An AI-based real-time logistics supply chain management system

Abstract

This invention belongs to the field of logistics supply chain management technology and discloses an artificial intelligence-based real-time logistics supply chain management system. It includes a mobile intelligent hardware terminal with a built-in YOLOv11 image recognition model, a self-learning module, and a navigation memory module, serving as the core data acquisition, analysis, and execution terminal; a cloud management platform for data storage, global management, and model optimization; and a mobile interactive terminal for user operation, information viewing, and early warning reception. This invention, utilizing the above system and relying on mobile intelligent hardware, combined with the YOLOv11 image recognition model and self-learning algorithm, achieves automatic quantity counting, real-time status monitoring, and early risk prediction throughout the entire process of loading, transportation, and unloading. It features convenient operation, flexible installation, information traceability, and reliable network security, fundamentally replacing core manual operations, solving existing technical pain points, and improving the intelligence level of logistics supply chain management.

Description

Technical Field

[0001] This invention relates to the field of logistics supply chain management technology, and in particular to a real-time logistics supply chain management system based on artificial intelligence. Background Technology

[0002] In the current less-than-truckload (LTL) freight transportation process, the counting, status monitoring, and safety management of goods mainly rely on manual operation by truck drivers. This results in numerous problems, including low efficiency, high error rates, high labor intensity, and significant safety hazards, as detailed below: Loading: Drivers must accompany the vehicle throughout the loading process, manually counting and checking the condition of the goods. This not only consumes a significant amount of the driver's time and energy but is also prone to human error, leading to issues such as missed or incorrect loading, or failure to detect damaged goods. This can result in a chain reaction of disputes and complicated claims. Transportation: Drivers must be highly vigilant and frequently patrol the interior of the vehicle to prevent loss, theft, or damage from tipping or pressure. Prolonged patrols exacerbate driver fatigue, increase driving safety risks, and cannot provide 24 / 7 monitoring, resulting in delayed detection of emergencies. Unloading: Similar to loading, drivers must manually count and verify the quantity and items of the goods, checking for damage or loss during transportation. This again suffers from the errors and inefficiencies of manual operation, and manually recorded information on the condition and quantity of goods is difficult to retain long-term, making subsequent inquiries and traceability inconvenient. Limitations of existing auxiliary equipment: Most logistics monitoring equipment on the market is currently fixed-installation type, which has poor flexibility and cannot be adapted to different vehicle models and different cargo stacking scenarios; and it lacks intelligent analysis capabilities, and can only realize simple video recording and positioning functions. It cannot automatically count the quantity, predict the risks of cargo tipping or being crushed, etc., and relies on manual interpretation of data, which fails to fundamentally solve the pain points of manual operation.

[0003] In view of the shortcomings of the existing technologies, there is an urgent need for a management system that is adaptable to the road less-than-truckload (LTL) freight transportation scenario, is mobile, highly intelligent, and can automatically complete cargo counting, status monitoring, risk prediction, and information retention and convenient query. This system would solve the pain points of manual operation, improve the efficiency, accuracy and security of logistics supply chain management, reduce the labor intensity of drivers, and reduce cargo damage disputes. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based real-time logistics supply chain management system. Relying on mobile intelligent hardware and combining the YOLOv11 image recognition model and autonomous learning algorithm, it can automatically count the quantity of goods, monitor their status in real time, and predict risks in advance throughout the entire process of loading, transportation, and unloading. It also features convenient operation, flexible installation, traceable information, and reliable network security. It fundamentally replaces core manual operations, solves the pain points of existing technologies, and improves the level of intelligence in logistics supply chain management.

[0005] To achieve the above objectives, the present invention provides an artificial intelligence-based real-time logistics supply chain management system, including a mobile intelligent hardware terminal, a cloud management platform, and a mobile interactive terminal, which establish a two-way data transmission connection through a wireless communication module. The mobile smart hardware terminal, with built-in YOLOv11 image recognition model, self-learning module, and navigation memory module, is the core data acquisition, analysis and execution terminal. A cloud-based management platform is used for data storage, global management, and model optimization. Mobile interactive terminal, used for user operation, information viewing and alarm reception.

[0006] Preferably, the mobile intelligent hardware terminal includes: an image acquisition module, a core processing module, an autonomous learning and risk prediction module, a navigation memory module, an early warning and communication module, a storage module, a power supply module, and an installation and fixing module.

[0007] Preferably, the image acquisition module consists of a high-definition wide-angle camera and an infrared camera. The high-definition wide-angle camera is used for acquiring images of goods and the environment of the carriage in well-lit environments, with a resolution of not less than 1080P and a frame rate of not less than 25fps. The infrared camera is used for image acquisition in low-light environments, achieving clear imaging in no-light conditions and ensuring uninterrupted acquisition 24 hours a day. The camera can be rotated 360° to cover the entire interior of the carriage and the loading and unloading area, avoiding blind spots.

[0008] Preferably, the core processing module adopts an embedded processor with a built-in YOLOv11 image recognition model, supporting computational tasks such as image preprocessing, cargo quantity recognition, cargo status detection, and abnormal identification of the carriage environment. The YOLOv11 image recognition model uses an end-to-end computation approach. The input is a preprocessed image, and the output is the cargo target box, category probability, and confidence score. The implementation process is as follows: (1) Feature extraction: CSPDarknet is used as the backbone network to extract multi-scale features from the input image. Low-level features, mid-level features and high-level features in the image are extracted through convolutional layers, pooling layers and residual connections. The convolutional layers use 3×3 convolutional kernels and 1×1 convolutional kernels, and the pooling layers use max pooling with a stride of 2 to ensure the efficiency and accuracy of feature extraction. (2) Feature fusion: PANet is used for feature fusion to fuse feature maps of different scales; during the fusion process, a combination of upsampling and downsampling is used; (3) Target detection and classification: A multi-scale detection head is used to detect the fused feature map. Each detection head corresponds to a target at a different scale. Each detection head outputs the target bounding box coordinates (x, y, w, h), class probability, and confidence score. Among them, x and y are the coordinates of the center point of the target bounding box; w and h are the width and height of the target bounding box. (4) Confidence screening and non-maximum suppression: Confidence screening is performed on the target boxes output by the detection head, and target boxes with confidence higher than the preset threshold are retained; non-maximum suppression algorithm is used to calculate the cross-union ratio between target boxes. Target boxes with cross-union ratio greater than the preset threshold are retained as the target boxes with the highest confidence, and duplicate target boxes are deleted to obtain the final recognition result.

[0009] Preferably, based on the recognition results of the YOLOv11 model, the following calculation method is used to achieve accurate counting of the quantity of goods: 1) Single frame counting: For each preprocessed image frame, all cargo targets are identified using the YOLOv11 model, and the number of target boxes is counted as the number of cargo in a single frame; if cargo is occluded, and the occlusion area is ≤30%, the features of the occluded cargo are restored using a feature completion algorithm to ensure accurate counting; if the occlusion area is >30%, it is marked as suspected occlusion, and supplementary identification is performed by combining adjacent frame images. 2) Multi-frame fusion counting: A sliding window fusion algorithm is used to select the single-frame counting results of consecutive frame images, calculate the average value, and use it as the quantity of goods at the current moment; if the counting result of a certain frame deviates from the average value by more than 5%, it is regarded as an abnormal frame, the counting result of that frame is discarded, and the average value is recalculated to ensure the stability and accuracy of the counting. 3) Loading and unloading count calibration: During the loading process, as goods are continuously loaded, the count is accumulated in real time. For each item loaded, the count is incremented by 1, and the result is compared with the total order quantity for calibration. During the unloading process, as goods are continuously unloaded, the count is decremented in real time. For each item unloaded, the count is decremented by 1, and the result is compared with the quantity at the time of loading for calibration. If a count deviation occurs, an alert is immediately triggered to remind the user to check. ; The error in cargo counting shall not exceed 1%.

[0010] The preferred self-learning and risk prediction module is electrically connected to the core processing module. It has a built-in reinforcement learning algorithm and risk prediction model. By continuously learning from historical image data, cargo stacking data, and cargo damage case data, it can autonomously optimize the recognition accuracy and risk prediction accuracy. Based on parameters such as cargo stacking height, angle, spacing, and weight distribution, it can predict the risk of cargo being damaged by pressure or tipping over, and issue early warnings. Based on the self-learning and risk prediction module and the reinforcement learning algorithm, combined with the physical parameters and stacking parameters of the goods, the risk prediction of goods tipping and compression damage is realized as follows: (1)Parameter extraction: From the images recognized by the YOLOv11 model, extract the stacking parameters of the goods, including the stacking height h, stacking angle θ, goods spacing d, and the coordinates of the center of gravity of the goods (x0, y0); the preset physical parameters of the goods, including the weight m, volume V, compressive strength σ, and friction coefficient μ; (2)Calculation of tipping risk: The calculation formula for the tipping risk probability P1 is: ; where α and β are weight coefficients obtained through training with historical cargo damage data; h is the stacking height, d is the goods spacing, μ is the friction coefficient, and θ is the stacking angle; when P1 > 30%, it is determined that there is a tipping risk and an alarm is triggered; (3)Calculation of compression damage risk: The calculation formula for the compression damage risk probability P2 is: ; where γ and δ are weight coefficients; F is the pressure exerted on the goods; σ is the compressive strength of the goods; t is the compression time; T is the maximum allowable compression time of the goods; when P2 > 25%, it is determined that there is a compression damage risk and an alarm is triggered; (4)Risk level classification: According to the values of P1 and P2, the risk level is divided into three levels: Low risk: P1 ≤ 30% and P2 ≤ 25%; Medium risk: 30% < P1 ≤ 50% or 25% < P2 ≤ 40%; High risk: P1 > 50% or P2 > 40%; Different risk levels correspond to different warning methods: Low risk only message push; Medium risk audible and visual alarm and message push; High risk audible and visual alarm, message push and emergency call.

[0011] Preferably, the navigation memory module: Built-in GPS or Beidou positioning module and path memory algorithm, real-time acquisition of vehicle position information, memory of transportation routes, loading and unloading locations; Combining historical route data, providing the driver with the optimal transportation route suggestion; At the same time, recording the movement trajectory of the intelligent hardware terminal for easy tracing of equipment usage; The navigation memory module uses the Dijkstra algorithm, combined with real-time traffic data and historical route data, to realize the calculation of the optimal transportation route, and the realization process is as follows: (1)Path node modeling: Regarding the transportation starting point, ending point, passing points, and traffic intersections as path nodes, constructing a road network graph G=(V,E), where V is the set of nodes and E is the set of paths between nodes; (2) Path weight calculation: A weight w is assigned to each path E. The weight w consists of the path length L, traffic congestion level C, road risk level R, and transportation cost C0, as shown below: ; in, , , , The weighting coefficients are obtained through training with historical route data; the traffic congestion level C is determined by real-time traffic data obtained from the cloud management platform, 0≤C≤1, C=0 indicates smooth traffic and C=1 indicates severe congestion; the road risk level R is determined based on road type and historical accident rate, 0≤R≤1, R=0 indicates low risk and R=1 indicates high risk. (3) Optimal route solution: Dijkstra's algorithm is used to find the path with the minimum weight w from the starting point to the ending point in the road network graph G, which is the optimal transportation route.

[0012] Preferably, the early warning and communication module includes an audible and visual early warning unit, a wireless communication unit, and an emergency call unit. The audible and visual early warning unit issues an audible and visual early warning signal when it detects an abnormal situation. The wireless communication unit is used to establish a data transmission connection with the cloud management platform and the mobile interactive terminal to realize the real-time transmission of image data, recognition results, early warning information, and location information. When a serious abnormality occurs, the emergency call unit automatically dials the preset driver's phone number and logistics management personnel's phone number to ensure that the abnormal situation is handled in a timely manner. Storage module: Employs solid-state storage chips for local storage of collected image data, video recordings, recognition results, and early warning log information, with a storage time of no less than 30 days; it also supports encrypted data storage to prevent data leakage and tampering, ensuring information security; Power supply module: It adopts a rechargeable lithium battery with a capacity of no less than 10000mAh and supports fast charging; it also supports vehicle charging and is compatible with 12V / 24V truck power supply to enable real-time charging during driving and ensure uninterrupted operation of the equipment; it has a built-in power monitoring unit to monitor the power status in real time and issue a low power warning when the power is below 20% to remind the user to charge. Installation and fixing module: It adopts a combination design of magnetic attraction and buckle, which eliminates the need for drilling and welding, and can be quickly fixed to the inner wall, top of the carriage or loading and unloading area; at the same time, it is easy to disassemble and can be flexibly transferred between different vehicles.

[0013] Preferably, the cloud management platform adopts a distributed architecture, is deployed on a cloud server, and establishes a two-way data transmission connection with mobile smart hardware terminals and mobile interactive terminals; (1) Data storage and management: Receive and store all data transmitted by the mobile smart hardware terminal, including images, videos, recognition results, early warning logs and location information. Use distributed database storage to realize data classification management, archive data in multiple dimensions, and facilitate user query and traceability. (2) Model optimization and update: The built-in model training server collects historical recognition data, misidentification cases and cargo damage data transmitted by mobile smart hardware terminals, and regularly iterates and trains the YOLOv11 image recognition model and the self-learning model to optimize the model recognition accuracy and risk prediction accuracy. After training, the model update package is pushed to all mobile smart hardware terminals through the wireless communication module to realize online model upgrade without manual intervention. (3) Global monitoring and management: Supports centralized monitoring of multiple terminals and multiple vehicles. Managers can view the real-time transportation status, cargo status and working status of smart hardware terminals of all vehicles through the cloud management platform; set early warning thresholds and access control; receive early warning information from all terminals and remotely issue instructions. (4) Data statistics and analysis: The built-in data statistics and analysis module performs statistical analysis on data such as cargo loading and unloading efficiency, cargo damage rate, abnormal situation occurrence rate and smart hardware terminal usage rate, generates visual reports, and provides data support for logistics supply chain management decisions; at the same time, it identifies the weak points in the logistics transportation process and provides optimization suggestions. (5) Network security protection: Built-in firewall and data encryption module, using SSL / TLS encryption protocol to encrypt the data transmission and storage process to prevent data leakage, tampering and attack; supports device access authentication, only authorized mobile smart hardware terminals and mobile interactive terminals can access the cloud management platform to ensure system security.

[0014] Preferred mobile interactive terminal: compatible with mobile devices such as mobile phones and tablets, supports iOS and Android systems, and establishes connection with cloud management platform and mobile smart hardware terminal through APP; (1) Operation control: Drivers and managers can remotely control the mobile smart hardware terminal through the APP, including starting and stopping the device, adjusting the camera angle, starting video recording, and querying historical records; the APP interface is simple and easy to operate, and adopts an icon-based design, so no professional skills are required to operate it. (2) Information viewing: Real-time viewing of cargo quantity identification results, cargo status, vehicle environment video, vehicle positioning information, warning logs, and historical records; supports multi-dimensional query of historical data by time range, order number, and cargo type, and query results can be exported; (3) Warning reception: Real-time reception of warning information pushed by mobile smart hardware terminals and cloud management platform. The warning information includes the type of abnormality, time of occurrence, vehicle location and cargo information, so that users can know and deal with abnormal situations in a timely manner; support the marking of warning information to avoid omissions; (4) Message interaction: Supports message interaction between drivers and managers, sending text, pictures and voice messages through the APP to facilitate communication and handling of abnormal situations; at the same time, it receives notifications and instructions issued by the cloud management platform.

[0015] Therefore, the present invention employs the above-mentioned real-time logistics supply chain management system based on artificial intelligence, and the beneficial effects are as follows: (1) This system combines a mobile intelligent hardware terminal with the YOLOv11 image recognition model to realize automatic counting of goods, item verification and status detection in the loading and unloading process, replacing the driver's manual operation throughout the process, freeing the driver from tedious counting and monitoring work, reducing the driver's labor intensity, solving the pain points of manual operation, and improving management efficiency and accuracy.

[0016] (2) The present invention has a built-in autonomous learning and risk prediction module, which, combined with reinforcement learning algorithm, can predict risks such as cargo tipping and damage under pressure based on cargo stacking parameters and physical characteristics, and issue early warnings in advance, thereby realizing early risk prediction, proactively preventing cargo damage risks, and significantly reducing cargo damage rate.

[0017] (3) The mobile smart hardware terminal of the present invention adopts magnetic and snap-fit ​​combination installation, which does not require drilling or welding, and is easy to install and disassemble. It can be flexibly transferred between different vehicle models, with strong adaptability and high practicality. The operating system is simple and efficient, and the mobile APP interface is concise and icon-based, which can be operated without professional skills. It also supports vehicle charging and fast charging, has strong battery life, and ensures that the device works 24 hours a day, adapting to the complex scenario of less-than-truckload (LTL) freight transportation on highways.

[0018] (4) This invention retains images, videos, recognition results, early warning logs, and location information throughout the entire loading, transportation, and unloading process for a period of no less than 30 days. It supports dual backup of local and cloud storage, and the data is encrypted to ensure information security and prevent loss. Users can query historical data by time range, order number, and cargo type through mobile APP and cloud management platform, which facilitates the handling of cargo damage disputes and liability determination, and improves the standardization and traceability of logistics supply chain management.

[0019] (5) The present invention adopts an architecture that combines edge computing with cloud collaboration to realize real-time data collection, local data analysis and autonomous decision-making, reduce network dependence and reduce data transmission delay; the cloud management platform realizes global monitoring, model iterative optimization and data statistical analysis, supports centralized management of multiple terminals and multiple vehicles, and expands functions such as temperature and humidity monitoring, open flame detection and anti-theft alarm according to the needs of logistics enterprises to adapt to the needs of different logistics scenarios.

[0020] (6) The present invention has a built-in firewall and data encryption module, and adopts the SSL / TLS encryption protocol to encrypt the data transmission and storage process throughout; it supports device access authentication, and only authorized devices can access the system to prevent data leakage, tampering and attack; it also has functions such as low power warning and device failure warning to ensure stable and reliable operation of the system.

[0021] 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

[0022] Figure 1 This is an architecture diagram of a real-time logistics supply chain management system based on artificial intelligence. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Example like Figure 1 As shown, this invention discloses an artificial intelligence-based real-time logistics supply chain management system, comprising a mobile intelligent hardware terminal, a cloud management platform, and a mobile interactive terminal. The three are connected via a wireless communication module to establish a bidirectional data transmission connection. The mobile intelligent hardware terminal has a built-in YOLOv11 image recognition model, an autonomous learning module, and a navigation memory module, and is the core data acquisition, analysis, and execution terminal of the system. The cloud management platform is used for data storage, global management, and model optimization. The mobile interactive terminal is used for user operation, information viewing, and early warning reception.

[0025] The system proposed in this invention adopts an architecture that combines edge computing with cloud collaboration. The mobile intelligent hardware terminal serves as an edge node, undertaking real-time data collection, local data analysis, and autonomous decision-making functions. The cloud management platform undertakes global data storage, model iteration optimization, and unified management functions. The mobile interactive terminal undertakes human-computer interaction functions. The three work together to achieve real-time management of the entire logistics supply chain.

[0026] 1. Mobile intelligent hardware terminal: As the core execution unit of the system, it adopts a mobile design to adapt to different vehicle models and different cargo stacking scenarios, making it easy to install, disassemble and move.

[0027] The mobile intelligent hardware terminal includes: an image acquisition module, a core processing module, an autonomous learning and risk prediction module, a navigation memory module, an early warning and communication module, a storage module, a power supply module, and an installation and fixing module.

[0028] 1. Image Acquisition Module: Composed of a high-definition wide-angle camera and an infrared camera; the high-definition wide-angle camera is used for acquiring images of goods and the environment of the carriage during the day or in well-lit environments, with a resolution of no less than 1080P and a frame rate of no less than 25fps; the infrared camera is used for image acquisition at night or in low-light environments, enabling clear imaging in no-light conditions and ensuring uninterrupted 24-hour acquisition; the camera can be rotated and adjusted 360°, covering the entire interior of the carriage and the loading and unloading area, avoiding blind spots.

[0029] 2. Core Processing Module: It adopts an embedded processor (such as NVIDIA Jetson Xavier NX) with a built-in YOLOv11 image recognition model, which undertakes core computing tasks such as image preprocessing, cargo quantity recognition, cargo status detection, and abnormal identification of the carriage environment; the processor's computing speed is no less than 200 TOPS, ensuring real-time processing of image data and recognition latency of no more than 500ms.

[0030] The YOLOv11 image recognition model employs an end-to-end computation approach. The input is a pre-processed 640×640 resolution image, and the output includes the object bounding box, category probability, and confidence score. The specific implementation process is as follows: (1) Feature extraction: CSPDarknet is used as the backbone network to extract multi-scale features from the input image. Low-level features (such as edges and textures), mid-level features (such as cargo outlines), and high-level features (such as cargo category features) in the image are extracted through convolutional layers, pooling layers, and residual connections. The convolutional layers use 3×3 convolutional kernels and 1×1 convolutional kernels, and the pooling layers use max pooling with a stride of 2 to ensure the efficiency and accuracy of feature extraction.

[0031] (2) Feature fusion: PANet (Path Aggregation Network) is used to fuse feature maps of different scales to improve the recognition accuracy of small-sized goods and occluded goods. During the fusion process, upsampling and downsampling are combined to ensure that the size of the feature maps is matched. The fused feature maps are used for subsequent target detection.

[0032] (3) Target detection and classification: Multi-scale detection heads are used to detect the fused feature map. Each detection head corresponds to a target of a different scale. Small-scale detection head detects small goods, medium-scale detection head detects medium-sized goods, and large-scale detection head detects large goods. Each detection head outputs the target box coordinates (x, y, w, h), class probability, and confidence score. Among them, x and y are the coordinates of the center point of the target box; w and h are the width and height of the target box.

[0033] (4) Confidence screening and non-maximum suppression: Confidence screening is performed on the target boxes output by the detection head, and target boxes with confidence higher than the preset threshold are retained; the non-maximum suppression (NMS) algorithm is used to calculate the intersection-union ratio (IoU) between target boxes. Target boxes with IoU greater than the preset threshold are retained as the target boxes with the highest confidence, and duplicate target boxes are deleted to obtain the final recognition result.

[0034] Based on the recognition results of the YOLOv11 model, the following calculation method is used to achieve accurate counting of the quantity of goods: 1) Single frame counting: For each preprocessed image frame, all cargo targets are identified using the YOLOv11 model, and the number of target boxes is counted as the number of cargo in a single frame; if there is cargo occlusion (occlusion area ≤ 30%), the features of the occluded cargo are restored using a feature completion algorithm to ensure accurate counting; if the occlusion area > 30%, it is marked as suspected occlusion, and supplementary identification is performed by combining adjacent frame images.

[0035] 2) Multi-frame fusion counting: A sliding window fusion algorithm is used to select the single-frame counting results of 10 consecutive frames of images, calculate the average value, and use it as the quantity of goods at the current moment; if the counting result of a certain frame deviates from the average value by more than 5%, it is regarded as an abnormal frame, the counting result of that frame is discarded, and the average value is recalculated to ensure the stability and accuracy of counting.

[0036] 3) Loading and unloading count calibration: During the loading process, as goods are continuously loaded, the count is accumulated in real time. For each item loaded, the count is incremented by 1, and the result is compared with the total quantity of the order for calibration. During the unloading process, as goods are continuously unloaded, the count is decremented in real time. For each item unloaded, the count is decremented by 1, and the result is compared with the quantity at the time of loading for calibration. If a count deviation occurs, an early warning is immediately triggered to remind the user to check.

[0037] ; The error in cargo counting in this system does not exceed 1%.

[0038] 3. Independent Learning and Risk Prediction Module: Electrically connected to the core processing module, it incorporates a reinforcement learning algorithm and a risk prediction model. By continuously learning historical image data, cargo stacking data, and cargo damage case data, it autonomously optimizes the recognition accuracy and risk prediction accuracy. Based on parameters such as the stacking height, angle, spacing, and weight distribution of the cargo, it predicts risks such as cargo compression damage and tipping, and issues early warnings.

[0039] Based on the independent learning and risk prediction module and the reinforcement learning algorithm, combined with the physical parameters and stacking parameters of the cargo, the risk prediction of cargo tipping and compression damage is realized. The specific implementation process is as follows: (1)Parameter extraction: From the images recognized by the YOLOv11 model, extract the cargo stacking parameters, including the stacking height h, stacking angle θ, cargo spacing d, and cargo center of gravity coordinates (x0, y0); the preset physical parameters of the cargo, including the cargo weight m, volume V, compressive strength σ, and friction coefficient μ.

[0040] (2)Calculation of tipping risk: The calculation formula for the tipping risk probability P1 is: ; where α and β are weight coefficients (α = 0.6, β = 0.4), obtained through training with historical cargo damage data; h is the stacking height, d is the cargo spacing, μ is the friction coefficient, and θ is the stacking angle (the angle with the horizontal plane); when P1 > 30%, it is determined that there is a tipping risk and an early warning is triggered.

[0041] (3)Calculation of compression damage risk: The calculation formula for the compression damage risk probability P2 is: ; where γ and δ are weight coefficients (γ = 0.7, δ = 0.3); F is the pressure exerted on the cargo ( , g is the acceleration due to gravity, and n is the number of upper-layer cargo); σ is the compressive strength of the cargo; t is the compression time; T is the maximum allowable compression time of the cargo; when P2 > 25%, it is determined that there is a compression damage risk and an early warning is triggered.

[0042] (4)Risk level classification: According to the values of P1 and P2, the risk level is classified into three levels: Low risk: P1 ≤ 30% and P2 ≤ 25%; Medium risk: 30% < P1 ≤ 50% or 25% < P2 ≤ 40%; High risk: P1 > 50% or P2 > 40%; Different risk levels correspond to different warning methods: Low risk only message push; Medium risk audible and visual warning and message push; High risk audible and visual warning, message push, and emergency call.

[0043] 4. Navigation Memory Module: Built-in GPS / BeiDou positioning module and route memory algorithm to obtain vehicle location information in real time, remember transportation routes and loading and unloading locations; combined with historical route data, to provide drivers with optimal transportation route suggestions; at the same time, to record the movement trajectory of the smart hardware terminal, so as to facilitate the tracking of equipment usage.

[0044] The navigation memory module uses Dijkstra's algorithm, combined with real-time traffic data and historical route data, to calculate the optimal transportation route. The specific implementation process is as follows: (1) Path node modeling: The starting point, ending point, transit points (such as gas stations, service areas) and traffic intersections are used as path nodes to construct a road network graph G=(V,E), where V is the set of nodes and E is the set of paths between nodes.

[0045] (2) Path weight calculation: A weight w is assigned to each path E. The weight w consists of the path length L, traffic congestion level C, road risk level R, and transportation cost C0, as shown below: ; in, , , , Weighting coefficients ( =0.3, =0.4, =0.2, =0.1), obtained through training with historical route data; the traffic congestion level C is determined by real-time traffic data obtained from the cloud management platform (0≤C≤1, C=0 indicates smooth traffic, C=1 indicates severe congestion); the road risk level R is determined based on road type and historical accident rate (0≤R≤1, R=0 indicates low risk, R=1 indicates high risk).

[0046] (3) Optimal route solution: Dijkstra's algorithm is used to find the path with the smallest weight w from the starting point to the destination in the road network map G, which is the optimal transportation route; at the same time, the route is memorized, and if there are subsequent transportation tasks with the same starting point and destination, the route is automatically pushed and dynamically adjusted in combination with real-time traffic data.

[0047] 5. Early Warning and Communication Module: Includes an audible and visual early warning unit, a wireless communication unit (5G / WiFi / Bluetooth), and an emergency call unit. When the audible and visual early warning unit detects abnormal situations (such as damaged, lost, or overturned goods, or misloading / omission), it issues an audible and visual early warning signal, with the light flashing red and the sound level not less than 80dB. The wireless communication unit establishes data transmission connections with the cloud management platform and mobile interactive terminals, enabling real-time transmission of image data, recognition results, early warning information, and location information. In the event of serious abnormalities (such as large-scale loss of goods or damage to the vehicle compartment), the emergency call unit automatically dials preset driver and logistics management personnel phone numbers to ensure timely handling of abnormal situations.

[0048] 6. Storage Module: It adopts a solid-state storage chip (capacity not less than 128GB, expandable to 1TB) for local storage of collected image data, video recordings, recognition results, early warning logs and other information, with a storage time of not less than 30 days; at the same time, it supports encrypted data storage to prevent data leakage and tampering and ensure information security.

[0049] 7. Power Supply Module: Uses a rechargeable lithium battery with a capacity of no less than 10000mAh, supports fast charging, charging time of no more than 2 hours, and battery life of no less than 8 hours; also supports on-board charging, compatible with 12V / 24V truck power supply, enabling real-time charging during driving to ensure uninterrupted operation of the equipment; built-in power monitoring unit monitors the power status in real time, and issues a low power warning when the power is below 20%, reminding the user to charge.

[0050] 8. Installation and fixing module: It adopts a combination design of magnetic attraction and buckle, which eliminates the need for drilling and welding. It can be quickly fixed to the inner wall, top of the truck body or loading and unloading area, with an installation time of no more than 5 minutes. It is also easy to disassemble and can be flexibly transferred between different vehicles, and is compatible with various vehicle types such as box trucks and flatbed trucks.

[0051] II. Cloud Management Platform: It adopts a distributed architecture, is deployed on a cloud server, and establishes a two-way data transmission connection with mobile smart hardware terminals and mobile interactive terminals.

[0052] 1. Data storage and management: Receives and stores all data transmitted by mobile smart hardware terminals, including images, videos, recognition results, early warning logs, location information, etc., using a distributed database (such as Hadoop) for storage, supporting efficient storage and fast retrieval of massive amounts of data; implements data classification management, archiving data according to dimensions such as transportation orders, vehicle numbers, time ranges, and cargo types, facilitating user query and traceability.

[0053] 2. Model Optimization and Updates: The built-in model training server can collect historical recognition data, misidentification cases, and cargo damage data transmitted from mobile smart hardware terminals. It regularly iterates and trains the YOLOv11 image recognition model and the self-learning model to optimize the model's recognition accuracy and risk prediction accuracy. After training, the model update package is pushed to all mobile smart hardware terminals through the wireless communication module to achieve online model upgrades without manual intervention.

[0054] 3. Global monitoring and management: Supports centralized monitoring of multiple terminals and vehicles. Managers can view the real-time transportation status of all vehicles, cargo status, and working status of smart hardware terminals through the cloud management platform; set early warning thresholds and manage permissions (such as driver permissions and manager permissions); receive early warning information from all terminals; and remotely issue commands, such as starting video recording, adjusting camera angles, and pausing early warnings.

[0055] 4. Data Statistics and Analysis: The built-in data statistics and analysis module performs statistical analysis on data such as cargo loading and unloading efficiency, cargo damage rate, occurrence rate of abnormal situations, and utilization rate of smart hardware terminals, generating visual reports (such as bar charts, line charts, and pie charts) to provide data support for logistics supply chain management decisions; at the same time, it identifies weak points in the logistics transportation process (such as high cargo damage rate in a certain area or low loading and unloading efficiency of a certain vehicle type) and provides optimization suggestions.

[0056] 5. Network security protection: Built-in firewall and data encryption module, using SSL / TLS encryption protocol to encrypt data transmission and storage process throughout, preventing data leakage, tampering and attacks; supports device access authentication, only authorized mobile smart hardware terminals and mobile interactive terminals can access the cloud management platform to ensure system security.

[0057] III. Mobile Interaction Terminal: Adapted to mobile devices such as smartphones and tablets, supporting iOS and Android systems, and connecting to cloud management platforms and mobile smart hardware terminals via an APP.

[0058] 1. Operation and Control: Drivers and managers can remotely control the mobile smart hardware terminal via the APP, including turning the device on / off, adjusting the camera angle, starting video recording, and querying historical records; the APP interface is simple and easy to use, with an icon-based design, and no professional skills are required to operate it.

[0059] 2. Information Viewing: Real-time viewing of cargo quantity identification results, cargo status (intact / damaged), vehicle compartment environment video, vehicle location information, warning logs, historical records, etc.; supports querying historical data by time range, order number, cargo type, etc., and query results can be exported (e.g., PDF, Excel format).

[0060] 3. Warning Reception: Receives warning information (simultaneous sound and light warnings, message push) from mobile smart hardware terminals and cloud management platforms in real time. The warning information includes the type of abnormality, the time of occurrence, the vehicle location, and cargo information, making it easy for users to be aware of and handle abnormal situations in a timely manner. It also supports marking warning information (such as processed or unprocessed) to avoid omissions.

[0061] 4. Message Interaction: Supports message interaction between drivers and managers, allowing them to send text, images, and voice messages via the APP to facilitate communication and handling of abnormal situations; at the same time, it receives notifications and instructions (such as transportation route adjustments and cargo damage handling requirements) issued by the cloud management platform.

[0062] Based on the YOLOv11 image recognition model, this system enables cargo quantity counting, status detection, autonomous learning, and risk prediction. Combined with its mobile design, it achieves intelligent management of the entire loading, unloading, and transportation process.

[0063] Step S1: System initialization.

[0064] Step S11, Installation of the mobile smart hardware terminal: Fix the terminal to a suitable position in the carriage using magnetic attraction and clips, adjust the camera angle to ensure that the collection range covers the inside of the carriage and the loading and unloading area; connect the vehicle power supply or fully charge the lithium battery, start the device, and the device will automatically complete a self-test to check whether the camera, processor, communication module, storage module, etc. are working properly.

[0065] Step S12, Device Activation and Binding: Scan the QR code on the mobile smart hardware terminal using the mobile APP to activate the device and bind it to the vehicle and transportation order (enter information such as vehicle number, order number, cargo type, and total quantity of cargo); the cloud management platform automatically records the device binding information and completes device access authentication.

[0066] Step S13, Model Loading and Parameter Setting: The mobile smart hardware terminal automatically loads the pre-trained YOLOv11 image recognition model, self-learning model, and navigation memory model; the user sets warning thresholds, including cargo tipping angle threshold, pressure warning threshold, and low battery threshold, preset emergency call numbers, and permission settings through the mobile APP and cloud management platform.

[0067] Step S2, the process of loading the vehicle.

[0068] Step S21, Image Acquisition: Start the loading mode. The driver starts it via the mobile APP, or the equipment starts it automatically after detecting the loading and unloading action. The high-definition wide-angle camera or infrared camera in the image acquisition module starts to collect real-time images and videos of the loading and unloading area and the inside of the truck in 360°. The acquisition frequency is 25fps and the image resolution is 1080P. The acquired images and videos are transmitted to the core processing module in real time and simultaneously stored in the local storage module and the cloud management platform.

[0069] Step S22, Image Preprocessing: The core processing module preprocesses the acquired image, including using a Gaussian filtering algorithm to remove noise points in the image, adjusting the image to a resolution of 640×640 to adapt to the input requirements of the YOLOv11 model, grayscale conversion, and using a histogram equalization algorithm to enhance contrast and improve image clarity for subsequent recognition; the preprocessing time does not exceed 100ms / frame.

[0070] Step S23, Cargo Quantity Counting and Status Detection: The core processing module inputs the preprocessed image into the YOLOv11 image recognition model. Based on a pre-trained cargo feature library (containing feature parameters of different types and specifications of cargo), the model performs target detection, counting, and status recognition on the cargo in the image. Step S231, Quantity Counting: The YOLOv11 model uses the anchor box clustering algorithm to locate and identify each item in the image, distinguish different types and specifications of items, and count the real-time quantity of each item; it uses the non-maximum suppression (NMS) algorithm to remove duplicate detection boxes and improve counting accuracy; the counting results are transmitted to the mobile APP and cloud management platform in real time, and users can view the loading quantity in real time.

[0071] Step S232, State Detection: The model detects the appearance of the goods and identifies whether the goods are damaged (such as damaged packaging, deformed goods, leakage), dirty, damp or other abnormal states; by comparing the normal appearance features of the goods with the appearance features of the goods in the real-time image, the similarity is calculated. When the similarity is less than 85%, it is determined that the state of the goods is abnormal and an early warning is triggered immediately.

[0072] Step S24, Misloading and Missing Loading Detection: The system compares the real-time count of goods quantity and type with the quantity and type of goods in the bound transportation order. If a quantity mismatch occurs (e.g., the actual number of goods loaded is less than the order quantity, which is judged as missing loading; the actual number of goods loaded is more than the order quantity, or a type of goods not in the order is loaded, which is judged as misloading), an audible and visual alarm is immediately triggered, and the alarm information is pushed through the mobile APP and cloud management platform to remind the driver and management personnel to check.

[0073] Step S25, Data Retention: After loading is completed, the system automatically generates a loading report, which includes loading time, location, quantity of goods, status of goods, video clips and recognition results. The report is stored in the local storage module and cloud management platform for easy subsequent query and traceability. The driver confirms the completion of loading through a mobile APP. The data after confirmation cannot be tampered with.

[0074] Step S3: The implementation process of the transportation link.

[0075] Step S31, Real-time monitoring: After loading is completed, the system automatically switches to transportation mode. The image acquisition module continuously acquires images of the interior of the carriage (the frame rate is adjusted to 15fps to save power and storage space). The core processing module monitors the status of the goods and the environment of the carriage in real time through the YOLOv11 model. The navigation memory module obtains the vehicle location information and driving route in real time and records the driving trajectory.

[0076] Step S32, Anomaly Detection and Early Warning: Step S321, Cargo Anomaly Detection: Real-time monitoring to check for abnormalities such as damage, loss, tipping, and displacement of cargo; if cargo loss is detected (reduction in the number of cargo in the image) or the damaged area is expanded, an audible and visual alarm is immediately triggered, and an alarm message is pushed to the mobile APP and cloud management platform, and a preset emergency call number is automatically dialed; if cargo displacement is detected, the risk prediction model of the self-learning module is used to determine whether there is a risk of tipping or being crushed.

[0077] Step S322, Risk Prediction: The autonomous learning and risk prediction module extracts parameters such as cargo stacking height, angle, spacing, and weight distribution based on real-time collected images of cargo stacking. Combined with historical cargo damage case data and cargo physical characteristics (such as weight, volume, and fragility), it calculates the probability of cargo tipping over or being damaged by pressure using a reinforcement learning algorithm. If the probability is higher than a preset threshold, an early warning is immediately issued, reminding the driver to stop, check, and adjust the cargo stacking status to achieve early risk prevention.

[0078] Step S323, Anomaly Detection of Carriage Environment: Monitor the temperature and humidity inside the carriage (can be extended to connect to temperature and humidity sensors), and whether there are any abnormal conditions such as open flames or smoke; if the temperature or humidity exceeds the storage requirements of the goods, or if open flames or smoke are detected, an early warning will be triggered immediately and a warning message will be pushed to remind the driver to take action and prevent accidents such as damage to goods or fires.

[0079] Step S33, Navigation and Route Optimization: The navigation memory module obtains vehicle location information and traffic conditions in real time (by acquiring real-time traffic data through a cloud management platform). Combined with historical transportation route data, it provides drivers with optimal transportation route suggestions to avoid congested and dangerous sections. At the same time, it memorizes transportation routes and automatically pushes route suggestions if there are subsequent transportation tasks on the same route. Drivers can view real-time navigation information and driving trajectory through a mobile APP.

[0080] Step S34: Real-time data transmission and storage: All data during transportation, including images, videos, cargo status, location information, and early warning logs, are transmitted to the cloud management platform in real time and simultaneously stored in the local storage module. If a network interruption occurs, the data is automatically stored locally and automatically synchronized to the cloud management platform after the network is restored, ensuring that the data is not lost.

[0081] Step S4, the process of unloading.

[0082] Step S41, Image Acquisition and Preprocessing: After arriving at the unloading location, the driver starts the unloading mode through the mobile APP. The image acquisition module begins to collect real-time images and videos of the unloading area and the inside of the truck, with the acquisition parameters being the same as those in the loading stage. The core processing module preprocesses the acquired images (denoising, resizing, contrast enhancement, etc.).

[0083] Step S42, Cargo Quantity Counting and Item Verification: The core processing module uses the YOLOv11 image recognition model to count and identify the cargo in real time during the unloading process, calculate the actual unloading quantity and items, and compare them with the quantity and items at the time of loading in the transportation order to verify whether there are any unloading errors, loss, or omissions.

[0084] Step S43, Cargo Status Detection: The model performs real-time detection of cargo status during the unloading process, identifying whether the cargo has been damaged, deformed, or damp during transportation. If damage is detected, the location and extent of the damage are recorded and compared with the cargo status at the time of loading to determine whether the damage occurred during loading or transportation, providing a basis for handling cargo damage disputes.

[0085] Step S44, Unloading Confirmation and Data Retention: After unloading is completed, the system automatically generates an unloading report including unloading time, location, quantity, items, condition, damage, video clips, and recognition results, which are stored in the local storage module and cloud management platform. The driver and the consignee can view the unloading report through the mobile APP, sign to confirm that it is correct, and the confirmed report can serve as a document for the handover of goods, which is convenient for subsequent traceability.

[0086] Step S5: Autonomous learning and model optimization implementation process.

[0087] Step S51, Data Collection: The mobile smart hardware terminal continuously collects historical image data (cargo loading and unloading images, transportation process images), recognition results (quantity, status), early warning logs, cargo damage case data, cargo stacking data, etc., and regularly uploads the data to the cloud management platform.

[0088] Step S52, Model Training: The model training server of the cloud management platform filters and labels the collected historical data (such as misidentified cases, damaged cases, and abnormal stacking cases) as training samples to iteratively train the YOLOv11 image recognition model and the self-learning model. During the training process, the gradient descent algorithm is used to optimize the model parameters, adjust the recognition threshold and risk prediction threshold, and improve the model's recognition accuracy, counting accuracy, and risk prediction accuracy.

[0089] Step S53, Model Update: After training is completed, the cloud management platform generates a model update package and pushes the update notification to all mobile smart hardware terminals through the wireless communication module. The terminals automatically download and install the update package to complete the online model upgrade. The upgrade process does not affect the normal operation of the terminals, ensuring continuous system optimization.

[0090] Step S54, Local Adaptive Optimization: The autonomous learning module of the mobile smart hardware terminal performs local model fine-tuning based on locally collected data (such as identification data of a specific type of goods) to adapt to local transportation scenarios and goods types, further improving recognition accuracy and risk prediction targeting; the fine-tuned model parameters are regularly uploaded to the cloud management platform to provide data support for global model optimization.

[0091] Therefore, the present invention adopts the above-mentioned real-time logistics supply chain management system based on artificial intelligence. Relying on mobile intelligent hardware and combining the YOLOv11 image recognition model and autonomous learning algorithm, it realizes automatic counting of goods, real-time status monitoring and early risk prediction throughout the entire process of loading, transportation and unloading. It has the characteristics of convenient operation, flexible installation, traceable information and reliable network security. It fundamentally replaces the core manual operation, solves the pain points of existing technology and improves the level of intelligence of logistics supply chain management.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A real-time logistics supply chain management system based on artificial intelligence, characterized in that: It includes a mobile smart hardware terminal, a cloud management platform, and a mobile interactive terminal, which establish a two-way data transmission connection through a wireless communication module; The mobile smart hardware terminal, with built-in YOLOv11 image recognition model, self-learning module, and navigation memory module, is the core data acquisition, analysis and execution terminal. A cloud-based management platform is used for data storage, global management, and model optimization. Mobile interactive terminal, used for user operation, information viewing and alarm reception.

2. The real-time logistics supply chain management system based on artificial intelligence according to claim 1, characterized in that, The mobile intelligent hardware terminal includes: an image acquisition module, a core processing module, an autonomous learning and risk prediction module, a navigation memory module, an early warning and communication module, a storage module, a power supply module, and an installation and fixing module.

3. The real-time logistics supply chain management system based on artificial intelligence according to claim 2, characterized in that, Image acquisition module: Composed of a high-definition wide-angle camera and an infrared camera; the high-definition wide-angle camera is used for acquiring images of goods and the environment of the carriage in well-lit environments, with a resolution of no less than 1080P and a frame rate of no less than 25fps; the infrared camera is used for image acquisition in low-light environments, achieving clear imaging in no-light environments and ensuring 24-hour uninterrupted acquisition; the camera can rotate 360° to adjust, and the acquisition range covers the entire interior of the carriage and the loading and unloading area, avoiding blind spots.

4. The real-time logistics supply chain management system based on artificial intelligence according to claim 2, characterized in that, Core processing module: It adopts an embedded processor, with a built-in YOLOv11 image recognition model, and supports computational tasks such as image preprocessing, cargo quantity recognition, cargo status detection, and anomaly recognition in the carriage environment; The YOLOv11 image recognition model uses an end-to-end computation approach. The input is a preprocessed image, and the output is the cargo target box, category probability, and confidence score. The implementation process is as follows: (1) Feature extraction: CSPDarknet is used as the backbone network to extract multi-scale features from the input image. Low-level features, mid-level features and high-level features in the image are extracted through convolutional layers, pooling layers and residual connections. The convolutional layers use 3×3 convolutional kernels and 1×1 convolutional kernels, and the pooling layers use max pooling with a stride of 2 to ensure the efficiency and accuracy of feature extraction. (2) Feature fusion: PANet is used for feature fusion to fuse feature maps of different scales; during the fusion process, a combination of upsampling and downsampling is used; (3) Target Detection and Classification: A multi-scale detection head is used to detect targets on the fused feature map. Each detection head corresponds to targets at different scales. Each detection head outputs the bounding box coordinates (x, y, w, h), class probability, and confidence score. y represents the coordinates of the center point of the target bounding box; w and h represent the width and height of the target bounding box, respectively. (4) Confidence screening and non-maximum suppression: Confidence screening is performed on the target boxes output by the detection head, and target boxes with confidence higher than the preset threshold are retained; non-maximum suppression algorithm is used to calculate the cross-union ratio between target boxes. Target boxes with cross-union ratio greater than the preset threshold are retained as the target boxes with the highest confidence, and duplicate target boxes are deleted to obtain the final recognition result.

5. The real-time logistics supply chain management system based on artificial intelligence according to claim 4, characterized in that, Based on the recognition results of the YOLOv11 model, the following calculation method is used to achieve accurate counting of the quantity of goods: 1) Single-frame counting: For each preprocessed image frame, all cargo targets are identified through the YOLOv11 model, and the number of target boxes is counted as the number of cargo in a single frame. If there is cargo occlusion and the occlusion area ≤ 30%, the features of the occluded cargo are restored through the feature completion algorithm to ensure accurate counting. If the occlusion area > 30%, it is marked as suspected occlusion, and supplementary recognition is performed in combination with adjacent frame images. 2) Multi-frame fusion counting: The sliding window fusion algorithm is adopted to select the single-frame counting results of consecutive frame images and calculate the average value as the number of cargo at the current moment. If the counting result of a certain frame deviates from the average value by more than 5%, it is regarded as an abnormal frame, and the counting result of this frame is excluded, and the average value is recalculated to ensure the stability and accuracy of counting. 3) Loading and unloading counting calibration: In the loading process, as the cargo is continuously loaded, the counting result is accumulated in real time. For each piece of cargo loaded, the counting result is incremented by 1, and at the same time, it is compared with the total order quantity for calibration. In the unloading process, as the cargo is continuously unloaded, the counting result is decremented in real time. For each piece of cargo unloaded, the counting result is decremented by 1, and it is compared with the quantity during loading for calibration. If there is a counting deviation, an alarm is immediately triggered to remind the user to check. ； The counting error of the cargo does not exceed 1%.

6. The real-time logistics supply chain management system based on artificial intelligence according to claim 2, characterized in that, Autonomous learning and risk prediction module: Electrically connected to the core processing module, with an embedded reinforcement learning algorithm and risk prediction model. By continuously learning historical image data, cargo stacking data, and cargo damage case data, it autonomously optimizes the recognition accuracy and risk prediction accuracy. Based on the parameters of the stacking height, angle, spacing, and weight distribution of the cargo, it predicts the risks of cargo being damaged by pressure and toppling, and issues early warnings. Based on the autonomous learning and risk prediction module and the reinforcement learning algorithm, combined with the physical parameters and stacking parameters of the cargo, the risk prediction of cargo toppling and being damaged by pressure is realized as follows: (1) Parameter extraction: From the images recognized by the YOLOv11 model, the cargo stacking parameters are extracted, including the stacking height h, stacking angle θ, cargo spacing d, and cargo center of gravity coordinates (x0, y0); the preset cargo physical parameters include cargo weight m, volume V, compressive strength σ, and friction coefficient μ. (2) Calculation of toppling risk: The calculation formula for the toppling risk probability P1 is: ； Where α and β are weight coefficients obtained through training with historical cargo damage data; h is the stacking height, d is the cargo spacing, μ is the friction coefficient, and θ is the stacking angle. When P1 > 30%, it is determined that there is a toppling risk and an alarm is triggered. (3) Calculation of compressive damage risk: The calculation formula for the compressive damage risk probability P2 is: ； Where γ and δ are weight coefficients; F is the pressure exerted on the cargo; σ is the compressive strength of the cargo; t is the compression time; T is the maximum allowable compression time of the cargo. When P2 > 25%, it is determined that there is a compressive damage risk and an alarm is triggered. (4) Risk level classification: According to the values of P1 and P2, the risk level is divided into three levels: Low risk: P1 ≤ 30% and P2 ≤ 25%; Medium risk: 30% < P1 ≤ 50% or 25% < P2 ≤ 40%; High risk: P1 > 50% or P2 > 40%. Different risk levels correspond to different early warning methods: low risk only push notifications; medium risk audio-visual warnings and push notifications; high risk audio-visual warnings, push notifications, and emergency calls.

7. The real-time logistics supply chain management system based on artificial intelligence according to claim 2, characterized in that, Navigation memory module: Built-in GPS or Beidou positioning module and path memory algorithm to obtain vehicle location information in real time, remember transportation routes and loading and unloading locations; combined with historical route data, to provide drivers with the best transportation route suggestions; at the same time, to record the movement trajectory of smart hardware terminals, so as to facilitate the tracking of equipment usage; The navigation memory module uses Dijkstra's algorithm, combined with real-time traffic data and historical route data, to calculate the optimal transportation route. The implementation process is as follows: (1) Path node modeling: The starting point, ending point, waypoint, and traffic intersection are taken as path nodes to construct a road network graph G=(V,E), where V is the set of nodes and E is the set of paths between nodes; (2) Path weight calculation: A weight w is assigned to each path E. The weight w consists of the path length L, traffic congestion level C, road risk level R, and transportation cost C0, as shown below: ； in, , , , The weighting coefficients are obtained through training with historical route data; the traffic congestion level C is determined by real-time traffic data obtained from the cloud management platform, 0≤C≤1, C=0 indicates smooth traffic and C=1 indicates severe congestion; the road risk level R is determined based on road type and historical accident rate, 0≤R≤1, R=0 indicates low risk and R=1 indicates high risk. (3) Optimal route solution: Dijkstra's algorithm is used to find the path with the minimum weight w from the starting point to the ending point in the road network graph G, which is the optimal transportation route.

8. The real-time logistics supply chain management system based on artificial intelligence according to claim 2, characterized in that, Early warning and communication module: including sound and light early warning unit, wireless communication unit, and emergency call unit; the sound and light early warning unit issues sound and light early warning signals when it detects an abnormal situation; the wireless communication unit is used to establish data transmission connection with the cloud management platform and mobile interactive terminal to realize real-time transmission of image data, recognition results, early warning information, and location information; the emergency call unit automatically dials the preset driver's phone number and logistics management personnel's phone number when a serious abnormality occurs to ensure timely handling of the abnormal situation; Storage module: Employs solid-state storage chips for local storage of collected image data, video recordings, recognition results, and early warning log information, with a storage time of no less than 30 days; it also supports encrypted data storage to prevent data leakage and tampering, ensuring information security; Power supply module: It adopts a rechargeable lithium battery with a capacity of no less than 10000mAh and supports fast charging; it also supports vehicle charging and is compatible with 12V / 24V truck power supply to enable real-time charging during driving and ensure uninterrupted operation of the equipment; it has a built-in power monitoring unit to monitor the power status in real time and issue a low power warning when the power is below 20% to remind the user to charge. Installation and fixing module: It adopts a combination design of magnetic attraction and buckle, which eliminates the need for drilling and welding, and can be quickly fixed to the inner wall, top of the carriage or loading and unloading area; at the same time, it is easy to disassemble and can be flexibly transferred between different vehicles.

9. The real-time logistics supply chain management system based on artificial intelligence according to claim 1, characterized in that, Cloud management platform: Adopting a distributed architecture, it is deployed on a cloud server and establishes a two-way data transmission connection with mobile smart hardware terminals and mobile interactive terminals; (1) Data storage and management: Receive and store all data transmitted by the mobile smart hardware terminal, including images, videos, recognition results, early warning logs and location information. Use distributed database storage to realize data classification management, archive data in multiple dimensions, and facilitate user query and traceability. (2) Model optimization and update: The built-in model training server collects historical recognition data, misidentification cases and cargo damage data transmitted by mobile smart hardware terminals, and regularly iterates and trains the YOLOv11 image recognition model and the self-learning model to optimize the model recognition accuracy and risk prediction accuracy. After training is completed, the model update package is pushed to all mobile smart hardware terminals through the wireless communication module to realize online model upgrade without manual intervention; (3) Global monitoring and management: Supports centralized monitoring of multiple terminals and multiple vehicles. Managers can view the real-time transportation status, cargo status and working status of smart hardware terminals of all vehicles through the cloud management platform. Set warning thresholds and access control; receive warning information from all terminals and remotely issue instructions. (4) Data statistics and analysis: The built-in data statistics and analysis module performs statistical analysis on data such as cargo loading and unloading efficiency, cargo damage rate, abnormal situation occurrence rate and smart hardware terminal usage rate, generates visual reports, and provides data support for logistics supply chain management decisions; at the same time, it identifies the weak points in the logistics transportation process and provides optimization suggestions. (5) Network security protection: Built-in firewall and data encryption module, using SSL / TLS encryption protocol to encrypt the data transmission and storage process to prevent data leakage, tampering and attack; supports device access authentication, only authorized mobile smart hardware terminals and mobile interactive terminals can access the cloud management platform to ensure system security.

10. A real-time logistics supply chain management system based on artificial intelligence according to claim 1, characterized in that, Mobile interactive terminal: Adapted to mobile devices such as mobile phones and tablets, supporting iOS and Android systems, and can establish connections with cloud management platforms and mobile smart hardware terminals through APP; (1) Operation control: Drivers and managers can remotely control the mobile smart hardware terminal through the APP, including starting and stopping the device, adjusting the camera angle, starting video recording, and querying historical records; the APP interface is simple and easy to operate, and adopts an icon-based design, so no professional skills are required to operate it. (2) Information viewing: Real-time viewing of cargo quantity identification results, cargo status, vehicle environment video, vehicle positioning information, warning logs, and historical records; supports multi-dimensional query of historical data by time range, order number, and cargo type, and query results can be exported; (3) Warning reception: Real-time reception of warning information pushed by mobile smart hardware terminals and cloud management platform. The warning information includes the type of abnormality, time of occurrence, vehicle location and cargo information, so that users can know and deal with abnormal situations in a timely manner. Supports marking of early warning information to avoid omissions; (4) Message interaction: Supports message interaction between drivers and managers, sending text, pictures and voice messages through the APP to facilitate communication and handling of abnormal situations; at the same time, it receives notifications and instructions issued by the cloud management platform.

Images