A neural network-based logistics transportation real-time early warning collaborative evaluation method and system

Abstract

The application provides a logistics transportation real-time early warning collaborative evaluation method and system based on a neural network, relates to the technical field of data processing, and comprises the following steps: optimizing time series data flow according to a dynamic adjustment strategy, inputting the optimized time series data flow into a pre-trained neural network model for analysis, identifying vehicle deviation from a common route event, and obtaining abnormal event data; matching the abnormal event data with a preset early warning rule set to generate early warning information; obtaining collaborative disposal instructions according to the early warning information, order data, cargo data and travel data, performing a notification operation based on the collaborative disposal instructions to obtain operation execution results; obtaining visual update data according to the operation execution results and the early warning information, feeding the visual update data back to a visual management page, and completing closed-loop management. The application can realize closed-loop management of abnormal early warning and collaborative disposal, and improves the monitoring efficiency and risk response capability of logistics transportation.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for real-time early warning and collaborative evaluation of logistics transportation based on neural networks. Background Technology

[0002] In road transport scenarios involving high-value bulk cargo (such as specialty minerals and high-purity metal ingots), existing technologies typically employ a security solution primarily based on satellite positioning trajectory monitoring, supplemented by electronic seals or door magnetic sensors. This type of solution can record vehicle routes, speeds, and the opening and closing status of cargo doors, forming a basic transport log. However, in actual operation, when carriers intend to carry out concealed partial cargo replacement (e.g., during a transport journey of hundreds of kilometers, briefly stopping on a remote, unmonitored section of road to partially open the cargo compartment and replace a small amount of cargo), existing technologies often face the following specific limitations: trajectory and status monitoring have perception blind spots and interpretability ambiguities; short-term deviations from the preset route or abnormal stops may be attributed by the system to routine rest, traffic congestion, or temporary detours, making it difficult to automatically distinguish whether it constitutes malicious behavior; simultaneously, existing electronic... Lead seals or magnetic door sensors primarily monitor the open or closed status of the truck doors. If a small amount of cargo is accessed or stored through non-primary door lock locations such as cargo vents or side panels, it may not trigger an effective open / closed status alarm, thus creating a vulnerability in the physical monitoring layer. Secondly, the warning logic relies on static rules, making it difficult to adapt to dynamic and complex transportation scenarios. Existing systems typically make warning judgments based on fixed thresholds (such as electronic fences or maximum dwell time limits), lacking the ability to fuse, analyze, and collaboratively model multi-dimensional data (such as order information, cargo characteristics, historical behavior patterns, and real-time environmental factors). For example, the system cannot correlate a brief parking event that does not trigger a door lock alarm with the high-value attributes of the cargo being transported or the frequent occurrence of historical anomalies on that route, thus making it difficult to provide effective warnings for potential, planned cargo replacement behaviors.

[0003] Due to the aforementioned limitations, current transportation monitoring systems often suffer from insufficient sensitivity in identifying small-scale goods quality substitutions or thefts that occur during transportation, and their early warnings are often delayed. Cargo owners usually only discover quality discrepancies or shortages after the goods arrive at their destination and undergo sampling or full inspection. By this time, not only have economic losses already occurred, but the lack of evidence during transit also makes it difficult to determine liability and pursue compensation. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a real-time early warning and collaborative evaluation method and system for logistics transportation based on neural networks, which can realize closed-loop management of abnormal early warning and collaborative handling, and improve the monitoring efficiency and risk response capability of logistics transportation.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a real-time early warning and collaborative evaluation method for logistics transportation based on neural networks, the method comprising:

[0007] Step 1: Collect multi-source data during the logistics and transportation process. The multi-source data includes trip data, cargo data, order data, and interaction data, and integrate them into a time-series data stream.

[0008] Step 2: Obtain the time series data point set based on the time series data stream; construct a multi-dimensional mapping surface based on the time series data point set; delineate the initial analysis region based on the multi-dimensional mapping surface; divide the initial analysis region according to the time series data point set to obtain multiple processing partitions;

[0009] Step 3: Map the time series data point set to the corresponding processing partition; calculate the Hu invariant moments of the time series data point set in each processing partition; normalize the central moments by calculating the central moments of the time series data point set; and obtain the Hu invariant moment set based on the normalized central moments to generate a dynamic adjustment strategy.

[0010] Step 4: Optimize the time-series data stream according to the dynamic adjustment strategy, input the optimized time-series data stream into the pre-trained neural network model for analysis, identify vehicle deviation from the common route events, and obtain abnormal event data;

[0011] Step 5: Match the abnormal event data with the preset set of early warning rules to generate early warning information;

[0012] Step 6: Based on the early warning information, order data, cargo data, and travel data, obtain the collaborative handling instruction, execute the notification operation based on the collaborative handling instruction, and obtain the operation execution result;

[0013] Step 7: Based on the operation execution results and warning information, obtain the visual update data and feed it back to the visual management page to complete the closed-loop management.

[0014] Secondly, a real-time early warning and collaborative evaluation system for logistics transportation based on neural networks includes:

[0015] The data acquisition module is used to collect multi-source data during the logistics and transportation process. The multi-source data includes trip data, cargo data, order data, and interaction data, and integrates them into a time-series data stream.

[0016] The partitioning module is used to obtain a time-series data point set based on the time-series data stream; construct a multi-dimensional mapping surface based on the time-series data point set; delineate the initial analysis region based on the multi-dimensional mapping surface; and partition the initial analysis region based on the time-series data point set to obtain multiple processing partitions.

[0017] The calculation module is used to map the time series data point set to the corresponding processing partition; calculate the Hu invariant moments of the time series data point set in each processing partition; normalize the central moments by calculating the central moments of the time series data point set; and obtain the Hu invariant moment set based on the normalized central moments to generate a dynamic adjustment strategy.

[0018] The optimization module is used to optimize the time-series data stream according to the dynamic adjustment strategy, input the optimized time-series data stream into the pre-trained neural network model for analysis, identify vehicle deviation from the common route events, and obtain abnormal event data.

[0019] The matching module is used to match abnormal event data with a set of preset early warning rules to generate early warning information;

[0020] The execution module is used to obtain collaborative handling instructions based on early warning information, order data, cargo data, and travel data, execute notification operations based on the collaborative handling instructions, and obtain the operation execution results;

[0021] The output module is used to obtain visual update data based on the operation execution results and early warning information, and to feed the visual update data back to the visual management page to complete closed-loop management.

[0022] The above-described solution of the present invention has at least the following beneficial effects:

[0023] By calculating and processing Hu invariant moments of time-series data points within a partition to generate a dynamic adjustment strategy, the essential characteristics and changing patterns of transportation data can be accurately captured, rather than being limited to fixed threshold judgments. At the same time, by combining a pre-trained neural network model to analyze the optimized time-series data stream, multi-dimensional data fusion modeling and collaborative evaluation are realized. The early warning logic can be dynamically adjusted according to the high-value attributes of goods, historical anomalies of road sections, and historical vehicle behavior patterns, effectively distinguishing normal events such as regular rest and traffic congestion from malicious abnormal behaviors, reducing the problems of early warning lag and high false alarm and missed alarm rates, and improving the reliability and pertinence of early warning results. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of a real-time early warning and collaborative evaluation method for logistics transportation based on neural networks, provided by an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of a real-time early warning and collaborative evaluation system for logistics transportation based on neural networks, provided by an embodiment of the present invention. Detailed Implementation

[0026] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0027] like Figure 1 As shown in the figure, an embodiment of the present invention proposes a real-time early warning and collaborative evaluation method for logistics transportation based on neural networks. The method includes the following steps:

[0028] Step 1: Collect multi-source data during the logistics and transportation process. The multi-source data includes trip data, cargo data, order data, and interaction data, and integrate them into a time-series data stream.

[0029] Step 2: Obtain the time series data point set based on the time series data stream; construct a multi-dimensional mapping surface based on the time series data point set; delineate the initial analysis region based on the multi-dimensional mapping surface; divide the initial analysis region according to the time series data point set to obtain multiple processing partitions;

[0030] Step 3: Map the time series data point set to the corresponding processing partition; calculate the Hu invariant moments of the time series data point set in each processing partition; normalize the central moments by calculating the central moments of the time series data point set; and obtain the Hu invariant moment set based on the normalized central moments to generate a dynamic adjustment strategy.

[0031] Step 4: Optimize the time-series data stream according to the dynamic adjustment strategy, input the optimized time-series data stream into the pre-trained neural network model for analysis, identify vehicle deviation from the common route events, and obtain abnormal event data;

[0032] Step 5: Match the abnormal event data with the preset set of early warning rules to generate early warning information;

[0033] Step 6: Based on the early warning information, order data, cargo data, and travel data, obtain the collaborative handling instruction, execute the notification operation based on the collaborative handling instruction, and obtain the operation execution result;

[0034] Step 7: Based on the operation execution results and warning information, obtain the visual update data and feed it back to the visual management page to complete the closed-loop management.

[0035] In this embodiment of the invention, by calculating the Hu invariant moments of the time-series data point set within the partition to generate a dynamic adjustment strategy, the essential characteristics and changing patterns of transportation data can be accurately captured, rather than being limited to fixed threshold judgments. At the same time, by combining a pre-trained neural network model to analyze the optimized time-series data stream, multi-dimensional data fusion modeling and collaborative evaluation are realized. The warning logic can be dynamically adjusted according to the high-value attributes of goods, historical anomalies of road sections, and historical vehicle behavior patterns, effectively distinguishing normal events such as regular rest and traffic congestion from malicious abnormal behaviors, reducing the problems of warning lag and high false alarm / missed alarm rates, and improving the reliability and pertinence of warning results.

[0036] In a preferred embodiment of the present invention, step 1 involves collecting multi-source data during the logistics and transportation process. This multi-source data includes trip data, cargo data, order data, and interaction data, and integrating them into a time-series data stream. This may include:

[0037] Step 101: Obtain raw trip data, raw cargo data, raw order data, and raw interaction data. Raw trip data includes location coordinates, speed, and timestamps obtained from the vehicle positioning unit. Raw cargo data includes cargo weight and image data inside and outside the cargo compartment obtained from the cargo monitoring unit. Raw order data includes order number, delivery address, and delivery deadline obtained from the order management unit. Raw interaction data includes call records, driver confirmation status, and abnormal feedback content obtained from the intelligent outbound calling unit. Specifically, this includes: real-time collection of vehicle location coordinates (latitude and longitude information) and real-time vehicle speed using GPS positioning devices installed on vehicles transporting high-value bulk cargo, while recording the timestamps of each data collection accurate to the second; these data collectively constitute the raw trip data; real-time weight data of the cargo is collected using weight sensors fixed inside the cargo compartment; high-definition cameras are installed around the cargo compartment, on the top, and around the doors; infrared cameras are installed at key locations inside the cargo compartment. These devices collect real-time image data inside and outside the cargo compartment. High-definition cameras capture images within the visible light range, while infrared cameras assist in capturing images at night or in low-light environments. The focus is on capturing images related to violations such as climbing on the cargo compartment, prying open compartment components, abnormally opening compartment gaps, and passing items inside or outside the compartment. This data collectively constitutes the original cargo data. Through the company's internal transportation order management platform, the system retrieves order information corresponding to current high-value bulk cargo transportation, including a unique order number automatically assigned by the platform, detailed delivery address information down to the province, city, district, street, and house number, and a delivery deadline specified in the year, month, day, hour, minute, and second. This information constitutes the original order data. Through intelligent outbound calling devices integrating voice call and recording functions, the system records complete recordings and corresponding text transcriptions of each call between the system and the driver. It records the driver's confirmation of key transportation nodes such as loading and unloading during the call, and records various abnormal situations encountered during transportation, such as traffic jams, queues, vehicle malfunctions, etc., reported by the driver. This content constitutes the interactive raw data.

[0038] Step 102: Based on the original itinerary data, obtain structured itinerary data; based on the original cargo data, obtain structured cargo data; based on the original order data, obtain structured order data; based on the original interaction data, obtain structured interaction data. Structured itinerary data includes formatted time, latitude and longitude, and instantaneous speed; structured cargo data includes cargo status tags and numerical physical quantities; structured order data includes order timestamps and address vectors; structured interaction data includes interaction time and classification codes. Specifically, this includes: standardizing the format of the original itinerary data, converting the timestamps to a unified standard format of year, month, day, hour, minute, and second as the formatted time, and converting the original latitude and longitude to degrees and minutes... The second format is converted to decimal format by adding the degree value to the minute value, dividing by 60, then adding the second value to the third thousand six hundred, retaining six decimal places as the standard latitude and longitude. The original speed in meters per second is converted to kilometers per hour by multiplying the meter per second value by 3.6, retaining one decimal place as the instantaneous speed. The formatted time, standard latitude and longitude, and instantaneous speed are integrated to form structured travel data. The raw cargo data is classified and quantified. Considering the safety requirements for transporting high-value bulk cargo (such as nickel iron and alumina), high-definition cameras are installed around the cargo compartment, on the top, and around the doors. Infrared cameras are installed in key locations inside the cargo compartment. These devices are used to... The system continuously collects video data from inside and outside the cargo compartment. During the data collection process, high-definition cameras capture images within the visible light range, while infrared cameras assist in capturing images at night or in low-light environments. The system focuses on monitoring violations such as climbing into the cargo compartment, prying open compartment components, abnormally opening compartment gaps, and passing items inside or outside the compartment. If no violations are detected in the images collected by all cameras, the cargo is labeled as normal. If any camera captures any of the aforementioned violations, the cargo is labeled as abnormal. Simultaneously, the weight data collected by the weight sensors is converted to tons and rounded to three decimal places. The effectiveness of the camera data collection is quantified, and the number of effective video frames per second is counted. (Default 30 frames per second), retaining integers; record the duration of abnormal actions, starting from the first detection of the violation and ending with the action, retaining two decimal places in seconds; integrate cargo status tags, weight quantification data, effective image frame count data, and abnormal action duration data to form structured cargo data; perform coordinate and standardization processing on the original order data, converting the delivery deadline in the order into a standard format of year, month, day, hour, minute, and second as the order time stamp, and using geocoding technology to convert the textual delivery address into corresponding decimal latitude and longitude coordinates to form an address vector containing both longitude and latitude values; integrate the order time stamp and address vector to form structured order data;The raw interactive data undergoes time standardization and encoding. The time of the call is converted into a standard format of year, month, day, hour, minute, and second as the interaction time. This time is then coded according to the company's pre-defined transportation anomaly classification standards: confirmed loading is coded as number one, unconfirmed loading as number two, confirmed unloading as number three, unconfirmed unloading as number four, traffic congestion as number five, queue congestion as number six, vehicle malfunction as number seven, and other anomalies as number eight. The interaction time and classification codes are then integrated to form structured interactive data.

[0039] Step 103: Based on the structured itinerary data, structured cargo data, structured order data, and structured interaction data, extract time information and align it to obtain a time-aligned multi-source data sequence. Specifically, this includes: extracting the formatted time corresponding to each data point from the structured itinerary data; extracting the collection time corresponding to each record from the structured cargo data; extracting the order timestamp from the structured order data; and extracting the interaction time from the structured interaction data. All these time information points are then uniformly adjusted to the same recording standard accurate to the second. A baseline timeline is generated based on the formatted time of the structured itinerary data, with the starting time of the baseline timeline being the minimum formatted time in the structured itinerary data. The values, with the end time being the maximum value of the formatted time, are arranged sequentially at one time point per second to form a continuous reference timeline. The time corresponding to each piece of data in the structured goods data, structured order data, and structured interaction data is compared one by one with each time point on the reference timeline to find a matching point where the time is completely consistent. All information corresponding to that data is then associated with that time point. If the time of a piece of data is not completely consistent with the time on the reference timeline, the time point on the reference timeline that is closest to that time is found, and the data is associated with that closest time point. This ensures that each time point on the reference timeline corresponds to one of the four types of data: travel, goods, orders, and interactions, forming a time-aligned multi-source data sequence.

[0040] Step 104: Based on the time-aligned multi-source data sequence, perform interpolation and merging to obtain a time-series data stream. Specifically, this includes: performing an integrity check on the time-aligned multi-source data sequence, examining each time point on the baseline timeline to see if it contains latitude and longitude, instantaneous velocity, cargo status tags, numerical physical quantities (cargo weight, number of valid image frames, duration of abnormal actions), address vectors, and full-dimensional data of classification codes; if a certain baseline time point is found to lack a certain type of data, such as missing valid image frames, then find the time point preceding and following that time point that contains valid image frame data, calculate the time interval between these two valid time points, and calculate the time interval between the time point with missing data and the previous valid time point. The time interval is calculated by adding the effective image frame count of the previous effective time point to the effective image frame count of the next effective time point and subtracting the effective image frame count of the previous effective time point. This result is then multiplied by the time interval between the missing data time point and the previous effective time point, divided by the ratio of the time interval between the two effective time points. This yields the supplementary data for the effective image frame count of the missing time point. Following the same method, all missing data at the baseline time point (including camera-related data such as the duration of abnormal actions) are supplemented to ensure that each time point has complete full-dimensional data. All the supplemented time point data are then arranged sequentially according to the time order of the baseline time axis, and the full-dimensional data of each time point are merged to form a continuous and uninterrupted time-series data stream.

[0041] This embodiment achieves structured processing of various types of raw data through explicit format conversion standards and specific numerical calculation methods. By using precise time alignment operations and missing data interpolation to supplement data, it ensures the integrity and consistency of the data, ultimately forming a continuous and complete time-series data stream, thus solving the problem of single-format, chaotic, and missing data in data acquisition.

[0042] In a preferred embodiment of the present invention, step 2 involves obtaining a time-series data point set based on the time-series data stream; constructing a multi-dimensional mapping surface based on the time-series data point set; defining an initial analysis region based on the multi-dimensional mapping surface; and dividing the initial analysis region according to the time-series data point set to obtain multiple processing partitions, which may include:

[0043] Step 201: Based on the time-series data stream, perform sliding sampling according to a preset time window to obtain a time-series data point set composed of multi-dimensional data vectors within each time window. Specifically, this includes: combining the typical transportation time and data change frequency of high-value bulk cargo transportation, setting the preset time window to five minutes (300 seconds), and setting the sliding step size to ten seconds; starting from the start time of the time-series data stream, extracting all data within the first time window. The start time of this window is the start time of the time-series data stream, and the end time is the start time plus 300 seconds. Then, extract the data for each time point within the window. The latitude and longitude, instantaneous speed, cargo weight, number of effective image frames, duration of abnormal actions, cargo status tags, order timestamps, address vectors, classification codes, and other multi-dimensional data corresponding to each point are integrated into a multi-dimensional data vector, with each dimension corresponding to one data item. Following a set ten-second sliding step, the time window is slid forward by ten seconds, capturing all data within the new window and integrating it into a multi-dimensional data vector. This sliding sampling operation is repeated until the time window covers the end time of the time-series data stream. All the multi-dimensional data vectors corresponding to the sliding windows together form the time-series data point set.

[0044] Step 202: Based on the time-series data point set, select key feature dimensions for projection to obtain a projection point set; based on the spatial distribution of the projection point set, select reference points as fitting references; based on the spatial location of the reference points, calculate the connection relationship between the reference points to generate a fitting grid; based on the distribution of the fitting grid and the projection point set, calculate the surface fitting parameters; construct a multi-dimensional mapping surface based on the surface fitting parameters, specifically including: selecting key feature dimensions closely related to the safety of high-value bulk cargo transportation from the multi-dimensional data vector of the time-series data point set, including latitude and longitude converted plane coordinates, instantaneous speed, cargo weight, number of effective image frames, duration of abnormal actions, distance from the preset route, and dwell time, and extracting the data of these key feature dimensions separately from each multi-dimensional data vector; The Gauss-Kruger projection method is used to convert latitude and longitude to plane coordinates. The Earth's radius is set to 6378.137 kilometers, the central meridian longitude is 110 degrees, and the projection zone is 3 degrees. First, the latitude and longitude values ​​are converted to radians by multiplying the longitude value by pi and then dividing by 180, and the latitude value by multiplying by pi and then dividing by 180. Then, the plane coordinates are calculated using the projection formula. The x-component is calculated by multiplying the Earth's radius by (radian longitude minus the central meridian radian value) and then by the cosine (radian latitude). The y-component is calculated by multiplying the Earth's radius by the natural logarithm (tangent (a quarter of pi plus a half of radian latitude)). Finally, both the x-component and y-component are retained to three decimal places. Longitude corresponds to the x-component of the plane coordinates, and latitude corresponds to the y-component of the plane coordinates.

[0045] The three core dimensions are defined as the x-component of the planar coordinate system, the y-component of the planar coordinate system, and the number of effective image frames. The values ​​of these three dimensions of each data vector are mapped to a specific point in three-dimensional space, with each data vector corresponding to a specific point in the three-dimensional space. All these points together form a projection point set. The distribution of the projection point set in three-dimensional space is observed. Points that are evenly distributed and completely cover the entire distribution range of the projection point set are selected as reference points at fixed intervals of ten projection points. Each reference point retains the corresponding values ​​of the three core dimensions. For any two reference points, their coordinates in three-dimensional space are calculated. The straight-line distance is calculated as follows: Subtract the x-component of the second reference point from the x-component of the first reference point, and square this difference. Subtract the y-component of the second reference point from the y-component of the first reference point, and square this difference. Subtract the z-component (effective image frame count) of the second reference point from the z-component of the first reference point, and square this difference. Add the results of these three square calculations together, and then take the square root of this sum to obtain the straight-line distance between the two reference points. The result is rounded to two decimal places. A distance threshold of 5 kilometers is set. If the straight-line distance between any two reference points is less than or equal to 5 kilometers, then these two reference points are connected by a straight line. All interconnected reference points interweave to form a mesh-like fitting grid. For each point in the projection point set, the distance from that point to the nearest grid line in the fitting grid is calculated. The calculation method is as follows: first, determine the foot of the perpendicular from the point to the nearest grid line using the perpendicular theorem; then, calculate the straight-line distance between the point and the foot of the perpendicular, following the same calculation method as for the straight-line distance between reference points. The distance values ​​from all projection points to their corresponding nearest grid lines are counted, and all distance values ​​are summed. Then, the sum is divided by the total number of points in the projection point set to obtain the final value. The average of these distance values ​​is calculated and rounded to three decimal places. The node positions of the fitted grid are adjusted based on this average distance. The x, y, and z components (corresponding to the effective image frame number) of each grid node are adjusted to the original value plus the average distance divided by 10. The adjusted data is rounded to three decimal places. At the same time, the curvature coefficient of the grid lines is calculated by dividing the average distance by the original spacing of the grid nodes. This yields surface fitting parameters that include the adjusted coordinates of the grid nodes and the curvature coefficient of the grid lines. Based on these surface fitting parameters, a multidimensional mapping surface that closely fits the distribution trend of the projection point set is constructed.

[0046] Step 203: Based on the boundary coordinates of the data points on the multidimensional mapping surface, calculate the minimum bounding region containing all data points to obtain the initial analysis region. Specifically, this includes: traversing all projection points on the multidimensional mapping surface sequentially according to their generation time, recording the x-axis, y-axis, and z-axis coordinates of each projection point in three-dimensional space, with all coordinate results rounded to three decimal places; after traversal, comparing the x-axis coordinate values ​​one by one to select the maximum and minimum values; using the same comparison method, selecting the maximum and minimum values ​​for the y-axis and z-axis coordinates respectively. These six extreme coordinates, selected from small values, collectively constitute the boundary coordinates of the data points on the multidimensional mapping surface. Based on these six boundary coordinates, a cubic region is constructed in three-dimensional space. The starting coordinate of the x-axis is the minimum boundary coordinate of the x-axis, and the ending coordinate is the maximum boundary coordinate of the x-axis; the starting coordinate of the y-axis is the minimum boundary coordinate of the y-axis, and the ending coordinate is the maximum boundary coordinate of the y-axis; the starting coordinate of the z-axis is the minimum boundary coordinate of the z-axis, and the ending coordinate is the maximum boundary coordinate of the z-axis. Since each edge of this cube is parallel to the three-dimensional coordinate axes, and each face exactly fits the extreme boundary containing all projection points, there is no smaller cube that can completely contain all projection points. Therefore, this cube is the smallest cube containing all projection points, and this cubic region is the initial analysis region.

[0047] Step 204: Based on the density distribution of the time-series data points within the initial analysis region, the initial analysis region is divided into multiple processing partitions. Specifically, this includes: determining the lengths of the x-axis, y-axis, and z-axis of the initial analysis region. The calculation method is as follows: the x-axis length equals the maximum boundary coordinate minus the minimum boundary coordinate; the y-axis length equals the maximum boundary coordinate minus the minimum boundary coordinate; and the z-axis length equals the maximum boundary coordinate minus the minimum boundary coordinate. The length results are rounded to three decimal places. Considering the anomaly monitoring accuracy requirements for high-value bulk cargo transportation, the number of segments is determined. The x-axis is divided into segments of 50 kilometers each. If the x-axis length... If the length is not divisible by 50 kilometers, the remaining value is used for the last segment. The y-axis is divided into segments of 50 kilometers each, following the same division rule as the x-axis. The z-axis is divided into segments of 40 kilometers each, following the same division rule as the x-axis. For example, if the initial analysis region has an x-axis length of 1200 kilometers, it is divided into 24 segments of 50 kilometers each. The y-axis length is 550 kilometers, divided into 11 segments, with the first 10 segments each being 50 kilometers and the last segment also being 50 kilometers. The z-axis length is 880 kilometers, divided into 22 segments of 40 kilometers each. Through the above segmentation method, the initial analysis region is divided into multiple small cubic subspaces of uniform size or with only slight differences in the end segments.

[0048] The number of data points in the time-series data point set contained within each small cubic subspace is counted one by one. Simultaneously, the volume of each small cubic subspace is calculated by multiplying the actual length of the small cube along the x-axis by its actual length along the y-axis, and then by its actual length along the z-axis. The volume unit is cubic kilometers, and the result is rounded to three decimal places. The data density of each subspace is obtained by dividing the number of data points in each small cubic subspace by the volume of that subspace. The data density unit is the number of data points per cubic kilometer, and the result is rounded to three decimal places. Six-digit; referencing the density distribution characteristics of historical data on high-value bulk cargo transportation, three distinct density thresholds are set: a high density threshold of 80 units per cubic kilometer, a medium density threshold of 30 units per cubic kilometer, and a low density threshold of 10 units per cubic kilometer; subspaces with a data density greater than 80 units per cubic kilometer are classified as high-density subspaces, subspaces with a data density between 30 and 80 units per cubic kilometer are classified as medium-density subspaces, and subspaces with a data density less than 10 units per cubic kilometer are classified as low-density subspaces.

[0049] All small cubic subspaces within the same category are integrated, prioritizing the merging of adjacent subspaces of the same category to form continuous regions. Isolated subspaces of the same category are retained separately. The integrated region corresponding to each category is a processing partition, ultimately forming high-density, medium-density, and low-density processing partitions. For high-value bulk cargo transportation areas with frequent data changes and high anomaly risks, such as high-density processing partitions corresponding to areas with frequent route deviations or risk areas, the segmentation standards are further refined. The segment lengths of the x, y, and z axes within this region are halved, i.e., the x-axis is divided into 25-kilometer segments, the y-axis into 25-kilometer segments, and the z-axis into 20-kilometer segments. This makes the small cubic subspaces within this region more densely divided, ensuring accurate capture of anomaly data characteristics. For low-density processing partitions with gradual data changes and low anomaly risks, the original segmentation standards remain unchanged, ensuring both analytical accuracy and overall analytical efficiency.

[0050] This embodiment constructs a multi-dimensional mapping surface that fits the data distribution by using a clear projection method, a reference point selection rule, a distance calculation method, and a grid generation logic. Through refined regional division and density calculation standards, it achieves differentiated partitioning based on data density, which not only ensures the analysis accuracy of concentrated areas of abnormal data but also reasonably controls the overall analysis cost.

[0051] In a preferred embodiment of the present invention, step 3 involves mapping the time-series data point set to the corresponding processing partition; calculating the Hu invariant moments of the time-series data point set within each processing partition; normalizing the central moments of the time-series data point set by calculating the central moments; and obtaining the Hu invariant moment set based on the normalized central moments to generate a dynamic adjustment strategy, which may include:

[0052] Step 301: Based on the time-series data point set and the spatial range of each processing partition, assign each data point to its corresponding processing partition to obtain a subset of data points for each processing partition. Specifically, this includes: retrieving the spatial range parameters of the high-density, medium-density, and low-density processing partitions through the logistics transportation monitoring system. Each partition's parameters include the x-axis start coordinate, x-axis end coordinate, y-axis start coordinate, y-axis end coordinate, z-axis start coordinate, and z-axis end coordinate. These coordinates are derived from the partition boundary data obtained in Step 204, and are retained to three decimal places. Then, iterate through each data point in the time-series data point set, extracting the x-axis, y-axis, and z-axis coordinates of that data point in sequence, retaining the coordinate values ​​to three decimal places. The coordinates of data points are compared with the spatial range of high-density processing zones. If the x-axis coordinate of a data point is greater than or equal to the starting x-axis coordinate and less than or equal to the ending x-axis coordinate of that zone, and the y-axis and z-axis coordinates also meet the same interval conditions, then the data point is directly determined to belong to the high-density processing zone. If not, the comparison continues to the medium-density processing zone. If the conditions are met, the data point is assigned to the medium-density processing zone. If the conditions are still not met, the data point is assigned to the low-density processing zone. For key data points in high-value bulk cargo transportation, such as data points near risk areas, data points with abnormal effective image frame counts, or data points with abnormal action durations exceeding the standard, the accuracy of the coordinates is checked first during the comparison process to ensure that their assignment is correct. All data points belonging to the same zone are aggregated and integrated to form a subset of data points corresponding to each processing zone.

[0053] Step 302: For each data point subset of the processing partition, calculate the mean coordinates of the data point subset to obtain the center point coordinates; based on the center point coordinates, calculate the zero-order central moment, first-order central moment, and second-order central moment of the data point subset in the two-dimensional spatial distribution. Specifically, this includes: for each data point subset of the processing partition, extract the x-axis and y-axis coordinates of all data points, keeping the coordinate values ​​to six decimal places; calculate the mean x-axis coordinates by summing the x-axis coordinate values ​​of all data points sequentially, then dividing the sum by the total number of data points in the subset, keeping the result to six decimal places; calculate the mean y-axis coordinates using the same method, also keeping six decimal places, and the two means together constitute the center point coordinates of the data point subset; since high-value bulk cargo transportation data needs to ensure the authenticity of the original features, the weight of each data point is set to one, and the zero-order central moment is the zero-order central moment of the data point subset. The total number of data points; when calculating the first-order central moment, first calculate the difference between the x-axis coordinate value of each data point and the mean x-axis coordinate of the center point to obtain the x-axis deviation value of each data point. Add all x-axis deviation values ​​sequentially to obtain the first-order central moment in the x-direction. Similarly, calculate the difference between the y-axis coordinate value of each data point and the mean y-axis coordinate of the center point to obtain the y-axis deviation value. Add all y-axis deviation values ​​to obtain the first-order central moment in the y-direction. When calculating the second-order central moment, the second-order central moment in the x-direction is obtained by multiplying the x-axis deviation value of each data point by itself to obtain the squared deviation value. Add all squared deviation values. The second-order central moment in the xy intersection direction is obtained by multiplying the x-axis deviation value of each data point by its y-axis deviation value. Add all products. The second-order central moment in the y-direction is obtained by multiplying the y-axis deviation value of each data point by itself to obtain the squared deviation value. Add all squared deviation values. All second-order central moment results are rounded to six decimal places.

[0054] Step 303: Based on the zero-order central moment, normalize the first-order and second-order central moments to obtain normalized first-order and normalized second-order central moments. Specifically, this includes: using the zero-order central moment calculated in step 302 as the normalization benchmark, since the zero-order central moment is the total number of data points, which is non-zero and a positive integer, ensuring the stability of the normalized data; the normalized first-order central moment in the x-direction is the value of the first-order central moment in the x-direction divided by the value of the zero-order central moment, and the normalized first-order central moment in the y-direction is the value of the first-order central moment in the y-direction divided by the value of the zero-order central moment; the normalized second-order central moment in the x-direction is the value of the second-order central moment in the x-direction divided by the value of the zero-order central moment, and the normalized second-order central moment in the xy intersection direction is the value of the second-order central moment in the xy intersection direction divided by the value of the zero-order central moment, and the normalized second-order central moment in the y-direction is the value of the second-order central moment in the y-direction divided by the value of the zero-order central moment.

[0055] Step 304: Based on the normalized first-order central moments and normalized second-order central moments, calculate according to the combination of the seven invariant formulas of Hu invariant moments to obtain seven invariant moment values, generating a set of Hu invariant moments. Specifically, the first invariant moment value is the sum of the values ​​of the normalized second-order central moments in the x-direction and y-direction; the second invariant moment value is the sum of (the value of the normalized second-order central moment in the x-direction minus the value of the normalized second-order central moment in the y-direction) multiplied by itself, plus (four times the value of the normalized second-order central moment in the xy intersection direction multiplied by itself); the third invariant moment value is the sum of (the value of the normalized second-order central moment in the x-direction minus the value of the normalized second-order central moment in the y-direction) multiplied by itself. The first invariant moment is calculated as follows: (the value of the normalized second-order central moment in the y-direction multiplied by three, minus the value of the normalized second-order central moment in the x-direction) multiplied by itself, plus (the value of the normalized second-order central moment in the y-direction multiplied by three, minus the value of the normalized second-order central moment in the x-direction) multiplied by itself; the second invariant moment is calculated as follows: (the value of the normalized second-order central moment in the x-direction plus the value of the normalized second-order central moment in the y-direction) multiplied by itself, plus (four times the value of the normalized second-order central moment in the xy intersection direction multiplied by itself); the third invariant moment is calculated as follows: (the value of the normalized second-order central moment in the x-direction minus the value of the normalized second-order central moment in the y-direction) multiplied by (the value of the normalized second-order central moment in the x ... The first product is obtained by multiplying the value of the central moment by three times the value of the normalized second central moment in the y-direction, then multiplying it by (three times the value of the normalized second central moment in the x-direction minus the value of the normalized second central moment in the y-direction). The second product is obtained by multiplying the value of the normalized second central moment in the xy-intersection direction by four times the value of the normalized second central moment in the x-direction plus the value of the normalized second central moment in the y-direction, and then multiplying the result by itself. The two products are then added together. The sixth invariant moment is obtained by multiplying (the value of the normalized second central moment in the x-direction minus the value of the normalized second central moment in the y-direction) by (three times the value of the normalized second central moment in the x-direction). The first invariant moment is obtained by subtracting the value of the normalized second central moment in the y-direction from the value of the first invariant moment, and then adding (four times the value of the normalized second central moment in the xy-intersection direction multiplied by itself). The second invariant moment is obtained by multiplying the value of the normalized second central moment in the xy-intersection direction by (three times the value of the normalized second central moment in the x-direction minus the value of the normalized second central moment in the y-direction), then multiplying by (the value of the normalized second central moment in the x-direction plus three times the value of the normalized second central moment in the y-direction), then adding (the value of the normalized second central moment in the x-direction minus the value of the normalized second central moment in the y-direction) multiplied by itself, and then multiplying by the value of the first invariant moment. All the results are added together. All seven invariant moment values ​​are rounded to eight decimal places and together form the Hu invariant moment set.

[0056] Step 305: Based on the Hu invariant moment set and the trend of data point density change over time in each partition, a dynamic adjustment strategy is generated. Specifically, this includes: analyzing the Hu invariant moment set of each processing partition one by one, setting the judgment threshold for the second to seventh invariant moment values, and setting the benchmark threshold for each invariant moment value as the mean of the invariant moment values ​​corresponding to the historical normal transportation data of the line plus twice the standard deviation. If four or more of the second to seventh invariant moment values ​​in a certain partition exceed the corresponding baseline threshold, and the absolute value of the rate of change of the same invariant moment value in two adjacent time intervals (the invariant moment value of the later time interval minus the invariant moment value of the previous time interval, then divided by the invariant moment value of the previous time interval) is greater than 0.3, it indicates that the spatial distribution of data points in that partition is highly irregular and likely contains anomalous data characteristics of high-value bulk cargo transportation. If two to three invariant moment values ​​exceed the baseline threshold, and the absolute value of the rate of change of one to two invariant moment values ​​is between 0.1 and 0.3, it indicates that the distribution is moderately irregular and there may be potential anomalies. If all invariant moment values ​​do not exceed the baseline threshold and the absolute value of the rate of change is less than 0.1, it indicates that the data point distribution is regular and there is no obvious abnormal trend. Using one hour as a fixed time interval, the data density of the partition in each time interval is calculated, and the absolute value of the rate of change of adjacent invariant moment values ​​is calculated. The density change rate between two time intervals is calculated by subtracting the density value of the previous time interval from the density value of the later time interval, and then dividing by the density value of the previous time interval. This yields the density change trend. If the Hu invariant moments of a high-density processing partition show severe or moderate irregularity, and the density change rate is positive, it indicates that the data in this area changes frequently and has a high risk of anomalies. The dynamic adjustment strategy is set to increase the sampling frequency of the corresponding time-series data stream in this partition from the original base frequency of once per second to twice per second. If the Hu invariant moments of a medium-density processing partition show a regular distribution, and the density change rate fluctuates slightly within the range of ±0.05, it indicates that the data changes smoothly. The strategy is set to maintain the original sampling frequency of once every five seconds. If the Hu invariant moments of a low-density processing partition show a regular distribution, and the density change rate is negative with an absolute value greater than 0.1, it indicates that the amount of data is gradually decreasing. The strategy is set to reduce the sampling frequency, from once every ten seconds to once every fifteen seconds.

[0057] The sensitivity of the warning threshold is adjusted based on the degree of distribution irregularity reflected by Hu invariant moments: for severely irregular zones, the sensitivity of the warning threshold is increased by 50%, that is, the original deviation distance threshold of 500 meters is adjusted to 375 meters, and the original time threshold of 30 minutes is adjusted to 22.5 minutes; for moderately irregular zones, the sensitivity of the warning threshold is increased by 20%, that is, the deviation distance threshold is adjusted to 400 meters, and the time threshold is adjusted to 24 minutes; for zones with regular distribution, the default warning threshold sensitivity is maintained, forming an adjustment strategy adapted to the dynamic data characteristics of high-value bulk cargo transportation.

[0058] This embodiment ensures targeted data processing by accurately allocating data points to corresponding processing partitions. It captures the essential characteristics of data spatial distribution based on the calculation of central moments and Hu invariant moments, and generates dynamic adjustment strategies by combining density time change trends. This enables data processing to flexibly adapt to the dynamic data characteristics of high-value bulk cargo transportation, avoiding the limitations of static processing methods.

[0059] In a preferred embodiment of the present invention, step 4, optimizing the time-series data stream according to a dynamic adjustment strategy, and inputting the optimized time-series data stream into a pre-trained neural network model for analysis to identify vehicle deviation events from the usual route and obtain abnormal event data, may include:

[0060] Step 401: According to the sampling frequency adjustment instruction in the dynamic adjustment strategy, the time-series data stream is resampled in real time to obtain a sampling-optimized time-series data stream. Specifically, this includes: extracting the sampling frequency parameters corresponding to each partition from the dynamic adjustment strategy; the high-density processing partition corresponds to a sampling frequency of once per second, the medium-density processing partition corresponds to a sampling frequency of once every five seconds, and the low-density processing partition corresponds to a sampling frequency of once every ten seconds; for the time-series data stream parts that need to increase the sampling frequency, linear interpolation is used to supplement the data. Assuming the original sampling frequency is once every five seconds, the original data points for a certain feature dimension are: 10.0 at second 0, 20.0 at second 5, and 25.0 at second 10. To increase the frequency to once per second, data points need to be added at seconds 1, 2, 3, 4, 6, and 9. The calculation for the data point added at second 1 is 10.0 + (20.0 - 10.0) × (1 second ÷ 5 seconds), resulting in 10.0 + 2.0 = 12.0. The ratio of the data points added at second 2 is 2 seconds ÷ 5 seconds, and the value... The first second is 10.0 + 4.0 = 14.0; the third second is 10.0 + 6.0 = 16.0; the fourth second is 10.0 + 8.0 = 18.0. All supplementary points are calculated using the same logic. For the time-series data stream portion that requires a reduced sampling frequency, data points are extracted at the new sampling frequency. Assuming the original sampling frequency is once per second, the data for a certain feature dimension from second 0 to second 19 are 5.2, 6.1, 5.8, 7.3, 6.5, 5.9, 7.1, 6.8, 5.7, 6.3, 8.0. The sampling times of 7.5, 6.9, 8.2, 7.8, 6.7, 7.4, 8.1, 7.6, and 6.9 need to be reduced to once every ten seconds. 5.2 at second 0 and 8.0 at second 10 should be extracted. The extraction principle is to use the starting point of each ten-second interval to ensure that the core trend of data change from low to high within that time period is reflected. For the portion maintaining the original sampling frequency, the original data points are directly retained. After resampling, all data points are rearranged in chronological order to form a continuous, uninterrupted time-series data stream after sampling optimization.

[0061] Step 402a: Based on the sampled and optimized time-series data stream, calculate the mean and standard deviation of each dimension. Normalize the time-series data using the mean and standard deviation to obtain standardized time-series data. Specifically, the sampled and optimized time-series data stream contains seven core feature dimensions, corresponding to key monitoring items in high-value bulk cargo transportation: x-component of planar coordinates, y-component of planar coordinates, instantaneous speed, cargo weight, number of effective image frames, distance from the preset route, and dwell time. Taking the x-component of planar coordinates as an example, assume there are 5 data points in this dimension, with values ​​of 100.123456, 100.234567, 100.34, and 100.34 respectively. The data points are 5678, 100.456789, and 100.567890. The total number of data points is 5. Adding all the values ​​sequentially: 100.123456 + 100.234567 + 100.345678 + 100.456789 + 100.567890 = 501.72838. Dividing this sum by the total number of data points (5) gives the mean for this dimension as 100.345676. To calculate the standard deviation, first calculate the deviation of each data point from the mean: 100.123456 - 100.345676 = -0.22222. 67 - 100.345676 = -0.111109, 100.345678 - 100.345676 = 0.000002, 100.456789 - 100.345676 = 0.111113, 100.567890 - 100.345676 = 0.222214; then multiply each deviation value by itself to obtain the squared deviation values, which are 0.0493817284, 0.012345209881, 0.000000000004, 0.012346098769, and 0.049379061796; all deviations The sum of the squared values ​​is 0.12345209885. Dividing the sum by the total number of data points (5) yields 0.02469041977. Taking the square root of this result gives a standard deviation of approximately 0.157131. The values ​​of each dimension for each data point are normalized. For example, subtracting the mean (100.345676) from the value of 100.123456 gives -0.22222. Dividing this by the standard deviation (0.157131) gives a standardized value of approximately -1.414233. All standardized values ​​are rounded to six decimal places. The standardized values ​​of the seven dimensions together constitute the standardized time series data.

[0062] Step 402b: Based on the standardized time-series data, convolution operations are performed through the convolutional layers of the neural network model to extract the spatial and temporal correlation features within each local time window, resulting in a primary spatiotemporal feature sequence. Specifically, this includes: when constructing the convolutional layers, the kernel size is set to 3x3, each kernel contains 9 weight parameters, and the initial values ​​are randomly selected from the range of -0.05 to 0.05 to ensure that the initial weights are appropriate; the number of kernels is set to 64 to cover all key feature combinations; the stride is set to 1, meaning that the kernel slides one data point at a time; the padding method uses the same dimension, with 0 values ​​added at the data edges to ensure that the dimension of the output data after convolution is consistent with the input data. Standardized time-series data is divided into local time windows of 60 seconds, with each window containing standardized data from 60 time points. The window sliding step is set to one second. Taking three consecutive time points within a local time window as an example, each time point contains seven feature dimensions, which are organized into two-dimensional data of 3 rows and 7 columns, and then expanded into a three-dimensional input format with dimensions of time dimension 3, feature dimension 7, and sample dimension 1. The three-dimensional input data is input into a convolutional layer. Taking the first convolutional kernel as an example, its weight parameters are [[0.02, -0.01, 0.03], [0.01, 0.04, -0.02], [0.03, -0.03, 0.02]]. When the convolutional kernel slides on the local data, the calculation of the first position is the multiplication of corresponding elements in the 3 rows and 3 columns: 0.02 × time 1 feature 1 value, -0.01 × time 1 feature 2 value, 0.03 × time 1 feature 3 value, 0.01 × time 2 feature 1 value, 0. The values ​​of 0.04×2 time feature 2, -0.02×2 time feature 3, 0.03×3 time feature 1, -0.03×3 time feature 2, and 0.02×3 time feature 3 are multiplied together, and then a bias term with an initial value of 0.01 is added to obtain the convolution result at that position. The convolution results of the sixty-four convolution kernels form a 60-row, 64-column feature set. Batch normalization is performed on the feature set. 100 feature values ​​are selected from the batch, and their mean is calculated to be 0.02 and the standard deviation is 0.15. Each feature value is reduced by 0.02, divided by 0.15, multiplied by a scaling factor of 1, and then the offset factor of 0 is added to obtain the normalized feature value. The ReLU activation function is used to set the values ​​less than zero in the normalized feature set to zero, while the values ​​greater than zero remain unchanged, resulting in a 60-row, 64-column primary spatiotemporal feature for each local time window. All features are arranged in chronological order to form a primary spatiotemporal feature sequence.

[0063] Step 402c: Based on the primary spatiotemporal feature sequence, the hidden states are updated step-by-step through the recurrent neural network layer of the neural network model to learn the temporal dependencies in the sequence and obtain a high-dimensional spatiotemporal feature vector. Specifically, this includes: when constructing the recurrent neural network layer, a long short-term memory network structure is adopted, with 128 hidden layer units. Each unit contains four core components: a forget gate, an input gate, a cell state, and an output gate; the activation function is the tanh function, calculated as tanh(x). The output range is [-1, 1]. The weights of the forget gate, input gate, and output gate of the Long Short-Term Memory network are initialized using the Xavier method. The input dimension is 64 (the feature dimension after convolution), the hidden layer dimension is 128, and the weight values ​​are randomly selected from a uniform distribution with a distribution range of [-1, 1]. The bias parameters of each gate are fixed at 0.01. A dropout layer is added with a dropout rate of 0.2, randomly discarding 20% ​​of the neuron outputs during training to prevent overfitting. The primary spatiotemporal feature sequence is input into the Long Short-Term Memory (LSTM) network layer step by step. Taking a certain time step as an example, the input feature value is 0.5, the hidden state of the previous time step is 0.3, the forget gate weight is 0.2 (within the range of -0.1768 to 0.1768), and the bias is 0.01. The forget weight is calculated as sigmoid(0.5×0.2+0.3×0.2+0.01)≈0.5425; the input gate weight is 0.3 (within the range), and the bias is 0.01. The update weight is calculated as sigmoid(0.5×0.3+0.3×0.3+0.01)≈ 0.5623; the candidate cell state is tanh(0.5×0.4+0.3×0.4+0.01)≈0.3233; the cell state is updated to the previous cell state 0.4×0.5425+0.3233×0.5623≈0.217+0.1818≈0.3988; the output gate weight is 0.25 (within the value range), the bias is 0.01, and the output weight is sigmoid(0.5×0.25+0.3×0.25+0.01)≈0.5523; the current hidden state is tanh(0.3988)×0.5523≈0.384×0.5523≈0.2121; after traversing all time steps, the 128-dimensional hidden state of the last time step is taken as the high-dimensional spatiotemporal feature vector, and each element is retained to six decimal places.

[0064] Step 402d: Based on the high-dimensional spatiotemporal feature vector and the preset common route pattern feature vector, calculate the cosine similarity between the two feature vectors to obtain the feature similarity value. Specifically, this includes: when constructing the preset common route pattern feature vector, collecting 100,000 normal transportation time-series data points for the transportation route over the past year, covering normal transportation scenarios under different seasons, weather, and road conditions. Each data point is processed according to steps 401, 402a, 402b, and 402c to obtain 100,000 128-dimensional high-dimensional spatiotemporal feature vectors; taking the first dimension of the feature vector as an example, iterate through the elements of that dimension of the 100,000 vectors, assuming that the sum of all element values ​​is 12345.6789, and divide the sum by 100,000 to obtain the mean of that dimension, 0.123456789. Calculate the mean of the remaining 127 dimensions in the same way. The mean of all dimensions constitutes the common route pattern feature vector; when calculating the cosine similarity, take a simplified two-dimensional vector as an example, and let the high-dimensional spatiotemporal feature vector be... = [0.8, 0.6], the feature vector of a common route pattern is = [1.0, 0.0]; Calculate the numerator, which is the dot product of the two vectors: =0.8×1.0+0.6×0.0=0.8; then calculate the denominator and find the magnitudes of the two vectors respectively.

[0065] Multiplying the two moduli gives Finally, the cosine similarity formula is used for calculation: The calculation logic for an actual 128-dimensional vector is the same. Let the high-dimensional spatiotemporal feature vector be... =[ , ,..., The feature vector of a commonly used route pattern is: =[ , ,..., The numerator is the sum of the product of elements in the corresponding dimension: The denominator is the product of the magnitudes of the two vectors. , The final cosine similarity is The result is rounded to four decimal places, and the value ranges from -1 to 1.

[0066] Step 402e involves comparing the feature similarity value with a preset deviation threshold to obtain the deviation judgment result. Specifically, the preset deviation threshold is determined based on the historical operation data of high-value bulk cargo transportation on this route, collecting 10,000 normal transportation records and 5,000 abnormal transportation records from the past two years. Normal transportation records cover different seasons, weather conditions (sunny, rainy, foggy, snowy), and road conditions (highways, national highways, rural roads), involving transportation scenarios of typical high-value bulk cargoes such as nickel iron, alumina, and copper ore. Abnormal transportation records include malicious detours of more than 500 meters from the designated route, stops for more than 20 minutes without reasonable reason, and unauthorized changes to the unloading location. All historical records were processed step by step 401 to 402d. First, the temporal data stream of each record was resampled and optimized. Then, the mean and standard deviation of each dimension were calculated and normalized. Primary spatiotemporal features were extracted through convolutional layers, and high-dimensional spatiotemporal feature vectors were obtained through long short-term memory (LSTM) networks. Finally, feature similarity values ​​were calculated with feature vectors of common route patterns. The feature similarity results of all historical records were statistically analyzed. It was found that the feature similarity values ​​of 10,000 normal transportation records were all 0.7 or higher, while the feature similarity values ​​of 5,000 abnormal transportation records were all below 0.7. Considering the safety monitoring requirements of high-value bulk cargo transportation, a balance was struck between anomaly identification accuracy and prevention. To address the issue of excessive misjudgment, the preset deviation threshold is ultimately set to 0.7. The current transportation feature similarity value calculated in step 402d is obtained. Assuming the current value is 0.66 or 0.73, this value is directly compared with the preset deviation threshold of 0.7. If the current feature similarity value is 0.73, which is greater than or equal to 0.7, it indicates that the spatiotemporal characteristics of the current transportation are highly consistent with the characteristics of the normally used route, and it is determined that it has not deviated from the normally used route. If the current feature similarity value is 0.66, which is less than 0.7, it indicates that the spatiotemporal characteristics of the current transportation are significantly different from the characteristics of the normally used route, and it is determined that it has deviated from the normally used route. In both cases, a clear deviation judgment result is generated.

[0067] Step 403: Based on the deviation determination result, calculate the shortest spatial distance between the vehicle's current position and the commonly used route, and record the duration of the deviation state to obtain abnormal event data containing the deviation distance and deviation time. Specifically, this includes: when the deviation determination result is a deviation from the commonly used route, the planar coordinates of the vehicle's current position are obtained in real time through a GPS positioning device installed on the transport vehicle, with the x-component being 120.128 and the y-component being 30.458, both rounded to three decimal places; the commonly used route consists of continuous coordinate points extracted from the normal transport trajectory of the route over the past year, selected at a fixed interval of 50 meters, with each coordinate point containing the corresponding x-component and y-component, for example, adjacent coordinate points are the first coordinate point. =120.123、 =30.456, the second coordinate point =120.133、 =30.466, the third coordinate point =120.143 =30.476, and all coordinate points are connected sequentially in the order of transportation to form a broken line shape.

[0068] When calculating the shortest spatial distance, each segment of the commonly used route polyline is traversed one by one. Each segment is determined by two adjacent coordinate points. Taking the segment formed by the first and second coordinate points as an example, the slope of the segment is first calculated. The calculation method is to subtract the y-component of the first coordinate point from the y-component of the second coordinate point, and divide the difference by the difference between the x-component of the second coordinate point and the x-component of the first coordinate point. Specifically, 30.466 minus 30.456 equals 0.01, and 120.133 minus 120 equals 0.01. 123 equals 0.01. Dividing 0.01 by 0.01 gives a slope of 1. Then, based on the slope and the coordinates of the first point, determine the linear constant term of the line segment. The calculation method is: subtract the slope multiplied by the x-component of the first point (y-component), i.e., 30.456 minus 1 multiplied by 120.123, resulting in a constant term of -89.667. To calculate the perpendicular distance from the vehicle's current coordinates to the line, first calculate the numerator: multiply the current x-component by the slope, subtract the current y-component, and add the constant term. The absolute value of the result is calculated as follows: 120.128 multiplied by 1, minus 30.458, plus (-89.667). This gives 120.128 minus 30.458 equals 89.67, and 89.67 minus 89.667 equals 0.003, so the absolute value is 0.003. Next, the denominator is calculated by multiplying the slope by itself, adding 1, and then taking the square root of the result: 1 multiplied by 1 plus 1 equals 2, resulting in a square root of 1.4142. Finally, the numerator is divided by the denominator to obtain the line segment. The corresponding vertical distance is 0.003 divided by 1.4142, which is approximately 0.00212 kilometers, or 2.12 meters. The vertical distances from the vehicle's current coordinates to all other segments of the frequently used route are calculated in the same way. For example, the vertical distance to the segment formed by the second and third coordinate points is 3.56 meters. The smallest value of 2.12 meters is selected from all vertical distances as the shortest spatial distance between the vehicle's current location and the frequently used route. The result is rounded to two decimal places and is in meters.

[0069] When recording the duration of deviation, timing begins from the time point when the deviation from the usual route is first determined through step 402e, 2025-10-09 14:30:00, with the timing device accurate to the second. The timing is automatically updated every second, recording the duration of the deviation in real time. At 14:30:01, the duration is updated to 1 second; at 14:30:02, it is updated to 2 seconds, and so on. If, at 14:35:20, the result is re-determined through step 402e and changes to no deviation from the usual route, timing stops immediately, and the total duration at this time is recorded as 320 seconds, which is the duration of this deviation. If the deviation continues without any change in the determination result, timing continues until the current time point, and the duration is updated in real time. For example, as of 14:40:00, the duration is recorded as 600 seconds. The shortest spatial distance and the duration of deviation together constitute the abnormal event data.

[0070] This embodiment, through dynamic resampling supported by specific numerical values, ensures data integrity while taking into account processing efficiency. Detailed feature standardization and convolution operations can accurately extract the core features of the data. The long short-term memory network layer deeply learns the time dependency relationship, and combined with cosine similarity comparison, it can accurately determine the route deviation, accurately calculate the deviation distance and duration, and effectively distinguish between malicious deviation and normal detour.

[0071] In a preferred embodiment of the present invention, step 5, matching the abnormal event data with a preset set of early warning rules to generate early warning information, may include:

[0072] Step 501: Based on the abnormal event data, extract the deviation distance and deviation duration values ​​to obtain matching feature pairs. Specifically, this includes: obtaining the abnormal event data generated in step 403, which contains core information about deviations in high-value bulk cargo transportation routes, namely the shortest spatial distance and the duration of the deviation. Accurately extract the deviation distance value from the abnormal event data, keeping it to two decimal places in meters, for example, 2.12 meters; extract the deviation duration value, accurate to the second, for example, 320 seconds; combine the extracted deviation distance and deviation duration values ​​in the order of distance and time to form matching feature pairs, for example, (2.12 meters, 320 seconds).

[0073] Step 502: Based on the matching feature pairs and the preset early warning rule set, compare the deviation distance value and deviation duration value with the distance threshold and time threshold in each early warning rule to obtain the comparison result for each early warning rule. Specifically, the preset early warning rule set is based on the safety risk level classification of high-value bulk cargo transportation and includes three core early warning rules, adapted to the transportation monitoring needs of high-value goods such as nickel iron and alumina; Rule 1 is a high-risk early warning rule with a distance threshold set at 500 meters and a time threshold set at 300 seconds; Rule 2 is a medium-risk early warning rule with a distance threshold set at 300 meters and a time threshold set at 200 seconds; Rule 3 is a low-risk early warning rule with a distance threshold set at 100 meters and a time threshold set at 10 seconds. 0 seconds; compare the deviation distance value in each matching feature pair with the distance threshold of each rule to determine if it is greater than or equal to the distance threshold; at the same time, compare the deviation duration value with the time threshold of the corresponding rule to determine if it is greater than or equal to the time threshold; for example, if the matching feature pair is (650 meters, 350 seconds), when compared with rule one, 650 meters is greater than 500 meters and 350 seconds is greater than 300 seconds, and both judgment results are yes; when compared with rule two, 650 meters is greater than 300 meters and 350 seconds is greater than 200 seconds, and both judgment results are yes; when compared with rule three, 650 meters is greater than 100 meters and 350 seconds is greater than 100 seconds, and both judgment results are yes. The comparison result of each rule includes both distance comparison result and time comparison result.

[0074] Step 503: Based on the comparison results of each warning rule, determine whether there exists a warning rule whose comparison results simultaneously meet the distance and time threshold conditions, and obtain the rule matching result. Specifically, this includes: combining the distance comparison result and time comparison result of each warning rule for judgment. Each rule has a clearly defined fixed threshold. The distance threshold for rule 1 is 500 meters and the time threshold is 300 seconds. The distance threshold for rule 2 is 300 meters and the time threshold is 200 seconds. The distance threshold for rule 3 is 100 meters and the time threshold is 100 seconds. If the distance comparison result of a rule is yes (i.e., the deviation distance value is greater than or equal to the distance threshold corresponding to the rule) and the time comparison result is yes (i.e., the deviation duration value is greater than or equal to the time threshold corresponding to the rule), then the rule is determined to meet the conditions and is a successfully matched warning rule. If either comparison result is no (the distance does not reach the corresponding threshold or the time does not reach the corresponding threshold), then the rule is determined to not meet the conditions and is a failed match warning rule. After sequentially traversing Rule 1, Rule 2, and Rule 3, the number of successfully matched rules is counted. If the count is greater than or equal to 1, it indicates that there is a warning rule that meets the conditions, and the rule matching result is considered successful. The names of all successfully matched rules and their corresponding thresholds are recorded. If the count is 0, it means that none of the three rules simultaneously meet their respective distance and time threshold conditions, and the rule matching result is considered unsuccessful. For example, when matching the feature pair (650 meters, 350 seconds), when compared with Rule 1, 650 meters is greater than 500 meters and 350 seconds is greater than 300 seconds, which meets the conditions. When compared with Rule 2, 650 meters is greater than 300 meters and 350 seconds is greater than 200 seconds, which also meets the conditions. When compared with Rule 3, 650 meters is greater than 100 meters and 350 seconds is greater than 100 seconds, which also meets the conditions. All three rules are successfully matched, and the rule matching result is considered successful. The successfully matched rules are Rule 1, Rule 2, and Rule 3.

[0075] Step 504: Based on the rule matching results, obtain the preset level identifier and handling suggestion text corresponding to the successfully matched warning rule, and obtain the warning level and handling suggestion content. Specifically, each warning rule corresponds to a unique level identifier and handling suggestion text. The level identifier corresponds to the risk level. Rule 1 corresponds to the high-risk level identifier, and the handling suggestion text is to immediately initiate intelligent outbound calls to contact the driver to verify the reason for the deviation, simultaneously notify the consignor and carrier's responsible persons, and arrange for a dedicated person to track the vehicle trajectory in real time. Rule 2 corresponds to the medium-risk level identifier, and the handling suggestion text is to initiate intelligent outbound calls to contact the driver, inquire about the deviation and record the feedback, and continuously monitor whether the vehicle's subsequent trajectory returns to the usual route. Rule 3 corresponds to a low-risk level, and the suggested action text is to send a reminder message to the driver via the APP, informing them of the current deviation and requesting them to return to the usual route in a timely manner. If the rule matching result is successful, the rule with the highest level among the successfully matched rules is selected as the core matching rule, its corresponding level identifier is extracted as the warning level, and its corresponding suggested action text is extracted as the suggested action content. For example, if the successfully matched rules are Rule 1, Rule 2, and Rule 3, the core matching rule is Rule 1, the warning level is high-risk, and the suggested action content is to immediately initiate intelligent outbound calls to contact the driver to verify the reason for the deviation, simultaneously notify the consignor and carrier's responsible persons, and arrange for a dedicated person to track the vehicle's trajectory in real time.

[0076] Step 505: Based on the warning level, the content of the handling suggestions, and the abnormal event data, generate warning information. Specifically, the warning information adopts a structured text format and contains four core parts. The first part is the basic information of the event, integrating the deviation distance and duration values ​​from the abnormal event data, and clearly identifying the name of the high-value bulk cargo, the transport order number, and the vehicle license plate number. The second part is the warning level indicator, directly presenting the warning level obtained in step 504. The third part is the handling suggestion content, fully referencing the handling suggestion text corresponding to the matching rule. The fourth part is the timestamp, recording the specific time the warning information was generated, accurate to the second. For example, the warning information is: transport order number PCDD20251007000576, license plate number Su A12345, transported cargo nickel iron, currently deviating from the usual route by 650 meters, duration 350 seconds, warning level high risk, handling suggestion immediately initiate intelligent outbound call to contact the driver to verify the reason for the deviation, simultaneously notify the consignor and carrier's responsible persons, arrange for a dedicated person to track the vehicle trajectory in real time, generation time 2025-10-09 14:35:20, ensuring that the warning information is clear and comprehensive, and can be directly used for subsequent handling operations.

[0077] This embodiment, through precise matching of a clear set of early warning rules with abnormal event data, clearly classifies risk levels and provides targeted handling suggestions, ensuring that deviations from high-value bulk cargo transportation routes can be quickly identified and located.

[0078] In a preferred embodiment of the present invention, step 6, obtaining a collaborative handling instruction based on the early warning information, order data, cargo data, and travel data, and executing a notification operation based on the collaborative handling instruction to obtain the operation execution result, may include:

[0079] Step 601: Based on the vehicle's current location coordinates in the trip data and the delivery address coordinates in the order data, determine the first and second circles with preset radii. Specifically, this includes: extracting the vehicle's current location planar coordinates from the trip data (latitude and longitude transformed by Gauss-Kruger projection, e.g., lateral coordinate 120.128, vertical coordinate 30.458); and extracting the delivery address planar coordinates from the order data (also transformed lateral and vertical coordinates, e.g., lateral coordinate 120.356, vertical coordinate 30.678). Considering the safety monitoring range requirements for high-value bulk cargo transportation, a preset radius of 5 kilometers is used. This radius is determined based on historical transportation emergency response experience and can effectively cover key areas around the vehicle and preparation areas around the delivery address. A circle is drawn with the vehicle's current location coordinates as the center and a radius of 5 kilometers to obtain the first circle, used to identify the core monitoring area where the vehicle is currently located. A circle is drawn with the delivery address coordinates as the center and a radius of 5 kilometers to obtain the second circle, used to identify the response area around the delivery address.

[0080] Step 602: Calculate the center-to-center distance based on the center coordinates of the first and second circles; calculate the intersection and union areas of the first and second circles based on the center-to-center distance and a preset radius. Specifically, when calculating the center-to-center distance, the difference in the horizontal and vertical coordinates of the two circles is used as the basis. The horizontal coordinate difference is 120.356 minus 120.128 equals 0.228, and the vertical coordinate difference is 30.678 minus 30.458 equals 0.22. Multiplying each difference by itself yields the square of the horizontal difference: 0.228 x 0.228 = 0.051984; the square of the vertical difference: 0.22 x 0.22 = 0.0484. Adding these two squares gives 0.051984 + 0.0484 = 0.100384. Taking the square root of this sum gives the distance between the centers of the circles as approximately 0.317 kilometers. When calculating the intersection and union areas, the area of ​​a single circle is first defined as the product of its preset radius and pi. That is, 5 times 5 times 3.1416 equals 78.54 square kilometers; To determine the relationship between the center distance and the sum and difference of the radii of the two circles, both circles have a radius of 5 kilometers, a sum of radii of 10 kilometers, and a difference of 0. The current center distance of 0.317 kilometers is less than the sum of the radii and greater than the difference of the radii, indicating that the two circles overlap; the intersection area is calculated by first calculating the ratio of the center distance to the radius, i.e., 0.317 divided by 5 equals 0.0634, and then using this ratio to query the preset... The table of arc coefficients shows that the corresponding arc coefficient is 0.0018. The arc area of ​​each circle is the area of ​​the circle multiplied by the arc coefficient, i.e., 78.54 multiplied by 0.0018 is approximately equal to 0.1414 square kilometers. The intersection area of ​​two arcs is approximately 0.2828 square kilometers. The union area is calculated by subtracting the intersection area from the sum of the areas of the two circles, i.e., 78.54 plus 78.54 equals 157.08, minus 0.2828 equals 156.7972 square kilometers.

[0081] Step 603: Based on the ratio of the intersection area to the union area, combined with the warning level and the cargo status tags in the cargo data, a comprehensive evaluation result is obtained. Specifically, this includes: first, calculating the ratio of the intersection area to the union area, i.e., 0.2828 divided by 156.7972 approximately equals 0.0018. This ratio reflects the proximity of the vehicle's current location to the delivery address; a smaller ratio indicates the vehicle is farther from the delivery address, and a larger ratio indicates it is closer; obtaining the warning level from the warning information, such as high risk; extracting the cargo status tags from the cargo data, which are the normal or abnormal tags generated in step 102, such as normal; the comprehensive evaluation result is divided into three levels: Level 1 is emergency handling, Level 2 is routine handling, and Level 3... The classification is as follows: Level 1 Emergency Response; Judgment Logic: If the ratio is less than 0.01 and the warning level is high risk, regardless of the cargo status label, it is judged as Level 1 Emergency Response; If the ratio is between 0.01 and 0.05, and the warning level is medium or high risk, and the cargo status label is abnormal, it is judged as Level 1 Emergency Response; If the ratio is between 0.01 and 0.05, the warning level is medium risk, and the cargo status label is normal, it is judged as Level 2 Routine Response; If the ratio is greater than 0.05, regardless of the warning level and the cargo status label, it is judged as Level 3 Simple Response; For example, if the current ratio is 0.0018, the warning level is high risk, and the cargo status label is normal, the comprehensive assessment result is judged as Level 1 Emergency Response.

[0082] Step 604: Based on the comprehensive assessment results, match instruction templates from the pre-set disposal instruction library, and fill in the warning level and delivery address coordinates into the instruction templates to generate collaborative disposal instructions. Specifically, this includes: the generation of the pre-set disposal instruction library is based on historical abnormal disposal cases of high-value bulk cargo transportation, safety control requirements, and the platform's actual operational capabilities; collecting abnormal disposal records of high-value goods such as nickel-iron and alumina on this route over the past three years, covering various scenarios such as route deviation, late delivery, and entry into risk areas; statistically analyzing effective disposal measures, response time, and participating roles under different scenarios; and combining the correlation analysis of transportation distance, cargo value, and risk level, sorting out three categories of instructions according to the three-level classification logic of the comprehensive assessment results. The template framework integrates existing platform functional modules such as intelligent outbound calling, APP push notifications, SMS notifications, and trajectory tracking. It clarifies the handling steps and execution entities for each template, ultimately forming a pre-set handling instruction library tailored to actual operations. The Level 1 emergency handling template is designed around the warning level XX, the vehicle's current location being far from the delivery address, and the transported goods being XX. The specific content of the emergency handling process is as follows: intelligent outbound calling the driver to verify the reason for the deviation; SMS notification to the consignor's representative XXX and the carrier's representative XXX; arranging for a monitoring specialist XXX to track the vehicle's trajectory in real time, providing status feedback every ten minutes, specifying the delivery address coordinates (120.356, 30.678), and requesting relevant parties to prepare for emergencies.

[0083] The Level 2 standard handling template is designed around the warning level XX, the vehicle's current location being at a suitable distance from the delivery address, and the transported goods being XX. The specific steps for initiating the standard handling process are as follows: The system intelligently calls the driver to understand the deviation; the system continuously monitors the vehicle's trajectory; if the vehicle does not return to the usual route within one hour, the system contacts the driver again to confirm the delivery address coordinates (XX, XX) and simultaneously records this information with the recipient. The Level 3 simplified handling template is designed around the warning level XX, the vehicle's current location being close to the delivery address, and the transported goods being XX. The specific steps for initiating the simplified handling process are as follows: The app pushes a reminder message to the driver, requiring them to return to the route promptly, confirming the delivery address coordinates (XX, XX), and informing the driver to unload at the scheduled time. Based on the comprehensive assessment results, Level 1 emergency handling is initiated by matching the corresponding Level 1 template, filling in the warning level (high risk), delivery address coordinates (120.356, 30.678), and the goods name (nickel iron) into the corresponding positions in the template, and generating a collaborative handling instruction.

[0084] Step 605: According to the collaborative processing instruction, the intelligent outbound calling unit performs telephone dialing and voice notification, and receives voice feedback during the call to obtain a notification feedback signal; based on the notification feedback signal, the confirmation status or failure reason is analyzed to obtain the operation execution result. Specifically, the intelligent outbound calling unit is deployed on a dedicated logistics transportation monitoring server and integrates functions related to speech synthesis, telephone dialing, and speech recognition. The telephone dialing function extracts the driver's reserved mobile phone number (e.g., 138XXXX1234) from the collaborative handling instruction and automatically initiates dialing; the voice synthesis function converts the core notification content in the instruction into voice, such as "Hello, the nickel-iron cargo being transported by your vehicle Su A12345 has deviated from the usual route by 650 meters for 350 seconds, triggering a high-risk warning. Please verify the reason for the deviation and return to the usual route as soon as possible. If there are any special circumstances, please report them promptly." During the call, the intelligent outbound call unit's voice recognition function collects the driver's voice feedback in real time, converts it into text signals, and forms a notification feedback signal; if the call is connected and the driver clearly responds that he will immediately adjust the route or explain the reasonable reason for the deviation (e.g., emergency avoidance, road closure), the parsing confirmation status is "confirmed"; if the call is connected but the driver does not respond clearly (e.g., vague answers like "I know" or "I'll talk later"), the confirmation status is "unclear"; if the call is not connected, it indicates a busy line or no one answers, and the parsing failure reason is "busy line" or "no one answers"; if the signal is interrupted during the call, the failure reason is "call interrupted". The operation execution result includes dialing status, confirmation status or failure reason, and call duration. For example, the dialing status is connected, the confirmation status is confirmed, and the call duration is 60 seconds. The execution status of the notification operation is fully recorded.

[0085] This embodiment combines vehicle location, delivery address, warning level, and cargo status for comprehensive evaluation, resulting in more targeted and operable collaborative handling instructions. The efficient execution of the intelligent outbound call unit ensures timely delivery of notifications, improving the handling efficiency and response speed of abnormal events in high-value bulk cargo transportation.

[0086] In a preferred embodiment of the present invention, step 7, obtaining visual update data based on the operation execution result and warning information, and feeding the visual update data back to the visual management page to complete closed-loop management, may include:

[0087] Step 701: Based on the operation execution result, extract the notification confirmation status or execution failure reason to obtain the handling status identifier; based on the handling status identifier and the warning level and handling suggestion content in the warning information, combined with the deviation distance value and deviation duration value, generate formatted text data containing an event summary and handling record, specifically including: the handling status identifier is divided into three categories: when the confirmation status is confirmed, the handling status identifier is handling in progress; when the confirmation status is unclear or the failure reason is busy line or no one answers, the handling status identifier is waiting to be retried; when the failure reason is call interruption and multiple dialing attempts (maximum 3 times) still fail, the handling status identifier is handling abnormal; for example, if the operation execution result is confirmed, the handling status identifier is handling in progress. The event summary includes the transport order number, license plate number, cargo name, warning level, deviation distance, and deviation duration. For example, the transport order number is PCDD20251007000576, the license plate number is Su A12345, the cargo name is nickel iron, the warning level is high risk, the deviation from the usual route is 650 meters, and the duration is 350 seconds. The handling record includes the handling instruction content, the notification execution result, and the handling status indicator. For example, the handling instruction is to initiate the emergency handling process, intelligently call the driver to verify the reason for the deviation, and simultaneously notify the relevant person in charge; the notification execution result is that the call is connected (mobile phone number 138XXXX1234), the driver confirms that the deviation was temporary due to road closure, and will detour back to the route as soon as possible; the handling status indicator is "handling in progress". The event summary and handling record are combined in a fixed format to form formatted text data.

[0088] Step 702: Based on the formatted text data, encapsulate the data according to the preset data interface specification to obtain a structured status data packet; send the structured status data packet to the backend service corresponding to the visual management page; based on the reception confirmation signal returned by the backend service, determine that the closed loop of the early warning event handling process is completed, specifically including: the preset data interface specification requires the structured status data packet to include field identifiers, data types, data lengths, and field values. Fields include order number, license plate number, goods name, early warning level, deviation distance, deviation duration, handling status identifier, event summary, handling record, and generation time; for example, the field identifier order number corresponds to a data type string with a data length of 20 and a field value PCDD20251007000576; the field identifier handling status identifier corresponds to a data type string with a data length of 10 and a field value of "handling in progress"; the field identifier generation time corresponds to a data type timestamp. The data length is 20, and the field value is 2025-10-09 14:36:00. The formatted text data is encapsulated according to the specification to generate a structured status data packet in JSON format. The data packet is sent to the backend service of the visual management page via HTTP protocol. After receiving the data packet, the backend service verifies the data integrity. The verification fields include four mandatory fields: order number, license plate number, warning level, and handling status identifier. If all mandatory fields exist and are in the correct format, a receipt confirmation signal is returned. If the data is missing or has an incorrect format, a verification failure signal is returned, and the data needs to be re-encapsulated and sent. When a receipt confirmation signal is received from the backend service, it indicates that the visual management page has obtained the updated data and is displaying it synchronously, indicating that the warning event handling process is complete. If a verification failure signal is still received after multiple (maximum 3) attempts, the data encapsulation problem needs to be investigated and the encapsulation and sending operation needs to be re-executed.

[0089] In this embodiment, formatted text data and structured status data packets ensure the accuracy and standardization of information transmission, while real-time updates of the visual management page allow relevant personnel to intuitively grasp the progress of events, realizing full-process management of abnormal events in high-value bulk cargo transportation from identification and handling to tracking, and improving the integrity and controllability of transportation safety management.

[0090] like Figure 2 As shown, embodiments of the present invention also provide a real-time early warning and collaborative evaluation system for logistics transportation based on neural networks, comprising:

[0091] The data acquisition module is used to collect multi-source data during the logistics and transportation process. The multi-source data includes trip data, cargo data, order data, and interaction data, and integrates them into a time-series data stream.

[0092] The partitioning module is used to obtain a time-series data point set based on the time-series data stream; construct a multi-dimensional mapping surface based on the time-series data point set; delineate the initial analysis region based on the multi-dimensional mapping surface; and partition the initial analysis region based on the time-series data point set to obtain multiple processing partitions.

[0093] The calculation module is used to map the time series data point set to the corresponding processing partition; calculate the Hu invariant moments of the time series data point set in each processing partition; normalize the central moments by calculating the central moments of the time series data point set; and obtain the Hu invariant moment set based on the normalized central moments to generate a dynamic adjustment strategy.

[0094] The optimization module is used to optimize the time-series data stream according to the dynamic adjustment strategy, input the optimized time-series data stream into the pre-trained neural network model for analysis, identify vehicle deviation from the common route events, and obtain abnormal event data.

[0095] The matching module is used to match abnormal event data with a set of preset early warning rules to generate early warning information;

[0096] The execution module is used to obtain collaborative handling instructions based on early warning information, order data, cargo data, and travel data, execute notification operations based on the collaborative handling instructions, and obtain the operation execution results;

[0097] The output module is used to obtain visual update data based on the operation execution results and early warning information, and to feed the visual update data back to the visual management page to complete closed-loop management.

[0098] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0099] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A real-time early warning and collaborative evaluation method for logistics transportation based on neural networks, characterized in that, The method includes: Step 1: Collect multi-source data during the logistics and transportation process. The multi-source data includes trip data, cargo data, order data, and interaction data, and integrate them into a time-series data stream. Step 2: Based on the time-series data stream, perform sliding sampling according to a preset time window to obtain a time-series data point set composed of multi-dimensional data vectors within each time window; based on the time-series data point set, select key feature dimensions for projection to obtain a projection point set; based on the spatial distribution of the projection point set, select reference points as fitting references; based on the spatial position of the reference points, calculate the connection relationship between the reference points to generate a fitting grid; based on the distribution of the fitting grid and the projection point set, calculate the surface fitting parameters; construct a multi-dimensional mapping surface based on the surface fitting parameters; based on the boundary coordinates of the data points on the multi-dimensional mapping surface, calculate the minimum bounding region containing all data points to obtain the initial analysis region; based on the density distribution of the time-series data point set within the initial analysis region, divide the initial analysis region to obtain multiple processing partitions; Step 3: Based on the time-series data point set and the spatial range of each processing partition, assign each data point to its corresponding processing partition, obtaining a subset of data points for each processing partition; for each subset of data points in each processing partition, calculate the mean coordinates of the subset to obtain the coordinates of the center point; calculate the zero-order central moment, first-order central moment, and second-order central moment of the subset of data points in the two-dimensional spatial distribution based on the center point coordinates; normalize the first-order and second-order central moments based on the zero-order central moment to obtain normalized first-order and normalized second-order central moments; calculate the seven invariant moment values ​​according to the combination of the seven invariant formulas of Hu invariant moments, generating a set of Hu invariant moments; based on the set of Hu invariant moments and the trend of data point density change over time in each partition, generate a dynamic adjustment strategy; Step 4: Optimize the time-series data stream according to the dynamic adjustment strategy, input the optimized time-series data stream into the pre-trained neural network model for analysis, identify vehicle deviation from the common route events, and obtain abnormal event data; Step 5: Match the abnormal event data with the preset set of early warning rules to generate early warning information; Step 6: Based on the warning information, order data, cargo data, and trip data, obtain collaborative handling instructions. Execute notification operations based on these instructions to obtain the operation execution results, including: determining a first circle and a second circle with preset radii based on the vehicle's current location coordinates in the trip data and the delivery address coordinates in the order data; calculating the center-to-center distance based on the center coordinates of the first and second circles; calculating the intersection and union areas of the first and second circles based on the center-to-center distance and the preset radius; obtaining a comprehensive evaluation result based on the ratio of the intersection and union areas, combined with the warning level and cargo status tags in the cargo data; matching instruction templates from the preset handling instruction library based on the comprehensive evaluation result, and filling the instruction templates with the warning level and delivery address coordinates to generate collaborative handling instructions; executing telephone dialing and voice notifications through the intelligent outbound call unit based on the collaborative handling instructions, and receiving voice feedback during the call to obtain notification feedback signals; and analyzing the confirmation status or failure reason based on the notification feedback signals to obtain the operation execution results. Step 7: Based on the operation execution results and warning information, obtain the visual update data and feed it back to the visual management page to complete the closed-loop management.

2. The real-time early warning and collaborative evaluation method for logistics transportation based on neural networks according to claim 1, characterized in that, Collect multi-source data during the logistics and transportation process, including trip data, cargo data, order data, and interaction data, and integrate them into a time-series data stream, including: The system acquires raw trip data, raw cargo data, raw order data, and raw interaction data. Raw trip data includes location coordinates, driving speed, and timestamps obtained from the vehicle positioning unit. Raw cargo data includes cargo weight and interior / exterior image data obtained from the cargo monitoring unit. Raw order data includes order number, delivery address, and delivery deadline obtained from the order management unit. Raw interaction data includes call records, driver confirmation status, and anomaly feedback content obtained from the intelligent outbound call unit. Based on the original trip data, structured trip data is obtained; based on the original cargo data, structured cargo data is obtained; based on the original order data, structured order data is obtained; based on the original interaction data, structured interaction data is obtained. The structured trip data includes formatted time, latitude and longitude, and instantaneous speed; the structured cargo data includes cargo status tags and numerical physical quantities; the structured order data includes order timestamps and address vectors; and the structured interaction data includes interaction time and classification codes. Based on structured itinerary data, structured cargo data, structured order data, and structured interaction data, time information is extracted and aligned to obtain a time-aligned multi-source data sequence. Interpolation and merging are performed on time-aligned multi-source data sequences to obtain a time-series data stream.

3. The real-time early warning and collaborative evaluation method for logistics transportation based on neural networks according to claim 2, characterized in that, The time-series data stream is optimized based on a dynamic adjustment strategy. The optimized time-series data stream is then input into a pre-trained neural network model for analysis to identify vehicle deviation events from the usual route, thus obtaining abnormal event data, including: Based on the sampling frequency adjustment instruction in the dynamic adjustment strategy, the time-series data stream is resampled in real time to obtain the sampling-optimized time-series data stream; The sampled and optimized time-series data stream is input into a pre-trained neural network model to extract the spatiotemporal features of the time-series data stream. The model is then matched and compared with a preset common route pattern to obtain the deviation judgment result between the vehicle driving status and the preset route. Based on the deviation determination results, the shortest spatial distance between the vehicle's current position and the usual route is calculated, and the duration of the deviation is recorded to obtain abnormal event data including deviation distance and deviation time.

4. The real-time early warning and collaborative evaluation method for logistics transportation based on neural networks according to claim 3, characterized in that, The optimized time-series data stream is input into a pre-trained neural network model to extract its spatiotemporal features. These features are then compared with preset common route patterns to obtain the deviation judgment result between the vehicle's driving state and the preset route, including: Based on the sampled and optimized time series data stream, the mean and standard deviation of each dimension are calculated. The time series data is then normalized based on the mean and standard deviation to obtain standardized time series data. Based on standardized time-series data, convolution operations are performed through the convolutional layers of a neural network model to extract spatial and temporal correlation features within each local time window, thus obtaining a primary spatiotemporal feature sequence. Based on the primary spatiotemporal feature sequence, the hidden state is updated step by step through the recurrent neural network layer of the neural network model to learn the temporal dependencies in the sequence and obtain a high-dimensional spatiotemporal feature vector. Based on the high-dimensional spatiotemporal feature vector and the preset common route pattern feature vector, the cosine similarity between the two feature vectors is calculated to obtain the feature similarity value; The deviation determination result is obtained by comparing the feature similarity value with the preset deviation threshold.

5. The real-time early warning and collaborative evaluation method for logistics transportation based on neural networks according to claim 4, characterized in that, The abnormal event data is matched with a set of preset early warning rules to generate early warning information, including: Based on the abnormal event data, the deviation distance and deviation duration values ​​are extracted to obtain matching feature pairs; Based on the matching feature pairs and the preset set of early warning rules, the deviation distance value and deviation duration value are compared with the distance threshold and time threshold in each early warning rule to obtain the comparison result of each early warning rule; Based on the comparison results of each warning rule, determine whether there are any warning rules whose comparison results simultaneously meet the distance and time threshold conditions, and obtain the rule matching results; Based on the rule matching results, obtain the preset level identifier and handling suggestion text corresponding to the successfully matched warning rule, and obtain the warning level and handling suggestion content; Warning information is generated by combining the warning level, the content of the handling suggestions, and the abnormal event data.

6. The real-time early warning and collaborative evaluation method for logistics transportation based on neural networks according to claim 5, characterized in that, Based on the operation execution results and early warning information, visualized update data is obtained and fed back to the visualization management page to complete closed-loop management, including: Based on the operation execution result, extract the notification confirmation status or execution failure reason to obtain the handling status identifier; based on the handling status identifier and the warning level and handling suggestion content in the warning information, combined with the deviation distance value and deviation duration value, generate formatted text data containing event summary and handling record; Based on the formatted text data, the data is encapsulated according to the preset data interface specifications to obtain a structured status data packet; the structured status data packet is sent to the backend service corresponding to the visual management page; based on the reception confirmation signal returned by the backend service, the closed loop of the early warning event handling process is determined to be completed.

7. A real-time early warning and collaborative evaluation system for logistics transportation based on neural networks, wherein the system implements the method as described in any one of claims 1 to 6, characterized in that, include: The data acquisition module is used to collect multi-source data during the logistics and transportation process. The multi-source data includes trip data, cargo data, order data, and interaction data, and integrates them into a time-series data stream. The partitioning module is used to obtain the set of time-series data points based on the time-series data stream; Constructing a multidimensional mapping surface based on time-series data point sets; Delineate the initial analysis region based on the multidimensional mapping surface; The initial analysis region is divided based on the time series data point set, resulting in multiple processing partitions; The calculation module is used to map the time series data point set to the corresponding processing partition; calculate the Hu invariant moments of the time series data point set in each processing partition; normalize the central moments by calculating the central moments of the time series data point set; and obtain the Hu invariant moment set based on the normalized central moments to generate a dynamic adjustment strategy. The optimization module is used to optimize the time-series data stream according to the dynamic adjustment strategy, input the optimized time-series data stream into the pre-trained neural network model for analysis, identify vehicle deviation from the common route events, and obtain abnormal event data. The matching module is used to match abnormal event data with a set of preset early warning rules to generate early warning information; The execution module is used to obtain collaborative handling instructions based on early warning information, order data, cargo data, and travel data, execute notification operations based on the collaborative handling instructions, and obtain the operation execution results; The output module is used to obtain visual update data based on the operation execution results and early warning information, and to feed the visual update data back to the visual management page to complete closed-loop management.

Images