Virtual waybill generation method, system and device based on multi-source data fusion

Abstract

The application discloses a virtual waybill generation method, system and device based on multi-source data fusion, relates to the technical field of intelligent transportation and logistics informatization, and is used for solving the problems of incomplete information, poor authenticity and low automation degree of the existing waybill generation mode. The method comprises the following steps: acquiring target multi-source data, and performing fusion processing on the target multi-source data to generate loading and unloading event recognition results; performing retrieval enhancement generation and semantic reasoning according to the loading and unloading event recognition results, vehicle information and public information related to map interest points to generate goods category speculation results; performing fusion reasoning according to the loading and unloading event recognition results, the goods category speculation results and multi-dimensional influence data to generate freight income estimation results; and performing confidence fusion and structured processing based on the loading and unloading event recognition results, the goods category speculation results and the freight income estimation results to generate a virtual waybill.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and logistics information technology, and in particular to a method, system and equipment for generating virtual waybills based on multi-source data fusion. Background Technology

[0002] In the road freight industry, the generation of traditional paper waybills relies primarily on manual entry, which suffers from inefficiency, error-proneness, difficulty in verifying data authenticity, and a high rate of missing waybills. Existing technologies, such as some platforms, attempt to determine the origin and destination of transport using GPS trajectory data. However, this method has the following technical limitations: relying solely on latitude and longitude clustering makes it difficult to accurately identify factory or industrial park boundaries; it lacks the ability to precisely match the actual business scope of an enterprise with the type of goods; freight cost estimation is coarse, failing to fully consider factors such as transport routes, vehicle types, and seasons; and it cannot generate complete and structurally sound virtual waybills. These technical deficiencies hinder the implementation of automated supervision and services, limiting applications that rely on real waybill data, such as tax deductions, insurance claims, and green and low-carbon certifications.

[0003] Therefore, there is an urgent need to provide a more reliable virtual waybill generation solution based on multi-source data fusion. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, and device for generating virtual waybills based on multi-source data fusion, in order to solve the problems of incomplete information, poor authenticity, and low automation in existing waybill generation methods.

[0005] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for generating virtual waybills based on multi-source data fusion, the method comprising: Acquire target multi-source data; the target multi-source data includes at least vehicle trajectory data, map point of interest data, and electronic fence data; The target multi-source data is fused to generate loading and unloading event identification results; Based on the loading and unloading event identification results, vehicle information, and publicly available information associated with map points of interest, retrieval enhancement generation and semantic reasoning are performed to generate cargo category inference results; Based on the loading and unloading event identification results, the cargo category inference results, and multi-dimensional impact data, a freight revenue forecast result is generated through fusion reasoning; the multi-dimensional impact data includes at least regional economic indicators, fuel price data, vehicle model data, and seasonal fluctuation factors. Based on the loading and unloading event identification results, the cargo category inference results, and the freight revenue prediction results, confidence fusion and structured processing are performed to generate a virtual waybill.

[0006] Optionally, the target multi-source data is fused to generate loading and unloading event identification results, including: The vehicle trajectory data is analyzed using the DBSCAN clustering algorithm to identify candidate stopping areas. The candidate stopping areas are those where the stopping time exceeds a first preset threshold and the speed variation is lower than a second preset threshold. For each candidate stop point in the candidate stop area, perform electronic fence matching to determine the associated points of interest; Based on the associated points of interest, semantic verification of POIs is performed to obtain semantic information; Based on the semantic information and the load change information in the vehicle trajectory data, a multi-source fusion decision model is used to output the loading and unloading event identification result; the multi-source fusion decision model is a Bayesian network; the load change information represents the load change amount collected by the load sensor, and the load change amount is regarded as a loading and unloading event when it exceeds a preset percentage threshold; the loading and unloading event identification result includes at least the location coordinates, event type, confidence level, truck license plate, vehicle type, load capacity, departure time, and arrival time.

[0007] Optionally, for each candidate stop point in the candidate stop area, perform electronic fence matching to determine associated points of interest, including: For each candidate stop point, check whether it has entered the preset fence: if the candidate stop point is inside the fence and the stay time meets the preset time, then match the electronic fence icon; For candidate stops that do not match the electronic fence identifier, the boundary range of the electronic fence is dynamically adjusted according to the spatial distribution variance of historical trajectory data; and a dynamic electronic fence is generated for the candidate stops that do not match the electronic fence identifier based on the convex hull algorithm.

[0008] Optionally, based on the loading and unloading event identification results, vehicle information, and publicly available information associated with map points of interest, retrieval enhancement generation and semantic reasoning are performed to generate cargo category inference results, including: Construct a RAG vector knowledge base; the RAG vector knowledge base shall include at least publicly available cargo data, historical waybill patterns, target multi-source data, vehicle information, and cargo category inference results. Based on the loading and unloading event identification results and the vehicle information, a query vector is generated, and related documents are retrieved from the RAG vector knowledge base through RAG retrieval enhancement; the RAG retrieval enhancement integrates trajectory data, vehicle information and POI public cargo data, and retrieves Top-K related documents, and the Top-K related documents are embedded with BERT vectors; The associated documents, the loading and unloading event identification results, and the vehicle information are input into the large language model to generate the cargo category inference result and confidence level; the confidence level is obtained by fusing company cargo information associated with POI, historical waybill data, and vehicle qualification information based on the Bayesian method.

[0009] Optionally, based on the loading and unloading event identification results, the cargo category inference results, and multi-dimensional impact data, a fusion reasoning is performed to generate a freight revenue prediction result, including: The actual transportation distance is calculated based on the location coordinates in the loading and unloading event identification results. Based on the actual transportation distance, the predicted cargo category, and historical transaction price data, the benchmark freight rate is calculated using a pre-trained random forest model. The benchmark freight rate, regional economic indicators, fuel price data, vehicle model data, and seasonal fluctuation factor are input into the large language model. Combined with the contextual information enhanced by RAG retrieval, interval reasoning is performed to generate freight revenue prediction results. The seasonal fluctuation factor is obtained by querying the fixed seasonal factor table and the dynamic event table and fusing them according to the preset weight ratio.

[0010] Optionally, the process for determining the seasonal fluctuation factor includes: Fixed seasonal factors are determined based on the basic data of the fixed seasonal factor table; the basic data of the fixed seasonal factor table includes at least event tags within a preset time period, basic floating range, regional weighting, and commodity type correction; The dynamic event factor is calculated by weighting and averaging dynamic event data and combining it with an LLM model; the dynamic event data includes at least the trigger time, event description, impact coefficient, duration in days, and event source. The seasonal fluctuation factor is calculated by weighting and summing the fixed seasonal factor and the dynamic event factor according to preset weights.

[0011] Optionally, based on the loading and unloading event identification results, the cargo category estimation results, and the freight revenue prediction results, confidence fusion and structured processing are performed to generate a virtual waybill, including: Assign preset weights to key fields in the loading and unloading event identification results, the cargo category estimation results, and the freight revenue estimation results; Based on the pre-defined weights assigned to key fields and the confidence levels of each key field, the overall confidence level of the virtual waybill is calculated using a weighted geometric average algorithm. The confidence scores of fields with logical conflicts are attenuated and corrected according to a correction rule. The correction rule is to multiply the confidence scores of fields with logical conflicts by a decay coefficient to obtain the corrected fields. The corrected fields are combined into a structured waybill according to a preset format; the structured waybill includes at least the waybill number, license plate number, carrier, place of origin, place of destination, cargo category, cargo weight, loading time, unloading time, mileage, estimated freight, and data confidence score.

[0012] Optionally, the method further includes: The vehicle trajectory data is subjected to anomaly pattern detection to identify abnormal behavior and to provide risk warnings for such abnormal behavior; the abnormal behavior includes at least detour behavior or abnormal unloading behavior midway. Virtual waybills with an overall confidence level below a preset threshold are marked, and a manual review process is triggered for the marked virtual waybills; all generated virtual waybills and abnormal behaviors are stored on the blockchain.

[0013] Compared with existing technologies, the virtual waybill generation method based on multi-source data fusion provided by this invention acquires target multi-source data and performs fusion processing to generate loading and unloading event identification results; based on the loading and unloading event identification results, vehicle information, and publicly available information associated with map points of interest, it performs retrieval enhancement generation and semantic reasoning to generate cargo category inference results; based on the loading and unloading event identification results, cargo category inference results, and multi-dimensional influence data, it performs fusion reasoning to generate freight revenue prediction results; based on the loading and unloading event identification results, cargo category inference results, and freight revenue prediction results, it performs confidence fusion and structured processing to generate virtual waybills. This invention achieves accurate identification of loading and unloading locations and improves the accuracy of location identification by fusing and intelligently analyzing multi-source target data; it enhances the accuracy and interpretability of cargo category reasoning by combining retrieval enhancement generation with a large language model; it improves the accuracy and adaptability of freight cost estimation on a national scale by using a freight cost estimation model that comprehensively considers multiple factors; and it ensures that the generated virtual waybills have high credibility and auditability through a confidence fusion mechanism, providing reliable data support for applications such as road freight supervision, tax compliance, and supply chain finance.

[0014] Secondly, the present invention provides a virtual waybill generation system based on multi-source data fusion, applied to the virtual waybill generation method based on multi-source data fusion provided in the first aspect above. The system includes: The target multi-source data acquisition module is configured to acquire target multi-source data; the target multi-source data includes at least vehicle trajectory data, map point of interest data, and electronic fence data; The data fusion module is configured to fuse the target multi-source data to generate loading and unloading event identification results; The cargo category estimation module is configured to perform retrieval enhancement generation and semantic reasoning based on the loading and unloading event identification results, vehicle information, and publicly available information associated with map points of interest, to generate cargo category estimation results. The freight revenue forecasting module is configured to perform fusion reasoning based on the loading and unloading event identification results, the cargo category inference results, and multi-dimensional impact data to generate a freight revenue forecasting result; the multi-dimensional impact data includes at least regional economic indicators, fuel price data, vehicle model data, and seasonal fluctuation factors. The waybill generation module is configured to perform confidence fusion and structured processing based on the loading and unloading event identification results, the cargo category inference results, and the freight revenue prediction results to generate a virtual waybill.

[0015] Thirdly, the present invention provides a virtual waybill generation device based on multi-source data fusion, the device comprising: The system includes a memory, a processor, and a communication interface coupled to the processor; the memory stores a computer program that can be run by the processor; when the processor runs the computer program, it executes the aforementioned virtual waybill generation method based on multi-source data fusion.

[0016] The technical effects achieved by the system-type solutions provided in the second aspect and the equipment-type solutions provided in the third aspect are the same as those achieved by the method-type solutions provided in the first aspect, and will not be repeated here. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating a virtual waybill generation method based on multi-source data fusion. Figure 2 This is a schematic diagram of the framework for a virtual waybill generation method based on multi-source data fusion; Figure 3 A flowchart illustrating the process for generating loading and unloading event identification results; Figure 4 A flowchart illustrating the process of generating goods category prediction results; Figure 5 A flowchart illustrating the process for generating freight revenue forecasts; Figure 6 This is a schematic diagram of the structure of a virtual waybill generation system based on multi-source data fusion; Figure 7 This is a schematic diagram of the equipment structure of a virtual waybill generation system based on multi-source data fusion. Detailed Implementation

[0018] To facilitate a clear description of the technical solutions in the embodiments of the present invention, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, the first threshold and the second threshold are merely used to distinguish different thresholds and do not limit their order. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" are not necessarily different.

[0019] It should be noted that in this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0020] In this invention, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, a combination of a and b, a combination of a and c, a combination of b and c, or a, b, and c, where a, b, and c can be single or multiple.

[0021] In the existing technology, traditional paper waybills have the following disadvantages in the road freight industry: (1) Manual filling is inefficient and prone to errors: Drivers or cargo owners need to manually fill in information such as place of origin, place of destination, type of goods, weight, and freight; (2) Data authenticity is difficult to verify: There is a lack of mechanisms to verify the authenticity and completeness of actual transportation routes and loading and unloading behaviors; (3) High rate of missing waybills: The phenomenon of "goods without waybills" is common, especially among small and medium-sized logistics enterprises; (4) Unable to achieve automated supervision and services: Applications such as tax deduction, insurance claims, and green and low-carbon certification that rely on real waybill data are limited.

[0022] To address the aforementioned shortcomings, some platforms have attempted to determine the origin and destination points using GPS tracks. However, existing solutions for determining origin and destination points using GPS tracks rely solely on latitude and longitude clustering, making it difficult to accurately identify factory / industrial park boundaries. They also lack the ability to accurately match the actual business scope of enterprises with the type of goods. Freight cost estimates are rough and do not take into account factors such as routes, vehicle types, and seasons. Furthermore, they cannot generate a complete and comprehensive "virtual waybill."

[0023] Therefore, to address the aforementioned technical deficiencies, this invention provides a technical solution capable of automatically reconstructing the entire transportation process from multi-dimensional heterogeneous data and generating reliable virtual waybills. Specifically, the virtual waybill generation method, system, and equipment based on multi-source data fusion provided by this invention utilize nationwide truck BeiDou trajectory data, map POIs (Points of Interest), enterprise freight yard electronic fence data, enterprise operating information, publicly available internet information, and Large Language Model (LLM), combined with a freight cost estimation model, to automatically generate virtual waybills with real transportation context information. This method and system are applicable to scenarios such as road freight supervision, tax compliance, supply chain finance, and carbon emission accounting. The following description, in conjunction with the accompanying drawings, illustrates the solutions provided in the embodiments of this specification: like Figure 1 As shown, the process may include the following steps: Step 110: Obtain target multi-source data.

[0024] Target multi-source data can represent a collection of data from different channels and types related to freight loading and unloading scenarios. It can include at least vehicle trajectory data, map point of interest data, and geofence data; data from various sources and in different forms together constitute target multi-source data, used to comprehensively describe various situations in the freight process.

[0025] Among them, map point of interest data (i.e., POI data) can be stored in a point of interest library, which can be the database corresponding to various navigation apps. Map point of interest data can include company addresses, warehouses, logistics parks, ports, etc., and attributes can cover latitude and longitude, category, business hours, etc.

[0026] Electronic fence data can represent preset virtual boundaries (such as circular / polygonal fences with a radius of 50-200m), data corresponding to the area generated based on the center coordinates of the POI, and thresholds including entry / exit time windows (e.g., 5-30min).

[0027] The trajectory data can be vehicle GPS sequences (time stamp, latitude and longitude, speed, load sensor, etc.) with a sampling frequency of ≥1Hz.

[0028] Step 120: Perform fusion processing on the target multi-source data to generate loading and unloading event identification results.

[0029] Loading and unloading events can include loading and unloading. Fusion processing includes techniques such as clustering, electronic fence matching, verification, and multi-source fusion decision models. By fusing target multi-source data, the resulting judgments on whether a vehicle has engaged in loading or unloading activities and related characteristics (such as loading / unloading location and time) clarify when and where the vehicle performed loading and unloading operations during freight transportation.

[0030] Step 130: Based on the loading and unloading event identification results, vehicle information, and publicly available information associated with map points of interest, perform retrieval enhancement generation and semantic reasoning to generate cargo category prediction results.

[0031] Vehicle information can include at least basic vehicle attributes, such as vehicle model, license plate number, load capacity, and vehicle type (e.g., van, flatbed truck, etc.).

[0032] Points of Interest (POIs) are various types of location information marked on a map, such as factories, warehouses, shopping malls, and logistics parks. Public information associated with these POIs may include the location's industry type, main types of goods handled, and business scale.

[0033] Step 140: Based on the loading and unloading event identification results, the cargo category inference results, and the multi-dimensional impact data, perform fusion reasoning to generate a freight revenue prediction result.

[0034] The multi-dimensional impact data includes at least regional economic indicators, fuel price data, vehicle type data, and seasonal fluctuation factors. Specifically, multi-dimensional impact data can represent data corresponding to multiple factors affecting freight revenue. These factors can include market conditions (such as market price fluctuations for different types of goods), transportation distance, transportation time, road conditions, and seasonal factors. Multi-dimensional impact data affects freight revenue from different perspectives, and these factors need to be comprehensively considered during the fusion reasoning process.

[0035] Five basic data indicators were identified for measuring regional economic indicators: regional GDP, total retail sales of consumer goods, freight turnover, highway density, and urbanization rate. These five basic data indicators were combined, and the entropy weight method was used to assign weights to the five dimensions of data. The weights were determined based on the degree of variation (i.e., dispersion) of each indicator data. The greater the variation in indicator values ​​(i.e., the higher the dispersion), the more information it provides, the greater its role in the comprehensive evaluation, and the higher its weight. Finally, standardized values ​​were weighted and summed to obtain the "regional economic indicator" between [0,1]. The larger the value, the higher the level of regional economic development. Specifically, the specific process for determining the regional economic indicator may include:

[0036] 1) Obtain basic data: Obtain the specific values ​​of the above five indicators for each region from the database of trusted institutions.

[0037] 2) Preprocess the acquired technical data: Preprocess the collected data to remove missing values, outliers, etc., and obtain valid data; for missing values, corresponding filling methods can be used, such as mean filling, median filling, or interpolation filling based on data trends; for outliers, it can be determined whether they are real data, and if they are erroneous data, they can be corrected or removed.

[0038] 3) Data Standardization: Since the dimensions and orders of magnitude of various indicators may differ, data standardization can be performed. For example, Z-score standardization (standard deviation standardization) and Min-Max standardization (deviation standardization) can be used. Taking Min-Max standardization as an example, as shown in formula (1): (1); in, This represents the original value of the j-th index in the i-th region. and Let represent the minimum and maximum values ​​of the j-th indicator across all regions, respectively. This represents the standardized value; through standardization, the values ​​of each indicator are mapped to the [0,1] interval.

[0039] 4) Use the entropy weight method to determine the index weights. The proportion of the i-th region under the j-th indicator is calculated using formula (2): (2); Where n is the number of regions.

[0040] The entropy value of the j-th index is calculated using formula (3): (3); Where k = ln(n)1, when When =0, =0, entropy value It reflects the amount of information provided by the j-th indicator; the larger the entropy value, the less information the indicator provides.

[0041] The difference coefficient of the j-th indicator is calculated using formula (4): (4); Among them, the coefficient of difference The coefficient of variation represents the degree of variation of the j-th indicator. The larger the coefficient of variation, the higher the degree of variation of the indicator and the more information it provides.

[0042] The weight of the j-th indicator is calculated using formula (5): (5); Where m represents the number of indicators (e.g., m=5). The weights of each indicator are obtained through the above calculations. The greater the weight, the greater the role of the indicator in the overall evaluation.

[0043] The comprehensive score of "Regional Economic Indicators" is calculated using formula (6): The standardized indicator values ​​are multiplied by their corresponding weights, and then summed to obtain the comprehensive score of the "Regional Economic Indicator" for each region. The formula is as follows: (6); Overall score The value ranges from [0,1]. The higher the score, the higher the economic level of the region.

[0044] Step 150: Based on the loading and unloading event identification results, the cargo category inference results, and the freight revenue prediction results, perform confidence fusion and structured processing to generate a virtual waybill.

[0045] Confidence fusion refers to the comprehensive evaluation and integration of the confidence levels of results obtained from different sources and methods (such as results of cargo loading and unloading event identification, cargo category inference, and freight revenue forecast) in order to determine the reliability of each result.

[0046] A virtual waybill is a digital, simulated freight business document generated based on the results of loading and unloading event identification, cargo category estimation, and freight revenue forecasting, after confidence fusion and structuring processing. It can include key information from the freight process, such as loading and unloading time and location, cargo category, and estimated revenue. Although not an actually signed paper waybill, it has similar functions and roles and can be used for recording, statistics, analysis, and internal management of freight operations.

[0047] Figure 1The method described herein acquires and fuses target multi-source data to generate loading and unloading event identification results. Based on the loading and unloading event identification results, vehicle information, and publicly available information associated with map points of interest, retrieval enhancement generation and semantic reasoning are performed to generate cargo category inference results. Based on the loading and unloading event identification results, cargo category inference results, and multi-dimensional influencing data, fusion reasoning is performed to generate freight revenue prediction results. Based on the loading and unloading event identification results, cargo category inference results, and freight revenue prediction results, confidence fusion and structured processing are performed to generate virtual waybills. This invention achieves accurate identification of loading and unloading locations through the fusion processing and intelligent analysis of target multi-source data, improving the accuracy of location identification. The combination of retrieval enhancement generation and a large language model enhances the accuracy and interpretability of cargo category inference. A freight prediction model that comprehensively considers multiple factors enhances the accuracy and adaptability of freight estimation at a national scale. The confidence fusion mechanism ensures that the generated virtual waybills have high credibility and auditability, providing reliable data support for applications such as road freight supervision, tax compliance, and supply chain finance.

[0048] based on Figure 1 In addition to the method described herein, this specification also provides some specific implementation methods of this method, which will be described below.

[0049] First, combine Figure 2 A brief overview of the overall framework of the method provided by this invention is as follows: Figure 2 As shown, the method flow framework may include: This diagram illustrates a logistics data processing workflow based on BeiDou trajectory data and a database of Points of Interest (POIs) and geofencing. The following are textual descriptions of all steps in the diagram: Data input: Collect BeiDou trajectory data (including real-time and historical data), POI and fence database information (which can cover relevant data such as factory areas and logistics parks).

[0050] Data fusion: Trajectory cleaning is performed to ensure data accuracy and integrity; DBSCAN clustering algorithm is used to detect stop points; POI matching and fence attribution are performed to determine the specific location and related attributes of trajectory points.

[0051] Cargo category estimation: Associate license plate information to determine the transportation company; match PIO (possibly a typo for POI) to obtain publicly available company information and historical waybill information; use LLM (Large Language Model) to infer cargo category by integrating diverse information.

[0052] Freight cost estimation: Input relevant information (including origin, destination, type of goods, vehicle type, actual mileage and fuel price index, etc.) and calculate the freight cost based on this input information.

[0053] Virtual waybill generation: Integrate all relevant information, including the results of data fusion, cargo category prediction results, and freight cost estimation results, to generate a virtual waybill with a standard waybill structure.

[0054] Figure 2 The process framework in this paper integrates multi-source data and utilizes data analysis and model inference techniques to automate the processing from trajectory data to virtual waybills. Next, regarding... Figure 2 Each step in the framework is explained in detail: For step 120, fusing the target multi-source data to generate loading and unloading event identification results may include: The vehicle trajectory data is analyzed using the DBSCAN clustering algorithm to identify candidate stopping areas. The candidate stopping areas are those where the stopping time exceeds a first preset threshold and the speed variation is less than a second preset threshold. For each candidate stop point in the candidate stop area, perform electronic fence matching to determine the associated points of interest; Based on the associated points of interest, semantic verification of POIs is performed to obtain semantic information; Based on the semantic information and the load change information in the vehicle trajectory data, a multi-source fusion decision model is used to output the loading and unloading event identification result; the multi-source fusion decision model is a Bayesian network; the load change information represents the load change amount collected by the load sensor, and the load change amount is regarded as a loading and unloading event when it exceeds a preset percentage threshold; the loading and unloading event identification result includes at least the location coordinates, event type, confidence level, truck license plate, vehicle type, load capacity, departure time, and arrival time.

[0055] Furthermore, performing geofence matching on each candidate stop point in the candidate stop area to determine associated points of interest may include: For each candidate stop point, check whether it has entered the preset fence: if the candidate stop point is inside the fence and the stay time meets the preset time, then match the electronic fence icon; For candidate stopping points that do not match the electronic fence identifier, the boundary range of the electronic fence is dynamically adjusted according to the spatial distribution variance of historical trajectory data; and a dynamic electronic fence is generated for the candidate stopping points that do not match the electronic fence identifier based on the convex hull algorithm.

[0056] The detailed discussion of step 120 above will be further illustrated with practical examples below, such as... Figure 3 As shown, the specific steps for fusing the target multi-source data to generate loading and unloading event identification results may include: (1) First, the collected input data (POI data, electronic fence data, and vehicle trajectory data; specifically, the trajectory data collected by the edge device or vehicle terminal may include license plate number, vehicle type, load, GPS location, and time, etc.) is loaded. This data is uploaded to the cloud server's data receiving and storage module in real time through the system. At the same time, POI data and electronic fence data are also uploaded to the cloud server as data sources. The cloud server receives and stores the uploaded trajectory data, POI data, and electronic fence data.

[0057] (2) Then, data preprocessing is performed (loading trajectory data and using a data preprocessing engine to clean and denoise the trajectory data to improve data quality). Specifically, data preprocessing may include: 1) POI filtering: Extract logistics-related categories (e.g., enterprise factories, warehousing and logistics), and use spatial index (KD-Tree) to match trajectory points.

[0058] 2) Track cleaning: The Douglas-Peucker algorithm is used to compress the track and remove noise (speed <5km / h is considered a potential stop).

[0059] 3) Fence generation: Dynamic fence = POI location + buffer (based on historical trajectory variance).

[0060] (3) Trajectory dwell analysis, i.e., determining candidate dwell points. The preprocessed trajectory data is analyzed to identify candidate dwell points, and the coordinates and dwell time of each dwell point are calculated. Specific steps may include: The DBSCAN clustering algorithm (ε=50m, minPts=5) was used to identify dwell points and calculate the dwell time: T. stay =t exit -t enter If T stay >θ time (e.g., 10 min) and velocity variation σ v <θ v (For example, at 2km / h) it is marked as a candidate stop point.

[0061] (4) Spatiotemporal correlation analysis (using a spatiotemporal correlation engine to correlate identified rest points with POI data and geofence information, which helps to determine the specific location and related information of rest points), specifically including: For each stop point, check if it has entered the preset fence: if the trajectory point is inside the fence and the dwell time is > θ fence (For example: 80% T) stay If a point is not matched, then the fence ID is associated. Unmatched points trigger dynamic fence generation (using the convex hull algorithm to surround the trajectory cluster).

[0062] (5) Loading and unloading event detection (using a loading and unloading event detection model to determine whether a loading and unloading event has occurred at a stop point based on the location, duration of the stop point, and associated POI / fence information), specifically including: Calculate the distance d between the stop point and the POI. poi <r poi (e.g., 100m), and use cosine similarity to match POI categories (e.g., Word2Vec embedding POI tags). If the matching score > 0.7, it is confirmed as a loading / unloading location; otherwise, it is combined with the load change Δ load (Sensor data) as behavioral evidence (Δ) load >20% is considered loading and unloading.

[0063] (6) Event type loading and multi-source fusion decision (the identified loading and unloading event types are loaded, and a multi-source fusion decision model is used to further optimize and confirm the identification results of loading and unloading events by combining extracted departure and arrival times, associated vehicle information, and confidence scores, etc.) Specifically, this includes: Using a Bayesian network to fuse the probabilities of the three: P(Loading and Unloading|POI, Fence, Stay)=P(POI)×P(Fence|Stay)×P(Stay).

[0064] Output recognition results when the threshold is >0.85, and support RAG (Retrieval-Augmented Generation) for historical trajectory optimization retrieval.

[0065] The final loading and unloading event identification results are formatted and output. These results can be stored in a database or provided to other systems via API, message queues, or other means. Output results include: location coordinates (departure / destination), type (loading / unloading), confidence level, truck license plate number, vehicle type, load capacity, departure time, and arrival time.

[0066] Through the detailed implementation scheme of step 120 above, the system can automatically and accurately identify loading and unloading events of vehicles during transportation and output the results in a structured form, which is convenient for subsequent data analysis and application.

[0067] For step 130, based on the loading and unloading event identification results, vehicle information, and publicly available information related to map points of interest, retrieval enhancement generation and semantic reasoning are performed to generate cargo category inference results, which may include: Construct a RAG vector knowledge base; the RAG vector knowledge base shall include at least publicly available cargo data, historical waybill patterns, target multi-source data, vehicle information, and cargo category inference results. Based on the loading and unloading event identification results and the vehicle information, a query vector is generated, and related documents are retrieved from the RAG vector knowledge base through RAG retrieval enhancement; the RAG retrieval enhancement integrates trajectory data, vehicle information and POI public cargo data, and retrieves Top-K related documents, and the Top-K related documents are embedded with BERT vectors; The associated documents, the loading / unloading event identification results, and the vehicle information are input into a large language model to generate the cargo category inference result and confidence score. The confidence score is obtained by fusing company cargo information associated with POIs, historical waybill data, and vehicle qualification information using a Bayesian method. Specifically, the cargo category inference result can be combined with... Figure 4 Please provide an explanation, such as Figure 4 As shown, the "Generate Goods Category Inference Results" engine adopts a layered architecture: a data acquisition layer, a RAG retrieval layer, an LLM (Large Language Model) inference layer, and an output layer. The specific implementation process can be divided into four main stages: data acquisition, information retrieval and processing, LLM (Large Language Model) inference, and result output. Specifically, it can include: (1) Data acquisition stage: 1) Trajectory Recognition Data Collection: Collect real-time or historical trajectory data of vehicles, including trajectory points, POI (Point of Interest) IDs, and other information. For example: location coordinates, loading / unloading status (loading / unloading), and confidence level (0-1).

[0068] 2) Vehicle Information Collection: Collect basic vehicle information, such as license plate number, vehicle type, and load capacity. For example: license plate, vehicle type, historical waybills (type of goods and route patterns of the past 10 orders), and company qualifications (food transportation permit, obtained from the enterprise credit system).

[0069] 3) POI Association Information Collection: Obtain POI association information related to enterprise registration information, business scope, comments, etc. For example: match POIs by coordinates, including location type (factory / freight yard) and publicly available online description information (e.g., "electronics factory").

[0070] 4) Public Goods Data Collection: Collect publicly available goods data such as industry classifications, price indices, and HS codes. This data is input as historical data into subsequent processing modules, for example, retrieving factory / freight yard goods (e.g., "textile factories mainly produce cotton yarn") from trusted website APIs or knowledge graphs.

[0071] (2) Information retrieval and processing stage 1) Real-time Query and Retrieval Enhanced Generation (RAG): Semantic retrieval is performed on real-time queries, retrieving the Top-K documents relevant to the query from the knowledge base; RAG optimization (such as reordering) improves the quality of retrieval results; for example, when inferring specific POI details, rules + LLM can be used for preliminary classification: if the coordinates match an industrial zone POI and historical waybills contain "manufacturing", it is marked as "factory address"; if the unloading confidence score is >0.9 and the company's qualifications contain "warehousing", it is marked as "freight yard address". Semantic representations are generated using embedding vectors (BERT, 768 dimensions).

[0072] 2) Data Standardization and Cleaning: The collected data is standardized and cleaned to improve its quality and consistency. The processed data is then used to build a knowledge base, vectorized, and stored in a vector database. For example, during data standardization, numerical values ​​(coordinates, confidence scores) are GeoHash encoded; text (waybill, qualification documents) is chunked. size =512 tokens). During knowledge base construction, the RAG vector database stores publicly available product data, with indexes including product type, associated POIs, and vehicle type. RAG retrieval enhancements can include: Query embedding: Embed the input data (Sentence-BERT) into a vector q.

[0073] Search Top-K (K=5) related documents: Retrieve POI cargo descriptions (e.g., "Coordinates (31.23,121.48) associated with a textile factory, cargo: cotton yarn, dye") and historical waybill patterns (e.g., "This vehicle type often transports textiles") from the knowledge base.

[0074] Contextualization: The search results C={doc1, doc2, ...} are combined with vehicle information V (vehicle type + qualifications).

[0075] RAG optimization: Post-retrieval re-ranking (BM25+ semantic similarity > 0.7) to reduce illusions.

[0076] (3) LLM reasoning stage 1) Input variable V and context C (from the results of RAG retrieval).

[0077] 2) LLM performs inference: The trained LLM model performs inference based on the input information and its own knowledge base to generate inference results about the product category.

[0078] Confidence calculation: Fusion Bayesian probability P(cargo|data) = P(POI) cargo )×P(V|hist)×P(RAG retrieval (), threshold > 0.8 confirmed.

[0079] (4) Result Output Stage 1) LLM extracts structured results, including information such as goods, confidence level, and inference chain. For example: structured results (JSON) include goods name, confidence level, and inference chain (e.g., "based on textile mill POI and historical waybills, inferred to be cotton yarn").

[0080] 2) Results Visualization and Storage: Structured results are displayed through visualization modules (such as Knowledge Graph) to facilitate user understanding and analysis. Simultaneously, the results are stored in a database or provided to other systems via API. For example, the knowledge graph (Neo4j) displays associations (vehicle → POI → goods), and the graph results are entered into the RAG enhanced vector database.

[0081] The training process for LLM (Large Language Model) can be illustrated with a specific example: 1) Data acquisition phase, this phase is related to the aforementioned... Figure 4 The explanation of "data acquisition stage" is the same, so it will not be repeated here.

[0082] 2) Constructing a RAG vector knowledge base: Convert the various types of data collected above into vector representations and construct a RAG vector knowledge base.

[0083] Enhanced training sample generation based on RAG retrieval: Simulates actual loading and unloading event recognition results and vehicle information to generate query vectors, and retrieves related documents from the RAG vector knowledge base. The retrieved related documents, simulated loading and unloading event recognition results, and vehicle information are combined as input samples.

[0084] Labeling of cargo category prediction results and confidence levels: Based on the company cargo information associated with the POI, historical waybill data, and vehicle qualification information, the confidence level of the cargo category prediction results is calculated using the Bayesian method, and the correct cargo category prediction results and confidence levels are used as the labeling information for the samples.

[0085] Construct training samples related to seasonal factors: Based on the basic data of the fixed seasonal factor table (event labels within a preset time period, basic floating range, regional weighting, and commodity type correction) and dynamic event data (trigger time, event description, impact coefficient, duration days, and event source), generate fixed seasonal factor and dynamic event factor calculation data for different scenarios, as well as the corresponding seasonal fluctuation factor results, as training samples.

[0086] Goods category inference task: Input samples (related documents, loading and unloading event recognition results, vehicle information) are fed into a large language model, and the model outputs the goods category inference result and confidence score. A suitable loss function, such as the cross-entropy loss function, is designed to measure the difference between the model's inference result and the labeled result. The model parameters are updated through backpropagation, enabling the model to accurately infer the goods category and provide a reasonable confidence score.

[0087] Seasonal fluctuation factor calculation task: Input data related to seasonal factors into the model, and let the model calculate the seasonal fluctuation factor by weighting and summing fixed seasonal factors and dynamic event factors according to preset weights. Design an appropriate loss function, such as the mean squared error loss function, to optimize the model's calculation results so that it can accurately calculate the seasonal fluctuation factor.

[0088] Multiple rounds of fine-tuning training are conducted: The large language model is trained multiple times using a pre-built fine-tuning dataset. In each round, the model parameters are adjusted based on the feedback from the loss function, gradually improving the model's performance in tasks such as cargo category inference and seasonal fluctuation factor calculation, resulting in an LLM model that meets the requirements. The specific implementation scheme for generating cargo category inference results includes a complex process of multi-source data collection, information retrieval and processing, large language model inference, and result output and storage. Through this process, the system can automatically and accurately infer the cargo category and output the results in a structured form, providing strong support for logistics, supply chain management, and other fields.

[0089] Step 140: Based on the loading and unloading event identification results, the cargo category inference results, and multi-dimensional impact data, perform fusion reasoning to generate a freight revenue prediction result, which may include: The actual transportation distance is calculated based on the location coordinates in the loading and unloading event identification results. Based on the actual transportation distance, the predicted cargo category, and historical transaction price data, the benchmark freight rate is calculated using a pre-trained random forest model. The benchmark freight rate, regional economic indicators, fuel price data, vehicle model data, and seasonal fluctuation factor are input into the large language model. Combined with the contextual information enhanced by RAG retrieval, interval reasoning is performed to generate freight revenue prediction results. The seasonal fluctuation factor is obtained by querying the fixed seasonal factor table and the dynamic event table and fusing them according to the preset weight ratio.

[0090] Furthermore, the process of determining the seasonal fluctuation factor may include: Fixed seasonal factors are determined based on the basic data of the fixed seasonal factor table; the basic data of the fixed seasonal factor table includes at least event tags within a preset time period, basic floating range, regional weighting, and commodity type correction; Through market research, we identified typical events in different months and their impact on prices. We collected and organized relevant information on common events each month, including event tags, base fluctuation ranges, regional weighting, and commodity type adjustments, to create a fixed seasonal factor table. For example, during the peak transportation season in January and February, the base fluctuation range was +15%-25%, the regional weighting was +5% for region A, and the commodity type adjustment was +5% for cold chain. Following the rule "Fixed Seasonal Factor = Base Fluctuation Range + Regional Weighting + Commodity Type Adjustment," we calculated the corresponding fixed seasonal factor for each month. For example, the fixed seasonal factor for the summer high temperatures and peak cold chain season from June to August was +10%-20% + +5% for region B + +15% for cold chain. This factor will serve as the numerical basis for the fixed factor portion in the calculation of the seasonal fluctuation factor. The dynamic event factor is calculated by weighting and averaging dynamic event data and combining it with an LLM model; the dynamic event data includes at least the trigger time, event description, impact coefficient, duration in days, and event source. Dynamic events can be obtained from major event information sources such as traffic information from the Ministry of Transport, weather warnings, and Double Eleven announcements from e-commerce platforms. A dynamic event table is constructed by daily web crawling of this information. LLM (Lightweight Leverage Model) is used to determine whether an event constitutes a new event. If so, the vinegar impact coefficient (which can be understood as a representation of price impact coefficient) is calculated by comparing the current shipping cost with the usual price and stored in the dynamic event table for vectorization. This results in a dynamic event table containing information such as trigger time, event description, impact coefficient, duration, and source. For example, event A has an impact coefficient of +12%-25%, a duration of 7 days, and its source is records from the Ministry of Transport's traffic information system. Sentence-BERT is used to generate a 128-dimensional vector for each event in the dynamic event table. Similar historical events are retrieved from the dynamic event table, and the impact coefficients of these similar events are weighted and averaged. The weighted average information is then fed into a lightweight LLM model for final fine-tuning, obtaining 20% ​​of the qualitative factors as dynamic event factors.

[0091] The seasonal fluctuation factor is calculated by weighting and summing the fixed seasonal factor and the dynamic event factor according to preset weights.

[0092] The seasonal fluctuation qualitative factor is calculated using the formula: "Seasonal fluctuation qualitative factor = Fixed seasonal factor × 80% + Dynamic event factor × 20%". This seasonal fluctuation factor, which comprehensively considers the influence of both fixed seasonal factors and dynamic events, can more comprehensively and accurately reflect the combined price fluctuations in different months due to seasonal and abnormal events.

[0093] The implementation process of step 140 above can be combined with Figure 5 Please provide an explanation, such as Figure 5As shown, in the specific process of generating freight revenue forecast results, the model adopts a layered architecture: a data preprocessing layer, a RAG retrieval layer, an LLM inference layer, and an output layer. The method flow may include: (1) Data input and preprocessing Point of Interest (POE) Input: The user or system provides the point of origin and destination information for the freight as the POE (Point of Interest / Origin-Destination) input.

[0094] Cargo Information Input: Enter detailed information about the cargo, which may include cargo type, weight, volume, etc.

[0095] Vehicle type and load input: Select the appropriate vehicle type and provide the vehicle's load information.

[0096] Market-related factor input: Consider market dynamics and input relevant market factors, such as supply and demand, seasonal changes, etc.

[0097] For example: Origin POI: origin coordinates, type (loading), confidence level, truck license plate, vehicle type and origin company information.

[0098] Destination POI: origin coordinates, type (loading), confidence level, truck license plate, vehicle type, and destination company information.

[0099] Goods category information: Goods category name, freight impact risk coefficient (e.g., "sensitive to moisture, sensitive to dirt, sensitive to fire, sensitive to delays, sensitive to mixed loading").

[0100] Regional economic level (GDP index): graded (high / medium / low), with high economic zones increasing by 1.2-1.5 times.

[0101] Oil price (RMB / liter): Real-time API data, fuel cost accounts for 25-35%.

[0102] Vehicle type axle count (2-6 axles): For heavy-duty agents, the higher the number of axles, the cost increases by 10-25%.

[0103] Seasonal fluctuations (month + event): Peak periods (such as the Spring Festival) rise by 15-30%, coded with qualitative descriptions (e.g., "rainy season delays").

[0104] Using a qualitative factor library that includes a fixed seasonal factor table and a dynamic event table, and employing a hybrid approach with 80% structured rules and 20% RagLLM predictions, the qualitative model maintains stability while retaining agile responses to abnormal events and market supply and demand conditions.

[0105] The fixed seasonal factor table is pre-constructed through market research, as shown in the following example: Table 1. Fixed Seasonal Factors The dynamic event table is generated daily by web crawlers, collecting data such as traffic, weather, and e-commerce. LLM (Local Leverage Management) is used to determine whether a new event is constituted. For events constituting new events, the impact coefficient (based on a comparison of the current shipping cost with weekday prices) is automatically stored in the dynamic event table (vectorized). A specific example is shown below: Table 2. Dynamic Event Table The specific rule is as follows: 80% of the qualitative factors are determined by querying a fixed seasonal factor table. The fixed seasonal factor = "base fluctuation range" + "regional weighting" + "goods category correction". Then, "transportation time + origin city + destination city + goods category" are concatenated into natural language to generate a 128-dimensional vector. Similar historical events are retrieved from the dynamic event table and weighted averaged. This information is then fed into the trained LLM model for final fine-tuning to obtain the remaining 20% ​​of the qualitative factors. The seasonal fluctuation qualitative factor = fixed seasonal factor × 80% + dynamic event factor × 20%.

[0106] Historical transaction prices (average / median): retrieved from the company database as a benchmark.

[0107] Data preprocessing: All input data (e.g., oil prices, vehicle-related factors, and market-related factors) undergoes standardization (factor normalization) in the data standardization (factor normalization) module to ensure data consistency and comparability. Subsequently, the data enters the feature calculation module for rule / model-based feature extraction and calculation. Specifically:

[0108] POI processing: Calculate the origin / destination POI, and use Word2Vec to embed the POI description to generate a semantic vector.

[0109] Base price calculation: Based on factors such as origin / destination POI, vehicle-related factors, and market-related factors, a pre-trained random forest (RF) model is used based on mileage / base cost, as shown in formula (7): (7) In this formula, This represents the benchmark intercept term, indicating the benchmark level of freight prices when all other independent variables are zero; This represents the calculated actual path distance; it also represents the distance of freight from the point of origin (POI). start ) to destination (POI) end The distance; This is a coefficient representing the distance, reflecting the degree to which distance affects freight prices; Product information is typically represented by vectors using embedding techniques (such as Word2Vec) in data standardization. It is a coefficient representing the information about goods, indicating the impact of different types of goods on freight prices; Economic factors can include macroeconomic indicators, market supply and demand, and other external economic factors that affect freight prices. It is the coefficient of the economic factor, reflecting the impact of economic conditions on freight prices; This represents the oil price factor, indicating the impact of changes in fuel prices on freight costs. The coefficient representing the oil price factor reflects the sensitivity of oil price changes to freight rates; This indicates the number of axles or vehicle type factor. Different vehicle types (such as single-axle, dual-axle, etc.) may have different transportation costs. It is a coefficient representing the number of vehicle axles, indicating the impact of vehicle type on freight prices; Seasonal factors represent the impact of different seasons on freight demand or costs, such as holidays and weather changes. It is the coefficient of the seasonal factor, reflecting the impact of seasonal changes on freight prices; This represents historical prices or trend factors used to capture dynamic changes in freight prices; It is the coefficient of the historical factor, representing the impact of historical price information on current freight price forecasts.

[0110] Training data: historical waybills nationwide (>500,000), ranked by feature importance (distance > goods > oil price).

[0111] (2) Information retrieval and enhancement RAG (Retrieval-Augmented Generation) retrieval leverages real-time query and semantic retrieval technologies to extract documents or information relevant to the current freight task from the knowledge base. The retrieved Top-K related documents, after RAG optimization (e.g., reordering), are provided as contextual information to subsequent processing modules. The knowledge base is built upon dynamic cost databases, dynamic market databases, historical transaction prices, and other information, and stored in a vector database using vectorized embedding technology. Specifically, the knowledge base is constructed as follows: the vector database stores publicly available data, including POI cargo associations (built based on public and historical information), historical waybill patterns, and national oil price / economic indices. The retrieval process involves inputting the embedded vector q (Sentence-BERT) to retrieve Top-K (K=5) related documents C (e.g., historical POI prices, cargo risk reports). Dynamic updates are achieved through a cyclical replenishment of the knowledge base by new waybills.

[0112] (3) LLM inference: The LLM that has been trained is used for inference output; the LLM inference process is the same as the above process, and will not be repeated here.

[0113] (4) Results output and application Results Visualization and Storage: Freight revenue forecasts are displayed through a visualization module for easy user understanding; simultaneously, the results are stored in a database or provided to other systems via API. Outputs: Transportation mileage, freight rate range, explanatory report, and confidence level.

[0114] Freight pricing and confidence level applications: Estimated freight revenue can serve as a reference for freight pricing. Confidence level information is used to assess the reliability of the estimates, helping users make more informed decisions.

[0115] The process of generating freight revenue forecast results includes multi-source data input, information retrieval and enhancement, large language model reasoning, and result output and application. Through these steps, freight revenue can be automatically and accurately forecasted.

[0116] For step 150, it may include: Assign preset weights to key fields in the loading and unloading event identification results, the cargo category inference results, and the freight revenue prediction results; Based on the pre-defined weights assigned to key fields and the confidence levels of each key field, the overall confidence level of the virtual waybill is calculated using a weighted geometric average algorithm. The confidence scores of fields with logical conflicts are attenuated and corrected according to a correction rule. The correction rule is to multiply the confidence scores of fields with logical conflicts by a decay coefficient to obtain the corrected fields. The corrected fields are combined into a structured waybill according to a preset format; the structured waybill includes at least the waybill number, license plate number, carrier, place of origin, place of destination, cargo category, cargo weight, loading time, unloading time, mileage, estimated freight, and data confidence score.

[0117] Combining the fields calculated above, a virtual electronic waybill is intelligently generated by integrating indicators from each step with confidence level as the core. The overall confidence level of the virtual electronic waybill is dynamically calculated using a method based on key field weights and a confidence level decay mechanism.

[0118] First, design the weights based on the importance of the fields, as shown in the table below: Table 3. Field Weight Design Table The confidence level of the virtual electronic waybill is calculated using the weighted geometric mean method to avoid the risk of high-confidence fields masking low-confidence key fields. The specific formula is shown in formula (8): (8) Here, Overall Confidence represents the overall confidence level, indicating the overall confidence level after comprehensively considering multiple factors; confidence represents the confidence level of key fields. represents the confidence level of the i-th factor, w represents the weight corresponding to the confidence level, and n represents the total number of factors involved in the calculation, i.e., how many different confidence levels are considered. Calculation example: With confidence levels of 0.85 for goods (weight 0.3), 0.92 for coordinates (weight 0.25), 0.90 for vehicle type (weight 0.2), 0.78 for freight costs (weight 0.15), and 0.95 for license plate (weight 0.1), the overall confidence score is 0.85. 0.3 ×0.92 0.25 ×0.90 0.2 ×0.78 0.15 ×0.95 0.1 It is approximately equal to 0.86.

[0119] Simultaneously, a confidence decay mechanism is used before weight calculation. By comparing and analyzing the inference chain of the explanatory fields in the above steps, the relevant confidence of conflicting (e.g., "goods = cotton" but "originating point = steel mill POI") is downgraded, and the rule is reduced to relevant confidence. A weighting adjustment of 0.8 is applied.

[0120] The virtual waybill generated using the above steps contains the fields shown in the following table: Table 4. Examples of Virtual Waybills

[0121] Output formats supported: JSON, XML, PDF, database table records, etc. The actual route of the waybill is generated into a real-time simulation playback animation on the map.

[0122] As an optional implementation, the technical solution provided by the present invention may further include: The vehicle trajectory data is subjected to anomaly pattern detection to identify abnormal behavior and to provide risk warnings for such abnormal behavior; the abnormal behavior includes at least detour behavior or abnormal unloading behavior midway. Virtual waybills with an overall confidence level below a preset threshold are marked, and a manual review process is triggered for the marked virtual waybills; all generated virtual waybills and abnormal behaviors are stored on the blockchain.

[0123] This invention also includes an anomaly detection and review mechanism: detecting whether there are serious detours or multiple stops for unloading along the way, and providing risk warnings if such situations occur; marking waybills with low confidence for manual review; and supporting blockchain evidence storage to ensure immutability.

[0124] Following the same approach, this invention also provides a virtual waybill generation system based on multi-source data fusion. This system is applied to the virtual waybill generation method based on multi-source data fusion provided in the foregoing embodiments of this specification, such as... Figure 6 As shown, the system may include: The target multi-source data acquisition module 610 is configured to acquire target multi-source data; the target multi-source data includes at least vehicle trajectory data, map point of interest data, and electronic fence data. Data fusion module 620 is configured to perform fusion processing on the target multi-source data to generate loading and unloading event identification results; The cargo category estimation module 630 is configured to perform retrieval enhancement generation and semantic reasoning based on the loading and unloading event identification results, vehicle information and publicly available information associated with map points of interest, to generate cargo category estimation results. The freight revenue forecasting module 640 is configured to perform fusion reasoning based on the loading and unloading event identification results, the cargo category inference results, and multi-dimensional impact data to generate a freight revenue forecasting result; the multi-dimensional impact data includes at least regional economic indicators, fuel price data, vehicle model data, and seasonal fluctuation factors. The waybill generation module 650 is configured to perform confidence fusion and structured processing based on the loading and unloading event identification results, the cargo category inference results, and the freight revenue prediction results to generate a virtual waybill.

[0125] based on Figure 6 In the system described above, the specific implementation of each module follows the same scheme as the virtual waybill generation method based on multi-source data fusion in the aforementioned embodiments, and will not be repeated here.

[0126] This invention provides a virtual waybill generation method, system, and device based on multi-source data fusion. The aim is to offer a technical method and system for automatically generating structured, auditable, and highly reliable virtual waybills without human intervention, addressing issues such as incomplete information, poor authenticity, and low automation in existing waybill generation methods. By integrating: nationwide truck BeiDou trajectory data, map POI and geofencing data, enterprise registration and business scope information, internet information, Large Language Model (LLM) semantic understanding and reasoning analysis capabilities, and a freight revenue prediction model, the invention achieves accurate reconstruction of vehicle transportation behavior and outputs standardized virtual waybills containing core fields such as origin, destination, cargo category, transportation mileage, estimated freight, loading and unloading time, and transportation duration.

[0127] Compared with the prior art, the innovative ideas and technical effects of this invention include at least the following: (1) A precise identification method for loading and unloading locations based on POI + fence + trajectory dwell analysis 1) Adaptive fence optimization: The fence radius is dynamically adjusted based on the trajectory variance to solve the boundary ambiguity problem of static fences and support the identification of temporary loading and unloading points.

[0128] 2) Enhanced behavioral features: By introducing load sensors and dwell time series analysis (HMM hidden Markov model to model trajectory states: driving-dwelling-loading / unloading), pseudo-dwellings (such as rest) are distinguished with high accuracy.

[0129] 3) Multimodal dynamic fusion: Unlike single trajectory analysis (triggered by a single dwell time), this invention embeds POI semantics into a Bayesian model for the first time, realizing three-dimensional verification of behavior, space and semantics, which significantly improves recognition accuracy.

[0130] (2) Intelligent mapping mechanism based on trajectory recognition data → business scope → goods category based on large model

[0131] 1) Multi-source logistics data RAG fusion: For the first time, trajectory recognition data (coordinates + loading and unloading + confidence level) and vehicle information (model + history + qualifications) are embedded into the RAG knowledge base to realize the whole link from POI cargo inference to cargo name reasoning, which is superior to single POI classification.

[0132] 2) Dynamic PIO Goods Category Inference: Combining coordinate association with public API (Tianyan Check), LLM is used to initially classify POI types (factory / freight yard), and RAG is used to retrieve goods, thus overcoming the limitations of static POI.

[0133] 3) Enhanced vehicle history model: Use historical waybills to build personalized knowledge graphs (e.g., HMM modeling route-cargo association), LLM matching vehicle type / license, distinguishing dangerous goods, etc., to improve the targeting of reasoning.

[0134] 4) Real-time Adaptation: The RAG enhanced retrieval vector database supports online updates (new waybill feedback update mechanism).

[0135] (3) National vehicle freight revenue prediction model integrating multiple factors 1) LLM+RAG multi-source fusion: RAG is used to retrieve publicly available goods and historical patterns of POIs, which enhances LLM's reasoning on geographic semantics (origin / endpoint POIs) and goods risk, outperforming traditional ML dynamic pricing models.

[0136] 2) POI-driven dynamic calculation: Integrating the embedding of origin / endpoint POIs, calculating path semantic risks (e.g., the high value of goods in industrial zones), significantly improving accuracy at the national scale.

[0137] 3) Adaptation between goods and seasons: LLM CoT handles non-linear interactions (such as the winter freight rate for perishable goods), and RAG supplements public data in real time to reduce illusions.

[0138] (4) Transportation process reconstruction and anomaly detection mechanism based on spatiotemporal trajectory: determine whether there is detour, unloading in the middle, false itinerary, etc., to enhance the credibility of virtual waybill.

[0139] (5) End-to-end virtual waybill generation capability: outputs standard JSON / XML format waybills, supporting integration with tax, finance, and government systems.

[0140] Based on the same approach, embodiments of this specification also provide a virtual waybill generation device based on multi-source data fusion. For example... Figure 7 As shown, the device includes: The system includes a memory, a processor, and a communication interface coupled to the processor; the memory stores a computer program that can be run by the processor; when the processor runs the computer program, it executes the aforementioned virtual waybill generation method based on multi-source data fusion.

[0141] like Figure 7 As shown, the memory stores a computer program that can be executed by the processor; when the processor executes the computer program, it implements the method provided in the embodiments of the present invention.

[0142] Optionally, the computer execution instructions in the embodiments of the present invention may also be referred to as application code, and the embodiments of the present invention do not specifically limit this.

[0143] In a specific implementation, as one example, such as Figure 7 As shown, a processor may include one or more CPUs, such as Figure 7 CPU0 and CPU1 in the CPU.

[0144] In a specific implementation, as one example, such as Figure 7 As shown, the terminal device may include multiple processors, such as Figure 7 The processors in the system. Each of these processors can be a single-core processor or a multi-core processor.

[0145] The foregoing mainly describes the solutions provided by the embodiments of the present invention from the perspective of the interaction between various modules. It is understood that each module, in order to achieve the above functions, includes corresponding hardware structures and / or software units for executing each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments disclosed herein, the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0146] The embodiments of the present invention can divide functional modules according to the above method examples. For example, each function can be divided into its own functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. It should be noted that the module division in the embodiments of the present invention is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0147] The processor described in this specification may also function as a memory. The memory stores computer execution instructions for carrying out the present invention, and its execution is controlled by the processor. The processor executes the computer execution instructions stored in the memory, thereby implementing the method provided in the embodiments of the present invention.

[0148] The memory can be read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions, random access memory (RAM) or other types of dynamic storage devices capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed discs, laser discs, optical discs, universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited to these. The memory can exist independently and be connected to the processor via communication lines. The memory can also be integrated with the processor.

[0149] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed invention. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0150] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely exemplary descriptions of the invention as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A virtual waybill generation method based on multi-source data fusion, characterized in that, The methods include: Acquire target multi-source data; The target multi-source data includes at least vehicle trajectory data, map point of interest data, and electronic fence data; The target multi-source data is fused to generate loading and unloading event identification results; Based on the loading and unloading event identification results, vehicle information, and publicly available information related to map points of interest, retrieval enhancement generation and semantic reasoning are performed to generate cargo category inference results; Based on the loading and unloading event identification results, the cargo category inference results, and multi-dimensional impact data, a freight revenue forecast result is generated through fusion reasoning; the multi-dimensional impact data includes at least regional economic indicators, fuel price data, vehicle model data, and seasonal fluctuation factors. Based on the loading and unloading event identification results, the cargo category inference results, and the freight revenue prediction results, confidence fusion and structured processing are performed to generate a virtual waybill; The target multi-source data is fused to generate loading and unloading event identification results, including: The vehicle trajectory data is analyzed using the DBSCAN clustering algorithm to identify candidate stopping areas. The candidate stopping areas are those where the stopping time exceeds a first preset threshold and the speed variation is lower than a second preset threshold. For each candidate stop point in the candidate stop area, perform electronic fence matching to determine the associated points of interest; Based on the associated points of interest, semantic verification of POIs is performed to obtain semantic information; Based on the semantic information and the load change information in the vehicle trajectory data, a multi-source fusion decision model is used to output the loading and unloading event identification result; the multi-source fusion decision model is a Bayesian network; the load change information represents the load change amount collected by the load sensor, and the load change amount is regarded as a loading and unloading event when it exceeds a preset percentage threshold; the loading and unloading event identification result includes at least the location coordinates, event type, confidence level, truck license plate, vehicle type, load capacity, departure time, and arrival time; Based on the loading and unloading event identification results, vehicle information, and publicly available information related to map points of interest, retrieval enhancement generation and semantic reasoning are performed to generate cargo category inference results, including: Construct a RAG vector knowledge base; the RAG vector knowledge base shall include at least publicly available cargo data, historical waybill patterns, target multi-source data, vehicle information, and cargo category inference results. Based on the loading and unloading event identification results and the vehicle information, a query vector is generated, and related documents are retrieved from the RAG vector knowledge base through RAG retrieval enhancement; RAG retrieval enhancement integrates trajectory data, vehicle information and POI public cargo data, and retrieves Top-K related documents, wherein the Top-K related documents are embedded with BERT vectors; The associated documents, the loading and unloading event identification results, and the vehicle information are input into the large language model to generate the cargo category inference result and confidence level; the confidence level is obtained by fusing company cargo information associated with POI, historical waybill data, and vehicle qualification information based on the Bayesian method.

2. The virtual waybill generation method based on multi-source data fusion according to claim 1, characterized in that, For each candidate stop point in the candidate stop area, perform electronic fence matching to determine associated points of interest, including: For each candidate stop point, check whether it has entered the preset fence: if the candidate stop point is inside the fence and the stay time meets the preset time, then match the electronic fence icon; For candidate stops that do not match the electronic fence identifier, the boundary range of the electronic fence is dynamically adjusted according to the spatial distribution variance of historical trajectory data; and a dynamic electronic fence is generated for the candidate stops that do not match the electronic fence identifier based on the convex hull algorithm.

3. The virtual waybill generation method based on multi-source data fusion according to claim 1, characterized in that, Based on the loading and unloading event identification results, the cargo category inference results, and multi-dimensional impact data, a fusion reasoning is performed to generate a freight revenue prediction result, including: The actual transportation distance is calculated based on the location coordinates in the loading and unloading event identification results. Based on the actual transportation distance, the predicted cargo category, and historical transaction price data, the benchmark freight rate is calculated using a pre-trained random forest model. The benchmark freight rate, regional economic indicators, fuel price data, vehicle model data, and seasonal fluctuation factor are input into the large language model. Combined with the contextual information enhanced by RAG retrieval, interval reasoning is performed to generate freight revenue prediction results. The seasonal fluctuation factor is obtained by querying the fixed seasonal factor table and the dynamic event table and fusing them according to the preset weight ratio.

4. The virtual waybill generation method based on multi-source data fusion according to claim 3, characterized in that, The process of determining the seasonal fluctuation factor includes: Fixed seasonal factors are determined based on the basic data of the fixed seasonal factor table; the basic data of the fixed seasonal factor table includes at least event tags within a preset time period, basic floating range, regional weighting, and commodity type correction; The dynamic event factor is calculated by weighting and averaging dynamic event data and combining it with an LLM model; the dynamic event data includes at least the trigger time, event description, impact coefficient, duration in days, and event source. The seasonal fluctuation factor is calculated by weighting and summing the fixed seasonal factor and the dynamic event factor according to preset weights.

5. The virtual waybill generation method based on multi-source data fusion according to claim 1, characterized in that, Based on the loading and unloading event identification results, the cargo category estimation results, and the freight revenue prediction results, confidence level fusion and structured processing are performed to generate a virtual waybill, including: Assign preset weights to key fields in the loading and unloading event identification results, the cargo category inference results, and the freight revenue prediction results; Based on the pre-defined weights assigned to key fields and the confidence levels of each key field, the overall confidence level of the virtual waybill is calculated using a weighted geometric average algorithm. The confidence scores of fields with logical conflicts are attenuated and corrected according to a correction rule. The correction rule is to multiply the confidence scores of fields with logical conflicts by a decay coefficient to obtain the corrected fields. The corrected fields are combined into a structured waybill according to a preset format; the structured waybill includes at least the waybill number, license plate number, carrier, place of origin, place of destination, cargo category, cargo weight, loading time, unloading time, mileage, estimated freight, and data confidence score.

6. The virtual waybill generation method based on multi-source data fusion according to claim 5, characterized in that, The method further includes: The vehicle trajectory data is subjected to anomaly pattern detection to identify abnormal behavior and to provide risk warnings for such abnormal behavior; the abnormal behavior includes at least detour behavior or abnormal unloading behavior midway. Virtual waybills with an overall confidence level below a preset threshold are marked, and a manual review process is triggered for the marked virtual waybills; all generated virtual waybills and abnormal behaviors are stored on the blockchain.

7. A virtual waybill generation system based on multi-source data fusion, characterized in that, The system is applied to the virtual waybill generation method based on multi-source data fusion as described in any one of claims 1-6, and the system comprises: The target multi-source data acquisition module is configured to acquire target multi-source data; the target multi-source data includes at least vehicle trajectory data, map point of interest data, and electronic fence data; The data fusion module is configured to fuse the target multi-source data to generate loading and unloading event identification results; The cargo category estimation module is configured to perform retrieval enhancement generation and semantic reasoning based on the loading and unloading event identification results, vehicle information, and publicly available information associated with map points of interest, to generate cargo category estimation results. The freight revenue forecasting module is configured to perform fusion reasoning based on the loading and unloading event identification results, the cargo category inference results, and multi-dimensional impact data to generate a freight revenue forecasting result; the multi-dimensional impact data includes at least regional economic indicators, fuel price data, vehicle model data, and seasonal fluctuation factors. The waybill generation module is configured to perform confidence fusion and structured processing based on the loading and unloading event identification results, the cargo category inference results, and the freight revenue prediction results to generate a virtual waybill.

8. A virtual waybill generation device based on multi-source data fusion, characterized in that the device... include: Memory, processor, and communication interface coupled to the processor; The memory stores computer programs that can be executed by the processor; When the processor runs the computer program, it executes the virtual waybill generation method based on multi-source data fusion as described in any one of claims 1 to 6.

Images