Intelligent logistics management system and method based on artificial intelligence and big data

Abstract

The invention discloses an intelligent logistics management system and method based on artificial intelligence and big data, and relates to the technical field of logistics management, and the method comprises the steps: collecting the motion data, shelf electromagnetic field data and environment data of an AMR cluster in a narrow roadway, and constructing an N * M-dimensional feature vector matrix through multi-dimensional feature extraction; the feature vector matrix is used as input, a three-dimensional potential energy field model is constructed, a multi-AMR path decision is converted into a non-cooperative game model, an improved particle swarm optimization algorithm is adopted for solving, and a smooth motion trail of AMR is generated; constructing a two-level federated learning architecture based on the edge computing unit and the federated server, and optimizing a three-dimensional potential energy field model through local training and global parameter aggregation; a distributed magneto-rheological damper is integrated on an AMR chassis, when it is detected that dispatching fails and extreme congestion occurs, deceleration or steering adjustment is achieved through differential control instructions, and efficient collaborative operation and emergency escape of an AMR cluster in the complex electromagnetic environment are achieved.

Description

Technical Field

[0001] This invention relates to the field of logistics management technology, specifically to a smart logistics management system and method based on artificial intelligence and big data. Background Technology

[0002] With the rapid development of the logistics industry, automated mobile robots (AMRs) have been widely used in warehousing and logistics management. In existing technologies, path planning for AMR clusters mainly relies on centralized scheduling algorithms or rule-based local obstacle avoidance strategies. For example, some systems use LiDAR and visual sensors to collect environmental data and generate static or dynamic paths for AMRs through traditional optimization algorithms such as the A* algorithm or dynamic window method. Some solutions utilize wireless communication networks to achieve information sharing among AMRs to coordinate the movement of multiple machines. In terms of data processing, existing technologies usually use cloud computing or edge computing to centrally process sensor data and make path decisions based on preset rules or simple machine learning models.

[0003] However, existing technologies have significant shortcomings in multi-AMR collaborative operation scenarios. First, traditional path planning algorithms are difficult to effectively solve the dynamic obstacle avoidance problem of high-density AMR clusters in narrow alleys, especially in complex electromagnetic environments or sudden congestion, which can easily lead to path conflicts or deadlocks. Second, centralized data processing architectures suffer from high communication latency and unbalanced computing load, which cannot meet real-time requirements and have weak data privacy protection capabilities.

[0004] Furthermore, existing technologies lack the ability to model the coupling effect between the motion state of AMRs and the environmental electromagnetic field, resulting in insufficient path planning accuracy. At the same time, the dynamic adjustment mechanism in emergency situations is relatively simple, usually relying only on deceleration or stopping strategies, which cannot achieve rapid escape from trouble, thereby reducing overall operational efficiency. These shortcomings limit the large-scale application of AMR clusters in smart logistics systems. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a smart logistics management system and method based on artificial intelligence and big data. Through multi-source data fusion, three-dimensional potential energy field modeling, non-cooperative game-theoretic path optimization, and a two-level federated learning architecture, it solves the problems of dynamic obstacle avoidance, path conflict, and deadlock in high-density AMR clusters in narrow alleyways. At the same time, it improves the system's real-time performance, computational load balancing, and data privacy protection capabilities, and realizes efficient collaborative operation and emergency escape of AMR clusters in complex electromagnetic environments.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a smart logistics management method based on artificial intelligence and big data, comprising: Motion data of AMR clusters, electromagnetic field data of shelves, and environmental data are collected in narrow aisles. Through multi-dimensional feature extraction, an N×M dimensional feature vector matrix is ​​constructed. Using the feature vector matrix as input, a three-dimensional potential energy field model is constructed, transforming the multi-AMR path decision into a non-cooperative game model. An improved particle swarm optimization algorithm is used to solve the problem and generate the smooth motion trajectory of the AMR. A two-level federated learning architecture is built based on edge computing units and federated servers. The three-dimensional potential energy field model is optimized through local training and global parameter aggregation, and the potential energy gradient is monitored in real time. The AMR chassis integrates a distributed magnetorheological damper. When a scheduling failure and extreme congestion are detected, an emergency escape mode is activated, which uses differentiated control commands to achieve deceleration or steering adjustment.

[0007] Furthermore, motion data, shelf electromagnetic field data, and environmental data of the AMR cluster within the narrow aisle are collected. This includes: deploying a combined LiDAR and millimeter-wave radar monitoring unit in the AMR cluster's operating area within the narrow aisle; LiDAR collecting static structural data of the aisle, and millimeter-wave radar capturing the real-time motion status of the AMRs; mounting three-axis accelerometers and incremental speed encoders on the four sides (front, rear, left, and right) of each AMR chassis to collect the AMR's three-dimensional spatial motion velocity vector and absolute position coordinates; integrating Hall effect sensors at the four corners of the bottom of all operating shelves to collect shelf electromagnetic field intensity distribution data; and receiving raw data transmitted through a distributed sensor network via an industrial-grade edge computing unit.

[0008] Furthermore, after data acquisition, a bidirectional alignment algorithm based on GPS timestamps is used to perform spatiotemporal synchronization calibration on the raw data collected by different sensors, using the local high-precision clock of the edge computing unit as a reference, and matching the AMR motion data with the electromagnetic field data of the corresponding location one by one; the db4 wavelet transform algorithm is used to perform noise reduction processing on the synchronized data.

[0009] Furthermore, multi-dimensional feature extraction includes velocity field gradient, electromagnetic field gradient, and the relative distance between each AMR and surrounding shelves and other AMRs. The velocity field gradient, electromagnetic field gradient, and relative distance features are arranged in time series to construct an N×M dimensional feature vector matrix, where N is the number of data sampling points and M is the feature dimension.

[0010] Furthermore, the three-dimensional potential energy field model adopts a three-layer perceptron architecture. The input layer dimension is consistent with the feature dimension M of the feature vector matrix. The hidden layer has 64 neurons, and the output layer outputs the superimposed field strength value. The narrow alleyway working space is discretized into three-dimensional grid units with a size of 10cm×10cm×5cm. Each grid unit stores the field state parameters. The Lorentz force equation is used as the motion constraint condition.

[0011] Furthermore, in the non-cooperative game model, each AMR is a player in the game, and the game objective is to optimize the overall operating efficiency of the AMR cluster; the payoff function is the weighted sum of the expected path time and the expected energy consumption; the Nash equilibrium constraint is that the superposition value of the field strength of the corresponding grid cells of any two adjacent AMRs in the three-dimensional potential energy field does not exceed the preset safety threshold.

[0012] Furthermore, the particle swarm size of the improved particle swarm optimization algorithm is set to 3 times the number of AMRs in the current job, and the inertia weight adopts a linear decreasing strategy; the inverse of the AMR reward function is used as the fitness function, and the position and velocity of the path particles are updated through the individual optimal solution and the global optimal solution. The Nash equilibrium point is used as the convergence criterion, and after convergence, a smooth motion trajectory is generated through the trajectory interpolation algorithm.

[0013] Furthermore, in the two-level federated learning architecture, the edge computing unit is a local node, bound to the AMR cluster and sensing devices within the coverage radius. The local node uses a stochastic gradient descent algorithm with a driving term for local training, and a cosine annealing strategy is used to adjust the learning rate. The samples are divided into training and validation sets in an 8:2 ratio, and the convergence of the validation set loss function is used as the training termination condition. The global parameter aggregation uses a weighted federated average algorithm. The federated server monitors the three-dimensional potential gradient of each local node in real time. When the potential gradient exceeds the potential anomaly threshold, the global model retraining process is triggered. Each local node retrains based on real-time data and uploads the parameters. The federated server aggregates the parameters and synchronously distributes the updated parameters.

[0014] Furthermore, the distributed magnetorheological dampers are integrated in a 2×2 matrix at the four corners of the AMR chassis. The front wheel side dampers focus on steering assistance for getting out of trouble, while the rear wheel side dampers focus on deceleration and braking for getting out of trouble. The dual triggering conditions for the emergency escape mode are: when the global model parameter update is delayed, it is determined to be a scheduling failure; when the three-dimensional potential energy gradient exceeds the preset abnormal threshold and lasts for several seconds, it is determined to be extreme congestion. After triggering, differentiated control commands are issued. During emergency deceleration, the current of the front and rear wheel dampers is adjusted to the target value for deceleration; during steering escape, the damper on one side is adjusted to the target current for steering adjustment.

[0015] A smart logistics management system based on artificial intelligence and big data includes: The multi-source data acquisition module collects motion data of the AMR cluster, electromagnetic field data of the shelves, and environmental data in the narrow aisle. Through multi-dimensional feature extraction, it constructs an N×M dimensional feature vector matrix. The path decision module uses the feature vector matrix as input to construct a three-dimensional potential energy field model, transforming the multi-AMR path decision into a non-cooperative game model. An improved particle swarm optimization algorithm is used to solve the problem and generate the smooth motion trajectory of the AMR. The federated learning module is based on edge computing units and federated servers to build a two-level federated learning architecture. It optimizes the three-dimensional potential energy field model through local training and global parameter aggregation, and monitors the potential energy gradient in real time. The scheduling execution module integrates a distributed magnetorheological damper in the AMR chassis. When scheduling failure is detected and there is extreme congestion, the emergency escape mode is activated, and deceleration or steering adjustment is achieved through differentiated control commands.

[0016] (III) Beneficial Effects This invention provides a smart logistics management system and method based on artificial intelligence and big data, which has the following beneficial effects: (1) Through multi-source data acquisition and processing, high-precision synchronous acquisition and fusion of AMR cluster motion data, shelf electromagnetic field data and environmental data in narrow aisles were achieved. A full-coverage perception network was constructed using lidar, millimeter-wave radar and Hall sensor. Combined with spatiotemporal synchronous calibration and wavelet denoising technology, the accuracy and real-time performance of the data were ensured. The feature vector matrix constructed by multi-dimensional feature extraction provided high-quality input for subsequent three-dimensional potential energy field modeling and path decision-making, significantly improving the system's perception capability and decision-making accuracy in complex environments.

[0017] (2) By constructing a three-dimensional potential energy field model and transforming the multi-AMR path decision into a non-cooperative game model, combined with an improved particle swarm optimization algorithm, efficient collaborative path planning for high-density AMR clusters in narrow alleys was realized. The coupling effect of AMR motion and electromagnetic field was incorporated into the modeling. Nash equilibrium constraints were used to ensure path safety and avoid conflicts and deadlocks. The improved particle swarm algorithm was optimized with dynamic weights and fitness functions to generate smooth motion trajectories, which significantly improved the real-time performance and global optimality of path planning, while reducing energy consumption. It is suitable for high-precision logistics scheduling needs in complex electromagnetic environments.

[0018] (3) By constructing a two-level federated learning architecture, the distributed optimization and real-time update of the three-dimensional potential energy field model are realized. The edge computing unit is used as a local node for local training. It only uploads model parameters instead of raw data, which effectively protects data privacy and security. The federated server dynamically optimizes model performance by weighted aggregation of global parameters, monitors potential energy gradient in real time and triggers retraining mechanism, ensuring that the model responds quickly in the event of sudden congestion or environmental changes, reducing communication latency, balancing computing load, and improving the system's adaptability and decision accuracy in complex scenarios. It provides efficient and secure collaborative scheduling capabilities for the AMR cluster.

[0019] (4) By integrating a distributed magnetorheological damper and setting a dual triggering mechanism, the emergency response capability of the AMR cluster in extreme congestion or scheduling failure is significantly improved. When the global model update delay or potential gradient exceeds the threshold is detected, the emergency escape mode is automatically started. The differential current control is used to achieve rapid deceleration or precise steering. Dynamic adjustment is completed within 50-100ms, effectively avoiding deadlock. Combining physical braking and intelligent decision-making, the operation interruption time is minimized while ensuring safety. The robustness and reliability of the system in complex scenarios are enhanced, providing an efficient emergency solution for high-density narrow aisle logistics scenarios. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the steps of the intelligent logistics management method based on artificial intelligence and big data according to the present invention. Figure 2 This is a schematic diagram of the intelligent logistics management method based on artificial intelligence and big data according to the present invention. Figure 3 This is a schematic diagram of the intelligent logistics management system based on artificial intelligence and big data of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1-2 This invention provides a smart logistics management method based on artificial intelligence and big data, comprising the following steps: Step 1: Collect motion data of the AMR cluster, electromagnetic field data of the shelving, and environmental data in the narrow aisle. Construct an N×M dimensional feature vector matrix through multi-dimensional feature extraction. Step one includes the following: Step 101: In the narrow aisle AMR cluster operation area, deploy combined LiDAR and millimeter-wave radar monitoring units at 3-meter intervals to form a fully covered environmental perception network without blind spots. The LiDAR is responsible for collecting static structural data of the aisle, such as aisle width and shelf spacing, while the millimeter-wave radar is responsible for dynamically capturing the real-time motion status of the AMRs. At the same time, high-precision triaxial accelerometers and incremental speed encoders are installed on the four sides of the chassis of each AMR to collect the AMR's motion velocity vector and absolute position coordinates in three-dimensional space in real time. High-sensitivity Hall sensors are integrated at the four corners of the bottom of all operating shelves to continuously collect the electromagnetic field intensity distribution data generated by the metal structure of the shelves, ensuring that the sensor collection range covers the entire aisle operation area without any data collection blind spots. Step 102: Deploy industrial-grade edge computing units at key nodes such as the entrances and exits of narrow aisles and areas with dense shelving. Each edge computing unit is responsible for receiving and initially processing sensor data within a 10-meter radius. The edge computing unit receives raw data transmitted from the distributed sensor network through a dual-mode communication protocol, including AMR motion data, shelf electromagnetic field data, and environmental perception data. Using a GPS timestamp-based bidirectional alignment algorithm, with the local high-precision clock of the edge computing unit as the reference, the raw data collected by different sensors are spatiotemporally synchronized and calibrated, and the AMR motion data is matched one-to-one with the electromagnetic field data at the corresponding location. Step 103: The original data after spatiotemporal synchronization is denoised using the db4 wavelet transform algorithm, with the wavelet decomposition layer set to 3 layers. During the denoising process, a soft threshold function is used to process high-frequency components. Multi-dimensional features are extracted from the denoised data. The velocity field gradient is calculated using the first-order partial derivative, and the electromagnetic field gradient is calculated using the spatial derivative of the electromagnetic field intensity. The relative distance between each AMR and surrounding shelves and other AMRs is calculated using the Euclidean distance formula. The velocity field gradient, electromagnetic field gradient, and relative distance features are arranged in time series to construct a feature vector matrix of dimension N×M, where N is the number of data sampling points and M is the feature dimension.

[0023] When using this method, refer to the content of steps 101 to 103: Through multi-source data acquisition and processing, high-precision synchronous acquisition and fusion of AMR cluster motion data, shelf electromagnetic field data, and environmental data in narrow aisles were achieved. A full-coverage perception network was constructed using lidar, millimeter-wave radar, and Hall sensors. Combined with spatiotemporal synchronous calibration and wavelet denoising technology, the accuracy and real-time performance of the data were ensured. The feature vector matrix constructed through multi-dimensional feature extraction provided high-quality input for subsequent three-dimensional potential energy field modeling and path decision-making, significantly improving the system's perception capability and decision-making accuracy in complex environments.

[0024] Step 2: Using the feature vector matrix as input, construct a three-dimensional potential energy field model, transform the multi-AMR path decision into a non-cooperative game model, and solve it using an improved particle swarm optimization algorithm to generate the smooth motion trajectory of the AMR. Step two includes the following: Step 201: Using an N×M dimensional feature vector matrix as the core input data, a three-dimensional potential energy field model is constructed. This model adopts a 3-layer perceptron (MLP) architecture. The input layer dimension is consistent with the feature dimension M of the feature vector matrix. The hidden layer has 64 neurons, and the output layer outputs the superimposed field strength value. The core parameters of the model include the weight matrix between the input layer and the hidden layer, the weight matrix between the hidden layer and the output layer, and the bias vectors of the corresponding two layers. A distributed parameterized expression is adopted, and the entire narrow tunnel working space is discretized into three-dimensional grid cells with a specification of 10cm×10cm×5cm. Each grid cell corresponds to storing a complete set of field state parameters, including velocity components, electromagnetic field components, and gradient values. The Lorentz force equation is used as the core motion constraint condition of this model. Among them, the equivalent charge coefficient of AMR needs to be accurately calibrated through historical operating data. 1000 sets of valid sample data under different loads and motion speeds within the past 3 months were selected. The least squares method was used to fit the mapping relationship between the equivalent charge coefficient and the load and speed. During the fitting process, the values ​​of fixed coefficients were determined. During real-time operation, the equivalent charge coefficient was dynamically calculated based on the actual load and real-time speed of the current AMR. The velocity vector was taken from the three-dimensional motion velocity data after spatiotemporal synchronization calibration, and the electromagnetic field intensity vector was taken from the synchronized three-dimensional electromagnetic field data. The force vector of the AMR in three-dimensional space was obtained through vector cross product operation, thereby constraining the motion direction and acceleration upper limit of the AMR. Step 202: Based on the constructed three-dimensional potential energy field model, the path decision-making problem of multiple AMRs in a narrow alleyway is transformed into a non-cooperative game model. This model includes: treating each AMR independently executing its task as a game participant, and taking the optimization of the overall operational efficiency of the AMR cluster as the core objective of the game; defining the payoff function of each AMR as the weighted sum of the expected path time and expected energy consumption, with the weighting coefficient dynamically adjusted according to the task priority. In emergency task scenarios, the weighting coefficient for path time is set to 0.7, and the weighting coefficient for energy consumption is set to 0.3; In the standard task scenario, the weighting coefficients for both are set to 0.5; Nash equilibrium constraint is set: for any two adjacent AMRs, the superposition value of the field strength of the corresponding grid cells in the three-dimensional potential energy field does not exceed the preset safety threshold. The superposition value of the field strength is the absolute value of the sum of the vector cross product of the velocity field gradient and the electromagnetic field gradient at the corresponding position of each of the two adjacent AMRs. It should be noted that this safety threshold is calculated and determined based on the fuselage size of the AMR and the preset safety distance. This constraint ensures that adjacent AMRs will not have path conflicts due to field interference during the movement. Step 203: An improved particle swarm optimization algorithm with dynamic inertia weight is used to solve the constructed non-cooperative game model in real time. The algorithm parameters are set as follows: the particle swarm size is set to 3 times the number of AMRs in the current task, the inertia weight adopts a linear decreasing strategy, with a value of 0.9 in the early stage of iteration and 0.4 in the later stage of iteration, the cognitive coefficient is set to 2.0, the social coefficient is set to 2.0, and the maximum number of iterations is set to 50. Step 204: Initialize the path particles for each AMR. Each path particle contains the node coordinates, real-time velocity, and steering angle on its trajectory. The fitness function is the reciprocal of the reward function for each AMR. During each iteration, the path particles update their position and velocity based on the individual optimal solution and the global optimal solution. Simultaneously, a field strength superposition constraint is introduced to filter particles; that is, if the field strength superposition value does not exceed a preset safety threshold, invalid particles that violate the constraint are directly removed. The Nash equilibrium point is used as the algorithm convergence criterion; that is, when the reward function values ​​of all AMRs no longer change significantly, such as when the difference in change over 5 consecutive iterations is less than 1%, the algorithm converges. When the algorithm converges, the path planning of the AMR cluster reaches Pareto optimality. The optimal path is smoothed by trajectory interpolation algorithm to generate continuous motion trajectory of each AMR, output real-time target speed and steering angle commands, and save the weight matrix and bias vector of the current model.

[0025] When using this method, refer to steps 201 to 203: By constructing a three-dimensional potential energy field model and transforming multi-AMR path decision-making into a non-cooperative game model, combined with an improved particle swarm optimization algorithm, efficient collaborative path planning for high-density AMR clusters in narrow alleyways was achieved. The coupling effect of AMR motion and electromagnetic field was incorporated into the modeling, and path safety was ensured through Nash equilibrium constraints to avoid conflicts and deadlocks. The improved particle swarm optimization algorithm was used to optimize the solution with dynamic weights and fitness functions to generate smooth motion trajectories, which significantly improved the real-time performance and global optimality of path planning, while reducing energy consumption. It is suitable for high-precision logistics scheduling requirements in complex electromagnetic environments.

[0026] Step 3: Construct a two-level federated learning architecture based on edge computing units and federated servers, optimize the three-dimensional potential energy field model through local training and global parameter aggregation, and monitor the potential energy gradient in real time; Step three includes the following: Step 301: Using edge computing units as local nodes for federated learning, each edge computing unit is bound to the AMR cluster and sensing devices within its coverage radius to form an independent local data domain. The core server of the AMR cluster scheduling center is set as the federated server, constructing a two-level federated learning architecture. The local nodes and the federated server establish an encrypted connection through a dual-mode communication protocol. The local nodes only store the field strength parameters, AMR motion data, and scheduling decision results within their own coverage area for the past 72 hours. The raw data is not uploaded to the cloud or transmitted across nodes. The federated server only stores global model parameters and node status information, and does not store any local raw data. Step 302: Each local node trains its local model using a stochastic gradient descent algorithm with momentum parameters, based on the constructed and solved 3D potential energy field model. The momentum parameter is set to 0.9, and the initial learning rate is set to 0.001. A cosine annealing strategy is used to dynamically adjust the learning rate, such as reducing it to 0.95 every 100 iterations, depending on the specific training situation. The training samples are field data and scheduling decision results from the past 5 minutes. The field data includes the velocity field gradient, electromagnetic field gradient, and relative distance feature vector extracted in Step 1. The scheduling decision results include the AMR target velocity, steering angle, and job completion efficiency data output in Step 2. The samples are divided into training and validation sets in an 8:2 ratio. During training, the validation set loss function is used for convergence, i.e., the mean squared error (MSE) is less than the preset value after 20 consecutive iterations. The preset value is set to ensure that the loss function converges, which serves as the local training termination condition. After training is completed, the local node only extracts the weight matrix and bias vector of the three-dimensional potential energy field model and uploads them to the federated server. The original field data and training samples are stored locally. Step 303: After receiving the 3D potential energy field model parameters uploaded by each local node, the federated server uses a weighted federated average algorithm to aggregate global parameters. The weight allocation is determined based on the proportion of sample data of each local node. Node weight = number of samples of that node / total number of samples of all nodes. During the aggregation process, the parameters of each node are standardized, and then the parameters are weighted and summed to generate a globally optimized 3D potential energy field model parameter set. At the same time, the federated server collects the 3D potential energy gradient fed back by each local node in real time. The potential energy gradient is calculated by the first-order partial derivative of the field strength superposition value. The potential energy anomaly is calibrated and set based on historical congestion data. The threshold is set at 1.2 times the average potential gradient before 100 typical congestion events in the past 6 months. When the potential gradient of any local node exceeds the potential anomaly threshold, the global model parameter retraining process is immediately triggered. The federated server issues a retraining instruction to all local nodes. Each node retrains locally based on real-time data and uploads the updated 3D potential field model parameters. The federated server performs aggregation again to generate new global model parameters and synchronously sends the updated parameters to each local node. After receiving the updated parameters, the local nodes replace the weight matrix and bias vector of the original model in real time.

[0027] When using this method, refer to the content of steps 301 to 303: By constructing a two-level federated learning architecture, distributed optimization and real-time updates of the three-dimensional potential energy field model are achieved. Edge computing units, as local nodes, perform local training and only upload model parameters instead of raw data, effectively ensuring data privacy and security. The federated server dynamically optimizes model performance by weighted aggregation of global parameters, monitors potential energy gradients in real time, and triggers retraining mechanisms to ensure that the model responds quickly in the event of sudden congestion or environmental changes. This reduces communication latency, balances the computational load, and improves the system's adaptability and decision-making accuracy in complex scenarios, providing efficient and secure collaborative scheduling capabilities for the AMR cluster.

[0028] Step 4: Integrate a distributed magnetorheological damper into the AMR chassis. When a scheduling failure and extreme congestion are detected, activate the emergency escape mode and use differentiated control commands to achieve deceleration or steering adjustment.

[0029] Step four includes the following: Step 401: Four sets of magnetorheological dampers are symmetrically integrated at the four corners of the chassis of each AMR to form a 2×2 distributed damper matrix. The front wheel side dampers focus on steering assistance to get out of trouble, while the rear wheel side dampers focus on deceleration and braking to get out of trouble. The global model parameters and three-dimensional potential energy gradient monitoring data are synchronized in real time, and dual triggering conditions are set: the first is parameter update delay triggering. When the global model parameter update cycle exceeds 500ms, it is determined to be a scheduling failure. The second is potential energy gradient threshold triggering. When the three-dimensional potential energy monitoring module detects that the three-dimensional potential energy gradient exceeds the potential energy anomaly threshold set in step 401, and the duration of this state reaches several seconds or more, such as 200ms, the specific setting depends on the situation. This is determined to be an extreme congestion deadlock scenario. Step 402: When both conditions are met simultaneously, the emergency escape mode is immediately activated. Differentiated control commands are issued to the four sets of damper actuators. The commands include the target current value, adjustment duration, and priority identifier. The priority identifier is divided into Level 1 emergency braking and Level 2 steering escape. The target current value and target damper are dynamically allocated according to the current movement direction of the AMR and the congestion location. Step 403: When emergency deceleration is required, the current of both the front and rear wheel dampers is adjusted to the target current value, reducing the speed of the AMR from the maximum of 1.5m / s to below 0.3m / s within 50ms; when steering to get out of trouble is required, the current of one side damper is adjusted to the target current value, while the other side maintains the base current. The steering torque is generated by the difference in damping force between the two sides, so that the AMR can complete a steering adjustment of 5°~15° within 100ms, avoiding the deadlock area formed by adjacent AMRs.

[0030] When using this method, please refer to the content of steps 401 to 403: By integrating distributed magnetorheological dampers and setting a dual triggering mechanism, the emergency response capability of the AMR cluster in the event of extreme congestion or scheduling failure is significantly improved. When a global model update delay or potential gradient exceeding the threshold is detected, an emergency escape mode is automatically activated. Differentiated current control is used to achieve rapid deceleration or precise steering, and dynamic adjustment is completed within 50-100ms, effectively avoiding deadlock. Combining physical braking and intelligent decision-making, the system minimizes operation interruption time while ensuring safety, enhancing the robustness and reliability of the system in complex scenarios and providing an efficient emergency solution for high-density narrow-lane logistics scenarios.

[0031] Please see Figure 3 This invention also provides a smart logistics management system based on artificial intelligence and big data, including: a multi-source data acquisition module, a route decision module, a federated learning module, and a scheduling execution module, wherein: The multi-source data acquisition module collects motion data of the AMR cluster, electromagnetic field data of the shelves, and environmental data in the narrow aisle. Through multi-dimensional feature extraction, it constructs an N×M dimensional feature vector matrix. The path decision module uses the feature vector matrix as input to construct a three-dimensional potential energy field model, transforming the multi-AMR path decision into a non-cooperative game model. An improved particle swarm optimization algorithm is used to solve the problem and generate the smooth motion trajectory of the AMR. The federated learning module is based on edge computing units and federated servers to build a two-level federated learning architecture. It optimizes the three-dimensional potential energy field model through local training and global parameter aggregation, and monitors the potential energy gradient in real time. The scheduling execution module integrates a distributed magnetorheological damper in the AMR chassis. When scheduling failure is detected and there is extreme congestion, the emergency escape mode is activated, and deceleration or steering adjustment is achieved through differentiated control commands.

[0032] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The coefficients in the formulas are set by those skilled in the art according to the actual situation.

[0033] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, and combinations thereof. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0034] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0035] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A smart logistics management method based on artificial intelligence and big data, characterized by: include: Motion data of AMR clusters, electromagnetic field data of shelves, and environmental data are collected in narrow aisles. Through multi-dimensional feature extraction, an N×M dimensional feature vector matrix is ​​constructed. Using the feature vector matrix as input, a three-dimensional potential energy field model is constructed, transforming the multi-AMR path decision into a non-cooperative game model. An improved particle swarm optimization algorithm is used to solve the problem and generate the smooth motion trajectory of the AMR. A two-level federated learning architecture is built based on edge computing units and federated servers. The three-dimensional potential energy field model is optimized through local training and global parameter aggregation, and the potential energy gradient is monitored in real time. The AMR chassis integrates a distributed magnetorheological damper. When a scheduling failure and extreme congestion are detected, an emergency escape mode is activated, which uses differentiated control commands to achieve deceleration or steering adjustment.

2. The intelligent logistics management method based on artificial intelligence and big data according to claim 1, characterized in that: The system collects motion data, shelf electromagnetic field data, and environmental data from AMR clusters within narrow aisles. This includes: deploying a combined LiDAR and millimeter-wave radar monitoring unit in the AMR cluster's operating area within the narrow aisle; using LiDAR to collect static structural data of the aisle, and millimeter-wave radar to capture the real-time motion status of the AMRs; mounting triaxial accelerometers and incremental speed encoders on the four sides (front, rear, left, and right) of each AMR chassis to collect the AMR's three-dimensional spatial motion velocity vector and absolute position coordinates; integrating Hall effect sensors at the four corners of the bottom of all operating shelves to collect shelf electromagnetic field intensity distribution data; and receiving raw data transmitted through a distributed sensor network via an industrial-grade edge computing unit.

3. The intelligent logistics management method based on artificial intelligence and big data according to claim 2, characterized in that: After data acquisition, a two-way alignment algorithm based on GPS timestamps is used to perform spatiotemporal synchronization calibration on the raw data collected by different sensors, with the local high-precision clock of the edge computing unit as the reference, and the AMR motion data is matched one by one with the electromagnetic field data of the corresponding location; the db4 wavelet transform algorithm is used to perform noise reduction processing on the synchronized data.

4. The intelligent logistics management method based on artificial intelligence and big data according to claim 3, characterized in that: Multi-dimensional feature extraction includes velocity field gradient, electromagnetic field gradient, and the relative distance between each AMR and surrounding shelves and other AMRs. The velocity field gradient, electromagnetic field gradient, and relative distance features are arranged in time series to construct an N×M dimensional feature vector matrix, where N is the number of data sampling points and M is the feature dimension.

5. The intelligent logistics management method based on artificial intelligence and big data according to claim 1, characterized in that: The three-dimensional potential energy field model adopts a three-layer perceptron architecture. The input layer dimension is consistent with the feature dimension M of the feature vector matrix. The hidden layer has 64 neurons, and the output layer outputs the superimposed field strength value. The narrow alleyway working space is discretized into three-dimensional grid units with a size of 10cm×10cm×5cm. Each grid unit stores the field state parameters. The Lorentz force equation is used as the motion constraint condition.

6. The intelligent logistics management method based on artificial intelligence and big data according to claim 5, characterized in that: In the non-cooperative game model, each AMR is a player in the game, and the game objective is to optimize the overall operating efficiency of the AMR cluster. The payoff function is the weighted sum of the expected path time and the expected energy consumption. The Nash equilibrium constraint is that the superposition value of the field strength of the corresponding grid cells of any two adjacent AMRs in the three-dimensional potential energy field does not exceed a preset safety threshold.

7. The intelligent logistics management method based on artificial intelligence and big data according to claim 6, characterized in that: The particle swarm size of the improved particle swarm optimization algorithm is set to 3 times the number of AMRs in the current job, and the inertia weight adopts a linear decreasing strategy. The inverse of the AMR reward function is used as the fitness function. The position and velocity of the path particles are updated by the individual optimal solution and the global optimal solution. The Nash equilibrium point is used as the convergence criterion. After convergence, a smooth motion trajectory is generated by the trajectory interpolation algorithm.

8. The intelligent logistics management method based on artificial intelligence and big data according to claim 1, characterized in that: In the two-level federated learning architecture, the edge computing unit is a local node, which is bound to the AMR cluster and sensing devices within the coverage radius. The local node uses the stochastic gradient descent algorithm with a driving term for local training, and adopts the cosine annealing strategy to adjust the learning rate. The samples are divided into training set and validation set in an 8:2 ratio, and the convergence of the validation set loss function is used as the training termination condition. The global parameter aggregation adopts the weighted federated average algorithm. The federated server monitors the three-dimensional potential energy gradient of each local node in real time. When the potential energy gradient exceeds the potential energy anomaly threshold, the global model retraining process is triggered. Each local node retrains based on real-time data and uploads the parameters. The federated server aggregates and synchronously distributes the updated parameters.

9. The intelligent logistics management method based on artificial intelligence and big data according to claim 1, characterized in that: Distributed magnetorheological dampers are integrated in a 2×2 matrix at the four corners of the AMR chassis. The front wheel side dampers focus on steering assistance for getting out of trouble, while the rear wheel side dampers focus on deceleration and braking for getting out of trouble. The dual triggering conditions for the emergency escape mode are: when the global model parameter update is delayed, it is determined to be a scheduling failure; when the three-dimensional potential energy gradient exceeds the preset abnormal threshold and lasts for several seconds, it is determined to be extreme congestion. After triggering, differentiated control commands are issued. During emergency deceleration, the current of the front and rear wheel dampers is adjusted to the target value for deceleration; during steering escape, the damper on one side is adjusted to the target current for steering adjustment.

10. A smart logistics management system based on artificial intelligence and big data, used to implement the method described in any one of claims 1 to 9, characterized in that: include: The multi-source data acquisition module collects motion data of the AMR cluster, electromagnetic field data of the shelves, and environmental data in the narrow aisle. Through multi-dimensional feature extraction, it constructs an N×M dimensional feature vector matrix. The path decision module uses the feature vector matrix as input to construct a three-dimensional potential energy field model, transforming the multi-AMR path decision into a non-cooperative game model. An improved particle swarm optimization algorithm is used to solve the problem and generate the smooth motion trajectory of the AMR. The federated learning module is based on edge computing units and federated servers to build a two-level federated learning architecture. It optimizes the three-dimensional potential energy field model through local training and global parameter aggregation, and monitors the potential energy gradient in real time. The scheduling execution module integrates a distributed magnetorheological damper in the AMR chassis. When scheduling failure is detected and there is extreme congestion, the emergency escape mode is activated, and deceleration or steering adjustment is achieved through differentiated control commands.

Images