A smart land port operation method and system based on artificial intelligence

By optimizing container storage locations and route planning using artificial intelligence, and combining it with real-time monitoring technology, the problems of unreasonable container placement and low loading and unloading efficiency at land port terminals have been solved, achieving efficient and safe automated management.

CN118674410BActive Publication Date: 2024-11-22SHANGHAI WENJING INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411163353.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-23
Publication Date
2024-11-22
Estimated Expiration
2044-08-23

AI Technical Summary

Technical Problem

Traditional land port terminals rely on manual operation and traditional data management methods, resulting in unreasonable container placement, low loading and unloading efficiency, insufficient safety, and low level of informatization, which affects operational efficiency and safety.

Method used

An AI-based smart land port operation method is adopted, which optimizes container storage location through a random forest model, calculates navigation path by combining Dijkstra's algorithm, and uses a dual-stream dilated 3D convolutional network and a generalized regression network for real-time monitoring. It integrates intelligent location selection, operation and monitoring modules to achieve automated and intelligent management.

Benefits of technology

It improved the space utilization and operational accuracy of the station, reduced human error and safety risks, increased the speed and safety of loading and unloading operations, and enhanced the real-time performance and efficiency of overall monitoring.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118674410B_ABST
    Figure CN118674410B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of smart land port management, and in particular to a smart land port operation method and system based on artificial intelligence, including step S1, collecting and preprocessing station data and information; step S2, constructing a key feature set, randomly selecting a feature subset for training to evaluate feature importance and generate corresponding prediction results, integrating all prediction results, and adjusting weight parameters according to adjustment factors to generate recommended storage locations; step S3, calculating a navigation path according to the actual location and target location of the terminal through the Dijkstra algorithm; step S4, analyzing real-time monitoring data through a dual-stream expansion 3D convolutional network and a generalized regression network, and automatically detecting and responding to abnormal behaviors. The random forest model is used to evaluate container characteristics and accurately predict storage locations in combination with real-time vacancy data to improve space utilization efficiency; real-time navigation and automatic path updates improve the operating efficiency and accuracy of manual and unmanned equipment, ensure smooth loading and unloading operations, and automatically detect abnormal behaviors.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of smart land port management, and in particular to an artificial intelligence-based smart land port operation method and system. Background Art

[0002] Land ports play an important hub role in modern logistics and transportation networks, and terminals are the core operating areas in land port operations. These terminals mainly rely on physical equipment and manual operations to manage and perform cargo loading, unloading and storage operations. Mechanical equipment widely used in terminals, such as forklifts and cranes, requires a large amount of manpower for operation and maintenance during operation, which not only increases operating costs but also affects operating efficiency. In terms of data management, most land port terminals still use traditional manual entry methods, relying on manual entry into computer systems or paper records. This processing method is inefficient and easily leads to untimely information updates and data errors. This affects the decision-making and efficiency of the entire operation.

[0003] In addition, traditional land ports have significant deficiencies in safety monitoring, such as the lack of advanced monitoring and identification technology, which makes it difficult to detect and resolve cargo damage in a timely manner. This not only increases operational risks, but may also lead to significant property losses. At the same time, the low level of informatization and insufficient information sharing hinder coordination between departments, further reducing work efficiency and service quality. In this context, the specific problems faced by traditional land ports include:

[0004] (1) Due to the lack of global monitoring and intelligent dispatching systems, unreasonable container location selection often occurs, resulting in waste of container terminal space and inconvenience in loading and unloading.

[0005] (2) When searching for the location of a container, staff can only rely on limited location information, which wastes more time for staff who are not familiar with the terminal and affects work efficiency.

[0006] (3) The loading and unloading of containers mainly relies on workers operating forklifts or gantry cranes. This manual operation method not only leads to the problem of incorrect storage of containers, but also affects the efficiency and safety of loading and unloading.

[0007] (4) The work monitoring method that relies on manual photography and recording is inefficient, lacks monitoring security, and is prone to monitoring blind spots. Summary of the invention

[0008] The purpose of the invention is to propose an intelligent land port operation method and system based on artificial intelligence in view of the above technical problems, which solves the problems of unreasonable container location selection, low efficiency of container picking, low efficiency of manual loading and unloading of containers, and low efficiency of operation monitoring, and realizes the rational use of terminal vacancies, improves the work efficiency of land port terminals, reduces the safety problems and work errors that may occur manually by reducing manual operations, realizes the operation guarantee of container storage and transportation through the overall monitoring of land port terminals, and greatly reduces the operation time. The purpose of the invention can be achieved through the following technical solutions:

[0009] The present invention provides an intelligent land port operation method based on artificial intelligence, comprising the following steps:

[0010] Step S1, collecting and preprocessing terminal data, container information and historical storage information;

[0011] Step S2: input the current container information into the intelligent location selection module to output the recommended storage location; wherein:

[0012] Based on the data collected from step S1, a key feature set is constructed, and the intelligent location selection module randomly selects a feature subset from the key feature set for training through a random forest model; the feature importance of each feature subset in the decision tree is evaluated, and the corresponding prediction results are generated; at the same time, the expected number of inbound containers of each type is calculated as an adjustment factor by combining the remaining number of empty containers in the current terminal and the historical demand for each type of container; the prediction results of all decision trees are integrated through a weighted average method, and the weight parameters in the random forest model are adjusted according to the adjustment factors to generate a recommended storage location;

[0013] Step S3: convert the pre-processed terminal data and container information into a real-time global map, and calculate the navigation path according to the actual location and target location of the terminal type through the Dijkstra algorithm; wherein,

[0014] When the terminal type is an unmanned automated device, the distance from the loading and unloading equipment to the positioning point is calculated offline as the path information, and a corresponding path database is constructed for pre-storage; the navigation path in the basic database is updated according to online task requirements or changes in equipment location;

[0015] Step S4, real-time monitoring of the real-time monitoring data of the field station data in step S1, analyze the real-time monitoring data through a dual-stream dilated 3D convolutional network and a generalized regression network, and automatically detect and respond to abnormal behaviors to ensure safety and operational efficiency within the field station.

[0016] Further, step S1 is specifically,

[0017] High-definition cameras and various sensors including temperature, humidity, vibration and smoke sensors are deployed in the operation areas and channels of the station to obtain station data; among them,

[0018] The station data is the station map, the spatial status of the operation area and the channel, including the layout, size, capacity, environmental data and real-time monitoring data;

[0019] Container information includes the size, type, and weight of the container;

[0020] Historical storage information includes the container's entry and exit times and storage locations.

[0021] Further, step S2 is specifically,

[0022] Step S21, obtaining basic features associated with the container, including terminal data and container information, and calculating derived features including the unit weight, expected departure time and movement frequency of the container, and forming a complete key feature set with the basic features and derived features;

[0023] Step S22, randomly selecting a feature subset from the key feature set for training each decision tree, for each decision tree, calculating the Gini impurity reduction of the selected feature at the split point, evaluating the importance of the feature in the location selection process, and generating a predicted storage location for each decision tree based on the importance;

[0024] Step S23, combining the number of remaining container spaces in the current terminal and the historical demand for each type of container, predicting the expected number of inbound containers of each type as an adjustment factor for optimizing the weight parameters in the random forest model;

[0025] Step S24, the prediction results of all decision trees are integrated by weighted average method, and adjusted in combination with adjustment factors to generate a recommended storage location.

[0026] Further, in step S22, for each decision tree, the Gini impurity reduction of the selected feature is calculated at the split point to evaluate the importance of the feature in the position selection process, specifically,

[0027] Step S221, for each decision tree, randomly select a feature subset from the key feature set for node segmentation in the decision tree construction process;

[0028] Step S222: For each node of the decision tree, the Gini impurity of each feature is calculated and the optimal feature is selected for segmentation. The optimal feature is the feature with the lowest Gini impurity. The Gini impurity calculation formula is:

[0029] ,

[0030] in, is the Gini impurity of node I, J is the total number of categories in the feature subset, and P is the proportion of samples of node I belonging to the jth category;

[0031] Step S223, for each usage feature Nodes for segmentation, calculation of features The Gini impurity reduction is calculated as follows:

[0032] ;

[0033] in, is the reduction of Gini impurity, is the impurity of node I, and is the Gini impurity of the left and right child nodes after segmentation;

[0034] Step S224, accumulating features The reduction of Gini impurity;

[0035] ;

[0036] in, To include features The segmentation set of

[0037] Step S224, normalization is performed based on the accumulated Gini impurity reduction of each feature to obtain a corresponding importance score.

[0038] Further, step S23 is specifically,

[0039] Step S231, collecting the number of remaining containers in the current terminal and the historical demand for various types of containers;

[0040] Step S232, define the objective function , where Q is the number of containers entering the warehouse, the historical number of containers entering the warehouse of each type is fitted into a random distribution D, and d is the actual demand sampled from the random distribution D;

[0041] Step S233, calculate and minimize the expected value of the objective function, and obtain the corresponding inventory quantity Q as the predicted quantity, that is:

[0042] ;

[0043] ;

[0044] Among them, I is the information set, including the remaining number of containers in the current terminal and the historical demand; D is a random distribution fitted with the demand d based on the historical demand; Q is the predicted number of containers of each type entering the warehouse, and d is the actual demand.

[0045] Further, step S3 is specifically,

[0046] Step S31: construct a real-time updated global map based on the terminal data and container data, and install positioning beacons at corresponding locations to obtain the location information of the positioning points;

[0047] Step S32: Based on the terminal type, determine the corresponding path information mode. When the terminal type is a manually operated device, the Dijkstra algorithm is used to calculate the recommended navigation path according to the actual position and target position of the manually operated device. When the terminal type is a loading and unloading device, the recommended path information from all loading and unloading devices to the positioning point is calculated and saved in the offline stage, and pre-stored in the path database. In the online stage, the corresponding path information is extracted from the path database and optimized and updated according to the task requirements or changes in the equipment position to generate a navigation path.

[0048] Further, step S4 is specifically,

[0049] Deploy cameras in stations to obtain real-time monitoring data;

[0050] The dual-stream expanded 3D convolutional network is used to extract features from the video stream collected by the camera, including spatial features and temporal features, to identify complex behavior patterns.

[0051] The extracted spatial and temporal features are input into the generalized regression network, which predicts the probability of abnormal behavior based on the training model. When the probability of abnormal behavior exceeds the preset threshold, an alarm is automatically triggered.

[0052] Based on the same inventive concept, the present invention also provides an artificial intelligence-based smart land port operation system, which adopts the above-mentioned smart land port operation method, including:

[0053] The intelligent location selection module collects and preprocesses terminal data, container information, and historical storage information, and constructs a key feature set. It randomly selects a feature subset from the key feature set for training through the random forest model; evaluates the feature importance of each feature subset in the decision tree and generates the corresponding prediction results; combines the number of remaining container spaces in the current terminal and the historical demand for each type of container to calculate the expected number of inbound containers of each type as an adjustment factor; combines the prediction results of all decision trees through the weighted average method, and adjusts the weight parameters in the random forest model according to the adjustment factors to generate a recommended storage location;

[0054] The operation module includes a box-finding unit and a loading and unloading unit. When the terminal type is a manually operated device, the navigation path is calculated in the box-finding unit according to the actual position and target position of the device through the Dijkstra algorithm. When the terminal type is an unmanned automated device, the distance from the loading and unloading device to the positioning point is calculated offline as the path information, and a corresponding path database is constructed for pre-storage. The navigation path in the basic database is updated according to the online task requirements or the change of the equipment position, and the automated loading and unloading is performed according to the navigation path navigation.

[0055] The monitoring module conducts all-round monitoring of the site through installed high-definition cameras and various environmental sensors; among them, video monitoring automatically detects abnormal behavior and automatically issues an alarm by analyzing the video in real time through a dual-stream expanded 3D convolutional network and a generalized regression network.

[0056] Furthermore, for the loading and unloading operation unit, when the terminal type is loading and unloading equipment, the recommended path information from all loading and unloading equipment to the positioning point is calculated and saved in the offline stage, and pre-stored in the path database; in the online stage, according to the task requirements or changes in the equipment position, the corresponding path information is extracted from the path database and optimized and updated to generate a navigation path, and the loading and unloading equipment performs automatic loading and unloading according to the navigation path.

[0057] Furthermore, in the monitoring module, video monitoring is specifically performed by extracting features from the video stream collected by the high-definition camera through a dual-stream expanded 3D convolutional network, including spatial features and temporal features, for identifying complex behavior patterns; the extracted spatial features and temporal features are input into a generalized regression network, and the generalized regression network predicts the probability of abnormal behavior based on the training model. When the probability of abnormal behavior exceeds a preset threshold, an alarm is automatically triggered.

[0058] Compared with the prior art, the present invention has at least one of the following technical effects:

[0059] The present invention improves the operational efficiency and safety level of the terminal through integrated modular design. The site selection evaluates the characteristics of containers through the random forest model and accurately predicts the storage location in combination with real-time vacancy data, thereby improving the space utilization efficiency and operational accuracy of the terminal. The operation module improves the operational efficiency and accuracy of manual and unmanned equipment through real-time navigation and automated path updates, ensuring the smooth progress of loading and unloading operations, thereby greatly improving the speed and safety of operations. The monitoring module performs real-time analysis of the video of the terminal through a dual-stream expansion 3D convolutional network and a generalized regression network, automatically detects abnormal behavior and responds quickly, thereby significantly improving the safety and operational efficiency of the terminal.

[0060] (1) During the site selection process, the random forest model is used to dynamically evaluate the characteristics of containers to ensure that each container is recommended to the most suitable storage location based on its actual size, weight and type, reducing unnecessary movement and rearrangement and optimizing the overall operation process of the terminal. At the same time, by combining the current vacancy situation and historical data in the terminal, the flow trend of containers can be predicted, and space resources can be further accurately adjusted and planned. Finally, the outputs of each decision tree are integrated through the weighted average method, so that the storage location of the container is more reasonable, reducing the waste of vacancy and inconvenience of loading and unloading in the terminal. The space utilization rate of the terminal is improved, the loading and unloading time is shortened, and the potential losses caused by improper site selection, such as cargo damage or transportation delays, are reduced, thereby reducing operating costs.

[0061] (2) In the operation module, it includes the container search unit and the loading and unloading operation unit. The container search unit enables the staff to quickly locate the container based on the real-time navigation and location information provided, reducing errors in human identification, improving the efficiency of container search, and also improving the work experience of the terminal staff. The loading and unloading operation unit can enable the unmanned loading and unloading equipment to automatically navigate to the destination through network optimization and re-optimization technology. At the same time, the automated loading and unloading equipment can work continuously without being restricted by the working hours of the operator, which significantly improves the efficiency of loading and unloading operations. At the same time, the intelligent and precise control of the automated equipment reduces the possibility of human operation errors and accidents. Improve safety. The automated operation of the equipment can also reduce energy consumption and maintenance costs.

[0062] (3) The monitoring module improves the safety and monitoring efficiency of the station through real-time monitoring and automatic alarm, and also enables managers to handle emergencies in the first place. The sensors on the monitoring system and equipment can collect real-time data from the station equipment, reducing the manual maintenance cost of the station equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solution of the embodiment of the present invention, the following is a brief introduction to the drawings required for describing the embodiment:

[0064] Figure 1 A flowchart of the steps of the AI-based smart dry port operation method;

[0065] Figure 2 This is a flow chart of the position selection calculation steps of Example 1. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. It should be understood that the specific implementation methods described herein are only used to explain the present invention and are not used to limit the present invention.

[0067] Example 1

[0068] like Figure 1 As shown, the present invention provides a smart land port operation method based on artificial intelligence, comprising the following steps:

[0069] Step S1, collecting and preprocessing terminal data, container information and historical storage information;

[0070] Step S2: input the current container information into the intelligent location selection module to output the recommended storage location; wherein:

[0071] Based on the data collected from step S1, a key feature set is constructed, and the intelligent location selection module randomly selects a feature subset from the key feature set for training through a random forest model; the feature importance of each feature subset in the decision tree is evaluated, and the corresponding prediction results are generated; at the same time, the expected number of inbound containers of each type is calculated as an adjustment factor by combining the remaining number of empty containers in the current terminal and the historical demand for each type of container; the prediction results of all decision trees are integrated through a weighted average method, and the weight parameters in the random forest model are adjusted according to the adjustment factors to generate a recommended storage location;

[0072] Step S3: convert the pre-processed terminal data and container information into a real-time global map, and calculate the navigation path according to the actual location and target location of the terminal type through the Dijkstra algorithm; wherein,

[0073] When the terminal type is an unmanned automated device, the distance from the loading and unloading equipment to the positioning point is calculated offline as the path information, and a corresponding path database is constructed for pre-storage; the navigation path in the basic database is updated according to online task requirements or changes in equipment location;

[0074] Step S4, real-time monitoring of the real-time monitoring data of the field station data in step S1, analyze the real-time monitoring data through a dual-stream dilated 3D convolutional network and a generalized regression network, and automatically detect and respond to abnormal behaviors to ensure safety and operational efficiency within the field station.

[0075] (1) The data collection process in step S1 includes:

[0076] High-definition cameras and various sensors including temperature, humidity, vibration and smoke sensors are deployed in the operation areas and channels of the station to obtain station data; among them,

[0077] The station data is the station map, the spatial status of the operation area and the channel, including the layout, size, capacity, environmental data and real-time monitoring data;

[0078] Container information includes the size, type, and weight of the container;

[0079] Historical storage information includes the container's entry and exit times and storage locations.

[0080] (2) Among them, Figure 2 As shown, the position selection calculation process in step S2 includes:

[0081] Step S21, obtaining basic features associated with the container, including terminal data and container information, and calculating derived features including the unit weight, expected departure time and movement frequency of the container, and forming a complete key feature set with the basic features and derived features;

[0082] Step S22, randomly select a feature subset from the key feature set for training each decision tree, for each decision tree, calculate the Gini impurity reduction of the selected feature at the split point, evaluate the importance of the feature in the location selection process, and generate a predicted storage location for each decision tree based on the importance; wherein, the calculation steps include,

[0083] Step S221, for each decision tree, randomly select a feature subset from the key feature set for node segmentation in the decision tree construction process;

[0084] Step S222: For each node of the decision tree, the Gini impurity of each feature is calculated and the optimal feature is selected for segmentation. The optimal feature is the feature with the lowest Gini impurity. The Gini impurity calculation formula is:

[0085] ,

[0086] in, is the Gini impurity of node I, J is the total number of categories in the feature subset, and P is the proportion of samples belonging to the jth category of node I; the purpose of squaring is to increase the weight of the category with a higher proportion. After squaring, the larger proportion (i.e., the majority category) contributes more to the total, thereby reducing the Gini impurity of the node.

[0087] Step S223, for each usage feature Nodes for segmentation, calculation of features The Gini impurity reduction is calculated as follows:

[0088] ;

[0089] in, is the reduction of Gini impurity, is the impurity of node I, and is the Gini impurity of the left and right child nodes after segmentation;

[0090] Step S224, accumulating features The reduction of Gini impurity;

[0091] ;

[0092] in, To include features The segmentation set of

[0093] Step S224, normalization is performed based on the accumulated Gini impurity reduction of each feature to obtain a corresponding importance score.

[0094] Step S23, combining the number of remaining container spaces in the current terminal and the historical demand for each type of container, predicting the expected number of inbound containers of each type as an adjustment factor for optimizing the weight parameters in the random forest model; wherein the calculation steps include:

[0095] Step S231, collecting the number of remaining containers in the current terminal and the historical demand for various types of containers;

[0096] Step S232, define the objective function , where Q is the number of containers entering the warehouse, and the historical number of containers entering the warehouse of each type is fitted into a random distribution D, and d is the actual demand sampled from the random distribution D;

[0097] Step S233, calculate and minimize the expected value of the objective function, and obtain the corresponding inventory quantity Q as the predicted quantity, that is:

[0098] ;

[0099] ;

[0100] Among them, I is the information set, including the remaining number of containers in the current terminal and the historical demand; D is a random distribution fitted with the demand d based on the historical demand; Q is the predicted number of containers of each type entering the warehouse, and d is the actual demand.

[0101] Step S24, the prediction results of all decision trees are integrated by weighted average method, and adjusted in combination with adjustment factors to generate a recommended storage location.

[0102] (3) The station operation process of step S3 is as follows:

[0103] Step S31: construct a real-time updated global map based on the terminal data and container data, and install positioning beacons at corresponding locations to obtain the location information of the positioning points;

[0104] Step S32: Based on the terminal type, determine the corresponding path information mode. When the terminal type is a manually operated device, the Dijkstra algorithm is used to calculate the recommended navigation path according to the actual position and target position of the manually operated device. When the terminal type is a loading and unloading device, the recommended path information from all loading and unloading devices to the positioning point is calculated and saved in the offline stage, and pre-stored in the path database. In the online stage, the corresponding path information is extracted from the path database and optimized and updated according to the task requirements or changes in the equipment position to generate a navigation path.

[0105] (4) The real-time monitoring process in step S4 is as follows:

[0106] Deploy cameras in stations to obtain real-time monitoring data;

[0107] The dual-stream expanded 3D convolutional network is used to extract features from the video stream collected by the camera, including spatial features and temporal features, to identify complex behavior patterns.

[0108] The extracted spatial and temporal features are input into the generalized regression network, which predicts the abnormal probability of behavior based on the training model. When the probability of abnormal behavior exceeds the preset threshold, an alarm is automatically triggered.

[0109] Example 2

[0110] Based on the same inventive concept, the present invention also provides an artificial intelligence-based smart land port operation system, which adopts the above-mentioned smart land port operation method, including:

[0111] The intelligent location selection module collects and preprocesses terminal data, container information, and historical storage information, and constructs a key feature set. It randomly selects a feature subset from the key feature set for training through the random forest model; evaluates the feature importance of each feature subset in the decision tree and generates the corresponding prediction results; combines the number of remaining container spaces in the current terminal and the historical demand for each type of container to calculate the expected number of inbound containers of each type as an adjustment factor; combines the prediction results of all decision trees through the weighted average method, and adjusts the weight parameters in the random forest model according to the adjustment factors to generate a recommended storage location;

[0112] The operation module includes a box-finding unit and a loading and unloading unit. When the terminal type is a manually operated device, the navigation path is calculated in the box-finding unit according to the actual position and target position of the device through the Dijkstra algorithm. When the terminal type is an unmanned automated device, the distance from the loading and unloading device to the positioning point is calculated offline as the path information, and a corresponding path database is constructed for pre-storage. The navigation path in the basic database is updated according to the online task requirements or the change of the device position, and the navigation is carried out according to the navigation path to perform automated loading and unloading;

[0113] The monitoring module conducts all-round monitoring of the site through installed high-definition cameras and various environmental sensors; video monitoring automatically detects abnormal behavior and automatically issues an alarm by analyzing the video in real time through a dual-stream expanded 3D convolutional network and a generalized regression network.

[0114] Furthermore, for the loading and unloading operation unit, when the terminal type is loading and unloading equipment, the recommended path information from all loading and unloading equipment to the positioning point is calculated and saved in the offline stage, and pre-stored in the path database; in the online stage, the corresponding path information is extracted from the path database according to task requirements or location changes, and the navigation path is optimized and updated to generate the navigation path, and the loading and unloading equipment performs automatic loading and unloading according to the navigation path.

[0115] Furthermore, in the monitoring module, video monitoring is specifically performed by extracting features from the video stream collected by the camera through a dual-stream expanded 3D convolutional network, including spatial features and temporal features, for identifying complex behavior patterns; the extracted spatial features and temporal features are input into a generalized regression network, and the generalized regression network predicts the probability of abnormal behavior based on the training model. When the probability of abnormal behavior exceeds a preset threshold, an alarm is automatically triggered.

[0116] Example 3

[0117] In order to enable those skilled in the art to have a deeper understanding of the present technical solution, the following will be combined with the third specific embodiment to elaborate on the various steps of site selection, operation process and real-time monitoring, and explain their functions and effects in practical applications.

[0118] 1. Calculation steps of the intelligent location selection process.

[0119] (1) Data collection and preprocessing

[0120] Collect spatial data of the terminal, including terminal maps, storage areas, channels and other information, and container information, including container size, type, weight, etc.;

[0121] Collect historical storage data of containers, including the entry and exit time and storage location of containers; pre-process the collected data, including cleaning missing values ​​and outliers, to ensure data quality and facilitate subsequent analysis and modeling.

[0122] (2) Feature construction

[0123] Build a set of key features from the preprocessed data, which includes features extracted directly from the data and generated derived features such as unit weight (weight divided by size), estimated departure time, movement frequency, etc.

[0124] (3) Model training and feature importance evaluation

[0125] A random forest model is used for training, which randomly selects a subset of features from the key feature set to train a decision tree. The importance of each feature subset in the decision tree is evaluated. The importance of features is evaluated by calculating the mean impurity reduction (MDI) caused by each feature in the decision tree.

[0126] That is, the model will analyze the features of each container and evaluate the importance of each feature. The importance of a feature can be obtained by calculating the average gain of each feature in each decision tree. Features with higher importance have a greater impact on the prediction results of the model, so they may be given priority when selecting the placement area.

[0127] Specifically, the importance of each feature can be measured by the average non-existence reduction index. In the decision tree, the Gini non-existence degree of each node I is ,in is the sample ratio of node I belonging to the jth class, J is the total number of classes. When the data is divided into two nodes, the reduction of the non-storage degree is For the feature The total impurity reduction when using this feature for splitting in all trees is calculated as ,in Indicates that it contains features The set of all partitions of .

[0128] Each decision tree will give a prediction result, and the model will use the weighted average calculation method to combine the results of all decision trees to get the most appropriate location selection result. The average value 1 / N (N is the number of decision trees) can be taken, or it can be adjusted according to the historical prediction accuracy. After the container is stored, it will be captured by the camera and sensor and used to update the input of the intelligent location selection model and inventory optimization algorithm. Among them, the weight adjustment is performed through the loss function, that is, in the process of optimizing the decision tree weight, the loss function is used to calculate the prediction error of each model, thereby guiding the adjustment of the weight.

[0129] The specific operation is as follows: ① First, define a loss function to measure the difference between the prediction of each decision tree and the actual data, such as mean square error (MSE) and cross entropy loss, etc. ② After the initial weight assignment, use the defined loss function to calculate the prediction loss of each tree. ③ Adjust the weights through optimization algorithms (such as gradient descent) to minimize the overall prediction loss. In this process, the algorithm will iteratively adjust the weights, and recalculate the loss of the integrated model obtained by weighted average after each iteration. The goal is to find the weight configuration that minimizes the overall loss. ④ Continue to monitor the prediction loss during operation and dynamically adjust the weights according to the change in loss. If the prediction performance of a tree decreases, resulting in an increase in the total loss, you can try to reduce the overall loss by reducing its weight. ⑤ Use the optimized weights to combine the prediction results of all decision trees through the weighted average method to calculate the final prediction output to ensure the accuracy and stability of the model output. ⑥ Combine the remaining container spaces and historical demand data in the current terminal to calculate the expected number of inbound containers as an important factor in model adjustment. The prediction results of all decision trees are combined using a weighted average method, and the weight parameters in the random forest model are further adjusted based on these adjustment factors to generate the final recommended storage location.

[0130] Therefore, through the intelligent location selection system, the storage location of containers is more reasonable, reducing the waste of space at the terminal and the inconvenience of loading and unloading. It improves the utilization rate of terminal space, shortens the loading and unloading time, and also reduces the potential losses caused by improper location selection, such as cargo damage or transportation delays, thereby reducing operating costs.

[0131] (II) Calculation steps for intelligent box search:

[0132] Install positioning beacons in the terminal to accurately locate the location information of containers, and use 3D real-scene modeling technology to draw a panoramic map of the terminal with the terminal data and container data collected by the intelligent terminal global monitoring system to display the location information of containers in real time. Use the Dijkstra shortest path algorithm to display the navigation path and the location of the target container according to the location and direction of the staff. Track and monitor the location and status of the container in real time. When an abnormality is detected in the container, such as not moving after a timeout, or the location deviates from the scheduled route, the system will issue an alarm to remind the staff to deal with it in time.

[0133] Therefore, the intelligent container search map system provides real-time navigation and location information, allowing staff to quickly locate containers, reducing errors in human identification and improving container search efficiency. It also improves the work experience of terminal staff.

[0134] (III) Implementation steps for unmanned automated loading and unloading equipment:

[0135] We selected unmanned forklifts and intelligent automated gantry cranes with autonomous driving functions, customized equipment transformation, and adjusted the lifting capacity and telescopic range of the equipment to accommodate containers of different sizes. At the same time, we installed high-definition cameras and sensors on the automated equipment to identify and capture images and data of containers, such as location, size, and potential obstacles, and combined the captured data with the control system of the automated loading and unloading equipment to ensure that the equipment can make correct operational decisions when encountering different containers in different locations.

[0136] Integrate intelligent driving technology into automated equipment, use vehicle path optimization algorithms and global maps of the site to provide accurate positioning and navigation capabilities for automated loading and unloading equipment, and realize automatic driving of the equipment. When planning vehicle paths, due to the real-time nature of demand and vehicle location, it is necessary to use online vehicle path planning algorithms to quickly respond to path planning requirements. Use online vehicle path planning algorithms to respond to real-time changes in demand and vehicle location, and quickly adjust pre-calculated paths through network optimization re-optimization technology to ensure that approximate optimal solutions are quickly provided when new demands arise. That is, when a new start-end pair appears online, use re-optimization technology to quickly update the shortest path learned offline, and obtain an approximate optimal solution to the online path planning problem to meet the requirements of fast calculation response. Collect on-site operation data of automated equipment and analyze the operation of the equipment. According to the results of data analysis, improve and upgrade the equipment to improve equipment performance and loading and unloading efficiency.

[0137] Therefore, through intelligent automated loading and unloading equipment, the automated loading and unloading equipment can work continuously without being restricted by the working hours of the operators, which significantly improves the efficiency of loading and unloading operations. At the same time, the intelligent and precise control of the automated equipment reduces the possibility of human operating errors and accidents. It improves safety. The automated operation of the equipment can also reduce energy consumption and maintenance costs.

[0138] (IV) Specific steps and processes for real-time monitoring:

[0139] To install high-definition cameras at a site, you first need to determine the monitoring area and key monitoring targets, including areas where people move around, equipment is running, and site operation areas. Analyze the safety risks and potential problems of the site, develop monitoring strategies, ensure the safety of the operation process, and prevent equipment from colliding with people or objects. High-definition cameras need to ensure that they cover the entire site and achieve 360-degree monitoring without blind spots. Deploy various sensors according to monitoring needs, such as temperature sensors, humidity sensors, smoke sensors, vibration sensors, etc.

[0140] High-definition cameras mainly collect real-time images and video streams, while sensors mainly collect parameter data, operation data, activity data, and environmental data of stations, containers, equipment, and staff. In order to reduce the pressure on the system to analyze data, edge computing gateways can be deployed near cameras and sensors to process the data preliminarily before being collected by the IoT platform and uniformly transmitted to the system.

[0141] Intelligent image recognition technology is used to analyze the videos and images collected by the camera to identify normal operating behaviors or abnormal events. The dual-stream expansion 3D convolutional network is used to extract features from the video. The dual-stream expansion 3D convolutional network combines spatial and temporal streams to simultaneously capture the spatial details and temporal dynamics of the video frame. This enables it to directly learn the dynamic changes between frames when processing time series video data without pre-calculating the optical flow, thereby effectively improving the accuracy and efficiency of action recognition and video content analysis. Then, a generalized regression network is used for training to determine the probability of abnormal behavior, such as a person falling or illegal intrusion. When the probability of abnormal behavior exceeds a given threshold of 75%, an abnormal warning is issued. At the same time, the data collected by the sensor, such as temperature and humidity, is processed and analyzed to detect whether the data is within the preset normal range and to promptly discover potential safety hazards. A monitoring center and an alarm system are set up to visualize the data collected by the camera and sensors in the monitoring center, so that operators can monitor the site in real time. An alarm system is set up in the monitoring center to issue an alarm in a timely manner when abnormal behavior is detected or environmental data exceeds the safety range. At the same time, the alarm system also has real-time response and automatic notification functions to ensure that abnormal situations are handled in a timely manner.

[0142] The specific identification process is as follows:

[0143] ①Use high-definition cameras installed in key areas to continuously collect real-time video streams within the station;

[0144] ② Extract key frames or continuous frame sequences from the video stream as the basis for analysis, while removing image noise and adjusting brightness and contrast to optimize image quality;

[0145] ③ Feature extraction through a dual-stream expanded 3D convolutional network, which specifically includes processing individual frames in the spatial stream to capture static visual features of the scene, such as shape and texture, analyzing changes in consecutive frames in the temporal stream to extract motion features, such as changes in direction and speed, and finally expanding the scope of the convolution kernel to capture dynamic changes in longer time series without sacrificing temporal resolution;

[0146] ④Behavior recognition and anomaly detection are performed through generalized regression networks, specifically

[0147] The network structure is set up to include an input layer, a pattern layer, a summation layer, and an output layer. Specifically, the input layer receives the feature vector obtained in the feature extraction stage (such as the output of a two-stream expanded 3D convolutional network); each training sample has a neuron in the pattern layer, which is used to store and process the relationship between the input data and the target output. The weighted sum of the pattern layer output is calculated through the summation layer, which is used to generate the average prediction output and the normalization factor, respectively, and finally produce a prediction result, which is mainly the probability of the behavior category.

[0148] The training dataset is processed, including instances of normal behavior and various abnormal behaviors, and the feature vector may include spatial and dynamic features extracted from video frames.

[0149] Use the labeled training data to train the GRNN, with the feature vector as input and the correct behavior label (normal or various abnormalities) as output. At the same time, adjust the kernel width (smoothing parameter) to adjust the sensitivity and generalization ability of the model. Finally, use the validation set to evaluate the model performance, adjust the network parameters, optimize the kernel width and training algorithm to reduce overfitting and improve prediction accuracy.

[0150] Deploy the trained GRNN in a real-time monitoring environment. The network receives the feature vectors from the video analysis module in real time, quickly outputs the probability of behavior classification, and determines whether there is abnormal behavior. When the abnormal probability of a behavior predicted by GRNN exceeds the preset threshold, the system automatically triggers an alarm, which can be displayed on the visualization platform and sent to relevant security personnel.

[0151] Therefore, the intelligent operation monitoring system improves the safety and monitoring efficiency of the station through real-time monitoring and automatic alarm, and also allows managers to deal with emergencies in the first place. The sensors on the monitoring system and equipment can collect real-time data of the station equipment, reducing the manual maintenance cost of the station equipment.

[0152] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications to the technical solution of the present invention by using the methods and technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the present invention.

Claims

1. A smart land port operation method based on artificial intelligence, characterized in that: The following steps are involved: Step S1, collecting and preprocessing terminal data, container information and historical storage information; wherein the terminal data is the terminal map, the spatial status of the operation area and the channel, including the layout, size, capacity, environmental data and real-time monitoring data; the container information includes the size, type and weight of the container; the historical storage information includes the time of entering and leaving the terminal and the storage location of the container; Step S2: input the current container information into the intelligent location selection module to output a recommended storage location; wherein: A key feature set is constructed based on the data collected from step S1, wherein the key feature set includes the terminal data, the container information, the unit weight of the container, the expected departure time and the movement frequency. The intelligent location selection module randomly selects a feature subset from the key feature set for training through a random forest model; the feature importance of each feature subset in the decision tree is evaluated, and a corresponding prediction result is generated; at the same time, the expected number of inbound containers of each type is calculated as an adjustment factor in combination with the remaining number of empty containers in the current terminal and the historical demand number of each type of container; the prediction results of all the decision trees are integrated through a weighted average method, and the weight parameters in the random forest model are adjusted according to the adjustment factor to generate the recommended storage location; The expected inbound quantity of each type of container is calculated by combining the number of remaining containers in the current terminal and the historical demand for each type of container. The adjustment factors include: Collect the remaining number of empty containers in the current terminal and the historical demand for each type of container; define the objective function ,in, For the expected inbound quantity, the historical inbound quantity data of various types of containers are fitted into a random distribution , For the random distribution The actual demand for sampling; Calculate the expected value of the objective function and minimize it to obtain the corresponding expected number of inbound ,Right now: ; ; in, is an information set, including the remaining number of containers in the current terminal and the historical demand; Based on the historical demand for each type of container, the actual demand A random distribution is fitted, is the actual demand; Step S3, converting the pre-processed terminal data and the container information into a real-time global map, and calculating the navigation path according to the actual position and target position of the terminal type through the Dijkstra algorithm; wherein, When the terminal type is an unmanned automated device, the distance from the unmanned automated device to the positioning point is calculated offline as path information, and a corresponding path database is constructed for pre-storage; the navigation path in the basic database is updated according to online task requirements or device position changes; Step S4, real-time monitoring of the station data in step S1, analyzing the real-time monitoring data through a dual-stream dilated 3D convolutional network and a generalized regression network, and automatically detecting and responding to abnormal behaviors to ensure safety and operational efficiency within the station.

2. The smart land port operation method according to claim 1 is characterized in that: Step S1, specifically, High-definition cameras and various sensors including temperature, humidity, vibration and smoke sensors are deployed in the working areas and channels of the station to obtain the station data.

3. The smart land port operation method according to claim 2 is characterized in that: Step S2 is specifically, Step S21, acquiring basic features associated with the container, including the terminal data and the container information, and calculating derived features including the unit weight of the container, the estimated departure time and the movement frequency, and forming a complete key feature set with the basic features and the derived features; Step S22, randomly selecting the feature subset from the key feature set for training each decision tree, for each decision tree, calculating the Gini impurity reduction of the selected feature at the split point, evaluating the importance of the feature in the location selection process, and generating a predicted storage location for each decision tree based on the importance; Step S23, combining the number of remaining container spaces in the current terminal and the historical demand for each type of container, predicting the expected inbound number of each type of container as the adjustment factor for optimizing the weight parameter in the random forest model; Step S24, combining the prediction results of all the decision trees by the weighted average method, and adjusting in combination with the adjustment factors to generate the recommended storage location.

4. The smart land port operation method according to claim 3 is characterized in that: In step S22, for each decision tree, the Gini impurity reduction of the selected feature is calculated at the split point to evaluate the importance of the feature in the position selection process, specifically, Step S221, for constructing each decision tree, randomly selecting a feature subset from the key feature set for node segmentation in the process of constructing the decision tree; Step S222: For each node of the decision tree, calculate the Gini impurity of each feature and select the optimal feature for segmentation, wherein the optimal feature is the feature with the lowest Gini impurity, and the Gini impurity calculation formula is: ; in, For Node The Gini impurity of is the total number of categories in the feature subset, For the node Belong to The sample ratio of the class; Step S223, for each use feature The node for segmentation is used to calculate the feature The Gini impurity reduction is calculated as follows: ; in, is the Gini impurity reduction, For the node The impurity of and For the node Gini impurity of left and right child nodes after segmentation; Step S224, accumulating the features The Gini impurity reduction of ; in, To include the features The segmentation set of Step S224, performing normalization processing based on the accumulated Gini impurity reduction of each feature to obtain a corresponding importance score.

5. The smart land port operation method according to claim 2, characterized in that: The step S3 is specifically: Step S31: construct the global map updated in real time based on the terminal data and container data, and install positioning beacons at corresponding positions to obtain the position information of the positioning points; Step S32: determining a corresponding path information mode based on the terminal type, and when the terminal type is a manually operated device, calculating the recommended navigation path according to the actual position and the target position of the manually operated device by using the Dijkstra algorithm; When the terminal type is the unmanned automated equipment, the recommended path information of all the unmanned automated equipment to the positioning point is calculated and saved in the offline stage, and pre-stored in the path database; The online stage is to extract the corresponding path information from the path database and optimize and update it according to the task requirements or the change of the device position to generate the navigation path.

6. The smart land port operation method according to claim 5 is characterized in that: The step S4 is specifically, Deploy the high-definition camera in the station to obtain the real-time monitoring data; Extracting features of the video stream collected by the high-definition camera through the dual-stream dilated 3D convolutional network, including spatial features and temporal features, for identifying complex behavior patterns; The extracted spatial features and the temporal features are input into the generalized regression network, and the generalized regression network predicts the probability of the abnormal behavior according to the training model. When the probability of the abnormal behavior exceeds a preset threshold, an alarm is automatically triggered.

7. An artificial intelligence-based smart land port operation system, using the smart land port operation method according to any one of claims 1 to 6, characterized in that: include, The intelligent location selection module collects and preprocesses terminal data, container information and historical storage information, and constructs a key feature set, and randomly selects a feature subset from the key feature set for training through a random forest model; evaluates the feature importance of each feature subset in the decision tree, and generates corresponding prediction results; combines the number of remaining container spaces in the current terminal and the historical demand for each type of container, calculates the expected number of inbound containers of each type as an adjustment factor; synthesizes the prediction results of all the decision trees through a weighted average method, and adjusts the weight parameters in the random forest model according to the adjustment factors to generate the recommended storage location; The expected inbound quantity of each type of container is calculated by combining the number of remaining empty spaces for containers in the current terminal and the historical demand for each type of container as an adjustment factor: Collect the remaining number of empty containers in the current terminal and the historical demand number of each type of containers; Define the objective function ,in, For the expected inbound quantity, the historical inbound quantity data of various types of containers are fitted into a random distribution , For the random distribution The actual demand for sampling; Calculate the expected value of the objective function and minimize it to obtain the corresponding expected number of inbound ,Right now: ; ; in, is an information set, including the remaining number of containers in the current terminal and the historical demand; Based on the historical demand for each type of container, the actual demand A random distribution is fitted, is the actual demand; The operation module includes a box-finding unit and a loading and unloading unit, wherein, when the terminal type is a manually operated device, the navigation path is calculated in the box-finding unit according to the actual position and target position of the device by using the Dijkstra algorithm, and when the terminal type is an unmanned automated device, the distance from the unmanned automated device to the positioning point is calculated offline as path information, and a corresponding path database is constructed for pre-storage, the navigation path in the basic database is updated according to online task requirements or changes in the device position, and automated loading and unloading is performed according to the navigation path navigation; The monitoring module conducts all-round monitoring of the site through installed high-definition cameras and various environmental sensors; among them, video monitoring automatically detects abnormal behavior and automatically issues an alarm by analyzing the video in real time through a dual-stream expanded 3D convolutional network and a generalized regression network.

8. The smart land port operation system according to claim 7, characterized in that: The loading and unloading operation unit, when the terminal type is the unmanned automated equipment, calculates and saves the recommended path information of all the unmanned automated equipment to the positioning point in an offline stage, and pre-stores it in the path database; The online stage is to extract the corresponding path information from the path database and optimize and update it to generate the navigation path according to the task requirements or the change of the equipment position. The unmanned automated equipment performs automatic loading and unloading according to the navigation path.

9. The smart land port operation system according to claim 8, characterized in that: In the monitoring module, video monitoring specifically includes: extracting features of the video stream collected by the high-definition camera through the dual-stream expanded 3D convolutional network, including spatial features and temporal features, for identifying complex behavior patterns; inputting the extracted spatial features and the temporal features into the generalized regression network, and the generalized regression network predicts the probability of the abnormal behavior according to the training model, and automatically triggers the alarm when the probability of the abnormal behavior exceeds a preset threshold.