Intelligent warehouse management system, method and program based on passive light-emitting RFID tag

Abstract

The invention relates to the technical field of warehouse management, in particular to an intelligent warehouse management system and method based on a passive light-emitting RFID tag and a computer program, and the system comprises an environment sensing unit, an integrated hardware module, the passive light-emitting RFID tag, an edge computing module, an AI algorithm unit, a cloud service platform and handheld terminal equipment. Storage environment data and personnel operation behaviors are collected, spatial-temporal feature expression is constructed in combination with a graph neural network and a time sequence prediction model, and bin allocation and material demand prediction are adaptively optimized based on a reinforcement learning mechanism. The system has the characteristics of battery-free visual indication, edge network disconnection fault tolerance, cloud edge cooperative intelligent scheduling and the like, can significantly improve the warehousing operation efficiency, accuracy and system reliability, and is suitable for multi-scene large-scale warehousing intelligent management.

Description

Technical Field

[0001] This application relates to the technical field of warehouse management, and in particular to an intelligent warehouse management system, method and program based on passive luminescent RFID tags, specifically integrating technical means such as environmental perception, edge computing, multi-antenna readers, and AI algorithm units to achieve comprehensive collection, analysis and dynamic optimization allocation of warehouse environmental data and material information, aiming to improve the efficiency, accuracy and sustainability of warehouse operations. Background Art

[0002] With the continuous expansion of the scale of modern warehouses and the rapid development of the logistics industry, enterprises' requirements for refined, automated and intelligent warehouse management are becoming increasingly prominent. Traditional warehouse operation processes usually rely on manual labor and paper documents, or only use primary identification means such as barcode scanning, and it is difficult to cope with complex and changeable inventory structures and frequent inbound and outbound operations. As a result, solutions using radio frequency identification (RFID) technology, Internet of Things (IoT) devices, and various automated guiding systems have gradually emerged in the market to improve warehouse efficiency and accuracy. However, the existing technologies still have deficiencies in the following aspects: insufficient dynamic perception and fusion ability of environmental data, warehouse management is vulnerable to network conditions, the consumables and maintenance costs of active RFID or light bar solutions are relatively high, and there is a lack of in-depth analysis of turnover rate and personnel operation behavior data, etc.

[0003] On the one hand, regarding the application of RFID technology in the warehouse environment, Chinese patent document CN119204951A proposes a warehouse management method, system and device based on RFID electronic tags. This solution sets active tags at the shelves or storage locations to facilitate real-time positioning and information update. However, this solution requires installing batteries or connecting to a power source on the tags. Once the warehouse scale is large, the replacement of batteries or the laying of lines will bring relatively high maintenance costs and hardware transformation difficulties, especially in the scenarios of high shelves or three-dimensional warehouses. In addition, the lifespan and power consumption problems of active tags will also cause them to be frequently replaced and unable to achieve long-term stable operation. At the same time, this patent document mainly focuses on transmitting position and identification information from the tags to the upper-level system, and lacks detailed description of the acquisition and comprehensive utilization of warehouse environmental data (such as temperature and humidity, illuminance, noise level) and personnel behavior data.

[0004] In addition, passive RFID tags are used to reduce the burden on energy supply, but most rely on fixed-position readers for batch scanning and identification. Although this method improves the picking efficiency to a certain extent, there are the following problems: First, due to the limited coverage radius of fixed readers and they are often installed in some local positions of the warehouse, it is difficult to effectively cope with complex warehouse layouts; Second, if the warehouse environment changes, such as the adjustment of shelf positions or heights, or the increase in the storage density of materials, signal blind spots or missed scans may occur.

[0005] In addition, in Chinese Patent Document CN118778471A, a method for designing an industrial Internet intelligent manufacturing system architecture is disclosed, which introduces a prototype system based on edge computing; establishes a virtual three-dimensional model corresponding to the real physical workshop; establishes a three-dimensional model of the production line; conducts simulation design on the production line; establishes a logistics system, and uses the production line after simulation design to plan the logistics transportation line; conducts simulation design on the logistics transportation line; conducts system analysis on the simulation design of the production line and the simulation design of the logistics transportation line; establishes a remote virtual monitoring system and a fault analysis system. However, this solution mainly focuses on how to perform caching and simple rule judgment at the edge nodes, and there is no sufficient description on how to perform more complex intelligent analysis (such as machine learning, deep learning or reinforcement learning) on the edge side and dynamically optimize the warehousing environment. More notably, the existing technical literature generally lacks the integration of the two major factors of "warehousing environment perception" and "personnel operation behavior": most systems only collect information on the incoming and outgoing of materials, lack the systematic integration of environmental parameters (temperature, humidity, illuminance, noise, space occupancy rate, etc.), and also ignore the role of such behavioral data as the picking path, picking preference and error rate of operators in optimizing the warehousing layout.

[0006] In addition, many current intelligent warehousing solutions also have the problem of "excessive network dependence". Once the cloud platform has an anomaly or the communication is interrupted, the edge-side devices often can only wait for the network to recover and cannot continue most of the core operations; once the operator needs to continue working in an offline environment, there may be a lack of necessary data support or real-time instructions. At the same time, in some application scenarios, even if the environmental perception data (noise sensors, temperature and humidity sensors, shelf load detection sensors, etc.) is collected, it is only uploaded to the cloud intact for simple monitoring, and there is no close linkage with the picking strategy, location allocation or material demand prediction of the warehouse.

[0007] In view of the defects or deficiencies in the above-mentioned background technology, how to further integrate the dynamic transmission coverage ability of multi-antenna readers on the basis of the low maintenance cost and convenience of passive RFID technology, combine the real-time monitoring of environmental data, and introduce a more intelligent algorithm framework (such as deep learning, reinforcement learning) on the edge side, so as to keep the core functions running under unstable or disconnected network conditions, has become an urgent problem to be solved in the industry.

[0008] In the exploration of this problem, some literature or solutions also attempt to use buzzers or warning lights in the warehouse to assist in addressing, but usually simply place independent prompting devices on the shelves (or issue simple warnings on handheld terminals). This approach cannot fully utilize the two-way communication characteristics between RFID tags and readers; moreover, the installation of conventional buzzers / light bars often requires additional wiring or battery power supply, which does not meet the requirements of low cost, low energy consumption, and sustainable operation. At the same time, readers mostly use fixed power and fixed directions, and it is difficult to balance the needs of full coverage and signal interference avoidance in tall or irregular three-dimensional warehouses, and phenomena such as unreadable tags, repeated tag readings, or incorrect tag readings are likely to occur, affecting the accuracy of warehousing data.

[0009] In addition, from the perspective of intelligent decision-making, traditional machine learning or simple rule engines mostly only target the historical turnover data of materials and are insufficient to consider various dynamic factors (seasonal demand fluctuations, environmental changes, personnel scheduling differences, etc.); even if some solutions introduce big data analysis modules in the cloud, due to insufficient real-time performance or scenario adaptability, it is still difficult to make a rapid response when facing sudden demands or network fluctuations at the warehousing site. More importantly, in complex warehousing scenarios, different materials have different sensitivities to environmental conditions such as temperature, humidity, and light, and the corresponding picking frequencies may also vary significantly due to personnel operation preferences. The lack of unified modeling of these factors often makes the warehouse location allocation strategy lack fineness and cannot continuously maintain the optimal or near-optimal state. Summary of the Invention

[0010] To overcome the above problems and further improve the efficiency and flexibility of warehousing management, this application proposes an intelligent warehousing management system based on passive luminous RFID tags, which combines core components such as multi-antenna readers, buzzers, environmental perception units, and edge computing modules, and integrates deep learning or graph neural networks and reinforcement learning mechanisms in the AI algorithm unit. This system can not only achieve battery-free visual positioning through passive luminous RFID tags to guide operators to quickly find the target storage location, but also dynamically optimize the transmission power or direction of the reader based on environmental sensors, operation behavior data, and material historical turnover records, and even perform warehouse location allocation and demand prediction on the edge side or in the cloud, greatly reducing the warehousing operation and maintenance costs and improving the operation accuracy and efficiency.

[0011] To achieve the above objectives, the present invention adopts the following technical solutions:

[0012] An intelligent warehousing management system based on passive luminous RFID tags, characterized in that it includes:

[0013] 1) A warehousing environment perception unit, including a number of sensors for collecting environmental data such as temperature, humidity, illuminance, noise level, and shelf vacancy rate in the warehouse;

[0014] 2) An integrated hardware module, which is set on the top of the warehouse or in key areas, and the integrated hardware module includes:

[0015] a) A multi-antenna card reader, which can dynamically adjust the transmission power or direction according to the real-time data provided by the environment perception unit to efficiently activate passive luminous RFID tags scattered in different areas;

[0016] b) A buzzer, which can automatically adjust the acoustic prompt intensity or frequency according to the noise level collected by the environment perception unit;

[0017] 3) Passive luminous RFID tags, which are distributed on each storage position or material package. The RFID tags rely on capturing the radio waves emitted by the multi-antenna card reader to achieve passive power supply and luminous indication, and can distinguish material status or warning information through different colors or LED flashing modes;

[0018] 4) An edge computing module, which is connected to the integrated hardware module and the environment perception unit, and is used for local data analysis and preprocessing based on the real-time collected environment and material information, and independently completes the execution and monitoring of warehousing operation instructions when the network is abnormal;

[0019] 5) An AI algorithm unit, which is deployed on the edge computing module or the cloud service platform, integrates material historical turnover records, environmental data and personnel operation behaviors, and adaptively optimizes the storage position allocation strategy and predicts material requirements through deep learning or machine learning models;

[0020] 6) A cloud service platform, which is communicatively connected to the edge computing module and the handheld terminal device, is used for collecting global warehousing data, performing big data analysis and generating a visual warehousing operation report, and forming a collaborative working mode with the edge computing;

[0021] 7) A handheld terminal device, which interacts with the cloud service platform and the edge computing module, displays the path planning, warehousing optimization instructions and the status information of passive luminous RFID tags provided by the AI algorithm unit through a graphical interface, and transmits the on-site execution results back to the cloud.

[0022] Preferably, the environment perception unit includes a noise sensor, a temperature and humidity sensor, a light intensity sensor and a ground load detection sensor to meet the various scenario requirements of refined warehousing management.

[0023] Preferably, the multi-antenna card reader adopts a rotatable or linear array antenna structure and can automatically adjust the angle according to the height or layout of the shelves in the warehouse to reduce signal blind spots.

[0024] Preferably, the acoustic prompt of the buzzer can be remotely configured by the handheld terminal device, and different tones or durations can be emitted for instructions with different priorities or urgencies.

[0025] Preferably, the passive luminous RFID tag is internally provided with a reflective layer or a high-brightness LED module, and a semi-transparent window can be designed at the edge of the tag to enhance the light scattering effect.

[0026] Preferably, the handheld terminal device has a touch screen interface and a voice prompt function, and can display the color, blinking mode or inventory information of the passive luminous RFID tag in real time, providing multimodal guidance for the operator.

[0027] Preferably, the edge computing module is equipped with local storage and anomaly detection functions. When there is a network failure or delay in the cloud service platform, it can independently complete the specified warehousing operations and cache the data locally;

[0028] Preferably, the AI algorithm unit can automatically switch the learning model according to the change of material demand in different time periods, predict seasonal or periodic fluctuations in advance and optimize the stocking strategy.

[0029] Preferably, the AI algorithm unit combines a graph neural network and a time series prediction model to extract material turnover and environmental features, and realizes the adaptive optimization of bin allocation and demand prediction through a reinforcement learning mechanism; the graph neural network adopts a graph convolutional network (GCN), and the time series prediction model is a long short-term memory network (LSTM); the graph neural network constructs a spatio-temporal graph based on the warehousing environment topology, regards the bins and channels as nodes and edges, and fuses time series data to capture the temporal changes of material demand and the interaction relationship between nodes; the reinforcement learning policy module adopts an Actor-Critic structure, and the reinforcement learning mechanism takes the bin rearrangement or material batch allocation as an action (Action), the current environmental state, node information and time series prediction result of the warehouse as a state (State), and sets a reward (Reward) according to the operation efficiency, error rate or cost target.

[0030] Preferably, within the time interval {1, 2, …, T}, the warehouse is discretely sampled or monitored, and various feature vectors of the bin and channel nodes are recorded at each sampling moment; where:

[0031] 1) Warehousing environment topology

[0032] Represents the spatio-temporal graph of the warehousing,

[0033] Where:

[0034] Denotes the set of nodes, including bin nodes and channel nodes, and N is the total number of nodes;

[0035] Denotes the set of edges, which is used to describe the adjacency or reachability relationship between bins or between bins and channels;

[0036] A ∈ R N×N : The adjacency matrix of the graph. If there is a path or a connection with a short distance between node v i and v j , then A i,j = 1, otherwise it is 0 or a corresponding weight is set according to distance attenuation;

[0037] 2) Nodes and time series features

[0038] Let x vi,t ∈ R dx denote the feature vector of node v i at time t, which includes:

[0039] Material turnover information, including the number of inbound and outbound times, inventory quantity, and residence duration;

[0040] Environmental data, including temperature, humidity, noise, and illuminance sensed by sensors, combined with the average value of the location or neighboring locations to which this node belongs;

[0041] Operation behavior data, including the operation duration, error rate, and picking frequency of the operator near this storage location;

[0042] At each moment t, the node feature matrix X t ∈ R N×dx is obtained; the length of the time series is T, that is, there is {X1, X2, …, X T};

[0043] 3) Graph neural network layer

[0044] In the graph neural network, a hidden representation is introduced for each node at each moment t where l represents the l-th layer of GNN; the graph convolution or graph attention mechanism is expressed as:

[0045]

[0046] where Aggregate(.) represents the neighborhood aggregation function;

[0047] 4) Temporal prediction module

[0048] After capturing the graph structure at the same moment, the changing trend across moments is then modeled; based on the output of the GNN at each moment t, the temporal model LSTM is stacked to express the dynamic evolution:

[0049] z t = RNN(Readout(H t ))

[0050] where represents the representation after concatenation or pooling of all node vectors output by the GNN at the L-th layer, RNN is LSTM, and z t represents the temporal hidden state of the entire graph at time t, which is used to predict the material requirements and environmental state at the next time or in several future time periods;

[0051] 5) Reinforcement learning

[0052] The reinforcement learning mechanism takes the re-layout of storage positions or the allocation of material batches as the actions of the Agent; it is defined as follows:

[0053] State: S t includes the overall storage state at the current moment, which can be specifically composed of the following information:

[0054] The global representation z extracted by GNN + temporal model t ;

[0055] The current environmental parameters, including average temperature, noise, and congestion degree of key channels;

[0056] Key performance indicators, including inventory turnover rate, error rate, and operation delay;

[0057] Action: A t includes scheduling decisions for storage positions or materials, including:

[0058] Re-layout of storage positions: Adjust material A from shelf X to shelf Y;

[0059] Recommendation of picking path: Plan the shortest or least conflict operation path for operators;

[0060] Material transfer: Pre-adjust high-frequency materials to high-heat areas;

[0061] Activation of buzzer and label status: Select which buzzers and passive luminous RFID tags emit light for auxiliary positioning;

[0062] Reward: Rt is determined by a pre-defined multi-objective function, comprehensively considering operation efficiency, cost, or reflecting requirements in terms of environmental safety or inventory risk;

[0063] Reinforcement learning updates the policy by maximizing the cumulative reward ∑tγtRt (γ ∈ (0,1) is the discount factor), so that the system gradually approaches the global optimal or sub-optimal layout.

[0064] Preferably, the system further includes a positioning cooperation unit for:

[0065] a) An acoustic positioning module that calculates the approximate area of the target based on the buzzer sound source intensity and the warehouse sound field model;

[0066] b) An optical positioning module that obtains the spatial coordinates or incident angles of the luminous RFID tags;

[0067] c) A fusion calculation module that performs weighted fusion of the acoustic area and the optical coordinates or performs triangulation to output the three-dimensional coordinates or shelf numbers of the target storage locations.

[0068] Preferably, the cloud service platform can be docked with the enterprise's ERP, MES, or WMS system to achieve cross-system collaboration and information sharing through APIs or data interfaces;

[0069] Preferably, the system supports modular expansion, and the edge computing module can flexibly upgrade the hardware configuration or add additional sensor interfaces according to the storage scale and business requirements.

[0070] Furthermore, the present invention also discloses an intelligent warehouse management method based on passive luminous RFID tags. This method uses the above-mentioned system and includes the following steps:

[0071] S1 Environmental data collection: The environmental perception unit is used to obtain the noise level, temperature and humidity, illuminance, and shelf vacancy rate in the warehouse in real time;

[0072] S2 Instruction reception: The integrated hardware module receives the material requisition or picking instructions sent by the warehouse management system, and the buzzer adapts the acoustic prompt according to the noise level;

[0073] S3 Wireless activation and tag identification: The multi-antenna reader dynamically adjusts the direction or power of the antenna according to the environmental data, emits radio waves to the passive luminous RFID tags in the warehouse, and completes the reading of the tag information;

[0074] S4 Edge computing processing: The edge computing module analyzes and preprocesses the data of the reader and the environmental perception unit, and independently executes the operation process in case of network anomalies;

[0075] S5 AI optimization decision-making: The AI algorithm unit trains and predicts the warehouse data, generates an operation optimization plan, and synchronizes the results to the handheld terminal device;

[0076] S6 Terminal execution and feedback: The handheld terminal device displays the status of the passive luminous RFID tags and the task instructions. After the operator completes the material operation, the results are transmitted back to the edge computing module and the cloud service platform;

[0077] S7 Data Recording and Visualization: The cloud service platform integrates the data from edge computing and handheld terminals, updates the warehousing database, and generates visual reports to achieve a continuous optimization closed-loop.

[0078] Preferably, in the step (S3), the light-emitting color or blinking frequency of the passive light-emitting RFID tag is dynamically set by the integrated hardware module according to different material types or risk levels.

[0079] Preferably, in the step (S6), the handheld terminal device can also perform further data verification with the passive light-emitting RFID tag by scanning the code or automatic identification to prevent incorrect picking or misaligned storage.

[0080] Preferably, in the step (S4), if the edge computing module detects that the cloud service platform is unavailable, it automatically enters the offline mode and records all operation logs for later synchronization to the cloud;

[0081] Preferably, in the step (S7), the cloud service platform performs big data analysis on the data accumulated by the edge computing module to discover potential operation bottlenecks and continuously iteratively optimize the AI algorithm.

[0082] Furthermore, the present invention also discloses a computer-readable storage medium, on which a computer program or instruction is stored, and when the computer program or instruction is executed by a processor, the method is implemented.

[0083] Furthermore, the present invention also discloses a computer program product, including a computer program or instruction, and when the computer program or instruction is executed by a processor, the method is implemented.

[0084] Due to the adoption of the above technical solution, the present invention combines passive light-emitting RFID technology with a multi-antenna reader, an environmental perception unit, edge computing, and an AI algorithm, and can achieve efficient, reliable, and low-power intelligent management in a warehousing environment, specifically reflected in the following aspects:

[0085] 1. Passive Visualization and Low-Cost Maintenance: The passive light-emitting RFID tag obtains energy by capturing the radio waves emitted by the reader, eliminating the need for wiring and battery replacement, reducing the equipment maintenance and transformation costs; using different LED colors or blinking modes can achieve rapid visual indication of the storage location, guiding the operators to quickly locate the target goods and significantly improving the operation efficiency.

[0086] 2. Multi-Antenna Reader and Environment Adaptation: The multi-antenna reader can adjust the transmission power or direction in real time according to the environmental perception unit (such as temperature and humidity, noise, illuminance, etc.), reducing signal blind spots and interference; in dynamic or high-density warehousing scenarios, the reading accuracy and coverage can be optimized, significantly improving the accuracy and stability of data collection.

[0087] 3. Edge Computing and Network Fault Tolerance: The edge computing module performs data preprocessing and intelligent decision-making locally. It can independently execute critical warehousing instructions and monitor anomalies when the network is abnormal, avoiding job stagnation in traditional systems when disconnected from the cloud. Through offline caching and local storage, it ensures that the core functions can still operate continuously in case of network disconnection or high latency, improving the reliability and resilience of the system.

[0088] 4. Deep Integration of AI Algorithms: Introduce material historical turnover records, environmental data, and personnel operation behaviors into deep learning or graph neural network models, and combine with reinforcement learning mechanisms to achieve adaptive optimization of bin allocation and material requirements. According to actual demand peaks, personnel picking habits, and environmental conditions, the AI algorithm can dynamically predict the optimal inventory provision and operation path, reducing misoperations, increasing turnover rates, and lowering operating costs.

[0089] 5. Multi-Mode Interaction and Full-Process Collaboration: Through the dual cooperation of the acoustic indication of the buzzer and the optical visualization of passive luminous RFID tags, it provides operators with a multi-modal picking navigation experience. The handheld terminal device links the cloud platform and the edge computing module to maintain the visualization and real-time nature of warehousing information, constructing a closed-loop management mode to ensure the smooth connection of the full process from data collection to execution feedback.

[0090] In summary, the present invention helps to quickly and accurately complete material tracking, scheduling, and demand forecasting in large-scale or complex warehousing scenarios. It not only effectively reduces maintenance costs and energy consumption but also significantly improves the overall operation efficiency and system reliability, fully demonstrating the technical advantages in modern intelligent warehousing applications. Brief Description of the Drawings

[0091] Figure 1 It is a schematic structural diagram of the system of the present invention.

[0092] Figure 2 It is a flow chart of the intelligent warehousing management method of the present invention

[0093] Figure 3 It is a flow chart of module data interaction of the present invention.

[0094] Figure 4 It is a structural diagram of the AI model of the present invention.

[0095] Figure 5 It is a logic diagram of bin layout and luminous indication of the present invention Detailed Embodiments

[0096] Combined with the embodiments of the present invention, the technical solutions in the embodiments will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present invention.

[0097] The present invention proposes an intelligent warehouse management system based on passive luminous RFID tags, which integrates a multi-antenna reader, an environmental perception unit, an edge computing module, and an AI algorithm unit to achieve intelligent tracking and scheduling of materials. Figure 1 (Schematic structural diagram) shows the interconnection relationship between the components of the system. Figure 5 It is a bin layout and luminous indication logic diagram. For the sake of convenience of description, Figure 1 only the main modules of the system and the data flow are schematically shown. When actually deployed, the components can be added, deleted, or improved according to the warehouse scale and operation mode.

[0098] 1. Warehouse environmental perception unit

[0099] The warehouse environmental perception unit is composed of several sensors, including temperature and humidity sensors, illuminance sensors, noise sensors, and ground load detection sensors, etc. The sensors are connected to the edge computing module or the integrated hardware module in a wired or wireless (such as ZigBee, Wi-Fi) manner to collect the environmental data inside the warehouse in real time. In a large-scale three-dimensional warehouse or multi-floor warehouse, several groups of sensors can be deployed on different floors or partitions to ensure that more comprehensive environmental information can be obtained.

[0100] 2. Integrated hardware module

[0101] The integrated hardware module mainly includes a multi-antenna reader and a buzzer, which can be installed on the top of the warehouse or key areas (such as the intersection of the main channels, the end of the shelf with a higher height).

[0102] Multi-antenna reader: Adopting a rotatable or linear array antenna structure, it can dynamically adjust the transmission power and angle according to the real-time data provided by the environmental perception unit (such as noise level, shelf distribution, etc.) to ensure maximum coverage and reduce signal blind spots.

[0103] Buzzer: Its acoustic prompt intensity or frequency can be automatically adjusted according to the noise level measured by the noise sensor, or can also be remotely configured by a handheld terminal to emit acoustic prompts with different tones or durations for instructions of different emergency levels (such as picking high-priority materials or error prompts).

[0104] 3. Passive luminous RFID tag

[0105] Passive luminous RFID tags are distributed on each storage bin or material package, obtaining energy by capturing the radio waves emitted by the reader, thereby achieving passive power supply. When activated, the tags can emit optical signals of different colors or blinking patterns to indicate the current bin status, material type, or risk warning. A reflective layer or a high-brightness LED module can be optionally configured inside the tag, and a translucent window is designed at the edge of the shell to enhance the light scattering effect, facilitating operators to observe the luminous status of the tag from different angles.

[0106] 4. Edge computing module

[0107] The edge computing module is connected to the integrated hardware module and the environmental perception unit. Its main functions include:

[0108] 1) Localized data processing and analysis: Initially aggregate, clean, and perform anomaly detection on the data collected by the reader and sensors.

[0109] 2) Network outage fault tolerance: When there are network failures or high latencies in the cloud service platform, it can still independently execute specified warehousing operations, such as activating luminous RFID tags and triggering acoustic prompts, and temporarily store the data locally.

[0110] 3) Edge inference: According to implementation requirements, a simplified AI model or rule engine can be deployed here to achieve rapid decision-making and instruction issuance for the operation scenario.

[0111] 4. AI algorithm unit

[0112] The AI algorithm unit can be deployed on the edge computing module or the cloud service platform. Using deep learning, graph neural network (GNN), and reinforcement learning (RL) mechanisms, comprehensively analyze the warehousing environment and historical material data:

[0113] 1) Graph neural network (GNN): Model the warehousing topology as a spatio-temporal graph, associate bin positions, channels, and sensor nodes, and capture the interaction relationships among material turnover rate, operation paths, and environmental changes in the spatio-temporal dimension.

[0114] 2) Reinforcement learning (RL): Consider "bin layout" or "material batch allocation" as actions, and the current environmental state of the warehouse (temperature, humidity, noise, personnel operation load, etc.), node information (shelf occupancy rate, historical turnover rate), and time series prediction results as states. Construct rewards by setting goals such as operation efficiency, error rate, energy consumption, or cost, thereby adaptively optimizing the warehousing scheduling strategy.

[0115] 5. Cloud service platform

[0116] The cloud service platform is communicatively connected to the edge computing module and the handheld terminal device, and is used to collect global warehousing data, perform big data analysis, and generate a visualized warehousing operation report. When the network is available, the cloud platform can integrate and iteratively upgrade the model parameters or rules of the edge nodes, and send the better training results to the edge module to form a cloud-edge collaborative working mode.

[0117] 6. Handheld Terminal Device

[0118] The handheld terminal device interacts with the cloud service platform and the edge computing module, and displays the path planning, warehousing optimization instructions, and passive optical RFID tag status information provided by the AI algorithm unit through a graphical interface. When performing picking or shelving operations, the terminal operator can quickly locate the target shelf or material according to the guidance of the terminal, and upload the actual execution result to the cloud or the edge side through the terminal device after the operation is completed to form a closed-loop management.

[0119] 7. Location Collaboration Unit (Optional)

[0120] Furthermore, the present invention may further include a location collaboration unit for:

[0121] a) An acoustic location module that calculates the approximate area of the target based on the sound source intensity of the buzzer and the warehouse sound field model;

[0122] b) An optical location module that obtains the spatial coordinates or incident angle of the optical RFID tag;

[0123] c) A fusion calculation module that weights and fuses the acoustic area and the optical coordinates or performs triangulation to output the three-dimensional coordinates or shelf number of the target warehousing location.

[0124] The location collaboration unit may include arranging more than four high-sensitivity microphones near each integrated hardware module (including the reader and the buzzer) to form a small microphone array. The relative coordinates of the array position and the buzzer are measured in advance and written into the system configuration. And installing more than two industrial cameras on the top of the warehouse or in the shelf interlayer. The cameras are pre-calibrated for internal and external parameters and aligned with the warehouse coordinate system. The camera orientations cover the target storage location area and can capture the LED-emitting RFID tags in real time.

[0125] The following combines Figure 2 with Figure 3 (Flow diagram and interaction diagram) to illustrate the implementation process of the system of the present invention in a typical operation scenario. Figure 2 Schematically shows the main steps of the warehousing management method of the present invention, Figure 3 while schematically showing the data interaction relationship between the modules.

[0126] (I) Environmental Data Collection and Initialization (S1)

[0127] When the system starts up, various sensors in the warehousing environment perception unit, such as temperature and humidity sensors, illuminance sensors, noise sensors, and ground load sensors, start to work and send the measured data to the edge computing module.

[0128] In the case of a large warehouse scale or many regions, integrated hardware modules and several environmental sensors can be separately set in multiple regions to build a partitioned and hierarchical data acquisition network.

[0129] In the initial stage, the cloud service platform can, according to historical warehousing records (such as the turnover status in the past quarter or year), and in coordination with the layout of on-site sensors, preliminarily set parameters such as the basic power of the multi-antenna reader and the antenna angle range.

[0130] (2) Instruction reception and operation trigger (S2)

[0131] When the warehousing management system (such as ERP, MES, or WMS) sends a material requisition or picking requirement to this system, the instruction can be transmitted to the edge computing module through the cloud platform or directly. If an acoustic prompt is required, the edge computing module controls the buzzer to adjust the volume or tone according to the real-time data of the noise sensor and the type of operation instruction. For high-priority and urgent alarms, a high-decibel or longer-duration mode can be adopted to ensure that it can be detected even in a noisy environment.

[0132] (3) Wireless activation and tag identification (S3)

[0133] According to the current warehouse environment conditions (such as shelf layout, noise level, temperature and humidity, illuminance, etc.), the multi-antenna reader can call the pre-set adaptive strategy to adjust the reading power or antenna angle for directional scanning of the target partition or shelf area. The reader emits ultra-high frequency radio waves to the target area, and once the passive luminous RFID tag captures enough energy, it is activated and sends a feedback signal.

[0134] The edge computing module compares the tag feedback information with the bin location or material ID recorded in the warehousing database to confirm the specific shelf location or material status. The LED on the tag or the different-colored outer shell can emit light synchronously to provide visual positioning and guide the operator to quickly lock the target.

[0135] (4) Edge computing processing and network disconnection fault tolerance (S4)

[0136] After receiving the tag identification result, the edge computing module locally stores and preliminarily analyzes the key data (tag ID, activation time, position coordinates, environmental parameters, etc.).

[0137] If the network communication is normal, the data can be synchronously uploaded to the cloud service platform for large-scale statistics and management; if high latency or a fault is detected in the cloud network, the edge computing module immediately enters the offline mode, queues the job instructions locally, and completes subsequent processing (such as activating other tags, guiding operators to replenish or relocate).

[0138] During the offline period, the system's job history, sensor data, personnel operation records, etc. will all be cached at the edge side and then batch-synchronized to the cloud after the network is restored, ensuring data integrity and continuity.

[0139] (V) AI Optimization Decision-making (S5)

[0140] When the system is in the online state, the AI algorithm unit can run a deep training model on the cloud service platform, comprehensively analyze a large amount of historical turnover data, environmental records, and personnel operation behaviors, and output a series of optimization strategies (such as the priority of material placement, suggestions for re-layout of storage positions, path planning, etc.).

[0141] As Figure 4 shown, for the problem of fusing multi-dimensional data (material turnover rate, environmental sensor data, personnel operation behaviors, warehouse layout, etc.) in a modern warehousing environment, the present invention uses a graph neural network to model the "warehousing environment topology", abstracts storage positions, channels, etc. as nodes and edges, and captures demand changes and state evolutions in the dimension of time series; subsequently, using reinforcement learning, "re-layout of storage positions or material batch allocation" is regarded as a decision-making action, and by balancing multiple objectives such as real-time operation efficiency, error rate, or cost, a reward signal (Reward) is obtained when the strategy is updated, so as to continuously iteratively optimize the storage position allocation strategy and the material demand prediction result. Through this method, adaptive regulation of the warehousing system in a complex environment can be achieved.

[0142] For the convenience of explanation, assume that within the time interval {1, 2, …, T}, we perform discrete sampling or monitoring on the warehouse, and record various feature vectors of storage position and channel nodes at each sampling moment. The following gives the definitions of the main mathematical symbols:

[0143] 1. Representation of Warehousing Environment Topology

[0144] represents the spatio-temporal graph of the warehouse, where:

[0145] represents the set of nodes (including storage position nodes and channel nodes), and N is the total number of nodes;

[0146] represents the set of edges, which is used to describe the adjacency or reachability relationship between storage positions or between storage positions and channels.

[0147] A ∈ R N×N : The adjacency matrix of the graph. If there is a path or a connection with a relatively short distance between nodes v i and v j , then A i,j = 1, otherwise it is 0 or a corresponding weight is set according to distance attenuation.

[0148] 2. Nodes and Time-Series Features

[0149] Let x vi,t ∈ R dx represent the feature vector of node v i at time t, which includes:

[0150] Material turnover information (such as the number of inbound and outbound times, inventory quantity, residence duration, etc.);

[0151] Environmental data (temperature, humidity, noise, illuminance, etc. sensed by sensors, combined with the average value of the location where this node belongs or the nearby locations);

[0152] Operation behavior data (operation duration, error rate, picking frequency, etc. of the operator near this storage location).

[0153] Therefore, we can obtain the node feature matrix X t ∈ R N×dx at each moment t; the length of the time series is T, that is, we have {X1, X2, …, X T}.

[0154] 3. Graph Neural Network Layer (GNN Layer)

[0155] In the graph neural network, we introduce a hidden representation for each node at each moment t, where l represents the l-th layer of GNN. The graph convolution or graph attention mechanism can be expressed as:

[0156]

[0157] where Aggregate(.) represents the neighborhood aggregation function, such as the implementation methods of GCN, GAT, GraphSAGE, etc.

[0158] GNN is used to learn the mutual influence relationship of nodes in the topological structure and capture the interaction features between each bin or between bin - channel at the same moment.

[0159] 4. Temporal Prediction Module

[0160] After capturing the graph structure at the same moment, we also need to model the change trend across moments. Based on the output of GNN at each moment t, a temporal model (LSTM) can be stacked to express the dynamic evolution:

[0161] z t = RNN(Readout(H t ))

[0162] where represents the representation after concatenating or pooling all node vectors output by the GNN at the L-th layer. RNN is LSTM, and z t represents the temporal hidden state of the entire graph at time t, which is used to predict the material requirements and environmental status at the next time or in several future time periods.

[0163] 5. Definition of Reinforcement Learning

[0164] In the present invention, the reinforcement learning mechanism takes "warehouse layout rearrangement or material batch allocation" as the action of the Agent. For clear description, the following are defined:

[0165] State: S t includes the overall warehouse state at the current moment, which can be specifically composed of the following information:

[0166] The global representation z extracted by the GNN + temporal model t ;

[0167] The current environmental parameters (such as average temperature, noise, congestion degree of key channels, etc.);

[0168] Key Performance Indicators (KPIs) such as inventory turnover rate, error rate, operation delay.

[0169] Action: A t includes scheduling decisions for the warehouse or materials, such as:

[0170] Which materials need to be moved to which warehouse;

[0171] Which batches of materials should be given priority for picking or replenishment;

[0172] Whether to perform emergency scheduling (such as prophylactically transferring sensitive materials to a more suitable temperature area);

[0173] Buzzer, label status activation: Select which buzzers and passive luminous RFID tags emit light for auxiliary positioning.

[0174] Reward: R t Is determined by a predefined multi-objective function, which can comprehensively consider operation efficiency (the number of picking tasks completed per unit time, error rate, etc.), cost or expense (handling distance, energy consumption, time overhead, etc.), and can also reflect requirements in aspects such as environmental safety or inventory risk.

[0175] Reinforcement learning maximizes the cumulative reward ∑t γ t R t (γ ∈ (0, 1) is the discount factor) to update the policy, making the system gradually approach the global optimal or sub - optimal layout.

[0176] 6. Network Structure

[0177] 1) Graph Neural Network Layer

[0178] To efficiently process the warehouse topology, the following GNN modules are often selected:

[0179] Graph Convolution Network (GCN): Uses regularized neighborhood aggregation so that the representation of each node contains features from neighboring nodes;

[0180] Graph Attention Network (GAT): Utilizes the attention mechanism to assign edge weights, giving higher weights to important neighbors, thus more flexibly capturing the influence of key nodes.

[0181] In the present invention, since the warehouse is large - scale and has a large number of nodes, a partition - based / sampling - based GNN (such as Graph SAGE) is preferentially adopted to improve the training and inference efficiency.

[0182] 2) Temporal Prediction Layer

[0183] The common method is that at each time step, the representation vectors obtained by passing all nodes through the GNN (or the vectors after a certain global pooling operation on them) are input into the LSTM / GRU network.

[0184] Input: Pool can be average pooling, max pooling, or attention pooling;

[0185] Taking LSTM as the update equation:

[0186] i t = σ(W i [Z t , h t-1 +b i ), f t = σ(W f [Z t , h t-1 +b f ),

[0187] o t = σ(W o [Z t , h t-1 +b o ),

[0188] h t = o t ☉tanh(c t ).

[0189] where h t and c t are the hidden state and cell state respectively, σ is the Sigmoid function, is the hyperbolic tangent function, W * and b* are learnable parameters.

[0190] 3) Reinforcement learning policy network

[0191] Combining GNN + time series model with RL, using the Actor-Critic structure:

[0192] In the AI algorithm unit of the present invention, the Actor-Critic structure is a dual-network architecture for optimizing reinforcement learning policies, suitable for dealing with complex policy problems in continuous state spaces, especially suitable for multi-objective and real-time decision-making tasks such as warehouse management. Its core idea is:

[0193] Actor network: Responsible for generating scheduling policies (i.e., outputting actions for bin allocation or job paths);

[0194] Critic network: Responsible for evaluating the quality of the current state or action (i.e., giving value estimates) and guiding the Actor to improve the policy.

[0195] This structure can balance the exploration and evaluation of policy optimization, has higher convergence efficiency and stability, and is suitable for deploying lightweight and large-scale training versions on the edge side and cloud side respectively.

[0196] a) State space S

[0197] The state vector is composed of the following information (extracted by GNN + LSTM / GRU):

[0198] S t = [z t , e t , k t ,

[0199] where:

[0200] z t : The global encoding output by the graph neural network and the time series model at time t;

[0201] e t : The current environmental state (temperature and humidity, noise level, shelf vacancy rate, etc.);

[0202] k t : Current key KPI status (inventory turnover rate, picking error rate, warehouse congestion, etc.);

[0203] b) Action space A

[0204] The action is an optimization suggestion for the warehousing task, including:

[0205] Warehouse location rearrangement: Adjust material A from shelf X to shelf Y;

[0206] Picking path recommendation: Plan the shortest or least conflict operation path for the operator;

[0207] Material transfer: Pre-transfer high-frequency materials to high-heat areas;

[0208] Buzzer and label status activation: Select which buzzers and passive RFID tags emit light for auxiliary positioning.

[0209] Mathematical form:

[0210] A t = π θ (S t ) → {a1, a2,..., a n},

[0211] where π θ is the Actor network parameter.

[0212] c) Reward function R t

[0213] The reward design needs to consider multi-objective optimization and can be defined as a weighted combination:

[0214] R t = ω1 · R eff + ω2 · R acc + ω3 · R cost

[0215] where:

[0216] R eff : Efficiency reward, such as the number of tasks completed per unit time;

[0217] R acc : Accuracy reward, such as a decrease in the picking error rate;

[0218] R cost : Cost negative reward, such as handling distance and congestion conflict.

[0219] d) Specific implementation process of Actor-Critic

[0220] Step1: Initialize the network structure

[0221] Actor network: Receive S using an MLP network t , and output the action distribution π θ (S t ).

[0222] Critic network: Receive S t or (S t , A t ), and output the state value V(S t ) or the action value Q(S t , A t ).

[0223] Step2: Interactive sampling

[0224] Generate the action A using the current policy π θ ; t

[0225] Execute the action in the simulation warehouse or the real environment to obtain the new state S t+1 and the reward R t ;

[0226] Store it as an experience tuple (S t , A t , R t , S t+1 ).

[0227] Step3: Network update

[0228] Update the Critic network (minimize the TD error):

[0229] L critic = (R t + γV(S t+1 ) - V(S t )) 2

[0230] Update the Actor network (using the advantage value A t ):

[0231] L actor = -logπ θ (A t |S t ) · A t

[0232] where the advantage value:

[0233] A t = R t + γV(S t+1 ) - V(S t )

[0234] ​An entropy term can be added to enhance exploration:

[0235]

[0236] Reinforcement learning updates the policy θ (including the parameters of GNN, time series model, and Actor-Critic) in each decision-making cycle. If advanced algorithms such as Proximal Policy Optimization (PPO) or Soft Actor-Critic (SAC) are adopted, better stability and convergence speed can be maintained in complex large-scale warehousing environments.

[0237] 7. Training and Inference Processes

[0238] The following takes the offline training + online deployment mode as an example to illustrate the operation process of this algorithm in actual implementation.

[0239] 1) Offline Training Phase

[0240] Data Collection: Through the warehousing operation process over a long period (such as several weeks or months), a large amount of sensor data, material turnover records, and operation behavior logs are collected to construct a spatio-temporal dataset {(X t , A t , R t )}. Where A t is the action taken in historical operations, and R t is the corresponding actual benefit or metric; if the historical data does not contain action / reward information, a simulation environment needs to be constructed for interactive sampling.

[0241] Initial Training: In a cloud server or high-performance computing environment, use the dataset to perform end-to-end training on the GNN + time series model and the reinforcement learning policy. Specifically, two methods can be adopted:

[0242] a) Behavior cloning based on real decision records + reinforcement learning fine-tuning;

[0243] b) Pure reinforcement learning + simulation environment (iteratively train in the simulator, and then fine-tune in the real environment after obtaining a better policy).

[0244] Model Verification: By comparing with a benchmark scheme (such as a fixed layout, scheduling based on simple rules), judge whether the policy has been improved in terms of metrics such as the accuracy of material demand prediction, operation efficiency, and error rate.

[0245] 2) Online Deployment Phase

[0246] Cloud-Edge Collaboration: Send the trained model and weights to the edge computing module to enable it to have independent inference capabilities. When the network is unobstructed, the edge-side data can be sent back to the cloud, and the global model can be updated in an incremental learning manner.

[0247] Real-time inference: After the warehouse enters the running state, at each time step (or detection period), the node information at the current moment is processed by the GNN, and the hidden state h of the past several moments is combined through the RNN / LSTM t-1 , c t-1 etc., to obtain the global state. The reinforcement learning policy network calculates the optimal action (selects the allocation plan) accordingly, and then executes it in the real warehouse.

[0248] Feedback update: After the action is executed, the environment, material distribution, and operation efficiency of the warehouse will generate new numerical changes, which are used as the input state for the next moment and update the reward to continuously correct or strengthen the existing policy. When offline or when the scheduling is relatively idle, periodic retraining or adaptive update of the model parameters can be performed.

[0249] The above AI algorithm unit adopts a technical route that combines graph neural network + time series prediction + reinforcement learning. It not only has high efficiency and scalability in large-scale warehousing scenarios, but also can make reasonable scheduling judgments when there are environmental fluctuations, sudden demand surges or network failures, thus significantly improving the robustness and actual operation performance of the warehousing system. Especially for those warehouses with a wide variety of materials, wide distribution areas, and frequent inventory dynamics, the multi-dimensional fusion ability of this solution is particularly prominent, and it can bring significant improvement in operation efficiency for enterprises under the comprehensive optimization of multiple objectives. Whether in the fields of manufacturing, retail, e-commerce, pharmaceutical cold chain, intelligent logistics hub, etc., the technical solution described in the present invention can be used as a key intelligent upgrade means to help achieve real intelligent warehousing management.

[0250] 8. Implementation of the positioning cooperation unit

[0251] 1. System hardware deployment

[0252] 1) Buzzer and acoustic positioning sensor array

[0253] Around each integrated hardware module (with known position, coordinates (x b , y b , z b ), evenly arrange 4 or more microphones (acoustic sensors), and the coordinates are The buzzer emits a known signal s(t) (such as a 20kHz sine pulse or a specific frequency spectrum), and the microphone sampling frequency f s ≥8 high frequencies to ensure good time resolution.

[0254] 2) Optical positioning hardware

[0255] Deploy at least two industrial cameras or optoelectronic sensors on the top or ceiling of the warehouse. The internal parameters of camera i are Ki, and the external parameters (rotation matrix R i , translation vector t i)It has been obtained through calibration. The passive RFID tag has a built-in LED, and its emission width is sufficient to be captured by a conventional lens. The camera can obtain a clear image at the moment of emission and identify the pixel coordinates u of the tag in the image i =(u i ,v i ).

[0256] 2. Acoustic positioning module

[0257] 2.1 Time difference measurement (TDOA) → Direction of arrival estimation

[0258] 1) Signal capture: Each microphone M i records the received signal r i (t).

[0259] 2) Cross-correlation calculation: Calculate the cross-correlation function for each pair of microphones (i, j)

[0260] R ij (τ) = ∫r i (t)r j (t + τ)dt,

[0261] The peak position of which is the time difference estimation between the two sensors.

[0262] 3) Time difference → Direction solution

[0263] Assume the speed of sound c, for a group of microphones, the direction of arrival (elevation angle θ, azimuth angle φ) of the sound source relative to the buzzer can be solved to satisfy:[[]]

[0264]

[0265] is the direction vector, and the direction vector is solved by the least squares method

[0266] 4) Acoustic area estimation

[0267] Taking the buzzer position b = (x b ,y b ,z b ) as the starting point, construct an "acoustic ray"

[0268]

[0269] On the warehouse floor or shelf level (assuming z = z shelf ), the intersection interval of the ray is the approximate acoustic area.

[0270] 3. Optical positioning module

[0271] 3.1 Image coordinates → Inverse projection of spatial ray

[0272] 1) Image capture and pixel recognition: Camera i captures an image at the moment of illumination and detects the tag point u i =(u i , v i ).

[0273] 2) Back-projection to camera coordinates

[0274] Let the homogeneous pixel coordinates Then the direction vector in the central coordinate system of camera i:

[0275]

[0276] 3) Spatial ray equation

[0277] The camera optical center c i (known), then the tag position should be located on

[0278] L op，i (s)=c i +sv i , s > 0,

[0279] 4) Multi-camera triangulation

[0280] For the set of rays {L op,i} from two or more cameras, find the point with the minimum distance:

[0281]

[0282] Solve using linear least squares or SVD to obtain the optically estimated coordinates,

[0283]

[0284] 4. Fusion calculation module

[0285] To balance the respective accuracies and coverage of acoustics and optics, weighted fusion or least squares joint estimation is adopted.

[0286] 4.1 Weighted optimization

[0287] Define the weighted loss between the point on the acoustic ray and the optical estimate:

[0288]

[0289] where d(p, L ac ) represents the shortest distance from the point to the acoustic ray, and α ∈ [0, 1] is the weight (which can be dynamically adjusted according to environmental noise and light intensity).

[0290] Optimal fusion position: Can be solved analytically or by gradient descent.

[0291] 4.2 Triangulation + Convergence

[0292] The search range can also be constrained to the acoustic ray neighborhood in optical triangulation to accelerate the calculation and improve robustness.

[0293] 5. Shelf Number Mapping

[0294] 1) Map the fused three-dimensional coordinates to the warehouse storage grid:

[0295] Pre-define the regional bounding box of each shelf in the database as {(X min , X max , Y min , Y max , Z min , Z max )}.

[0296] 2) Search for the shelf number that satisfies

[0297] X min ≤ x ≤ X max , Y min ≤ y ≤ Y max , Z min ≤ z ≤ Z max , which is the final positioning result.

[0298] (VI) Terminal Execution and Feedback (S6)

[0299] After receiving the instructions sent by the edge computing module or the cloud service platform, the handheld terminal device will notify the operator in the form of a graphical interface, voice guidance, or vibration prompt. The interface can display the floor and area of the target shelf or material, as well as the color or blinking mode currently displayed by the passive luminous RFID tag.

[0300] The operator performs operations such as picking, shelving, inventory taking, or stock transfer according to the prompts, and scans the code or automatically identifies through the handheld terminal device for further data verification with the RFID tag. When the tag does not match the system expectation, the handheld terminal can issue a warning to prevent incorrect operations. After the operation is completed, the handheld terminal device will send the feedback information (picking completion time, actual quantity, personnel ID, etc.) to the edge computing module and finally archive it in the cloud service platform.

[0301] (VII) Data Recording and Visualization (S7)

[0302] After receiving the combined data from the edge computing module and the handheld terminal, the cloud service platform forms a complete operation log and the updated content of the warehousing database.

[0303] Managers can access the cloud service platform through a browser or a client to view visual reports, such as the picking efficiency, storage location occupancy rate, error rate, and label activation status on the same day. If the AI algorithm unit detects areas for improvement in this operation (for example, repeatedly activating the same area multiple times but with poor results), a new optimization plan can be generated in the next cycle.

[0304] At the same time, the cloud service platform will feedback the key performance indicators (KPIs) to the AI algorithm unit, which serves as a reference for the reward function of the reinforcement learning module, laying a foundation for the next round of scheduling optimization.

[0305] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0306] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of processes and / or blocks in the flowchart and / or block diagram can also be implemented. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0307] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured product including an instruction device, and the instruction device implements the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0308] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Therefore, the instructions executed on the computer or other programmable device provide for implementing the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1Steps of the functions specified in one or more boxes.

[0309] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0310] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM), and / or non-volatile memory such as read-only memory (ROM) or flash RAM. The memory is an example of computer-readable media.

[0311] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can store information by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0312] Test Examples

[0313] The following are experimental examples and results based on the intelligent warehouse management system described in the present invention to illustrate the technical effects of the present invention in practical applications. This data can be used to compare the differences between the prior art and the present invention in key indicators such as efficiency, accuracy rate, and maintenance cost, thereby proving that the present invention has significant improvement effects.

[0314] I. Experimental Scenario and Design

[0315] 1. General Situation of the Experimental Warehouse

[0316] With an area of approximately 3000 square meters and a maximum shelf height of 7 meters, it mainly stores electronic accessories, general materials, and several temperature and humidity sensitive materials. Multiple antenna readers and several environmental sensors (temperature and humidity, noise, illuminance, load detection, etc.) are installed in the warehouse, and passive luminous RFID tags are deployed at key storage locations. The edge computing device is connected to the cloud server to achieve online / offline collaboration of the AI algorithm unit.

[0317] 2. Comparison Scheme

[0318] Control group: Traditional hand-held scanning + fixed-power RFID reader mode; without environmental adaptive adjustment and AI dynamic bin allocation, relying only on regular manual adjustment of material positions.

[0319] This invention: Adopt the intelligent warehousing management system of this invention, including:

[0320] Visual positioning of passive luminous RFID tags;

[0321] The multi-antenna reader dynamically adjusts the power and angle according to the environmental sensor data;

[0322] Edge computing + cloud AI (GNN + LSTM + reinforcement learning) adaptively optimizes bin allocation and material requirements.

[0323] Measurement period

[0324] 3. The experiment lasts for 4 weeks, and the two schemes are executed alternately or in partitions: In the first and second weeks, the control group management is enabled; in the third and fourth weeks, the scheme of this invention is enabled, and indicators such as picking efficiency, error rate, maintenance cost, and latency are tracked and statistically analyzed.

[0325] II. Key indicators and calculation methods

[0326] 1. Picking efficiency

[0327] Statistically count the number of picking tasks completed per unit time, either per capita or the total system volume.

[0328] 2. Picking error rate

[0329] The ratio of picking the wrong materials, missing picks, or over-picking.

[0330] 3. Maintenance cost

[0331] Including the comprehensive costs of hardware transformation of readers and tags, wiring, manual inspection, battery replacement (if any), etc.

[0332] 4. Average activation / recognition latency

[0333] The average time from when the reader issues an activation instruction to when the RFID tag is recognized and gives feedback.

[0334] 5. Bin utilization rate

[0335] The reasonable degree of material turnover within the unit shelf space, which can be used to evaluate the effectiveness of warehouse space optimization.

[0336] 6. Operational rate during network disconnection

[0337] Refers to the proportion of core functions that the system can independently execute on the edge side when the network is disconnected. If the warehousing operation is not affected, it indicates strong edge fault tolerance.

[0338] III. Comparison of Experimental Data

[0339] The following table shows the comparison results of the control group and the present invention in the main indicators:

[0340] Index Control Group The Present Invention Improvement or Change Order Picking Efficiency (pieces / person·hour) 45～50 58～64 ↑By approximately 20 - 30% Order Picking Error Rate (%) 1.5～2.0 0.5～0.8 ↓By approximately 50 - 70% Maintenance Cost (yuan / week) 1200～1500 700～900 ↓By approximately 35 - 40% Average Activation / Recognition Delay (seconds) 2.5～3.0 1.2～1.6 ↓By approximately 50% Warehouse Space Utilization Rate (%) 70～75 80～85 ↑By approximately 10 - 15 percentage points Operability Rate during Disconnection Period (%) 30～40 80～90 ↑By approximately 2 times Correct Recognition Rate when Noise > 80dB (%) 85～88 95～97 ↑By more than approximately 10% Material Loss Rate in Temperature and Humidity Sensitive Area (%) 2.5～3.0 1.0～1.5 ↓By approximately 50%

[0341] IV. Experimental Conclusions

[0342] 1. Efficiency improvement: Through AI-driven adaptive scheduling and visual guidance, the present invention improves the picking efficiency by 20-30% compared with the conventional mode under the same personnel configuration, bringing considerable benefits to large-scale warehousing operations.

[0343] 2. Accuracy improvement: The mis-picking rate is significantly reduced, which not only reduces subsequent rework, but also improves customer satisfaction and material safety.

[0344] 3. Cost and sustainability: The combination of passive tags and multi-antenna readers avoids excessive wiring and battery replacement expenses; the edge-cloud collaboration mode is outstanding in reducing network load and improving fault tolerance.

[0345] 4. Wide range of applicable scenarios: This solution can be extended and applied to temperature-controlled warehousing, high-noise operation areas, and multi-floor stereoscopic warehouses, with good portability and scalability.

[0346] The above experimental data show that the present invention has obvious improvements in terms of efficiency, accuracy, availability, sustainability, etc. compared with the traditional warehousing management mode, which is sufficient to prove its significant technical and economic value in actual production operations.

[0347] The above is the description of the embodiments of the present invention. Through the above description of the disclosed embodiments, those skilled in the art can implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel points disclosed herein.

Claims

1. An intelligent warehouse management system based on passive luminous RFID tags, characterized in that, Including: 1) A warehousing environment perception unit, including several sensors, which are used to collect environmental data such as temperature and humidity, illuminance, noise level, and shelf vacancy rate in the warehouse; 2) An integrated hardware module, which is set on the top of the warehouse or in key areas. The integrated hardware module includes: a) A multi-antenna reader, which can dynamically adjust the transmission power or direction according to the real-time data provided by the environment perception unit, to efficiently activate passive luminous RFID tags scattered in different areas; b) A buzzer, which can automatically adjust the acoustic prompt intensity or frequency according to the noise level collected by the environment perception unit; 3) Passive luminous RFID tags, which are distributed on each storage location or material packaging. The RFID tags rely on capturing radio waves emitted by the multi-antenna reader to achieve passive power supply and luminous indication, and can distinguish material status or warning information through different colors or LED flashing modes; 4) An edge computing module, which is connected to the integrated hardware module and the environment perception unit, and is used to perform local data analysis and preprocessing based on the real-time collected environment and material information, and independently complete the execution and monitoring of warehousing operation instructions when the network is abnormal; 5) An AI algorithm unit, which is deployed on the edge computing module or the cloud service platform, integrates material historical turnover records, environmental data, and personnel operation behaviors, and adaptively optimizes the storage location allocation strategy and predicts material requirements through deep learning or machine learning models; 6) A cloud service platform, which is communicatively connected to the edge computing module and the handheld terminal device, and is used to collect global warehousing data, perform big data analysis and generate a visual warehousing operation report, forming a collaborative working mode with the edge computing; 7) A handheld terminal device, which interacts with the cloud service platform and the edge computing module, displays the path planning, warehousing optimization instructions, and passive luminous RFID tag status information provided by the AI algorithm unit through a graphical interface, and uploads the on-site execution results to the cloud.

2. The intelligent warehousing management system according to claim 1, wherein, The environment perception unit includes a noise sensor, a temperature and humidity sensor, an illuminance sensor, and a ground load detection sensor to meet the various scenario requirements of refined warehousing management; And / or, the multi-antenna reader adopts a rotatable or linear array antenna structure, and can automatically adjust the angle according to the height or layout of the shelves in the warehouse to reduce signal blind spots; And / or, the acoustic prompt of the buzzer can be remotely configured by the handheld terminal device to emit different tones or durations for instructions with different priorities or urgencies; And / or, the passive luminous RFID tag is internally provided with a reflective layer or a high-brightness LED module, and a semi-transparent window can be designed at the edge of the tag to enhance the light scattering effect; And / or, the handheld terminal device has a touch screen interface and a voice prompt function, and can display the color, flashing mode, or inventory information of the passive luminous RFID tag in real time, providing multimodal guidance for operators; And / or, the edge computing module is equipped with local storage and anomaly detection functions. When the cloud service platform has a network failure or delay, it can independently complete the specified warehousing operations and cache the data locally; And / or, the AI algorithm unit can automatically switch the learning model according to the material demand changes in different time periods, make advance predictions on seasonal or periodic fluctuations, and optimize the stocking strategy.

3. The intelligent warehousing management system according to claim 1 or 2, characterized in that The AI algorithm unit combines a graph neural network with a time series prediction model to extract material turnover and environmental features, and realizes the adaptive optimization of warehouse location allocation and demand prediction through a reinforcement learning mechanism; the graph neural network uses a graph convolutional network (GCN), and the time series prediction model is a long short-term memory network (LSTM); the graph neural network constructs a spatio-temporal graph based on the warehouse environment topology, regards the warehouse locations and channels as nodes and edges, and integrates time series data to capture the temporal changes in material demand and the interaction relationships between nodes; The reinforcement learning policy module adopts an Actor-Critic structure. The reinforcement learning mechanism takes the rearrangement of warehouse locations or the allocation of material batches as an action, and the current environmental state, node information, and time series prediction results of the warehouse as the state. The reward is set according to operation efficiency, error rate, or cost objectives.

4. The intelligent warehousing management system according to claim 3, characterized in that Within the time interval {1, 2, …, T}, discrete sampling or monitoring is performed on the warehouse, and various feature vectors of the warehouse location and channel nodes are recorded at each sampling moment; where: 1) Warehouse environment topology A spatio-temporal diagram representing warehousing Where: Denotes a set of nodes, including bin nodes and channel nodes, where N is the total number of nodes; Indicates an edge set, used to describe the adjacency or reachability relationship between storage positions or between a storage position and a channel; A ∈ R N×N : Adjacency Matrix of the graph. If there is a path or a connection with a relatively short distance between node v i and v j , then A i,j = 1, otherwise it is 0 or a corresponding weight is set according to distance attenuation; 2) Node and time series features Let x vi,t ∈R dx represent the feature vector of node v i at time t, which contains: Material turnover information, including the number of inbound and outbound times, inventory quantity, and residence duration; Environmental data, including temperature, humidity, noise, and illuminance sensed by sensors, combined with the average value of the location where the node is located or its neighboring locations; Operation behavior data, including the operation duration, error rate, and picking frequency of the operator near the goods location; Obtain the node feature matrix X at each moment t t ∈R N×dx ; The length of the time series is T, that is, there are {X1, X2, …, X T}; 3) Graph neural network layer In a graph neural network, a hidden representation is introduced for each node at each time step t where l represents the l-th layer of the GNN; the graph convolution or graph attention mechanism is expressed as: Where Aggregate(.) represents the neighborhood aggregation function; 4) Time series prediction module After capturing the graph structure at the same moment, the changing trend across moments is then modeled; based on the output of the GNN at each moment t, the time series model LSTM is stacked to express the dynamic evolution: z t = RNN(Readout(H t )) Among them represents the representation after concatenating or pooling all node vectors output by the GNN at the L-th layer. The RNN is an LSTM, and z t represents the temporal hidden state of the entire graph at time t, which is used to predict the material requirements and environmental state at the next moment or several future time periods; 5) Reinforcement learning To combine the GNN + time series model with RL, an Actor-Critic structure is used; the reinforcement learning mechanism takes the rearrangement of warehouse locations or the allocation of material batches as the actions of the agent; defined as follows: State: S t Contains the overall warehousing state at the current moment, which can be specifically composed of the following information: The global representation z extracted by the GNN+ time series model t ; Current environmental parameters, including average temperature, noise, and congestion degree of key channels; Key performance indicators, including inventory turnover rate, error rate, and operation delay; Action: A t Including scheduling decisions for positions or materials, including: Rearrangement of warehouse locations: Adjust material A from shelf X to shelf Y; Recommendation of picking path: Plan the shortest or least conflicting operation path for the operator; Material transfer: Pre-transfer high-frequency materials to high-heat areas; Activation of buzzer and label status: Select which buzzers and passive RFID tags emit light for auxiliary positioning; Reward (Reward): Rt is determined by a pre-defined multi-objective function, comprehensively considering operation efficiency, cost, or expense, or reflecting requirements in terms of environmental safety or inventory risk; Reinforcement learning updates the policy by maximizing the cumulative reward ∑ t γ t R t . γ ∈ (0, 1) is the discount factor, which enables the system to gradually approach the global optimal or sub-optimal layout.

5. The intelligent warehousing management system according to claim 1, characterized in that, The system also includes a positioning cooperation unit for: a) An acoustic positioning module, which calculates the approximate area of the target based on the sound source intensity of the buzzer and the warehouse sound field model; b) An optical positioning module to obtain the spatial coordinates or incident angles of the luminescent RFID tags; c) A fusion calculation module to perform weighted fusion of the acoustic regions and optical coordinates or perform triangulation to output the three-dimensional coordinates or shelf numbers of the target storage locations.

6. The intelligent warehousing management system according to claim 1, wherein The cloud service platform can be docked with the enterprise's ERP, MES, or WMS systems to achieve cross-system collaboration and information sharing through APIs or data interfaces; And / or, the system supports modular expansion, and the edge computing module can flexibly upgrade the hardware configuration or add additional sensor interfaces according to the storage scale and business requirements.

7. An intelligent warehousing management method based on passive luminous RFID tags, characterized in that, This method uses the system described in any one of claims 1-5, and includes the following steps: S1 Environmental data collection: Real-time obtain the noise level, temperature and humidity, illuminance, and shelf vacancy rate in the warehouse through the environmental perception unit; S2 Instruction reception: The integrated hardware module receives the material requisition or picking instructions sent by the warehouse management system, and the buzzer adapts the acoustic prompt according to the noise level; S3 Wireless activation and tag identification: The multi-antenna reader dynamically adjusts the direction or power of the antenna according to the environmental data, emits radio waves to the passive luminescent RFID tags in the warehouse, and completes the reading of the tag information; S4 Edge computing processing: The edge computing module analyzes and preprocesses the data of the reader and the environmental perception unit, and independently executes the operation process when the network is abnormal; S5 AI optimization decision-making: The AI algorithm unit trains and predicts the warehouse data, generates an operation optimization plan, and synchronizes the results to the handheld terminal device; S6 Terminal execution and feedback: The handheld terminal device displays the status of the passive luminescent RFID tags and the task instructions, and the operator uploads the results to the edge computing module and the cloud service platform after completing the material operation; S7 Data recording and visualization: The cloud service platform integrates the data of the edge computing and the handheld terminal, updates the warehouse database, and generates a visualization report to achieve a continuous optimization closed-loop.

8. The method according to claim 6, characterized in that, In the step (S3), the luminescent color or blinking frequency of the passive luminescent RFID tags is dynamically set by the integrated hardware module according to different material types or risk levels; And / or, in the step (S4), if the edge computing module detects that the cloud service platform is unavailable, it automatically enters the offline mode and records all operation logs for later synchronization to the cloud; And / or, in the step (S6), the handheld terminal device can also perform further data verification with the passive luminescent RFID tags through barcode scanning or automatic identification to prevent incorrect picking or misaligned storage; And / or, in the step (S7), the cloud service platform performs big data analysis on the accumulated data of the edge computing module to discover potential operation bottlenecks and continuously iterate and optimize the AI algorithm.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When this computer program or instruction is executed by a processor, it implements the method described in any one of claims 7-8.

10. A computer program product comprising a computer program or instructions, characterized in that, When this computer program or instruction is executed by a processor, it implements the method described in any one of claims 7-8.

Images