Intelligent AI full-automatic bag storing and taking system

Abstract

The invention discloses an intelligent AI full-automatic package storing and taking system, particularly relates to the field of intelligent logistics and automatic storing and taking, and comprises a multi-mode sensing acquisition module, a dynamic feature recognition module, a real-time space construction module, an intelligent decision module, an automatic scheduling execution module and a data management monitoring module. According to the intelligent AI full-automatic packet access system, through the multi-mode perception acquisition module, the fine perception capability of physical attributes of a packet is improved, and the problem of insufficient space optimization caused by lack of fine perception of the physical attributes of the packet is directly relieved; a dynamic feature recognition module is used for modeling a sensor space relation and image semantic association, deep space optimization storage based on package features is directly supported, and the space utilization rate is increased; the intelligent decision-making module adopts a reinforcement learning and autonomous learning optimal decision-making strategy, so that the storage efficiency and the energy consumption can be dynamically balanced when concurrent requests are processed, and the path planning efficiency and the system throughput in a high-concurrency scene are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent logistics and automated storage and retrieval technology, and more specifically, to an intelligent AI fully automatic storage and retrieval system for packages. Background Art

[0002] With the development of intelligent warehousing and automated control technologies, the requirements for efficiency, intelligence, and space utilization of package storage and retrieval systems are becoming increasingly higher. In scenarios such as logistics, warehousing, and large-scale events, traditional package storage and retrieval methods rely on fixed storage architectures dominated by manual operations, consisting of physical, static cabinets and manual registration systems. This leads to problems such as manual sorting and searching making single operations too time-consuming, a sharp drop in system throughput during peak periods, and a high accident rate caused by manual operations.

[0003] To overcome the above-mentioned defects, existing technologies integrate automated execution technologies and introduce automated handling equipment to replace manual labor in the physical movement and basic sorting of packages; establish a basic digital management system to record simple information about packages and their corresponding fixed storage locations, which to a certain extent improves access efficiency, reduces some human errors, and enables the system to have a certain processing capability.

[0004] However, in actual use, it still has some shortcomings, such as the lack of detailed perception and intelligent analysis capabilities of the physical properties of the package, as well as real-time dynamic evaluation of the storage space status. Existing technologies make it difficult to achieve deep spatial optimization storage based on package characteristics, resulting in room for improvement in storage space utilization; it is difficult to cope with complex and changing package types and access requirements, especially when processing concurrent requests or non-standard packages, resulting in inefficient path planning and insufficient coordination of automated equipment, which in turn affects the overall access efficiency and system throughput. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides an intelligent AI fully automatic bag storage and retrieval system, which solves the problems raised in the above-mentioned background technology through the following scheme.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] An intelligent AI fully automatic storage and retrieval system, comprising:

[0008] Multimodal sensing and acquisition module: used for pre-deployed package collection equipment to obtain multimodal information of the target package;

[0009] Dynamic feature recognition module: used to analyze and process the multimodal information using a graph convolutional neural network to obtain the geometric and physical features corresponding to the target package;

[0010] Real-time spatial construction module: used to fuse storage space data based on voxelized occupancy grid maps and dynamically generate three-dimensional spatial heat maps;

[0011] Intelligent decision-making module: a deep reinforcement learning algorithm embedded in a spatiotemporal conflict prediction mechanism, and synchronously optimizes storage location, energy consumption, and path conflict probability based on the three-dimensional spatial heat map;

[0012] Automated Scheduling Execution Module: This module uses a multi-agent collaborative scheduling algorithm to decompose concurrent access requests into discrete subtasks and drives the robotic arm through digital threads to achieve access to target packages.

[0013] Data management and monitoring module: used to store and manage the multimodal information, package characteristics, space status, users and access requests, and use an intelligent data bus to realize information interaction between the modules and with external systems.

[0014] Preferably, the multimodal sensing and acquisition module, the multimodal information of the target package includes the size, shape, weight, center of gravity position, surface texture and physical identification of the target package.

[0015] Preferably, the dynamic feature recognition module uses a graph convolutional neural network to extract package features from the multimodal information, specifically including:

[0016] Receiving a graph structure composed of preprocessed multimodal information, wherein node features in the graph structure include the number of nodes and corresponding feature dimensions;

[0017] Based on the spatial position correlation of sensors and the similarity between image features provided by the multimodal perception acquisition module, the adjacency matrix A of the graph is dynamically constructed;

[0018] Stack multiple GCN layers, and each GCN layer is followed by a ReLU activation function, and feature aggregation operations are performed in each GCN layer;

[0019] Output the geometric features and physical features corresponding to the target package through parallel branches respectively;

[0020] The geometric features corresponding to the target package include the length, width, height of the package, a shape regularity score, and a contour curvature feature describing the surface details of the package;

[0021] The physical feature branch corresponding to the target package includes the material classification probability distribution, density estimation value and fragility score of the package.

[0022] Preferably, the dynamic feature recognition module dynamically constructs an adjacency matrix of the graph based on the sensor spatial position correlation and the similarity between image features provided by the multimodal perception acquisition module, specifically including:

[0023] Based on the spatial position of the sensors provided by the multimodal perception acquisition module, the spatial correlation weight w between the sensors is calculated. space , specifically expressed as:

[0024]

[0025] in, It is expressed as the spatial correlation weight between sensor i and sensor j, p i With p j They are the coordinates of sensor i and sensor j in three-dimensional space, σ is the scale parameter, and ‖·‖2 is the Euclidean distance;

[0026] Based on the features of each node in the graph structure, calculate the similarity weight w between image features image , specifically expressed as:

[0027]

[0028] in, Expressed as the similarity weight between node m and node n, f m With f n They are represented as the image feature vectors of node m and node n respectively, and ‖·‖2 is represented as the Euclidean distance;

[0029] The edge weights aw of the adjacency matrix of the graph are fused by geometric averaging, which is specifically expressed as:

[0030]

[0031] Preferably, in the intelligent decision-making module, the spatiotemporal conflicts in the spatiotemporal conflict prediction mechanism include storage location conflicts, path conflicts and sequence conflicts.

[0032] Preferably, the intelligent decision-making module synchronously optimizes the storage location based on the three-dimensional spatial heat map, specifically including:

[0033] The dynamic feature recognition module obtains the precise geometric and physical features of the package, and combines them with the storage density, access frequency, and device accessibility data in the 3D spatial heat map to analyze storage requirements and suitable storage area types.

[0034] The real-time storage space status and the characteristics of the package to be accessed are used as the environment state. Through the learned decision strategy, an execution signal is sent, and the execution signal determines the placement of the target package in a specific storage unit.

[0035] The reward is calculated according to the preset reward function, and the environment state and reward are fed back.

[0036] Technical effects and advantages of the present invention:

[0037] 1. This invention improves the ability to finely perceive the physical properties of packages through a multimodal sensing and acquisition module, directly alleviating the problem of insufficient spatial optimization caused by the lack of fine perception of the physical properties of packages;

[0038] 2. This invention uses a dynamic feature recognition module to model the spatial relationship between sensors and the semantic association of images, alleviating the robustness problem of feature extraction for irregular packages. It directly supports deep spatial optimization storage based on package features, improving space utilization.

[0039] 3. The present invention adopts reinforcement learning to autonomously learn the optimal decision-making strategy through the intelligent decision-making module, and combines spatiotemporal conflict prediction to avoid device coordination conflicts in advance. It can dynamically balance storage efficiency and energy consumption when processing concurrent requests, and significantly improve path planning efficiency and system throughput in high-concurrency scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a module block diagram of an intelligent AI fully automatic package storage and access system provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0042] The terms used in the following embodiments of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification of the present application, the singular expressions "one", "a kind of", "said", "above", "the" and "this" are intended to also include plural expressions, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in the present application refers to and includes any or all possible combinations of one or more listed items. In the description of the embodiments of the present application, unless otherwise stated, the meaning of "multiple" is two or more.

[0043] As attached Figure 1 The intelligent AI fully automatic package storage and access system shown includes a multimodal perception and acquisition module, a dynamic feature recognition module, a real-time space construction module, an intelligent decision-making module, an automatic scheduling execution module, and a data management and monitoring module.

[0044] The multimodal sensing and acquisition module is used by a pre-deployed package acquisition device to obtain multimodal information of the target package.

[0045] Specifically, the multimodal perception and acquisition module obtains the multi-dimensional physical properties and identification information of the target package through a variety of pre-deployed sensors, multispectral cameras and 3D laser scanners; the multimodal information of the target package includes the size, shape, weight, center of gravity position, surface texture and physical identification of the target package, where the physical identification includes but is not limited to barcodes, QR codes, etc.

[0046] In Example 1, four sets of 360-degree panoramic fisheye cameras were deployed on the warehouse entrance dome. These cameras have a high resolution of 3840×2160 and a frame rate of 60fps, forming a wide-range stereoscopic acquisition area with a coverage height of 0.2 meters to 3 meters. Linear scan industrial cameras were installed on both sides of the conveyor belt in the sorting area. These industrial cameras are configured with a pixel size of 3.45μm and capture images of moving packages at a linear rate of up to 2 meters / second, making them suitable for package identification on high-speed assembly lines.

[0047] It should be noted that in Example 1, in addition to the camera, various types of sensors are deployed to obtain the physical property information of the package; the sensors include but are not limited to piezoelectric weighing platforms, 3D laser scanners, RFID read-write arrays, etc.; the piezoelectric weighing platforms are intermittently embedded in key positions of the conveyor belt, with a range of 0.1kg to 50kg and an accuracy of up to ±1g; the 3D laser scanners are deployed at multiple points on the arched gantry above the package channel, with a resolution of 1000 points / cm 2 The package surface is scanned with a high point cloud density and a Z-axis accuracy of 0.1mm to generate high-precision three-dimensional point cloud data; the RFID read-write array is arranged in a ring at the boundary of the collection area, with a reading speed of 200 tags / second.

[0048] In Example 2, a dual sensing mechanism of laser light curtains and millimeter-wave radar is used to detect incoming packages. When any package enters the preset collection area, the dual sensing mechanism immediately triggers a multi-device synchronous startup protocol. The multi-device synchronous startup protocol instructs all collection devices such as cameras, laser scanners, and weighing platforms to start working synchronously in a very short time.

[0049] Furthermore, the camera starts multi-angle shooting, including multiple preset angles such as 0°, 45°, and 90°, to obtain images of different sides of the package and transmit RGB-D image streams. The laser scanner 0 emits a scanning beam to scan the surface of the package and uploads the generated point cloud data to the point cloud server. At the same time, the piezoelectric weighing platform activates the pressure sensing function, starts collecting the weight data of the package, and submits the weight distribution matrix to the data center. The RFID read-write array is activated synchronously to attempt to read the RFID tag information on the package.

[0050] The dynamic feature recognition module is used to analyze and process the multimodal information using a graph convolutional neural network to obtain the geometric and physical features corresponding to the target package.

[0051] Specifically, the dynamic feature recognition module accurately extracts the geometric and physical features of the target package by real-time analysis of the multimodal information of the multimodal perception and acquisition module using a graph convolutional neural network. In this embodiment, the packages are multi-dimensionally classified based on the geometric and physical features extracted by the dynamic feature recognition module, and associated with corresponding storage strategies, including but not limited to regular cubes and non-fragile items being classified into high-density stacking areas to maximize space utilization; long strips and metal materials are classified into magnetic shelf partitions and utilize special storage structures; irregular bodies and fragile items are classified into independent air cushion positions and stored and retrieved by low-speed robotic arms; flexible packages and low-density packages are classified into compressible cargo holds to further improve space utilization.

[0052] Furthermore, the multimodal information included in the target is preprocessed, and the preprocessing operation includes but is not limited to the processing of images, point cloud data, sensor data and sensor data; wherein, the image preprocessing includes but is not limited to image scaling to a uniform size, pixel value normalization, denoising, contrast enhancement, etc.; the point cloud data preprocessing includes but is not limited to point cloud downsampling, denoising, normal vector estimation, and conversion of point cloud data into a graph structure, etc.; the sensor data standardization includes but is not limited to standardizing numerical sensor data such as weight and size to have zero mean and unit variance, eliminating the influence of data of different dimensions; the identification information processing includes but is not limited to verifying and formatting the decoded barcode and QR code.

[0053] In one possible embodiment, constructing a feature extraction model corresponding to a target package based on the graph convolutional neural network includes: receiving a graph structure composed of preprocessed multimodal information, wherein the node features in the graph structure include the number of nodes and the corresponding feature dimensions, and dynamically constructing an adjacency matrix A of the graph based on the sensor spatial position correlation and the similarity between image features provided by the multimodal perception acquisition module; stacking multiple GCN layers, and each GCN layer is followed by a ReLU activation function, and performing a feature aggregation operation on each GCN layer; outputting the geometric features and physical features corresponding to the target package through parallel branches; the geometric features corresponding to the target package include the length, width, height of the package, a shape regularity score, and a contour curvature feature describing the surface details of the package; the physical features corresponding to the target package include the material classification probability distribution, density estimation value, and fragility score of the package.

[0054] It should be noted that the node features include image features, point cloud features, sensor values, etc.; the adjacency matrix describes the connection relationship between each node and its weight; the shape regularity score ranges from 0 to 1, with 1 indicating complete regularity; the material classification probability distribution indicates the probability that the target package belongs to categories such as cardboard, plastic, and metal; the fragility score ranges from 0 to 1, with 1 indicating extremely fragile.

[0055] In a possible embodiment, constructing a feature extraction model corresponding to the target package based on the graph convolutional neural network further includes: calculating the spatial correlation weight w between sensors based on the sensor spatial position provided by the multimodal perception acquisition module space , specifically expressed as:

[0056]

[0057] in, It is expressed as the spatial correlation weight between sensor i and sensor j, p i With p j are the coordinates of sensor i and sensor j in three-dimensional space, and σ is a scale parameter, which in this embodiment is taken as 1 / 2 of the average distance of the sensor network, and ‖·‖2 is the Euclidean distance. Based on the features of each node in the graph structure, the similarity weight w between the image features is calculated. image , specifically expressed as:

[0058]

[0059] in, Expressed as the similarity weight between node m and node n, f m With f n are respectively represented as the image feature vectors of node m and node n, ‖·‖2 is represented as the Euclidean distance; the edge weights aw of the adjacency matrix of the graph are fused by geometric averaging, specifically expressed as:

[0060]

[0061] It should be noted that the calculation result of the similarity weight between the image features has a value range of [-1, 1] and needs to be mapped to [0, 1] through linear transformation; the elements of the adjacency matrix are determined by setting a similarity threshold. When the similarity exceeds the preset similarity threshold, it is judged that there is an association between the packages and the value is 1, otherwise it is 0.

[0062] Furthermore, the calculation formula for the feature aggregation operation performed by each layer of GCN is specifically expressed as:

[0063]

[0064] Among them, H (l) Represented as the node feature matrix of the lth layer, H (l+1) Represented as the node feature matrix of the l+1 layer, W (l) Represented as a preset weight matrix, Represented as an adjacency matrix with self-connection added, Expressed as The corresponding degree matrix.

[0065] The real-time space construction module is used to fuse storage space data based on the voxelized occupancy grid map and dynamically generate a three-dimensional space heat map.

[0066] In the second embodiment, the voxelized occupancy grid map of the real-time space construction module uses 10cm×10cm×10cm as a basic unit, and updates the unit occupancy status in real time, wherein the unit occupancy status includes occupied, idle, and unknown.

[0067] Specifically, the three-dimensional heat map is not a single-dimensional representation, but a superposition of multiple key heat indicators. Each indicator is obtained through a preset calculation formula and distinguished by different visual codes. The three-dimensional heat map contains at least three core thermal dimensions, which include but are not limited to storage density heat, access frequency heat, device accessibility heat, etc.

[0068] In the third embodiment, the core thermal dimensions of the three-dimensional spatial heat map include storage density heat, access frequency heat, and device accessibility heat. The storage density heat is calculated by calculating the sum of the ratios of the volumes of all packages in a preset area to the total volume of voxels in the area, reflecting the degree of filling of the preset area space. The storage density heat is visually encoded using a red gradient, with darker red indicating higher storage density. The access frequency heat represents the total number of times packages in the preset area are accessed per unit time. The access frequency heat is visually encoded using a blue gradient, with darker blue indicating more frequent package access in the area. The device accessibility heat measures the difficulty of automated equipment reaching the preset area. The shorter the path and the fewer turns, the higher the accessibility. The device accessibility heat is visually encoded using a green gradient, with darker green indicating better physical accessibility of the area to the equipment.

[0069] The intelligent decision-making module is used to embed a deep reinforcement learning algorithm with a spatiotemporal conflict prediction mechanism, and to analyze time, energy consumption and path conflict probability.

[0070] Specifically, the intelligent decision-making module receives the package features from the dynamic feature recognition module and the three-dimensional spatial heat map provided by the real-time space construction module, and based on the deep reinforcement learning algorithm embedded with the spatiotemporal conflict prediction mechanism, simultaneously optimizes the storage location of the package, the operating energy consumption of the automated equipment, and the potential path conflict probability.

[0071] Furthermore, the implementation of the intelligent decision-making module includes the application of deep reinforcement learning algorithm and spatiotemporal conflict prediction mechanism, optimization target setting based on three-dimensional spatial heat map, storage location planning process, and evaluation and optimization of energy consumption and path conflict probability.

[0072] Furthermore, the application of the deep reinforcement learning algorithm and the spatiotemporal conflict prediction mechanism includes: selecting a deep reinforcement learning algorithm to handle decision-making problems, and learning the optimal decision-making strategy through interaction with the environmental state without the need to pre-set complex rules; before planning any storage or package retrieval operation, the spatiotemporal conflict prediction mechanism predicts the spatiotemporal conflicts that will occur within a preset time based on the current real-time posture of all automated equipment, the expected motion trajectory, and the number and type of packages to be processed.

[0073] It should be noted that the environmental status includes package characteristics, storage space status, device location, etc., and the decision-making strategy is to select the storage location and plan the package retrieval path; through a large number of trial and error and reward signals, the reward signals include but are not limited to successful access, improved space utilization, reduced energy consumption, conflict avoidance, etc.; the spatiotemporal conflicts include storage location conflicts, path conflicts and sequence conflicts; the location conflict is represented by the possibility of predicting that multiple packages are planned to the same storage unit, or the possibility of grasping and placing operations of different devices in the same storage area; the path conflict is represented by the prediction that the motion paths of multiple robotic arms will intersect or overlap in the same time period, resulting in collision or deadlock; the sequence conflict means predicting the execution order of different tasks, resulting in inefficiency or resource competition.

[0074] Furthermore, the optimization goal setting based on the three-dimensional spatial heat map includes: using the three-dimensional spatial heat map to guide decision-making and optimization goals, the optimization goals include simultaneously optimizing storage location, energy consumption and path conflict probability, and constructing a reward function through a weighted combination of optimization goals. By maximizing the reward function, the deep reinforcement learning algorithm can learn.

[0075] In the fourth embodiment, the storage location optimization reward includes a positive reward for successfully placing a package in an area with low storage density or a high degree of match with the package characteristics; a penalty will be imposed for failing to effectively utilize the space; the energy consumption optimization reward includes a positive reward for selecting a storage location or path with a short distance and a smooth movement trajectory; a penalty will be imposed for long-distance movement or complex operations; the path conflict avoidance reward includes a positive reward for a planned path that is predicted to have a low conflict probability; a penalty will be imposed for a high conflict probability.

[0076] Furthermore, the storage location planning process based on package characteristics and spatial heat maps includes: obtaining the precise geometric and physical characteristics of the package provided by the dynamic feature recognition module, combining the storage density heat, access frequency heat and device accessibility heat in the three-dimensional spatial heat map to analyze storage requirements and the type of area suitable for storage; the storage requirements include but are not limited to independent storage, stacking, proximity to entrances and exits, etc.; using the real-time storage space status and the characteristics of the package to be accessed as the environmental state, sending an execution signal through the learned decision-making strategy, and the execution signal determines the placement of the target package in a specific storage unit; calculating the reward according to the preset reward function, and feeding back the environmental state and reward.

[0077] Furthermore, the assessment and optimization of energy consumption and path conflict probability involves establishing an energy consumption evaluation model to quantify the impact of different decision-making options on equipment energy consumption. Factors influencing equipment energy consumption include the type of automated equipment, travel distance, acceleration, deceleration, robot joint motion angles and torques, and the complexity of the grasping or placing operation. By combining a spatiotemporal conflict prediction mechanism with path information from a three-dimensional heat map, the conflict probability of different path planning solutions is evaluated. Conflict factors include analyzing the likelihood of multiple robots passing through the same area within the same time period, path intersections, and possible sequence conflicts.

[0078] The automated scheduling execution module is used to decompose concurrent access requests into discrete subtasks using a multi-agent collaborative scheduling algorithm, and drive the robotic arm through digital threads to achieve access to the target package.

[0079] Furthermore, the multi-agent collaborative scheduling algorithm of the automated scheduling execution module includes: a Hungarian algorithm to solve the robotic arm task allocation matrix, a spatiotemporal conflict resolver using time window constraint programming, and an activation of a deformation compensation controller for non-standard packages.

[0080] Specifically, the Hungarian algorithm for solving the robot arm task allocation matrix includes: decomposing multiple concurrent access requests from the intelligent decision-making module into discrete subtasks; using the Hungarian algorithm to construct and solve the robot arm task allocation matrix, the elements of the task allocation matrix represent the benefits of a specific robot arm combination performing subtasks; the spatiotemporal conflict resolver uses time window constraint planning, including: using the time window constraint planning method to plan an initial path for each assigned subtask based on the device's motion capabilities and environmental map, and identifying potential conflict points through spatiotemporal analysis, and allocating a time window for each device in the conflict area that is only allowed to pass within a specified time period; the activation of the deformation compensation controller for non-standard packages includes: when there is a package with a high deformation coefficient and a package classified as requiring special treatment, during the task allocation and path planning stage, activating the deformation compensation controller for the robot arm performing the target task.

[0081] The data management and monitoring module is used to store and manage the multimodal information, package characteristics, space status, users and access requests, and uses an intelligent data bus to realize information interaction between the modules and with external systems.

[0082] Secondly: The drawings of the embodiments disclosed in the present invention only involve structures related to the embodiments disclosed in the present invention. Other structures may refer to conventional designs. The same embodiment and different embodiments of the present invention may be combined with each other without conflict.

[0083] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent AI fully automatic bag storage and retrieval system, characterized by: include: Multimodal sensing and acquisition module: used for pre-deployed package collection equipment to obtain multimodal information of the target package; Dynamic feature recognition module: used to analyze and process the multimodal information using a graph convolutional neural network to obtain the geometric and physical features corresponding to the target package; Real-time spatial construction module: used to fuse storage space data based on voxelized occupancy grid maps and dynamically generate three-dimensional spatial heat maps; Intelligent decision-making module: a deep reinforcement learning algorithm embedded in a spatiotemporal conflict prediction mechanism, and synchronously optimizes storage location, energy consumption, and path conflict probability based on the three-dimensional spatial heat map; Automated Scheduling Execution Module: This module uses a multi-agent collaborative scheduling algorithm to decompose concurrent access requests into discrete subtasks and drives the robotic arm through digital threads to achieve access to target packages. Data management and monitoring module: used to store and manage the multimodal information, package characteristics, space status, users and access requests, and use an intelligent data bus to realize information interaction between the modules and with external systems.

2. The intelligent AI fully automatic bag storage and retrieval system according to claim 1, characterized in that: The multimodal sensing and acquisition module collects the multimodal information of the target package, including the size, shape, weight, center of gravity position, surface texture, and physical identification of the target package.

3. The intelligent AI fully automatic bag storage and retrieval system according to claim 1, characterized in that: The dynamic feature recognition module uses a graph convolutional neural network to extract package features from the multimodal information, specifically including: Receiving a graph structure composed of preprocessed multimodal information, wherein node features in the graph structure include the number of nodes and corresponding feature dimensions; Based on the spatial position correlation of sensors and the similarity between image features provided by the multimodal perception acquisition module, the adjacency matrix A of the graph is dynamically constructed; Stack multiple GCN layers, and each GCN layer is followed by a ReLU activation function, and feature aggregation operations are performed in each GCN layer; Output the geometric features and physical features corresponding to the target package through parallel branches respectively; The geometric features corresponding to the target package include the length, width, height of the package, a shape regularity score, and a contour curvature feature describing the surface details of the package; The physical feature branch corresponding to the target package includes the material classification probability distribution, density estimation value and fragility score of the package.

4. The intelligent AI fully automatic bag storage and retrieval system according to claim 3, characterized in that: The dynamic feature recognition module dynamically constructs the adjacency matrix of the graph based on the sensor spatial position correlation and the similarity between image features provided by the multimodal perception acquisition module, specifically including: Based on the spatial position of the sensors provided by the multimodal perception acquisition module, the spatial correlation weight w between the sensors is calculated. space , specifically expressed as: in, It is expressed as the spatial correlation weight between sensor i and sensor j, p i With p j They are the coordinates of sensor i and sensor j in three-dimensional space, σ is the scale parameter, and ‖·‖2 is the Euclidean distance; Based on the features of each node in the graph structure, calculate the similarity weight w between image features image , specifically expressed as: in, Expressed as the similarity weight between node m and node n, f m With f n They are represented as the image feature vectors of node m and node n respectively, and ‖·‖2 is represented as the Euclidean distance; The edge weights aw of the adjacency matrix of the graph are fused by geometric averaging, which is specifically expressed as:

5. The intelligent AI fully automatic bag storage and retrieval system according to claim 1, characterized in that: In the intelligent decision-making module, the spatiotemporal conflicts in the spatiotemporal conflict prediction mechanism include storage location conflicts, path conflicts and sequence conflicts.

6. The intelligent AI fully automatic bag storage and retrieval system according to claim 1, characterized in that: The intelligent decision-making module synchronously optimizes the storage location based on the three-dimensional spatial heat map, specifically including: The dynamic feature recognition module obtains the precise geometric and physical features of the package, and combines them with the storage density, access frequency, and device accessibility data in the 3D spatial heat map to analyze storage requirements and suitable storage area types. The real-time storage space status and the characteristics of the package to be accessed are used as the environment state. Through the learned decision strategy, an execution signal is sent, and the execution signal determines the placement of the target package in a specific storage unit. The reward is calculated according to the preset reward function, and the environment state and reward are fed back.

Images