Storage automated guided vehicle positioning and formation control method

Abstract

The invention relates to a storage automated guided vehicle positioning and formation control method. Comprising the following steps: arranging a plurality of UWB radar base stations in a storage environment, installing a UWB responder on each automatic guided vehicle, and outputting original UWB echo signal time sequence data; performing data processing on the original UWB echo signal time sequence data; based on a three-dimensional convolutional neural network, spatial-temporal characteristics of the automated guided vehicle are extracted from the high-quality UWB echo data set; inputting the spatial-temporal characteristics into a bidirectional gating circulation unit network, and outputting a predicted three-dimensional position and a predicted three-dimensional speed of the automatic guided vehicle; and based on an intelligent scheduling algorithm, outputting an optimal task allocation scheme and a global path plan including a formation instruction according to batch task instructions of the warehouse management system, current positions of all the automatic guided vehicles, predicted three-dimensional positions, a speed adjustment instruction, a steering adjustment instruction and respective battery power. According to the invention, formation control of the automatic guided vehicles can be realized.

Description

Technical Field

[0001] This application relates to the field of positioning and formation control technology, and in particular to a method for positioning and formation control of unmanned warehouse transport vehicles. Background Technology

[0002] With the booming development of global e-commerce and the deepening of digital transformation of supply chains, intelligent warehousing has become a core infrastructure of modern logistics systems. Against this backdrop, miniature unmanned transport vehicles (UGVs), as key execution units in warehouse automation, undertake core tasks such as sorting, handling, and delivery of goods in high-density racking environments. Their positioning accuracy, platooning coordination capabilities, and operational efficiency directly determine the operational efficiency, throughput capacity, and economic benefits of the entire warehouse management system. With the surge in demand for fragmented orders and on-demand delivery, warehousing operations are evolving towards multi-vehicle collaboration, high-density deployment, and complex path planning, placing unprecedentedly stringent requirements on the real-time positioning accuracy, platooning control stability, and environmental adaptability of miniature UGVs.

[0003] Among related technologies, the positioning technology of miniature unmanned transport vehicles can be mainly divided into optical vision positioning, lidar positioning, and exploratory wireless signal positioning methods.

[0004] Optical vision positioning relies on vehicle-mounted cameras to capture environmental feature points or pre-set QR codes and reflective markers, and then uses image processing algorithms to estimate pose. This method can achieve a certain level of positioning accuracy in well-lit and simple environments, but its performance is severely limited in practical applications of smart warehousing: frequent occlusion caused by high-density shelves and complex stacking of goods can lead to the loss of feature points; low light and drastic changes in lighting caused by nighttime operations, backlighting, and shelf shadows can drastically reduce image quality; lens fogging and sensor drift caused by temperature fluctuations can introduce additional errors; and the high computational complexity of visual algorithms makes it difficult to achieve high-frequency real-time positioning on the limited computing power platform of a micro unmanned transport vehicle.

[0005] LiDAR positioning works by emitting laser beams and receiving reflected signals from the environment to construct a two-dimensional or three-dimensional point cloud map for localization. LiDAR has advantages such as high ranging accuracy and immunity to lighting conditions, but it also faces many challenges in high-density warehouse environments: stacked goods and shelving obstructions can create numerous holes and artifacts in the point cloud; environmental changes caused by dynamic obstacles can lead to map mismatch; the coexistence of specular and diffuse reflections on metal shelving surfaces can cause ranging jitter; and high-performance LiDAR is expensive, making it uneconomical for large-scale deployment of miniature automated guided vehicles (AGVs).

[0006] Exploratory methods based on wireless signals, such as Wi-Fi, Bluetooth, and UWB, for indoor positioning have become a research hotspot. However, existing methods are mostly limited to the static positioning verification stage for single vehicles, and still have many shortcomings when applied to multi-vehicle platooning and dynamic operation scenarios: most studies are based on simplified two-dimensional planar models, which cannot handle the three-dimensional positioning requirements caused by the three-dimensional structure of shelves and the vertical movement of miniature unmanned transport vehicles; signal fading and time delay jitter caused by multipath reflection and non-line-of-sight propagation can amplify positioning errors to the decimeter or even meter level; and existing algorithms mostly rely on empirical models or simple geometric calculations, lacking the ability to deeply model the electromagnetic propagation characteristics of complex environments, and even more so lacking the ability to intelligently perceive and predict multi-vehicle platooning cooperative modes and dynamic interaction with obstacles.

[0007] Although the methods described above in the existing technology have made progress in the field of positioning of miniature unmanned transport vehicles, the following main drawbacks still exist in the practical application of miniature transport vehicles in smart warehousing: (1) Inadequate positioning accuracy and insufficient real-time performance: Traditional methods are difficult to achieve high-precision real-time positioning with lateral and longitudinal errors ≤5cm and update frequency ≥10Hz in complex environments such as obstruction, low light, and multipath interference. They cannot meet the needs of passage through dense shelving gaps and precise docking of goods. (2) Lack of coordination in formation control and unstable formation: Existing methods are mostly single-vehicle positioning, lacking in-depth modeling of the spatial relationship, motion coupling and coordination mode of multiple vehicles, resulting in slow response to formation switching, and even the risk of collision and loss of control. (3) Poor environmental adaptability and insufficient robustness: Existing algorithms have weak ability to suppress interference factors such as metal shelf obstruction, multipath reflection, and temperature fluctuation, and lack online adaptive and anomaly correction mechanisms, resulting in severe performance degradation when the storage environment changes dynamically. (4) Unintelligent obstacle perception and non-proactive obstacle avoidance: Most existing positioning warehouse management systems only output location coordinates and lack the ability to identify, predict trajectories and actively avoid static and dynamic obstacles in real time. They cannot complete path replanning within 50ms, which poses a safety hazard.

[0008] Therefore, it is necessary to provide a new technical solution to improve one or more of the problems existing in the above solutions.

[0009] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0010] The purpose of this application is to provide a method for positioning and platooning control of unmanned warehouse transport vehicles, thereby overcoming, to at least some extent, one or more problems caused by the limitations and defects of related technologies.

[0011] A method for positioning and grouping control of unmanned transport vehicles in warehouses, according to an embodiment of this application, includes: Multiple UWB radar base stations are deployed in the warehouse environment, and a UWB responder is installed on each unmanned transport vehicle. The UWB responder obtains distance information by initiating pulse signal exchange and performing time difference ranging with multiple radar base stations. It also collects the echo signals of each UWB radar base station, records multiple sets of distance data corresponding to the unmanned transport vehicle at each moment, and outputs the original UWB echo signal time sequence data and the calculated distance-time matrix. Data processing is performed on the original UWB echo signal timing data to obtain a high-quality UWB echo dataset; Based on a three-dimensional convolutional neural network, the spatiotemporal features of the unmanned transport vehicle are extracted from the high-quality UWB echo dataset; The spatiotemporal features are input into a bidirectional gated recurrent unit network, which outputs the predicted three-dimensional position and predicted three-dimensional velocity of the unmanned transport vehicle. The spatiotemporal features are input into a convolutional attention mechanism to output an obstacle state probability map. Based on the predicted three-dimensional position and the predicted three-dimensional velocity, a dynamic interaction diagram of the unmanned transport vehicle is obtained; The dynamic interaction relationship diagram and the obstacle state probability diagram are used to output the speed adjustment command and steering adjustment command of the unmanned transport vehicle. Based on the intelligent scheduling algorithm, according to the batch task instructions of the warehouse management system, the current position of all unmanned transport vehicles, the predicted three-dimensional position, the speed adjustment instructions, the steering adjustment instructions, and the battery level of each vehicle, the optimal task allocation scheme and the global path planning including the formation instructions are output.

[0012] In the embodiments of this application, the step of extracting the spatiotemporal features of the unmanned transport vehicle from the high-quality UWB echo dataset based on a three-dimensional convolutional neural network includes: The high-quality UWB echo dataset is constructed into a three-dimensional data cube according to the base station ID, time window, and signal characteristics, and then the three-dimensional data cube is input into the three-dimensional convolutional neural network. The aforementioned three-dimensional convolutional neural network extracts the spatiotemporal features of the unmanned transport vehicle's movement across different radar base stations in space and continuous time dimensions to capture its movement patterns through spatiotemporal convolution operations.

[0013] In the embodiments of this application, the step of inputting the spatiotemporal features into a bidirectional gated recurrent unit network and outputting the predicted three-dimensional position and predicted three-dimensional velocity of the unmanned transport vehicle includes: The spatiotemporal features are input into the bidirectional gated loop unit network. Through the update gate and reset gate of the bidirectional gated loop unit network, the motion data sequence in the spatiotemporal features of the unmanned transport vehicle is processed forward and backward to capture the time dependence in the motion trajectory of the unmanned transport vehicle. The output layer of the gated loop unit network regresses and calculates the predicted 3D position and predicted 3D velocity of the unmanned transport vehicle based on the hidden state of the bidirectional gated loop unit network.

[0014] In embodiments of this application, the step of obtaining the dynamic interaction diagram of the unmanned transport vehicle based on the predicted three-dimensional position and the predicted three-dimensional velocity includes: The unmanned transport vehicle is regarded as a node, and a feature vector is constructed for each node. The feature vector includes state information, which includes the predicted three-dimensional position, predicted three-dimensional velocity, and heading angle of the unmanned transport vehicle. Edges are dynamically constructed based on communication range and collision risk. The Euclidean distance between any two unmanned transport vehicles is calculated. When the Euclidean distance is less than a preset interaction radius, a connecting edge is established between the two unmanned transport vehicles. An adjacency matrix describing the topology at the current moment is generated. The adjacency matrix is ​​a dynamic interaction relationship graph generated according to time evolution. The connecting edge includes a distance vector and a relative velocity vector. The adjacency matrix represents the connection relationship, and the feature vectors at the nodes in the adjacency matrix represent state information.

[0015] In embodiments of this application, the step of outputting the speed adjustment command and steering adjustment command of the unmanned transport vehicle from the dynamic interaction relationship diagram and the obstacle state probability diagram includes: The dynamic interaction graph is input into the graph neural network, which aggregates the information of neighboring nodes through multi-layer graph convolution, enabling each unmanned transport vehicle node to perceive the status of its surrounding unmanned transport vehicles and obstacles in real time. The graph neural network learns dynamic interaction patterns between multiple unmanned transport vehicles, including following, avoiding, and maintaining formation. Based on the macroscopic path, formation instructions, and the obstacle state probability map, speed adjustment instructions and steering adjustment instructions for the unmanned transport vehicle are generated; wherein, the speed adjustment instructions and the steering adjustment instructions are used for maintaining formation and dynamic obstacle avoidance.

[0016] In the embodiments of this application, the step of outputting an optimal task allocation scheme and a global path planning including formation instructions based on the intelligent scheduling algorithm, according to the batch task instructions of the warehouse management system, the current positions of all unmanned transport vehicles, the predicted three-dimensional positions, the speed adjustment instructions, the steering adjustment instructions, and their respective battery levels, includes: Based on the batch task instructions of the warehouse management system, the current position of all unmanned transport vehicles, the predicted three-dimensional position, the speed adjustment instructions, the steering adjustment instructions, the battery power of each vehicle, and the distance factor of the Euclidean distance from the current position of the unmanned transport vehicle to the task target point, the remaining battery power factor, and the probability factor of the unmanned transport vehicle being assigned a simple task, a multi-objective cost function is constructed. A genetic algorithm is used to solve the multi-objective cost function to find the optimal task allocation scheme that minimizes the total global cost. Based on the optimal task allocation scheme, a macro-level global optimal path is planned for the formation of unmanned transport vehicles, key path nodes and desired formations are set, and a global path plan containing formation instructions is output.

[0017] In embodiments of this application, the method further includes: The residual sequence between the predicted 3D position and the current position of the unmanned transport vehicle is reconstructed based on the variational autoencoder in order to detect whether any anomalies occur. If an anomaly occurs, a self-calibration trigger command and a high-quality UWB echo dataset will be output.

[0018] In the embodiments of this application, after the step of outputting the optimal task allocation scheme and global path planning including formation instructions based on the intelligent scheduling algorithm according to the batch task instructions of the warehouse management system, the current position of all unmanned transport vehicles, the predicted three-dimensional position, the speed adjustment instructions, the steering adjustment instructions, and their respective battery levels, the method further includes: The warehouse management system updates its real-time tasks and closed-loop control instructions based on the optimal task allocation scheme and the global path planning. The hyperparameters in the graph neural network are then optimized to obtain the optimized hyperparameters.

[0019] The technical solutions provided by the embodiments of this application may include the following beneficial effects: In one embodiment of this application, the above method achieves high-precision real-time positioning in complex warehousing environments by deploying multiple UWB base stations to form a coverage network and installing UWB responders on each unmanned transport vehicle (ARTV). Time-difference ranging (TDAR) generates raw UWB echo signal timing data. Processing the raw UWB echo signal timing data yields a high-quality dataset, which is then used for feature extraction and state prediction based on a 3D convolutional neural network, a bidirectional gated recurrent unit network, and a convolutional attention mechanism. This effectively reduces the impact of complex environments such as shelf obstruction and multipath interference on positioning and control, eliminating the need for additional reflective markers or manual intervention, and is applicable to warehousing scenarios of different layouts and scales. Furthermore, the convolutional attention mechanism outputs an obstacle state probability map, which, combined with the dynamic interaction graph of the ARTV, generates speed and steering adjustment commands, enabling dynamic obstacle avoidance and collaborative operation between vehicles. Simultaneously, an intelligent scheduling algorithm integrates multi-dimensional information such as task commands, vehicle positions, and battery levels to output optimal task allocation and global path planning, avoiding collision risks within the platoon and improving the stability and safety of platoon operation. The intelligent scheduling algorithm achieves optimal allocation and path planning for batch tasks. Combined with the formation control strategy, it can coordinate the operation process of multiple unmanned transport vehicles, reduce vehicle empty running rate and waiting time, improve the overall efficiency of warehouse cargo handling, and reduce labor costs and operating costs.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0022] Figure 1 This schematically illustrates a flowchart of the steps of a warehouse unmanned transport vehicle positioning and formation control method in an exemplary embodiment of this application; Figure 2 This illustration schematically shows a comparison curve of high-precision positioning errors in an exemplary embodiment of this application; Figure 3 This illustration schematically shows a bar chart comparing the stability of unmanned transport vehicle formation control in an exemplary embodiment of this application; Figure 4 The diagram illustrates the anomaly detection effect of the variational autoencoder in an exemplary embodiment of this application. Detailed Implementation

[0023] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0024] Furthermore, the accompanying drawings are merely illustrative of this application and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0025] This example implementation provides a method for positioning and platooning control of unmanned transport vehicles in warehouses. (Reference) Figure 1 As shown, the method may include: Step S101: Deploy multiple UWB radar base stations in the warehouse environment and install a UWB responder on each unmanned transport vehicle. The UWB responder initiates pulse signal exchange to obtain distance information by performing time difference ranging with multiple radar base stations, and collects the echo signals of each UWB radar base station. It records multiple sets of distance data corresponding to the unmanned transport vehicle at each moment, outputs the original UWB echo signal time sequence data, and the calculated distance-time matrix.

[0026] Step S102: Process the original UWB echo signal timing data to obtain a high-quality UWB echo dataset.

[0027] Step S103: Extract the spatiotemporal features of the unmanned transport vehicle from a high-quality UWB echo dataset based on a three-dimensional convolutional neural network.

[0028] Step S104: Input the spatiotemporal features into the bidirectional gated recurrent unit network and output the predicted 3D position and predicted 3D velocity of the unmanned transport vehicle.

[0029] Step S105: Input the spatiotemporal features into the convolutional attention mechanism and output the obstacle state probability map.

[0030] Step S106: Based on the predicted three-dimensional position and predicted three-dimensional velocity, obtain the dynamic interaction relationship diagram of the unmanned transport vehicle.

[0031] Step S107: Output the speed adjustment command and steering adjustment command of the unmanned transport vehicle from the dynamic interaction relationship diagram and obstacle state probability diagram.

[0032] Step S108: Based on the intelligent scheduling algorithm, according to the batch task instructions of the warehouse management system, the current position of all unmanned transport vehicles, the predicted three-dimensional position, speed adjustment instructions, steering adjustment instructions and their respective battery levels, output the optimal task allocation scheme and global path planning including formation instructions.

[0033] In one embodiment of this application, the above method achieves high-precision real-time positioning in complex warehousing environments by deploying multiple UWB base stations to form a coverage network and installing UWB responders on each unmanned transport vehicle (ARTV). Time-difference ranging (TDAR) generates raw UWB echo signal timing data. Processing the raw UWB echo signal timing data yields a high-quality dataset, which is then used for feature extraction and state prediction based on a 3D convolutional neural network, a bidirectional gated recurrent unit network, and a convolutional attention mechanism. This effectively reduces the impact of complex environments such as shelf obstruction and multipath interference on positioning and control, eliminating the need for additional reflective markers or manual intervention, and is applicable to warehousing scenarios of different layouts and scales. Furthermore, the convolutional attention mechanism outputs an obstacle state probability map, which, combined with the dynamic interaction graph of the ARTV, generates speed and steering adjustment commands, enabling dynamic obstacle avoidance and collaborative operation between vehicles. Simultaneously, an intelligent scheduling algorithm integrates multi-dimensional information such as task commands, vehicle positions, and battery levels to output optimal task allocation and global path planning, avoiding collision risks within the platoon and improving the stability and safety of platoon operation. The intelligent scheduling algorithm achieves optimal allocation and path planning for batch tasks. Combined with the formation control strategy, it can coordinate the operation process of multiple unmanned transport vehicles, reduce vehicle empty running rate and waiting time, improve the overall efficiency of warehouse cargo handling, and reduce labor costs and operating costs.

[0034] Below, we will refer to Figure 1 The steps of the method described above in this example embodiment will be explained in more detail.

[0035] In step S101, when deploying UWB radar base stations, they can be placed at key locations on the top or around the perimeter of the warehouse to form a radar array that fully covers the unmanned transport vehicle's operating area. UWB responders are installed on the unmanned transport vehicles to cooperate with the UWB radar base stations, collecting echo signals from each UWB radar base station, recording multiple sets of distance data corresponding to the unmanned transport vehicle at each moment, and outputting the raw UWB echo signal timing data. UWB stands for Ultra Wide Band.

[0036] This application can acquire high-precision time-series data by deploying UWB radar base stations and unmanned transport vehicle responders.

[0037] In step S102, the original UWB echo signal time series data is processed to improve its data quality. The data processing includes denoising and dataset augmentation.

[0038] In noise reduction processing, the channel impulse response of the acquired raw UWB echo timing data is generally preprocessed first. Then, a digital bandpass filter is used to filter out out-of-band electromagnetic interference from the environment, removing low-frequency drift caused by hardware instability and high-frequency environmental noise. Specifically, the adaptive weighting algorithm dynamically allocates weights based on the ratio of the first path power to the total received signal power: when the difference is small, it is determined to be line-of-sight propagation, and the base station signal is given a higher weight; when the difference is large, it is determined to be non-line-of-sight propagation, dominated by multipath effects, and the data management system automatically reduces the weight of this data, thereby improving the data quality of the effective signal. In the data augmentation part, virtual multipath simulation based on ray tracing is used: During data augmentation, virtual multipath simulation technology is used to augment the denoised original UWB echo signal. Specifically, a deterministic ray tracing algorithm is employed, based on a known 3D CAD layout of the warehouse, to simulate the propagation process of the UWB signal from the transmitter to the receiver. The algorithm calculates the direct path of the signal in space, as well as the first-order and second-order reflection and transmission paths occurring on metal shelves and concrete walls. It calculates the time delay based on the propagation distance and the energy attenuation based on the material's reflectivity. Finally, the signals from multiple paths are coherently superimposed to generate virtual channel impulse response waveform data containing complex multipath characteristics. This generates massive amounts of simulation data to augment the denoised original UWB echo signal (i.e., the training dataset), improving the robustness of the warehouse management system in real, complex environments. The final output is a high-quality UWB echo dataset, which includes cleaned real data and simulation data generated by ray tracing.

[0039] In step S103, the spatiotemporal features of the unmanned transport vehicle are extracted from the high-quality UWB echo dataset using a three-dimensional convolutional neural network. Specifically, step S103 includes the following: A high-quality UWB echo dataset is constructed into a three-dimensional data cube according to the base station ID, time window, and signal characteristics. Then, the three-dimensional data cube is input into a three-dimensional convolutional neural network. This three-dimensional convolutional neural network extracts the spatiotemporal features of the unmanned transport vehicle's motion in different radar base stations in space and in continuous time dimensions through spatiotemporal convolution operations, in order to capture its motion patterns.

[0040] Understandably, the 3D convolutional neural network aggregates signal strength differences from different UWB radar base stations in the spatial dimension and analyzes the evolution trend of signals over time in the temporal dimension. Through this spatiotemporal joint convolution, the spatiotemporal features extracted by the 3D convolutional neural network include the spatial complementarity features of multiple UWB radar base station signals. Simultaneously, the 3D convolutional neural network captures motion patterns by analyzing the gradient changes of spatiotemporal features along the time axis. These motion patterns include stationary motion, uniform speed, rapid acceleration / deceleration, and trajectory geometric behavior, which includes straight-line driving, left turns, right turns, and stationary rotation. Finally, the spatiotemporal features of the unmanned transport vehicle are output, providing input for subsequent predictions of 3D position and 3D velocity.

[0041] In step S104, the predicted 3D position and predicted 3D velocity of the unmanned transport vehicle can be accurately predicted based on spatiotemporal characteristics through a bidirectional gated recurrent unit network. Specifically, step S104 includes the following: Spatiotemporal features are input into a bidirectional gated recurrent unit network. The motion data sequence in the spatiotemporal features of the unmanned transport vehicle is processed forward and backward through the update gate and reset gate of the bidirectional gated recurrent unit network to capture the time dependence in the motion trajectory of the unmanned transport vehicle. The output layer of the gated recurrent unit network regresses and calculates the predicted 3D position and predicted 3D velocity of the unmanned transport vehicle based on the hidden state of the bidirectional gated recurrent unit network.

[0042] Understandably, this application effectively captures the forward and backward time dependencies in the motion trajectory of the unmanned transport vehicle through the update gate and reset gate mechanism of the bidirectional gated cyclic unit network, thereby achieving accurate prediction of three-dimensional position and three-dimensional velocity.

[0043] In step S105, a convolutional attention mechanism is used to identify static and dynamic obstacles around the unmanned transport vehicle. By inputting spatiotemporal features into the convolutional attention mechanism, an obstacle state probability map is output.

[0044] Furthermore, step S105 includes the following: The convolutional attention mechanism consists of a convolutional sub-network and an activation function. The convolutional sub-network processes the input feature map and finally generates a spatial attention weight map through the logistic function. Then, the weight map is multiplied element-wise with the spatiotemporal features to obtain a weighted feature map. This operation enables the convolutional attention mechanism to automatically assign high weights (values ​​close to 1) to the regions on the feature map corresponding to the obstacle echo features, and suppress irrelevant background noise features (values ​​close to 0) in the weight map. Finally, the obstacle state probability map is output.

[0045] In step S106, a dynamic interaction diagram of the unmanned transport vehicle is obtained based on the predicted three-dimensional position and predicted three-dimensional velocity.

[0046] Furthermore, step S106 includes the following: The unmanned transport vehicle is regarded as a node, and a feature vector is constructed for each node. The feature vector includes state information, including the predicted three-dimensional position, predicted three-dimensional velocity, and heading angle of the unmanned transport vehicle. Edges are dynamically constructed based on communication range and collision risk. The Euclidean distance between any two unmanned transport vehicles is calculated. When the Euclidean distance is less than the preset interaction radius, a connecting edge is established between the two unmanned transport vehicle nodes. An adjacency matrix describing the topology at the current moment is generated. The adjacency matrix is ​​a dynamic interaction graph generated according to the evolution over time. The connecting edge includes a distance vector and a relative velocity vector. The adjacency matrix represents the connection relationship, and the eigenvectors at the nodes in the adjacency matrix represent the state information.

[0047] Understandably, the connecting edge not only indicates that two automated guided vehicles (AGVs) have a communication connection, but also includes the relative distance vector and the relative velocity vector. The warehouse management system generates an adjacency matrix that describes the topology at the current moment. The adjacency matrix represents the connection relationship, and the eigenvectors at the nodes in the adjacency matrix represent the state information. This adjacency matrix is ​​dynamically changing. Every time a time step has passed, the warehouse management system updates the adjacency matrix according to the position changes of the AGVs, thus forming a dynamic interaction relationship graph that evolves over time.

[0048] In step S107, a graph neural network is used to learn the dynamic interaction patterns between multiple unmanned transport vehicles, generating specific speed adjustment commands and steering adjustment commands. These two commands are used for local, real-time formation keeping and dynamic obstacle avoidance.

[0049] Furthermore, step S107 includes the following: The dynamic interaction graph is input into the graph neural network. The graph neural network aggregates the information of neighboring nodes through multi-layer graph convolution, enabling each unmanned transport vehicle node to perceive the status of its surrounding unmanned transport vehicles and obstacles in real time. Graph neural networks learn dynamic interaction patterns between multiple unmanned transport vehicles, including following, avoiding, and maintaining formation. Based on the macroscopic path, formation instructions, and obstacle state probability map, speed adjustment instructions and steering adjustment instructions for the unmanned transport vehicle are generated; among them, speed adjustment instructions and steering adjustment instructions are used for formation maintenance and dynamic obstacle avoidance.

[0050] Understandably, graph neural networks aggregate information from neighboring nodes through multi-layer graph convolutions, enabling each Automated Guided Vehicle (AGV) node to perceive the movement trends of surrounding AGVs and the state of obstacles in real time. This allows for accurate capture of the dynamic interactions between multiple AGVs and the warehouse environment, enhancing perception capabilities. Graph neural networks can autonomously learn dynamic interaction patterns such as following, avoiding, and maintaining formation, eliminating the need for manually pre-setting complex interaction rules. For scenarios involving formation changes during platooning (such as straight-line formations and echelon formations) and differences in AGV acceleration and deceleration, the control strategy can be dynamically adjusted based on the learned interaction patterns. This solves the problem of insufficient flexibility caused by rigid rules in traditional platooning control, ensuring that multiple AGVs maintain their preset formation even under complex operating conditions and reducing the risk of platoon collapse.

[0051] Combining macro-path planning, formation commands, and obstacle state probabilities Figure 3 The system generates control commands based on information, namely speed and steering adjustments. When generating these commands, it satisfies both the need to maintain formation and prioritizes obstacle avoidance safety. Compared to the traditional single-choice decision logic of "obstacle avoidance first" or "formation first," it achieves coordinated optimization of obstacle avoidance and formation, avoiding formation chaos caused by obstacle avoidance or collision risks caused by maintaining formation. Here, the macroscopic path refers to the globally optimal path.

[0052] In step S108, the intelligent scheduling algorithm can be a genetic algorithm. This algorithm can plan a macroscopically optimal global path for the formation of unmanned transport vehicles receiving tasks.

[0053] Step S108 includes the following: Based on the batch task instructions of the warehouse management system, the current position, predicted 3D position, speed adjustment instructions, steering adjustment instructions, individual battery power, as well as the distance factor of the Euclidean distance from the current position of the unmanned transport vehicle to the task target point, the remaining battery power factor, and the probability factor of the unmanned transport vehicle being assigned a simple task, a multi-objective cost function is constructed. A genetic algorithm is used to solve the multi-objective cost function and find the optimal task allocation scheme that minimizes the total global cost. Based on the optimal task allocation scheme, a macro-level global optimal path is planned for the formation of unmanned transport vehicles, setting critical path nodes and desired formations to output a global path plan containing formation instructions.

[0054] Understandably, when constructing the multi-objective function, it's necessary to consider factors such as the distance factor (Euclidean distance from the current position of the Automated Guided Vehicle (AGV) to the task objective point), the remaining battery power factor, and the probability factor (the AGV being assigned a simple task). A genetic algorithm is used to solve the multi-objective cost function, seeking the optimal task allocation scheme that minimizes the global total cost. This optimal task allocation scheme is also the task-vehicle matching scheme that minimizes the global total cost. Then, macro-optimal path planning is performed. For the AGVs receiving tasks, a heuristic search algorithm is used on the global topology map of the warehouse, using path length and the current path's congestion index as weights, to plan a collision-free path consisting of a series of critical path stages. Here, segmented expected formations are set. Based on the geometric constraints of the warehouse environment in the global map, including aisle width and turning radius, formation instructions are automatically added to different road segments based on a rule base. Finally, the optimal task allocation scheme and the global path plan containing formation instructions are output. This global path plan containing formation instructions can serve as the macro-optimal path.

[0055] It should be noted that the expression for the multi-objective cost function is as follows: Multi-objective cost function = Standardized distance factor + (1 - Standardized Remaining Energy Factor) + (1 - standardized probability factor).

[0056] in, These are the weighting coefficients of the standardized distance factor. These are the weighting coefficients for the standardized remaining energy factor. These are all weighting coefficients of the standardized probability factors.

[0057] The optimal task allocation scheme refers to a table that establishes a one-to-one correspondence between all tasks to be assigned and all available unmanned transport vehicles at the current moment. This scheme clearly specifies which particular unmanned transport vehicle should be responsible for performing each specific transport task, aiming to achieve the highest overall operational efficiency and lowest energy consumption cost for the entire fleet.

[0058] The specific genetic algorithm solution process is as follows: First, encoding and population initialization are performed. Each possible task allocation scheme is encoded as a chromosome sequence. Each gene position in the sequence represents a task, and the value on the gene position represents the number of the unmanned transport vehicle assigned to perform the task. Several such sequences are randomly generated to form the initial population. Next, a fitness assessment is performed, in which each individual in the population is fed into a pre-constructed multi-objective cost function for calculation, and the total global cost value of each allocation scheme is obtained. The scheme with the lower total cost has the higher fitness. The selection process is repeated, and high-quality individuals are selected from the current population based on their fitness to serve as parents, thus preserving a low-cost allocation pattern through natural selection. Then, crossover and mutation operations are performed. New individuals are generated by exchanging segments of the gene sequence of the parent individuals, and the vehicle number on individual gene loci is randomly changed with a certain probability, thereby generating a diverse offspring population and avoiding getting trapped in local optima. Finally, a termination condition is determined. If the algorithm reaches the preset number of iterations or the cost of the obtained solution has converged and stabilized, the iteration stops and the decoding sequence corresponding to the individual with the highest fitness is output as the optimal task allocation scheme. Otherwise, the evaluation, selection, crossover and mutation steps are repeated.

[0059] In one embodiment, the method further includes: The residual sequence between the predicted 3D position and the current position of the unmanned transport vehicle is reconstructed based on the variational autoencoder to detect whether any anomalies occur. If an anomaly occurs, a self-calibration trigger command and a high-quality UWB echo dataset will be output.

[0060] The process involves calculating the residual sequence between the predicted 3D position and the current position of the automated guided vehicle (AGV). This residual sequence is then input into a variational autoencoder (VAE), which learns the probability distribution of the residuals under normal operating conditions. The VAE reconstructs the input residual sequence and calculates the reconstruction error. When the reconstruction error exceeds a preset dynamic threshold, an anomaly is detected. At this point, an automatic correction trigger command and a high-quality UWB echo dataset are output. Anomalies include sensor drift and positioning failure. It should be noted that the VAE is a pre-trained variational autoencoder.

[0061] Based on automatic correction trigger commands and high-quality UWB echo datasets, the bidirectional gated recurrent unit network and graph neural network are fine-tuned and incrementally trained to make them more adaptable to the characteristics of the current warehousing environment.

[0062] In one embodiment, after step S108, the method further includes: The warehouse management system updates real-time tasks and closed-loop control commands based on the optimal task allocation scheme and global path planning. The hyperparameters in the graph neural network are then optimized to obtain the optimized hyperparameters.

[0063] Understandably, based on the optimal task allocation scheme and global path planning, the automated guided vehicle (AGV) executes its tasks. The warehouse management system integrates the AGV's predicted 3D position, predicted 3D speed, and AGV formation execution status (speed adjustment commands and steering adjustment commands) into a high-level operational status. These operational statuses, such as task progress, path execution status, AGV trajectory, speed, and obstacle avoidance records, are fed back to the warehouse management system in real time as feedback information. The warehouse management system updates its real-time tasks and closed-loop control commands based on this feedback information. The AGV trajectory, speed, and obstacle avoidance records in the feedback information serve as historical task execution data.

[0064] First, calculate the key performance indicators. The warehouse management system defines a composite performance function, which is determined by the formation holding error e. form Obstacle avoidance response smoothness s smooth and average throughput r task Weighted components: Formation holding error, used to calculate the average Euclidean distance deviation between the actual formation and the desired formation node positions; Obstacle avoidance response smoothness, used to calculate the L2 norm of acceleration changes during obstacle avoidance, with a smaller value indicating smoother action; Average task throughput, used to represent the number of transport tasks completed per unit time.

[0065] The formula for calculating the composite performance function is as follows: J score =w4×e form +w5×s smooth -w6×r task .

[0066] Where w4 is the weighting coefficient for formation holding error, w5 is the weighting coefficient for obstacle avoidance response smoothness, and w6 is the weighting coefficient for average task throughput, the objective of calculating the key performance indicators is to minimize the composite performance function. Next, online optimization and adjustment are performed. The warehouse management system employs a Bayesian optimization algorithm. The calculated key performance indicators are used as the objective function, and the hyperparameters in the graph neural network are used as optimization variables. Specifically, the hyperparameters include the aggregation weighting coefficients of the graph convolutional layers and the penalty factor of the obstacle avoidance potential field. The aggregation weighting coefficients determine the extent to which the automated guided vehicle (AGV) is influenced by neighboring nodes, while the penalty factor of the obstacle avoidance potential field determines the sensitivity to obstacles. The Bayesian optimizer uses a Gaussian process to model the hyperparameter-key performance indicator relationship and strikes a balance between exploration and utilization to find the optimal hyperparameters that minimize the composite performance function, i.e., the optimized hyperparameters. These parameters are updated in real-time to the graph neural network, thereby enabling the warehouse management system to exhibit better formation stability and operational efficiency in subsequent operations.

[0067] The following is a further explanation of this application.

[0068] The method described in this application is compared with pure UWB solution, extended Kalman filter (EKF) based on UWB + inertial measurement unit (IMU) fusion.

[0069] like Figure 2 The figure illustrates the superiority of the bidirectional gated recurrent unit network (GURN)-based method compared to traditional methods. The figure shows the real-time positioning error curves of three different positioning methods over time in the same complex warehouse environment. While pure UWB solution provides basic location information, its signal is severely interfered with by multipath effects and non-line-of-sight propagation, resulting in drastic fluctuations in the error curve and a large peak error. EKF, by fusing UWB and IMU, smooths out noise to some extent, but due to the highly nonlinear and sudden motion of Automated Guided Vehicles (AGVs) in the warehouse environment, EKF struggles to accurately track these dynamic changes, leading to significant error jumps. The method in this application extracts spatiotemporal features from high-quality UWB echo datasets using a three-dimensional convolutional neural network and leverages a bidirectional gated recurrent unit network to deeply mine the temporal dependence of the motion trajectory, enabling accurate prediction of the AGV's motion state. Figure 2 As shown, the positioning error curve of the method in this application is the smoothest and remains at an extremely low level, which proves the high accuracy and high robustness of the method in complex dynamic environments.

[0070] like Figure 3 As shown, average formation deviation and average obstacle avoidance response time are used as evaluation metrics to compare the performance of different formation control algorithms in standard test scenarios. While the traditional leader-follower model has a simple structure, its long reaction chain amplifies errors at each stage, resulting in the highest average deviation. Furthermore, follower vehicles experience severe lag in response when the leader vehicle attempts to avoid an obstacle. Centralized scheduling algorithms, while theoretically capable of calculating the globally optimal path, require high-bandwidth communication, and all micro-decision-making relies on a central server, leading to high computational latency, the longest average obstacle avoidance response time, and a single point of failure risk. The proposed method constructs the unmanned transport vehicle (UGV) cluster as a dynamic graph, using a graph neural network for distributed and reactive coordination. Each UGV can perceive the status of surrounding UGVs and obstacles in real time and autonomously fine-tune its actions. Experimental results show that the proposed method has the lowest average formation deviation and the shortest average obstacle avoidance response time, demonstrating the significant advantages of graph neural networks in achieving efficient, flexible, and robust formation control.

[0071] like Figure 4As shown, the anomaly detection method based on variational autoencoder in this application and the traditional statistical thresholding method are compared in their effectiveness in distinguishing between normal and anomaly localization residuals. The horizontal axis represents the reconstruction error, and the vertical axis represents the frequency of data points. In actual operation, the distributions of normal and anomaly residuals overlap to some extent. The threshold set by the traditional statistical threshold is too rigid, either too sensitive leading to a large number of false alarms or too insensitive leading to missed alarms. The method in this application learns the probability distribution under normal residuals, and its set dynamic threshold (i.e., Figure 4 The anomaly detection threshold in the model can more accurately fit the boundary of the normal distribution. For example... Figure 4 As shown, the method of this application can maximize the differentiation of two overlapping distributions, and while maintaining an extremely low false alarm rate, it significantly improves the recall rate of real location anomalies, ensuring the long-term stability of the system operation.

[0072] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

Claims

1. A method for positioning and platooning control of unmanned warehouse transport vehicles, characterized in that, The method includes: Multiple UWB radar base stations are deployed in the warehouse environment, and a UWB responder is installed on each unmanned transport vehicle. The UWB responder initiates pulse signal exchange, performs time difference ranging with multiple radar base stations to obtain distance information, collects the echo signals of each UWB radar base station, records multiple sets of distance data corresponding to the unmanned transport vehicle at each moment, and outputs the original UWB echo signal timing data. Data processing is performed on the original UWB echo signal timing data to obtain a high-quality UWB echo dataset; Based on a three-dimensional convolutional neural network, the spatiotemporal features of the unmanned transport vehicle are extracted from the high-quality UWB echo dataset. The spatiotemporal features are input into a bidirectional gated recurrent unit network, which outputs the predicted three-dimensional position and predicted three-dimensional velocity of the unmanned transport vehicle. The spatiotemporal features are input into a convolutional attention mechanism to output an obstacle state probability map. Based on the predicted three-dimensional position and the predicted three-dimensional velocity, a dynamic interaction diagram of the unmanned transport vehicle is obtained; The dynamic interaction relationship diagram and the obstacle state probability diagram are used to output the speed adjustment command and steering adjustment command of the unmanned transport vehicle. Based on the intelligent scheduling algorithm, according to the batch task instructions of the warehouse management system, the current position of all unmanned transport vehicles, the predicted three-dimensional position, the speed adjustment instructions, the steering adjustment instructions, and the battery level of each vehicle, the optimal task allocation scheme and the global path planning including the formation instructions are output.

2. The method for positioning and platooning control of unmanned warehouse transport vehicles according to claim 1, characterized in that, The step of extracting the spatiotemporal features of the unmanned transport vehicle from the high-quality UWB echo dataset based on a three-dimensional convolutional neural network includes: The high-quality UWB echo dataset is constructed into a three-dimensional data cube according to the base station ID, time window and signal characteristics, and then the three-dimensional data cube is input into the three-dimensional convolutional neural network. The aforementioned three-dimensional convolutional neural network extracts the spatiotemporal features of the unmanned transport vehicle's movement across different radar base stations in space and continuous time dimensions to capture its movement patterns through spatiotemporal convolution operations.

3. The method for positioning and platooning control of unmanned warehouse transport vehicles according to claim 1, characterized in that, The step of inputting the spatiotemporal features into a bidirectional gated recurrent unit network and outputting the predicted three-dimensional position and predicted three-dimensional velocity of the unmanned transport vehicle includes: The spatiotemporal features are input into the bidirectional gated loop unit network. Through the update gate and reset gate of the bidirectional gated loop unit network, the motion data sequence in the spatiotemporal features of the unmanned transport vehicle is processed forward and backward to capture the time dependence in the motion trajectory of the unmanned transport vehicle. The output layer of the gated loop unit network regresses and calculates the predicted 3D position and predicted 3D velocity of the unmanned transport vehicle based on the hidden state of the bidirectional gated loop unit network.

4. The method for positioning and platooning control of unmanned warehouse transport vehicles according to claim 1, characterized in that, The step of inputting the spatiotemporal features into a convolutional attention mechanism and outputting an obstacle state probability map includes: The convolutional attention mechanism includes a convolutional sub-network and an activation function. The convolutional sub-network processes the input feature map and finally generates a spatial attention weight map through the logistic function. Then, the weight map is multiplied element-wise with the spatiotemporal features to obtain a weighted feature map. This allows the convolutional attention mechanism to automatically assign high weights to regions on the feature map that correspond to the obstacle echo features and suppress irrelevant background noise features, ultimately outputting an obstacle state probability map.

5. The method for positioning and platooning control of unmanned warehouse transport vehicles according to claim 1, characterized in that, The step of obtaining the dynamic interaction diagram of the unmanned transport vehicle based on the predicted three-dimensional position and the predicted three-dimensional velocity includes: The unmanned transport vehicle is regarded as a node, and a feature vector is constructed for each node. The feature vector includes state information, which includes the predicted three-dimensional position, predicted three-dimensional velocity, and heading angle of the unmanned transport vehicle. Edges are dynamically constructed based on communication range and collision risk. The Euclidean distance between any two unmanned transport vehicles is calculated. When the Euclidean distance is less than a preset interaction radius, a connecting edge is established between the two unmanned transport vehicles. An adjacency matrix describing the topology at the current moment is generated. The adjacency matrix is ​​a dynamic interaction relationship graph generated according to time evolution. The connecting edge includes a distance vector and a relative velocity vector. The adjacency matrix represents the connection relationship, and the feature vectors at the nodes in the adjacency matrix represent state information.

6. The method for positioning and platooning control of unmanned warehouse transport vehicles according to claim 1, characterized in that, The step of outputting the speed adjustment command and steering adjustment command of the unmanned transport vehicle from the dynamic interaction relationship diagram and the obstacle state probability diagram includes: The dynamic interaction graph is input into the graph neural network, which aggregates the information of neighboring nodes through multi-layer graph convolution, enabling each unmanned transport vehicle node to perceive the status of its surrounding unmanned transport vehicles and obstacles in real time. The graph neural network learns dynamic interaction patterns between multiple unmanned transport vehicles, including following, avoiding, and maintaining formation. Based on the macroscopic path, formation instructions, and the obstacle state probability map, speed adjustment instructions and steering adjustment instructions for the unmanned transport vehicle are generated; wherein, the speed adjustment instructions and the steering adjustment instructions are used for maintaining formation and dynamic obstacle avoidance.

7. The method for positioning and platooning control of unmanned warehouse transport vehicles according to claim 1, characterized in that, The step of outputting the optimal task allocation scheme and global path planning including formation instructions based on the intelligent scheduling algorithm, according to the batch task instructions of the warehouse management system, the current position of all unmanned transport vehicles, the predicted three-dimensional position, the speed adjustment instructions, the steering adjustment instructions, and their respective battery levels, includes: Based on the batch task instructions of the warehouse management system, the current position of all unmanned transport vehicles, the predicted three-dimensional position, the speed adjustment instructions, the steering adjustment instructions, the battery power of each vehicle, and the distance factor of the Euclidean distance from the current position of the unmanned transport vehicle to the task target point, the remaining battery power factor, and the probability factor of the unmanned transport vehicle being assigned a simple task, a multi-objective cost function is constructed. A genetic algorithm is used to solve the multi-objective cost function to find the optimal task allocation scheme that minimizes the total global cost. Based on the optimal task allocation scheme, a macro-level global optimal path is planned for the formation of unmanned transport vehicles, key path nodes and desired formations are set, and a global path plan containing formation instructions is output.

8. The method for positioning and platooning control of unmanned warehouse transport vehicles according to claim 1, characterized in that, The method further includes: The residual sequence between the predicted 3D position and the current position of the unmanned transport vehicle is reconstructed based on the variational autoencoder in order to detect whether any anomalies occur. If an anomaly occurs, a self-calibration trigger command and a high-quality UWB echo dataset will be output.

9. The method for positioning and platooning control of unmanned warehouse transport vehicles according to claim 1, characterized in that, Following the step of outputting the optimal task allocation scheme and global path planning including formation instructions based on the intelligent scheduling algorithm, according to the batch task instructions of the warehouse management system, the current position of all unmanned transport vehicles, the predicted three-dimensional position, the speed adjustment instructions, the steering adjustment instructions, and their respective battery levels, the method further includes: The warehouse management system updates its real-time tasks and closed-loop control instructions based on the optimal task allocation scheme and the global path planning. The hyperparameters in the graph neural network are then optimized to obtain the optimized hyperparameters.

Images