Intelligent logistics-oriented path perception federated learning collaborative optimization method and system

Abstract

The application provides a path perception federated learning collaborative optimization method and system for intelligent logistics, which comprises the following steps: system initialization and collection of the computing capacity, flight trajectory and data quality parameters of the unmanned aerial vehicle; based on the path perception mechanism, the LSTM model is used to predict the communicable time of the unmanned aerial vehicle in the coverage range of the base station, and the computing capacity, path redundancy and base station load are combined to dynamically screen the unmanned aerial vehicle nodes participating in the federated learning; the selected unmanned aerial vehicle uses the locally collected logistics scene data to train the model, and uploads the gradient to the edge base station; the edge base station adopts a strategy based on data scarcity to perform lower-layer aggregation, and the cloud server adopts a strategy based on Jaccard similarity to perform upper-layer aggregation on the aggregation results of multiple base stations, and updates the global model. The method and system effectively improve the convergence speed, scene adaptability and system resource utilization efficiency of the federated learning model.

Description

Technical Field

[0001] This invention relates to the fields of smart logistics management, drone last-mile delivery, federated learning, and edge computing, specifically to a path-aware federated learning collaborative optimization method and system for smart logistics scenarios. Background Technology

[0002] With the rapid development of e-commerce and on-demand delivery services, smart logistics systems are evolving towards automation, intelligence, and precision. Drones, with their advantages of high mobility, flexible deployment, and high transportation efficiency, have been widely used in last-mile delivery, warehouse inspection, emergency supplies delivery, and logistics status monitoring, becoming an important infrastructure for building a new generation of smart logistics systems.

[0003] In practical applications, drones can continuously collect images, location information, and environmental perception data while performing delivery tasks, providing data support for logistics scheduling optimization, route planning, and anomaly identification. However, as the number of drones and the scale of delivery areas continue to expand, the traditional centralized data processing model is gradually revealing significant limitations. On the one hand, the centralized uploading of raw data from large-scale drones to the cloud will cause a surge in communication link load, making it difficult to meet the requirements of real-time delivery systems for low latency and high reliability. On the other hand, delivery data usually involves sensitive information such as user location and business trajectory, and centralized transmission and storage also bring potential data security and privacy risks.

[0004] To address the aforementioned issues, federated learning, a distributed modeling technique that eliminates the need for data to travel to the remote terminal, is increasingly being introduced into the field of smart logistics. This technology allows drones to train models locally, only uploading updated parameters to a server for aggregation, thereby reducing communication load while improving system privacy and security. However, most existing federated learning methods are based on static or low-mobility terminal designs, and their direct application to drone last-mile delivery scenarios still faces several challenges: First, when drones perform tasks in different delivery areas, the data they collect varies significantly in terms of scenario type, target distribution, and business attributes, resulting in strong non-independent and identically distributed characteristics in the local data of each drone. The geographical and functional differences between delivery areas cause significant differences in the gradient update direction of different nodes during model training, which can easily lead to slow convergence or even performance degradation of the federated model.

[0005] Secondly, drones are characterized by high-speed mobility, frequently entering and exiting the communication coverage area of ​​edge base stations during flight. The downloading and uploading of model parameters are often limited by brief communication windows. If a drone participating in the training experiences a communication interruption during federated learning, the local training results cannot be effectively aggregated, which not only reduces the stability of system training but also wastes computing resources and energy.

[0006] Furthermore, in real-world logistics systems, drones exhibit heterogeneity in terms of computing power, flight speed, battery capacity, and communication capabilities. Traditional federated learning typically organizes the training process using fixed rounds or synchronous aggregation, and the training time is often limited by the weaker nodes, leading to the "falling behind" problem and further reducing overall training efficiency.

[0007] While some existing studies have attempted to select nodes through random client selection or based solely on communication coverage, such methods typically ignore the spatial distribution information contained in the drone's flight path. This makes it difficult to ensure both communication reliability and data diversity, and can easily lead to redundant training samples or insufficient regional coverage, making it difficult to adapt to the complex and ever-changing smart logistics and delivery environment.

[0008] Therefore, in the scenario of smart logistics drone last-mile delivery, how to make full use of drone flight path information and achieve federated client selection and efficient model aggregation based on data distribution characteristics while ensuring communication stability and system resource constraints has become a key technical problem that urgently needs to be solved. Summary of the Invention

[0009] The purpose of this invention is to provide a path-aware federated learning collaborative optimization method and system for smart logistics. By deeply integrating the flight path information of UAVs into the node selection and model aggregation process of federated learning, a cloud-edge-device collaborative intelligent logistics perception and decision-making framework is constructed to solve the problems of low model training efficiency, slow convergence and insufficient generalization ability caused by data spatiotemporal heterogeneity, unstable communication connection and limited equipment resources in the last-mile delivery scenario of smart logistics.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A path-aware federated learning collaborative optimization method for smart logistics, the method being executed based on a cloud-edge-device system consisting of cloud servers, edge base stations, and delivery drones, includes the following steps: S1. Initialization and Resource Request: The cloud server sends the initial global model to the edge base station. The edge base station broadcasts resource requests to the delivery drones within its coverage area and collects the drones' computing power, flight trajectory, speed, and data quality parameters. S2, Dynamic UAV Selection: The edge base station executes a dynamic UAV selection algorithm based on the collected information; S3. Local Training and Aggregation: The selected drone performs model training based on locally collected logistics scenario data and uploads the trained model gradients to the edge base station. The edge base station uses a federated aggregation algorithm to perform lower-level aggregation on the received gradients and then uploads the aggregation results to the cloud server. The cloud server performs upper-level aggregation on the aggregation results from each edge base station and updates the global model. S4. Repeat the dynamic UAV selection and local training and aggregation steps until the global model converges.

[0011] Furthermore, the dynamic UAV selection algorithm includes: S21. Trajectory Prediction: Based on the historical movement trajectory of the UAV, predict its dwell time in the coverage area of ​​the edge base station and exclude UAVs that do not meet the communication time constraint; the communication time constraint is that the dwell time of the UAV in the coverage area of ​​the edge base station must complete at least one round of model gradient transmission. S22. Resource Filtering: Based on the bandwidth constraints of the edge base stations, the computing power of the remaining UAVs is sorted, and the top N% of nodes are selected as the candidate set; the sorting of computing power is based on the UAV's processor performance, available memory resources, and local training time overhead. S23. Pruning Optimization: Calculate the path redundancy based on the drone's flight trajectory, and combine it with the current load status of the edge base stations to prune nodes with high redundancy and load, and supplement high-value drone nodes covering data-scarce areas to improve the diversity and balance of training data.

[0012] Furthermore, the lower-level aggregation adopts a weighted aggregation strategy based on data scarcity. By calculating the global scarcity of data in different delivery areas and the local proportion of data from each drone node, higher weights are assigned to model updates from data-scarce areas to mitigate the impact of data heterogeneity on model convergence.

[0013] Furthermore, the upper-layer aggregation adopts a weighted aggregation strategy based on Jaccard similarity and data scale. By calculating the similarity between the data category set of each edge base station and the global category set, and combining it with its data scale, the contribution weight of each base station in the global model update is dynamically adjusted, thereby enhancing the model's adaptability to different logistics regional environments.

[0014] Furthermore, the method also includes communication and optimization mechanisms: the drone and the edge base station transmit compressed and encrypted model gradients wirelessly; the edge base station and the cloud server transmit aggregation results via wired communication; during the node selection phase, drones with strong computing power and sufficient power are given priority to participate in training; and during the local training phase, sensitive logistics data is desensitized and features are extracted, and only irreversible feature vectors or model gradients are transmitted to ensure data privacy and system energy efficiency.

[0015] A path-aware federated learning collaborative optimization system for smart logistics is provided to implement the above-mentioned method. The system includes a cloud server, an edge base station, and a delivery drone. The cloud server is used to perform drone path prediction and global model aggregation. The edge base station is used to perform dynamic drone selection and edge-side model aggregation. The drone is used to perform local data acquisition, model training, and gradient uploading.

[0016] As can be seen from the above technical solutions, the present invention has the following advantages compared with the prior art: (1) This invention constructs a cloud-edge-device collaborative federated learning architecture for the smart logistics last-mile delivery scenario, realizing distributed modeling of logistics data "without leaving the terminal", and significantly reducing the system communication load while protecting the privacy of sensitive information such as user location and delivery trajectory; (2) By introducing a path-aware mechanism, this invention uses the UAV flight trajectory as a common basis for communication reliability assessment and data value measurement, thereby achieving accurate screening and optimization of participating nodes in a dynamic mobile environment and improving the stability and continuity of the training process. (3) The present invention designs a two-layer weighted aggregation strategy based on data scarcity and Jaccard similarity, which effectively alleviates the problem of heterogeneous data distribution caused by differences in delivery areas, accelerates model convergence, and improves the generalization performance and state recognition accuracy of the global model in multi-regional and multi-scenario logistics environments. (4) The system of the present invention has good scalability and practicality, supports the dynamic addition and removal of drone nodes, and can be applied to various smart logistics scenarios such as urban last-mile delivery, park logistics, and emergency delivery, providing efficient and reliable technical support for real-time perception, intelligent scheduling and abnormal response of logistics systems. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system architecture of the present invention.

[0018] Figure 2 This is a simulation scenario diagram of the implementation and operation of this invention; Figure 3 This is a comparison chart of the accuracy of the present invention and the baseline method on Dirichlet distribution data; Figure 4 yes Figure 3 Line graph showing the number of drones participating in each round of training in the corresponding experiment; Figure 5 This is a comparison chart of the accuracy of the present invention and the baseline method on geographically distributed data; Figure 6 yes Figure 5 Line graph showing the number of drones participating in each round of training in the corresponding experiment. Detailed Implementation

[0019] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0020] like Figure 1 The system architecture diagram of the present invention shown illustrates that the path-aware federated learning collaborative optimization system for smart logistics includes a cloud server, edge base stations (BS), and delivery drones. The cloud server performs drone delivery path prediction and global model aggregation, and distributes the global model to the edge base stations. The edge base stations collect flight trajectories, speeds, computing power, and data quality parameters of drones within their coverage area, execute a dynamic drone selection algorithm, aggregate model gradients uploaded by the drones, and upload the aggregation results to the cloud server. During last-mile delivery, the drones perform model training based on locally collected logistics scenario data and transmit model gradients to the edge base stations. In this preferred embodiment, the drones have a built-in local model training unit, a data anonymization module, and a gradient encryption module.

[0021] Based on the aforementioned path-aware federated learning collaborative optimization system for smart logistics, this preferred embodiment also proposes a path-aware federated learning collaborative optimization method for smart logistics, comprising the following steps: S1. The cloud server sends the initial global model to the edge base station. The edge base station broadcasts resource requests to the delivery drones within its coverage area and collects the drones' computing power, flight trajectory, speed, and data quality parameters.

[0022] In practical operation, after the system starts, the cloud server constructs an initial global model (e.g., a logistics package recognition, stacking status detection, or delivery route planning model based on YOLOv8 or ResNet) based on historical logistics and delivery data (such as package status images, delivery route information, and warehousing environment data), and distributes it to registered delivery drones via edge base stations. The edge base stations broadcast resource request packets to drones within their wireless communication coverage area, collecting multiple parameters including computing power (processor performance, memory capacity), data quality (image clarity, annotation completeness, service category distribution), and flight status (real-time location, flight speed, planned trajectory). The edge base stations standardize these parameters to generate a drone status information table, providing data support for subsequent dynamic selection.

[0023] S2. Dynamic UAV Selection: The edge base station executes a dynamic UAV selection algorithm based on the collected information; specifically, the dynamic UAV selection algorithm includes: S21. Trajectory Prediction: Based on the historical flight trajectory and real-time status of the UAV, an LSTM model is used to predict its path and dwell time within the coverage area of ​​the edge base station in the future. The following formula (1) is used to select effective UAV nodes that can remain within the communication coverage area during the next round of federated learning training (from model download, local training to gradient upload), excluding nodes that may lose connection midway: (1); In formula (1): Indicates drone A binary indicator of whether the conditions for participation in federal aggregation are met (1 = valid, 0 = invalid). Indicates drone At any moment A binary indicator of whether it is within the coverage area of ​​an edge base station (1 = within range, 0 = not within range); Indicates drone At any moment Location prediction confidence; Indicates the end time of the federated upload process; The start time of the upload process; This represents the confidence threshold parameter.

[0024] S22. Resource Filtering: Among the set of drones that meet the communication constraints, the computing power of the drones (considering processor computing power, memory, and single-round training time) is sorted in descending order based on the available bandwidth of the edge base stations. For example, the top 20% are selected as the core candidate set. The first 20% to 40% are considered as an alternative set. This ensures that the nodes participating in the training have sufficient local computing efficiency.

[0025] S23. Pruning and Optimization: Analyze the delivery areas flown by candidate drones and calculate their path overlap (redundancy). Considering the current processing load of edge base stations, remove nodes with highly overlapping flight paths and duplicate data contributions. Simultaneously, supplement the candidate set with "high-value" drones whose flight paths cover data-scarce areas (such as remote delivery points, special business warehouses, and newly opened delivery routes) to improve the spatial diversity and business balance of the training data.

[0026] S3. Local Training and Aggregation: The selected drone performs model training based on locally collected logistics scenario data, and uploads the trained model gradients to the edge base station. The edge base station uses a federated aggregation algorithm to perform lower-level aggregation of the received gradients, and then uploads the aggregation results to the cloud server. The cloud server performs upper-level aggregation of the aggregation results from each edge base station and updates the global model. Specifically, this includes the following steps: S31. The selected drone receives the current global model and uses the logistics data (such as images and sensor data, which have been anonymized) collected locally during the delivery task to train the model. The training parameters can be dynamically adjusted according to the device's computing power: drones with high computing power execute more iteration rounds, while drones with low computing power appropriately reduce the number of iterations or increase the batch size to ensure that training is completed within the predicted communication time window; S32. Edge Base Station Layer Aggregation (Lower Layer Aggregation): The drone acts as a client, uploading the trained model gradients to the edge base stations it connects to. The edge base stations employ a weighted aggregation strategy based on data scarcity. First, the global scarcity of each delivery area (or business data type) is calculated. Node scarcity weights for each drone The calculations are performed using the following formulas (2) and (3) respectively: (2); (3); In formulas (2) and (3): This indicates the number of drones participating in the aggregation. The total number of classpaths; This represents a smoothing factor to prevent the denominator from being zero; it is set to 0.01 in the code. Indicates the first Global data scarcity in the classpath; For drones The Middle The percentage of classpaths; Indicates drone Number of data categories contained; Indicates client Scarcity weights; Indicates the total number of path categories; Furthermore, the aggregation weight of each UAV is calculated using the following formula (4): (4); In formula (4): Indicates the first The weight value of the drone in the lower-level aggregation; This represents the total number of drones participating in edge base station aggregation; and They represent the first frame and the first The amount of data generated locally by the drone; Indicates the first The scarcity of the routes for drones.

[0027] S33, Cloud Server Layer Aggregation (Upper Layer Aggregation): Each edge base station uploads the aggregated model gradients to the cloud server. The cloud server employs a weighted aggregation strategy based on Jaccard similarity and data size. It calculates the data category set for each edge base station. With global category collection Jaccard similarity between And taking into account the size of its data Determine its weight in the global aggregation. : (5); (6); In formulas (5) and (6): Indicates the first Jaccard similarity between each edge base station and the global category set; Indicates the first A set of categories for each edge base station; This represents the global category set obtained by aggregating all edge base stations; Indicates the first The weight values ​​of each edge base station in the upper-layer aggregation; This represents the total number of edge base stations participating in the aggregation; and They represent the first The and the first The size of the data aggregated from the edge base stations.

[0028] S4. Iteration and Convergence: Repeat steps S2 and S3. After each round of aggregation, the cloud server updates the global model and distributes it to the edge base stations until the performance of the global model on the validation set tends to stabilize (convergence).

[0029] Specifically, such as Figure 2 The simulation diagram shown represents the implementation operation. The dots represent drones, the red squares represent no-fly zones, and the curves passing through the dots represent the predicted path and the historical path of the drones.

[0030] Furthermore, the method also includes optimizations for communication, energy consumption, and data privacy: During communication, the drone and the edge base station use a 5G / dedicated wireless network for gradient transmission, with compression and a retransmission mechanism implemented before transmission. The edge base station monitors the drone's status in real time, pausing the mission for nodes that go offline midway to avoid resource waste.

[0031] In terms of energy consumption, drones with strong computing power and sufficient remaining battery power are given priority during the node selection phase. The drone can dynamically adjust the processor frequency according to the task load, enabling high-performance mode during training and switching to low-power mode when idle.

[0032] Regarding data privacy protection, the drone blurs and extracts features from the sensitive information it collects (such as recipient address identifiers and package detail images) locally, and only uploads irreversible feature vectors or encrypted model gradients, ensuring that the original data never leaves the device.

[0033] This system and method have been deployed and validated in simulated smart logistics scenarios (such as large-scale warehousing parks and urban last-mile delivery networks). Experimental results show that, compared with traditional "cover-only screening" (BS-cover) and "random selection" strategies, this invention performs better in highly non-independent and identically distributed (Non-IID) logistics data environments: The model convergence speed is significantly improved: such as Figure 3 and Figure 5 As shown, under two typical data distributions (Dirichlet distribution simulating business volume differences and geographical distribution simulating regional characteristics), the accuracy of the model of this invention increases faster with the number of training rounds.

[0034] Choose a more stable and efficient client: such as Figure 4 and Figure 6 As shown, the number of drones selected in each round of this invention is more stable and can continuously cover high-value nodes, avoiding the selection fluctuations and redundancy of traditional methods.

[0035] High system resource utilization: Through path-aware filtering and gradient compression, the pressure on wireless bandwidth and the energy consumption of drones are effectively reduced, while the security of logistics data is ensured through privacy protection mechanisms.

[0036] This invention supports the dynamic addition and removal of delivery drones, and new devices can be quickly integrated into the federated training process after certification. Furthermore, the system can be linked with warehouse management systems (WMS) and logistics scheduling platforms to achieve closed-loop decision-making from "environmental awareness" to "intelligent sorting" and "dynamic route optimization," thereby improving overall logistics operational efficiency.

[0037] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A path-aware federated learning collaborative optimization method for smart logistics, characterized in that, The method is based on a cloud-edge-device system consisting of cloud servers, edge base stations, and delivery drones, and includes the following steps: S1. Initialization and Resource Request: The cloud server sends the initial global model to the edge base station. The edge base station broadcasts resource requests to the delivery drones within its coverage area and collects the drones' computing power, flight trajectory, speed, and data quality parameters. S2, Dynamic UAV Selection: The edge base station executes a dynamic UAV selection algorithm based on the collected information; S3. Local Training and Aggregation: The selected drone performs model training based on locally collected logistics scenario data and uploads the trained model gradients to the edge base station. The edge base station uses a federated aggregation algorithm to perform lower-level aggregation on the received gradients and then uploads the aggregation results to the cloud server. The cloud server performs upper-level aggregation on the aggregation results from each edge base station and updates the global model. S4. Repeat the dynamic UAV selection and local training and aggregation steps until the global model converges.

2. The path-aware federated learning collaborative optimization method for smart logistics according to claim 1, characterized in that, The dynamic UAV selection algorithm includes: S21. Trajectory prediction: Based on the historical movement trajectory of the UAV, predict its dwell time in the coverage area of ​​the edge base station and exclude UAVs that do not meet the communication time constraints. The communication time constraint requires that the UAV stay in the coverage area of ​​the edge base station for at least one round of model gradient transmission. S22. Resource Filtering: Based on the bandwidth constraints of the edge base stations, sort the computing power of the remaining drones and select the top N% of nodes as the candidate set. The ranking of computing power is based on the UAV's processor performance, available memory resources, and local training time overhead; N is a preset value. S23. Pruning Optimization: Calculate the path redundancy based on the drone's flight trajectory, and combine it with the current load status of the edge base station to prune nodes with high redundancy and load, and supplement high-value drone nodes to cover areas with scarce data.

3. The path-aware federated learning collaborative optimization method for smart logistics according to claim 2, characterized in that, The trajectory prediction uses an LSTM model to predict the future path of the UAV, and the effectiveness of the UAV is determined by the following formula (1): （1）； In formula (1): Indicates drone A binary indicator of whether the conditions for participation in federal aggregation are met (1 = valid, 0 = invalid). Indicates drone At any moment A binary indicator of whether it is within the coverage area of ​​an edge base station (1 = within range, 0 = not within range); Indicates drone At any moment Location prediction confidence; Indicates the end time of the federated upload process; The start time of the upload process; This represents the confidence threshold parameter.

4. The path-aware federated learning collaborative optimization method for smart logistics according to claim 1, characterized in that, The lower-level aggregation is the aggregation from drones to edge base stations, and adopts an aggregation strategy based on data scarcity weighting.

5. The path-aware federated learning collaborative optimization method for smart logistics according to claim 4, characterized in that, The data scarcity-weighted aggregation strategy includes: Calculate the global data scarcity of all participating drones within the delivery area; Calculate the node scarcity weight based on the proportion of each drone in different logistics scenario data. The local model update of the UAV is weighted and aggregated based on the aforementioned scarcity weights; The global data scarcity and node scarcity weights are calculated using the following formulas (2) and (3), respectively: （2）； （3）； In formulas (2) and (3): This indicates the number of drones participating in the aggregation. The total number of classpaths; Indicates the smoothing factor; Indicates the first Global data scarcity in the classpath; For drones The Middle The percentage of classpaths; Indicates drone Number of data categories contained; Indicates client Scarcity weights; Indicates the total number of path categories; The aggregate weight of each UAV is calculated using the following formula (4): （4）； In formula (4): Indicates the first The weight value of the drone in the lower-level aggregation; This represents the total number of drones participating in edge base station aggregation; and They represent the first frame and the first The amount of data generated locally by the drone; Indicates the first The scarcity of the routes for drones.

6. The path-aware federated learning collaborative optimization method for smart logistics according to claim 1, characterized in that, The upper-layer aggregation is an aggregation from edge base stations to cloud servers, using a weighted aggregation strategy based on Jaccard similarity and data scale.

7. The path-aware federated learning collaborative optimization method for smart logistics according to claim 6, characterized in that, The Jaccard similarity weighted aggregation specifically includes the following steps: Construct a set of logistics data categories contained in all edge base stations; Calculate the Jaccard similarity between the local category set and the global category set of each edge base station; The model aggregation weights are determined based on the similarity results and the data scale of each edge base station. The aggregation weight of each edge base station is calculated using the following formulas (5) and (6): （5）； （6）； In formulas (5) and (6): Indicates the first Jaccard similarity between each edge base station and the global category set; Indicates the first A set of categories for each edge base station; This represents the global category set obtained by aggregating all edge base stations; Indicates the first The weight values ​​of each edge base station in the upper-layer aggregation; This represents the total number of edge base stations participating in the aggregation; and They represent the first The and the first The size of the data aggregated from the edge base stations.

8. The path-aware federated learning collaborative optimization method for smart logistics according to claim 1, characterized in that, The method also includes communication and optimization steps: The drone transmits model gradients to the edge base station via a wireless communication network, and the gradients are compressed and encrypted. The edge base station and the cloud server transmit the aggregation results through a wired communication network; When selecting dynamic drones, prioritize drones with strong computing power and sufficient remaining battery power; When drones are trained locally, sensitive logistics data collected is anonymized and features are extracted, and only irreversible feature vectors or model gradients are transmitted.

9. A path-aware federated learning collaborative optimization system for smart logistics, used to implement the path-aware federated learning collaborative optimization method for smart logistics as described in any one of claims 1-8, characterized in that, This includes cloud servers, edge base stations, and delivery drones; The cloud server is used to perform UAV path prediction and global model aggregation, and to distribute the global model to the edge base station; The edge base station is used to collect the trajectory, speed, computing power and data quality parameters of drones within its coverage area, and to execute the dynamic drone selection algorithm, as well as to aggregate the model gradients uploaded by drones and upload the aggregation results to the cloud server. The delivery drone is used to perform model training based on locally collected traffic data and transmit the model gradient to the edge base station.

10. The path-aware federated learning collaborative optimization system for intelligent logistics according to claim 9, characterized in that, The delivery drone is equipped with a local model training unit, a data anonymization module, and a gradient encryption module.

Images