Intelligent storage method and device based on cloud warehouse

Abstract

The present invention relates to an intelligent storage method and device based on cloud warehouse. The method: perform all-round scanning and cargo feature recognition on the goods entering the cloud warehouse to obtain cargo feature data; assign and bind RFID tags to each piece of cargo and perform initial positioning to obtain the three-dimensional coordinate value of the cargo; construct a multi-dimensional cargo feature vector and calculate the storage location allocation plan to perform path planning and multi-agent task scheduling on the AGV to obtain the storage task execution instruction of the AGV; obtain the RFID scanning data of the stored goods and the AGV on-board sensor data and perform fusion positioning to obtain the real-time and accurate location and storage status information of the goods; input the real-time and accurate location and storage status information of the goods into the encoder-decoder model to perform AGV path planning and sorting decision calculation to obtain the optimal sorting strategy and AGV motion trajectory. The implementation of the present invention improves the accuracy of AGV path planning and sorting decision of the cloud warehouse.

Description

Technical Field

[0001] The present invention relates to the field of cloud warehouse storage technology, and in particular to an intelligent storage method and device based on cloud warehouse. Background Art

[0002] With the rapid development of e-commerce and logistics industries, traditional warehousing management models have been unable to meet the growing demand for efficient, accurate and flexible storage. As a new type of intelligent warehousing model, cloud warehouses provide enterprises with more efficient warehousing solutions by integrating advanced information technology and automation equipment. However, the current cloud warehouse system still faces many challenges in terms of cargo identification, positioning accuracy, storage optimization and sorting efficiency.

[0003] Traditional cargo identification and positioning methods often rely on manual operations or simple barcode systems, which are prone to errors and inefficient. At the same time, complex environmental factors in the warehouse, such as occlusion and signal interference, make it difficult to accurately locate cargo. In addition, the allocation of storage locations is usually based on simple rules or experience, which makes it difficult to fully utilize warehouse space and optimize pickup efficiency. In terms of multi-AGV collaborative operations, existing path planning and task scheduling algorithms are difficult to effectively handle dynamic obstacles and real-time task changes, resulting in interference and efficiency loss between AGVs. At the same time, how to effectively integrate the real-time location and status information of cargo into sorting decisions to achieve the optimal sorting strategy is also a major challenge facing the current cloud warehouse intelligent storage system. Summary of the invention

[0004] The main purpose of the present invention is to provide an intelligent storage method and device based on a cloud warehouse, so as to improve the accuracy of AGV path planning and sorting decisions in the cloud warehouse.

[0005] To achieve the above object, the present invention provides a cloud warehouse-based intelligent storage method, comprising the following steps:

[0006] Carry out all-round scanning and cargo feature recognition on the goods entering the cloud warehouse to obtain cargo feature data;

[0007] Assign and bind an RFID tag to each piece of cargo according to the cargo characteristic data, and perform initial positioning through the neural XGB three-dimensional positioning system to obtain the three-dimensional coordinate value of the cargo;

[0008] Based on the cargo feature data and the three-dimensional coordinate values ​​of the cargo, a multi-dimensional cargo feature vector is constructed, and a storage location optimization calculation is performed through a deep reinforcement learning model to obtain a storage location allocation plan;

[0009] According to the storage location allocation scheme and the digital twin model of the cloud warehouse, path planning and multi-agent task scheduling are performed on the AGV to obtain storage task execution instructions for the AGV;

[0010] Obtain the RFID scanning data of the stored goods and the AGV vehicle-mounted sensor data and perform fusion positioning to obtain the real-time accurate location and storage status information of the goods;

[0011] The real-time accurate location and storage status information of the goods are input into the encoder-decoder model for AGV path planning and sorting decision calculation to obtain the optimal sorting strategy and AGV motion trajectory.

[0012] The present invention also provides a cloud warehouse-based intelligent storage device, comprising:

[0013] The scanning module is used to perform all-round scanning and cargo feature recognition on the cargo entering the cloud warehouse to obtain cargo feature data;

[0014] A binding module is used to assign and bind an RFID tag to each piece of cargo according to the cargo characteristic data, and perform initial positioning through the neural XGB three-dimensional positioning system to obtain the three-dimensional coordinate value of the cargo;

[0015] A calculation module, used to construct a multi-dimensional cargo feature vector based on the cargo feature data and the three-dimensional coordinate values ​​of the cargo, and perform storage location optimization calculation through a deep reinforcement learning model to obtain a storage location allocation plan;

[0016] A planning module, used to perform path planning and multi-agent task scheduling for the AGV according to the storage location allocation scheme and the digital twin model of the cloud warehouse, and obtain storage task execution instructions for the AGV;

[0017] The positioning module is used to obtain the RFID scanning data of the stored goods and the AGV vehicle-mounted sensor data and perform fusion positioning to obtain the real-time accurate location and storage status information of the goods;

[0018] The processing module is used to input the real-time accurate location and storage status information of the goods into the encoder-decoder model to perform AGV path planning and sorting decision calculation to obtain the optimal sorting strategy and AGV motion trajectory.

[0019] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0020] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0021] In summary, the technical solution provided by the present invention improves the accuracy and comprehensiveness of cargo identification through omnidirectional scanning and multimodal feature fusion. The neural XGB three-dimensional positioning system is adopted, combined with RFID technology and multi-sensor fusion, which significantly improves the accuracy and reliability of cargo positioning, and effectively solves the positioning problem in complex warehouse environments. The storage location optimization is carried out using a deep reinforcement learning model, taking into account multi-dimensional factors such as cargo characteristics, correlation and warehouse status, and achieving more efficient space utilization and a more convenient pickup process. The AGV path planning method based on the digital twin model and time-varying artificial potential field can effectively deal with dynamic obstacles and improve the obstacle avoidance ability and operation safety of the AGV. The multi-agent task scheduling algorithm takes into account the real-time status and task priority of the AGV, improves the overall scheduling efficiency, and reduces interference and conflict between AGVs. The fusion positioning technology combines multiple sensor data to realize real-time and accurate monitoring of cargo location and status. The encoder-decoder model combined with the attention mechanism can better capture the key information of cargo location and warehouse status, and improve the accuracy of AGV path planning and sorting decisions. The combined design of the classification head and the regression head realizes the joint optimization of discrete actions and continuous parameters, making the motion control of AGV more precise and flexible. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of the steps of a cloud warehouse-based intelligent storage method in one embodiment of the present invention;

[0023] Figure 2 It is a structural block diagram of a cloud warehouse-based intelligent storage device in one embodiment of the present invention;

[0024] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0025] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0027] Reference Figure 1 This embodiment provides a cloud warehouse-based intelligent storage method, comprising the following steps:

[0028] S1, conduct all-round scanning and cargo feature recognition on the cargo entering the cloud warehouse to obtain cargo feature data;

[0029] Among them, multi-angle high-resolution images of the goods entering the cloud warehouse are collected to obtain the appearance information of the goods and obtain multi-view image data. Image denoising, illumination compensation and geometric correction are performed on the multi-view image data. Image denoising effectively eliminates the noise caused by environmental interference, illumination compensation is used to balance the image brightness difference caused by uneven light, and geometric correction ensures the accurate proportion of the image and generates standardized image data. Deep convolution operation is performed on the standardized image data, and visual features are extracted from the image through deep learning convolutional neural network technology to obtain visual feature vectors. At the same time, the goods are scanned in three dimensions to obtain their point cloud data. The point cloud data contains the three-dimensional shape information of the goods. By performing denoising and registration processing on these data, unnecessary noise is eliminated, and the scanning data at different angles are aligned to obtain a complete three-dimensional shape model. Based on the three-dimensional shape model, the geometric features of the goods are calculated, including volume, surface area and main axis direction. The goods are weighed by an electronic scale to obtain the weight data of the goods. The weight data is combined with the geometric feature data to calculate the density distribution of the goods and obtain the physical feature data. The calculation of density distribution not only provides information on the material distribution inside the goods, but can also be used to determine whether the goods have abnormal or irregular internal structures. The visual feature vectors, geometric feature data, and physical feature data are fused and classified, and feature information of different dimensions is combined to form a complete cargo description model to obtain the category, quantity, and batch information of the goods. Category recognition helps the system automatically determine the type of goods, and quantity information is used to determine the specific quantity of goods in the same batch, thereby improving the accuracy and management efficiency of warehousing. The category, quantity, and batch information of the goods are encoded and combined with the visual feature vectors, geometric feature data, and physical feature data to generate complete cargo feature data.

[0030] S2, assign and bind RFID tags to each piece of cargo according to cargo feature data, and perform initial positioning through the neural XGB 3D positioning system to obtain the 3D coordinate value of the cargo;

[0031] Specifically, a unique RFID identification code is generated for each piece of goods according to the category, quantity and batch information of the goods. The identification code is the exclusive identity of each piece of goods, ensuring the uniqueness and traceability between different goods. The RFID identification code is written into the RFID tag for subsequent identification and tracking of the goods. The RFID tag is affixed to the surface of the goods to physically bind it to the goods. At the same time, an association relationship between the goods and the RFID identification code is established in the cloud warehouse management system, and the specific information of the goods is data-bound to the RFID tag to ensure that the system can trace all relevant information of the goods through the identification code. The RFID reader array is used to collect multi-point signal strength of the goods with RFID tags affixed. These reader arrays are distributed at multiple locations in the cloud warehouse and can collect signal strength data emitted by RFID tags from different directions. The collected signal strength distribution data reflects the approximate position of the tag in three-dimensional space. The signal strength distribution data is subjected to outlier removal and signal smoothing to effectively eliminate noise interference and obtain standardized signal strength data. The standardized signal strength data is input into the neural network module in the neural XGB three-dimensional positioning system. This module uses the feature extraction capability of deep learning to analyze the spatial distribution characteristics of signal strength data, thereby making a preliminary estimate of the location of the goods and obtaining an initial position estimate. The XGBoost regressor is used to refine the initial position. The XGBoost algorithm eliminates positioning errors through regression calculation of the input data and generates more accurate corrected three-dimensional coordinate prediction values. The corrected three-dimensional coordinate prediction values ​​are constrained and matched with the warehouse layout information. The coordinates are constrained based on the warehouse layout information to ensure that the coordinates fall within the valid area of ​​the warehouse, avoid positioning results in unreasonable locations, and obtain a reasonable initial three-dimensional coordinate that conforms to the actual scene. The initial three-dimensional coordinates are subjected to time series filtering to eliminate short-term fluctuations caused by environmental changes or system errors, and obtain the three-dimensional coordinate values ​​of the goods.

[0032] S3, based on the cargo feature data and the three-dimensional coordinate values ​​of the cargo, construct a multi-dimensional cargo feature vector, and perform storage location optimization calculation through a deep reinforcement learning model to obtain a storage location allocation plan;

[0033] It should be noted that the visual feature vector, geometric feature data and physical feature data of the goods are normalized to eliminate the dimensional differences between different feature dimensions and obtain standardized feature data. The standardized feature data is spliced ​​with the three-dimensional coordinate values ​​of the goods to obtain a multi-dimensional feature vector of the goods. In order to reduce the computational complexity, the spliced ​​feature vector is subjected to dimensionality reduction processing, and the feature vector is made more compact by reducing redundant information. According to the multi-dimensional feature vector of the goods, the similarity matrix between the goods is calculated to reflect the similarity between different goods in the feature space. Through the calculation of the similarity matrix, similar goods are identified and grouped, and the clustering algorithm is used to automatically classify similar goods to generate the result of the goods grouping. The result of the goods grouping is associated with the historical storage data to extract the combination of goods that frequently co-occur and obtain the goods association rules. Based on the association rules, it is identified which goods should be stored together in space first. Combined with the storage status of the current warehouse, a storage location candidate set is constructed, which represents the possible storage location selection. The candidate set is input into the deep reinforcement learning model as the state space. The policy network of the deep reinforcement learning model evaluates the storage location candidate set and generates storage location allocation actions by learning and weighing the pros and cons of various storage locations. These actions represent different storage location selection schemes, and the expected benefits of each allocation action are calculated through the value function network. The value function network estimates the benefits and efficiency of each storage location selection scheme in future operations by analyzing historical data and the current warehouse status. Based on the expected benefits, the optimal storage location allocation action is selected and applied to the current warehouse storage status, thereby updating the warehouse storage status and arranging the most reasonable storage location for the goods. The updated storage status is evaluated and its corresponding reward value is calculated. The reward value reflects the efficiency and optimization degree of the storage operation. If the storage result meets the expectations, the reward value is high, otherwise it is low. Based on the reward value, the policy network and value function network in the deep reinforcement learning model are updated through temporal difference learning, and the decision-making ability of the model is gradually optimized to ensure that more accurate storage location allocation can be made in future storage operations. Through the iterative learning mechanism, the optimization ability of the storage location of goods is continuously improved, and finally an efficient storage location allocation plan is generated.

[0034] S4, based on the storage location allocation scheme and the digital twin model of the cloud warehouse, performs path planning and multi-agent task scheduling for the AGV to obtain the storage task execution instructions of the AGV;

[0035] Specifically, a rasterized map of the warehouse environment is constructed based on the digital twin model of the cloud warehouse. The digital twin model fully reflects the physical layout, equipment location and dynamic state of the cloud warehouse. The actual space of the warehouse is divided into discrete grid units, so that each grid unit represents a location in the warehouse. At the same time, the target location in the storage location allocation plan is marked on the rasterized map to generate a specific task map, showing the location where the AGV needs to go and other key locations in the warehouse. The task map is subjected to obstacle expansion processing to generate an expanded map. Obstacle expansion refers to expanding the occupation range of actual obstacles in the map to ensure that the AGV can reserve enough safety space during path planning to avoid collision with obstacles. Based on the expanded map, the initial path of the AGV from the current position to the target storage location is calculated. The initial path is generated based on the spatial layout of the warehouse and the location of obstacles, and represents the basic route of the AGV. The initial path is smoothed by Bezier curves to effectively smooth the path, so that the AGV does not need to frequently turn or stop during driving, improves its stability and efficiency, and obtains a smooth path. The smooth path is input into the time-varying artificial potential field model to deal with the obstacle avoidance problem of dynamic obstacles. The time-varying artificial potential field model can perceive the position changes of other moving objects in the warehouse in real time, and make local adjustments to the path of the AGV according to these changes to ensure that it can smoothly avoid dynamic obstacles and form a safe local path. Based on the current state of the AGV queue, the tasks of multiple AGVs are collaboratively scheduled. A task allocation matrix for multi-AGV collaborative scheduling is constructed. By solving the matrix, the initial task scheduling results are generated to ensure that multiple AGVs can collaborate efficiently in the same warehouse environment to avoid path conflicts and resource waste. The initial task scheduling results are converted into specific operation instructions that can be executed by the AGV, including the operation sequences of AGV movement, steering, grabbing and placement. Each instruction corresponds to an action of the AGV in executing the task. These operation instructions are output to form a complete AGV storage task execution instruction.

[0036] The smooth path is discretized into a series of continuous path point sequences, each of which represents a specific position of the AGV in space. In order to ensure that the AGV can successfully avoid obstacles in the actual environment, the gravitational field and repulsive field are calculated for each path point. The gravitational field is the attraction pointing to the target position, while the repulsive field is the repulsive force generated by dynamic obstacles. By calculating the superposition of these two force fields, the initial artificial potential field distribution is obtained, which reflects the combined force of the AGV on the path. Based on the real-time position information of the AGV and dynamic obstacles, the initial artificial potential field distribution is dynamically updated to construct a time-varying artificial potential field model. The model is continuously adjusted over time and can reflect the changes in the environment in the warehouse in real time, especially the position and state of dynamic obstacles. In order to calculate the specific movement direction and speed of the AGV, the gradient descent method is used in the time-varying artificial potential field model. By moving along the gradient direction of the potential field, the AGV avoids obstacles and moves towards the target. Through this process, the local path of the AGV is generated, that is, the safe route it should follow at the current moment. The local path is segmented and each segment is regarded as an independent subtask. These subtasks together constitute the set of tasks that AGV needs to perform, and a list of tasks to be assigned is generated based on this set of tasks. In order to achieve the collaborative scheduling of multiple AGVs, the current state of the AGV queue is considered, including key factors such as the position, power level, and load of each AGV. These state information is integrated into an AGV state matrix to characterize the current capacity and availability of each AGV. The AGV state matrix is ​​combined with the list of tasks to be assigned to generate a task allocation matrix for multi-AGV collaborative scheduling, which reflects the tasks that each AGV can perform and possible matching relationships. In order to optimize the task allocation, the Hungarian algorithm is used to solve the task allocation matrix. The Hungarian algorithm obtains the optimal task allocation plan by maximizing the overall benefit or minimizing the task cost, ensuring that each AGV can complete its assigned task in the most efficient way. The optimal task allocation plan is analyzed in time windows to detect and resolve potential interference and conflict problems between AGVs. Since multiple AGVs collaborate in the same environment, there may be path overlap or task conflict. Through time window analysis, the task execution order or path planning of each AGV is adjusted to avoid collisions or resource competition between AGVs, and finally the initial task scheduling result is generated.

[0037] S5, obtain the RFID scanning data of the stored goods and the AGV vehicle-mounted sensor data and perform fusion positioning to obtain the real-time accurate location and storage status information of the goods;

[0038] Among them, the RFID tags of the stored goods are periodically scanned by the RFID reader array. The RFID reader regularly collects the signal strength data of the RFID tags of the goods and uses this data as the initial input for the positioning of the goods. The collected RFID signal strength data is preprocessed, including the removal of noise and outliers, to obtain the cleaned RFID data. At the same time, the surrounding environment is scanned based on the AGV on-board laser radar to obtain environmental point cloud data. The point cloud data is filtered to remove unnecessary interference information, and then the point cloud data is registered to align it with the existing environmental model to obtain accurate three-dimensional spatial information around the AGV. The inertial measurement unit on the AGV collects the acceleration and angular velocity data of the AGV. Based on these data, the initial position and posture information of the AGV are calculated by the dead reckoning method. The cleaned RFID data is input into the neural XGB three-dimensional positioning system to calculate the preliminary position estimate of the goods. The positioning system uses the prediction ability of the machine learning model and combines multi-dimensional features to infer the approximate position of the goods corresponding to the RFID tag. The 3D spatial information of the AGV's surrounding environment is feature extracted and matched, and the AGV's own position is accurately located in combination with the AGV's initial position and posture information, ensuring that the AGV's real-time position can be accurately tracked and providing a more refined coordinate reference for subsequent cargo positioning. Based on the AGV's real-time position, the RFID-based cargo position estimate is transformed and corrected. By combining the AGV's precise position in the local space with the RFID's preliminary position estimate, the global coordinates of the cargo are obtained, indicating the exact position of the cargo in the entire warehouse layout. The global coordinates of the cargo are matched with the warehouse's layout information to identify the specific shelf location where the cargo is located. The matching process combines the shelf distribution in the warehouse with information on other storage units to ensure that the cargo's location can be accurately mapped to the actual physical location. The AGV's onboard temperature and humidity sensors continuously monitor the temperature and humidity data of the surrounding environment, associate the environmental data with the global position of the cargo, and obtain the cargo's storage status information. Temperature and humidity data are particularly important for stored cargo in some cases, such as commodities that are susceptible to environmental influences, which use this information to determine whether the storage environment is suitable. Perform time series analysis on the global coordinates and storage status information of the goods, and continuously monitor the changes in these data to identify any abnormal change patterns. If there is an abnormal fluctuation in the location or storage status of the goods, the system will respond in a timely manner.

[0039] S6, input the real-time accurate location and storage status information of the goods into the encoder-decoder model for AGV path planning and sorting decision calculation to obtain the optimal sorting strategy and AGV motion trajectory.

[0040] Among them, the real-time accurate location and storage status information of the goods are converted into vector form to form a goods state vector. The goods state vector contains the current spatial location, storage status and interaction information with the environment of the goods. The goods state vector is spliced ​​with the current AGV location information and warehouse layout data to form a complete input feature vector. The input feature vector is normalized to ensure that the data is calculated at the same scale to avoid the imbalance problem caused by different data magnitudes. The normalized input data is input to the encoder part of the encoder-decoder model, which consists of five convolutional layers. Each convolutional layer uses a 3x3 convolution kernel with a step size of 2. The dimension of the input features is gradually reduced through convolution operations, and high-level features in the data are extracted at the same time. In order to enhance the generalization ability of the model, each convolutional layer is followed by a BatchNormalization layer to normalize the convolution output. The ReLU activation function is applied after each convolution operation to ensure that the model has nonlinear mapping capabilities and generates encoded feature representations. The encoded feature representations are input to the attention mechanism layer. This layer uses a multi-head self-attention structure, which contains 8 independent attention heads. Each head independently calculates the attention weight and performs weighted operations on the features. Through this mechanism, the model can capture important information in the input data and ignore the minor parts. The outputs of the 8 heads are concatenated to obtain the feature representation with enhanced attention. The feature representation with enhanced attention is input to the decoder part. The decoder consists of five transposed convolutional layers. Each transposed convolutional layer uses a 3x3 convolution kernel and gradually restores the dimension of the feature to the dimension of the original input through a deconvolution operation with a step size of 2. Each transposed convolutional layer is connected to a BatchNormalization layer to ensure the stability of the network, and the LeakyReLU activation function is used to enhance the nonlinear ability of the model. Through the decoding operation, a predicted path planning result is finally obtained, which reflects the possible route of the AGV in the warehouse. The predicted path planning result is input to the classification head and regression head for processing. The classification head outputs the probability distribution of the action category of the AGV through the Softmax function, thereby predicting the specific actions taken by the AGV in each step, such as moving, turning, stopping, etc. The regression head outputs continuous AGV motion parameters, including precise displacement, steering angle, and speed that the AGV needs to perform on the path. The outputs of the classification head and regression head together form the motion trajectory of the AGV, which fully describes the action plan of the AGV from the current state to the target position. Based on the generated AGV motion trajectory and the overall state of the current warehouse, multiple candidate sorting strategies are generated. By setting a reward function, the expected return of each candidate sorting strategy is evaluated. The reward function scores based on factors such as the efficiency of the warehouse, the movement cost of the AGV, and the task completion time, and finally selects the strategy with the highest expected return as the optimal sorting strategy.

[0041] The predicted path planning results are mapped to the action space dimension through a fully connected layer, and the features related to the AGV's execution operation are extracted to form an action feature vector. The action feature vector can fully reflect all the key information required by the AGV when performing tasks, including movement mode, speed change, and direction adjustment. The action feature vector is dimensionally segmented and decomposed into a discrete action part and a continuous action part. The discrete action part represents the discrete operations taken by the AGV, such as specific action types such as turning, stopping, or accelerating, while the continuous action part corresponds to the precise motion parameters of the AGV, such as the angle of its movement direction, the distance traveled, or the speed. The discrete action features are input into the classification head network for processing. The classification head network consists of two fully connected layers, each of which is followed by a ReLU activation function to ensure that the network has sufficient nonlinear representation capabilities to accurately capture the complex relationship between different discrete actions. After being processed by two fully connected layers, the classification head network finally outputs a probability distribution normalized by the Softmax function, which represents the probability weights of the AGV in different action categories. For example, an action category is "turn left" or "stop", and the output of the Softmax function indicates the possibility of the AGV performing each action in the current situation. At the same time, the continuous action features are input into the regression head network. The regression head network also consists of two fully connected layers, each of which is followed by the Tanh activation function. The Tanh activation function can limit the output between [-1, 1], thereby effectively expressing the continuous changes in the AGV motion parameters, such as precise speed adjustment or slight direction changes. After processing these two fully connected layers, the regression head network outputs continuous AGV motion parameters, including the fine operation information required by the AGV when performing tasks, such as the turning angle, forward speed, etc. Based on the probability distribution of AGV action categories, discrete action sampling is performed. The sampling process randomly selects a specific discrete action as the execution action of the AGV in the current situation according to the probability weight of each action category, ensuring that the AGV can make adaptive decisions in an uncertain environment. The discrete action selection results after sampling are combined with the continuous motion parameters output by the regression head network to form a complete AGV motion trajectory.

[0042] In one example, the goods entering the cloud warehouse are scanned in all directions and their characteristics are identified to obtain the characteristic data of the goods, including:

[0043] Capture high-resolution images of goods entering the cloud warehouse from multiple angles to obtain multi-view image data, and perform image denoising, illumination compensation, and geometric correction on the multi-view image data to obtain standardized image data;

[0044] Perform deep convolution operations and feature extraction on standardized image data to obtain visual feature vectors, and perform three-dimensional scanning on the goods to obtain point cloud data;

[0045] Perform noise reduction and registration processing on the point cloud data to obtain a three-dimensional shape model, and calculate the volume, surface area and main axis direction of the goods based on the three-dimensional shape model to obtain geometric feature data;

[0046] Weigh the goods using an electronic scale to obtain weight data, and combine the weight data with the geometric characteristic data to calculate the density distribution of the goods and obtain physical characteristic data;

[0047] Perform feature fusion and classification on visual feature vectors, geometric feature data and physical feature data to obtain the category, quantity and batch information of the goods;

[0048] The category, quantity and batch information of the goods are encoded and combined with the visual feature vector, geometric feature data and physical feature data to generate goods feature data.

[0049] In this example, multiple high-resolution industrial cameras are arranged in the warehouse entry area to simultaneously shoot the goods from different angles to generate multi-view image data containing information about the goods in all directions. Assuming there are n cameras, the image resolution of each camera is H×W (height is H pixels, width is W pixels), and the collected multi-view image data is represented as a four-dimensional tensor I ijkm , where i represents the camera number, j and k represent the vertical and horizontal pixel coordinates of the image respectively, and m represents the color channel (such as RGB three channels). Image denoising is performed on multi-view image data to remove random noise in the image. In order to eliminate the brightness difference caused by the uneven distribution of light sources, the image data is subjected to illumination compensation. Illumination compensation uses histogram equalization or adaptive histogram equalization to adjust the brightness distribution of the image to a standard range to improve the uniformity of image details. Through the calibration of the internal and external parameters of the camera and the registration of feature points, the image is geometrically corrected to obtain a set of standardized image data. A deep convolution operation is performed on the standardized image data, and a convolutional neural network is used to extract deep features in the image. Assuming the convolution kernel size is k×k, the result of the convolution operation is expressed as:

[0050]

[0051] Among them, F ijkl represents the convolutional feature map, W pqml is the convolution kernel weight, b lis the bias term, u and v are the moving steps of the convolution. Through multi-layer convolution and pooling operations, a high-dimensional vector representing the visual features of the goods is extracted, namely the visual feature vector V. At the same time, the goods are scanned in three dimensions to obtain point cloud data. The point cloud data contains the three-dimensional coordinate information of the surface of the goods, represented as a three-dimensional point set Where N is the number of points (x i ,y i ,z i ) represents the three-dimensional coordinates of each point. In order to eliminate the noise in the scanning process, the point cloud data is de-noised to obtain the processed point cloud data. The de-noised point cloud data is registered to accurately align the point cloud data from different perspectives to obtain a complete three-dimensional shape model of the goods. Based on the three-dimensional shape model, the geometric feature data of the goods, such as the volume V geo , surface area A and main axis direction θ. The volume of the cargo is calculated by voxelization of the three-dimensional model:

[0052]

[0053] Among them, δV i is the volume of each voxel. The surface area A is calculated by adding the area of ​​each triangle on the triangular mesh surface. The principal axis direction θ is extracted from the point cloud data by principal component analysis. The goods are weighed using an electronic scale to obtain the weight data W of the goods. In order to analyze the physical properties of the goods, the weight data is combined with the geometric feature data to calculate the density distribution ρ of the goods, which is calculated as follows:

[0054]

[0055] Among them, ρ represents the density of the cargo, W is the weight, V geo is the volume of the cargo. Density distribution helps determine whether there are abnormalities inside the cargo, such as gaps or areas with uneven density. geo ,A,θ) and physical feature data ρ are fused. Multilayer perceptron is used for feature splicing and classification. Classification operation is performed on the fused feature data to obtain the category, quantity and batch information of the goods. The category, quantity and batch information of the goods are encoded and combined with the previously extracted visual feature vector, geometric feature data and physical feature data to generate a complete cargo feature data set.

[0056] In one example, an RFID tag is assigned and bound to each piece of cargo according to the cargo feature data, and the initial positioning is performed through the neural XGB 3D positioning system to obtain the 3D coordinate values ​​of the cargo, including:

[0057] Generate a unique RFID identification code based on the category, quantity and batch information in the cargo characteristic data, and write the RFID identification code into the RFID tag;

[0058] Attach RFID tags to the surface of goods, establish an association between goods and RFID identification codes in the cloud warehouse management system, and collect multi-point signal strength of RFID tags through the RFID reader array to obtain signal strength distribution data;

[0059] The signal strength distribution data is subjected to outlier removal and signal smoothing to obtain standardized signal strength data, and the standardized signal strength data is input into the neural network module in the neural XGB three-dimensional positioning system for feature extraction and preliminary position estimation to obtain an initial position estimation value;

[0060] Based on the initial position estimate, the XGBoost regressor is used to refine the position to obtain the corrected three-dimensional coordinate prediction value, and the corrected three-dimensional coordinate prediction value is constrained and matched with the warehouse layout information to ensure that the coordinate falls within the valid area to obtain the initial three-dimensional coordinate;

[0061] Perform time series filtering on the initial three-dimensional coordinates to obtain the three-dimensional coordinate values ​​of the goods.

[0062] In this example, a unique RFID identification code is generated based on the category, quantity and batch information in the cargo feature data. Assume that the cargo feature data includes the category C, quantity N and batch information B of the cargo, which can uniquely identify the identity of each cargo. In order to generate a unique RFID identification code, the above three variables are encoded and combined, for example, using string concatenation or hash function to convert them into a unique identifier ID. The unique RFID identification code ID is written into the RFID tag, and the RFID tag is affixed to the surface of each cargo. The identification code in each tag is used to identify the identity of the cargo in the warehouse. When the RFID tag is bound to the cargo, an association relationship between the cargo and the RFID identification code is established in the cloud warehouse management system, that is, the detailed information of each cargo and its corresponding RFID identification code are recorded in the database. The RFID reader array arranged in the warehouse collects multi-point signal strength of the cargo with RFID tags. The RFID reader array consists of multiple readers distributed at different locations in the warehouse, and each reader can detect the radio signal emitted by the RFID tag within the range. The signal strength S is usually related to the location of the RFID tag, the distance from the reader and the surrounding environment. The data collected by multiple readers forms a signal strength distribution matrix S ij, where i represents the reader number and j represents the RFID tag number. This matrix can reflect the signal strength distribution of each tag at different reader locations. The signal strength distribution data is processed by outlier removal and signal smoothing. Outlier removal is performed through statistical analysis methods, such as removing abnormal data exceeding three times the standard deviation, or using a dynamic threshold method based on historical data for judgment. Signal smoothing uses methods such as moving average or exponential smoothing to minimize short-term fluctuations. After processing, the standardized signal strength data S′ is obtained ij The standardized signal strength data S′ ij The neural network module is input into the neural XGB 3D positioning system. The neural network module consists of a multi-layer neural network that can extract spatial features from the input signal strength data and preliminarily estimate the location of the RFID tag. Assume that the output of the neural network is the estimated coordinate value of the cargo tag in 3D space. in are the preliminary position estimates of cargo j in three-dimensional space. The coordinate estimate can give the approximate location of the cargo in the warehouse, but due to signal reflection and multipath effects, the preliminary estimate may have a certain error. In order to improve the positioning accuracy, based on the preliminary position estimate Use XGBoost regressor to refine the position. XGBoost is an ensemble learning method that uses multiple regression trees to model and correct errors to obtain more accurate position prediction values. Let the output of XGBoost be the corrected three-dimensional coordinate prediction value P j ′=(x′ j ,y′ j ,z′ j ), then its calculation formula is expressed as:

[0063]

[0064] Among them, f XGBoost represents the XGBoost model, Θ is the set of model parameters, including the number of trees, depth, learning rate, etc. j ′ is the corrected coordinate prediction value, combined with the initial estimate The characteristics of the signal strength are used to obtain a more accurate location. The corrected three-dimensional coordinate prediction value is constrained and matched with the warehouse layout information. The warehouse layout information includes the boundaries of each area inside the warehouse, the location and size of the shelves, etc., which are used to verify and correct whether the predicted coordinates fall within a reasonable storage area. Assume that the warehouse layout information is represented by a set of spatial constraints R = {r1, r2, ..., r k}, each r k Represents a valid storage area, then the corrected coordinate P j The following conditions must be met:

[0065] P j ′∈R;

[0066] If the predicted coordinates exceed the valid area, the constraints are adjusted to make them fall within the nearest valid storage location to ensure that the final initial three-dimensional coordinates P j0 is a legal coordinate value. In order to smooth the positioning deviation caused by short-term fluctuations, the initial three-dimensional coordinate P j0 Perform time series filtering. Using Kalman filtering, by introducing state prediction and measurement update steps, the predicted position of the goods is dynamically adjusted so that the final three-dimensional coordinates of the goods P j =(x j ,y j ,z j ) is more accurate and stable.

[0067] In one example, a multi-dimensional cargo feature vector is constructed based on cargo feature data and cargo three-dimensional coordinate values, and a storage location optimization calculation is performed through a deep reinforcement learning model to obtain a storage location allocation plan, including:

[0068] Normalizing the visual feature vector, geometric feature data and physical feature data in the cargo feature data to obtain standardized feature data;

[0069] The standardized feature data and the three-dimensional coordinate values ​​of the goods are spliced ​​and reduced in dimension to obtain a multi-dimensional feature vector of the goods. Based on the multi-dimensional feature vector of the goods, the similarity matrix between the goods is calculated, and the clustering algorithm is used to group similar goods to obtain the result of the goods grouping.

[0070] The cargo grouping results are analyzed in association with historical storage data to extract frequently co-occurring cargo combinations and obtain cargo association rules. Based on the cargo association rules and the current warehouse storage status, a storage location candidate set is constructed and input into the state space of the deep reinforcement learning model.

[0071] Use the policy network of the deep reinforcement learning model to evaluate the storage location candidate set, generate storage location allocation actions, and calculate the expected benefit of each action through the value function network;

[0072] Based on the expected return, the optimal storage location allocation action is selected and applied to the current warehouse storage status to obtain the updated storage status;

[0073] The updated storage state is evaluated, the reward value is calculated, and temporal difference learning is used to update the policy network and value function network of the deep reinforcement learning model to obtain the storage location allocation plan.

[0074] In this example, the visual feature vector, geometric feature data, and physical feature data in the cargo feature data are normalized so that the value of each dimension is within a standardized interval, such as [-1, 1] or [0, 1], to obtain standardized feature data. The standardized feature data S is concatenated with the three-dimensional coordinate value C = (x, y, z) of the cargo to form a complete multi-dimensional cargo feature vector F. In order to reduce computational complexity and feature redundancy, the multi-dimensional cargo feature vector F is subjected to dimensionality reduction processing to map the high-dimensional feature vector to a lower-dimensional feature space while retaining the main information features. The similarity between the cargoes is calculated based on the multi-dimensional cargo feature vector after dimensionality reduction. The cosine similarity is used, expressed as:

[0075]

[0076] Among them, F i ′ and F j ′ represent the characteristic vectors of the two goods, sin(F i ′,F j ′) represents the similarity value between two goods, ranging from [-1, 1]. Based on the similarity between all goods, a similarity matrix M is generated, and each element M ij Represents the similarity between goods i and goods j. Based on the similarity matrix M, a clustering algorithm is used to group similar goods, and goods with high similarity are assigned to the same cluster to form a grouping result of goods. For example, in K-means clustering, goods are divided into k clusters according to the similarity matrix, and each cluster represents a class of similar goods. Let the clustering result be {C1, C2, ..., C k}, where C i Represents the set of goods of the i-th category. Perform association analysis on the grouping results of goods and the historical storage data, extract the frequently co-occurring goods combinations, and form the association rules of goods. For example, the combination of goods that are often stored together in a warehouse includes electronic components and their supporting installation tools, then these goods are regarded as a high-frequency co-occurrence combination. Association rules are mined using the Apriori algorithm or the FP-Growth algorithm. The mined rules are expressed as R = {(A→B, support, confidence)}, where A→B means that if goods A appears, then goods B is also likely to appear at the same time, support is the support, indicating the frequency of occurrence of the rule, and confidence is the confidence, indicating the reliability of the rule. Based on the association rules of goods and the storage status of the current warehouse, a storage location candidate set H = {h1, h2, ..., h m}, each candidate position h iis a potential storage location. The storage location candidate set is input into the state space of the deep reinforcement learning model. The state space contains all influencing factors such as the current warehouse layout, storage status, cargo grouping information, and candidate locations. The policy network of the deep reinforcement learning model evaluates the candidate set and generates a storage location allocation action a, that is, where each cargo should be assigned. The policy network uses a deep neural network structure to encode the state and output the probability distribution of each possible action. The value function network calculates the expected benefit Q(s, a) of each storage location allocation action a, where s represents the current state and the Q value represents the long-term benefit of performing action a in state s. Based on the expected benefit, select the storage location allocation action a with the highest benefit. * , and apply it to the storage state of the current warehouse to obtain the updated storage state s′. The updated storage state is evaluated and the reward value r is calculated. The reward value represents the degree of impact of the current allocation action on the warehouse efficiency. For example, if a certain cargo allocation scheme significantly reduces the picking time, the reward value is high. Through temporal difference learning, the reward value r is used to update the policy network and value function network of the deep reinforcement learning model. The updated policy network can better guide future storage location allocation decisions.

[0077] In one example, based on the storage location allocation scheme and the digital twin model of the cloud warehouse, the AGV is path planned and multi-agent task scheduled to obtain the storage task execution instructions of the AGV, including:

[0078] Based on the digital twin model of the cloud warehouse, a rasterized map of the warehouse environment is constructed, and the target locations in the storage location allocation plan are marked on the rasterized map to obtain a task map.

[0079] Perform obstacle expansion processing on the task map to obtain an expanded map;

[0080] Based on the expansion map, the initial path of the AGV from the current position to the target storage position is calculated to obtain the initial path planning result, and the initial path planning result is smoothed by Bezier curve to obtain a smooth path;

[0081] The smooth path is input into the time-varying artificial potential field model, and the dynamic obstacle avoidance calculation is performed to obtain the local path. Based on the local path and the current AGV queue status, the task allocation matrix for multi-AGV collaborative scheduling is constructed and the initial task scheduling result is obtained by solving.

[0082] The initial task scheduling results are converted into a sequence of moving, turning, grabbing and placing operation instructions executable by the AGV to obtain the storage task execution instructions of the AGV.

[0083] In this example, a rasterized map of the warehouse environment is constructed based on the digital twin model of the cloud warehouse. The digital twin model reflects information such as the internal layout of the warehouse, shelf distribution, aisle location, equipment configuration, and dynamic changes. Rasterized mapping is a way of discretizing a continuous warehouse environment into a representation consisting of many small cells. Each cell (grid) represents a small area in the warehouse. The size of the grid is adjusted according to the sophistication of the warehouse operation, usually ranging from tens of centimeters to one meter. Assuming the grid size is d×d square meters, the rasterized map of the entire warehouse area is represented as a two-dimensional matrix G, where each element G ij represents the state of the grid cell at row i and column j. The target locations in the storage location allocation plan are marked on the grid map to form a task map. The storage location allocation plan specifies a target location for each item to be stored. These target locations are usually certain cells in the grid map. For example, suppose the target location is (x t ,y t ), then the corresponding cell G in the rasterized map xy are marked as target points. These marked points constitute the task map, which indicates the final storage location of each item in the warehouse. The task map is expanded to allow the AGV to have a sufficient safety distance during path planning to prevent it from colliding when approaching obstacles. Several cells around the obstacle in the grid map are also marked as obstacle states, thereby expanding the influence range of the obstacle. Assuming the obstacle expansion radius is r, any obstacle grid G ij All grids G within the surrounding r range mn (satisfy All are marked as inaccessible. After expansion processing, the expanded map G′ is obtained. Based on the expanded map G′, the initial path of the AGV from the current position to the target storage location is calculated, and an optimal route from the starting position to the target position is found for the AGV. Assume that the starting position of the AGV is (x s ,y s ), the target position is (x t ,y t ), the result of path planning is a sequence of path points {(x1, y1), (x2, y2), ..., (x t ,y t )}, these path points represent the grid cells that the AGV needs to pass through in sequence, and the initial path planning result is obtained. Due to the resolution limitation of the rasterized map, there are discontinuous turning points in the initial path, which will cause the AGV to frequently stop and turn during driving, affecting driving efficiency and stability. The initial path planning result is smoothed by Bezier curve to generate a smooth and continuous path. Assume that the path point sequence is {P1, P2, ..., P n}, each path point P i =(x i ,y i ), the parametric equation of the Bezier curve is:

[0084]

[0085] Among them, B(t) represents the position of the Bezier curve at parameter t (0≤t≤1), b i (t) is the Bessel basis function, the specific form is:

[0086]

[0087] After Bezier curve processing, a smooth path B(t) is generated. This path is geometrically smoother, avoids sharp changes in direction, and helps improve the driving stability of the AGV. The smooth path B(t) is input into the time-varying artificial potential field model to perform obstacle avoidance calculations for dynamic obstacles. The time-varying artificial potential field model can take into account the changes of dynamic obstacles in real time, such as other AGVs, temporary obstacles, etc. In the artificial potential field model, the target position exerts a gravitational force on the AGV, and the obstacle exerts a repulsive force on the AGV. Assuming that the total potential field force on the AGV at a certain moment is F(t), its calculation formula is:

[0088] F(t)=F att (t)+F rep (t);

[0089] Among them, F att (t) is the gravitational field generated by the target position, pointing to the target point, and is usually defined as:

[0090] F att (t) = -k att ·(PP t );

[0091] Among them, k att is the gravitational coefficient, P is the current position of the AGV, and P t is the target position. rep (t) is the direction of the repulsive field generated by the obstacle away from the obstacle. Its calculation usually depends on the distance between the AGV and the obstacle, such as:

[0092]

[0093] Among them, k rep is the repulsion coefficient, is the position of the ith obstacle. When the obstacle approaches, the repulsive force increases significantly, forcing the AGV to avoid the obstacle. Under the action of the dynamic potential field force, a new local path L(t) is generated, which dynamically avoids the obstacle. Based on the local path L(t) and the current state of the AGV queue (including the position, task status, power, etc. of each AGV), a task allocation matrix for multi-AGV collaborative scheduling is constructed. Each element A of the task allocation matrix A ij It represents the priority or cost of the i-th AGV to perform the j-th task. By solving this matrix, we can optimize the task allocation between AGVs and reduce task conflicts and resource competition. When solving the task allocation matrix, we use the Hungarian algorithm or linear programming to obtain the initial task scheduling result D = {d1, d2, ..., d m}, each d i represents the specific task allocation plan for the i-th AGV. The task scheduling result is converted into a sequence of operation instructions that can be executed by the AGV, including operations such as moving, turning, grabbing and placing. Assume that an AGV currently needs to move from position P s Move to the target position P t , the operation instruction sequence is expressed as: move to the next path point P i+1 ; At path point P i+1 Turn to angle θ i ; Arrive at the target point P t After that, the grabbing or placing operation is performed. These operation instruction sequences will be sent to the AGV control system to instruct the AGV to perform tasks according to the predetermined path and sequence, thereby completing the handling and storage operations of the goods.

[0094] In one example, a smooth path is input into a time-varying artificial potential field model, and obstacle avoidance calculation is performed on dynamic obstacles to obtain a local path. Based on the local path and the current AGV queue status, a task allocation matrix for multi-AGV collaborative scheduling is constructed, and the initial task scheduling results are obtained by solving the matrix, including:

[0095] The smooth path is discretized into a sequence of path points, and the gravitational field and repulsive field are calculated for each path point to obtain the initial artificial potential field distribution. The initial artificial potential field distribution is dynamically updated based on the real-time position information of the AGV and dynamic obstacles to obtain a time-varying artificial potential field model.

[0096] In the time-varying artificial potential field model, the gradient descent method is used to calculate the movement direction and speed of the AGV to obtain the local path, and the local path is segmented. Each segment of the path is taken as a subtask to construct the AGV task set and obtain the task list to be assigned.

[0097] Based on the current AGV queue status, including the location, power and load of the AGV, an AGV status matrix is ​​constructed and combined with the task list to be assigned to construct a task allocation matrix for multi-AGV collaborative scheduling;

[0098] The Hungarian algorithm is used to solve the task allocation matrix of multi-AGV collaborative scheduling to obtain the optimal task allocation plan. The optimal task allocation plan is then analyzed in time windows to detect and resolve potential interference and conflicts between AGVs to obtain the initial task scheduling result.

[0099] In this example, the smooth path is converted into a discrete sequence of path points. Assume that the smooth path is represented by a Bezier curve as B(t), where t is a parameter with a value range of [0, 1]. By discretizing the Bezier curve, it is divided into N evenly distributed path point sequences {P1, P2, ..., P N}, where each path point P i =(x i ,y i ) represents a position point that the AGV needs to pass through in sequence. In order to calculate the gravitational field and repulsive field of each path point, the forces of the target point and the obstacle are defined. The gravitational field represents the attraction of the target point to the AGV, and the repulsive field represents the repulsive force of the obstacle to the AGV. For the path point P i , the gravitational force F of the target point att，i Defined as:

[0100] F att，i =-k att ·(P i -P goal );

[0101] Among them, k att is the gravitational coefficient, which controls the strength of attraction, P goal =(x goal ,y goal ) represents the target position. The direction of this gravitational force is toward the target position. As the path point gets closer to the target position, the strength of the gravitational force gradually decreases. At the same time, the obstacle has an impact on the path point P. i The repulsive force F rep，i Defined as:

[0102]

[0103] Among them, k rep is the repulsion coefficient, controlling the strength of the repulsive force, P obs，jrepresents the position of the jth obstacle, and M is the total number of obstacles. The direction of the repulsive force is away from the center of the obstacle, and as the distance between the path point and the obstacle decreases, the repulsive force increases sharply. By calculating the attraction and repulsion of each path point, the initial artificial potential field distribution is obtained. The total potential field force F at each path point i It is the superposition of attraction and repulsion:

[0104] F i =F att，i +F rep,i ;

[0105] The initial artificial potential field distribution reflects the combined force field of the target point and obstacles on the AGV. In order to enable the artificial potential field to dynamically adapt to the real-time position information of the AGV and obstacles, the initial artificial potential field distribution is dynamically updated based on the real-time position of the AGV and dynamic obstacles. Whenever the position of the AGV or obstacle changes, the gravitational and repulsive forces of each path point are recalculated to reflect changes in the current environment. The resulting time-varying artificial potential field model reflects changes in the warehouse environment in real time. In the time-varying artificial potential field model, the gradient descent method is used to calculate the movement direction and speed of the AGV. The gradient descent method enables the AGV to move along the descending direction of the potential field force, thereby gradually approaching the target point. For each path point P i , its moving direction θ i and speed v i Solved by the following formula:

[0106]

[0107] Among them, F i , x and F i , y are the potential force F i The components in the x and y directions, α is the speed proportional coefficient, which is used to adjust the driving speed of the AGV. Through the gradient descent method, the AGV can automatically adjust the driving direction and speed according to the current potential field force, so as to avoid obstacles and move to the target position. i ,y i ,θ i , v i )}, the path is segmented and each segment is regarded as a subtask. Each subtask represents the driving goal and motion parameters of the AGV in a short period of time. For example, a certain segment of the path may indicate that the AGV should turn left and slow down at the current moment, while the next segment of the path may indicate that the AGV needs to go straight and speed up. By decomposing the local path into multiple subtasks, the motion state and task objectives of the AGV can be adjusted more flexibly, thereby effectively responding to changes in the dynamic environment. These subtasks are organized into a complete AGV task set T = {t1, t2, ..., t k}, where each task t i Contains the motion instructions and target states of the AGV in this path. The AGV state matrix is ​​constructed based on the current state of the AGV queue. The AGV state matrix S is a matrix containing the current state information of all AGVs. Each element S ij Represents the value of the i-th AGV in the j-th state, such as the AGV's position, power, load, etc. Assume there are n AGVs, and the state information of each AGV is represented by a vector s i =[p i , e i , l i ] indicates that p i is the location, e i is the amount of electricity, l i is the load condition, then the dimension of the AGV state matrix S is n×3. The AGV state matrix S is combined with the task list T to be assigned to construct the task allocation matrix A for multi-AGV collaborative scheduling. Each element A of the task allocation matrix ij Represents the cost or priority of the i-th AGV performing the j-th task. This cost is determined by analyzing the AGV's status information and the task requirements. For example, if a task requires a high amount of power, and a certain AGV has insufficient power, then the cost of the AGV performing the task will be very high, and it will be at a disadvantage in task allocation. In order to solve the task allocation matrix A for multi-AGV collaborative scheduling, the Hungarian algorithm is used. The Hungarian algorithm is a classic algorithm for solving bipartite graph matching problems, which effectively finds the optimal matching solution between a task and an AGV. Assume that the dimension of the task allocation matrix A is n×m, that is, there are n AGVs and m tasks. The algorithm will look for an optimal match π={(i1, j1), (i2, j2),..., (i k , j k )}, so that the total cost is minimized

[0108] min∑ (i，j)∈π A ij ;

[0109] Among them, A ijrepresents the cost of the i-th AGV performing the j-th task, and π is a matching set of AGVs and tasks. By solving this optimization problem, the optimal task allocation solution is obtained. A time window analysis is performed on the optimal task allocation solution to detect and resolve potential interference and conflicts between AGVs. The purpose of the time window analysis is to determine the time order of each AGV when performing tasks to ensure that they do not occupy the same resources or paths at the same time. For example, if two AGVs need to pass through the same narrow passage in adjacent time periods, the system will adjust their schedules so that one AGV passes first while the other AGV waits to pass to avoid collisions. Through this analysis and adjustment, a more robust and effective initial task scheduling result is obtained.

[0110] In one example, the RFID scanning data of the stored goods and the AGV onboard sensor data are obtained and integrated for positioning, and the real-time accurate location and storage status information of the goods are obtained, including:

[0111] Periodically scanning the RFID tags of the stored goods through the RFID reader array to obtain RFID signal strength data, and preprocessing the RFID signal strength data to obtain cleaned RFID data;

[0112] The AGV on-board laser radar scans the surrounding environment to obtain point cloud data of the environment, and then filters and aligns the point cloud data of the environment to obtain the three-dimensional spatial information of the AGV surrounding environment;

[0113] Based on the acceleration and angular velocity data of the AGV's onboard inertial measurement unit, dead reckoning is performed to obtain the AGV's initial position and posture information, and the cleaned RFID data is input into the Neural XGB 3D positioning system to obtain the RFID-based cargo position estimate;

[0114] Extract and match the features of the three-dimensional spatial information of the AGV's surrounding environment, combine the AGV's initial position and posture information, accurately locate the AGV itself, and obtain the AGV's real-time position;

[0115] Based on the real-time position of the AGV, coordinate transformation and correction are performed on the estimated position of the goods based on RFID to obtain the global coordinates of the goods;

[0116] Match the global coordinates of the goods with the warehouse layout information to identify the specific shelf location of the goods, and combine the environmental data of the AGV vehicle-mounted temperature and humidity sensor to obtain the storage status information of the goods;

[0117] Perform time series analysis on the global coordinates and storage status information of the goods, identify abnormal change patterns, and obtain the real-time and accurate location and storage status information of the goods.

[0118] In this example, multiple RFID readers are arranged inside the warehouse to periodically scan all RFID tags in the warehouse and obtain the signal strength data of each tag. Each RFID tag will emit a signal within the scanning range of the reader, and the signal strength data S received by the reader is ij Reflects the signal strength of tag j at the position of reader i. The original RFID signal strength data is preprocessed, including outlier removal and signal smoothing. Outlier removal uses statistical methods, such as removing signal data points that exceed three times the standard deviation, or using a dynamic threshold method to identify data points that deviate significantly from the average as outliers and remove them. Signal smoothing uses a moving average or exponential smoothing method to eliminate short-term fluctuations in signal strength and obtain more stable signal strength data S′ ij That is, the RFID data after cleaning. At the same time, the laser radar on the AGV scans its surrounding environment in real time to generate three-dimensional point cloud data. Point cloud data Each point (x i ,y i , z i ) represents the three-dimensional coordinates of the surface of an object scanned by the radar. The point cloud data is filtered and registered. Filtering removes outliers and noise points by statistical filtering or radius filtering. Registration processing is to align the point cloud data obtained by multiple scans into the same coordinate system to form a complete three-dimensional environmental model. The obtained three-dimensional spatial information reflects the actual environment around the AGV. The AGV is equipped with an inertial measurement unit (IMU) to collect the acceleration and angular velocity data of the AGV in real time. Based on these data, the initial position and attitude information of the AGV are calculated by dead reckoning. Dead reckoning continuously calculates the position and attitude of the AGV at each moment based on the known initial position (x0, y0, θ0) and acceleration and angular velocity information. Assuming the time interval is Δt, the position and attitude of the AGV at time t are calculated by the following formula:

[0119] x t =x t-1 +v t-1 ·cos(θ t-1 )·Δt;

[0120] y t =y t-1 +v t-1 ·sin(θ t-1 )·Δt;

[0121] θ t =θ t-1 +ω t-1 Δt;

[0122] Among them, (x t-1 ,yt-1 ,θ t-1 ) is the position information at the previous moment, v t-1 and ω t-1 are the speed and angular velocity of the previous moment respectively. Through the above recursive relationship, the initial position and posture information of the AGV at each moment are obtained. ij Input into the neural XGB 3D positioning system. The neural XGB 3D positioning system combines the feature extraction capability of the neural network and the regression capability of XGBoost to learn complex nonlinear relationships from the cleaned RFID signal strength data and infer the position of the RFID tag in 3D space. Using the signal strength data of each tag as input features, the neural network extracts the features and inputs them into the XGBoost regressor, which ultimately outputs the estimated 3D coordinates of the tag. Obtain the estimated value of the cargo position based on RFID. Perform feature extraction and matching based on the three-dimensional spatial information of the AGV's surrounding environment, and accurately locate the AGV's own position by combining the AGV's initial position and posture information. The feature extraction and matching process includes extracting key feature points (such as corner points and boundary points) from the point cloud data and matching them with the known warehouse environment model. Based on the matching results, correct the AGV's initial position information to obtain the AGV's real-time accurate position (x′ t ,y′ t ,θ′ t ). Real-time location based on AGV, estimated location of goods based on RFID Perform coordinate transformation and correction to obtain the global coordinates of the goods. Assume that the current position of the AGV in the global coordinate system is (x′ t ,y′ t ,θ′ t ), and the position of the cargo in the local coordinate system of the AGV is Then the coordinates of the cargo in the global coordinate system are j =(x j ,y j ,z j ) is obtained by the following coordinate transformation formula:

[0123]

[0124] Among them, θ′ tis the current orientation angle of the AGV. Through this transformation, the local coordinates of the goods are converted into global coordinates to obtain the exact location of the goods in the warehouse. The global coordinates of the goods are matched with the warehouse layout information to identify the specific shelf location of the goods. The warehouse layout information includes data such as the number, location and size of the shelf. By comparing the global coordinates of the goods with this information, it is determined that the goods are stored in a specific shelf unit. At the same time, combined with the environmental data collected by the AGV on-board temperature and humidity sensor, including information such as temperature and humidity in the warehouse, the environmental data is associated with the location of the goods to obtain the storage status information of the goods. Time series analysis is performed on the global coordinates and storage status information of the goods to identify potential abnormal change patterns. Time series analysis methods include moving average, exponential smoothing or Kalman filtering, etc., which detect abnormal movement or environmental changes by analyzing the trend of the location and status of the goods over time. For example, if the location of a certain product suddenly changes significantly, or the temperature and humidity data exceed the normal range, the system will issue an alarm to prompt the warehouse manager to check. After analysis and processing, the real-time accurate location and storage status information of the goods are finally obtained.

[0125] In one example, the real-time accurate location and storage status information of the goods are input into the encoder-decoder model for AGV path planning and sorting decision calculation to obtain the optimal sorting strategy and AGV motion trajectory, including:

[0126] Convert the real-time accurate location and storage status information of the goods into vector form to obtain the goods status vector, and then concatenate the goods status vector with the current AGV location information and warehouse layout data to obtain the input feature vector;

[0127] The input feature vector is normalized to obtain standardized input data, and the standardized input data is fed into the encoder part of the encoder-decoder model; the encoder part contains 5 convolutional layers, each of which uses a 3x3 convolution kernel and a stride of 2, followed by BatchNormalization and ReLU activation functions, which successively reduce the input feature dimension and extract high-level features to obtain the encoded feature representation;

[0128] The encoded feature representation is input into the attention mechanism layer. The attention mechanism layer adopts a multi-head self-attention structure, which includes 8 attention heads. Each head independently calculates the attention weight and performs feature weighting. Finally, the outputs of the 8 heads are spliced ​​to obtain the attention-enhanced feature representation.

[0129] The attention-enhanced feature representation is input into the decoder part, which contains 5 transposed convolutional layers. Each transposed convolutional layer uses a 3x3 convolution kernel and a deconvolution operation with a step size of 2, followed by BatchNormalization and LeakyReLU activation functions to gradually restore the feature dimension and obtain the predicted path planning result.

[0130] The predicted path planning results are input into the classification head and regression head. The classification head uses the Softmax function to output the probability distribution of AGV action categories, and the regression head outputs continuous AGV motion parameters. The combined results are used to obtain the AGV motion trajectory.

[0131] Based on the AGV motion trajectory and the current warehouse status, multiple candidate sorting strategies are generated, and the expected benefit of each candidate sorting strategy is evaluated through the reward function, and the strategy with the highest benefit is selected as the optimal sorting strategy.

[0132] In this example, the relevant information of the goods is represented as a specific vector structure. Assuming that the real-time accurate location of the goods is represented by three-dimensional coordinates (x, y, z), and its storage state information includes environmental parameters such as temperature and humidity (T, H), then the state vector V of the goods is h Defined as:

[0133] V h =[x,y,z,T,H];

[0134] The current AGV location information (x AGv ,y AGV ,θ AGV ) is spliced ​​into the cargo state vector, and the warehouse layout data is encoded into a vector form. For example, discrete information such as shelf location and channel location is converted into a vector form L through One-Hot encoding to obtain a complete input feature vector V:

[0135] V = [x, y, z, T, H, x AGV ,y AGV ,θ AGV , L];

[0136] This vector contains all the key attributes of the goods, the current status of the AGV, and warehouse layout information. The input feature vector V is normalized to map all features to a standard value range (such as [-1, 1] or [0, 1]) to avoid some features from having an adverse effect on the model due to excessive values ​​when training the model. Normalization usually uses the following formula:

[0137]

[0138] Among them, V i represents the original value of the i-th feature, μ iand σ i are the mean and standard deviation of the feature, V′ i is the normalized feature value. After normalization, the standardized input data V′ is obtained. The standardized input data V′ is input to the encoder part of the encoder-decoder model. The encoder part contains 5 convolutional layers, each of which uses a 3x3 convolution kernel and performs convolution operations with a step size of 2. The convolution operation can extract local features of the input data, while reducing the feature dimension layer by layer, and improving the model's ability to abstract high-level features. The convolution operation is expressed as:

[0139] h l =f(W l *h l-1 +b l );

[0140] Among them, h l is the output of the l-th convolution operation, W l is the convolution kernel weight matrix of layer l, b l is the bias term, * indicates the convolution operation, and f is the ReLU activation function. The ReLU activation function can introduce nonlinear transformations and improve the expressiveness of the model. Each convolution layer is followed by a BatchNormalization layer to normalize the convolution output and improve the training stability and convergence speed of the model. After 5 layers of convolution and normalization, the dimension of the input features is gradually reduced, and more abstract and meaningful high-level features are extracted to obtain the encoded feature representation H. enc The encoded features are represented by H enc Input to the attention mechanism layer. The attention mechanism layer adopts a multi-head self-attention structure, which contains 8 attention heads. Each attention head independently calculates the attention weight and weights the features. For each attention head k, its output is expressed as:

[0141]

[0142] Among them, Q (k) , K (k) and V (k) They are the encoded feature representations H enc The query matrix, key matrix and value matrix obtained by linear transformation are d k is a scaling factor used to prevent the attention weight from being too large, which may cause the gradient to disappear or explode. The weighted attention output A (k) Capture different patterns and important information in the input features. Concatenate the outputs of the eight attention heads to obtain the attention-enhanced feature representation H att . Attention-enhanced feature representation H attIt will be input to the decoder part, which is symmetrical with the encoder part and also contains 5 transposed convolution layers. Each transposed convolution layer uses a 3x3 convolution kernel and performs deconvolution operation with a step size of 2 to gradually restore the spatial dimension of the feature. The transposed convolution operation is understood as the inverse process of convolution, which is used to map low-dimensional features back to high-dimensional space while retaining the high-level features extracted during the encoding process. Each transposed convolution layer is also followed by a BatchNormalization layer and a LeakyReLU activation function. LeakyReLU can give a slight negative slope when the input is less than zero, thereby preventing the gradient vanishing problem. After 5 layers of deconvolution operations, the decoder outputs a feature representation with the same dimension as the original input, represented as the predicted path planning result H dec . The predicted path planning result H dec The classification head contains multiple fully connected layers and uses the Softmax function to output the probability distribution P of the AGV action category. class The Softmax function converts the score of each action category into a probability distribution, which is used to represent the probability of AGV choosing different actions, such as moving, turning, stopping, and other action categories. The output formula of the classification head is:

[0143] P class =softmax(W class ·H dec +b class );

[0144] Among them, W class is the weight matrix of the classification head, b class is the bias term. The regression head outputs the continuous motion parameter Y of the AGV reg , including speed, steering angle and other information. The output formula of the regression head is:

[0145] Y reg =W reg ·H dec +b reg ;

[0146] Among them, W reg is the weight matrix of the regression head, b reg is the bias term. The motion trajectory of the AGV is generated by combining the probability distribution of the action categories output by the classification head with the continuous motion parameters output by the regression head. After the AGV motion trajectory is generated, multiple candidate sorting strategies are generated in combination with the current state of the warehouse. Each sorting strategy represents a different cargo picking and transportation path. By setting the reward function R, the expected benefit of each candidate sorting strategy is evaluated. The design of the reward function takes into account multiple factors of warehouse operation, such as picking efficiency, energy consumption, path conflict, etc. Let the expected benefit of the i-th sorting strategy be E i, and its calculation formula is:

[0147]

[0148] Among them, γ is the discount factor, which controls the decay of future rewards, R(s t ,a t ) means in state s t Next, perform action a t The immediate reward obtained. Select the sorting strategy E with the highest expected return best =max(E i ) as the optimal sorting strategy, thereby guiding AGV to efficiently complete the sorting and transportation tasks of goods.

[0149] In one example, the predicted path planning result is input into the classification head and regression head. The classification head uses the Softmax function to output the probability distribution of AGV action categories, and the regression head outputs continuous AGV motion parameters. The AGV motion trajectory is obtained by combining:

[0150] The predicted path planning result is mapped to the action space dimension through a fully connected layer to obtain an action feature vector, and the action feature vector is dimensionally segmented to divide the action feature vector into a discrete action part and a continuous action part to obtain discrete action features and continuous action features;

[0151] The discrete action features are input into the classification head network. The classification head network consists of two fully connected layers, each followed by a ReLU activation function, and the last layer uses a Softmax function to obtain the probability distribution of AGV action categories;

[0152] The continuous motion features are input into the regression head network, which contains two fully connected layers, each followed by a Tanh activation function to obtain continuous AGV motion parameters.

[0153] Based on the probability distribution of AGV action categories, discrete action sampling is performed to obtain discrete action selection results, and the discrete action selection results are combined with continuous AGV motion parameters to obtain the AGV motion trajectory.

[0154] In this example, the predicted path planning result is converted into a feature vector of the action space through a fully connected layer. Assume that the predicted path planning result is H dec , and maps it to the action space dimension through a fully connected layer to obtain an action feature vector A. The function of the fully connected layer is to compress the high-dimensional features into a low-dimensional representation in the action space for subsequent processing. The specific calculation formula is:

[0155] A=W act ·H dec +b act ;

[0156] Among them, W act is the weight matrix of the fully connected layer, representing the linear transformation in the feature mapping process, b act is a bias term, which is used to translate and adjust the mapping result. The obtained action feature vector A contains all the information related to the action performed by the AGV. The action feature vector A is dimensionally split into a discrete action part and a continuous action part. The discrete action part contains information about the AGV in different action categories, such as forward, backward, turning, stopping, etc.; while the continuous action part represents specific action parameters, such as forward speed, turning angle, etc. Assume that the dimension of the action feature vector A is d, the dimension of the discrete action part is d1, and the dimension of the continuous action part is d2, satisfying d=d1+d2. Then the discrete action feature A dis and continuous action feature A con It is expressed as:

[0157]

[0158] Among them, a i Represents the i-th element in the action feature vector A. By segmenting, different types of action information are processed independently. dis Input to the classification head network. The classification head network consists of two fully connected layers, each of which is followed by a ReLU activation function, and the last layer uses a Softmax function. The first fully connected layer maps the discrete action features from the d1 dimension to the hidden layer dimension h1, and its output is:

[0159] h dis,1 =ReLU(W dis,1 ·A dis +b dis,1 );

[0160] Among them, W dis,1 is the weight matrix of the first fully connected layer, b dis,1 is the bias term. The ReLU activation function ReLU(x)=max(0,x) is used to introduce nonlinearity, enabling the model to learn more complex feature representations. The second fully connected layer converts the output h of the hidden layer into dis,1 Mapped to discrete action category space, the output is:

[0161] h dis,2 =W dis,2 ·h dis,1 +b dis,2 ;

[0162] System dis,2 Perform Softmax operation to obtain the probability distribution P of AGV in different action categories classThe calculation formula of the Softmax function is:

[0163]

[0164] Among them, h dis,2,i is the score of the ith action category, and k is the total number of action categories. The Softmax function converts the score of each action category into a probability value, which represents the probability of AGV choosing each action category. In this way, the selection tendency of AGV on different discrete actions is obtained. At the same time, the continuous action feature A con Input to the regression head network. The regression head network also consists of two fully connected layers, but each layer is followed by a Tanh activation function. The first fully connected layer maps the continuous action features from the d2 dimension to the hidden layer dimension h2, and its output is:

[0165] h con,1 =Tanh(W con,1 ·A con +b con,1 );

[0166] Among them, W con,1 is the weight matrix of the first fully connected layer, b con,1 is the bias term. Tanh activation function The function of is to limit the output value to the range of [-1, 1] to avoid the gradient explosion problem caused by too much data. The second fully connected layer converts the output h of the hidden layer into con，1 Mapped to the continuous action parameter space, the output is:

[0167] h con，2 =W con，2 ·h con，1 +b con，2 ;

[0168] The output of the regression head h con，2 Represents the predicted value Y of AGV on the continuous action parameter reg , such as the forward speed, the steering angle, etc. In this way, corresponding continuous motion parameters are provided for each discrete action category. class Discrete action sampling is performed to determine the specific discrete actions that the AGV should perform in the current situation. The sampling process uses a random sampling strategy, randomly selecting an action category a according to the probability value of each action category. Let the discrete action category set be {a1, a2, ..., a k}, each action category a i The probability of selection is P class (i) The system samples based on these probability values ​​and obtains the discrete action selection result a selected . The discrete action selection result aselected With continuous action parameter Y reg Combine them to generate a complete AGV motion trajectory. The motion trajectory contains the action category (such as forward, turn, stop) that the AGV should take at a certain moment and the corresponding motion parameters (such as forward speed, turn angle, etc.). Assuming that the current state of the AGV is S = (x, y, θ), based on the selected discrete action a selected and continuous action parameter Y reg , the next state of the AGV is obtained by S′=(x′, y′, θ′). For example, if the discrete action is selected as "forward" and the continuous action parameter is speed v, the next position of the AGV is calculated by the following formula:

[0169] x′=x+v·cos(θ)·Δt;

[0170] y′=y+v·sin(θ)·Δt;

[0171] Among them, Δt is the time interval, and θ is the current orientation angle of the AGV. In this way, a continuous AGV motion trajectory is generated.

[0172] Reference Figure 2 This embodiment provides a cloud warehouse-based intelligent storage device, including:

[0173] Scanning module 1 is used to perform all-round scanning and cargo feature recognition on the cargo entering the cloud warehouse to obtain cargo feature data;

[0174] Binding module 2 is used to assign and bind RFID tags to each piece of cargo according to cargo feature data, and perform initial positioning through the neural XGB 3D positioning system to obtain the 3D coordinate value of the cargo;

[0175] Calculation module 3, used to construct a multi-dimensional cargo feature vector based on cargo feature data and cargo three-dimensional coordinate values, and perform storage location optimization calculation through a deep reinforcement learning model to obtain a storage location allocation plan;

[0176] Planning module 4 is used to perform path planning and multi-agent task scheduling for AGV according to the storage location allocation scheme and the digital twin model of the cloud warehouse, and obtain the storage task execution instructions of AGV;

[0177] Positioning module 5, used to obtain the RFID scanning data of the stored goods and the AGV vehicle-mounted sensor data and perform fusion positioning to obtain the real-time accurate location and storage status information of the goods;

[0178] Processing module 6 is used to input the real-time accurate location and storage status information of the goods into the encoder-decoder model for AGV path planning and sorting decision calculation to obtain the optimal sorting strategy and AGV motion trajectory.

[0179] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0180] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0181] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0182] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0183] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0184] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0185] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation based on the contents of the present invention description and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An intelligent storage method based on cloud warehouse, characterized in that: The following steps are involved: Carry out all-round scanning and cargo feature recognition on the goods entering the cloud warehouse to obtain cargo feature data; Assign and bind an RFID tag to each piece of cargo according to the cargo characteristic data, and perform initial positioning through the neural XGB three-dimensional positioning system to obtain the three-dimensional coordinate value of the cargo; Based on the cargo feature data and the three-dimensional coordinate values ​​of the cargo, a multi-dimensional cargo feature vector is constructed, and a storage location optimization calculation is performed through a deep reinforcement learning model to obtain a storage location allocation plan; specifically, the following steps are performed: the visual feature vector, geometric feature data and physical feature data in the cargo feature data are normalized to obtain standardized feature data; the standardized feature data and the three-dimensional coordinate values ​​of the cargo are concatenated and dimensionally reduced to obtain a multi-dimensional cargo feature vector; based on the multi-dimensional cargo feature vector, a similarity matrix between the cargoes is calculated, and a clustering algorithm is used to group similar cargoes to obtain a cargo grouping result; the cargo grouping result is associated with the historical storage data to extract frequently shared cargoes. The current cargo combination is obtained to obtain cargo association rules, and based on the cargo association rules and the current warehouse storage status, a storage location candidate set is constructed, and the storage location candidate set is input into the state space of the deep reinforcement learning model; the storage location candidate set is evaluated using the policy network of the deep reinforcement learning model, a storage location allocation action is generated, and the expected benefit of each action is calculated through the value function network; based on the expected benefit, the optimal storage location allocation action is selected and applied to the current warehouse storage status to obtain an updated storage status; the updated storage status is evaluated to calculate the reward value, and the policy network and value function network of the deep reinforcement learning model are updated using temporal difference learning to obtain a storage location allocation plan; According to the storage location allocation scheme and the digital twin model of the cloud warehouse, the AGV is path planned and multi-agent task scheduled to obtain the storage task execution instruction of the AGV; specifically, the method comprises: based on the digital twin model of the cloud warehouse, a rasterized map of the warehouse environment is constructed, and the target location in the storage location allocation scheme is marked on the rasterized map to obtain a task map; the task map is subjected to obstacle expansion processing to obtain an expanded map; based on the expanded map, the initial path of the AGV from the current position to the target storage position is calculated to obtain the initial path planning result, and the initial path planning result is subjected to Bezier curve processing; The smoothed path is processed by smoothing the line to obtain a smooth path; the smoothed path is input into the time-varying artificial potential field model, and the obstacle avoidance calculation is performed on the dynamic obstacles to obtain a local path, and based on the local path and the current AGV queue state, a task allocation matrix for multi-AGV collaborative scheduling is constructed, and the initial task scheduling result is obtained by solving; the initial task scheduling result is converted into a sequence of movement, steering, grabbing and placing operation instructions executable by the AGV to obtain the storage task execution instruction of the AGV, and the storage task execution instruction will be sent to the control system of the AGV to guide the AGV to perform the task according to the predetermined path and sequence, thereby completing the handling and storage operations of the goods; Obtain the RFID scanning data of the stored goods and the AGV vehicle-mounted sensor data and perform fusion positioning to obtain the real-time accurate location and storage status information of the goods; The real-time accurate location and storage status information of the goods are input into the encoder-decoder model for AGV path planning and sorting decision calculation to obtain the optimal sorting strategy and AGV motion trajectory.

2. The cloud warehouse-based intelligent storage method according to claim 1 is characterized in that: The goods entering the cloud warehouse are scanned in all directions and their characteristics are identified to obtain the characteristic data of the goods, including: Capture high-resolution images of goods entering the cloud warehouse from multiple angles to obtain multi-view image data, and perform image denoising, illumination compensation and geometric correction on the multi-view image data to obtain standardized image data; Performing deep convolution operation and feature extraction on the standardized image data to obtain a visual feature vector, and performing three-dimensional scanning on the goods to obtain point cloud data; Performing noise reduction and registration processing on the point cloud data to obtain a three-dimensional shape model, and calculating the volume, surface area and main axis direction of the cargo based on the three-dimensional shape model to obtain geometric feature data; Weighing the goods using an electronic scale to obtain weight data, and combining the weight data with the geometric characteristic data to calculate the density distribution of the goods to obtain physical characteristic data; Performing feature fusion and classification on the visual feature vector, the geometric feature data and the physical feature data to obtain category, quantity and batch information of the goods; The category, quantity and batch information of the goods are encoded and combined with the visual feature vector, the geometric feature data and the physical feature data to generate goods feature data.

3. The cloud warehouse-based intelligent storage method according to claim 2 is characterized in that: The method of allocating and binding an RFID tag to each piece of cargo according to the cargo feature data, and performing initial positioning through the neural XGB three-dimensional positioning system to obtain the three-dimensional coordinate value of the cargo includes: Generate a unique RFID identification code based on the category, quantity and batch information in the cargo characteristic data, and write the RFID identification code into the RFID tag; The RFID tag is attached to the surface of the goods, an association relationship between the goods and the RFID identification code is established in the cloud warehouse management system, and multi-point signal strength collection of the RFID tag is performed through an RFID reader array to obtain signal strength distribution data; Perform outlier removal and signal smoothing on the signal intensity distribution data to obtain standardized signal intensity data, and input the standardized signal intensity data into a neural network module in the neural XGB three-dimensional positioning system to perform feature extraction and preliminary position estimation to obtain an initial position estimation value; Based on the initial position estimate, the XGB regressor is used to refine the position to obtain a revised three-dimensional coordinate prediction value, and the revised three-dimensional coordinate prediction value is constrained and matched with the warehouse layout information to ensure that the coordinate falls within the valid area to obtain the initial three-dimensional coordinate; The initial three-dimensional coordinates are subjected to time series filtering to obtain three-dimensional coordinate values ​​of the goods.

4. The cloud warehouse-based intelligent storage method according to claim 1 is characterized in that: The smooth path is input into the time-varying artificial potential field model, and the dynamic obstacle avoidance calculation is performed to obtain a local path. Based on the local path and the current AGV queue state, a task allocation matrix for multi-AGV collaborative scheduling is constructed to obtain the initial task scheduling result by solving the problem, including: The smooth path is discretized into a sequence of path points, and the gravitational field and the repulsive field are calculated for each path point to obtain an initial artificial potential field distribution, and the initial artificial potential field distribution is dynamically updated based on the real-time position information of the AGV and the dynamic obstacle to obtain a time-varying artificial potential field model; In the time-varying artificial potential field model, the gradient descent method is used to calculate the movement direction and speed of the AGV to obtain a local path, and the local path is segmented. Each segment of the path is used as a subtask to construct an AGV task set and obtain a task list to be assigned. Based on the current AGV queue status, including the position, power and load of the AGV, an AGV status matrix is ​​constructed and combined with the task list to be assigned to construct a task allocation matrix for multi-AGV collaborative scheduling; The Hungarian algorithm is used to solve the task allocation matrix of the multi-AGV collaborative scheduling to obtain the optimal task allocation scheme, and a time window analysis is performed on the optimal task allocation scheme to detect and resolve potential interference and conflicts between AGVs to obtain the initial task scheduling result.

5. The cloud warehouse-based intelligent storage method according to claim 1 is characterized in that: The method of acquiring the RFID scanning data of the stored goods and the AGV vehicle-mounted sensor data and performing fusion positioning to obtain the real-time accurate location and storage status information of the goods includes: Periodically scanning the RFID tags of the stored goods through an RFID reader array to obtain RFID signal strength data, and preprocessing the RFID signal strength data to obtain cleaned RFID data; Scan the surrounding environment based on the AGV on-board laser radar to obtain point cloud data of the environment, and filter and align the point cloud data of the environment to obtain three-dimensional spatial information of the AGV surrounding environment; Based on the acceleration and angular velocity data of the AGV on-board inertial measurement unit, dead reckoning is performed to obtain the initial position and posture information of the AGV, and the cleaned RFID data is input into the neural XGB three-dimensional positioning system to obtain the RFID-based cargo position estimation value; Extract and match the three-dimensional spatial information of the AGV's surrounding environment, and accurately locate the AGV itself based on the AGV's initial position and posture information to obtain the AGV's real-time position. Based on the real-time position of the AGV, coordinate transformation and correction are performed on the estimated value of the cargo position based on the RFID to obtain the global coordinates of the cargo; Match the global coordinates of the goods with the warehouse layout information, identify the specific shelf location of the goods, and combine the environmental data of the AGV vehicle-mounted temperature and humidity sensor to obtain the storage status information of the goods; The global coordinates and storage status information of the goods are analyzed in time series to identify abnormal change patterns and obtain the real-time accurate location and storage status information of the goods.

6. The cloud warehouse-based intelligent storage method according to claim 1, characterized in that: The real-time accurate location and storage status information of the goods are input into the encoder-decoder model for AGV path planning and sorting decision calculation to obtain the optimal sorting strategy and AGV motion trajectory, including: Convert the real-time accurate location and storage status information of the goods into vector form to obtain a goods status vector, and concatenate the goods status vector with the current AGV location information and warehouse layout data to obtain an input feature vector; Normalizing the input feature vector to obtain standardized input data, and inputting the standardized input data into the encoder part of the encoder-decoder model; the encoder part comprises 5 convolutional layers, each of which uses a convolution operation with a 3x3 convolution kernel and a step size of 2, followed by BatchNormalization and ReLU activation functions, which successively reduce the input feature dimension and extract high-level features to obtain encoded feature representations; Input the encoded feature representation into the attention mechanism layer, which adopts a multi-head self-attention structure and includes 8 attention heads. Each head independently calculates the attention weight and performs feature weighting. Finally, the outputs of the 8 heads are spliced ​​to obtain an attention-enhanced feature representation; The attention-enhanced feature representation is input into a decoder part, which includes 5 transposed convolutional layers, each of which uses a 3x3 convolution kernel and a deconvolution operation with a step size of 2, followed by BatchNormalization and LeakyReLU activation functions to gradually restore the feature dimension and obtain a predicted path planning result; The predicted path planning result is input into the classification head and the regression head. The classification head uses the Softmax function to output the probability distribution of the AGV action category, and the regression head outputs the continuous AGV motion parameters, which are combined to obtain the AGV motion trajectory; Based on the AGV motion trajectory and the current warehouse status, multiple candidate sorting strategies are generated, and the expected benefit of each candidate sorting strategy is evaluated through a reward function, and the strategy with the highest benefit is selected as the optimal sorting strategy.

7. The cloud warehouse-based intelligent storage method according to claim 6 is characterized in that: The predicted path planning result is input into the classification head and the regression head. The classification head uses the Softmax function to output the probability distribution of the AGV action category, and the regression head outputs continuous AGV motion parameters. The AGV motion trajectory is obtained by combining them, including: Mapping the predicted path planning result to the action space dimension through a fully connected layer to obtain an action feature vector, and performing dimension segmentation on the action feature vector to divide the action feature vector into a discrete action part and a continuous action part to obtain discrete action features and continuous action features; The discrete action features are input into a classification head network, which includes two fully connected layers, each layer is followed by a ReLU activation function, and the last layer uses a Softmax function to obtain the probability distribution of AGV action categories; The continuous motion features are input into a regression head network, wherein the regression head network comprises two fully connected layers, each layer is followed by a Tanh activation function, and continuous AGV motion parameters are obtained; Based on the probability distribution of the AGV action categories, discrete action sampling is performed to obtain discrete action selection results, and the discrete action selection results are combined with the continuous AGV motion parameters to obtain the AGV motion trajectory.

8. A cloud warehouse-based intelligent storage device, characterized in that: For implementing the steps of the method according to any one of claims 1 to 7, the device comprises: The scanning module is used to perform all-round scanning and cargo feature recognition on the cargo entering the cloud warehouse to obtain cargo feature data; A binding module is used to assign and bind an RFID tag to each piece of cargo according to the cargo characteristic data, and perform initial positioning through the neural XGB three-dimensional positioning system to obtain the three-dimensional coordinate value of the cargo; A calculation module, used to construct a multi-dimensional cargo feature vector based on the cargo feature data and the three-dimensional coordinate values ​​of the cargo, and perform storage location optimization calculation through a deep reinforcement learning model to obtain a storage location allocation plan; A planning module, used to perform path planning and multi-agent task scheduling for the AGV according to the storage location allocation scheme and the digital twin model of the cloud warehouse, and obtain storage task execution instructions for the AGV; The positioning module is used to obtain the RFID scanning data of the stored goods and the AGV vehicle-mounted sensor data and perform fusion positioning to obtain the real-time accurate location and storage status information of the goods; The processing module is used to input the real-time accurate location and storage status information of the goods into the encoder-decoder model to perform AGV path planning and sorting decision calculation to obtain the optimal sorting strategy and AGV motion trajectory.

Images