A multi-agent data collection autonomous navigation method for indoor scenes

Abstract

This invention discloses a multi-agent data acquisition autonomous navigation method for indoor scenarios, comprising the following steps: constructing a multi-agent edge intelligent system model; constructing and pre-training multiple neural network classification models; constructing a radio electromagnetic map; calculating the channel propagation gain mapping function for each sub-map; fitting a high-order nonlinear fitting model; using minimizing the squared error of the channel propagation gain mapping function and the high-order nonlinear fitting model as the optimization objective, solving for the optimal parameter set, and calculating the sensor spectral efficiency; using minimizing the weighted sum of the Euclidean distance between the mobile intelligent vehicle's own trajectory and the corresponding reference path point, the communication throughput cost, and the classification error of the neural network classification model as the optimization objective, constructing an optimization problem and transforming it into a convex problem, solving for a local optimum, and obtaining the optimal navigation path based on the local optimum. This invention can improve the flexibility and efficiency of path planning while ensuring safety.

Description

Technical Field

[0001] This invention relates to the field of multi-agent autonomous navigation technology, specifically to a multi-agent data acquisition autonomous navigation method for indoor scenarios. Background Technology

[0002] Sensor devices transmit data to edge servers via wireless channels for AI model training, enabling real-time monitoring and intelligent decision-making. However, in real-world industrial environments, sensors are typically statically deployed in complex spaces. Their fixed positions and severe non-line-of-sight propagation issues result in low wireless signal transmission efficiency, limiting data acquisition quality and model training performance. This poses challenges to the data-driven path planning and decision-making aspects of autonomous navigation technology. Existing robot autonomous navigation technologies often focus on single objectives such as obstacle avoidance and path tracking, neglecting the impact of communication channel quality on data acquisition and ignoring the varying sensitivities of different AI models to training data. This leads to unreasonable resource allocation, making it difficult to meet the demands of complex models. Static systems lack dynamic adjustment capabilities and cannot adapt to channel changes, exacerbating communication blind spots.

[0003] In existing robot autonomous navigation technologies, such as the ground robot autonomous navigation method and robot based on plane fitting (patent publication number CN114993316A), a map is built and a path is planned by receiving LiDAR point cloud data and combining it with odometry information. This can accurately simulate robot posture and integrate terrain analysis for planning to save resources. Another example is the global path planning method and system for UAV data collection (patent publication number CN114167898A), which aims to maximize data collection benefits and minimize data collection time. It uses decomposition strategies, pointer network deep learning neural models, etc., to determine the sensor cluster set and access order, achieving the goal of minimizing collection time and maximizing data collection benefits. However, these publicly disclosed robot autonomous navigation technologies focus only on the navigation algorithm itself, without considering using data to improve model performance, or only consider data collection benefits and time. They neglect the overall coordination between edge intelligence learning performance and communication needs in complex communication environments such as factories, failing to achieve integrated optimization of communication and learning performance. Summary of the Invention

[0004] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a multi-agent data acquisition autonomous navigation method for indoor scenarios. This method addresses the problem that existing technologies fail to simultaneously meet data acquisition requirements and optimize model training performance during motion planning. By rationally planning the robot's motion path while satisfying kinematic equations and safety obstacle avoidance constraints, it maximizes data acquisition and model training performance while ensuring safe obstacle avoidance, thereby improving the overall performance of the industrial IoT system. This method has advantages such as low complexity, high practicality, high flexibility, and wide applicability.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a multi-agent data acquisition and autonomous navigation method for indoor scenarios, comprising the following steps: Construct a multi-agent edge intelligence system model, including edge servers, mobile intelligent vehicles, and sensors; Construct multiple neural network classification models and pre-train them; Obtain signal strength and propagation loss in indoor communication scenarios to construct a wireless electromagnetic map; The radio electromagnetic map is divided into multiple sub-maps based on the signal strength distribution, and the channel propagation gain mapping function for each sub-map is calculated. A high-order nonlinear fitting model of the signal propagation characteristics of the entire radio electromagnetic map is obtained by fitting the channel propagation gain mapping function of all sub-maps. The optimal parameter set is obtained by minimizing the error square of the channel propagation gain mapping function and the high-order nonlinear fitting model as the optimization objective, and the sensor spectral efficiency is calculated. The optimization problem is constructed by calculating the weighted sum of the Euclidean distance between the mobile intelligent vehicle's own trajectory and the corresponding reference path point, the communication throughput cost, and the classification error of the neural network classification model, and taking the minimization of the weighted sum as the optimization objective. The optimization problem is transformed into a convex problem, and the local optimum of the path planning control input is obtained by iterative solution. The optimal navigation path is obtained based on the local optimal solution, and the mobile intelligent vehicle travels along the optimal navigation path.

[0006] As a preferred technical solution, the signal strength and propagation loss of the indoor communication scenario are obtained to construct a radio electromagnetic map, specifically including: Use the Blender tool to build an indoor communication scenario, and then import the built indoor communication scenario into the Sionna library; The Sionna library performs wireless communication simulations on imported indoor communication scenarios, calculates signal strength and propagation loss, and constructs a wireless electromagnetic map.

[0007] As a preferred technical solution, the channel propagation gain mapping function for each sub-map is calculated, specifically including: The channel propagation gain mapping function for line-of-sight propagation is expressed as: ; Where RM represents the channel propagation gain mapping function, Indicates time The position of the mobile intelligent vehicle in the planar coordinate system. Indicates the first The location of each sensor Indicates path loss. This represents the path loss index. Describes the Euclidean norm. Indicates the calculation position and location The Euclidean distance between them; The channel propagation gain mapping function for non-line-of-sight propagation is expressed as: ; in, This represents a correction factor that takes into account the propagation effects of radio signals. This represents the path loss index for non-line-of-sight propagation.

[0008] As a preferred technical solution, a high-order nonlinear fitting model of the entire radio electromagnetic map signal propagation characteristics is obtained by fitting the channel propagation gain mapping function of all sub-maps, specifically expressed as: ; in, This represents a high-order nonlinear fitting model. Indicates time The position of the mobile intelligent vehicle in the planar coordinate system. Indicates the first The location of each sensor Indicates the first Gain coefficient of each sub-map Indicates the first The propagation index of a submap, where L represents the number of submaps. This indicates the region indication function, used to distinguish the current location of the mobile intelligent vehicle. Sub-map , This represents the Euclidean norm.

[0009] As a preferred technical solution, minimizing the squared error of the channel propagation gain mapping function and the high-order nonlinear fitting model is taken as the optimization objective. The optimal parameter set is obtained by solving the problem, and the sensor spectral efficiency is calculated, specifically including: By solving an optimization problem with the objective of minimizing the squared error of the channel propagation gain mapping function and the higher-order nonlinear fitting model, the optimal parameter set of the higher-order nonlinear fitting model is obtained. ; Optimal parameter set based on high-order nonlinear fitting model The sensor spectral efficiency is calculated using the channel propagation gain mapping function, and is expressed as: ; in, Indicates the first Each sensor in the time slot Spectral efficiency, This represents a high-order nonlinear fitting model. Indicates time The position of the mobile intelligent vehicle in the planar coordinate system. Indicates the first The location of each sensor For complex Gaussian noise power, Indicates the first The transmit power of each sensor, and These represent the propagation index and the gain coefficient, respectively.

[0010] As a preferred technical solution, the weighted sum of the Euclidean distance between the mobile intelligent vehicle's own trajectory and the corresponding reference path point, the communication throughput cost, and the classification error of the neural network classification model is calculated. Minimizing the weighted sum is used as the optimization objective, and the optimization problem is constructed as follows: ; in, Indicates the current state. This indicates that the mobile intelligent vehicle is at any time The control input vector, , , Indicates hyperparameters, This represents the distance cost function, used to calculate its own trajectory. The sum of Euclidean distances between the reference path point and the corresponding reference path point. The communication throughput cost function is represented by the following: This represents the mean classification error of the model. This represents the number of discrete time slots within the field of view of the mobile intelligent vehicle. This represents the time offset within the prediction time domain. Represents a state evolution model. , , Represents the coefficient matrix. and These represent the minimum and maximum values ​​of the control input speed for the mobile intelligent vehicle, respectively. This represents the control input vector at the h-th time after the current time t. This represents the control input vector at the (h+1)th time after the current time t. and These represent the minimum and maximum values ​​of acceleration, respectively. This represents the collision avoidance function. express Time of the first The state vector of each obstacle This indicates the safe distance between the mobile intelligent vehicle and obstacles. This represents a set of obstacles.

[0011] As a preferred technical solution, the distance cost function is specifically expressed as follows: ; in, Indicates a reference path point; The communication throughput cost function is specifically expressed as follows: ; in, Indicates the location of the mobile intelligent vehicle With the Each sensor in time Spectral efficiency during communication. Indicates the number of sensors; ; in, This represents the total number of neural network classification models. Harmony Hyperparameters representing the learning difficulty of the model Indicates the relationship with the first A set of sensors associated with a neural network classification model. This indicates the bandwidth allocated to each sensor group. Indicates the duration of the time slot. Indicates the first The number of bits required for each sample in each model Indicates the first The number of historical samples accumulated by each model.

[0012] As a preferred technical solution, a collision avoidance optimization step is also included, specifically: safe distance Replace with time-varying safe distance sequence ,in, Indicates the safety distance parameter; Collision avoidance constraints are expressed as: ; in, The time step variable represents the time-varying safety distance, corresponding to the minimum allowable distance between the moving intelligent vehicle and the obstacle at time t in the prediction time domain. This indicates the calculation of Euclidean distance. Indicates in Time of the first Geometry modeling of an obstacle Indicating that the mobile intelligent vehicle is in Geometric modeling at time step, and These represent the minimum distance and the maximum distance, respectively. A regularization penalty term based on the L1 norm is added to the optimization objective. ; The collision avoidance constraint is transformed into a linear dual form, expressed as: ; in, and Let T represent the dual variables of the geometric constraints of the mobile intelligent vehicle and the obstacle surface, respectively, with the superscript T indicating transpose. This represents a slack variable, used to convert inequality constraints into equality constraints. , , Represents the number of geometric surfaces on the obstacle's surface. Represents the set of real numbers. Indicates the location of the mobile intelligent vehicle The relevant translation matrix, , , This indicates the number of geometric faces on the surface of the mobile intelligent vehicle. This is a rotation matrix related to the orientation angle.

[0013] As a preferred technical solution, the optimization problem is transformed into a convex problem, specifically including: A surrogate function for the spectral efficiency of communication between the mobile intelligent vehicle and sensors is constructed, transforming the optimization problem into a convex problem, expressed as: ; ; in, Indicates the current state. This indicates that the mobile intelligent vehicle is at any time The control input vector, and Let T represent the dual variables of the geometric constraints of the mobile intelligent vehicle and the obstacle surface, respectively, with the superscript T indicating transpose. Represents slack variables. , , Indicates hyperparameters, This represents the distance cost function, used to calculate its own trajectory. The sum of Euclidean distances between the reference path point and the corresponding reference path point. This represents the number of discrete time slots within the field of view of the mobile intelligent vehicle. This represents the time offset within the prediction time domain. This represents a proxy function. This represents the total number of neural network classification models. Harmony Hyperparameters representing the learning difficulty of the model Indicates the relationship with the first A set of sensors associated with a neural network classification model. This indicates the bandwidth allocated to each sensor group. Indicates time slot, Indicates the first The number of bits required for each sample in each model Indicates the first The number of historical samples accumulated by each model This represents the regularization penalty term. The time step variable represents the time-varying safety distance. , , Represents the coefficient matrix. and These represent the minimum and maximum values ​​of the control input speed for the mobile intelligent vehicle, respectively. This represents the control input vector at the h-th time after the current time t. This represents the control input vector at the (h+1)th time after the current time t. and These represent the minimum and maximum values ​​of acceleration, respectively. and These represent the minimum distance and the maximum distance, respectively. , , Represents the number of geometric surfaces on the obstacle's surface. Represents the set of real numbers. Indicates the location of the mobile intelligent vehicle The relevant translation matrix, , , This indicates the number of geometric faces on the surface of the mobile intelligent vehicle. This is a rotation matrix related to the orientation angle.

[0014] As a preferred technical solution, a surrogate function for the spectral efficiency of communication between the mobile intelligent vehicle and the sensors is constructed, expressed as: ; in, This represents a proxy function. Indicates the current state. Let L represent a feasible solution to the optimization problem, and let L represent the number of submaps. This indicates the region indication function, used to distinguish the current location of the mobile intelligent vehicle. Sub-map , Indicates the first Gain coefficient of each sub-map Indicates the first The propagation index of each sub-map For complex Gaussian noise power, Indicates the first The transmit power of each sensor, , This indicates the position of the mobile intelligent vehicle in a planar coordinate system. This indicates the reference orientation of the mobile intelligent vehicle at time t. This represents the Euclidean norm.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) The present invention adopts a multi-agent collaborative modeling technology scheme with a mobile intelligent vehicle as the main intelligent agent and sensors as the slave intelligent agents, which solves the technical problems of static sensor data being scattered and single devices being unable to efficiently collect and upload data, and achieves the technical effect of realizing dynamic data collection and edge server collaboration, thereby improving the overall data transmission and processing efficiency of the system.

[0016] (2) The present invention adopts a technical solution based on Blender and Sionna library to build electromagnetic map and combine it with high-order nonlinear model fitting. It solves the technical problem that the non-line-of-sight propagation problem of static sensors deployed in complex industrial environment leads to low wireless signal transmission efficiency and limited data acquisition and model training performance. It achieves the technical effect of accurately simulating line-of-sight and non-line-of-sight propagation characteristics, providing reliable channel quality basis for robot path planning, and improving data acquisition efficiency and model training performance.

[0017] (3) The present invention adopts an obstacle avoidance constraint technology solution that combines an arbitrary convex non-point mass model with dynamic safety distance adjustment. This solves the technical problem that the traditional point mass model ignores the actual shape of the robot and the obstacle, resulting in high obstacle avoidance conservatism and low navigation efficiency. It achieves the technical effect of accurately depicting the spatial relationship between the robot and the obstacle, and improving the flexibility and efficiency of path planning while ensuring safety.

[0018] (4) This invention can adapt well to indoor environments with multiple obstacles and dynamic changes in IoT load. By tightly coupling collision avoidance, communication throughput and learning performance, it takes minimizing the weighted comprehensive cost of path tracking, communication throughput and model learning performance as the optimization objective. It constructs an optimization problem that maximizes robot communication efficiency and classification model learning performance while minimizing comprehensive cost. Based on L1 regularization, linear duality and MM algorithm, the optimization problem is transformed into a smooth convex problem. The local optimal solution of the path planning control input is obtained by iterative solution. The optimal navigation path is obtained based on the local optimal solution of the control input. While ensuring collision avoidance, it enhances data acquisition efficiency and model learning performance.

[0019] (5) The present invention adopts a resource allocation technology scheme that introduces a neural network classification model learning performance regularizer and combines it with model hyperparameter adjustment. This solves the technical problem of low resource utilization efficiency caused by data collection not taking into account the differences in model training difficulty. It achieves the technical effect of prioritizing data collection for models with high training requirements and improving the overall model learning performance.

[0020] (6) This invention introduces a weight factor when constructing the optimization problem, so that path tracking, communication throughput and learning performance can be balanced by adjusting the size of the weight factor. Moreover, by adjusting the weight factor, it can flexibly adapt to different application scenarios and meet different task requirements, and has a wider range of applications.

[0021] (7) Through simulation experiments, compared with existing robot autonomous navigation schemes, the present invention has shown obvious advantages in terms of the accuracy of autonomous navigation and the overall performance of the system. It can adaptively adjust the path in a variable environment, and take into account obstacle avoidance, communication and learning performance, and has broad application prospects. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the multi-agent data acquisition and autonomous navigation method for indoor scenarios according to the present invention. Figure 2 A schematic diagram of the overall architecture of the multi-agent edge intelligence system model of this invention; Figure 3 The path gain electromagnetic map generated by the sensor at different locations in this invention; Figure 4 This diagram illustrates the autonomous navigation trajectory of a robot in an indoor scenario, comparing the autonomous navigation method of this invention with existing solutions such as PMM+Commu, OBCA, RDA, and RDA+Commu. Figure 5 A diagram showing the performance comparison of different path planning methods in terms of communication data acquisition efficiency; Figure 6This diagram illustrates the comparison of classification errors of neural network classification models under different path planning methods. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] Example like Figure 1 As shown, this embodiment provides a multi-agent data acquisition autonomous navigation method for indoor scenarios. A mobile intelligent vehicle (i.e., a robot) autonomously navigates in a complex environment containing M obstacles, collecting data from various sensors. Upon reaching the destination, the data is transmitted to an edge server via an uplink for training different models. This embodiment is applicable to edge intelligence and robot data gathering (RDG) scenarios. It integrates communication data acquisition with learning performance optimization, achieving the goal of maximizing data acquisition and model training performance while ensuring safe obstacle avoidance. Furthermore, it allows for adjustment of weight factors to flexibly adapt to the specific needs of different application scenarios. The specific steps include: S1: Construct a multi-agent edge intelligence system model, build multiple neural network classification models, and pre-train the neural network classification models based on historical sample data; like Figure 2 As shown, the multi-agent edge intelligence system model includes an intelligent edge server with N antennas, a mobile intelligent vehicle, and K fixed-position sensors. This model is used in complex indoor environments in industrial settings. In this environment, the mobile intelligent vehicle acts as the master agent (MA), the K sensors as slave agents (SA), and there are multiple obstacles of different types. One edge server is paired with one indoor scene. The mobile intelligent vehicle autonomously navigates in a complex environment containing M obstacles. These obstacles can be internal machinery, fixed facilities, or other static / dynamic obstacles in indoor scenes such as factories, and are represented by a set of obstacles. In the simulation experiment, N=16 and K=2 are preferred. The data collected by each sensor is used to train the corresponding neural network classification model. The corresponding neural network classification models include CNN model and SVM model. These two neural network classification models are built at the edge server and trained using data samples collected by different sensors. In order to be more realistic, these neural network classification models will first use historical sample data for model pre-training. This allows the trained model parameters (such as weights, biases, learning rates, etc.) to be fed back to the local user in the subsequent model feedback. In this way, the local user can directly use this model to classify the new data they have collected. In this embodiment, the current visible range of the mobile intelligent vehicle is divided into... A discrete time slot, i.e. The time step is At any time step Inside, the robot's state is represented as ,in , This represents the robot's position in a planar coordinate system. This represents the robot's orientation angle at that location, and the robot's orientation angle in each time slot. control input vector Defined as ,in , This indicates the robot's current linear velocity. Represents the robot's current angular velocity, the first in the scene. The position of each sensor is represented as ,in The navigation environment of a mobile intelligent vehicle mainly consists of a set of reference waypoints. With obstacles It consists of two parts, the first The state vectors of the path points are: ,in This indicates the coordinates of the path point, while This indicates the recommended forward direction angle at that point, the first... The state vectors of the obstacles are: ,in Represents the center position of the obstacle; In this embodiment, the robot and obstacles use a non-point mass model, taking into account the actual shapes of the robot and obstacles. Therefore, the robot... The shape of time is modeled as a geometric set Specifically, This describes the spatial region occupied by the robot at a given moment; the set is determined by the robot's state. and its own geometry The decision is made jointly, and its expression is as follows: ; in, For direction angle The relevant rotation matrix, and This is related to the robot's position. Related translation matrices, set Used to characterize the geometry of the robot, where , , The number of geometric surfaces representing the robot's surface. Representing the set of real numbers, specifically, It is a coefficient matrix that defines the boundary constraint directions of the robot's initial shape. It is a real number vector; So in Time of the first Each obstacle is modeled as a geometric set. , No. An obstacle can be represented as: ; in , , Represents the number of geometric surfaces on the obstacle's surface. The dimension is The matrix, where each row corresponds to a geometric face of the obstacle, and the three elements of the row are related to the x, y, and z coordinate axes in three-dimensional space, respectively; It is a dimension of A vector, where each element of the vector corresponds to a geometric surface of the obstacle, used to determine the specific location of that geometric surface in space; In this embodiment, the robot must satisfy the collision avoidance constraint function during its movement. For any time and any obstacle The constraint can be expressed as: ; in, express Time of the first The state vector of each obstacle Indicates the safe distance between the robot and the obstacle; This function calculates the robot's... Time relative to the first Accurately determine the distance to each obstacle and always maintain a safe distance; S2: Obtain signal strength and propagation loss in indoor communication scenarios, and construct a wireless electromagnetic map, specifically including: Use the Blender tool to build indoor communication scenes such as factories and use the rendering function to generate scene visualization effects. Import the built indoor scenes such as factories into the Sionna library. The Sionna library performs wireless communication simulation on imported 3D scenes, calculates signal strength and propagation loss. Specifically, it sets the location and number of base stations and the signal propagation model through the Sionna library's API, and calculates the signal propagation loss based on obstacles, distance and other factors in the environment to obtain a wireless electromagnetic map. S3: Divide the radio electromagnetic map based on the characteristics of signal strength distribution, and calculate the channel propagation gain mapping function RM for each small-scale sub-map; In this embodiment, the channel propagation gain mapping function RM is divided into line-of-sight (LOS) propagation and non-line-of-sight (NLOS) propagation based on whether there are obstacles between the transmitter and receiver that obstruct the direct propagation of the signal. Specifically, it includes: The channel propagation gain mapping function RM in the case of line-of-sight (LOS) propagation is specifically expressed as follows: ; in, Represents the moment The robot's position in a planar coordinate system Representing the The location of each sensor This represents the path loss at a distance of 1 meter. This is the path loss index. Describes the Euclidean norm. Indicates the calculation position and location The Euclidean distance between them; The channel propagation gain mapping function RM in the case of non-line-of-sight (NLOS) propagation is specifically expressed as follows: ; in, Represents the path loss exponent of NLOS, when Indicates no distance dependency. This represents a correction factor that takes into account the propagation effects of complex radio signals.

[0025] In this embodiment, based on the distribution characteristics of signal strength, image segmentation technology is used to divide the radio electromagnetic map into regions, dividing the entire environment into multiple small-scale sub-maps. Each sub-map corresponds to an independent propagation model and has relatively consistent propagation characteristics within the sub-map, that is, a sub-map is either line-of-sight (LOS) or non-line-of-sight (NLOS) propagation.

[0026] S4: A high-order nonlinear fitting model for the propagation characteristics of the entire radio electromagnetic map signal is obtained by fitting the RM function of the propagation gain mapping function of all sub-map channels. Specifically, it is expressed as: ; in, Indicates the first The propagation index of each sub-map region Indicates the first Gain coefficient for each sub-map region Used to separate the input variables of a function from the set of parameters on which the model depends. Representing the electromagnetic map, the first The RM is essentially a channel gain mapping function, but because it can be pre-generated and stored as a map covering the entire area to describe the electromagnetic propagation characteristics at any location in space, it is also called an electromagnetic map. In this embodiment, the radio electromagnetic map is divided into L sub-maps based on the signal strength distribution characteristics. The sub-map regions are constrained by a set of inequality constraints. definition, and The area was jointly defined The boundary, Represents the number of submap boundaries. The region indication function is used to distinguish the robot's current position. Sub-map The indicator function is 1 when the robot is located in this sub-map area, and 0 otherwise. The indicator function is specifically expressed as follows: ; S5: Obtaining a high-order nonlinear fitting model The actual set of parameters will minimize the original radio electromagnetic map channel propagation gain mapping function. and higher-order nonlinear fitting models Using the squared error under the same input as the optimization objective, the optimal parameter set is obtained by solving for the sensor spectral efficiency, which specifically includes: In this embodiment, the propagation characteristics of each sub-map are determined by a set of parameters. Characterization is performed in the LOS region. middle, and Different NLOS subregions Different transmission indices may exist. and gain coefficient By solving an optimization problem that aims to minimize the squared difference between the actual function RM and the higher-order nonlinear model, a higher-order nonlinear fitting model is obtained. Optimal parameter set , making To approximate RM as closely as possible, the optimization problem can be specifically expressed as: ; in, and These represent the propagation index and gain coefficient in the LOS region, respectively. In this embodiment, the robot in the first... Each time slot moves to position The specific expression for the received signal-to-noise ratio is: ; in, Representative from the first Uplink channel gain from sensor to robot receiver The channel gain is the complex Gaussian noise power including inter-user interference. Relying on the wireless propagation characteristics between the robot and the sensor, it can... Model estimation, then the first Each sensor in the time slot The spectral efficiency can be expressed as: ; Use the method to find the optimal set of parameters higher-order nonlinear fitting model Let RM represent the original electromagnetic map channel propagation gain mapping function, then the... Each sensor in the time slot The spectral efficiency can ultimately be expressed as: ; in, Indicates the first The transmit power of each sensor is the complex Gaussian noise power including inter-user interference.

[0027] S6: Construct an optimization problem. The optimization problem aims to maximize data acquisition and model training performance while ensuring accurate obstacle avoidance. The optimization objective is to minimize the weighted comprehensive cost of path tracking, communication throughput, and model learning performance. In this embodiment, to maximize the amount of communication data collected through autonomous navigation and improve the model's learning performance while ensuring the robot can safely avoid obstacles and successfully reach the target location, and to minimize the deviation between the robot's actual path and the preset reference path, a CLD (Continuous Communication Data Acquisition) scheme integrating communication data acquisition and learning performance optimization is constructed to achieve the goal. Therefore, the objective of the optimization problem can be understood as: minimizing the weighted comprehensive cost of path tracking, communication throughput, and model learning performance while ensuring the robot safely avoids obstacles and successfully reaches the target location. The optimization problem is specifically expressed as: ; ; ; ; in, This represents the number of discrete time slots within the robot's field of view. Represents the time offset within the prediction time domain and , , , Indicates that the robot is The orientation angle of the position, Indicates the robot in the time slot The control input vector, and These represent the robot's current linear velocity and angular velocity, respectively. , , Hyperparameters are used to maintain... , , The numerical balance between these parameters is used to prevent any single cost item from having an excessive impact on the optimization process. These hyperparameters can also be used to adjust the weights of different performance metrics; for example, when path tracing requirements are higher, they can be appropriately increased. Similarly, the corresponding value can be adjusted appropriately if there are requirements for communication throughput and model learning performance. and The value, and These represent the minimum and maximum allowable values ​​for the robot control input speed, respectively. and Represents the minimum and maximum acceleration limits. The distance-cost function calculates the trajectory of itself. and corresponding reference path points The sum of the Euclidean distances between them The communication throughput cost function is represented by the following: This represents the regularizer for AI model learning performance, specifically the mean of the model's classification error. Indicates a local reference path. Indicates the robot's location With the Each sensor in time Spectral efficiency during communication. Indicates the number of static sensors. This represents the total number of AI models. Indicates the relationship with the first A set of sensors associated with a learning model. This indicates the bandwidth allocated to each sensor group. Indicates the duration of the time slot. Indicates the first The number of bits required for each sample in each model Indicates the first The number of historical samples accumulated by each model, and and Hyperparameters, which characterize the learning difficulty of a model, have corresponding hyperparameters for different classification models. , , , Represents the coefficient matrix. Representing the state evolution model and , Indicates the robot's acceleration. express Time of the first The state vector of each obstacle Represents a set of obstacles. This represents the collision avoidance function. Indicates the safe distance between the robot and the obstacle. This represents the control input vector at the h-th time after the current time t. This represents the control input vector at the (h+1)th time after the current time t; In this embodiment, there are constraints for all time steps t. Under this constraint, a state evolution model is used. It is possible to predict the future trajectory under the current state. That is, the robot's next state can be determined by its current state. Current control input and coefficient matrix , and Calculated; In the simulation experiment, the "PMM+Commu scheme" failed to enable the robot to successfully reach the target point in a complex indoor environment and failed to complete the basic data collection task. Therefore, it was not compared with its data collection efficiency and classification error value. S7: Based on L1 regularization and linear duality techniques, nonlinear constraint problems are solved. At the same time, the MM algorithm is used to transform the optimization problem into a convex problem, and the local optimal solution of the path planning control input is obtained by iterative solution. In this embodiment, the nonlinear constraint problem is solved based on L1 regularization and linear duality techniques, specifically including: Due to nonsmooth constraints The existence of this means that existing methods typically employ a fixed safety distance. While it can handle all obstacles, it cannot accurately reflect the spatial distribution characteristics of collision risk. To address this issue, a method based on prediction step size is proposed. Built-in time-varying safe distance sequence The dynamic safety distance strategy, after introducing this time-varying safety distance mechanism, the collision avoidance constraint is restated as follows: ; in, The time step variable representing the time-varying safety distance is a variable that changes with time step t, corresponding to the minimum permissible distance between the robot and the obstacle at time t in the prediction time domain. This represents any time step t and any obstacle m; In this embodiment, the dynamic safety distance strategy requires the safety distance to gradually decrease as the prediction time extends; that is, a larger safety distance should be allocated for the upcoming time step. To enhance conservatism in short-term obstacle avoidance; while for longer time steps, smaller safety distances are allocated. To enhance the flexibility of trajectory planning, thereby better balancing local safety and global optimization capabilities; In this embodiment, in order to automatically generate this non-uniformly distributed safety distance sequence during the optimization process, a regularization penalty term based on the L1 norm is introduced into the objective function P1. This regularization term utilizes the sparsity constraint property of L1 regularization to make the safe distance parameter... It exhibits a non-uniform distribution over time, thus adaptively focusing on nearby moments; In this embodiment, to address the non-smoothness of the collision avoidance constraint, the original collision avoidance constraint is... This is equivalently transformed into a smooth linear dual form, specifically expressed as: ; in, and The dual variables introduced by the geometric constraints of the robot body and the obstacle surface. This represents a slack variable, used to convert inequality constraints into equality constraints; In this embodiment, the optimization problem is transformed into a convex problem based on the MM algorithm, and the local optimal solution of the path planning control input is obtained by iterative solution. To address this comprehensive cost optimization problem, this embodiment employs the Majorize-Minimize (MM) algorithm to find... The surrogate function, the MM algorithm, transforms the optimization of the original problem into solving a series of suboptimal problems, thereby gradually approaching the optimal solution of the original problem by constructing a lower bound function. Replace P1 This leads to a series of proxy problems, specifically, given any feasible solution to P1. The constructed proxy function is defined as follows: ; in, , Let represent a feasible solution to the optimization problem P1, corresponding to the robot state at time t. This represents the robot's reference orientation at time t, which is related to the feasible solution. Corresponding orientation information, proxy function The following constraints need to be met: (1) Lower bound constraint: ; (2) Concave constraint: about It is a concave function; (3) Local consistency constraint: The following conditions must be met at the optimal solution: and ; To prove that the constructed surrogate function satisfies the lower bound constraint, concavity constraint, and local consistency constraint, then for the lower bound constraint (1), based on the inequality... right Established, resulting in: ; Since the logarithmic function is monotonically increasing, it can be proven that... Satisfy the lower bound constraint, and then pass the test. Proof of the positive semidefiniteness of the Hessian matrix Satisfying the concavity constraint, and finally through calculation and exist The fact that the function value and gradient value are equal at a given point confirms the surrogate function. It satisfies the local consistency constraint.

[0028] Based on the lower bound constraint, if a surrogate function is used near the feasible point... Replace the original function This allows us to obtain a lower bound estimate of the objective function. Therefore, when iteratively constructing a new surrogate function using the current solution as the initial point, we can gradually tighten the lower bound until the convergence condition is met. Therefore, assume the first The solution of the next iteration is... The original optimization problem P1 is in In the next iteration, it can be transformed into an optimization problem P2: ; ; in, Indicates the first The solution of the next iteration This represents a weighting factor used to adjust collision avoidance capabilities in different scenarios. , It is a penalty function used to automatically generate a non-uniformly distributed sequence of safe distances during the optimization process. This represents a sequence of safe distances with a time step from h=0 to h=H. At this point, the optimization problem P2 has been transformed into a convex optimization problem, which can be solved directly using the CVXPY tool in MATLAB.

[0029] S8: Obtain the optimal navigation path based on the local optimal solution of the control input, acquire the data collected by the intelligent agent device, and calculate the data acquisition amount and classification error value of multiple neural network classification models; In this embodiment, the mobile intelligent vehicle plans the optimal navigation path based on the obtained optimal control input, obtains the amount of data collected for training the neural network classification model, calculates the classification error value corresponding to the neural network classification model, and adjusts devices or parameters such as antennas. Thus, while ensuring safe arrival at the destination, it minimizes distance costs while maximizing robot communication efficiency and neural network classification model learning performance.

[0030] like Figure 3 As shown, electromagnetic maps generated by two sensors at different locations are displayed. The sensor locations are marked with a cross symbol. The horizontal and vertical axes represent the row and column indices of the spatial grid. The color bars next to the two electromagnetic maps are scales for signal strength. The brightness of the colors in the electromagnetic maps corresponds to the signal strength, reflecting the propagation loss of electromagnetic waves in space. The brighter the color, the stronger the signal at that location and the smaller the propagation loss of electromagnetic waves. The autonomous navigation scheme proposed in this invention is the CLD scheme. In simulation experiments, the CLD scheme is compared with the traditional PMM+Commu scheme, the Optimization-based Collision Avoidance (OBCA) scheme, the Regularized Dual Alternating (RDA) scheme, and the RDA+Commu scheme. Among them, the PMM+Commu scheme is a communication-aware motion planning method that simplifies and models each obstacle and robot as a point mass. The OBCA scheme is a traditional collision avoidance method based on model predictive control (MPC). The RDA scheme is an advanced MPC obstacle avoidance strategy that models obstacles as convex sets and achieves high-precision collision avoidance by accurately calculating the minimum distance between objects. The RDA+Commu scheme is a method that further introduces communication-aware factors on the basis of RDA. like Figure 4 As shown, the robot's autonomous navigation trajectories under different schemes are illustrated. The horizontal and vertical axes represent horizontal and vertical distances in space, respectively. Black circles and rectangles represent obstacles of different types and sizes. Two sensors are represented by black triangles. The center line spanning the vertical axis and parallel to the horizontal axis represents the reference path generated from the predetermined pathpoint list. The curve represents the robot's actual trajectory, and the rectangles or circles covering the curve represent the robot body. Figure 4 It can be seen that the autonomous navigation scheme proposed in this invention avoids all obstacles and successfully reaches the target position. It is worth noting that, in order to improve data acquisition efficiency, compared with other schemes, the autonomous navigation strategy of the CLD method prioritizes the path that is closer to sensor 1 and sensor 2. At the same time, the CLD method further optimizes the model training process by introducing a learning performance regularization term, so that the path selection can not only improve communication throughput, but also take into account the model training difficulty, and achieve adaptive path adjustment in a variable environment, taking into account obstacle avoidance, communication and learning performance. like Figure 5 , Figure 6 As shown, in the same indoor scene, the comparison results of communication data collection volume and classification error of the autonomous navigation scheme proposed in this invention and the traditional scheme for different classification models are obtained. The classification models include CNN and SVM. Since the PMM+Commu method cannot enable the robot to successfully reach the target point in a complex indoor environment and fails to complete the basic task, it is not included in the comparison of data collection volume and classification error. The traditional schemes are PMM+Commu, Optimization-based Collision Avoidance (OBCA), Regularized Dual Alternating (RDA), and RDA+Commu, respectively. like Figure 5 As shown, the data acquisition volume of each scheme at the two sensors is displayed. The horizontal axis represents different autonomous navigation schemes, and the vertical axis represents the communication throughput of the robot at the corresponding sensors. Figure 5 It can be found that the autonomous navigation CLD scheme proposed in this invention has a higher amount of communication data acquisition at both sensors than other schemes, indicating that the scheme of this invention can enable the robot to always navigate the path within the optimal signal area, effectively improving the data acquisition efficiency. like Figure 6 As shown, this diagram compares the classification errors of neural network classification models for robots under different path planning methods. The horizontal axis represents different autonomous navigation schemes, and the vertical axis represents the model's classification error. Figure 6 The simulation results show that the classification error values ​​of the autonomous navigation CLD scheme proposed in this invention are all smaller than those of other schemes, demonstrating the best performance in improving learning performance. This indicates that the scheme not only prioritizes areas with stronger signals when selecting paths, but also collects more high-quality data, thereby significantly improving the model's learning effect and training efficiency. The simulation results also prove that when faced with the dual requirements of communication and learning tasks, the robot can prioritize paths closer to the sensors without significantly deviating from the reference path, effectively enhancing communication throughput and model learning performance.

[0031] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A multi-agent data collection autonomous navigation method for an indoor scene, characterized in that, Includes the following steps: Construct a multi-agent edge intelligence system model, including edge servers, mobile intelligent vehicles, and sensors; Construct multiple neural network classification models and pre-train them; Obtain signal strength and propagation loss in indoor communication scenarios to construct a wireless electromagnetic map; The radio electromagnetic map is divided into multiple sub-maps based on the signal strength distribution, and the channel propagation gain mapping function for each sub-map is calculated. A high-order nonlinear fitting model of the signal propagation characteristics of the entire radio electromagnetic map is obtained by fitting the channel propagation gain mapping function of all sub-maps. The optimization objective is to minimize the squared error of the channel propagation gain mapping function and the high-order nonlinear fitting model. The optimal parameter set is obtained by solving this problem, and the sensor spectral efficiency is calculated, specifically including: The optimal parameter set of the high-order nonlinear fitting model is obtained by solving an optimization problem with an optimization objective of minimizing error squares of a channel propagation gain mapping function and the high-order nonlinear fitting model ; Optimal parameter set based on high-order nonlinear fitting model The sensor spectral efficiency is calculated using the channel propagation gain mapping function, and is expressed as: ； in, Indicates the first Each sensor in the time slot Spectral efficiency, This represents a high-order nonlinear fitting model. Indicates time The position of the mobile intelligent vehicle in the planar coordinate system. Indicates the first The location of each sensor For complex Gaussian noise power, Indicates the first The transmit power of each sensor, and These represent the propagation index and the gain coefficient, respectively. The optimization problem is constructed by calculating the weighted sum of the Euclidean distance between the mobile intelligent vehicle's own trajectory and the corresponding reference path point, the communication throughput cost, and the classification error of the neural network classification model. Minimizing this weighted sum is then used as the optimization objective. The problem is specifically expressed as follows: ； in, Indicates the current state. This indicates that the mobile intelligent vehicle is at any time The control input vector, , , Indicates hyperparameters, This represents the distance cost function, used to calculate its own trajectory. The sum of Euclidean distances between the reference path point and the corresponding reference path point. The communication throughput cost function is represented by the following: This represents the mean classification error of the model. This represents the number of discrete time slots within the field of view of the mobile intelligent vehicle. This represents the time offset within the prediction time domain. Represents a state evolution model. , , Represents the coefficient matrix. and These represent the minimum and maximum values ​​of the control input speed for the mobile intelligent vehicle, respectively. This represents the control input vector at the h-th time after the current time t. This represents the control input vector at the (h+1)th time after the current time t. and These represent the minimum and maximum values ​​of acceleration, respectively. This represents the collision avoidance function. express Time of the first The state vector of each obstacle This indicates the safe distance between the mobile intelligent vehicle and obstacles. Represents a set of obstacles; The distance cost function is specifically expressed as: ； in, Indicates a reference path point; The communication throughput cost function is specifically expressed as follows: ； in, Indicates the location of the mobile intelligent vehicle With the Each sensor in time Spectral efficiency during communication. Indicates the number of sensors; ； in, This represents the total number of neural network classification models. and Hyperparameters that characterize the learning difficulty of the model Indicates the relationship with the first A set of sensors associated with a neural network classification model. This indicates the bandwidth allocated to each sensor group. Indicates the duration of the time slot. Indicates the first The number of bits required for each sample in each model Indicates the first The number of historical samples accumulated by each model; The optimization problem is transformed into a convex problem, and the local optimum of the path planning control input is obtained by iterative solution. The optimal navigation path is obtained based on the local optimal solution, and the mobile intelligent vehicle travels along the optimal navigation path.

2. The multi-agent data acquisition and autonomous navigation method for indoor scenarios according to claim 1, characterized in that, Obtain signal strength and propagation loss in indoor communication scenarios, and construct a radio electromagnetic map, specifically including: Use the Blender tool to build an indoor communication scenario, and then import the built indoor communication scenario into the Sionna library; The Sionna library performs wireless communication simulations on imported indoor communication scenarios, calculates signal strength and propagation loss, and constructs a wireless electromagnetic map.

3. The multi-agent data acquisition and autonomous navigation method for indoor scenarios according to claim 1, characterized in that, The calculation of the channel propagation gain mapping function for each sub-map includes: The channel propagation gain mapping function for line-of-sight propagation is expressed as: ； Where RM represents the channel propagation gain mapping function, Indicates time The position of the mobile intelligent vehicle in the planar coordinate system. Indicates the first The location of each sensor Indicates path loss. This represents the path loss index. Denotes the Euclidean norm. Indicates the calculation position and location The Euclidean distance between them; The channel propagation gain mapping function for non-line-of-sight propagation is expressed as: ； in, This represents a correction factor that takes into account the propagation effects of radio signals. This represents the path loss index for non-line-of-sight propagation.

4. The multi-agent data acquisition and autonomous navigation method for indoor scenarios according to claim 1, characterized in that, A high-order nonlinear fitting model of the entire radio electromagnetic map signal propagation characteristics is obtained by fitting the channel propagation gain mapping function of all sub-maps, specifically expressed as: ； in, This represents a high-order nonlinear fitting model. Indicates time The position of the mobile intelligent vehicle in the planar coordinate system. Indicates the first The location of each sensor Indicates the first Gain coefficient of each sub-map Indicates the first The propagation index of a submap, where L represents the number of submaps. This indicates the region indication function, used to distinguish the current location of the mobile intelligent vehicle. Sub-map , This represents the Euclidean norm.

5. The multi-agent data acquisition and autonomous navigation method for indoor scenarios according to claim 1, characterized in that, It also includes collision avoidance optimization steps, specifically including: safe distance Replace with time-varying safe distance sequence ,in, Indicates the safety distance parameter; Collision avoidance constraints are expressed as: ； in, The time step variable represents the time-varying safety distance, corresponding to the minimum allowable distance between the moving intelligent vehicle and the obstacle at time t in the prediction time domain. This indicates the calculation of Euclidean distance. Indicates in Time of the first Geometry modeling of an obstacle Indicating that the mobile intelligent vehicle is in Geometric modeling at time step, and These represent the minimum distance and the maximum distance, respectively. A regularization penalty term based on the L1 norm is added to the optimization objective. ; The collision avoidance constraint is transformed into a linear dual form, expressed as: ； in, and Let T represent the dual variables of the geometric constraints of the mobile intelligent vehicle and the obstacle surface, respectively, with the superscript T indicating transpose. This represents a slack variable, used to convert inequality constraints into equality constraints. , , Represents the number of geometric surfaces on the obstacle's surface. Represents the set of real numbers. The dimension is A matrix, where each row corresponds to a geometric face of the obstacle. It is a dimension of A vector, where each element of the vector corresponds to a geometric surface of the obstacle. Indicates the location of the mobile intelligent vehicle The relevant translation matrix, , , This indicates the number of geometric faces on the surface of the mobile intelligent vehicle. It is a coefficient matrix that defines the boundary constraint directions of the robot's initial shape. It is a real number vector. This is a rotation matrix related to the orientation angle.

6. The multi-agent data acquisition and autonomous navigation method for indoor scenarios according to claim 1, characterized in that, Transforming the optimization problem into a convex problem specifically includes: A surrogate function for the spectral efficiency of communication between the mobile intelligent vehicle and sensors is constructed, transforming the optimization problem into a convex problem, expressed as: ； ； in, Indicates the current state. This indicates that the mobile intelligent vehicle is at any time The control input vector, and Let T represent the dual variables of the geometric constraints of the mobile intelligent vehicle and the obstacle surface, respectively, with the superscript T indicating transpose. Represents slack variables. , , Indicates hyperparameters, This represents the distance cost function, used to calculate its own trajectory. The sum of Euclidean distances between the reference path point and the corresponding reference path point. This represents the number of discrete time slots within the field of view of the mobile intelligent vehicle. This represents the time offset within the prediction time domain. This represents a proxy function. This represents the total number of neural network classification models. and Hyperparameters that characterize the learning difficulty of the model Indicates the relationship with the first A set of sensors associated with a neural network classification model. This indicates the bandwidth allocated to each sensor group. Indicates time slot, Indicates the first The number of bits required for each sample in each model Indicates the first The number of historical samples accumulated by each model This represents the regularization penalty term. The time step variable represents the time-varying safety distance. , , Represents the coefficient matrix. and These represent the minimum and maximum values ​​of the control input speed for the mobile intelligent vehicle, respectively. This represents the control input vector at the h-th time after the current time t. This represents the control input vector at the (h+1)th time after the current time t. and These represent the minimum and maximum values ​​of acceleration, respectively. and These represent the minimum distance and the maximum distance, respectively. , , Represents the number of geometric surfaces on the obstacle's surface. Represents the set of real numbers. The dimension is A matrix, where each row corresponds to a geometric face of the obstacle. It is a dimension of A vector, where each element of the vector corresponds to a geometric surface of the obstacle. Indicates the location of the mobile intelligent vehicle The relevant translation matrix, , , This indicates the number of geometric faces on the surface of the mobile intelligent vehicle. It is a coefficient matrix that defines the boundary constraint directions of the robot's initial shape. It is a real number vector. This is a rotation matrix related to the orientation angle.

7. The multi-agent data acquisition and autonomous navigation method for indoor scenarios according to claim 6, characterized in that, A surrogate function for constructing the spectral efficiency of communication between the mobile intelligent vehicle and sensors is expressed as: ; in, This represents a proxy function. Indicates the current state. Let L represent a feasible solution to the optimization problem, and let L represent the number of submaps. This indicates the region indication function, used to distinguish the current location of the mobile intelligent vehicle. Sub-map , Indicates the first Gain coefficient of each sub-map Indicates the first The propagation index of each sub-map For complex Gaussian noise power, Indicates the first The transmit power of each sensor, , This indicates the position of the mobile intelligent vehicle in a planar coordinate system. This indicates the reference orientation of the mobile intelligent vehicle at time t. This represents the Euclidean norm.

Images