Intelligent cargo checking and carrying device and carrying system for supply chain management

Abstract

The invention relates to the technical field of cargo handling, and particularly discloses an intelligent cargo checking and handling device and system for supply chain management, and the device comprises a mobile platform which is a four-wheel-drive automatic guided vehicle and has infrared obstacle avoidance and navigation functions, an RGB-D camera is fixedly installed at the top of the front end of the mobile platform, and the RGB-D camera is fixedly connected with the mobile platform. Three-dimensional data of the artwork is obtained through an image acquisition module, after data preprocessing and recognition, a generative adversarial network is adopted for pressure distribution estimation, a dynamic model is constructed through finite element analysis and a recurrent neural network based on the pressure data, a dynamic pressure peak value is predicted, and finally an optimal carrying path is generated through a multi-target optimization algorithm. The technical problems that the stacking pressure of artworks is concentrated, the dynamic load intensifies damage in the carrying process and the pressure minimization is not considered in path planning are solved, the whole process optimization from image recognition to intelligent carrying is achieved, and the safety and efficiency of artwork carrying are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of cargo handling technology, specifically to an intelligent cargo inventory and handling device and system for supply chain management. Background Technology

[0002] Intelligent inventory and handling systems in supply chain management are core technologies in modern logistics and warehousing. They combine automated devices with intelligent algorithms to achieve efficient identification, location, handling, and status monitoring of inventory goods. These systems typically include mobile handling devices, image acquisition equipment, a central processing unit, and a communication module, enabling basic functions such as goods picking, transportation, and data collection. With the development of artificial intelligence technology, existing systems have gradually incorporated machine vision and automated control technologies to improve the accuracy of goods identification and the efficiency of handling operations.

[0003] For example, application number "CN202510556491.6" discloses a handling system that does not require a separate release request to the control module, effectively reducing the communication frequency between the handling vehicle and the control module. Furthermore, it can adopt different release strategies based on different situations, ensuring that the exclusive use of cells is released as early as possible, while also providing a fallback option for changing the travel route in special circumstances. However, in application scenarios involving high-value, fragile artworks and other special goods, existing systems have significant shortcomings: First, traditional image recognition technology can only complete basic shape recognition and cannot accurately analyze the pressure distribution under stacked artwork conditions, potentially leading to damage to vulnerable parts due to concentrated pressure. Second, existing dynamic models are mostly based on the assumption of rigid objects, making it difficult to accurately predict the dynamic load changes of irregularly shaped artworks during handling and effectively assess the risk of breakage under motion. Finally, conventional path planning algorithms only consider conventional factors such as distance and time, failing to minimize dynamic pressure as an optimization objective, making it difficult to balance handling safety and operational efficiency in complex warehousing environments. These problems severely restrict the application effectiveness of intelligent inventory and handling systems in the management of high-value, fragile goods. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] This invention provides an intelligent inventory and handling device and system for supply chain management, which solves the problems mentioned in the background art.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution: a smart inventory and handling device for supply chain management, comprising a mobile platform, wherein the mobile platform is a four-wheel drive automated guided vehicle, and the mobile platform has infrared obstacle avoidance and navigation functions, and an RGB-D camera is fixedly installed on the top front end of the mobile platform;

[0008] The mobile platform has an embedded GPU processor, which is directly connected to the RGB-D camera via a data cable to run a lightweight AI model for data preprocessing.

[0009] The bottom of the mobile platform is fixedly connected to an electric fork, which supports lifting and extension. The operation of the electric fork is controlled by the GPU processor in the mobile platform through a control line.

[0010] The mobile platform has an embedded communication module, which is a Wi-Fi / 5G wireless transmission unit. The communication module is connected to the GPU processor in the mobile platform via a circuit board and is used to synchronize data with the central management system in real time.

[0011] A smart inventory and handling system for supply chain management includes:

[0012] The image acquisition module's core function is to acquire images and depth data of high-value, fragile artworks in real time using an RGB-D camera on the handling device, including stacked shapes, surface textures, and three-dimensional coordinate information, providing raw data input for subsequent analysis.

[0013] The core function of the data preprocessing module is to perform noise filtering, coordinate calibration, and data format standardization on the raw data transmitted by the image acquisition module to ensure data quality and consistency, making it easier for AI models to process.

[0014] The core function of the artwork recognition module is to use a convolutional neural network to extract features and classify the preprocessed data, identify the type, shape and key contact points of the artwork, and output the recognition results for stress analysis.

[0015] The core function of the pressure distribution estimation module is to calculate the pressure distribution at the contact points of stacked artworks based on the output of the artwork recognition module, using a generative adversarial network algorithm to generate data on pressure concentration areas and reveal potential damage risks.

[0016] The core function of the dynamics simulation module is to use finite element analysis and recurrent neural networks to build dynamic models, simulate acceleration, vibration and load changes during the transportation process, and predict the evolution of pressure distribution under different motion states.

[0017] The core function of the dynamic pressure prediction module is to combine the output of the dynamic simulation module with the time series analysis algorithm to calculate the dynamic pressure peak and assess the probability of damage to the artwork during handling.

[0018] The core function of the path planning module is to use reinforcement learning and multi-objective particle swarm optimization algorithms to generate the optimal transport path and speed curve with the goal of minimizing the maximum dynamic pressure, while taking into account environmental constraints and cargo priority.

[0019] The core function of the control and execution module is to control the movement of the handling device and the operation of the forks in real time according to the instructions of the path planning module, execute the handling task, and monitor the execution status to ensure the coordinated operation of the system.

[0020] Furthermore, the image acquisition module further includes:

[0021] The RGB-D camera captures images at 30 frames per second, covering a 120-degree field of view. It can capture multi-angle images and depth point cloud data of artworks. The captured data includes color information, depth distance, and 3D coordinates. The built-in sensor synchronizes the timestamp to ensure real-time data. The image acquisition module also integrates an ambient light adaptive function, automatically adjusting exposure and gain according to warehouse lighting conditions to avoid overexposure or dark areas affecting data quality. At the same time, the embedded processor performs preliminary data compression to reduce transmission load. All captured raw data is uploaded to the central management system's storage unit in real time via a wireless communication module, providing a complete and reliable input source for subsequent modules and ensuring efficient operation even in special environments such as low light or high humidity.

[0022] Furthermore, the data preprocessing module further includes a data cleaning submodule and a coordinate transformation submodule. The data cleaning submodule uses median filtering and wavelet transform algorithms to remove noise and outliers from image and depth data, improving the signal-to-noise ratio. The coordinate transformation submodule converts the acquired 2D images and 3D point cloud data into a standardized format in the world coordinate system, facilitating subsequent AI model processing. The data preprocessing module also includes a data verification function, checking data integrity through checksums and hash values ​​to prevent transmission errors. The preprocessed data is divided into training and test sets for model iteration and optimization. Simultaneously, the module integrates a caching mechanism to temporarily store processing results to cope with network latency, ensuring smooth data flow to the artwork recognition module, reducing system response time, and improving overall efficiency.

[0023] Furthermore, the artwork recognition module employs a deep convolutional neural network model pre-trained on the ResNet architecture and adapted to high-value artwork datasets through transfer learning. This model enables feature extraction of shape and texture, and the recognition process includes bounding box detection, semantic segmentation, and classification output. It can distinguish different types of artworks such as sculptures, paintings, and ceramics, and output the coordinates and confidence scores of key contact points. The artwork recognition module also integrates an adversarial example defense mechanism, which enhances the robustness of the model through generative adversarial networks to prevent misidentification. The recognition results are stored in a structured JSON format, including the artwork ID, shape parameters, and risk level, and are updated in real time to a central database. This provides accurate input to the stress distribution estimation module, ensuring high accuracy even in complex stacking scenarios.

[0024] Furthermore, the pressure distribution estimation module constructs a pressure mapping model based on a generative adversarial network algorithm. This model takes the contact point data output by the artwork recognition module as input, simulates the pressure distribution map through a generator network, and verifies its authenticity through a discriminator network. The output includes a pressure value matrix and a heat map, identifying areas of high pressure concentration. The pressure distribution estimation module also introduces physical constraints, such as the material elastic modulus and stacking angle, and uses finite element analysis to assist in calculations, improving estimation accuracy. The generated pressure data is compared with historical records for iterative training of the model. The output results are stored in vector form and trigger an alarm when the pressure exceeds a threshold. This is then directly used as input to the dynamic simulation module, realizing the transition from static recognition to dynamic analysis.

[0025] Furthermore, the dynamic simulation module employs a hybrid model combining finite element analysis and recurrent neural networks. Finite element analysis processes the material properties and geometry of the artwork to simulate stress distribution under static load, while the recurrent neural network learns time-series data to predict the impact of acceleration and vibration on pressure during handling. The dynamic simulation module inputs pressure distribution data, solves the dynamic response through numerical integration and differential equations, and outputs curves including velocity, acceleration, and pressure changes. It also simulates extreme situations such as sudden stops or turns. All simulation results are stored in time steps for real-time prediction updates. Simultaneously, the module integrates sensitivity analysis to evaluate the impact of parameter changes on the results, ensuring that the input to the dynamic pressure prediction module is reliable and comprehensive.

[0026] Furthermore, the dynamic pressure prediction module is based on time series analysis algorithms, including an autoregressive integral moving average model and a long short-term memory network. This module takes the dynamic data output by the dynamic simulation module as input, extracts features through sliding window technology, and predicts the pressure peak and fluctuation trend at future time points. The dynamic pressure prediction module also includes a risk assessment submodule, which calculates the probability of damage based on historical pressure data and the artwork fragility index. The output includes risk level and recommended measures. All prediction results are visualized in the form of charts and reports and pushed to the path planning module in real time for decision optimization. At the same time, the module supports online learning, adjusting model parameters according to new data to adapt to environmental changes such as uneven ground or load variations.

[0027] Furthermore, the path planning module employs a multi-objective particle swarm optimization algorithm combined with a reinforcement learning framework. This module takes the output of the dynamic pressure prediction module as its main input, while also considering the environmental map, obstacle locations, and cargo priorities to generate a Pareto optimal solution set, including path coordinates and velocity curves. The path planning module also integrates a real-time replanning function, dynamically adjusting the path through sensor feedback to avoid collisions and vibration hotspots. The optimization objectives include minimizing maximum dynamic pressure, shortest time, and lowest energy consumption. The output results are transmitted to the control execution module in the form of path point sequences and control commands, and stored in a central log for subsequent analysis, ensuring safe and efficient operation even in challenging environments such as narrow passages or stacked artworks.

[0028] Furthermore, the control execution module further includes an instruction parsing submodule and an execution monitoring submodule. The instruction parsing submodule converts the path and speed data output by the path planning module into executable commands for the handling device, including directional control of the moving platform and lifting and extending / retracting operations of the forks. The execution monitoring submodule collects the device status, such as position, speed, and load, in real time through sensors, compares it with expected values, and adjusts parameters to eliminate deviations. The control execution module also integrates a fault detection and recovery mechanism. When an anomaly, such as overpressure or communication interruption, is detected, an emergency protocol, such as pause or rollback, is automatically triggered to ensure the safety of the workpiece. All execution data is recorded and fed back to the central management system for system optimization and report generation.

[0029] (III) Beneficial Effects

[0030] This invention provides an intelligent inventory and handling device and system for supply chain management. It offers the following advantages:

[0031] (I) This intelligent inventory and handling device and system for supply chain management collects images and depth data of high-value, fragile artworks in real time through an image acquisition module, including stacking shape and three-dimensional coordinate information. Then, the data preprocessing module performs noise filtering and coordinate calibration on the raw data to ensure consistent data quality. Next, the artwork recognition module uses a convolutional neural network for feature extraction and classification to accurately identify the type of artwork and key contact points. Finally, the pressure distribution estimation module calculates the pressure distribution at the contact points based on the recognition results using a generative adversarial network algorithm, generating a heat map and data matrix of pressure concentration areas. This effectively solves the technical problem that irregular shapes of artworks can cause surface damage or structural damage due to pressure concentration. It achieves rapid processing from raw data to pressure estimation, accurately assesses the risks under static stacking conditions, significantly reduces potential damage to artworks, and improves the safety and reliability of warehouse management.

[0032] (II) The intelligent inventory and handling device and system for supply chain management, based on pressure distribution data, firstly, the dynamic simulation module uses finite element analysis and recurrent neural network to construct a hybrid model to simulate the influence of acceleration and vibration on pressure distribution during handling, and outputs dynamic response curves. Then, the dynamic pressure prediction module uses time series analysis algorithms, including autoregressive integral moving average model and long short-term memory network, to predict the pressure peak and fluctuation trend at future time points and calculate the probability of breakage. This effectively solves the technical problem of damage to artworks caused by the increased pressure concentration points due to dynamic load generated by motion during the handling of artworks. It realizes the transition from static pressure analysis to dynamic risk prediction, can identify high-risk links in handling in advance, reduce the probability of damage to artworks, and enhance the adaptability and robustness of the system in complex environments.

[0033] (III) The intelligent inventory and handling device and system for supply chain management, based on dynamic pressure prediction data, firstly, the path planning module adopts a multi-objective particle swarm optimization algorithm combined with a reinforcement learning framework to generate the optimal handling path and speed curve with the goal of minimizing the maximum dynamic pressure. At the same time, environmental constraints and cargo priority are considered. Then, the control execution module parses the path planning output into executable commands, controls the movement of the handling device and the operation of the forks in real time, and adjusts parameters by monitoring the status through sensors. It integrates fault detection and recovery mechanisms to deal with abnormal situations, thereby effectively solving the technical problem of high risk of damage and low efficiency of existing handling systems due to path planning not considering the minimization of dynamic pressure. It realizes intelligent decision-making from prediction to execution, optimizes the overall handling process, significantly improves operational efficiency and safety level, and is suitable for special warehousing environments of high-value and fragile goods. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the structure of the conveying device of the present invention;

[0035] Figure 2 This is a system flowchart of the material handling system of the present invention.

[0036] In the picture: 1. Mobile platform; 2. RGB-D camera; 3. Electric forklift. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] First embodiment: as follows Figures 1 to 2 As shown, the present invention provides a technical solution: a smart inventory and handling device for supply chain management, including a mobile platform 1, which is a four-wheel drive automatic guided vehicle, and the mobile platform 1 has infrared obstacle avoidance and navigation functions, and an RGB-D camera 2 is fixedly installed on the top of the front end of the mobile platform.

[0039] The mobile platform 1 has an embedded GPU processor, which is directly connected to the RGB-D camera 2 via a data cable to run a lightweight AI model for data preprocessing.

[0040] The bottom of the mobile platform 1 is fixedly connected to an electric fork 3, which supports lifting and extension. The operation of the electric fork 3 is controlled by the GPU processor in the mobile platform 1 through a control line.

[0041] The mobile platform 1 has an embedded communication module, which is a Wi-Fi / 5G wireless transmission unit. The communication module is connected to the GPU processor in the mobile platform 1 via a circuit board and is used to synchronize data with the central management system in real time.

[0042] A smart inventory and handling system for supply chain management includes:

[0043] The image acquisition module's core function is to acquire images and depth data of high-value, fragile artworks in real time using an RGB-D camera on the handling device, including stacked shapes, surface textures, and three-dimensional coordinate information, providing raw data input for subsequent analysis.

[0044] The core function of the data preprocessing module is to perform noise filtering, coordinate calibration, and data format standardization on the raw data transmitted by the image acquisition module to ensure data quality and consistency, making it easier for AI models to process.

[0045] The core function of the artwork recognition module is to use a convolutional neural network to extract features and classify the preprocessed data, identify the type, shape and key contact points of the artwork, and output the recognition results for stress analysis.

[0046] The core function of the pressure distribution estimation module is to calculate the pressure distribution at the contact points of stacked artworks based on the output of the artwork recognition module, using a generative adversarial network algorithm to generate data on pressure concentration areas and reveal potential damage risks.

[0047] The core function of the dynamics simulation module is to use finite element analysis and recurrent neural networks to build dynamic models, simulate acceleration, vibration and load changes during the transportation process, and predict the evolution of pressure distribution under different motion states.

[0048] The core function of the dynamic pressure prediction module is to combine the output of the dynamic simulation module with the time series analysis algorithm to calculate the dynamic pressure peak and assess the probability of damage to the artwork during handling.

[0049] The core function of the path planning module is to use reinforcement learning and multi-objective particle swarm optimization algorithms to generate the optimal transport path and speed curve with the goal of minimizing the maximum dynamic pressure, while taking into account environmental constraints and cargo priority.

[0050] The core function of the control and execution module is to control the movement of the handling device and the operation of the forks in real time according to the instructions of the path planning module, execute the handling task, and monitor the execution status to ensure the coordinated operation of the system.

[0051] The image acquisition module further includes:

[0052] The RGB-D camera captures images at 30 frames per second, covering a 120-degree field of view. It can capture multi-angle images and depth point cloud data of artworks. The acquired data includes color information, depth distance, and 3D coordinates. The built-in sensor synchronizes the timestamp to ensure real-time data. The image acquisition module also integrates an ambient light adaptive function, automatically adjusting exposure and gain according to warehouse lighting conditions to avoid overexposure or dark areas affecting data quality. At the same time, the embedded processor performs preliminary data compression to reduce transmission load. All acquired raw data is uploaded to the central management system's storage unit in real time via a wireless communication module, providing a complete and reliable input source for subsequent modules and ensuring efficient operation even in special environments such as low light or high humidity.

[0053] The data preprocessing module further includes a data cleaning submodule and a coordinate transformation submodule. The data cleaning submodule uses median filtering and wavelet transform algorithms to remove noise and outliers from image and depth data, improving the signal-to-noise ratio. The coordinate transformation submodule converts the acquired 2D images and 3D point cloud data into a standardized format in the world coordinate system, facilitating subsequent AI model processing. The data preprocessing module also includes a data verification function, checking data integrity through checksums and hash values ​​to prevent transmission errors. The preprocessed data is divided into training and test sets for model iteration and optimization. At the same time, the module integrates a caching mechanism to temporarily store processing results to cope with network latency, ensuring smooth data flow to the artwork recognition module, reducing system response time, and improving overall efficiency.

[0054] The art recognition module further employs a deep convolutional neural network model, which is pre-trained based on the ResNet architecture and adapted to high-value art datasets through transfer learning. This model achieves feature extraction of shape and texture, and the recognition process includes bounding box detection, semantic segmentation, and classification output. It can distinguish different art types such as sculptures, paintings, and ceramics, and output the coordinates and confidence scores of key contact points. The art recognition module also integrates an adversarial example defense mechanism, which enhances the robustness of the model through generative adversarial networks to prevent misidentification. The recognition results are stored in a structured JSON format, including the art ID, shape parameters, and risk level, and are updated in real time to the central database, providing accurate input to the stress distribution estimation module and ensuring high accuracy even in complex stacking scenarios.

[0055] The pressure distribution estimation module further constructs a pressure mapping model based on a generative adversarial network algorithm. This model takes the contact point data output by the artwork recognition module as input, simulates the pressure distribution map through a generator network, and verifies its authenticity through a discriminator network. The output includes a pressure value matrix and a heat map, identifying areas of high pressure concentration. The pressure distribution estimation module also introduces physical constraints, such as the material elastic modulus and stacking angle, and uses finite element analysis to assist in the calculation, improving the estimation accuracy. The generated pressure data is compared with historical records for iterative training of the model. The output results are stored in vector form and trigger an alarm when the pressure exceeds a threshold. This is then directly used as input to the dynamic simulation module, realizing the transition from static recognition to dynamic analysis.

[0056] During operation, the image acquisition module first uses an RGB-D camera on the handling device to acquire real-time images and depth data of high-value, fragile artworks, including stacking shape, surface texture, and 3D coordinate information. Next, the data preprocessing module performs noise filtering, coordinate calibration, and format standardization on the raw data to ensure consistent data quality. Then, the artwork recognition module uses a convolutional neural network to extract features and classify the preprocessed data, identifying the type, shape, and key contact points of the artworks and outputting the recognition results. Finally, the pressure distribution estimation module calculates the pressure distribution at the contact points of the stacked artworks based on the recognition results using a generative adversarial network algorithm, generating data on pressure concentration areas. This reveals and solves the technical problem of surface or structural damage caused by pressure concentration due to irregular shapes when artworks are stacked. Specifically:

[0057] First, the RGB-D camera in the image acquisition module acquires images and depth data of high-value, fragile artworks in real time, including stacked shapes, surface textures, and 3D coordinate information. The acquisition process is performed at 30 frames per second, covering a 120-degree field of view. An ambient light adaptive function automatically adjusts exposure and gain to ensure stable data quality in low-light or high-humidity warehouse environments. Simultaneously, a built-in sensor synchronizes timestamps, generating raw point cloud and image datasets. Next, the data preprocessing module filters noise from the raw data, using median filtering to remove image noise and wavelet transform to eliminate outliers in the depth data. Then, coordinate calibration is performed, unifying the 2D image pixel coordinates and 3D point cloud data to the world coordinate system. A data verification function calculates checksums and hash values ​​to ensure integrity. Finally, the format is standardized into a unified tensor structure for easy processing by subsequent AI models. The artwork recognition module then uses a deep convolutional neural network model based on the ResNet architecture. Feature extraction and classification involve several computational steps. First, multiple convolutional layers are used to apply filters to extract shape and texture features. Then, the ReLU activation function is used to introduce non-linearity. Max pooling layers are used to reduce dimensionality through downsampling. Finally, fully connected layers are used to integrate the features, and a softmax function is applied to output a probability distribution, classifying the type of artwork (e.g., sculpture, painting, or ceramic). Simultaneously, a semantic segmentation algorithm is used to identify the coordinates of key contact points, outputting structured data including artwork ID, shape parameters, and risk level. Finally, a pressure distribution estimation module uses the key contact point data from the identification results and a generative adversarial network (GAN) algorithm to calculate the contact point pressure distribution of stacked artworks. Specifically, the generator network takes contact point coordinates and a random noise vector as input, generating a pressure value matrix through fully connected and deconvolutional layers. The discriminator network takes real pressure data or generated data as input, outputs a probability of authenticity through convolutional layers, and optimizes the model by minimizing the adversarial loss function. The key pressure distribution calculation formula is as follows:

[0058]

[0059] Where G is the generator, D is the discriminator, x is the actual pressure data, and z is the noise vector. It is the actual data distribution. The noise distribution is regularized by combining physical constraints such as the elastic modulus of materials during the training process, generating a heat map and data matrix of the pressure concentration area, thereby revealing and solving the technical problem that the pressure concentration caused by the irregular shape of the artwork during stacking can easily cause surface damage or structural damage. The beneficial effect of this embodiment is that it can accurately estimate the pressure distribution and reduce the risk of damage to artworks in the static stacking state.

[0060] Second embodiment: as follows Figures 1 to 2 As shown, the dynamic simulation module further employs a hybrid model combining finite element analysis and recurrent neural networks. Finite element analysis processes the material properties and geometry of the artwork, simulating stress distribution under static load. The recurrent neural network learns time-series data to predict the impact of acceleration and vibration on pressure during handling. The dynamic simulation module inputs pressure distribution data, solves the dynamic response through numerical integration and differential equations, and outputs curves including velocity, acceleration, and pressure changes. It also simulates extreme situations such as sudden stops or turns. All simulation results are stored in time steps for real-time prediction updates. Simultaneously, the module integrates sensitivity analysis to evaluate the impact of parameter changes on the results, ensuring that the input to the dynamic pressure prediction module is reliable and comprehensive.

[0061] The dynamic stress prediction module is further based on time series analysis algorithms, including an autoregressive integral moving average model and a long short-term memory network. This module takes the dynamic data output from the dynamic simulation module as input, extracts features through sliding window technology, and predicts the stress peak and fluctuation trend at future time points. The dynamic stress prediction module also includes a risk assessment submodule, which calculates the probability of failure based on historical stress data and the artwork's fragility index. The output includes risk level and recommended measures. All prediction results are visualized in the form of charts and reports and pushed to the path planning module in real time for decision optimization. At the same time, the module supports online learning, adjusting model parameters according to new data to adapt to environmental changes such as uneven ground or load variations.

[0062] During operation, based on the pressure distribution data output by the first embodiment, the dynamic simulation module first constructs a dynamic model using finite element analysis and recurrent neural networks to simulate the impact of acceleration, vibration, and load changes on pressure distribution during handling, outputting a dynamic response curve. Then, the dynamic pressure prediction module uses time series analysis algorithms, including an autoregressive integral moving average model and a long short-term memory network, to predict the pressure peak and fluctuation trend at future time points using the dynamic simulation data as input, and calculates the probability of breakage. This reveals and solves the technical problem of art damage caused by dynamic loads amplified by motion during art handling, leading to increased pressure concentration points. Specifically:

[0063] Based on the pressure distribution data output from the first embodiment, namely the pressure value matrix and thermogram generated by the pressure distribution estimation module, the dynamic simulation module uses finite element analysis and recurrent neural network to construct a hybrid dynamics model. The finite element analysis part discretizes the geometric mesh of the artwork and applies material properties such as elastic modulus and Poisson's ratio to solve the static stress distribution based on Hooke's law and equilibrium equations. The specific calculation steps include dividing the element mesh, applying boundary conditions such as fixed supports or contact constraints, and then assembling the stiffness matrix and solving for nodal displacements and stress values. The recurrent neural network part takes time series data as input, including the acceleration of the handling device, velocity sensor readings, and environmental vibration frequencies. It uses the gating mechanism of LSTM elements to remember previous states, the forget gate determines the information to be retained, the input gate updates the cell state, and the output gate controls the output of the hidden state to simulate the evolution of pressure distribution under dynamic load. The key dynamic response calculation formula is based on Newton's second law and the strain energy principle.

[0064]

[0065] in It is a time-dependent stress, and E is the elastic modulus. It is a response. This is the viscous damping coefficient. The differential equation is solved using numerical integration methods such as the Newmark-β method, outputting a dynamic response curve including data on velocity, acceleration, and pressure changes over time. Then, the dynamic pressure prediction module takes the dynamic response curve output by the dynamic simulation module as input and performs predictions using time series analysis algorithms, including an autoregressive integral moving average model and a long short-term memory network. The autoregressive integral moving average model first performs d-order differencing on the data to make it stationary, and then fits the parameters of the autoregressive and moving average terms. The model expression is:

[0066]

[0067] Where B is the shift operator. and It is a polynomial. It is an observation sequence. It uses white noise. The Long Short-Term Memory (LSTM) network learns long-term dependencies through sigmoid (a sigmoid / logic function) and tanh (a hyperbolic tangent function) activation functions at the input, forget, and output gates, predicting future pressure peaks and fluctuation trends. Specific calculation steps include data normalization, sliding window feature extraction, and optimization of the loss function (such as mean squared error) through backpropagation. Finally, the probability of breakage is calculated based on the pressure peak and the artwork's fragility index. The key formula for calculating the probability of breakage is:

[0068]

[0069] in This is the predicted peak pressure. is the material damage threshold, and k is the sensitivity parameter, which is obtained through training with historical damage data. This reveals and solves the technical problem that dynamic loads generated by movement during the handling of artworks exacerbate pressure concentration points, leading to artwork damage. The beneficial effect of this embodiment is that it can accurately predict dynamic pressure changes and damage risks, reducing the probability of artwork damage during handling.

[0070] Third embodiment: as follows Figures 1 to 2 As shown, the path planning module further employs a multi-objective particle swarm optimization algorithm combined with a reinforcement learning framework. This module takes the output of the dynamic pressure prediction module as its main input, while also considering the environmental map, obstacle locations, and cargo priorities to generate a Pareto optimal solution set, including path coordinates and velocity curves. The path planning module also integrates a real-time replanning function, dynamically adjusting the path through sensor feedback to avoid collisions and vibration hotspots. The optimization objectives include minimizing maximum dynamic pressure, shortest time, and lowest energy consumption. The output results are transmitted to the control execution module in the form of path point sequences and control commands, and stored in a central log for subsequent analysis, ensuring safe and efficient operation even in challenging environments such as narrow passages or stacked artworks.

[0071] The control execution module further includes an instruction parsing submodule and an execution monitoring submodule. The instruction parsing submodule converts the path and speed data output by the path planning module into executable commands for the handling device, including directional control of the moving platform and lifting and extending / retracting operations of the forks. The execution monitoring submodule collects the device status, such as position, speed, and load, in real time through sensors, compares it with expected values, and adjusts parameters to eliminate deviations. The control execution module also integrates a fault detection and recovery mechanism. When an anomaly, such as overpressure or communication interruption, is detected, emergency protocols such as pause or rollback are automatically triggered to ensure the safety of the workpiece. All execution data is recorded and fed back to the central management system for system optimization and report generation.

[0072] During operation, based on the dynamic pressure prediction data output from the second embodiment, including the pressure peak time series, failure probability curve, and environmental constraint parameters, the path planning module first uses a multi-objective particle swarm optimization algorithm combined with a reinforcement learning framework to optimize the path and velocity. The multi-objective particle swarm optimization algorithm initializes the particle swarm, with each particle representing a potential path encoded as a position vector including coordinates and velocity values. A fitness function is defined based on multiple objectives: minimizing maximum dynamic pressure, minimizing transport time, and minimizing energy consumption. The maximum dynamic pressure objective function directly uses the failure probability data from the second embodiment, and the calculation formula is as follows:

[0073]

[0074] Where X is the particle position vector, and T is the time range. It is the probability of damage at time point t; other objective functions include transportation time. and energy consumption Where m is the mass of the goods. The velocity is used, and then the Pareto optimal solution set is constructed through non-dominated sorting and crowding distance calculation. During particle swarm optimization, particle velocity and position are updated using key formulas:

[0075]

[0076] in It is a particle The speed in the k-th iteration Here, w represents position, and w represents inertia weight. and It is a learning factor. and It is a random number. The best position in the particle's history is gbest, which is the globally optimal position selected from the Pareto front. The reinforcement learning part uses the particle swarm output as the initial policy and employs the Q-learning algorithm to iteratively optimize through a state-action value function. The state includes the current position, the predicted pressure value, and the environment map, while the actions include movement direction and speed adjustment. The Q-value update formula is:

[0077]

[0078] Where s is the current state, a is the action, and r is the reward function based on the objective function value. It's the learning rate. It is a discount factor. The next state involves outputting the optimal path coordinates and velocity curve after multiple iterations and convergence. Next, the control execution module receives the output from the path planning module. First, the instruction parsing submodule converts the path coordinates and velocity data into executable commands for the transport device, including directional control commands for the moving platform such as steering angle and speed setpoints, and lifting and extending commands for the cargo such as height and displacement. Then, the execution monitoring submodule collects real-time device status data such as encoder position, IMU acceleration, and load pressure through sensors, compares it with expected values, and uses a PID controller to adjust parameters to eliminate deviations. The control law is:

[0079]

[0080] Where u(t) is the control output and e(t) is the error. It uses proportional, integral, and derivative gains, and integrates fault detection and recovery mechanisms. When the pressure sensor detects overpressure or communication interruption, it automatically triggers emergency protocols such as pausing movement or returning to a safe point. All execution data is recorded and fed back to the central management system for closed-loop optimization. This reveals and solves the technical problem that existing handling systems do not consider dynamic pressure minimization in their path planning, resulting in high risk of damage to artworks and low efficiency. The beneficial effect of this embodiment is that it can generate safe and efficient handling paths, significantly reducing the risk of artwork damage and improving operational efficiency.

[0081] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0082] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart inventory and handling device for supply chain management, characterized in that: The mobile platform (1) is a four-wheel drive automated guided vehicle, and the mobile platform (1) has infrared obstacle avoidance and navigation functions. An RGB-D camera (2) is fixedly installed on the top front end of the mobile platform. The mobile platform (1) is internally embedded with a GPU processor, wherein the GPU processor in the mobile platform (1) is directly connected to the RGB-D camera (2) via a data cable; The bottom of the mobile platform (1) is fixedly connected to an electric fork (3). The electric fork (3) supports lifting and extension. The operation of the electric fork (3) is controlled by the GPU processor in the mobile platform (1) through a control line. The mobile platform (1) is internally embedded with a communication module, which is a Wi-Fi / 5G wireless transmission unit. The communication module is connected to the GPU processor in the mobile platform (1) via a circuit board and is used to synchronize data with the central management system in real time.

2. A smart inventory and handling system for supply chain management, using the smart inventory and handling device for supply chain management as described in claim 1, characterized in that: include: The image acquisition module's core function is to acquire images and depth data of high-value, fragile artworks in real time using an RGB-D camera on the handling device, including stacked shapes, surface textures, and three-dimensional coordinate information, providing raw data input for subsequent analysis. The core function of the data preprocessing module is to perform noise filtering, coordinate calibration, and data format standardization on the raw data transmitted by the image acquisition module to ensure data quality and consistency, making it easier for AI models to process. The core function of the artwork recognition module is to use a convolutional neural network to extract features and classify the preprocessed data, identify the type, shape and key contact points of the artwork, and output the recognition results for stress analysis. The core function of the pressure distribution estimation module is to calculate the pressure distribution at the contact points of stacked artworks based on the output of the artwork recognition module, using a generative adversarial network algorithm to generate data on pressure concentration areas and reveal potential damage risks. The core function of the dynamics simulation module is to use finite element analysis and recurrent neural networks to build dynamic models, simulate acceleration, vibration and load changes during the transportation process, and predict the evolution of pressure distribution under different motion states. The core function of the dynamic pressure prediction module is to combine the output of the dynamic simulation module with the time series analysis algorithm to calculate the dynamic pressure peak and assess the probability of damage to the artwork during handling. The core function of the path planning module is to use reinforcement learning and multi-objective particle swarm optimization algorithms to generate the optimal transport path and speed curve with the goal of minimizing the maximum dynamic pressure, while taking into account environmental constraints and cargo priority. The core function of the control and execution module is to control the movement of the handling device and the operation of the forks in real time according to the instructions of the path planning module, execute the handling task, and monitor the execution status to ensure the coordinated operation of the system.

3. The intelligent inventory and handling system for supply chain management according to claim 2, characterized in that: The image acquisition module further includes: The RGB-D camera captures images at 30 frames per second, covering a 120-degree field of view. It can capture multi-angle images and depth point cloud data of artworks. The captured data includes color information, depth distance, and 3D coordinates. The built-in sensor synchronizes the timestamp to ensure real-time data. The image acquisition module also integrates an ambient light adaptive function, automatically adjusting exposure and gain according to warehouse lighting conditions to avoid overexposure or dark areas affecting data quality. At the same time, the embedded processor performs preliminary data compression to reduce transmission load. All captured raw data is uploaded to the central management system's storage unit in real time via a wireless communication module, providing a complete and reliable input source for subsequent modules.

4. The intelligent inventory and handling system for supply chain management according to claim 3, characterized in that: The data preprocessing module further includes a data cleaning submodule and a coordinate transformation submodule. The data cleaning submodule uses median filtering and wavelet transform algorithms to remove noise and outliers from image and depth data, improving the signal-to-noise ratio. The coordinate transformation submodule converts the acquired 2D images and 3D point cloud data into a standardized format in the world coordinate system, facilitating subsequent AI model processing. The data preprocessing module also includes a data verification function, which checks data integrity through checksums and hash values ​​to prevent transmission errors. The preprocessed data is divided into training and test sets for model iteration and optimization. The module also integrates a caching mechanism to temporarily store processing results to cope with network latency.

5. A smart inventory and handling system for supply chain management according to claim 4, characterized in that: The artwork recognition module further employs a deep convolutional neural network model, which is pre-trained based on the ResNet architecture and adapted to high-value artwork datasets through transfer learning. This model achieves feature extraction of shape and texture, and the recognition process includes bounding box detection, semantic segmentation, and classification output. It can distinguish different types of artworks such as sculptures, paintings, and ceramics, and output the coordinates and confidence scores of key contact points. The artwork recognition module also integrates an adversarial example defense mechanism, which enhances the robustness of the model by generating adversarial networks to prevent misidentification. The recognition results are stored in a structured JSON format, including the artwork ID, shape parameters, and risk level, and are updated in real time to a central database, providing accurate input for the stress distribution estimation module.

6. The intelligent inventory and handling system for supply chain management according to claim 5, characterized in that: The pressure distribution estimation module further constructs a pressure mapping model based on a generative adversarial network algorithm. This model takes the contact point data output by the artwork recognition module as input, simulates the pressure distribution map through a generator network, and verifies its authenticity through a discriminator network. The output includes a pressure value matrix and a heat map, identifying areas of high pressure concentration. The pressure distribution estimation module also introduces physical constraints, such as the material elastic modulus and stacking angle, and uses finite element analysis to assist in calculations to improve estimation accuracy. The generated pressure data is compared with historical records for iterative training of the model. The output results are stored in vector form and trigger an alarm when the pressure exceeds a threshold, which is then directly used as input to the dynamic simulation module.

7. A smart inventory and handling system for supply chain management according to claim 6, characterized in that: The dynamics simulation module further employs a hybrid model combining finite element analysis and recurrent neural networks. Finite element analysis processes the material properties and geometry of the artwork to simulate stress distribution under static load, while the recurrent neural network learns time-series data to predict the effects of acceleration and vibration on pressure during handling. The dynamics simulation module takes pressure distribution data as input, solves the dynamic response through numerical integration and differential equations, and outputs curves including velocity, acceleration, and pressure changes. It also simulates extreme situations such as sudden stops or turns. All simulation results are stored in time steps for real-time prediction updates. The module also integrates sensitivity analysis to evaluate the impact of parameter changes on the results.

8. A smart inventory and handling system for supply chain management according to claim 7, characterized in that: The dynamic pressure prediction module is further based on time series analysis algorithms, including an autoregressive integral moving average model and a long short-term memory network. This module takes the dynamic data output by the dynamic simulation module as input, extracts features through sliding window technology, and predicts the pressure peak and fluctuation trend at future time points. The dynamic pressure prediction module also includes a risk assessment submodule, which calculates the probability of damage based on historical pressure data and the artwork fragility index. The output includes risk level and recommended measures. All prediction results are visualized in the form of charts and reports and pushed to the path planning module in real time for decision optimization. At the same time, the module supports online learning, adjusting model parameters according to new data to adapt to environmental changes such as uneven ground or load variations.

9. A smart inventory and handling system for supply chain management according to claim 8, characterized in that: The path planning module further employs a multi-objective particle swarm optimization algorithm combined with a reinforcement learning framework. This module takes the output of the dynamic pressure prediction module as its main input, while also considering the environmental map, obstacle locations, and cargo priorities to generate a Pareto optimal solution set, including path coordinates and velocity curves. The path planning module also integrates a real-time replanning function, dynamically adjusting the path through sensor feedback to avoid collisions and vibration hotspots. The optimization objectives include minimizing maximum dynamic pressure, shortest time, and lowest energy consumption. The output results are transmitted to the control execution module in the form of path point sequences and control commands, and stored in a central log for subsequent analysis.

10. A smart inventory and handling system for supply chain management according to claim 9, characterized in that: The control execution module further includes an instruction parsing submodule and an execution monitoring submodule. The instruction parsing submodule converts the path and speed data output by the path planning module into executable commands for the handling device, including directional control of the mobile platform and lifting and extending / retracting operations of the forks. The execution monitoring submodule collects the device status, such as position, speed, and load, in real time through sensors, compares it with expected values, and adjusts parameters to eliminate deviations. The control execution module also integrates a fault detection and recovery mechanism. When an anomaly, such as overpressure or communication interruption, is detected, an emergency protocol, such as pause or rollback, is automatically triggered. All execution data is recorded and fed back to the central management system for system optimization and report generation.

Images