Railway logistics whole-process management and control system and platform based on internet of things

Abstract

The application discloses a railway logistics whole-process management and control system and platform based on Internet of Things, and belongs to the technical field of railway logistics intelligent monitoring. The system comprises a data acquisition device, a score calculation device, a model construction device, a forward calculation device, a deviation acquisition device, a projection judgment device, a reverse tracing device and an instruction feedback device. The application establishes a three-level correlation mapping model of operation parameters, operation quality score and cargo loss risk score, obtains the deviation between the theoretical value and the actual value through forward calculation, projects the deviation to a multi-dimensional evaluation space and compares it with a dynamic qualified boundary area, adopts a deep learning back propagation mechanism for reverse calculation when the boundary is exceeded, decouples the contribution degree of each parameter dimension layer by layer to locate the risk source, and generates and feeds back operation adjustment instructions. The application realizes accurate tracing and closed-loop management and control of reloading operation risks, and effectively reduces the cargo loss rate.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for railway logistics, and in particular to a railway logistics end-to-end control system and platform based on the Internet of Things. Background Technology

[0002] In railway logistics transportation, container transshipment between railways, ports, and highways is a high-risk link for cargo damage. Existing transshipment monitoring systems typically use threshold alarms, triggering an alarm when sensor data exceeds a preset threshold. Such systems have been deployed at some transshipment nodes and can achieve real-time monitoring of single indicators such as vibration, tilt, and impact.

[0003] However, the existing system suffers from fundamental technical flaws: it can only make independent threshold judgments based on single sensor indicators, failing to establish a deep correlation model between operational parameters and cargo damage risk. This results in the inability to pinpoint the specific operational step or parameter dimension that caused the cargo damage risk to exceed the limit. Specifically, the system can detect anomalies but cannot identify their source. After receiving an alarm, operators still need to manually check multiple steps such as lifting, lowering, and stacking of the lifting equipment, which is inefficient, relies heavily on experience, and makes it difficult to form closed-loop control. In addition, the existing system is not connected to an industrial internet platform, lacking cross-node control capabilities through cloud collaboration. Data from each transshipment operation node is isolated, making it impossible to achieve integrated intelligent scheduling and risk prevention throughout the entire railway logistics process.

[0004] Therefore, this invention proposes a railway logistics end-to-end control system and platform based on the Internet of Things. Summary of the Invention

[0005] This invention provides a railway logistics end-to-end control system and platform based on the Internet of Things. To address the aforementioned shortcomings, it establishes a three-level correlation mapping model and combines forward calculation and reverse tracing mechanisms to achieve accurate location of risk sources and closed-loop adjustment of operating parameters, significantly reducing manual inspection costs and cargo damage rates.

[0006] This invention provides an Internet of Things-based railway logistics end-to-end control system, comprising: Data acquisition devices are used to collect the operating parameters and status parameters of containers during transshipment operations; The scoring calculation device is used to calculate the operation quality score based on the operation parameters and the cargo damage risk score based on the status parameters. The model building device is used to establish a three-level correlation mapping model among operating parameters, operating quality scores, and cargo damage risk scores. The forward extrapolation device is used to input the operation parameters into the three-level correlation mapping model for forward extrapolation to obtain the theoretical value of the operation quality score and the predicted value of the cargo damage risk score. The deviation acquisition device is used to compare the theoretical value of the operation quality score obtained by the forward calculation device with the actual value of the operation quality score calculated by the scoring calculation device to obtain a quality deviation sequence, compare the predicted value of the cargo damage risk score obtained by the forward calculation device with the actual value of the cargo damage risk score calculated by the scoring calculation device to obtain a risk deviation sequence, and concatenate the quality deviation sequence and the risk deviation sequence into a joint deviation vector. The projection determination device is used to project the joint deviation vector onto the multidimensional evaluation space, dynamically delineate the qualified operation boundary area in the multidimensional evaluation space based on historical operation data, and calculate the shortest distance from the projection point of the joint deviation vector to the qualified operation boundary area. The reverse tracing device is used to input the current operation parameters into the three-level association mapping model and use the deep learning backpropagation mechanism to perform reverse calculation when the shortest distance is greater than zero. In the reverse calculation process, a deviation propagation weight matrix is ​​introduced to decouple the contribution of each operation parameter dimension to the joint deviation vector layer by layer, so as to obtain the risk source identification result. The instruction feedback device is used to generate operation adjustment instructions based on the risk source identification results and to feed the operation adjustment instructions back to the replacement execution mechanism.

[0007] Furthermore, the projection judgment device dynamically delineates the qualified operation boundary area in the multidimensional evaluation space based on historical operation data, including: Transshipment operations whose actual cargo damage risk score is lower than the preset risk threshold are marked as qualified operations. Extract the actual values ​​of the operation quality score and cargo damage risk score of all qualified operations from historical operation data to form a set of qualified data points in the multi-dimensional evaluation space; A density-based spatial clustering algorithm is used to cluster qualified data points, identify the core set of qualified points that are density-connected, and remove isolated outliers. Perform convex hull calculation on the core set of qualified points to obtain the initial qualified boundary polygon; Extract the actual values ​​of the operation quality score and cargo damage risk score of all non-conforming operations from historical operation data to form a set of non-conforming data points in the multi-dimensional evaluation space; Calculate the shortest distance from each point in the set of unqualified data points to the initial qualified boundary polygon, and statistically fit all the shortest distances to obtain the distance distribution function; The boundary contraction coefficient is determined based on the quantiles of the distance distribution function. The boundary contraction coefficient is then used to shrink the initial qualified boundary polygon inward to obtain a dynamic qualified boundary region.

[0008] Furthermore, after obtaining the dynamic qualified boundary region, the projection determination device is also used for: Real-time acquisition of the joint deviation vector projection points of newly added costume change operations; When a newly added projection point is located within the dynamic qualified boundary area, the actual value of the operation quality score and the actual value of the cargo damage risk score corresponding to the newly added projection point will be added to the qualified data point set as incremental data. At preset time intervals, density clustering and convex hull calculation are performed again on the qualified data point set to obtain the updated dynamic qualified boundary region. When a newly added projection point is located outside the dynamic qualified boundary area but is marked as qualified by subsequent manual review, the newly added projection point will be forcibly added to the qualified data point set and trigger a local expansion correction of the boundary. Specifically, the dynamic qualified boundary area will be locally expanded outward with the newly added projection point as the center according to the preset expansion radius, so that the newly added projection point is included in the expanded dynamic qualified boundary area.

[0009] Furthermore, the reverse tracing device inputs the current operating parameters into a three-level correlation mapping model and uses a deep learning backpropagation mechanism for reverse calculation. During the reverse calculation process, a deviation propagation weight matrix is ​​introduced to decouple the contribution of each operating parameter dimension to the joint deviation vector layer by layer, obtaining the risk source identification results, including: The three-level correlation mapping model is expressed as a composite function of the first mapping function and the second mapping function. The first mapping function describes the mapping relationship from the operation parameters to the operation quality score, and the second mapping function describes the mapping relationship from the operation quality score to the cargo damage risk score. Perform a first-order Taylor expansion on the first mapping function at the current position to obtain the first Jacobian matrix. The elements of the first Jacobian matrix represent the local sensitivity of each operation parameter dimension to each operation quality score dimension. Perform a first-order Taylor expansion on the second mapping function at the current position to obtain the second Jacobian matrix. The elements of the second Jacobian matrix represent the local sensitivity of each operation quality score dimension to the cargo damage risk score dimension. Multiplying the first Jacobian matrix by the second Jacobian matrix yields the bias propagation weight matrix, whose elements represent the overall sensitivity of each operational parameter dimension to the cargo damage risk scoring dimension. Multiply the joint bias vector by the pseudo-inverse of the bias propagation weight matrix to obtain the contribution vector of each operational parameter dimension. The contribution vector is normalized, and one or more operational parameter dimensions with the highest absolute value of contribution after normalization are identified as sources of risk.

[0010] Furthermore, the reverse tracing device uses ridge regression regularization estimation instead of direct pseudo-inverse calculation, specifically including: Obtain the bias propagation weight matrix using the above method; Extract multiple historical operation parameter samples and corresponding historical joint deviation vector samples from historical operation data; The ridge regression algorithm is used to perform regularized estimation of the bias propagation weight matrix to obtain the regularized bias propagation weight matrix. Multiply the joint bias vector by the pseudo-inverse of the regularized bias propagation weight matrix to obtain the regularized contribution vector.

[0011] Furthermore, the data acquisition device includes: The six-axis accelerometer integrated in the spreader collects the three components of lifting acceleration and the three components of placement speed at a sampling frequency of no less than 100 Hz. Four pressure sensors are installed at the four bottom corners of the container to collect data on the stacked pressure distribution. The sampling times of the four pressure sensors are kept synchronized. A nine-axis inertial measurement unit is installed at the geometric center of the container to collect three-axis vibration data, three-axis angular velocity data, and container tilt angle data. Two photosensitive sensors are symmetrically installed on the upper and lower sides of the container door frame to collect data on changes in light intensity when the door is opened. The difference in light intensity between the two photosensitive sensors is used to eliminate ambient light interference.

[0012] Furthermore, the scoring calculation device calculates an operation quality score based on the operating parameters, including: The composite acceleration curve is extracted from the three components of the lifting acceleration. The peak value of the composite acceleration curve is detected during the lifting period. The ratio of the peak value to the standard lifting acceleration benchmark value is used as the original lifting impact score. The original lifting impact score is mapped by an S-shaped function to obtain the normalized lifting impact score. The vertical velocity curve is extracted from the three components of the landing velocity. The ratio of the standard deviation to the mean of the vertical velocity curve during the landing period is calculated as the velocity variation coefficient. The velocity variation coefficient is then mapped to the landing stability score through an exponential decay function. The pressure values ​​collected by the pressure sensors at the four bottom corners are used to form a four-dimensional pressure vector. After normalizing the four-dimensional pressure vector, the Shannon entropy is calculated to obtain the stacked pressure distribution entropy. The stacked pressure distribution entropy is then mapped to the stacked uniform score through an inverse proportional function. The lifting impact score, the smooth placement score, and the uniform stacking score are weighted and summed according to preset weights to obtain the operation quality score.

[0013] Furthermore, the scoring calculation device calculates a cargo damage risk score based on the status parameters, including: The vertical vibration component is extracted from the triaxial vibration data. The vertical vibration component is integrally squared during the transshipment operation period to obtain the vertical vibration dose value. At the same time, the two vibration components in the horizontal plane are integrally squared to obtain the horizontal vibration dose value. The vertical vibration dose value and the horizontal vibration dose value are weighted and fused according to the height of the cargo center of gravity to obtain the cumulative vibration dose value. From the box tilt angle data collected by the nine-axis inertial measurement unit, the cumulative duration of pitch angle exceeding the preset pitch threshold and the cumulative duration of roll angle exceeding the preset roll threshold are counted, and the maximum value of the two is taken as the tilt over-limit duration value. The light intensity difference between the two photosensitive sensors is calculated from the light intensity data collected by the two photosensitive sensors. The light intensity difference sequence is subjected to differential operation to obtain the light intensity change rate. The number of times the light intensity change rate exceeds the preset change threshold is counted to obtain the number of abnormal openings of the cabinet door. Peak angular velocity is extracted from the three-axis angular velocity data collected by the nine-axis inertial measurement unit. When the peak angular velocity exceeds the preset angular velocity threshold, an impact event marker is generated, and the number of impact events during the changing operation period is counted. After normalizing the cumulative vibration dose value, tilt over-limit duration value, number of abnormal door openings and number of impact events by maximum and minimum values, the data are then weighted and fused according to the risk weight vector corresponding to the cargo type to obtain the cargo damage risk score.

[0014] Furthermore, it also includes a liability tracing device for performing the following steps in the event of a cargo damage dispute: The historical operation records generated and stored by the projection determination device and the reverse tracing device are obtained and stored in the historical database. Obtain the operational quality score sequence and cargo damage risk score sequence for disputed transshipment operations, and concatenate the two sequences into a joint score time series matrix; Retrieve multiple historical job records from the historical database that have a multidimensional Frescher distance less than a preset distance threshold with the joint scoring time series matrix. Each historical job record contains the historical joint scoring time series matrix and the historical responsibility determination result. The responsibility reference weight for each historical operation record is calculated based on the reciprocal of the multidimensional Frescher distance between the historical joint scoring time series matrix of each historical operation record and the joint scoring time series matrix of the disputed change operation. The historical responsibility determination results of all historical operation records are weighted and summed according to the corresponding responsibility reference weights to obtain the confidence vector of dispute responsibility attribution; The carrier with the highest confidence level is output as the responsible party, and a responsibility determination matrix is ​​also output. The responsibility determination matrix includes the three operational parameter dimensions with the highest contribution and the three state parameter dimensions with the highest contribution.

[0015] This invention provides an Internet of Things-based railway logistics end-to-end management and control platform, comprising: Multiple IoT-based railway logistics end-to-end management and control systems, such as any one of the above, are deployed at a transshipment operation node; A cloud-based collaborative server communicates and connects with all changing operation nodes. The cloud-based collaborative server includes: A cross-node data aggregator is used to collect operation quality scores, cargo damage risk scores, risk source identification results, and operation adjustment instructions from each transshipment operation node, and aggregate and store them according to time and space dimensions. The cross-node anomaly pattern miner is used to perform time-series clustering analysis on aggregated data, identify common anomaly patterns that recur in multiple transshipment operation nodes, and mark the operation parameter dimensions corresponding to the common anomaly patterns as global risk factors. A global model updater is used to update the three-level association mapping model of each transshipment operation node by using the global risk factor as an additional regularization constraint through transfer learning. The cross-node responsibility tracer is used to construct a cross-node responsibility chain by connecting the joint scoring time series matrix of each node in chronological order when a cargo damage event involves multiple transshipment operation nodes. It then performs segmented Fraser distance retrieval on the cross-node responsibility chain to obtain the cross-node responsibility attribution determination result. The early warning broadcaster is used to broadcast early warning information to all transshipment operation nodes when the cross-node anomaly pattern miner identifies a new common anomaly pattern. The early warning information includes the feature vector of the common anomaly pattern and recommended preventive operation parameter adjustment values.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: Existing transshipment monitoring systems can only make independent threshold judgments for single sensor indicators such as vibration, tilt, and impact, and cannot establish a deep correlation model between operating parameters and cargo damage risk. This results in the inability to pinpoint the specific operational step or parameter dimension that caused the cargo damage risk to exceed the limit. After receiving an alarm, operators still need to manually check multiple steps such as lifting, lowering, and stacking of the lifting equipment, which is inefficient, relies on experience, and makes it difficult to form closed-loop control. This invention establishes a three-level correlation mapping model between operating parameters, operating quality scores, and cargo damage risk scores. It combines forward calculation to obtain the deviation between theoretical and actual values, projects the deviation onto a multi-dimensional evaluation space for comparison with the dynamic acceptable boundary area, and uses a deep learning backpropagation mechanism to perform back calculation when the boundary is exceeded, decoupling the contribution of each parameter dimension layer by layer. This achieves accurate location of risk sources and closed-loop adjustment of operating parameters, significantly reducing manual investigation costs and effectively reducing the cargo damage rate. In addition, this invention relies on the industrial internet platform to achieve full-domain data aggregation, and completes cross-node anomaly mining, global model update and full-link responsibility traceability through cloud collaboration, so that railway logistics management and control can be upgraded from single-node independent monitoring to full-platform collaborative management and control, which greatly improves the intelligence and collaboration of the whole process management and control.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a module architecture diagram of the Internet of Things-based railway logistics end-to-end control system in an embodiment of the present invention; Figure 2 This is an internal flowchart of the reverse tracing device in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] refer to Figure 1 and Figure 2 This invention provides an embodiment of an Internet of Things-based railway logistics end-to-end management and control system, comprising: Data acquisition devices are used to collect the operating parameters and status parameters of containers during transshipment operations; The scoring calculation device is used to calculate the operation quality score based on the operation parameters and the cargo damage risk score based on the status parameters. The model building device is used to establish a three-level correlation mapping model among operating parameters, operating quality scores, and cargo damage risk scores. The forward extrapolation device is used to input the operation parameters into the three-level correlation mapping model for forward extrapolation to obtain the theoretical value of the operation quality score and the predicted value of the cargo damage risk score. The deviation acquisition device is used to compare the theoretical value of the operation quality score obtained by the forward calculation device with the actual value of the operation quality score calculated by the scoring calculation device to obtain a quality deviation sequence, compare the predicted value of the cargo damage risk score obtained by the forward calculation device with the actual value of the cargo damage risk score calculated by the scoring calculation device to obtain a risk deviation sequence, and concatenate the quality deviation sequence and the risk deviation sequence into a joint deviation vector. The projection determination device is used to project the joint deviation vector onto the multidimensional evaluation space, dynamically delineate the qualified operation boundary area in the multidimensional evaluation space based on historical operation data, and calculate the shortest distance from the projection point of the joint deviation vector to the qualified operation boundary area. The reverse tracing device is used to input the current operation parameters into the three-level association mapping model and use the deep learning backpropagation mechanism to perform reverse calculation when the shortest distance is greater than zero. In the reverse calculation process, a deviation propagation weight matrix is ​​introduced to decouple the contribution of each operation parameter dimension to the joint deviation vector layer by layer, so as to obtain the risk source identification result. The instruction feedback device is used to generate operation adjustment instructions based on the risk source identification results and to feed the operation adjustment instructions back to the replacement execution mechanism.

[0022] In this embodiment, transshipment operation refers to the complete process of transferring and unloading containers between railway freight cars, port cranes, and road freight vehicles. This process includes four consecutive operation stages: lifting the container with a spreader, moving the container in the air, placing the container at the target location with a spreader, and stacking the containers.

[0023] In this embodiment, the operating parameters refer to quantifiable data that reflect the operating status of the transshipment actuator during the transshipment operation, including the lifting acceleration of the spreader, the lowering speed of the spreader, and the stacking pressure of the containers.

[0024] In this embodiment, the state parameters refer to quantifiable data that reflect the changes in the container's own state during the transshipment process, including container three-axis vibration data, container tilt angle data, and container door light intensity change data.

[0025] In this embodiment, establishing a three-level correlation mapping model among operating parameters, operating quality scores, and cargo damage risk scores refers to using a multiple linear regression algorithm to construct mathematical mapping relationships between the three levels. The first-level mapping takes operating parameters as input and operating quality scores as output; the second-level mapping takes operating quality scores as input and cargo damage risk scores as output; and the third-level mapping is a composite function of the first two levels. When constructing this model, at least one hundred sets of historical transshipment operation sample data need to be collected. Each set of sample data includes the actual values ​​of operating parameters, operating quality scores, and cargo damage risk scores. The regression coefficients of each mapping relationship are determined by fitting using the least squares method.

[0026] In this embodiment, inputting the operation parameters into the three-level correlation mapping model for forward extrapolation means that the currently collected operation parameters are sent as input to the three-level correlation mapping model. The model calculates and outputs the theoretical value of the operation quality score according to the first-level mapping relationship. Then, the theoretical value of the operation quality score is used as the input of the second-level mapping relationship to calculate and output the predicted value of the cargo damage risk score. The forward extrapolation process does not involve any parameter adjustment or feedback.

[0027] In this embodiment, the quality deviation sequence obtained by comparing the theoretical value of the operation quality score obtained by the forward calculation device with the actual value of the operation quality score calculated by the scoring calculation device refers to the difference obtained by subtracting the actual value of the operation quality score from the theoretical value of the operation quality score. This difference is the deviation of the operation quality dimension of the changeover operation. For multiple consecutive changeover operations, the difference obtained from each calculation is arranged in chronological order to form a quality deviation sequence.

[0028] In this embodiment, the risk deviation sequence obtained by comparing the predicted value of the cargo damage risk score obtained by the forward calculation device with the actual value of the cargo damage risk score calculated by the scoring calculation device refers to the difference obtained by subtracting the actual value of the cargo damage risk score from the predicted value of the cargo damage risk score. This difference is the deviation of the cargo damage risk dimension of the transshipment operation. For multiple transshipment operations, the difference obtained from each calculation is arranged in chronological order to form a risk deviation sequence.

[0029] In this embodiment, concatenating the quality deviation sequence and the risk deviation sequence into a joint deviation vector means connecting each deviation value in the quality deviation sequence and each deviation value in the risk deviation sequence in a fixed order, combining them into a higher-dimensional vector. The dimension of this vector is equal to the sum of the length of the quality deviation sequence and the length of the risk deviation sequence. The concatenation order is to first place all the elements of the quality deviation sequence, and then place all the elements of the risk deviation sequence.

[0030] In this embodiment, the multidimensional evaluation space refers to a mathematical coordinate system with two or more dimensions. Each coordinate axis of the coordinate system corresponds to a certain dimension of the operation quality score or a certain dimension of the cargo damage risk score. The origin of the coordinate axis represents zero deviation, the positive direction of the coordinate axis represents positive deviation, and the negative direction of the coordinate axis represents negative deviation.

[0031] In this embodiment, projecting the joint deviation vector onto the multidimensional evaluation space means mapping each element of the joint deviation vector to a coordinate value on the corresponding coordinate axis in the multidimensional evaluation space, thereby determining a unique projection point in the multidimensional evaluation space. This projection point represents the comprehensive deviation status of the current transshipment operation in terms of operational quality and cargo damage risk.

[0032] In this embodiment, historical operation data refers to all transshipment operation records that have been completed and stored in the historical database during the deployment and operation of this system. Each record includes the actual value of the operation quality score, the actual value of the cargo damage risk score, and the judgment label that the operation is marked as a qualified operation or an unqualified operation.

[0033] In this embodiment, the qualified operation boundary region refers to a continuous region defined in a high-dimensional space. Any change operation corresponding to any projection point within this region is considered a qualified operation. The boundary line of this region is obtained by clustering analysis and geometric calculation of qualified data points in historical operation data. Projection points located inside the boundary region represent qualified operations, while projection points located outside the boundary region represent unqualified operations.

[0034] In this embodiment, calculating the shortest distance from the joint deviation vector projection point to the qualified operation boundary area means, in the multidimensional evaluation space, starting from the joint deviation vector projection point, drawing perpendicular lines to each boundary line segment of the qualified operation boundary area, calculating the length of all perpendicular line segments and the straight-line distance from the projection point to each boundary vertex, and taking the minimum value as the shortest distance. If the shortest distance is zero, it means that the projection point is located inside the boundary area or on the boundary line. If the shortest distance is greater than zero, it means that the projection point is located outside the boundary area.

[0035] In this embodiment, the deviation propagation weight matrix is ​​a two-dimensional matrix. The number of rows in the matrix is ​​equal to the number of dimensions of the operating parameters, and the number of columns in the matrix is ​​equal to the number of dimensions of the cargo damage risk score. Each element in the matrix represents the comprehensive sensitivity of a certain operating parameter dimension to a certain cargo damage risk score dimension. This comprehensive sensitivity is obtained by combining the local sensitivity of the operating parameters to the operating quality score and the local sensitivity of the operating quality score to the cargo damage risk score through matrix multiplication.

[0036] In this embodiment, the risk source identification result refers to the identifier of one or more operation parameter dimensions that cause the joint deviation vector projection point of the current transposition operation to be located outside the qualified operation boundary area. Each identifier corresponds to a specific operation parameter type, including abnormal lifting acceleration, abnormal placement speed, or abnormal stacking pressure distribution.

[0037] In this embodiment, generating an operation adjustment instruction based on the risk source identification result means converting the operation parameter dimensions contained in the risk source identification result into corresponding control parameter adjustment values, encapsulating them into an instruction format that the replacement execution mechanism can parse and execute, and feeding the instruction back to the replacement execution mechanism. After receiving the instruction, the replacement execution mechanism modifies the corresponding operation parameters in the next replacement operation according to the adjustment values, thereby achieving closed-loop control.

[0038] Furthermore, the projection judgment device dynamically delineates the qualified operation boundary area in the multidimensional evaluation space based on historical operation data, including: Transshipment operations whose actual cargo damage risk score is lower than the preset risk threshold are marked as qualified operations. Extract the actual values ​​of the operation quality score and cargo damage risk score of all qualified operations from historical operation data to form a set of qualified data points in the multi-dimensional evaluation space; A density-based spatial clustering algorithm is used to cluster qualified data points, identify the core set of qualified points that are density-connected, and remove isolated outliers. Perform convex hull calculation on the core set of qualified points to obtain the initial qualified boundary polygon; Extract the actual values ​​of the operation quality score and cargo damage risk score of all non-conforming operations from historical operation data to form a set of non-conforming data points in the multi-dimensional evaluation space; Calculate the shortest distance from each point in the set of unqualified data points to the initial qualified boundary polygon, and statistically fit all the shortest distances to obtain the distance distribution function; The boundary contraction coefficient is determined based on the quantiles of the distance distribution function. The boundary contraction coefficient is then used to shrink the initial qualified boundary polygon inward to obtain a dynamic qualified boundary region.

[0039] In this embodiment, the preset risk threshold is a pre-set numerical limit used to determine whether the cargo damage risk of the transshipment operation is within an acceptable range. When the actual value of the cargo damage risk score is lower than the preset risk threshold, the transshipment operation is marked as a qualified operation. When the actual value of the cargo damage risk score is higher than or equal to the preset risk threshold, the transshipment operation is marked as a non-qualified operation. The specific value of the preset risk threshold is pre-configured by the system administrator based on the safety standards of railway logistics transportation and historical cargo damage statistics.

[0040] In this embodiment, extracting the actual values ​​of the operation quality score and cargo damage risk score of all qualified operations from historical operation data to form the qualified data point set in the multidimensional evaluation space refers to traversing all transshipment operation records stored in the historical database, filtering out all records marked as qualified operations, and combining the actual values ​​of the operation quality score and cargo damage risk score of each qualified operation record into a multidimensional coordinate point. The multidimensional coordinate points corresponding to all qualified operations together form the qualified data point set, which is used to subsequently determine the distribution range and boundary characteristics of qualified operations in the multidimensional evaluation space.

[0041] In this embodiment, clustering the qualified data point set using a density-based spatial clustering algorithm refers to using the DBSCAN clustering algorithm to perform cluster analysis on all coordinate points in the qualified data point set. This algorithm takes each coordinate point as the center and a preset radius as the search range, counts the number of coordinate points contained in the search range, and when the number exceeds a preset density threshold, the coordinate point and all coordinate points in its search range are grouped into the same cluster. The above process is repeated until all coordinate points have been traversed, and finally the core set of qualified points with density connection is identified. At the same time, isolated coordinate points that do not belong to any cluster are removed as outliers.

[0042] In this embodiment, convex hull calculation of the core set of qualified points refers to using the Graham scan method or the Andrew algorithm to calculate the smallest convex polygon that can contain all coordinate points in the core set of qualified points. The boundary of the convex polygon is composed of several straight line segments connected end to end. The line connecting any two points inside the convex polygon is entirely inside the convex polygon. This convex polygon is the initial qualified boundary polygon, representing the minimum convex hull range of qualified operations in the multidimensional evaluation space.

[0043] In this embodiment, non-conforming operations refer to transshipment operations where the actual value of the cargo damage risk score is higher than or equal to the preset risk threshold. Such operations indicate that events such as vibration, tilting, impact, or abnormal opening of the container door occurred during the transshipment process, which may lead to damage to the cargo inside the container. These operations need to be identified and analyzed to prevent similar situations from happening again.

[0044] In this embodiment, extracting the actual values ​​of the operational quality score and cargo damage risk score of all non-conforming operations from historical operation data to form the non-conforming data point set in the multi-dimensional evaluation space refers to traversing all transshipment operation records stored in the historical database, filtering out all records marked as non-conforming operations, and combining the actual values ​​of the operational quality score and cargo damage risk score of each non-conforming operation record into a multi-dimensional coordinate point. The multi-dimensional coordinate points corresponding to all non-conforming operations together form the non-conforming data point set, which is used to analyze the degree of deviation of non-conforming operations from the boundary area of ​​conforming operations.

[0045] In this embodiment, calculating the shortest distance from each point in the set of non-conforming data points to the initial qualified boundary polygon means that for each multi-dimensional coordinate point in the set of non-conforming data points, the perpendicular distance from the point to the straight line containing each edge of the initial qualified boundary polygon and the straight-line distance from the point to each vertex of the initial qualified boundary polygon are calculated respectively. The minimum value among all the calculation results is taken as the shortest distance from the point to the initial qualified boundary polygon. This shortest distance reflects the degree of deviation between the boundary area of ​​non-conforming operations and qualified operations. The larger the distance, the more serious the deviation.

[0046] In this embodiment, statistical fitting of all shortest distances means taking the shortest distances corresponding to all points in the set of non-conforming data points as a set of sample data, and using the kernel density estimation method to fit the probability distribution of this set of sample data to obtain a distance distribution function. This distance distribution function describes the probability distribution characteristics of the distance between non-conforming operations and the boundary area of ​​conforming operations.

[0047] In this embodiment, the quantile of the distance distribution function refers to the position of the probability distribution corresponding to a specific cumulative probability after dividing the probability distribution into several equal parts in the order of numerical size. For example, the 90th percentile means that 90% of the sample data in the distance distribution function are less than or equal to this value. The quantile is used to quantify the statistical distribution characteristics of the degree of deviation of non-conforming operations.

[0048] In this embodiment, determining the boundary shrinkage coefficient based on the quantile of the distance distribution function means selecting a preset quantile from the distance distribution function as a reference value, and calculating the ratio of this reference value to the upper limit of the tolerance for non-compliance distance of the initial qualified boundary polygon to obtain the boundary shrinkage coefficient. The boundary shrinkage coefficient is a value between zero and one. The larger the value, the smaller the degree of inward shrinkage of the qualified operation boundary area, and the smaller the value, the greater the degree of shrinkage.

[0049] In this embodiment, the boundary shrinkage coefficient is a dimensionless value between zero and one, used to control the extent of inward shrinkage of the initial qualified boundary polygon. The specific value of the coefficient is calculated based on the quantile of the distance distribution function. When the quantile is large, the boundary shrinkage coefficient is close to one, and the initial qualified boundary polygon hardly shrinks. When the quantile is small, the boundary shrinkage coefficient is close to zero, and the initial qualified boundary polygon shrinks significantly inward.

[0050] In this embodiment, shrinking the initial qualified boundary polygon inward using the boundary shrinkage coefficient means subtracting the product of the direction vector from the vertex to the geometric center of the polygon and the boundary shrinkage coefficient from the coordinates of each vertex of the initial qualified boundary polygon to obtain the new vertex coordinates after shrinkage. Connecting all the new vertex coordinates in sequence forms a new polygon, which is the dynamic qualified boundary region. The dynamic qualified boundary region is located inside the initial qualified boundary polygon, and its size and shape are dynamically adjusted according to the statistical distribution of the deviation degree of non-qualified operations.

[0051] Furthermore, after obtaining the dynamic qualified boundary region, the projection determination device is also used for: Real-time acquisition of the joint deviation vector projection points of newly added costume change operations; When a newly added projection point is located within the dynamic qualified boundary area, the actual value of the operation quality score and the actual value of the cargo damage risk score corresponding to the newly added projection point will be added to the qualified data point set as incremental data. At preset time intervals, density clustering and convex hull calculation are performed again on the qualified data point set to obtain the updated dynamic qualified boundary region. When a newly added projection point is located outside the dynamic qualified boundary area but is marked as qualified by subsequent manual review, the newly added projection point will be forcibly added to the qualified data point set and trigger a local expansion correction of the boundary. Specifically, the dynamic qualified boundary area will be locally expanded outward with the newly added projection point as the center according to the preset expansion radius, so that the newly added projection point is included in the expanded dynamic qualified boundary area.

[0052] In this embodiment, a new garment replacement operation refers to a garment replacement operation that was not included in the historical operation data before the current time during the deployment and operation of this system, and is currently being executed or has just been completed. Data of this type of operation is collected in real time and used to dynamically update the qualified operation boundary area.

[0053] In this embodiment, real-time acquisition of the joint deviation vector projection point of the new transshipment operation refers to obtaining the actual value of the operation quality score and the actual value of the cargo damage risk score of the operation during the execution of the new transshipment operation, according to the same acquisition frequency and calculation method as the historical operation data, and then calculating the corresponding joint deviation vector, and projecting the joint deviation vector into the multidimensional evaluation space to obtain the projection point coordinates of the new transshipment operation in the multidimensional evaluation space.

[0054] In this embodiment, the preset time period is a pre-set time interval length used to control the update frequency of the qualified data point set and the dynamic qualified boundary region. This time interval length is pre-configured by the system administrator according to the busyness of the replacement operation and the data accumulation speed. For example, it can be set to 24 hours or 7 days. Every time interval, the system automatically performs a re-clustering of the qualified data point set and a recalculation of the dynamic qualified boundary region.

[0055] In this embodiment, when a new projection point is located inside the dynamic qualified boundary area, adding the actual value of the operation quality score and the actual value of the cargo damage risk score corresponding to the new projection point as incremental data to the qualified data point set means that the system determines whether the coordinates of the new projection point are within the space enclosed by the current dynamic qualified boundary area. If it is inside, the new transshipment operation is regarded as a qualified operation, and its actual value of the operation quality score and the actual value of the cargo damage risk score are added as a new multi-dimensional coordinate point to the qualified data point set, thereby realizing the incremental expansion of the qualified data point set.

[0056] In this embodiment, the density clustering and convex hull calculation of the qualified data point set are re-performed every preset time period to obtain the updated dynamic qualified boundary region. This means that every time a preset time period has elapsed, the system automatically takes all the coordinate points in the current qualified data point set as input, re-executes the density-based spatial clustering algorithm to cluster the qualified data point set to identify the core set of qualified points with density connections and remove isolated outliers. Then, the convex hull calculation is performed again on the qualified point core set obtained after clustering to obtain a new initial qualified boundary polygon. Finally, the new initial qualified boundary polygon is shrunk inward according to the boundary shrinkage coefficient to obtain the updated dynamic qualified boundary region.

[0057] In this embodiment, when a newly added projection point is located outside the dynamic qualified boundary area but is subsequently marked as qualified by manual review, the newly added projection point is forcibly added to the qualified data point set, and the local expansion correction of the boundary is triggered. This means that the system first determines that the coordinates of the newly added projection point are outside the current dynamic qualified boundary area. Then, the manual reviewer reviews the newly added garment operation. If the review result determines that the operation is qualified, the manual reviewer marks the operation as qualified in the system. After receiving the mark, the system forcibly adds the actual value of the operation quality score and the actual value of the cargo damage risk score of the operation to the qualified data point set, and at the same time triggers the boundary local expansion correction process to avoid misjudging qualified operations as unqualified due to overly strict boundary delineation. The specific process of local expansion correction is as follows: taking the location of the newly added projection point as the expansion center point and the preset expansion radius as the expansion distance, the boundary in the dynamic qualified boundary area towards the direction of the newly added projection point is pushed outward, so that the newly added projection point is included in the expanded dynamic qualified boundary area. The expanded boundary maintains continuity and convexity.

[0058] In this embodiment, the preset expansion radius is a pre-set distance value used to control the outward expansion of the dynamic qualified boundary area during local expansion correction. This distance value represents the distance of outward expansion from the location of the newly added projection point. The system takes the newly added projection point as the center and moves the boundary line segment closest to the newly added projection point in the dynamic qualified boundary area outward by the preset expansion radius, so that the newly added projection point is included in the expanded boundary area. The specific value of the preset expansion radius is pre-configured by the system administrator according to the normal deviation fluctuation range of the replacement operation.

[0059] like Figure 2 As shown, further, the reverse tracing device inputs the current operating parameters into the three-level association mapping model and uses a deep learning backpropagation mechanism for reverse calculation. During the reverse calculation process, a deviation propagation weight matrix is ​​introduced to decouple the contribution of each operating parameter dimension to the joint deviation vector layer by layer, obtaining the risk source identification result, including: The three-level correlation mapping model is expressed as a composite function of the first mapping function and the second mapping function. The first mapping function describes the mapping relationship from the operation parameters to the operation quality score, and the second mapping function describes the mapping relationship from the operation quality score to the cargo damage risk score. Perform a first-order Taylor expansion on the first mapping function at the current position to obtain the first Jacobian matrix. The elements of the first Jacobian matrix represent the local sensitivity of each operation parameter dimension to each operation quality score dimension. Perform a first-order Taylor expansion on the second mapping function at the current position to obtain the second Jacobian matrix. The elements of the second Jacobian matrix represent the local sensitivity of each operation quality score dimension to the cargo damage risk score dimension. Multiplying the first Jacobian matrix by the second Jacobian matrix yields the bias propagation weight matrix, whose elements represent the overall sensitivity of each operational parameter dimension to the cargo damage risk scoring dimension. Multiply the joint bias vector by the pseudo-inverse of the bias propagation weight matrix to obtain the contribution vector of each operational parameter dimension. The contribution vector is normalized, and one or more operational parameter dimensions with the highest absolute value of contribution after normalization are identified as sources of risk.

[0060] In this embodiment, expressing the three-level correlation mapping model as a composite function of the first and second mapping functions means splitting the entire three-level correlation mapping model into two independent functions. The first mapping function describes the mapping relationship between operating parameters and operating quality scores, and the second mapping function describes the mapping relationship between operating quality scores and cargo damage risk scores. The entire three-level correlation mapping model is equal to the composite of the second and first mapping functions. That is, the first mapping function first maps the operating parameters to the operating quality scores, and then the second mapping function maps the operating quality scores to the cargo damage risk scores.

[0061] In this embodiment, the current position refers to the specific coordinate point corresponding to the operation parameters of the current transshipment operation in the multi-dimensional operation parameter space. Each coordinate value of this coordinate point corresponds to the specific values ​​of operation parameters such as the lifting acceleration of the spreader, the placement speed of the spreader, and the stacking pressure of the container. The current position is used to determine the specific deployment point of the Taylor unfolding.

[0062] In this embodiment, the local sensitivity of each operation parameter dimension to each operation quality score dimension refers to the magnitude of the change in a certain operation quality score dimension when a certain operation parameter dimension undergoes a small change. This magnitude reflects the degree of influence of the operation parameter dimension on the operation quality score dimension. The larger the absolute value of the local sensitivity, the more significant the influence.

[0063] In this embodiment, performing a first-order Taylor expansion of the first mapping function at the current position means expanding the first mapping function at the position of the operation parameter of the current changing operation into a sum of constant terms and first-order derivative terms, ignoring second-order and higher-order terms. The first Jacobian matrix obtained after expansion is a two-dimensional matrix. The number of rows of this matrix is ​​equal to the number of dimensions of the operation quality score, and the number of columns of this matrix is ​​equal to the number of dimensions of the operation parameter. The element in the i-th row and i-th column of the matrix represents the local sensitivity of the i-th operation quality score dimension to the i-th operation parameter dimension.

[0064] In this embodiment, the local sensitivity of each operation quality scoring dimension to the cargo damage risk scoring dimension refers to the magnitude of the change in a certain cargo damage risk scoring dimension when a certain operation quality scoring dimension undergoes a small change. This magnitude reflects the degree of influence of the operation quality scoring dimension on the cargo damage risk scoring dimension. The larger the absolute value of the local sensitivity, the more significant the influence.

[0065] In this embodiment, performing a first-order Taylor expansion of the second mapping function at the current position means expanding the second mapping function at the position of the operation quality score of the current transshipment operation into a sum of constant terms and first-order derivative terms, ignoring second-order and higher-order terms. The resulting second Jacobian matrix is ​​a two-dimensional matrix. The number of rows in this matrix is ​​equal to the number of dimensions of the cargo damage risk score, and the number of columns in this matrix is ​​equal to the number of dimensions of the operation quality score. The element in the i-th row and i-th column of the matrix represents the local sensitivity of the i-th cargo damage risk score dimension to the i-th operation quality score dimension.

[0066] In this embodiment, multiplying the first Jacobian matrix and the second Jacobian matrix means, according to the rules of matrix multiplication, using the second Jacobian matrix as the left multiplication matrix and the first Jacobian matrix as the right multiplication matrix, calculating their product to obtain the deviation propagation weight matrix. The meaning of this product operation is to compositely transfer the local sensitivity of the operation parameter dimension to the operation quality score dimension and the local sensitivity of the operation quality score dimension to the cargo damage risk score dimension, thereby obtaining the comprehensive sensitivity of the operation parameter dimension to the cargo damage risk score dimension.

[0067] In this embodiment, the overall sensitivity of each operational parameter dimension to the cargo damage risk scoring dimension refers to the magnitude of the change in a cargo damage risk scoring dimension after being passed through the intermediate operational quality scoring dimension when a certain operational parameter dimension undergoes a slight change. This overall sensitivity is equal to the sum of the products of the local sensitivity of the operational parameter dimension to all operational quality scoring dimensions and the local sensitivity of the corresponding operational quality scoring dimension to the cargo damage risk scoring dimension.

[0068] In this embodiment, the pseudo-inverse of the deviation propagation weight matrix refers to the pseudo-inverse operation of a non-square matrix deviation propagation weight matrix with unequal number of rows and columns, which yields a new matrix. The result of multiplying this new matrix with the deviation propagation weight matrix is ​​the identity matrix. The pseudo-inverse operation is used to solve a system of linear equations when the deviation propagation weight matrix is ​​not invertible, and to obtain an approximate solution that minimizes the error.

[0069] In this embodiment, multiplying the joint deviation vector by the pseudo-inverse of the deviation propagation weight matrix means treating the joint deviation vector as a column vector and left-multiplying the column vector by the pseudo-inverse of the deviation propagation weight matrix to obtain the contribution vector of each operational parameter dimension. The dimension of the contribution vector is equal to the number of operational parameter dimensions. Each element in the vector represents the contribution of the corresponding operational parameter dimension to the current joint deviation vector. A positive contribution indicates that the operational parameter dimension is biased towards the positive deviation direction, and a negative contribution indicates that the operational parameter dimension is biased towards the negative deviation direction. The larger the absolute value of the contribution, the greater the influence of the operational parameter dimension on the overall deviation.

[0070] In this embodiment, normalizing the contribution vector means dividing each element of the contribution vector by the sum of the absolute values ​​of all elements in the vector to obtain a normalized contribution value, making the sum of the absolute values ​​of all normalized contribution values ​​equal to one. Then, the operation parameter dimensions are sorted from largest to smallest according to the absolute values ​​of the normalized contribution values, and one or more operation parameter dimensions at the top of the sort are identified as sources of risk. The operation parameter dimensions identified as sources of risk are the main reasons why the current changing operation deviates from the qualified operation boundary area.

[0071] Furthermore, the reverse tracing device uses ridge regression regularization estimation instead of direct pseudo-inverse calculation, specifically including: Obtain the bias propagation weight matrix using the above method; Extract multiple historical operation parameter samples and corresponding historical joint deviation vector samples from historical operation data; The ridge regression algorithm is used to perform regularized estimation of the bias propagation weight matrix to obtain the regularized bias propagation weight matrix. Multiply the joint bias vector by the pseudo-inverse of the regularized bias propagation weight matrix to obtain the regularized contribution vector.

[0072] In this embodiment, extracting multiple historical operation parameter samples and corresponding historical joint deviation vector samples from historical operation data means traversing all the changing operation records stored in the historical database. For each changing operation record, the operation parameters of that operation are extracted as historical operation parameter samples, and the joint deviation vector corresponding to that operation is extracted as historical joint deviation vector samples. The two are paired to form a set of samples. The samples of all changing operation records together constitute the training dataset for ridge regression estimation.

[0073] In this embodiment, the ridge regression algorithm is used to perform regularized estimation of the bias propagation weight matrix. This involves using historical operating parameter samples from the training dataset as input features and historical joint bias vector samples as output targets. A regularized estimation matrix is ​​calculated using the ridge regression method. A penalty term is introduced during the calculation of this regularized estimation matrix. This penalty term is equal to the sum of the squares of all elements in the bias propagation weight matrix multiplied by a regularization parameter. By minimizing the sum of the prediction error on the training data and the penalty term, the regularized bias propagation weight matrix is ​​obtained. This matrix can effectively reduce the risk of overfitting and improve robustness to noisy data.

[0074] In this embodiment, the pseudo-inverse of the regularized bias propagation weight matrix refers to performing a pseudo-inverse operation on the regularized bias propagation weight matrix obtained after ridge regression regularization estimation. Since the regularization process improves the condition number of the matrix, the numerical stability of the pseudo-inverse operation is higher than that of directly performing a pseudo-inverse operation on the original bias propagation weight matrix. The calculated pseudo-inverse matrix can more accurately reflect the inverse mapping relationship between the dimension of the operating parameters and the joint bias vector.

[0075] In this embodiment, multiplying the joint deviation vector by the pseudo-inverse of the regularized deviation propagation weight matrix means treating the joint deviation vector of the current alternation operation as a column vector, and multiplying this column vector on the left by the pseudo-inverse of the regularized deviation propagation weight matrix to obtain the regularized contribution vector. Since the regularized contribution vector has undergone ridge regression regularization, its value is more stable and less sensitive to noise and outliers in historical operation data, and it can more reliably reflect the true contribution of each operation parameter dimension to the current joint deviation vector.

[0076] Furthermore, the data acquisition device includes: The six-axis accelerometer integrated in the spreader collects the three components of lifting acceleration and the three components of placement speed at a sampling frequency of no less than 100 Hz. Four pressure sensors are installed at the four bottom corners of the container to collect data on the stacked pressure distribution. The sampling times of the four pressure sensors are kept synchronized. A nine-axis inertial measurement unit is installed at the geometric center of the container to collect three-axis vibration data, three-axis angular velocity data, and container tilt angle data. Two photosensitive sensors are symmetrically installed on the upper and lower sides of the container door frame to collect data on changes in light intensity when the door is opened. The difference in light intensity between the two photosensitive sensors is used to eliminate ambient light interference.

[0077] In this embodiment, the six-axis accelerometer integrated in the lifting device refers to the six-axis accelerometer installed on the lifting device of the changing actuator. This sensor can simultaneously measure linear acceleration in three mutually perpendicular directions and angular velocity in three mutually perpendicular directions. The sensor collects the three components of lifting acceleration and the three components of positioning velocity at a sampling frequency of not less than 100 Hz. The sampling frequency of 100 Hz means that the sensor collects 100 sets of data per second, which can capture the rapid dynamic changes during the changing operation.

[0078] In this embodiment, the three components of lifting acceleration and the three components of placement speed refer to the linear acceleration values ​​of the spreader in three mutually perpendicular directions during the lifting stage and the motion speed values ​​of the spreader in three mutually perpendicular directions during the placement stage. The three directions are defined as the vertical direction perpendicular to the horizontal plane, the horizontal direction parallel to the long side of the container, and the horizontal direction parallel to the short side of the container. The vertical component of the lifting acceleration reflects the impact intensity of the spreader during lifting, and the vertical component of the placement speed reflects the buffering control effect of the spreader during placement.

[0079] In this embodiment, the installation of four pressure sensors at the four bottom corners of the container means that the four pressure sensors are fixed at the four bottom corners of the container, with one pressure sensor installed at each bottom corner. The synchronous sampling time of the four pressure sensors means that at each sampling time, the four pressure sensors simultaneously collect the pressure value at their respective positions, ensuring that the four pressure data collected at the same time can accurately reflect the instantaneous pressure distribution at the four bottom corners of the container. Synchronous sampling is achieved by triggering a unified clock signal.

[0080] In this embodiment, the stacking pressure distribution data refers to the data set consisting of four pressure values ​​collected by the pressure sensors at the four bottom corners of the container at the sampling time. This data set reflects the force distribution at the four bottom corners of the container. When the container is lifted away from the lower support surface, the four pressure values ​​return to zero at the same time. When the container is placed on the lower support surface or on top of other containers, the four pressure values ​​rise from zero and eventually stabilize. By analyzing the relative magnitude of the four pressure values, it can be determined whether the stacking posture of the containers after they are placed is uniform.

[0081] In this embodiment, installing the nine-axis inertial measurement unit at the geometric center of the container means fixing the nine-axis inertial measurement unit at the exact center of the container's internal space. This position can most evenly reflect the overall attitude changes and vibration transmission characteristics of the container. The nine-axis inertial measurement unit is a combined measurement device that integrates multiple sensors to simultaneously collect three-axis vibration data, three-axis angular velocity data, and container tilt angle data.

[0082] In this embodiment, the three-axis vibration data refers to the numerical sequence of linear acceleration of the container in three mutually perpendicular directions over time, reflecting the impact and vibration intensity experienced by the container during the transshipment operation; the three-axis angular velocity data refers to the numerical sequence of angular velocity of the container rotating around three mutually perpendicular directions over time, reflecting the rotational motion state of the container during the hoisting process; and the container tilt angle data refers to the values ​​of the pitch and roll angles of the container relative to the horizontal plane, reflecting the degree of attitude deviation of the container during hoisting and placement.

[0083] In this embodiment, the two photosensitive sensors are symmetrically installed on the upper and lower sides of the container door frame. This means that one photosensitive sensor is installed on the inner side of the upper frame of the container door frame, and the other photosensitive sensor is installed on the inner side of the lower frame. The two sensors are aligned vertically and symmetrical with respect to the horizontal center line of the door frame. The two photosensitive sensors are used to collect light intensity change data when the door is opened. The light intensity difference between the two photosensitive sensors is used to eliminate ambient light interference. This means that the light intensity value collected by the upper frame photosensitive sensor is subtracted from the light intensity value collected by the lower frame photosensitive sensor. Since the influence of ambient light on the upper and lower sensors is basically the same, the ambient light components cancel each other out after the subtraction. The remaining difference mainly reflects the intensity change of external light entering through the door gap when the door is opened, thereby accurately detecting the door opening event.

[0084] Furthermore, the scoring calculation device calculates an operation quality score based on the operating parameters, including: The composite acceleration curve is extracted from the three components of the lifting acceleration. The peak value of the composite acceleration curve is detected during the lifting period. The ratio of the peak value to the standard lifting acceleration benchmark value is used as the original lifting impact score. The original lifting impact score is mapped by an S-shaped function to obtain the normalized lifting impact score. The vertical velocity curve is extracted from the three components of the landing velocity. The ratio of the standard deviation to the mean of the vertical velocity curve during the landing period is calculated as the velocity variation coefficient. The velocity variation coefficient is then mapped to the landing stability score through an exponential decay function. The pressure values ​​collected by the pressure sensors at the four bottom corners are used to form a four-dimensional pressure vector. After normalizing the four-dimensional pressure vector, the Shannon entropy is calculated to obtain the stacked pressure distribution entropy. The stacked pressure distribution entropy is then mapped to the stacked uniform score through an inverse proportional function. The lifting impact score, the smooth placement score, and the uniform stacking score are weighted and summed according to preset weights to obtain the operation quality score.

[0085] In this embodiment, extracting the composite acceleration curve from the three components of lifting acceleration means calculating the composite acceleration by taking the square root of the acceleration values ​​in the three mutually perpendicular directions during the lifting process. That is, the square value of each of the three acceleration components at each sampling time is calculated and summed, and then the square root of the sum is taken to obtain the composite acceleration value at that time. The composite acceleration values ​​at all sampling times are arranged in chronological order to form the composite acceleration curve.

[0086] In this embodiment, the lifting period refers to the continuous time interval between when the spreader starts lifting the container upwards during the transshipment operation and when the container is completely separated from the support surface below. The start time of this period is determined by the increase of the spreader's lifting acceleration from zero and the decrease of the pressure sensor value at the bottom of the container. The end time of this period is determined by the return of all four pressure sensors at the bottom of the container to zero.

[0087] In this embodiment, peak detection of the synthetic acceleration curve during the lifting period refers to traversing every numerical point on the synthetic acceleration curve within the time range of the lifting period, finding the maximum value of the synthetic acceleration during that period as the peak value, and using the ratio of the peak value to the standard lifting acceleration benchmark value as the original score of the lifting impact. When the ratio is greater than one, it indicates that the actual lifting impact exceeds the benchmark value, and when the ratio is less than one, it indicates that the actual lifting impact is lower than the benchmark value.

[0088] In this embodiment, the standard lifting acceleration benchmark value is a pre-set reference acceleration value. This value is determined based on the railway logistics transportation safety regulations and the statistical average value of lifting acceleration in qualified operations from historical transshipment operation data. It represents the impact control level that should be achieved during the lifting process under normal transshipment operation conditions and is used as a reference standard to judge whether the current lifting impact is too large.

[0089] In this embodiment, mapping the original lifting impact score to a S-shaped function means inputting the original lifting impact score into a S-shaped function, and converting the input value into an output value between zero and one through the nonlinear transformation of the S-shaped function. This output value is the normalized lifting impact score. When the original lifting impact score is much less than one, the normalized lifting impact score is close to zero. When the original lifting impact score is much greater than one, the normalized lifting impact score is close to one. When the original lifting impact score is equal to one, the normalized lifting impact score is equal to 0.5.

[0090] In this embodiment, the S-shaped function mapping is a nonlinear transformation method that converts any real input value into an output value between zero and one. The mathematical form of the S-shaped function used in this method is that the output value is equal to one divided by one plus the negative power of the input value with the natural constant as the base. The output value of this function is 0.5 when the input value is zero, the output value approaches zero when the input value decreases towards negative infinity, and the output value approaches one when the input value increases towards positive infinity.

[0091] In this embodiment, extracting the vertical velocity curve from the three components of the placement speed means selecting the vertical velocity component separately from the three mutually perpendicular components of the placement speed, arranging the values ​​of the vertical velocity component at each sampling time in chronological order to form a vertical velocity curve. This curve reflects the change law of the vertical motion speed of the lifting device over time during the placement process.

[0092] In this embodiment, the placement period refers to the continuous time interval between when the container begins to contact the support surface below and when it is completely and stably placed during the transshipment operation. The start time of this period is determined by the increase of the pressure sensor value at the bottom of the container from zero, and the end time of this period is determined by the stabilization of the values ​​of the four pressure sensors at the bottom of the container and the absence of significant changes.

[0093] In this embodiment, the ratio of the standard deviation to the mean of the vertical velocity curve during the landing period is calculated by first calculating the arithmetic mean of all numerical points on the vertical velocity curve during the landing period, then calculating the square root of the sum of squares of the deviations of each numerical point from the arithmetic mean, dividing it by the number of numerical points to obtain the standard deviation, and finally dividing the standard deviation by the arithmetic mean to obtain the velocity variation coefficient. The velocity variation coefficient reflects the degree of fluctuation of the vertical velocity during the landing process. The larger the value, the more uneven the velocity.

[0094] In this embodiment, the exponential decay function is a nonlinear transformation method that converts the input value into the output value through exponential operation. The output value of this function is equal to the negative power of the input value with the natural constant as the base. When the input value is zero, the output value is one. When the input value increases, the output value approaches zero at an exponential rate. This function has the characteristic that the output value is always positive and decreases monotonically as the input value increases.

[0095] In this embodiment, mapping the velocity variation coefficient to the landing stability score through the exponential decay function means using the calculated velocity variation coefficient as the input value of the exponential decay function, and the output value obtained after calculation by the exponential decay function is the landing stability score. When the velocity variation coefficient is zero, the landing stability score is one, indicating that the landing process is completely stable. When the velocity variation coefficient increases, the landing stability score decreases rapidly, indicating that there is obvious velocity fluctuation in the landing process.

[0096] In this embodiment, constructing a four-dimensional pressure vector from the pressure values ​​collected by the pressure sensors at the four bottom corners means arranging the four pressure values ​​collected by the four bottom corner pressure sensors at the same sampling time according to a fixed bottom corner order to form a vector containing four elements. Normalizing this four-dimensional pressure vector means dividing each element in the vector by the sum of the four elements so that the sum of the four normalized elements equals one. Calculating the Shannon entropy after normalization means multiplying each normalized element by the logarithm of that element to the base 2, then taking the negative value of the four products and summing them to obtain the stacking pressure distribution entropy. This entropy value reflects the uniformity of the pressure distribution at the four bottom corners of the container; the larger the entropy value, the more uniform the pressure distribution.

[0097] In this embodiment, mapping the stack pressure distribution entropy to the stack uniformity score through an inverse proportional function means using the stack pressure distribution entropy as the input value of the inverse proportional function. The output value obtained after calculation by the inverse proportional function is the stack uniformity score. When the stack pressure distribution entropy is close to its maximum value, the stack uniformity score is close to one, indicating that the pressure distribution is very uniform. When the stack pressure distribution entropy is small, the stack uniformity score decreases significantly, indicating that the pressure distribution is concentrated in one or two bottom corners.

[0098] In this embodiment, the inverse proportional function is a nonlinear transformation method in which the output value is inversely proportional to the input value. Specifically, the output value is equal to the input value divided by the sum of the input value and the preset half-saturation constant. When the input value is zero, the output value is zero. When the input value is much larger than the preset half-saturation constant, the output value is close to one. This function maps the range of input values ​​from zero to infinity to the range of output values ​​from zero to one.

[0099] In this embodiment, the preset weights refer to three coefficients that are set in advance for the lifting impact score, the smooth placement score, and the uniform stacking score. These three coefficients are all values ​​between zero and one, and the sum of the three is equal to one. The specific values ​​of the preset weights are configured by the system administrator according to the different levels of importance attached to the three indicators of lifting impact, smooth placement, and uniform stacking in railway logistics transportation.

[0100] In this embodiment, the weighted summation of the lifting impact score, the placement stability score, and the stacking uniformity score according to preset weights means multiplying the lifting impact score by the preset weight corresponding to the lifting impact score, multiplying the placement stability score by the preset weight corresponding to the placement stability score, and multiplying the stacking uniformity score by the preset weight corresponding to the stacking uniformity score, and then adding the three product results together. The summation result is the operation quality score, which comprehensively reflects the overall operation quality level of the transshipment operation in the three stages of lifting, placement, and stacking.

[0101] Furthermore, the scoring calculation device calculates a cargo damage risk score based on the status parameters, including: The vertical vibration component is extracted from the triaxial vibration data. The vertical vibration component is integrally squared during the transshipment operation period to obtain the vertical vibration dose value. At the same time, the two vibration components in the horizontal plane are integrally squared to obtain the horizontal vibration dose value. The vertical vibration dose value and the horizontal vibration dose value are weighted and fused according to the height of the cargo center of gravity to obtain the cumulative vibration dose value. From the box tilt angle data collected by the nine-axis inertial measurement unit, the cumulative duration of pitch angle exceeding the preset pitch threshold and the cumulative duration of roll angle exceeding the preset roll threshold are counted, and the maximum value of the two is taken as the tilt over-limit duration value. The light intensity difference between the two photosensitive sensors is calculated from the light intensity data collected by the two photosensitive sensors. The light intensity difference sequence is subjected to differential operation to obtain the light intensity change rate. The number of times the light intensity change rate exceeds the preset change threshold is counted to obtain the number of abnormal openings of the cabinet door. Peak angular velocity is extracted from the three-axis angular velocity data collected by the nine-axis inertial measurement unit. When the peak angular velocity exceeds the preset angular velocity threshold, an impact event marker is generated, and the number of impact events during the changing operation period is counted. After normalizing the cumulative vibration dose value, tilt over-limit duration value, number of abnormal door openings and number of impact events by maximum and minimum values, the data are then weighted and fused according to the risk weight vector corresponding to the cargo type to obtain the cargo damage risk score.

[0102] In this embodiment, extracting the vertical vibration component from the triaxial vibration data means selecting the vibration value in the vertical direction perpendicular to the horizontal plane from the vibration data in three mutually perpendicular directions collected by the nine-axis inertial measurement unit, and arranging the values ​​of the vertical vibration value at each sampling time in chronological order to form a vertical vibration component sequence.

[0103] In this embodiment, the transshipment operation period refers to the complete continuous time interval from the start to the end of the transshipment operation. The start time of this period is the moment when the spreader begins to lift the container, and the end time of this period is the moment when the container is fully positioned and the stacking pressure distribution data is stable.

[0104] In this embodiment, the square integration of the vertical vibration component during the transshipment operation period means squaring each value in the vertical vibration component sequence and then summing all the squared values ​​over the time length of the transshipment operation period. The obtained integral result is the vertical vibration dose value, which reflects the total amount of vibration energy accumulated in the vertical direction of the container.

[0105] In this embodiment, the sum of squares of the two vibration components in the horizontal plane is to extract the vibration components in two mutually perpendicular directions parallel to the horizontal plane from the triaxial vibration data, square the values ​​of the two vibration components at the same sampling time, add them together, and then integrate the sum over the time length of the transshipment operation. The obtained integral result is the horizontal vibration dose value, which reflects the total amount of vibration energy accumulated by the container in the horizontal plane.

[0106] In this embodiment, the cargo center of gravity height refers to the vertical height of the overall center of gravity of the cargo loaded inside the container from the bottom surface of the container. This value is calculated based on the cargo type, the way the cargo is stacked inside the container, and the weight distribution of the cargo. Different types of cargo have different center of gravity height characteristics.

[0107] In this embodiment, the weighted fusion of vertical and horizontal vibration dose values ​​based on the cargo's center of gravity height means first calculating the vertical and horizontal vibration weights based on the cargo's center of gravity height. The higher the cargo's center of gravity height, the greater the vertical vibration weight; the lower the cargo's center of gravity height, the greater the horizontal vibration weight. Then, the vertical vibration dose value is multiplied by the vertical vibration weight, and the horizontal vibration dose value is multiplied by the horizontal vibration weight. Finally, the two products are added together to obtain the cumulative vibration dose value. This cumulative vibration dose value reflects the overall vibration damage risk after comprehensively considering the cargo's center of gravity height.

[0108] In this embodiment, the preset pitch threshold is a pre-set pitch angle limit value used to determine whether the tilt of the container in the front-to-back direction exceeds the safe range. When the absolute value of the pitch angle in the container tilt angle data exceeds the preset pitch threshold, it indicates that the tilt of the container in the front-to-back direction has reached a level that needs attention.

[0109] In this embodiment, the preset roll threshold is a pre-set roll angle limit value used to determine whether the tilt of the container in the left and right directions exceeds the safe range. When the absolute value of the roll angle in the container tilt angle data exceeds the preset roll threshold, it indicates that the tilt of the container in the left and right directions has reached a level that needs attention.

[0110] In this embodiment, from the container tilt angle data collected by the nine-axis inertial measurement unit, the cumulative duration of pitch angle exceeding a preset pitch threshold and the cumulative duration of roll angle exceeding a preset roll threshold are statistically analyzed. The maximum value of the two is taken as the tilt over-limit duration value. This means that during the transshipment operation period, all moments when the absolute value of the pitch angle is greater than the preset pitch threshold are recorded, and the total duration of these moments is calculated as the pitch over-limit duration. At the same time, all moments when the absolute value of the roll angle is greater than the preset roll threshold are recorded, and the total duration of these moments is calculated as the roll over-limit duration. The pitch over-limit duration and the roll over-limit duration are compared, and the larger value is taken as the tilt over-limit duration value. This duration value reflects the most serious degree of deviation of the container's attitude from the safe range during the transshipment operation.

[0111] In this embodiment, the light intensity difference sequence refers to the difference obtained by subtracting the light intensity values ​​collected by two photosensitive sensors at the same sampling time, and arranging the differences at all sampling times in chronological order to form a light intensity difference sequence. This sequence reflects the change pattern of the intensity of external light entering during the opening of the cabinet door.

[0112] In this embodiment, obtaining the light intensity change rate by performing differential operation on the light intensity difference sequence refers to subtracting the difference between two adjacent sampling times in the light intensity difference sequence to obtain the change in the difference, and then dividing the change by the time interval between adjacent sampling times to obtain the light intensity change rate within that time interval. The light intensity change rate reflects the degree of drastic change in light intensity when the door is opened.

[0113] In this embodiment, the preset change threshold is a pre-set limit value for the rate of change of light intensity, which is used to determine whether the door opening event has actually occurred. When the rate of change of light intensity exceeds the preset change threshold, it indicates that a valid door opening event has been detected.

[0114] In this embodiment, calculating the light intensity difference between the two photosensitive sensors from the light intensity data collected by the two photosensitive sensors involves subtracting the light intensity value collected by the upper frame photosensitive sensor from the light intensity value collected by the lower frame photosensitive sensor to obtain the light intensity difference at that sampling moment. Performing a differential operation on the light intensity difference sequence to obtain the light intensity change rate involves subtracting the light intensity difference between two adjacent sampling moments and dividing by the sampling time interval. The number of times the light intensity change rate exceeds the preset change threshold refers to the number of times the cumulative light intensity change rate exceeds the preset change threshold during the changing operation period. This cumulative number is the number of times the cabinet door is abnormally opened.

[0115] In this embodiment, extracting the peak angular velocity from the three-axis angular velocity data collected by the nine-axis inertial measurement unit means that during the changing operation period, all angular velocity values ​​in the three directions of the three-axis angular velocity data are traversed, and the angular velocity value with the largest absolute value in the three directions is found as the peak angular velocity.

[0116] In this embodiment, the preset angular velocity threshold is a pre-set angular velocity limit value used to determine whether the container has suffered a violent impact rotational motion. When the peak angular velocity exceeds the preset angular velocity threshold, it indicates that the container has experienced an impact event that may cause damage to the cargo during the transshipment operation.

[0117] In this embodiment, the peak angular velocity is extracted from the three-axis angular velocity data collected by the nine-axis inertial measurement unit. When the peak angular velocity exceeds the preset angular velocity threshold, an impact event marker is generated. The number of impact events during the changing operation period is calculated by dividing the time into multiple consecutive time windows, extracting the peak angular velocity in each time window, and generating an impact event marker if the peak angular velocity exceeds the preset angular velocity threshold. After traversing all time windows, the total number of impact event markers is counted, and this total number is the number of impact events.

[0118] In this embodiment, normalizing the cumulative vibration dose value, tilt over-limit duration value, number of abnormal door openings, and number of impact events by maximum and minimum values ​​means that for each of the four indicators, the maximum and minimum values ​​of the indicator are found from historical operation data. The minimum value of the indicator in the current transshipment operation is subtracted from the minimum value and then divided by the difference between the maximum and minimum values ​​to obtain a normalized value between zero and one. Weighted fusion according to the risk weight vector corresponding to the cargo type means that the four normalized values ​​are multiplied by the weight coefficient of the corresponding position in the risk weight vector corresponding to the cargo type, and then the four products are added together. The sum is the cargo damage risk score.

[0119] In this embodiment, the risk weight vector corresponding to the cargo type is a vector containing four weight coefficients. The four weight coefficients of this vector correspond to the cumulative vibration dose value, the duration of tilt exceeding the limit value, the number of abnormal openings of the container door, and the number of impact events, respectively. The risk weight vector values ​​are different for different cargo types. The weight coefficients for the cumulative vibration dose value and the number of impact events are higher for fragile cargo, the weight coefficient for the duration of tilt exceeding the limit value is higher for liquid cargo, and the four weight coefficients are relatively average for ordinary dry goods. The sum of the weight coefficients in the risk weight vector of each cargo type is equal to one.

[0120] Furthermore, it also includes a liability tracing device for performing the following steps in the event of a cargo damage dispute: The historical operation records generated and stored by the projection determination device and the reverse tracing device are obtained and stored in the historical database. Obtain the operational quality score sequence and cargo damage risk score sequence for disputed transshipment operations, and concatenate the two sequences into a joint score time series matrix; Retrieve multiple historical job records from the historical database that have a multidimensional Frescher distance less than a preset distance threshold with the joint scoring time series matrix. Each historical job record contains the historical joint scoring time series matrix and the historical responsibility determination result. The responsibility reference weight for each historical operation record is calculated based on the reciprocal of the multidimensional Frescher distance between the historical joint scoring time series matrix of each historical operation record and the joint scoring time series matrix of the disputed change operation. The historical responsibility determination results of all historical operation records are weighted and summed according to the corresponding responsibility reference weights to obtain the confidence vector of dispute responsibility attribution; The carrier with the highest confidence level is output as the responsible party, and a responsibility determination matrix is ​​also output. The responsibility determination matrix includes the three operational parameter dimensions with the highest contribution and the three state parameter dimensions with the highest contribution.

[0121] In this embodiment, obtaining the historical operation records generated and stored by the projection determination device and the reverse tracing device refers to reading all data records generated and saved by the projection determination device and the reverse tracing device during the previous transshipment operation from the historical database. Each historical operation record includes the operation quality score sequence, cargo damage risk score sequence, joint score time series matrix, multidimensional Frescher distance comparison result, and final responsibility determination result for the transshipment operation.

[0122] In this embodiment, obtaining the operational quality score sequence and cargo damage risk score sequence of the disputed transshipment operation refers to extracting the numerical sequences of the operational quality score and cargo damage risk score over time from the complete record of the transshipment operation in which the cargo damage dispute occurred. Concatenating the two sequences into a joint scoring time series matrix after aligning them by time means combining the two values ​​of the operational quality score sequence and cargo damage risk score sequence under the same time index into a two-dimensional vector, arranging all the two-dimensional vectors under the time index in chronological order to form a matrix with two rows and multiple columns. This matrix is ​​the joint scoring time series matrix.

[0123] In this embodiment, the multidimensional Frescher distance between the joint scoring time series matrix and the joint scoring time series matrix is ​​a metric used to measure the similarity between the two joint scoring time series matrices. This metric takes into account the similarity in shape and the alignment on the time axis of the two multidimensional time series curves represented by the two matrices. The smaller the multidimensional Frescher distance, the higher the similarity between the two curves, and the larger the multidimensional Frescher distance, the greater the difference between the two curves.

[0124] In this embodiment, the preset distance threshold is a pre-set distance limit value used to determine whether the similarity between two joint scoring time series matrices reaches the reference standard. When the multidimensional Frescher distance is less than the preset distance threshold, it indicates that the changing operations corresponding to the two joint scoring time series matrices have a sufficiently high similarity, and the responsibility determination result of one operation can be used as a reference basis for the responsibility determination of the other operation.

[0125] In this embodiment, retrieving multiple historical operation records in the historical database whose multidimensional Frescher distance between the historical operation record and the joint scoring time series matrix is ​​less than a preset distance threshold means traversing all historical operation records stored in the historical database. For each historical operation record, the multidimensional Frescher distance between the historical joint scoring time series matrix in the historical operation record and the joint scoring time series matrix of the disputed clothing change operation is calculated. All historical operation records whose calculated multidimensional Frescher distance is less than the preset distance threshold are selected. Each selected historical operation record contains the historical joint scoring time series matrix and the historical responsibility determination result.

[0126] In this embodiment, the historical responsibility determination result refers to the final responsibility attribution conclusion for the cargo damage dispute of this transshipment operation stored in the historical operation record. The conclusion includes the name of the responsible party, the basis for the responsibility determination, and the proportion of responsibility. The name of the responsible party includes one or more of the railway carrier, port carrier, or road carrier.

[0127] In this embodiment, the responsibility reference weight of each historical operation record is calculated based on the reciprocal of the multidimensional Frescher distance between the historical joint scoring time series matrix of each historical operation record and the joint scoring time series matrix of the disputed equipment change operation. This means that for each selected historical operation record, the reciprocal of its multidimensional Frescher distance is first calculated. The smaller the multidimensional Frescher distance, the larger the reciprocal. Then, the reciprocals of all selected historical operation records are summed. The reciprocal of each historical operation record is divided by the summation result to obtain the responsibility reference weight of that historical operation record. The responsibility reference weight is between zero and one, and the sum of the responsibility reference weights of all historical operation records is equal to one. The historical operation record with the smaller the multidimensional Frescher distance receives a higher responsibility reference weight.

[0128] In this embodiment, the weighted summation of the historical responsibility determination results of all historical operation records according to the corresponding responsibility reference weights means that for each carrier category, the number of times or responsibility percentage of that carrier appears in the historical responsibility determination results of each historical operation record is multiplied by the responsibility reference weight of that historical operation record, and then the product results of all historical operation records are added together to obtain the confidence level of the carrier's dispute responsibility attribution. The confidence levels of the dispute responsibility attribution of all carriers are combined into a vector, which is the dispute responsibility attribution confidence vector. The value of each element in the vector is between zero and one, and the sum of all elements is equal to one.

[0129] In this embodiment, outputting the carrier with the highest confidence level as the responsible party means finding the element with the largest value from the confidence vector of dispute responsibility attribution, and outputting the carrier corresponding to that element as the final responsibility determination result. At the same time, a responsibility determination basis matrix is ​​output. The responsibility determination basis matrix is ​​a two-dimensional matrix. The rows of the matrix correspond to the selected historical operation records, and the columns of the matrix correspond to the operation parameter dimension and the status parameter dimension. Each element in the matrix represents the specific value of the three operation parameter dimensions and the three status parameter dimensions with the highest contribution in the corresponding historical operation record. This responsibility determination basis matrix is ​​used to show the user the detailed reasons and supporting data for the responsibility determination.

[0130] This invention provides an embodiment of an Internet of Things-based railway logistics end-to-end management and control platform, comprising: Multiple IoT-based railway logistics end-to-end management and control systems, such as any one of the above, are deployed at a transshipment operation node; A cloud-based collaborative server communicates and connects with all changing operation nodes. The cloud-based collaborative server includes: A cross-node data aggregator is used to collect operation quality scores, cargo damage risk scores, risk source identification results, and operation adjustment instructions from each transshipment operation node, and aggregate and store them according to time and space dimensions. The cross-node anomaly pattern miner is used to perform time-series clustering analysis on aggregated data, identify common anomaly patterns that recur in multiple transshipment operation nodes, and mark the operation parameter dimensions corresponding to the common anomaly patterns as global risk factors. A global model updater is used to update the three-level association mapping model of each transshipment operation node by using the global risk factor as an additional regularization constraint through transfer learning. The cross-node responsibility tracer is used to construct a cross-node responsibility chain by connecting the joint scoring time series matrix of each node in chronological order when a cargo damage event involves multiple transshipment operation nodes. It then performs segmented Fraser distance retrieval on the cross-node responsibility chain to obtain the cross-node responsibility attribution determination result. The early warning broadcaster is used to broadcast early warning information to all transshipment operation nodes when the cross-node anomaly pattern miner identifies a new common anomaly pattern. The early warning information includes the feature vector of the common anomaly pattern and recommended preventive operation parameter adjustment values.

[0131] In this embodiment, deploying each system at a transshipment operation node means that the IoT-based railway logistics full-process control system as described in any one of claims 1 to 9 is installed and run as a complete software and hardware integrated unit at each physical location where transshipment operation monitoring is required. Each transshipment operation node independently performs data collection, scoring calculation, projection judgment, reverse tracing and instruction feedback functions, and independently maintains its own historical operation database and three-level association mapping model.

[0132] In this embodiment, collecting operation quality scores, cargo damage risk scores, risk source identification results, and operation adjustment instructions from each transshipment operation node means that the cloud-based collaborative server periodically sends data requests to each transshipment operation node through a network communication protocol. After receiving the request, each transshipment operation node packages and uploads the operation quality score sequence, cargo damage risk score sequence, risk source identification result list, and operation adjustment instruction list generated by the node in the most recent period to the cloud-based collaborative server. Aggregated storage according to time and spatial dimensions means that the cloud-based collaborative server records the timestamp of each data point and the geographical location identifier of the transshipment operation node that generated the data when storing the data, and organizes the data of all nodes into a multidimensional dataset according to time order and spatial location relationship.

[0133] In this embodiment, performing time-series clustering analysis on the aggregated data means that the cloud-based collaborative server takes the operation quality score sequence and cargo damage risk score sequence of all the transshipment operation nodes obtained after aggregation as input, and uses a time-series clustering algorithm to divide these sequences into multiple clusters according to shape similarity. Identifying common abnormal patterns that recur in multiple transshipment operation nodes means analyzing the common features of the sequences in each cluster. When a certain abnormal feature pattern appears simultaneously in the sequences of at least two different transshipment operation nodes, the abnormal feature pattern is marked as a common abnormal pattern, and the operation parameter dimension corresponding to the common abnormal pattern is marked as a global risk factor. The global risk factor indicates that the abnormality in the operation parameter dimension is not a problem unique to individual nodes, but a systemic risk that is prevalent in multiple nodes.

[0134] In this embodiment, using the global risk factor as an additional regularization constraint means adding a penalty term to the original model parameter optimization objective function when updating the three-level association mapping model of a certain transshipment operation node. This penalty term makes the model parameters tend to maintain a direction consistent with the global risk factor during the update process. The transfer learning update of the three-level association mapping model of each transshipment operation node means that the cloud collaborative server uses the global model parameters trained after aggregating all node data as the source model and the local three-level association mapping model of each transshipment operation node as the target model. The parameter layers related to the global risk factor are extracted from the source model, these parameter layers are transferred to the target model, and then the historical operation data of the node is used to fine-tune the transferred target model so that the target model can adapt to the local features of the node while retaining the global commonality.

[0135] In this embodiment, when a cargo damage event involves multiple transshipment operation nodes, it means that a batch of containers, during the complete transportation process from origin to destination, passed through two or more different transshipment operation nodes for loading and unloading operations, and cargo damage was discovered upon final receipt, but it was impossible to determine which transshipment operation node the damage occurred at. The joint scoring time series matrix of each node in chronological order refers to splicing the joint scoring time series matrices generated by each transshipment operation node through which the batch of containers passed during the transshipment period in chronological order of transshipment occurrence, forming a complete cross-node joint scoring time series matrix sequence. Constructing a cross-node responsibility chain means using the time interval and spatial distance between two adjacent matrices in the sequence as connection weights to form a directed graph structure, which is the cross-node responsibility chain. Segmented Fraser distance retrieval of the cross-node responsibility chain means comparing each segment in the cross-node responsibility chain with the cross-node responsibility chain segments stored in the historical database using Fraser distance, finding the historical segment with the highest similarity and its corresponding responsibility attribution judgment result, and obtaining the cross-node responsibility attribution judgment result.

[0136] In this embodiment, when the cross-node anomaly pattern miner identifies a new common anomaly pattern, it means that the cloud collaboration server, during the periodic execution of time-series clustering analysis of aggregated data, discovers an anomaly feature pattern that has not been recorded before, and this anomaly feature pattern appears in at least two garment operation nodes. Broadcasting the warning information to all garment operation nodes means that the cloud collaboration server pushes the relevant data of the newly discovered common anomaly pattern to all deployed garment operation nodes simultaneously through the communication network. The warning information includes the feature vector of the common anomaly pattern and the recommended preventive operation parameter adjustment value.

[0137] In this embodiment, the feature vector of the common anomaly pattern is a multi-dimensional numerical vector. Each dimension of the vector corresponds to an operation parameter dimension or a state parameter dimension. Each element in the vector represents the typical deviation degree of that dimension in the common anomaly pattern. This feature vector serves as a digital identifier of the common anomaly pattern and is used by each transshipment operation node to quickly match and identify the anomaly pattern locally.

[0138] In this embodiment, the recommended preventive operation parameter adjustment value refers to a set of operation parameter correction amounts calculated by the cloud collaboration server based on the feature vector of the common anomaly pattern and the operation parameter adjustment records of successfully correcting such anomalies in historical data. After receiving the adjustment value, each garment operation node can modify the corresponding operation parameters in advance when performing the same type of garment operation next time, thereby avoiding the recurrence of the same anomaly.

[0139] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A railway logistics end-to-end control system based on the Internet of Things, characterized in that, include: Data acquisition devices are used to collect the operating parameters and status parameters of containers during transshipment operations; The scoring calculation device is used to calculate the operation quality score based on the operation parameters and the cargo damage risk score based on the status parameters. The model building device is used to establish a three-level correlation mapping model among operating parameters, operating quality scores, and cargo damage risk scores. The forward extrapolation device is used to input the operation parameters into the three-level correlation mapping model for forward extrapolation to obtain the theoretical value of the operation quality score and the predicted value of the cargo damage risk score. The deviation acquisition device is used to compare the theoretical value of the operation quality score obtained by the forward calculation device with the actual value of the operation quality score calculated by the scoring calculation device to obtain a quality deviation sequence, compare the predicted value of the cargo damage risk score obtained by the forward calculation device with the actual value of the cargo damage risk score calculated by the scoring calculation device to obtain a risk deviation sequence, and concatenate the quality deviation sequence and the risk deviation sequence into a joint deviation vector. The projection determination device is used to project the joint deviation vector onto the multidimensional evaluation space, dynamically delineate the qualified operation boundary area in the multidimensional evaluation space based on historical operation data, and calculate the shortest distance from the projection point of the joint deviation vector to the qualified operation boundary area. The reverse tracing device is used to input the current operation parameters into the three-level association mapping model and use the deep learning backpropagation mechanism to perform reverse calculation when the shortest distance is greater than zero. In the reverse calculation process, a deviation propagation weight matrix is ​​introduced to decouple the contribution of each operation parameter dimension to the joint deviation vector layer by layer, so as to obtain the risk source identification result. The instruction feedback device is used to generate operation adjustment instructions based on the risk source identification results and to feed the operation adjustment instructions back to the replacement execution mechanism.

2. The railway logistics end-to-end control system based on the Internet of Things as described in claim 1, characterized in that, The projection judgment device dynamically delineates the qualified operation boundary area in a multi-dimensional evaluation space based on historical operation data, including: Transshipment operations whose actual cargo damage risk score is lower than the preset risk threshold are marked as qualified operations. Extract the actual values ​​of the operation quality score and cargo damage risk score of all qualified operations from historical operation data to form a set of qualified data points in the multi-dimensional evaluation space; A density-based spatial clustering algorithm is used to cluster qualified data points, identify the core set of qualified points that are density-connected, and remove isolated outliers. Perform convex hull calculation on the core set of qualified points to obtain the initial qualified boundary polygon; Extract the actual values ​​of the operation quality score and cargo damage risk score of all non-conforming operations from historical operation data to form a set of non-conforming data points in the multi-dimensional evaluation space; Calculate the shortest distance from each point in the set of unqualified data points to the initial qualified boundary polygon, and statistically fit all the shortest distances to obtain the distance distribution function; The boundary contraction coefficient is determined based on the quantiles of the distance distribution function. The boundary contraction coefficient is then used to shrink the initial qualified boundary polygon inward to obtain a dynamic qualified boundary region.

3. The railway logistics end-to-end control system based on the Internet of Things as described in claim 2, characterized in that, After obtaining the dynamic qualified boundary region, the projection determination device is also used for: Real-time acquisition of the joint deviation vector projection points of newly added costume change operations; When a newly added projection point is located within the dynamic qualified boundary area, the actual value of the operation quality score and the actual value of the cargo damage risk score corresponding to the newly added projection point will be added to the qualified data point set as incremental data. At preset time intervals, density clustering and convex hull calculation are performed again on the qualified data point set to obtain the updated dynamic qualified boundary region. When a newly added projection point is located outside the dynamic qualified boundary area but is marked as qualified by subsequent manual review, the newly added projection point will be forcibly added to the qualified data point set and trigger a local expansion correction of the boundary. Specifically, the dynamic qualified boundary area will be locally expanded outward with the newly added projection point as the center according to the preset expansion radius, so that the newly added projection point is included in the expanded dynamic qualified boundary area.

4. The railway logistics end-to-end control system based on the Internet of Things as described in claim 1, characterized in that, The reverse tracing device inputs the current operating parameters into a three-level correlation mapping model and uses a deep learning backpropagation mechanism for reverse calculation. During the reverse calculation process, a deviation propagation weight matrix is ​​introduced to decouple the contribution of each operating parameter dimension to the joint deviation vector layer by layer, obtaining the risk source identification results, including: The three-level correlation mapping model is expressed as a composite function of the first mapping function and the second mapping function. The first mapping function describes the mapping relationship from the operation parameters to the operation quality score, and the second mapping function describes the mapping relationship from the operation quality score to the cargo damage risk score. Perform a first-order Taylor expansion on the first mapping function at the current position to obtain the first Jacobian matrix. The elements of the first Jacobian matrix represent the local sensitivity of each operation parameter dimension to each operation quality score dimension. Perform a first-order Taylor expansion on the second mapping function at the current position to obtain the second Jacobian matrix. The elements of the second Jacobian matrix represent the local sensitivity of each operation quality score dimension to the cargo damage risk score dimension. Multiplying the first Jacobian matrix by the second Jacobian matrix yields the bias propagation weight matrix, whose elements represent the overall sensitivity of each operational parameter dimension to the cargo damage risk scoring dimension. Multiply the joint bias vector by the pseudo-inverse of the bias propagation weight matrix to obtain the contribution vector of each operational parameter dimension. The contribution vector is normalized, and one or more operational parameter dimensions with the highest absolute value of contribution after normalization are identified as sources of risk.

5. The railway logistics end-to-end control system based on the Internet of Things according to claim 4, characterized in that, The reverse source tracing device uses ridge regression regularization estimation instead of direct pseudo-inverse calculation, specifically including: The deviation propagation weight matrix is ​​obtained in accordance with the method described in claim 4; Extract multiple historical operation parameter samples and corresponding historical joint deviation vector samples from historical operation data; The ridge regression algorithm is used to perform regularized estimation of the bias propagation weight matrix to obtain the regularized bias propagation weight matrix. Multiply the joint bias vector by the pseudo-inverse of the regularized bias propagation weight matrix to obtain the regularized contribution vector.

6. The railway logistics end-to-end control system based on the Internet of Things according to claim 1, characterized in that, The data acquisition device includes: The six-axis accelerometer integrated in the spreader collects the three components of lifting acceleration and the three components of placement speed at a sampling frequency of no less than 100 Hz. Four pressure sensors are installed at the four bottom corners of the container to collect data on the stacked pressure distribution. The sampling times of the four pressure sensors are kept synchronized. A nine-axis inertial measurement unit is installed at the geometric center of the container to collect three-axis vibration data, three-axis angular velocity data, and container tilt angle data. Two photosensitive sensors are symmetrically installed on the upper and lower sides of the container door frame to collect data on changes in light intensity when the door is opened. The difference in light intensity between the two photosensitive sensors is used to eliminate ambient light interference.

7. The railway logistics end-to-end control system based on the Internet of Things according to claim 1, characterized in that, The scoring calculation device calculates an operation quality score based on operating parameters, including: The composite acceleration curve is extracted from the three components of the lifting acceleration. The peak value of the composite acceleration curve is detected during the lifting period. The ratio of the peak value to the standard lifting acceleration benchmark value is used as the original lifting impact score. The original lifting impact score is mapped by an S-shaped function to obtain the normalized lifting impact score. The vertical velocity curve is extracted from the three components of the landing velocity. The ratio of the standard deviation to the mean of the vertical velocity curve during the landing period is calculated as the velocity variation coefficient. The velocity variation coefficient is then mapped to the landing stability score through an exponential decay function. The pressure values ​​collected by the pressure sensors at the four bottom corners are used to form a four-dimensional pressure vector. After normalizing the four-dimensional pressure vector, the Shannon entropy is calculated to obtain the stacked pressure distribution entropy. The stacked pressure distribution entropy is then mapped to the stacked uniform score through an inverse proportional function. The lifting impact score, the smooth placement score, and the uniform stacking score are weighted and summed according to preset weights to obtain the operation quality score.

8. The railway logistics end-to-end control system based on the Internet of Things according to claim 1, characterized in that, The scoring calculation device calculates a cargo damage risk score based on status parameters, including: The vertical vibration component is extracted from the triaxial vibration data. The vertical vibration component is integrally squared during the transshipment operation period to obtain the vertical vibration dose value. At the same time, the two vibration components in the horizontal plane are integrally squared to obtain the horizontal vibration dose value. The vertical vibration dose value and the horizontal vibration dose value are weighted and fused according to the height of the cargo center of gravity to obtain the cumulative vibration dose value. From the box tilt angle data collected by the nine-axis inertial measurement unit, the cumulative duration of pitch angle exceeding the preset pitch threshold and the cumulative duration of roll angle exceeding the preset roll threshold are counted, and the maximum value of the two is taken as the tilt over-limit duration value. The light intensity difference between the two photosensitive sensors is calculated from the light intensity data collected by the two photosensitive sensors. The light intensity difference sequence is subjected to differential operation to obtain the light intensity change rate. The number of times the light intensity change rate exceeds the preset change threshold is counted to obtain the number of abnormal openings of the cabinet door. Peak angular velocity is extracted from the three-axis angular velocity data collected by the nine-axis inertial measurement unit. When the peak angular velocity exceeds the preset angular velocity threshold, an impact event marker is generated, and the number of impact events during the changing operation period is counted. After normalizing the cumulative vibration dose value, tilt over-limit duration value, number of abnormal door openings and number of impact events by maximum and minimum values, the data are then weighted and fused according to the risk weight vector corresponding to the cargo type to obtain the cargo damage risk score.

9. The railway logistics end-to-end control system based on the Internet of Things according to claim 1, characterized in that, It also includes a liability tracing device to perform the following steps when a cargo damage dispute occurs: The historical operation records generated and stored by the projection determination device and the reverse tracing device are obtained and stored in the historical database. Obtain the operational quality score sequence and cargo damage risk score sequence for disputed transshipment operations, and concatenate the two sequences into a joint score time series matrix; Retrieve multiple historical job records from the historical database that have a multidimensional Frescher distance less than a preset distance threshold with the joint scoring time series matrix. Each historical job record contains the historical joint scoring time series matrix and the historical responsibility determination result. The responsibility reference weight for each historical operation record is calculated based on the reciprocal of the multidimensional Frescher distance between the historical joint scoring time series matrix of each historical operation record and the joint scoring time series matrix of the disputed change operation. The historical responsibility determination results of all historical operation records are weighted and summed according to the corresponding responsibility reference weights to obtain the confidence vector of dispute responsibility attribution; The carrier with the highest confidence level is output as the responsible party, and a responsibility determination matrix is ​​also output. The responsibility determination matrix includes the three operational parameter dimensions with the highest contribution and the three state parameter dimensions with the highest contribution.

10. A railway logistics end-to-end management and control platform based on the Internet of Things, characterized in that, include: Multiple IoT-based railway logistics end-to-end control systems as described in any one of claims 1 to 9, each system being deployed at a transshipment operation node; A cloud-based collaborative server communicates and connects with all changing operation nodes. The cloud-based collaborative server includes: A cross-node data aggregator is used to collect operation quality scores, cargo damage risk scores, risk source identification results, and operation adjustment instructions from each transshipment operation node, and aggregate and store them according to time and space dimensions. The cross-node anomaly pattern miner is used to perform time-series clustering analysis on aggregated data, identify common anomaly patterns that recur in multiple transshipment operation nodes, and mark the operation parameter dimensions corresponding to the common anomaly patterns as global risk factors. A global model updater is used to update the three-level association mapping model of each transshipment operation node by using the global risk factor as an additional regularization constraint through transfer learning. The cross-node responsibility tracer is used to construct a cross-node responsibility chain by connecting the joint scoring time series matrix of each node in chronological order when a cargo damage event involves multiple transshipment operation nodes. It then performs segmented Fraser distance retrieval on the cross-node responsibility chain to obtain the cross-node responsibility attribution determination result. The early warning broadcaster is used to broadcast early warning information to all transshipment operation nodes when the cross-node anomaly pattern miner identifies a new common anomaly pattern. The early warning information includes the feature vector of the common anomaly pattern and recommended preventive operation parameter adjustment values.

Images