change events, customs node state change events, and warehouse node state change events.

[0010] In an optional implementation, the first logistics data is classified, cleaned, and format-standardized to obtain second logistics data, including: removing drift points and jump points in the GPS positioning data in the time series data by using a Kalman filter algorithm, and marking missing data segments corresponding to communication blind areas; unifying waybill coding rules and data formats for the business system data, and interpolating and completing missing state fields based on context business logic; performing spatial interpolation processing on the external environment data, and matching grid environment data to longitude and latitude points corresponding to logistics paths.

[0011] In an optional implementation, the second logistics data is spatio-temporally aligned by using a sliding time window based on a unified waybill number as a primary key to obtain third logistics data, including: sorting the second logistics data in an event time dimension; stitching the sorted second logistics data based on a handover event pair composed of unloading completion events and loading completion events; checking entity state transitions in the stitched second logistics data by using a preset state machine, and marking illegal state jump data; performing time dimension alignment on the second logistics data after the checking by using a sliding time window to obtain the third logistics data.

[0012] In an optional implementation, the third logistics data is feature-extracted and attention-mechanism-weighted fused to generate a first fused feature vector, including: extracting time series features, statistical features, cross features, and event sequence features from the third logistics data, the event sequence features being generated by using an event vectorization method for coding; calculating fusion weights corresponding to the time series features, the statistical features, the cross features, and the event sequence features based on a current logistics scene context by using a feedforward neural network; performing weighted summation on the time series features, the statistical features, the cross features, and the event sequence features based on the fusion weights to generate the first fused feature vector.

[0013] In one optional implementation, a logistics digital twin model is constructed based on a first fusion feature vector, including: extracting entity attribute data and state data from the first fusion feature vector; digitally abstracting cargo logistics entities, transportation vehicle logistics entities, facility logistics entities, and personnel logistics entities; defining attribute fields and state sets for each entity; establishing standardized state transition paths for entities; and generating digital twin entity objects; extracting entity association data from the first fusion feature vector; constructing an association network between cargo logistics entities, transportation vehicle logistics entities, facility logistics entities, and personnel logistics entities based on an attribute graph model; forming an entity association graph of the logistics digital twin model; the association network includes loading relationships, docking relationships, responsibility relationships, and batch relationships; and synchronously updating the real-time states of cargo logistics entities, transportation vehicle logistics entities, facility logistics entities, and personnel logistics entities based on real-time update data from the first fusion feature vector; handling concurrent state updates through a version vector mechanism; establishing a full-link state mirror corresponding to the physical logistics process; and obtaining the logistics digital twin model.

[0014] In one optional implementation, a multi-model fusion prediction algorithm is used to dynamically calculate the predicted cargo status and corresponding confidence interval based on a logistics digital twin model. This includes: extracting historical time-delivery data and route attribute data from the logistics digital twin model; using a gradient boosting tree model to calculate a standard time-delivery baseline based on the historical time-delivery data and route attribute data; outputting a basic time-delivery prediction value; extracting real-time time-series status data from the logistics digital twin model; using a long short-term memory network and attention mechanism to output a time-delivery deviation correction value based on the real-time time-series status data; extracting external environment feature data from the logistics digital twin model; using a Bayesian network model to output an external factor time-delivery adjustment amount based on the external environment feature data; and generating the predicted cargo status and corresponding confidence interval based on the basic time-delivery prediction value, the time-delivery deviation correction value, and the external factor time-delivery adjustment amount.

[0015] In one optional implementation, actual logistics status is collected, and based on the actual logistics status, the predicted cargo status, and the corresponding confidence interval, abnormal events, abnormal types, and risk levels are identified. This includes: collecting actual logistics status data; calculating the status deviation based on the actual logistics status and the predicted cargo status; matching and verifying the status deviation with the confidence interval; determining an abnormal event when the status deviation exceeds the confidence interval; and classifying the risk level based on the deviation magnitude; and matching the abnormal type based on the status deviation, where abnormal types include time delays, location loss, exceeding temperature and humidity limits, abnormal vibration, and illegal status jumps.

[0016] Secondly, this invention provides a logistics end-to-end visualization system based on multi-source data fusion. The system includes: an acquisition module for acquiring first logistics data, which includes time-series data, business system data, external environment data, and discrete event data; a standardization module for performing classification, cleaning, and format standardization on the first logistics data to obtain second logistics data; an alignment module for using a unified waybill number as the primary key and a sliding time window to perform spatiotemporal alignment on the second logistics data to obtain third logistics data; a fusion module for performing feature extraction and attention-based weighted fusion on the third logistics data to generate a first fusion feature vector; and a logistics digital twin model constructed based on the first fusion feature vector; a prediction module for dynamically calculating the predicted cargo status and corresponding confidence intervals based on the logistics digital twin model using a multi-model fusion prediction algorithm; an identification module for collecting actual logistics status and identifying abnormal events, abnormal types, and risk levels based on the actual logistics status, the predicted cargo status, and the corresponding confidence intervals; and a visualization module for acquiring the current user's role data, permission data, and behavioral characteristic data, and generating a personalized visualization view based on the role data, permission data, behavioral characteristic data, actual logistics status, abnormal events, abnormal types, and risk levels. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a specific example of a logistics end-to-end visualization method based on multi-source data fusion, according to embodiments of the present invention. Figure 2 This is an overall architecture diagram of a full-process visualization system based on the multi-source data fusion-based logistics full-process visualization method according to embodiments of the present invention; Figure 3 This is a flowchart of a full-process visualization method based on multi-source data fusion in logistics, according to an embodiment of the present invention. Figure 4 This is a flowchart of the multi-source data fusion processing method based on the multi-source data fusion-based end-to-end logistics visualization method according to embodiments of the present invention; Figure 5 This is a diagram of the architecture of the digital twin modeling layer and ETA prediction engine based on the multi-source data fusion-based logistics end-to-end visualization method in this invention. Figure 6This is a flowchart of the closed-loop process for anomaly warning and one-click intervention based on the multi-source data fusion-based full-process logistics visualization method according to embodiments of the present invention. Figure 7 This is a five-layer architecture data flow and time-series interaction diagram based on the multi-source data fusion-based logistics end-to-end visualization method according to embodiments of the present invention; Figure 8 This is a principle block diagram of a specific example of a logistics end-to-end visualization system based on multi-source data fusion, according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] As modern logistics networks evolve towards multimodal transport collaboration, the efficiency of connections between various modes of transportation, including road, rail, sea, and air, directly determines the overall effectiveness of the supply chain. Transparency and traceability throughout the logistics process have become core demands for both cargo owners and operators. While current logistics visualization systems are widely used in single-transport scenarios, significant technological shortcomings remain in complex cross-modal, end-to-end logistics scenarios.

[0021] Logistics visualization data is scattered across multiple heterogeneous platforms, including transportation management systems, warehouse management systems, customs declaration systems, and port information systems. Data formats, reporting frequencies, and waybill coding rules are inconsistent, lacking unified data access standards and a spatiotemporal alignment mechanism based on event timestamps. This makes it impossible to piece together a continuous and complete digital logistics trajectory, resulting in low data reliability. Furthermore, data blind spots of several hours to several days are common in the transshipment and handover stages of multimodal transport. Simultaneously, existing systems mostly provide descriptive historical status retrospective displays, lacking the dynamic prediction capabilities to integrate multi-dimensional information such as real-time road conditions, weather, and port congestion. They cannot output accurate estimated arrival times for goods, leaving operators only able to react passively after delays occur, losing the window of opportunity for early intervention.

[0022] Furthermore, the visualization and business execution processes are disconnected. Abnormal events are only marked on the interface and require manual cross-system coordination, resulting in low response efficiency. The system has not built a logistics digital twin model, so it cannot support the simulation and deduction of scheduling schemes. It adopts a unified view for all user roles, resulting in poor information matching. Moreover, it mostly uses batch data synchronization mechanisms, which have significant delays in status updates and are difficult to adapt to the fast-paced needs of modern logistics operations.

[0023] In summary, the relevant logistics visualization technologies suffer from shortcomings such as insufficient data integration, lack of predictive capabilities, fragmented anomaly response, incomplete coverage of multimodal transport, and poor real-time performance and role adaptability, thus failing to meet the actual needs of intelligent management and control throughout the modern logistics process.

[0024] To address the technical problems mentioned in the background section, this application provides a method for end-to-end logistics visualization based on multi-source data fusion, as detailed below. Figure 1 As shown, the method includes: Step S101: Obtain first logistics data, which includes time-series data, business system data, external environment data, and discrete event data.

[0025] Specifically, step S101 includes: Step a1: Collect time-series data using IoT device access protocols. The time-series data includes GPS positioning data and vehicle sensor data, and supports offline caching and breakpoint retransmission mechanisms.

[0026] Furthermore, Figure 2 This is an overall architecture diagram of a logistics end-to-end visualization system based on the multi-source data fusion method of this invention. The architecture is divided into a data access layer, a data processing and fusion layer, a digital twin modeling layer, a visualization rendering layer, and an interactive intervention layer from bottom to top. It is equipped with a Kafka message bus and a closed-loop feedback mechanism to realize the end-to-end data flow and continuous optimization. Figure 3This is a flowchart of a full-process visualization method for logistics based on multi-source data fusion, as described in this invention. The main process proceeds from top to bottom: First, multi-protocol data access is performed, real-time access to four types of heterogeneous data from multiple sources: time series, business systems, external environment, and events. Next, data cleaning is carried out, including GPS drift removal, format standardization, and missing value filling. Third, spatiotemporal alignment and feature fusion are performed, using a unified waybill number as the primary key to complete spatiotemporal alignment of multi-source data and generate multi-dimensional feature vectors for each logistics entity. Fourth, digital twin modeling is performed, completing entity modeling, state synchronization, and association graph construction to establish a full-link state mirror. Fifth, dynamic ETA prediction is performed, outputting predicted values ​​and confidence intervals based on a multi-model fusion algorithm. Sixth, anomaly detection and tiered early warning are performed, comparing the actual state with the predicted state to determine anomalies. The process includes anomaly judgment nodes; if no anomaly is detected, a personalized visualization view is generated directly; if anomaly is detected, a decision dashboard and intervention entry are simultaneously displayed, and an emergency procedure is executed. Finally, all operational data enters the feedback optimization stage, updating model parameters in reverse based on prediction deviations and intervention effects, forming a continuously iterative closed-loop optimization system. The data access layer consists of four types of acquisition modules, which respectively connect to terminal devices, business systems, meteorological and traffic interfaces, and node event sources to achieve unified access of time-series data, business data, environmental data, and discrete event data. The data processing and fusion layer sequentially performs data cleaning, spatiotemporal alignment, feature engineering, and attention-weighted fusion to convert heterogeneous data into standardized fused features. The digital twin modeling layer constructs digital mirrors of four types of logistics entities, establishes a full-link state mirror, and calculates the predicted state and confidence interval of goods through multi-model fusion algorithms to support simulation and deduction. The visualization and rendering layer generates multi-form visualization content based on user roles and distributes status updates and anomaly annotations through real-time push channels. The interactive intervention layer provides decision dashboards, one-click intervention, and effect feedback capabilities. Data is used to optimize model parameters through a right-side closed-loop process, forming a self-iterative operating system. Communication connections with terminal devices are established based on various IoT device access protocols, including message queue telemetry transmission protocols and restricted application protocols, adapting to the communication standards of various vehicle terminals. The collected time-series data includes GPS positioning data and vehicle sensor data. GPS positioning data covers information such as longitude, latitude, speed, direction of travel, and altitude, while vehicle sensor data covers information such as temperature, humidity, vibration status, and door opening / closing status. It is compatible with industry-standard positioning protocols and proprietary protocols from major manufacturers, using a protocol parsing component to uniformly convert binary data from different manufacturers into a standard structured data format. For scenarios with communication dead zones, an offline caching mechanism is configured on the terminal device side. When the device is in a signal interruption state, the collected data is temporarily stored in the local cache area. Once the communication link is restored, the cached historical data is automatically retransmitted, carrying a corresponding timestamp for subsequent spatiotemporal alignment.The data reporting frequency can be configured and adjusted according to actual business needs to adapt to the different requirements of data accuracy in different transportation scenarios.

[0027] Step a2 involves using application programming interfaces (APIs) and message queues to connect with business system data, which includes business status data for transportation, warehousing, orders, and customs clearance.

[0028] Furthermore, data connections with various business systems are established based on application programming interface (API) gateways and message queue components. Business status data is acquired using both descriptive state passing interfaces and message queues. The integrated business systems include transportation management systems, warehouse management systems, order management systems, and customs declaration systems. The collected data covers order status change information, waybill node update information, warehouse operation status information, and customs clearance progress information. Warehouse operation status information includes the execution status of inbound, outbound, shelving, and picking operations. For legacy business systems that do not support API integration, a database log change capture method is provided for integration. Business status change data is extracted by capturing incremental logs from the database, achieving non-intrusive data access. All integrated business data is distributed and transmitted through a unified message bus, ensuring the real-time performance and reliability of data transmission. This provides a stable business data source for subsequent data cleaning and fusion processing, avoiding increased management costs caused by fragmented data transmission links between different business systems.

[0029] Step a3: Collect external environmental data using standard service interfaces. External environmental data includes meteorological data, real-time road operation data, real-time port operation data, and real-time railway operation data.

[0030] Furthermore, data connections are established with various external public service platforms based on standardized service interfaces to collect various external environmental data along the logistics chain routes as needed. This external environmental data includes meteorological data, real-time road operation data, real-time port operation data, and real-time railway operation data. Meteorological data covers meteorological elements such as precipitation probability, wind force, and visibility in the affected areas. Real-time road operation data covers real-time road conditions and congestion levels on the affected road sections. Real-time port operation data covers ship schedules, berthing and departure events, and port congestion levels. Real-time railway operation data covers train timetables and on-time / delay status information. An additional holiday calendar service is integrated to collect time information for statutory holidays and major events. The collection frequency is configured according to the data update characteristics of different data sources. Data sources with high timeliness requirements are acquired through real-time push, while data sources with gradual changes are acquired through periodic polling. This ensures data validity while controlling resource consumption during data collection, providing external data support for subsequent timeliness prediction and anomaly detection.

[0031] Step a4: Collect discrete event data using the domain event format, where the discrete event data includes port node status change events, customs node status change events, and warehouse node status change events.

[0032] Further, standardize and encapsulate discretely occurring logistics node events based on the domain event schema, and collect status change event data of each node in the full link of the logistics network. The discrete event data includes port node status change events, customs node status change events, and warehouse node status change events, wherein the port node events include a ship berthing event, a ship unberthing event, and a loading / unloading operation completion event; the customs node events include a customs declaration form acceptance event, a customs declaration form examination and approval completion event, and a customs release event; the warehouse node events include a warehousing operation completion event, a outbound operation start event, and an inventory check completion event. Discrete events in the transportation link are additionally collected, including a pickup completion event, a transit arrival event, and a sign-off completion event. Each event includes an event type, an occurrence timestamp, an associated entity identifier, an event source system, and a business attribute field. All events are defined in format according to a unified domain event specification, ensuring that event data from different sources have a consistent parsing standard, providing a standardized event data source for subsequent link splicing and state flow verification, and supporting trajectory stitching and state machine verification in the multimodal transport scenario.

[0033] Step S102: Perform classification cleaning and format standardization processing on the first logistics data to obtain second logistics data.

[0034] Specifically, step S102 includes: Step b1: Use a Kalman filter algorithm to eliminate drift points and jump points from GPS positioning data in time series data, and mark the missing data segments corresponding to communication blind areas.

[0035] Further, Figure 4This is a flowchart of the multi-source data fusion processing method for end-to-end logistics visualization based on multi-source data fusion, as described in this invention. The time-series data acquisition module connects to GPS devices to collect positioning and sensor time-series data; the business system data acquisition module connects to the transportation management system, warehouse management system, and order management system to collect end-to-end business status data; the external environment data acquisition module connects to meteorological, transportation, and port open interfaces to collect external environment operation data; and the event data acquisition module connects to port, customs, and warehouse nodes to collect discrete state change event data. All four types of collected data are uniformly integrated into the Kafka message bus, forming a standardized and unified data flow for downward distribution. The data processing and fusion layer is set up with four functional modules according to the processing order, sequentially completing data cleaning, spatiotemporal alignment, feature engineering, and data fusion. The data cleaning module performs drift removal, format standardization, and missing value filling; the spatiotemporal alignment module completes event time alignment and state machine verification based on a sliding time window and watermarking mechanism; the feature engineering module extracts three types of features—time-series, statistical, and cross-features—and completes event sequence vectorization encoding; the data fusion engine achieves multi-feature weighted fusion based on an attention mechanism, ultimately outputting a standardized multi-dimensional logistics feature vector, providing high-quality feature input for subsequent digital twin modeling. This system processes GPS positioning data from time-series data using a Kalman filter algorithm. The filter parameters are dynamically adjusted based on the vehicle's motion state, distinguishing between different motion modes such as stationary, moving, and turning, and adjusting the state estimation weights accordingly. The system analyzes the position, velocity, and direction information of the GPS positioning data frame by frame, identifying position jumps and drift points that do not conform to the vehicle's motion patterns, and removing abnormal positioning data that deviates from the normal trajectory. Time periods without valid positioning data are identified as missing data segments corresponding to communication blind spots. The start and end times of these missing data segments are marked as the basis for subsequent data interpolation. For example, when a vehicle enters a tunnel area with signal obstruction or a densely built-up port area, the positioning signal is interrupted. The system automatically marks the missing data segments for the corresponding time period. Simultaneously, based on the valid positioning data before and after, the trajectory is smoothed using a filtering algorithm to eliminate trajectory spikes caused by signal fluctuations, ensuring the continuity and accuracy of the positioning data.

[0036] Step b2 involves standardizing the waybill coding rules and data format for the business system data, and interpolating and completing the missing status fields based on the contextual business logic.

[0037] Furthermore, a mapping relationship is established between waybill codes from different business systems and a unified waybill number. The waybill identifiers inherent in each business system are uniformly converted into globally consistent waybill coding rules, eliminating coding differences between different systems. Format standardization processing is performed on various fields of business system data. Date fields are uniformly converted to international standard time formats, and address text fields are uniformly converted to latitude and longitude coordinates, ensuring that business data from different sources has a consistent parsing format. For missing status fields in business data, interpolation is performed based on adjacent business node data and historical business operation logic. The missing status time and value are estimated by referring to the historical operation patterns of similar business scenarios. For example, when the warehousing business data lacks the inbound completion time field, it is estimated and supplemented based on the inbound start time record and the historical average inbound operation time for the same type of goods, ensuring the integrity of the business status flow chain and preventing missing fields from affecting subsequent spatiotemporal alignment and feature extraction results.

[0038] Step b3 involves performing spatial interpolation on the external environment data to match the gridded environmental data to the latitude and longitude points corresponding to the logistics path.

[0039] Furthermore, external environmental data is often published in a gridded format, with each grid cell covering a certain geographical area and corresponding environmental values ​​for that area. This makes it impossible to directly correlate these values ​​with precise latitude and longitude points along the logistics path. Spatial interpolation is performed to extract a series of latitude and longitude sampling points along the logistics path. Based on multiple gridded environmental data surrounding each sampling point, a spatial interpolation algorithm calculates the corresponding environmental parameter values ​​for each sampling point, transforming the discretely distributed gridded environmental data into continuously distributed environmental data along the logistics path. For example, gridded precipitation and wind data published by meteorological services can be spatially interpolated to calculate the precipitation probability and wind force level corresponding to each coordinate point on the logistics transportation path, establishing a one-to-one correspondence between environmental data and logistics trajectory points. This provides accurately matched external environmental characteristic data for subsequent timeliness prediction and anomaly analysis.

[0040] Step S103: Using the unified waybill number as the primary key, the second logistics data is spatiotemporally aligned using a sliding time window to obtain the third logistics data.

[0041] Specifically, step S103 includes: Step c1: Sort the second logistics data according to the event time dimension.

[0042] Furthermore, event time refers to the actual time when data is generated in the physical world, distinct from the processing time at the data arrival and processing stage. Using a unified waybill number as the grouping basis, all secondary logistics data corresponding to the same waybill are aggregated and sorted in ascending order according to the event timestamp carried by each data entry. The sorting process is entirely based on event time, not processing time. For data arriving out of order or with delays, it is still inserted into the corresponding position on the timeline according to the actual time of the event, without using the order of arrival as the sorting basis. For example, if port berthing event data arrives at the processing stage later than subsequent loading / unloading completion event data due to network transmission delays, it is still placed before the loading / unloading completion event according to the actual time of the berthing event, ensuring the accuracy of the data timing throughout the entire chain and providing a correct timing foundation for subsequent trajectory stitching and status verification.

[0043] Step c2: Construct a handover event pair consisting of unloading completion event and loading completion event, and stitch and sort the second logistics data based on the handover event pair.

[0044] Furthermore, the handover event pairs correspond to the handover nodes in multimodal transport scenarios where goods are transferred from one mode of transport to another. They consist of an unloading completion event from the previous transport stage and a loading completion event from the next transport stage. In the sorted second logistics data, unloading completion events corresponding to the same waybill are retrieved, along with time-adjacent loading completion events, to form valid handover event pairs. Based on the time and location information of these handover event pairs, the trajectory data of the previous transport mode is connected to the trajectory data of the next transport mode, generating a virtual transport segment for in-yard transfer within the interval between the two trajectories, filling the data gaps during the switch between the two transport modes. For example, a port unloading completion event in the sea transport segment and a railway loading completion event in the railway transport segment are matched as a handover event pair, smoothly connecting the ship transport trajectory and the railway transport trajectory in the port area, eliminating data blind spots during multimodal transport and ensuring the continuity of the logistics trajectory.

[0045] Step c3: Use a preset state machine to verify the entity state transitions in the stitched second logistics data and mark illegal state transition data.

[0046] Furthermore, the preset state machine defines the legal state transition paths for various logistics entities, clarifies the allowed preceding and succeeding states for each entity state, and restricts the legal order of entity state transitions. It iterates through all entity state change records in the stitched second logistics data, verifying each state transition against the legal transition path defined by the preset state machine, and checking if the preceding state for each state change is an allowed preceding state. State transitions that do not conform to the legal transition path are identified as illegal state transition data and marked to prevent erroneous state data from entering subsequent processing stages. For example, if a record jumps directly from the "signed" state to the "in transit" state, it is determined to be inconsistent with the state transition logic of the goods entity and marked as illegal state transition data, ensuring the legality and consistency of entity state data and providing a reliable state data foundation for the construction of the digital twin model.

[0047] Step c4: Align the time dimension of the inspected second logistics data using a sliding time window to obtain the third logistics data.

[0048] Furthermore, the sliding time window has a preset window size and sliding step size, continuously sliding along the time axis at a fixed step size to achieve segmented aggregation of time-series data. The verified second logistics data is grouped according to a unified waybill number, with each group corresponding to the full-link data of one waybill. Within each sliding time window, all types of data from the same waybill within that window's time range are aggregated, matching data from different sources and with different reporting frequencies at the same time granularity, achieving unified alignment of multi-source data in the time dimension. A watermarking mechanism is configured, setting a watermark threshold based on the historical maximum delay of each data source. Late data exceeding the watermark window is individually marked, triggering correction processing of the already output alignment results, ultimately generating third logistics data with unified time-dimensional alignment, providing a time-consistent standardized dataset for subsequent feature extraction and data fusion.

[0049] Step S104: Feature extraction and attention-based weighted fusion of the third-party logistics data are performed to generate a first fused feature vector. A logistics digital twin model is then constructed based on the first fused feature vector.

[0050] Specifically, step S104 includes: Step d1 involves extracting time-series features, statistical features, cross-features, and event sequence features from the third-party logistics data. The event sequence features are generated using an event vectorization method.

[0051] Furthermore, temporal features, statistical features, cross-features, and event sequence features are extracted from the third-party logistics data. Temporal features are calculated using a sliding window to measure temporal statistics such as speed change rate, dwell time, acceleration, historical average speed, and yaw distance from the planned route, comprehensively capturing the operational sequence patterns of logistics entities. Statistical features are calculated based on historical operational data to measure historical delay rates for each route segment, average customs clearance times for each port, on-time rates for each carrier, and warehouse turnover rates for each warehouse, reflecting the overall operational patterns of the logistics chain. Cross-features are calculated to measure higher-order interaction features such as combinations of cargo type and weather conditions, transportation mode and time period, and origin and destination, uncovering the coupled influence of factors across different dimensions. The event vectorization method uses a jump-word model to map discrete event sequences to dense vectors, embedding and encoding the event sequences experienced by logistics entities, making semantically similar events close in distance in the vector space, generating event sequence features. For example, for a batch of multimodal transport of precision equipment, the system extracts the temporal characteristics of speed fluctuations during transportation, the historical delay statistics of the corresponding ports, the cross-risk characteristics of cargo type and typhoon weather, and the vectorized characteristics of all-link node events, comprehensively covering the multi-dimensional representation of logistics operations.

[0052] Step d2: Using a feedforward neural network, calculate the fusion weights corresponding to time-series features, statistical features, cross features, and event sequence features based on the current logistics scenario context.

[0053] Furthermore, a feedforward neural network is a unidirectional, multi-layered neural network consisting of an input layer, hidden layers, and an output layer. It can learn nonlinear mapping relationships based on input data. Contextual information of the current logistics scenario is extracted, including the current transportation stage, the surrounding environment, and the operational status of the logistics link. This contextual data is input into the feedforward neural network. The hidden layers of the feedforward neural network undergo nonlinear transformations through activation functions to learn the importance relationships of various feature dimensions under different scenarios. The output layer outputs the fusion weights corresponding to temporal features, statistical features, cross-features, and event sequence features. The weight calculation process is dynamically generated based entirely on the current scenario, and the weight ratio of each feature is automatically adjusted under different scenarios. For example, when the logistics scenario is in a port congestion stage, the feedforward neural network automatically increases the fusion weight of statistical features and event sequence features, and decreases the weight ratio of temporal features, making the fusion result more focused on the core influencing factors related to port congestion.

[0054] Step d3: Based on the fusion weights, the temporal features, statistical features, cross features, and event sequence features are weighted and summed to generate the first fusion feature vector.

[0055] Furthermore, the extracted temporal feature vector, statistical feature vector, cross feature vector, and event sequence feature vector are multiplied element-wise with the corresponding fusion weights calculated in step d2 to complete the weighting of single-dimensional features. The weighted temporal feature vector, statistical feature vector, cross feature vector, and event sequence feature vector are then concatenated and accumulated according to their dimensional order. All feature dimensions are aligned according to a unified vector dimension rule to generate a first fusion feature vector with unified dimensions. During the fusion process, all feature dimensions retain their original physical meaning; only the contribution of different features to the final result is adjusted through weighting to ensure the interpretability of the fused features. For example, in a typhoon-affected maritime shipping scenario, the weighted temporal operation features, port delay statistics, weather-cargo cross-risk features, and port event sequence features are accumulated and fused to generate a unified feature vector that comprehensively reflects the current status of the waybill, providing standardized feature input for subsequent digital twin model construction.

[0056] Step d4: Extract entity attribute data and state data from the first fusion feature vector, digitally abstract the cargo logistics entity, transportation vehicle logistics entity, facility logistics entity and personnel logistics entity, define the attribute fields and state sets of the cargo logistics entity, transportation vehicle logistics entity, facility logistics entity and personnel logistics entity respectively, establish standardized state transition paths for the entities, and generate digital twin entity objects.

[0057] Furthermore, attribute and status data corresponding to cargo logistics entities, transportation vehicle logistics entities, facility logistics entities, and personnel logistics entities are extracted from the first fusion feature vector. These entities in the physical world are mapped to digital objects in the digital space, completing the digital abstraction of the entities. For cargo logistics entities, attribute fields such as cargo type, weight, volume, and temperature requirements are defined, along with status sets such as en route, in warehouse, at port, and signed for, establishing a standardized status transition path from order placed to signed for. For transportation vehicle logistics entities, attribute fields such as vehicle type, load capacity, and temperature control capability are defined, along with status sets such as idle, driving, and parked, establishing corresponding status flow rules. For facility logistics entities, attribute fields such as warehouse capacity and berth utilization rate are defined, along with status sets such as normal operation and congestion. For personnel logistics entities, attribute fields such as role and department are defined, along with status sets such as idle and performing tasks. Each type of entity generates an independent digital twin entity object according to the defined attributes and state rules. For example, a batch of cold chain electronic products is mapped to a digital twin entity object corresponding to the cargo logistics entity, and the corresponding container ship is mapped to a digital twin entity object corresponding to the transportation logistics entity, with corresponding attribute fields and state transition rules configured respectively.

[0058] Step d5: Extract entity association data from the first fused feature vector, and construct an association network between cargo logistics entities, transportation vehicle logistics entities, facility logistics entities, and personnel logistics entities based on the attribute graph model, forming an entity association graph of the logistics digital twin model. The association network includes loading relationships, docking relationships, responsibility relationships, and batch relationships.

[0059] Furthermore, the attribute graph model is a graph data storage model where nodes represent entities and attributed edges represent the relationships between entities, supporting efficient graph querying and path tracing. Interactive relationship data between entities is extracted from the first fused feature vector, identifying the relationship types and attributes between cargo logistics entities, transportation vehicle logistics entities, facility logistics entities, and personnel logistics entities. Using the digital twin entity objects corresponding to cargo logistics entities, transportation vehicle logistics entities, facility logistics entities, and personnel logistics entities as graph nodes, loading relationships, docking relationships, responsibility relationships, and batch relationships are constructed as graph edges. The loading relationship corresponds to the carrying relationship between cargo logistics entities and transportation vehicle logistics entities; the docking relationship corresponds to the dwelling relationship between transportation vehicle logistics entities and facility logistics entities; the responsibility relationship corresponds to the duty relationship between cargo logistics entities and personnel logistics entities; and the batch relationship corresponds to the same order relationship between different cargo logistics entities. All relationship edges carry corresponding business attributes, fully recording the time, status, and other information of the relationship, ultimately forming a structured entity relationship graph. For example, through the entity association graph, one can directly query all the transport vehicles and logistics entities docked at a port facility logistics entity, as well as all the cargo logistics entities loaded on each transport vehicle and logistics entity and the corresponding responsible personnel logistics entities, supporting full-chain association tracking and query.

[0060] Step d6: Based on the real-time updated data of the first fused feature vector, the real-time status of the cargo logistics entity, transportation vehicle logistics entity, facility logistics entity, and personnel logistics entity is updated synchronously. The concurrent status updates are processed through the version vector mechanism to establish a full-link status mirror corresponding to the physical logistics process, thus obtaining a logistics digital twin model.

[0061] Furthermore, the version vector mechanism is a consistency mechanism for handling distributed concurrent updates. Each entity maintains a corresponding version identifier vector, and each state update carries version information to identify concurrent updates and handle conflicts according to eventual consistency rules. It receives real-time update data streams from the first fused feature vector, extracts the latest state data for each of the cargo logistics entity, transportation vehicle logistics entity, facility logistics entity, and personnel logistics entity, and synchronously updates the attribute and state fields of the corresponding digital twin entity objects. For multi-source concurrent update requests for the same entity, the version vector mechanism compares the order of updates and merges state changes according to eventual consistency rules, resolving state conflicts caused by simultaneous updates from multiple sources. It continuously synchronizes the state changes of all entities, constructing a full-link state mirror in the digital space that completely corresponds to the physical logistics process. The mirror data remains synchronized with the state changes in the physical world. Incremental snapshots of the full-link state are generated at fixed time intervals, recording only state changes to reduce storage overhead, supporting historical state backtracking and batch queries, ultimately forming a complete logistics digital twin model. For example, when vehicle location data and warehouse operation data update the status of the same cargo logistics entity simultaneously, the order of status changes is confirmed through version vectors to ensure accurate cargo status flow. After continuous synchronization, a digital mirror covering the entire chain is formed, completing the construction of a logistics digital twin model.

[0062] Step S105: Based on the logistics digital twin model, use a multi-model fusion prediction algorithm to dynamically calculate the predicted status of the goods and the corresponding confidence interval.

[0063] Specifically, step S105 includes: Step e1: Extract historical timeliness data and road segment attribute data from the logistics digital twin model, use the gradient boosting tree model to calculate the standard timeliness baseline based on the historical timeliness data and road segment attribute data, and output the basic timeliness prediction value.

[0064] Furthermore, Figure 5This is an architecture diagram of the digital twin modeling layer and ETA prediction engine based on the multi-source data fusion-based end-to-end logistics visualization method of this invention. The upper-layer digital twin modeling layer contains four functional modules. The logistics entity modeling module completes the digital abstraction of four core logistics entities: goods, vehicles, facilities, and personnel. It is equipped with a state machine and association graph to support the definition of entity state transitions and relationships. The end-to-end state mirroring module generates an end-to-end state mirror corresponding to the physical logistics through real-time state synchronization and concurrent state merging processing, using a global state machine. The ETA prediction engine module integrates the basic timeliness model, real-time correction model, and external influence model, and outputs the estimated arrival time and confidence interval of goods through fusion calculation. The simulation and deduction module sequentially completes simulation modeling, scheme arrangement, and comparative evaluation, outputting a simulation comparison report of multiple schemes to support the pre-simulation of scheduling schemes. The lower layer is the internal structure of the ETA prediction engine, with three independent inputs and corresponding models. Static features are input into the basic timeliness model, which outputs a baseline timeliness value based on the gradient boosting tree algorithm. Real-time time-series features are input into the real-time correction model, which outputs a timeliness deviation correction value based on a long short-term memory network and an attention mechanism. External environment features are input into the external influence model, which outputs the timeliness adjustment amount for external factors based on a Bayesian network. The three types of results are merged into the fusion calculation module, which calculates the final estimated arrival time and the corresponding confidence interval through a preset formula, realizing dynamic timeliness prediction of multi-dimensional data. The gradient boosting tree model is an ensemble learning algorithm that gradually fits the data pattern through iterative training of multiple decision trees, and is suitable for regression prediction tasks of structured data. The entire historical timeliness data and attribute data of each road segment are extracted from the logistics digital twin model. The historical timeliness data includes historical running time and historical delay records for each road segment under different transportation modes, while the road segment attribute data includes road segment distance, transportation mode type, carrier information, and cargo type information. Historical timeliness data and route attribute data are used as training samples input into the gradient boosting tree model. The model learns the timeliness distribution patterns under different route segments and transportation scenarios through multiple rounds of iterative training, fitting the mapping relationship between route attributes and timeliness duration, and calculating the standard timeliness baseline for each route and transportation mode. For the current waybill to be predicted, the model matches the standard timeliness baseline for the corresponding route and transportation scenario, outputting the basic timeliness prediction value for the corresponding waybill. For example, for the Shenzhen-Los Angeles sea freight route, historical sailing time data and route attribute data are extracted, and the standard sea freight timeliness baseline for this route is calculated using the gradient boosting tree model, outputting the basic timeliness prediction value for the corresponding batch of goods.

[0065] Step e2: Extract real-time time-series status data from the logistics digital twin model, and output timeliness deviation correction values ​​based on the real-time time-series status data using long short-term memory networks and attention mechanisms.

[0066] Furthermore, Long Short-Term Memory (LSTM) networks are recurrent neural networks suitable for processing long-term time-series data. They can capture long-term dependencies in time-series data, and their attention mechanisms can automatically identify the time-series segments that have the greatest impact on prediction results and increase their corresponding weights. Real-time time-series status data within a preset time period is extracted from the logistics digital twin model. This real-time time-series status data includes continuous time-series information such as speed sequence data, dwell time event data, and location update frequency data. The real-time time-series status data is input into the LSM network, which performs sequence encoding on the time-series data, capturing the trend of deviation changes between the current transportation status and the standard timeliness baseline. Simultaneously, the attention mechanism assigns different attention weights to different time segments, focusing on capturing time-series features that significantly affect timeliness, such as recent long dwell times and abnormal speed fluctuations. Finally, it calculates and outputs a timeliness deviation correction value. For example, when a shipping vessel experiences a long berthing wait in a port area, the LSM network captures the time-series feature of a dwell time exceeding the normal level. The attention mechanism increases the weight of this dwell time period, calculating the corresponding timeliness deviation correction value, which is used to adjust the basic timeliness prediction results.

[0067] Step e3: Extract external environment feature data from the logistics digital twin model, and use a Bayesian network model to output the timeliness adjustment amount of external factors based on the external environment feature data.

[0068] Furthermore, the Bayesian network model is a causal analysis model based on probabilistic inference. It represents the causal relationships between variables using a directed acyclic graph (DAG), enabling the quantification of the causal impact of different external factors on timeliness. External environmental feature data corresponding to the current transportation route is extracted from the logistics digital twin model. This data includes various external influencing factors such as weather forecasts, traffic congestion, port congestion, and holiday markers. These external environmental feature data are input into the Bayesian network model. Based on a pre-trained probabilistic network structure and conditional probability parameters, the model infers and calculates the independent impact of each type of external factor on transportation timeliness. It then integrates the combined effect of all external factors and outputs the timeliness adjustment caused by the external factors. For example, if heavy rainfall occurs in the area traversed by the transportation route and the corresponding port is under high congestion, the Bayesian network model calculates the impact of the rainfall and port congestion on timeliness separately, and outputs the corresponding timeliness adjustment based on the combined effect of these two factors.

[0069] Step e4: Based on the basic timeliness forecast value, the timeliness deviation correction value, and the timeliness adjustment amount of external factors, generate the cargo forecast status and the corresponding confidence interval.

[0070] Furthermore, following pre-defined fusion calculation rules, the basic time-delivery forecast value, the time-delivery deviation correction value, and the time-delivery adjustment amount due to external factors are superimposed to obtain the final predicted arrival time of the goods, which serves as the core component of the predicted cargo status. Simultaneously, the confidence interval corresponding to the prediction result is calculated. The confidence interval is calculated based on basic uncertainty parameters and scenario risk factors. Scenario risk factors are dynamically calculated across multiple dimensions, including weather risk, traffic risk, and historical delay rates. The higher the risk level, the wider the confidence interval, and the higher the uncertainty of the corresponding prediction result. The predicted cargo status also includes multi-dimensional information such as predicted arrival time, predicted transit points, and predicted operational status, which are output along with the confidence interval. For example, the basic time-delivery forecast value for a shipping route is superimposed with the time-delivery deviation correction value caused by port stays, and then superimposed with the time-delivery adjustment amount due to external factors caused by heavy rainfall to obtain the final predicted arrival time of the goods. Simultaneously, a corresponding confidence interval range is generated by combining this with the comprehensive risk level of the current shipping scenario.

[0071] Step S106: Collect actual logistics status, and based on the actual logistics status, the predicted status of goods and the corresponding confidence interval, identify abnormal events, abnormal types and risk levels.

[0072] Specifically, step S106 includes: Step f1: Collect actual logistics status data and calculate the status deviation based on the actual logistics status and the predicted cargo status.

[0073] Furthermore, real-time actual logistics status data is collected through a full-link data acquisition channel. This data includes real-time location data, status node data, sensor monitoring data, event-triggered data, and other comprehensive operational data. Predicted cargo status data for the same time dimension of the corresponding waybill is extracted. The actual logistics status data and predicted cargo status data are compared item by item according to their respective dimensions. Deviation values ​​are calculated for timeliness, location, environmental parameters, and status nodes. The results of these deviations are then integrated to obtain a comprehensive status deviation. This status deviation quantifies the degree of deviation between the actual and predicted operational status. Each dimension corresponds to an independent deviation value, and the overall comprehensive deviation value is generated. For example, for goods in transit, the actual arrival time data and predicted arrival time data for a port node are collected to calculate the time dimension deviation. Simultaneously, the actual location coordinates and predicted location coordinates are collected to calculate the location deviation. The multi-dimensional deviations are then integrated to obtain the overall status deviation.

[0074] Step f2 involves matching and verifying the state deviation against the confidence interval. When the state deviation exceeds the confidence interval, it is determined to be an abnormal event, and the risk level is classified according to the deviation magnitude.

[0075] Furthermore, pre-defined rules for determining confidence intervals are established. A confidence interval represents the normal fluctuation range of the prediction result; deviations falling within the interval are considered acceptable normal fluctuations. The calculated state deviations for each dimension are matched and verified against the corresponding confidence intervals for the prediction results to determine whether the state deviation for each dimension falls within the confidence interval. When the state deviation for any dimension exceeds the upper or lower boundary of the confidence interval, the current operating state is determined to constitute an abnormal event. Simultaneously, based on the magnitude of the deviation exceeding the confidence interval, the abnormal event is classified into multiple risk levels: smaller deviations correspond to lower risk levels, and larger deviations correspond to higher risk levels. For example, when the actual arrival time of goods exceeds the upper limit of the predicted confidence interval, it is determined to be a timeliness anomaly event. Based on the magnitude of the excess time, it is classified into the corresponding risk level: smaller deviations correspond to lower risk levels, and larger deviations correspond to higher risk levels.

[0076] Step f3: Match the anomaly type based on the state deviation amount. The anomaly types include time delay, location loss, temperature and humidity exceeding limits, vibration anomaly, and illegal state jump.

[0077] Furthermore, based on the corresponding data dimensions generated by the state deviation, the corresponding anomaly type is matched. When the deviation in the timeliness dimension exceeds the confidence interval and exhibits lag characteristics, it is matched as a timeliness delay anomaly type. When the location dimension data has not been updated for a long time and the location deviation continues to increase, it is matched as a location disconnection anomaly type. When the actual values ​​collected by the temperature and humidity sensors deviate from the predicted normal range and exceed the threshold, it is matched as a temperature and humidity exceeding limit anomaly type. When the vibration amplitude monitored by the vibration sensor exceeds the normal threshold and the corresponding deviation exceeds the standard, it is matched as a vibration anomaly type. When the flow sequence of the state node does not match the preset flow path and a reverse jump occurs, it is matched as an illegal state jump anomaly type. For example, in cold chain transportation, if the actual temperature collected by the temperature and humidity sensors is consistently higher than the preset normal range, the corresponding state deviation is reflected in the temperature and humidity dimension, and it is matched as a temperature and humidity exceeding limit anomaly type, which is used for subsequent anomaly labeling and intervention processing.

[0078] Step S107: Obtain the current user's role data, permission data, and behavioral characteristic data, and generate a personalized visualization view based on the role data, permission data, behavioral characteristic data, actual logistics status, abnormal events, abnormal types, and risk levels.

[0079] In some alternative embodiments, Figure 6This is a closed-loop flowchart of anomaly warning and one-click intervention based on a multi-source data fusion-based logistics end-to-end visualization method according to embodiments of the present invention. It takes the actual collected GPS, AIS, and sensor status data and the predicted status data output by the digital twin model as input. First, it performs a step-by-step comparison of the two types of statuses and calculates the deviation. If the deviation does not exceed the threshold, it maintains normal monitoring mode and continues monitoring; if the deviation exceeds the threshold, it is judged as an abnormal event. Subsequently, based on the deviation magnitude and impact range, it divides the warning levels into four levels from low to high. Anomaly identification rendering is simultaneously completed in three carriers of the visualization rendering layer: map view, Gantt chart, and dashboard. The warnings are then communicated via SMS, email, and App. Anomaly notifications are sent simultaneously through multiple channels. Next, the decision dashboard in the interactive intervention layer is displayed, showing an anomaly overview, cause analysis, alternative solutions, impact assessment, and relevant contact information to support handling decisions. Then, a one-click intervention executor provides three standardized handling paths: rescheduling, contacting customers, and switching lines, supporting one-click triggering to automatically complete the entire handling process. Finally, the feedback optimization module iterates backward from the prediction model, strategy library, and detection threshold from three dimensions: prediction deviation, intervention effect, and anomaly detection, and feeds back to the front-end model along the closed loop, forming a sustainable and self-optimizing full-link anomaly control closed loop.

[0080] Specifically, step S107 includes: Step g1: Use the interactive interface to obtain the current user's role data, permission data, and behavioral characteristic data.

[0081] Furthermore, when a user completes login authentication through the interactive interface, the current user's role and permission data are simultaneously extracted from the user's account configuration information. Role data corresponds to the user's identity and positioning within the logistics business system, while permission data corresponds to the user's accessible data range and executable operation permissions. Behavioral feature data is generated based on the user's historical operation logs, extracting user actions such as function clicks, filter setting records, information viewing mode records, and dwell time records for each functional module within the interactive interface, organizing these into structured behavioral feature data. The interactive interface automatically triggers the data acquisition process after successful user login, retrieving the current user's role and permission data in real time, while simultaneously retrieving the user's historical operation records to calculate and generate the latest behavioral feature data. For example, when a cargo owner user logs into the visual interactive interface, the cargo owner's role data and corresponding order viewing permission data are automatically extracted from the account configuration, while simultaneously retrieving the user's operation records from past periods to generate behavioral feature data, completing the synchronous acquisition of these three types of data.

[0082] Step g2 involves matching a preset basic view template based on the current user's role and permission data to limit the data range and information granularity displayed in the view.

[0083] Furthermore, the preset basic view templates are standardized view component sets pre-configured for different business roles. Different templates contain differentiated functional components, data fields, and display layouts. Based on the current user's role data, the preset view template library is traversed to match the corresponding basic view template, generating an initial view framework. Based on the current user's permission data, the data display scope within the initial view framework is filtered and limited, eliminating all data content outside the permission range. Simultaneously, the granularity of information display is adjusted according to role type: management roles correspond to macro-level summary information granularity, while frontline operational roles correspond to detailed operational information granularity. After matching and filtering, a basic view adapted to the current role and permissions is formed, ensuring that different roles can only view logistics data within their permission range, and that the information granularity matches the role's work requirements. For example, the dispatch supervisor role matches a dispatch supervisor-specific basic view template. Based on the jurisdictional area permission limitation, the view only displays waybill data and vehicle resource data within the responsible area, hiding operational details outside the jurisdiction. The information granularity is set to regional-level summary with drill-down capability for single waybills.

[0084] Step g3: Based on the current user's behavioral characteristic data, perform behavioral sequence modeling using a Hidden Markov Model to dynamically adjust the component layout and functional priority within the view.

[0085] Furthermore, Hidden Markov Models (HMMs) are statistical models used for probabilistic modeling of time series sequences. They learn state transition patterns based on historical behavioral sequences and infer user operational tendencies. By inputting current user behavior feature data into the HMM, the user's historical operation sequences are encoded into observation sequences recognizable by the model. Based on a pre-trained state transition probability matrix, the model infers the user's current behavioral state and the most likely subsequent operation, identifying frequently used functional components and information viewing preferences. Based on the model's output of behavioral preferences, the arrangement of functional components within the view is dynamically adjusted. Frequently used components are moved to prominent positions and given higher priority, while infrequently used components are collapsed, optimizing the view's operational path. For example, a frontline dispatcher might prioritize viewing a list of pending exceptions and performing batch approvals during fixed daily time slots. After recognizing this regular behavioral sequence, the HMM automatically adjusts the pending exception components to the top area of ​​the view, preloads the corresponding exception list data, and reduces the user's operational steps.

[0086] Step g4: Based on the actual logistics status, abnormal events, and risk levels, generate multi-format visualization content, including map trajectory view, Gantt chart timeline view, operation dashboard view, and topology network view.

[0087] Furthermore, based on real-time actual logistics status, abnormal event data, and risk level data, four types of visualization content are generated, and the display can be switched according to the user's view scenario. The map trajectory view, rendered by a geographic information system map engine, dynamically marks in-transit cargo logistics entities on the map at their corresponding locations. Completed transportation routes are drawn with solid lines, while unexecuted predicted routes are drawn with dashed lines. The display style is adjusted accordingly for abnormal transportation segments. The Gantt chart timeline view uses a timeline as a benchmark to compare the planned and actual times of each logistics node, with different display styles corresponding to different operating states. The operations dashboard view summarizes and displays core logistics operation indicators, presenting an overall operational overview. The topology network view displays the connection relationships between facility nodes and transportation routes in the logistics network. Node size corresponds to throughput, and the thickness of connecting lines corresponds to cargo volume. Different levels of congestion correspond to different display styles. For example, when management views the overall operational situation, an operations dashboard view is generated to display the total number of in-transit waybills and on-time rate indicators for the entire region. It also supports switching to the topology network view, which intuitively presents the congestion distribution of each port and terminal. Figure 7 This invention presents a five-layer architecture data flow and time-series interaction diagram based on the multi-source data fusion-based logistics end-to-end visualization method. It divides the data flow into nine participating entities using vertical swimlanes: data source, data access layer, data processing and fusion layer, digital twin modeling layer, ETA prediction engine, visualization rendering layer, anomaly detection module, interactive intervention layer, and feedback optimization module. It fully presents the time-series interaction logic of the positive data flow and the reverse optimization closed loop. In the forward flow, raw data generated by GPS devices, business systems, and external interfaces is first fed into the data access layer to aggregate multi-source data. This data is then transmitted down the message bus to the data processing and fusion layer, where it undergoes data cleaning, spatiotemporal alignment, and feature fusion processing to output standardized feature vectors. These vectors are then fed into the digital twin modeling layer to complete logistics entity modeling and full-link state mirroring. Subsequently, the ETA prediction engine dynamically calculates the estimated arrival time of goods and corresponding confidence intervals based on the digital twin model. The prediction results are synchronously transmitted to the visualization rendering layer, where corresponding view templates are matched based on user roles, generating multi-format visualization content that is pushed to the front end. Simultaneously, the anomaly detection module compares the actual state with the predicted state to determine and classify anomalies. When no anomalies are detected, continuous monitoring and visualization updates are maintained. When an anomaly is triggered, the anomaly data is pushed to the interactive intervention layer, where a decision dashboard is displayed and one-click intervention is executed. In the reverse closed loop, the feedback optimization module collects prediction deviation data and intervention execution effect data, transmitting them back along the dotted line to the digital twin modeling layer and the ETA prediction engine. This completes iterative optimization of model parameters and detection strategies, forming a self-iteratory full-link data flow interaction system.

[0088] The present invention provides the following embodiments to adapt to different implementation scenarios: Specific Implementation Example 1 addresses the need for end-to-end visualization of cold chain cargo across maritime transport links, implementing a full-process visualization processing flow. In the data access phase, GPS data from transport vehicles and containers, along with temperature and humidity sensor data, are collected via message queue telemetry transmission protocol. For port-inherited warehouse management systems, database log change capture is used to access warehouse operation data. Meteorological data from the transit sea areas and operational data from the destination port are collected through standard service interfaces. Port loading / unloading events and customs clearance events are collected using a domain event format. In the data cleaning phase, an extended Kalman filter algorithm is used to process GPS data, adapting it to the nonlinear motion models of ships and vehicles. Position jumps caused by signal obstruction are removed, missing data segments corresponding to signal blind spots are marked and completed using linear interpolation, and outlier detection is performed on temperature and humidity sensor data to remove abnormal readings caused by temporary sensor malfunctions. In the spatiotemporal alignment phase, spatiotemporal alignment is implemented based on a stream processing framework. Using a unified waybill number as the primary key, multi-source data is sorted by event time dimension through a sliding time window, and a watermarking mechanism is configured to handle delayed port event data. In the feature fusion stage, temporal features, statistical features, cross features, and event sequence features are extracted. A two-layer feedforward neural network is used to calculate the fusion weights for each feature dimension, generating a weighted first fused feature vector. In the digital twin modeling stage, four types of digital twin entities—cargo, transport vessels, port facilities, and personnel—are constructed. An attribute graph model is used to establish a relationship graph between entities. Entity states are synchronized in real-time, and concurrent updates are handled through a version vector mechanism, creating a full-link state mirror. In the timeliness prediction stage, a gradient boosting tree model outputs the basic timeliness prediction value for the sea transport segment. A long short-term memory network combined with an attention mechanism outputs the timeliness deviation correction value caused by port delays. A Bayesian network model outputs the timeliness adjustment amount corresponding to external factors such as typhoon weather. These are fused to generate the predicted cargo status and confidence interval. In the anomaly identification stage, the actual logistics status is compared with the predicted cargo status to calculate the status deviation. When the deviation exceeds the confidence interval, it is judged as an abnormal event. Risk levels are classified based on the deviation magnitude, and two anomaly types are matched: temperature and humidity exceeding limits and timeliness delays. In the visualization generation phase, corresponding view templates are matched based on user role data to generate personalized visualization views that include temperature and humidity monitoring curves and estimated arrival times. Status updates are pushed in real time through a long-connection channel. In the feedback and optimization phase, after the goods are signed for, the predicted and actual values ​​are compared to calculate the deviation, and the deviation is fed back to the prediction model for parameter fine-tuning.

[0089] Specific Implementation Example 2 addresses multimodal transport scenarios involving road-to-sea-to-rail-to-road transport, implementing a full-process visualization and anomaly intervention handling workflow. In the data access phase, time-series data and business system data from the entire transport chain are accessed through multiple protocols, collecting discrete event data from each node and external environment data. In the data cleaning and spatiotemporal alignment phase, after cleaning and standardizing multi-source data, data is sorted using the event time dimension. Handover event pairs consisting of unloading completion events and loading completion events are constructed. Valid handover event pairs are matched based on heuristic matching rules using time windows and spatial distance, automatically stitching together the transport chain data from different modes of transport. The legality of entity state transitions is verified through a preset state machine, and finally, time dimension alignment is achieved through a sliding time window to obtain third-party logistics data. In the feature fusion and digital twin modeling phase, multi-dimensional features are extracted and weighted through an attention mechanism to generate a first fusion feature vector. Based on this first fusion feature vector, a full-chain digital twin model is constructed, synchronizing the entity states of each link in the multimodal transport. In the timeliness prediction phase, the predicted cargo status and confidence interval for each transport segment are calculated using a multi-model fusion algorithm. In the anomaly identification phase, when handover delays occur in railway transfer links, the deviation between the actual and predicted states is calculated. If the deviation exceeds the confidence interval, it is identified as an anomaly event. Based on the deviation magnitude, corresponding risk levels are assigned, and the time delay anomaly type is matched. In the decision-making and intervention phase, a decision dashboard is generated after the anomaly event is triggered, aggregating an anomaly overview, root cause analysis, recommended solutions, impact assessment, and relevant contact information. Integer programming is used to solve for the optimal solution among alternative solutions, generating three alternative solutions: rescheduling, switching stations, and waiting for recovery. A workflow engine enables automatic orchestration of a one-click intervention process. After the user triggers the intervention command, the entire process of solution verification, command generation, system synchronization, and notification to relevant parties is automatically executed. In the feedback and optimization phase, after the intervention is completed, the intervention effect data is recorded and fed back to the anomaly handling strategy library for optimization and updates.

[0090] Specific Implementation Example 3 addresses the scenario of a logistics operation center where multiple roles are used simultaneously, implementing a personalized visualization generation and multi-terminal synchronization process. In the user data acquisition phase, after a user logs in through the interactive interface, their role data, permission data, and behavioral characteristic data are obtained. The behavioral characteristic data includes four dimensions: function click frequency, filter preference, information viewing mode, and module dwell time. In the basic view matching phase, a hybrid mapping strategy combining rules and machine learning is adopted. First, a preset basic view template is matched based on the role data and permission data, limiting the data display scope and information granularity of the view. The shipper role corresponds to an order progress view, the dispatcher role to a waybill management view, the driver role to a task navigation view, and the warehouse manager role to a job execution view. In the dynamic view adjustment phase, a Hidden Markov Model is used to model the user behavior sequence. The Baum-Welch algorithm is used to learn the state transition probability matrix online, predicting the user's operational tendencies and next needs. The layout of components and functional priorities within the view are dynamically adjusted, with frequently used components automatically placed on top and infrequently used components automatically collapsed. In the multi-format visualization generation stage, based on the actual logistics status, abnormal events, and risk levels, map trajectory views, Gantt chart timeline views, operation dashboard views, and topology network views are generated. Figure Four For visualization-like content, grid aggregation and viewport clipping strategies are employed to ensure rendering performance in scenarios with large data volumes. During the real-time push phase, a publish-subscribe message distribution mechanism is used, establishing a long-connection push channel for each online user. Environments that do not support long connections automatically downgrade to a server-side event push channel. Incremental update protocols are used to push state change data, ensuring that multiple devices receive consistent view updates synchronously. In the anomaly synchronization phase, after an anomaly event is triggered, views from different roles synchronously receive anomaly annotation information within their respective permission ranges, achieving real-time collaboration on anomaly states across all roles.

[0091] The logistics end-to-end visualization method based on multi-source data fusion provided in this embodiment has the following beneficial effects: First, by collecting first-stage logistics data covering time-series data, business system data, external environment data, and discrete event data from multiple dimensions, it fully covers the equipment perception dimension, business process dimension, external environment dimension, and node event dimension of the entire logistics chain. This builds a multi-dimensional data foundation for subsequent end-to-end visualization, avoiding the problem of incomplete information coverage caused by a single data source. The four types of data correspond to the equipment operating status, business flow nodes, external influencing factors, and discrete event markers in the logistics process, respectively. This allows the logistics operation process in the physical world to be fully mapped through multi-dimensional data, ensuring that the state changes of each link in the logistics process have corresponding data carriers. Second, by performing classification cleaning and format standardization processing on the first-stage logistics data, targeted abnormal data removal and missing value completion are carried out according to the characteristics of data from different sources. The encoding rules and format specifications of various types of data are unified, eliminating data format differences and encoding barriers between different business systems and equipment. Classification cleaning can specifically handle typical anomalies of different types of data, avoiding invalid data from interfering with subsequent processing. Format standardization can provide a unified parsing basis for multi-source heterogeneous data, improving the accuracy and consistency of the original data. Then, by using a unified waybill number as the primary key and employing a sliding time window to perform spatiotemporal alignment on the second logistics data, data from different sources and with different reporting frequencies are systematically integrated according to the event time dimension. This achieves precise matching of multi-source data in the time dimension, resolving the issues of asynchronous data reporting times and out-of-order data arrival across different systems. Using a unified waybill number as the primary key allows data scattered across various systems to aggregate around the same logistics object. The sliding time window alignment method adapts to differences in data reporting frequencies, ensuring that data with different update frequencies can form a coherent time series on a unified time axis. Furthermore, by performing feature extraction and attention-based weighted fusion on the third logistics data, deep representational information of the logistics data can be mined from multiple dimensions, including temporal features, statistical features, cross features, and event sequence features. Simultaneously, the fusion weights of different feature dimensions are dynamically allocated based on the current logistics scenario context, allowing features more relevant to the current scenario to occupy a higher proportion during the fusion process. Multi-dimensional feature extraction can comprehensively capture the temporal patterns, statistical characteristics, cross-influences, and event sequence logic of logistics operations. The weighting method of the attention mechanism can improve the adaptability of feature vectors to different logistics scenarios, generating a first fused feature vector that accurately reflects the actual operational status of logistics. Subsequently, by constructing a logistics digital twin model based on the first fused feature vector, various entities in the entire logistics chain can be digitally abstracted, the attribute fields, state sets, and standardized state transition paths of various entities can be clearly defined, a network of relationships between various entities can be built, and the real-time synchronization of entity states and the establishment of a full-chain state mirror can be achieved.Digital twin models can completely replicate the operation process and entity relationships of physical logistics in digital space, allowing for synchronous mapping of changes in the state, flow, and interaction logic of logistics entities. Furthermore, based on the logistics digital twin model, multi-model fusion prediction algorithms dynamically calculate the predicted state of goods and corresponding confidence intervals. Timeliness calculations are performed from multiple dimensions, including baseline timeliness, real-time deviation, and external influences. The computational advantages of different models are combined to output the final prediction result, while generating corresponding confidence intervals to reflect the uncertainty of the prediction results. This multi-model fusion approach balances the stability of baseline timeliness, the sensitivity of real-time status, and the correlation with external factors. This ensures that the prediction results are supported by historical statistical patterns while responding to dynamic changes in real-time operation and the external environment. The output of the confidence interval quantifies the reliability of the prediction results. Furthermore, by collecting actual logistics data, and based on the actual logistics status, predicted cargo status, and corresponding confidence intervals, abnormal events, anomaly types, and risk levels can be identified. This allows for a quantitative comparison between the real-time collected physical logistics status and the predicted status output by the model. Using the confidence interval as the baseline boundary for anomaly judgment, it accurately identifies abnormal events deviating from the normal operating range, while simultaneously matching corresponding anomaly types and classifying risk levels. The comparison between real-time and predicted status provides a dynamic reference standard for anomaly judgment, rather than relying on fixed thresholds. The classification of anomaly types and risk levels enables precise location and hierarchical control of abnormal events. Finally, by acquiring the current user's role data, permission data, and behavioral characteristic data, and based on this data, actual logistics status, abnormal events, anomaly types, and risk levels, personalized visualization views are generated. These views are adapted to the different user identities, permissions, and operating habits, adjusting the data display scope and information granularity accordingly, while simultaneously integrating real-time logistics status and anomaly information to generate multi-format visualization content. Role and permission matching ensures that different users only access logistics information within their authorized scope. The introduction of behavioral features allows view layout and functional priorities to align with user habits. Multi-format visualization content adapts to information viewing needs in different scenarios, enabling different users to obtain logistics information matching their own needs, avoiding information overload or insufficiency, and improving the adaptability and usability of the visualization system. By implementing this solution, the shortcomings of existing logistics visualization technologies—such as insufficient data integration, lack of predictive capabilities, fragmented anomaly response, incomplete multimodal transport coverage, and poor real-time performance and role adaptability—are addressed, thus failing to meet the actual needs of intelligent management and control throughout the modern logistics process.

[0092] The above are embodiments of the logistics end-to-end visualization method based on multi-source data fusion provided in this application. Other embodiments of the logistics end-to-end visualization method based on multi-source data fusion provided in this application are described below.

[0093] This invention also discloses a logistics end-to-end visualization system based on multi-source data fusion, such as... Figure 8 As shown, the system includes: The acquisition module is used to acquire the first logistics data, which includes time-series data, business system data, external environment data, and discrete event data.

[0094] The standardization module is used to perform classification, cleaning, and format standardization processing on the first logistics data to obtain the second logistics data.

[0095] The alignment module is used to perform spatiotemporal alignment of the second logistics data using a unified waybill number as the primary key and a sliding time window to obtain the third logistics data.

[0096] The fusion module is used to extract features from third-party logistics data and perform weighted fusion using an attention mechanism to generate a first fused feature vector. A logistics digital twin model is then constructed based on this first fused feature vector.

[0097] The prediction module is used to dynamically calculate the predicted status of goods and the corresponding confidence interval based on the logistics digital twin model and using a multi-model fusion prediction algorithm.

[0098] The identification module is used to collect actual logistics status and, based on the actual logistics status, the predicted status of goods, and the corresponding confidence interval, to identify abnormal events, abnormal types, and risk levels.

[0099] The visualization module is used to acquire the current user's role data, permission data, and behavioral characteristic data, and generate personalized visualization views based on the role data, permission data, behavioral characteristic data, actual logistics status, abnormal events, abnormal types, and risk levels.

[0100] The logistics end-to-end visualization system based on multi-source data fusion provided in this embodiment has the following beneficial effects: First, by collecting first-stage logistics data covering time-series data, business system data, external environment data, and discrete event data from multiple dimensions, it fully covers the equipment perception dimension, business process dimension, external environment dimension, and node event dimension of the entire logistics chain. This provides a multi-dimensional data foundation for subsequent end-to-end visualization, avoiding the problem of incomplete information coverage caused by a single data source. The four types of data correspond to the equipment operating status, business flow nodes, external influencing factors, and discrete event markers in the logistics process, respectively. This allows the logistics operation process in the physical world to be fully mapped through multi-dimensional data, ensuring that the state changes of each link in the logistics process have corresponding data carriers. Second, by performing classification cleaning and format standardization processing on the first-stage logistics data, targeted abnormal data removal and missing value completion are carried out according to the characteristics of data from different sources. The coding rules and format specifications of various types of data are unified, eliminating data format differences and coding barriers between different business systems and equipment. Classification cleaning can specifically handle typical anomalies of different types of data, avoiding invalid data from interfering with subsequent processing. Format standardization enables multi-source heterogeneous data to have a unified parsing basis, improving the accuracy and consistency of the original data. Then, by using a unified waybill number as the primary key and employing a sliding time window to perform spatiotemporal alignment on the second logistics data, data from different sources and with different reporting frequencies are systematically integrated according to the event time dimension. This achieves precise matching of multi-source data in the time dimension, resolving the issues of asynchronous data reporting times and out-of-order data arrival across different systems. Using a unified waybill number as the primary key allows data scattered across various systems to aggregate around the same logistics object. The sliding time window alignment method adapts to differences in data reporting frequencies, ensuring that data with different update frequencies can form a coherent time series on a unified time axis. Furthermore, by performing feature extraction and attention-based weighted fusion on the third logistics data, deep representational information of the logistics data can be mined from multiple dimensions, including temporal features, statistical features, cross features, and event sequence features. Simultaneously, the fusion weights of different feature dimensions are dynamically allocated based on the current logistics scenario context, allowing features more relevant to the current scenario to occupy a higher proportion during the fusion process. Multi-dimensional feature extraction can comprehensively capture the temporal patterns, statistical characteristics, cross-influences, and event sequence logic of logistics operations. The weighting method of the attention mechanism can improve the adaptability of feature vectors to different logistics scenarios, generating a first fused feature vector that accurately reflects the actual operational status of logistics. Subsequently, by constructing a logistics digital twin model based on the first fused feature vector, various entities in the entire logistics chain can be digitally abstracted, the attribute fields, state sets, and standardized state transition paths of various entities can be clearly defined, a network of relationships between various entities can be built, and the real-time synchronization of entity states and the establishment of a full-chain state mirror can be achieved.Digital twin models can completely replicate the operation process and entity relationships of physical logistics in digital space, allowing for synchronous mapping of changes in the state, flow, and interaction logic of logistics entities. Furthermore, based on the logistics digital twin model, multi-model fusion prediction algorithms dynamically calculate the predicted state of goods and corresponding confidence intervals. Timeliness calculations are performed from multiple dimensions, including baseline timeliness, real-time deviation, and external influences. The computational advantages of different models are combined to output the final prediction result, while generating corresponding confidence intervals to reflect the uncertainty of the prediction results. This multi-model fusion approach balances the stability of baseline timeliness, the sensitivity of real-time status, and the correlation with external factors. This ensures that the prediction results are supported by historical statistical patterns while responding to dynamic changes in real-time operation and the external environment. The output of the confidence interval quantifies the reliability of the prediction results. Furthermore, by collecting actual logistics data, and based on the actual logistics status, predicted cargo status, and corresponding confidence intervals, abnormal events, anomaly types, and risk levels can be identified. This allows for a quantitative comparison between the real-time collected physical logistics status and the predicted status output by the model. Using the confidence interval as the baseline boundary for anomaly judgment, it accurately identifies abnormal events deviating from the normal operating range, while simultaneously matching corresponding anomaly types and classifying risk levels. The comparison between real-time and predicted status provides a dynamic reference standard for anomaly judgment, rather than relying on fixed thresholds. The classification of anomaly types and risk levels enables precise location and hierarchical control of abnormal events. Finally, by acquiring the current user's role data, permission data, and behavioral characteristic data, and based on this data, actual logistics status, abnormal events, anomaly types, and risk levels, personalized visualization views are generated. These views are adapted to the different user identities, permissions, and operating habits, adjusting the data display scope and information granularity accordingly, while simultaneously integrating real-time logistics status and anomaly information to generate multi-format visualization content. Role and permission matching ensures that different users only access logistics information within their authorized scope. The introduction of behavioral features allows view layout and functional priorities to align with user habits. Multi-format visualization content adapts to information viewing needs in different scenarios, enabling different users to obtain logistics information matching their own needs, avoiding information overload or insufficiency, and improving the adaptability and usability of the visualization system. By implementing this solution, the shortcomings of existing logistics visualization technologies—such as insufficient data integration, lack of predictive capabilities, fragmented anomaly response, incomplete multimodal transport coverage, and poor real-time performance and role adaptability—are addressed, thus failing to meet the actual needs of intelligent management and control throughout the modern logistics process.

[0101] In this embodiment, the logistics end-to-end visualization system based on multi-source data fusion is presented in the form of functional units. Here, a unit refers to an ASIC (Application Specific Integrated Circuit), a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

Claims

1. A method for end-to-end logistics visualization based on multi-source data fusion, characterized in that, The method includes: Acquire first logistics data, which includes time-series data, business system data, external environment data, and discrete event data; The first logistics data is classified, cleaned, and standardized to obtain the second logistics data. Using the unified waybill number as the primary key, the second logistics data is spatiotemporally aligned using a sliding time window to obtain the third logistics data. The third logistics data is subjected to feature extraction and weighted fusion using an attention mechanism to generate a first fused feature vector, and a logistics digital twin model is constructed based on the first fused feature vector; Based on the aforementioned logistics digital twin model, a multi-model fusion prediction algorithm is used to dynamically calculate the predicted status of goods and the corresponding confidence interval. Collect actual logistics status, and based on the actual logistics status, the predicted status of goods, and the corresponding confidence interval, identify abnormal events, abnormal types, and risk levels; Obtain the current user's role data, permission data, and behavioral characteristic data, and generate a personalized visualization view based on the role data, permission data, behavioral characteristic data, actual logistics status, abnormal events, abnormal types, and risk levels.

2. The method based on claim 1, characterized in that, The acquisition of the first logistics data includes: The system collects time-series data using IoT device access protocols. The time-series data includes GPS positioning data and vehicle sensor data, and supports offline caching and breakpoint retransmission mechanisms. The system utilizes application programming interfaces (APIs) and message queues to connect with business system data, which includes business status data for transportation, warehousing, orders, and customs clearance. External environmental data is collected using standard service interfaces, including meteorological data, real-time road operation data, real-time port operation data, and real-time railway operation data. Discrete event data is collected using a domain event format, including port node status change events, customs node status change events, and warehouse node status change events.

3. The method based on claim 2, characterized in that, The process of classifying, cleaning, and standardizing the format of the first logistics data to obtain the second logistics data includes: Kalman filtering algorithm is used to remove drift points and jump points from GPS positioning data in time series data, and missing data segments corresponding to communication blind spots are marked. Standardize waybill coding rules and data format for business system data, and interpolate missing status fields based on the context of business logic; Spatial interpolation is performed on the external environment data to match the gridded environmental data to the latitude and longitude points corresponding to the logistics path.

4. The method based on claim 3, characterized in that, The process of using a unified waybill number as the primary key and employing a sliding time window to perform spatiotemporal alignment on the second logistics data to obtain the third logistics data includes: The second logistics data is sorted according to the event time dimension; Construct a handover event pair consisting of unloading completion event and loading completion event, and stitch and sort the second logistics data based on the handover event pair; The state transitions of entities in the stitched second logistics data are verified using a preset state machine, and illegal state transition data are marked. By aligning the second logistics data with the time dimension after inspection using a sliding time window, the third logistics data is obtained.

5. The method based on claim 4, characterized in that, The step of extracting features from the third logistics data and weighting and fusing them using an attention mechanism to generate a first fused feature vector includes: Temporal features, statistical features, cross features, and event sequence features are extracted from the third logistics data, and the event sequence features are generated by encoding using an event vectorization method. Based on the current logistics scenario context, a feedforward neural network is used to calculate the fusion weights corresponding to time-series features, statistical features, cross features, and event sequence features. Based on the fusion weights, the temporal features, statistical features, cross features, and event sequence features are weighted and summed to generate the first fusion feature vector.

6. The method based on claim 5, characterized in that, The construction of the logistics digital twin model based on the first fused feature vector includes: Entity attribute data and state data are extracted from the first fused feature vector. Cargo logistics entities, transportation vehicle logistics entities, facility logistics entities and personnel logistics entities are digitally abstracted. Attribute fields and state sets of cargo logistics entities, transportation vehicle logistics entities, facility logistics entities and personnel logistics entities are defined respectively. Standardized state transition paths of entities are established to generate digital twin entity objects. Entity association data is extracted from the first fused feature vector, and an association network between cargo logistics entities, transportation vehicle logistics entities, facility logistics entities and personnel logistics entities is constructed based on the attribute graph model to form an entity association graph of the logistics digital twin model. The association network includes loading relationship, docking relationship, responsibility relationship and batch relationship. Based on the real-time updated data of the first fused feature vector, the real-time status of cargo logistics entities, transportation logistics entities, facility logistics entities, and personnel logistics entities is updated synchronously. Concurrent status updates are processed through a version vector mechanism to establish a full-link status mirror corresponding to the physical logistics process, thereby obtaining a logistics digital twin model.

7. The method based on claim 6, characterized in that, The method of dynamically calculating the predicted status of goods and the corresponding confidence interval based on the logistics digital twin model using a multi-model fusion prediction algorithm includes: Historical timeliness data and road segment attribute data are extracted from the logistics digital twin model. The standard timeliness baseline is calculated based on the historical timeliness data and road segment attribute data using the gradient boosting tree model, and the basic timeliness prediction value is output. Real-time time-series status data is extracted from the logistics digital twin model, and timeliness deviation correction values ​​are output based on the real-time time-series status data using long short-term memory networks and attention mechanisms. External environment feature data is extracted from the logistics digital twin model, and the timeliness adjustment amount of external factors is output based on the external environment feature data using a Bayesian network model. Based on the basic timeliness prediction value, the timeliness deviation correction value, and the timeliness adjustment amount of external factors, the cargo prediction status and the corresponding confidence interval are generated.

8. The method based on claim 7, characterized in that, The process of collecting actual logistics status data, and based on this actual logistics status, predicted cargo status, and corresponding confidence intervals, identifying abnormal events, anomaly types, and risk levels, includes: Collect actual logistics status data, and calculate the status deviation based on the actual logistics status and the predicted cargo status. The state deviation is matched and verified with the confidence interval. When the state deviation exceeds the confidence interval, it is determined to be an abnormal event, and the risk level is classified according to the deviation magnitude. Based on the state deviation quantity matching anomaly type, the anomaly type includes time delay, location loss, temperature and humidity exceeding limits, vibration anomaly, and illegal state jump.

9. The method based on claim 8, characterized in that, The process of acquiring the current user's role data, permission data, and behavioral characteristic data, and generating a personalized visualization view based on the role data, permission data, behavioral characteristic data, actual logistics status, abnormal events, and risk levels, includes: Use the interactive interface to obtain the current user's role data, permission data, and behavioral characteristic data; Based on the current user's role and permission data, a preset basic view template is matched to limit the data range and information granularity displayed in the view. Based on the current user's behavioral characteristics data, a hidden Markov model is used to model behavioral sequences and dynamically adjust the component layout and functional priority within the view. Based on the actual logistics status, abnormal events, and risk levels, multi-format visualization content is generated, including map trajectory view, Gantt chart timeline view, operation dashboard view, and topology network view.

10. A logistics end-to-end visualization system based on multi-source data fusion, characterized in that, The system includes: The acquisition module is used to acquire first logistics data, which includes time-series data, business system data, external environment data, and discrete event data. The standardization module is used to perform classification cleaning and format standardization processing on the first logistics data to obtain the second logistics data; The alignment module is used to perform spatiotemporal alignment of the second logistics data using a unified waybill number as the primary key and a sliding time window to obtain the third logistics data. The fusion module is used to extract features from the third logistics data and perform weighted fusion using an attention mechanism to generate a first fused feature vector; and to construct a logistics digital twin model based on the first fused feature vector. The prediction module is used to dynamically calculate the predicted status of goods and the corresponding confidence interval based on the logistics digital twin model and using a multi-model fusion prediction algorithm. The identification module is used to collect the actual logistics status and, based on the actual logistics status, the predicted cargo status, and the corresponding confidence interval, identify abnormal events, abnormal types, and risk levels. The visualization module is used to acquire the current user's role data, permission data, and behavioral characteristic data, and generate personalized visualization views based on the role data, permission data, behavioral characteristic data, actual logistics status, abnormal events, abnormal types, and risk levels.

Images