Automatic walking and positioning method and system of signal fixed measurement robot in complex environment

Abstract

The invention provides an automatic walking and positioning method and system for a signal positioning and measuring robot in a complex environment, and relates to the field of technical positioning and measuring robots, and the method comprises the steps: obtaining an indoor three-dimensional map and multi-mode perception data, carrying out the semantic segmentation through a graph neural network, and achieving the environment matching through the combination of a point cloud registration technology; and generating robot motion track planning data. Multi-dimensional signal features are collected through a phased-array antenna system, and a signal feature distribution diagram is constructed in combination with a robot pose and an environment electromagnetic propagation model. A partitioning strategy and a multi-constraint optimization algorithm are adopted to calculate three-dimensional coordinates of a signal source, and finally, the signal source is accurately positioned through a depth calibration network.

Description

Technical Field

[0001] The invention relates to a signal measurement robot technology, and in particular to an automatic movement and positioning method and system for a signal measurement robot in a complex environment. Background Art

[0002] Traditional signal measurement robots have limited autonomous navigation capabilities in complex environments. They usually rely on single sensor data for environmental perception and cannot fully and accurately capture various obstacles and dynamic changes in complex environments, which makes them prone to collisions or blind spots in complex scenes.

[0003] Existing signal source positioning methods are not accurate enough to meet the needs of precise positioning in complex environments. Most methods only use single features such as signal strength or angle of arrival for positioning, and fail to fully utilize the multi-dimensional feature information of the signal, resulting in low positioning accuracy in complex indoor environments with severe multipath effects.

[0004] Existing technologies lack an effective multi-source information fusion mechanism. Signal measurement robots often separate functions such as environmental perception, autonomous navigation, and signal positioning, and fail to achieve information sharing and collaborative optimization among modules, making it difficult to achieve stable and reliable automatic movement and precise positioning in complex and changing environments. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for automatic movement and positioning of a signal measurement robot in a complex environment, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provided is a method for automatic movement and positioning of a signal measurement robot in a complex environment, including:

[0008] Acquire indoor three-dimensional map data, acquire environmental multimodal perception data through binocular vision sensors and three-dimensional laser radar scanning, perform feature extraction on the multimodal perception data to obtain environmental feature data, use graph neural network to perform semantic segmentation on the environmental feature data to obtain scene semantic information, combine point cloud registration technology to perform real-time matching of the scene semantic information with the three-dimensional map data to obtain environmental matching data, input the environmental matching data into a pre-trained trajectory planning model, integrate obstacle position prediction information and robot motion constraint parameters in the scene semantic information, and generate robot motion trajectory planning data;

[0009] The robot movement is controlled based on the robot motion trajectory planning data, broadband parallel acquisition of the measured signal is performed through the orthogonal polarization directional antenna array in the phased array antenna system to obtain multi-dimensional signal feature data including signal strength, arrival angle, phase difference and polarization characteristics, the multi-dimensional signal feature data and the robot real-time posture data and the preset environmental electromagnetic propagation model parameters are input into the tensor decomposition model to construct a signal feature distribution map, the signal feature distribution map is input into the deep neural network for feature extraction to obtain a signal feature vector, and the signal feature vector is processed based on the multi-source information fusion algorithm to obtain the signal source initial position data and signal emission characteristic data;

[0010] According to the initial position data of the signal source and the signal emission characteristic data, a zoning strategy is adopted to divide the measurement area, sampling grid points are constructed in the measurement area, the robot is controlled to perform signal measurement at the sampling grid points, a space-time two-dimensional adaptive beamforming algorithm is used to process the multi-dimensional signal feature data, a signal propagation model is constructed in combination with multipath effect compensation, the three-dimensional spatial coordinates of the signal source are calculated by a multi-constraint optimization algorithm, the three-dimensional spatial coordinates of the signal source are corrected by a particle filter algorithm, the corrected three-dimensional spatial coordinates of the signal source are fused and calibrated with the indoor three-dimensional map data based on a deep calibration network, and the final signal source position coordinates are output.

[0011] In an optional embodiment,

[0012] The environment feature data is semantically segmented using a graph neural network to obtain scene semantic information, and the scene semantic information is matched with the three-dimensional map data in real time using point cloud registration technology to obtain environment matching data, including:

[0013] Adaptive sampling is used to calculate the local feature entropy of each point in the environmental feature data, a sampling density threshold is set based on the local feature entropy to perform downsampling to obtain a feature point sequence, a multi-scale covariance matrix of the feature point sequence is constructed to extract a geometric feature vector, the geometric feature vector is fused with a depth feature vector to generate a vertex feature vector, a dynamic threshold K nearest neighbor algorithm is used to construct a graph structure network, and the spatial propagation characteristics and geometric similarity of the connecting edges in the graph structure network are calculated to obtain an edge feature vector;

[0014] Input the vertex feature vector and the edge feature vector into a multi-head cross attention network, obtain the attention weight coefficient through feature channel recalibration and spatial position encoding, adaptively aggregate the features in the graph structure network based on the attention weight coefficient to obtain local fusion features, input the local fusion features into the residual feature enhancement network to extract discriminative features, perform spectral domain transformation on the discriminative features and input them into the dynamic graph convolution network to obtain scene semantic information;

[0015] Based on the scene semantic information, a similarity measurement model with semantic consistency constraints is constructed, reliable matching point pairs are screened through bidirectional cross-validation, the matching point pairs are probabilistically sampled and the transformation matrix is ​​calculated in combination with the local geometric structure consistency constraints, and the optimal initial transformation matrix is ​​selected using a robustness evaluation criterion;

[0016] A hierarchical KD tree is used to perform local feature matching on the area covered by the optimal initial transformation matrix to obtain corresponding point pairs, and a Mahalanobis distance metric space that considers the semantic similarity between point pairs is constructed. The transformation matrix is ​​iteratively optimized in the metric space, and the registration accuracy and scene consistency of the optimized transformation matrix are evaluated in real time. When the evaluation index meets the convergence condition, the environment matching data is output.

[0017] In an optional embodiment,

[0018] A hierarchical KD tree is used to perform local feature matching on the area covered by the optimal initial transformation matrix to obtain corresponding point pairs, and a Mahalanobis distance metric space considering the semantic similarity between point pairs is constructed. The transformation matrix is ​​iteratively optimized in the metric space, and the registration accuracy and scene consistency of the optimized transformation matrix are evaluated in real time. When the evaluation index meets the convergence condition, the output environment matching data includes:

[0019] Performing feature extraction on the point cloud data within the coverage area of ​​the optimal initial transformation matrix, extracting local geometric structure features and semantic label features of the point cloud data, and performing feature fusion on the local geometric structure features and the semantic label features to obtain a multi-dimensional feature vector;

[0020] Calculating feature similarity for the multidimensional feature vector, constructing a feature similarity matrix, performing feature decomposition based on the feature similarity matrix to obtain feature main directions, using the feature main directions to perform spatial division on the point cloud data and construct a hierarchical KD tree;

[0021] Calculate the geometric information entropy and semantic information entropy of the hierarchical KD tree nodes, construct a node splitting cost function by combining the geometric information entropy, semantic information entropy and position constraint, determine the optimal splitting parameter based on the minimization result of the node splitting cost function, organize similar feature points into the same subtree to obtain a hierarchical feature index;

[0022] Calculating the correlation strength of feature points in the source point cloud and the target point cloud based on the hierarchical feature index, establishing a corresponding relationship between the feature points according to the correlation strength, and constructing edge connections according to the corresponding relationship to obtain a matching graph structure;

[0023] Iteratively updating node features and edge weights in the matching graph structure, calculating a reconstruction error of the node features, constructing a metric learning objective function based on the reconstruction error, and optimizing the metric learning objective function to obtain a discriminative metric matrix;

[0024] Calculating similarity scores of feature point pairs in the metric space defined by the discriminant metric matrix, constructing a transformation matrix optimization objective function by combining the similarity scores with local geometric constraints and global semantic constraints, and solving the transformation matrix optimization objective function to obtain an optimal transformation matrix;

[0025] Calculate the registration error of the optimal transformation matrix at multiple feature scales, combine the registration error with the semantic label consistency to obtain an environment matching evaluation score, output the final environment feature matching data when the environment matching evaluation score meets the preset convergence condition, and when the environment matching evaluation score does not meet the preset convergence condition, optimize and update the feature extraction parameters, feature matching parameters and metric calculation parameters according to the environment matching evaluation score, and return to the feature extraction step to continue iterative optimization.

[0026] In an optional embodiment,

[0027] Inputting the environment matching data into a pre-trained trajectory planning model, fusing the obstacle position prediction information and robot motion constraint parameters in the scene semantic information, and generating robot motion trajectory planning data includes:

[0028] Mapping the environment matching data to the grid map to obtain an initial grid map, calculating the logarithmic value of the obstacle occupancy probability in the initial grid map to construct a static cost map, extracting the dynamic obstacle position information to perform Gaussian kernel density calculation to construct a dynamic cost map, converting the grid point semantic labels based on the semantic cost mapping function to construct a semantic cost map, and performing weighted combination of the static cost map, the dynamic cost map and the semantic cost map to obtain an environmental comprehensive cost map;

[0029] Based on the comprehensive cost map of the environment, a state vector is constructed, wherein the state vector includes the robot position coordinates, motion speed, target position coordinates and local cost information, the state vector is input into the policy neural network to obtain the action parameters of linear velocity and angular velocity, the state-action value function is calculated according to the action parameters, a reward function including a target distance term, an obstacle avoidance term and a trajectory smoothing term is constructed based on the state-action value function, the parameters of the policy neural network are optimized by using a proximal policy algorithm, and the optimized action parameters are subjected to kinematic constraint processing to obtain a planned trajectory;

[0030] The planned trajectory is interpolated with a cubic spline function to obtain a smooth trajectory, an elastic potential field is constructed based on the smooth trajectory, the trajectory curvature is set as an internal potential energy function, the obstacle repulsion force in the environmental comprehensive cost map is set as an external potential energy function, the weighted sum of the internal potential energy function and the external potential energy function is minimized based on the gradient descent method to obtain an optimized trajectory, the optimized trajectory is kinematically verified, and the final robot motion trajectory planning data is output.

[0031] In an optional embodiment,

[0032] Acquiring multidimensional signal feature data including signal strength, arrival angle, phase difference and polarization characteristics, inputting the multidimensional signal feature data, the robot's real-time posture data and preset environmental electromagnetic propagation model parameters into a tensor decomposition model, and constructing a signal feature distribution map includes:

[0033] The power spectral density of the received signal is estimated by using a periodogram averaging method and signal strength data is obtained by constant false alarm threshold detection. An array manifold is constructed based on the antenna directional diagram to locate the array response and a multiple signal classification algorithm is used to calculate the spatial spectrum to obtain the angle of arrival data. Phase difference data is extracted by multi-channel coherent sampling and phase unwrapping processing. Polarization matrix is ​​reconstructed based on dual orthogonal polarization channels to obtain polarization characteristic data. The signal strength data, angle of arrival data, phase difference data and polarization characteristic data are combined to form multi-dimensional signal feature data.

[0034] According to the multi-dimensional signal characteristic data, an environmental electromagnetic propagation model including distance attenuation, atmospheric absorption, reflection diffraction and frequency shift is established, terrain shielding and building shielding losses are calculated by ray tracing method, and delay spread characteristics are analyzed based on power delay spectrum to obtain environmental electromagnetic propagation model parameters;

[0035] Rearranging the multidimensional signal feature data and the environmental electromagnetic propagation model parameters according to the signal feature dimension, spatial dimension, time dimension and posture dimension to construct a feature tensor, performing tensor decomposition on the feature tensor to obtain a core tensor and a factor matrix, optimizing the core tensor and the factor matrix using an alternating least squares method, introducing regularization constraints to achieve iterative convergence, and obtaining an optimized core tensor and factor matrix;

[0036] The optimized core tensor and factor matrix are modally projected to obtain compressed feature data, and the compressed feature data are processed by grid resampling and interpolation smoothing to generate an initial feature distribution map. The noise in the initial feature distribution map is suppressed by singular value threshold processing and norm regularization method, and the boundary features of the initial feature distribution map are optimized in combination with variational regularization and gradient preserving filtering method, and the target feature distribution map is output.

[0037] In an optional embodiment,

[0038] The space-time two-dimensional adaptive beamforming algorithm is used to process multi-dimensional signal feature data, and the signal propagation model is constructed in combination with multipath effect compensation. The three-dimensional spatial coordinates of the signal source are calculated through a multi-constraint optimization algorithm, including:

[0039] Calculate the time-frequency spectrum of multidimensional signal feature data to obtain a frequency domain feature matrix, perform singular value decomposition on the frequency domain feature matrix to extract principal component feature vectors, construct a two-dimensional covariance matrix containing spatial correlation and time correlation, perform feature decomposition on the two-dimensional covariance matrix using regularization constraints to obtain a signal subspace and a noise subspace, calculate a projection operator based on the orthogonality of the signal subspace and the noise subspace, calculate the expected signal steering vector in combination with a wave direction estimation algorithm, and use a recursive least squares criterion to optimize and obtain an optimal weight vector;

[0040] Establish an indoor environment model including the geometric characteristics of the reflecting surface and the dielectric characteristics of the material, use a layered ray tracing method to identify the direct path and the multipath propagation path, calculate the reflection point coordinate sequence and the diffraction edge position sequence of each propagation path in combination with the optimal weight vector, calculate the path loss based on the reflection coefficient and the unified diffraction theory, and construct a multipath propagation channel model considering frequency selective fading based on the path loss, path delay and phase offset;

[0041] Constructing a multi-constraint optimization objective function based on the multipath propagation channel model, wherein the multi-constraint optimization objective function includes an observation matrix constraint term, a signal power constraint term, a propagation delay constraint term, and a carrier phase constraint term, introducing an adaptive weight coefficient for each constraint term in the multi-constraint optimization objective function, and using a variable splitting strategy to decompose the multi-constraint optimization objective function into a sub-problem sequence;

[0042] The sub-problem sequence is constructed as an augmented Lagrangian function, wherein the augmented Lagrangian function includes a main variable, an auxiliary variable and a dual variable, the main variable is iteratively updated by a proximal gradient descent algorithm, the auxiliary variable is iteratively updated by a soft threshold operation, the dual variable is iteratively updated in combination with a dynamic step size strategy, and a preprocessing matrix is ​​constructed by online estimating curvature to improve the convergence speed;

[0043] The three-dimensional spatial coordinates of the initial signal source are obtained according to the iterative update result of the augmented Lagrangian function, and an error evaluation function including geometric consistency and energy consistency is constructed. When the calculation result of the error evaluation function is less than a preset threshold, the final three-dimensional spatial coordinates of the signal source are output.

[0044] In an optional embodiment,

[0045] The calculation formula of the augmented Lagrangian function is as follows:

[0046]

[0047] Among them, L rd (·) represents the augmented Lagrangian function, x represents the main variable, y represents the auxiliary variable, λ represents the dual variable, β represents the step size adjustment factor, f(x) represents the objective function, gi (x) represents the i-th constraint, and ρ represents the augmentation factor.

[0048] According to a second aspect of the embodiments of the present invention,

[0049] Provides an automatic walking and positioning system for a signal measurement robot in a complex environment, including:

[0050] The first unit is used to obtain indoor three-dimensional map data, obtain environmental multimodal perception data through binocular vision sensors and three-dimensional laser radar scanning, perform feature extraction on the multimodal perception data to obtain environmental feature data, use graph neural network to perform semantic segmentation on the environmental feature data to obtain scene semantic information, combine point cloud registration technology to perform real-time matching of the scene semantic information with the three-dimensional map data to obtain environmental matching data, input the environmental matching data into a pre-trained trajectory planning model, integrate obstacle position prediction information and robot motion constraint parameters in the scene semantic information, and generate robot motion trajectory planning data;

[0051] The second unit is used to control the movement of the robot based on the robot motion trajectory planning data, perform broadband parallel acquisition of the measured signal through the orthogonal polarization directional antenna array in the phased array antenna system, obtain multi-dimensional signal feature data including signal strength, arrival angle, phase difference and polarization characteristics, input the multi-dimensional signal feature data and the robot's real-time posture data and preset environmental electromagnetic propagation model parameters into a tensor decomposition model, construct a signal feature distribution map, input the signal feature distribution map into a deep neural network for feature extraction to obtain a signal feature vector, and process the signal feature vector based on a multi-source information fusion algorithm to obtain signal source initial position data and signal emission characteristic data;

[0052] The third unit is used to divide the measurement area by adopting a partitioning strategy according to the initial position data of the signal source and the signal emission characteristic data, construct sampling grid points in the measurement area, control the robot to perform signal measurement at the sampling grid points, process the multi-dimensional signal feature data by adopting a space-time two-dimensional adaptive beamforming algorithm, construct a signal propagation model in combination with multipath effect compensation, calculate the three-dimensional spatial coordinates of the signal source by a multi-constraint optimization algorithm, correct the three-dimensional spatial coordinates of the signal source by adopting a particle filter algorithm, fuse and calibrate the corrected three-dimensional spatial coordinates of the signal source with the indoor three-dimensional map data based on a deep calibration network, and output the final signal source position coordinates.

[0053] According to a third aspect of the embodiments of the present invention,

[0054] An electronic device is provided, comprising:

[0055] processor;

[0056] a memory for storing processor-executable instructions;

[0057] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0058] According to a fourth aspect of the embodiments of the present invention,

[0059] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0060] In this embodiment, accurate semantic segmentation and real-time matching of complex indoor environments are achieved by fusing multi-sensor data and advanced graph neural network algorithms. Combined with pre-trained trajectory planning models and obstacle prediction, the robot's autonomous navigation capability and positioning accuracy in complex environments are greatly improved, enabling it to complete signal determination tasks more intelligently and safely. The phased array antenna system and multi-dimensional signal feature extraction technology, combined with tensor decomposition models and deep neural networks, significantly improve the initial positioning accuracy of the signal source. Through the multi-source information fusion algorithm, the signal characteristics, robot posture and environmental electromagnetic propagation model are effectively integrated, laying the foundation for subsequent precise positioning. The present invention innovatively uses partitioning strategies and sampling grid point technology, combined with space-time two-dimensional adaptive beamforming algorithms and multipath effect compensation, to construct an accurate signal propagation model. Through the comprehensive application of multi-constraint optimization algorithms, particle filter algorithms and deep calibration networks, high-precision calculation and correction of the three-dimensional spatial coordinates of the signal source are achieved, and finally fused with indoor three-dimensional map data to output extremely high-precision signal source position coordinates. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a flow chart of the automatic walking and positioning method of the signal determination robot in a complex environment according to an embodiment of the present invention;

[0062] Figure 2 It is a schematic diagram of the structure of the automatic walking and positioning system of the signal determination robot in a complex environment according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0065] Figure 1 FIG. 1 is a flow chart of a method for automatically moving and positioning a signal determination robot in a complex environment according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0066] S101. Acquire indoor three-dimensional map data, acquire environmental multimodal perception data through binocular vision sensors and three-dimensional lidar scanning, perform feature extraction on the multimodal perception data to obtain environmental feature data, use graph neural network to perform semantic segmentation on the environmental feature data to obtain scene semantic information, combine point cloud registration technology to perform real-time matching of the scene semantic information with the three-dimensional map data to obtain environmental matching data, input the environmental matching data into a pre-trained trajectory planning model, integrate the obstacle position prediction information and robot motion constraint parameters in the scene semantic information, and generate robot motion trajectory planning data.

[0067] First, obtain indoor 3D map data. Use laser scanners and other equipment to perform 3D scanning and modeling of the indoor environment to obtain accurate 3D point cloud data of objects such as walls, doors, windows, and furniture. At the same time, use binocular vision sensors and 3D laser radars to scan the current environment in real time to obtain multimodal perception data such as RGB images and depth images.

[0068] The acquired multimodal perception data is subjected to feature extraction, including geometric features such as edges, corners, planes, and visual features such as color and texture. The extracted features are semantically segmented using a pre-trained graph neural network model, and objects in the environment are divided into different semantic categories such as walls, floors, tables and chairs to obtain scene semantic information.

[0069] Combined with point cloud registration technology, the scene semantic information obtained by segmentation is matched with the pre-acquired 3D map data in real time. Specifically, the optimal rigid body transformation between point clouds is calculated by the iterative closest point (ICP) algorithm to achieve precise alignment of semantic information and map, and obtain environmental matching data.

[0070] The environment matching data is input into the pre-trained trajectory planning model. The model uses a deep reinforcement learning method, with the environment matching data as the observation state and the reaching target position as the reward function for training. At the same time, the obstacle positions extracted from the scene semantic information and the robot's own kinematic and dynamic constraint parameters are integrated to comprehensively generate the robot's motion trajectory planning data.

[0071] In an optional embodiment,

[0072] The environment feature data is semantically segmented using a graph neural network to obtain scene semantic information, and the scene semantic information is matched with the three-dimensional map data in real time using point cloud registration technology to obtain environment matching data, including:

[0073] Adaptive sampling is used to calculate the local feature entropy of each point in the environmental feature data, a sampling density threshold is set based on the local feature entropy to perform downsampling to obtain a feature point sequence, a multi-scale covariance matrix of the feature point sequence is constructed to extract a geometric feature vector, the geometric feature vector is fused with a depth feature vector to generate a vertex feature vector, a dynamic threshold K nearest neighbor algorithm is used to construct a graph structure network, and the spatial propagation characteristics and geometric similarity of the connecting edges in the graph structure network are calculated to obtain an edge feature vector;

[0074] Input the vertex feature vector and the edge feature vector into a multi-head cross attention network, obtain the attention weight coefficient through feature channel recalibration and spatial position encoding, adaptively aggregate the features in the graph structure network based on the attention weight coefficient to obtain local fusion features, input the local fusion features into the residual feature enhancement network to extract discriminative features, perform spectral domain transformation on the discriminative features and input them into the dynamic graph convolution network to obtain scene semantic information;

[0075] Based on the scene semantic information, a similarity measurement model with semantic consistency constraints is constructed, reliable matching point pairs are screened through bidirectional cross-validation, the matching point pairs are probabilistically sampled and the transformation matrix is ​​calculated in combination with the local geometric structure consistency constraints, and the optimal initial transformation matrix is ​​selected using a robustness evaluation criterion;

[0076] A hierarchical KD tree is used to perform local feature matching on the area covered by the optimal initial transformation matrix to obtain corresponding point pairs, and a Mahalanobis distance metric space that considers the semantic similarity between point pairs is constructed. The transformation matrix is ​​iteratively optimized in the metric space, and the registration accuracy and scene consistency of the optimized transformation matrix are evaluated in real time. When the evaluation index meets the convergence condition, the environment matching data is output.

[0077] Exemplarily, first, the environmental feature data is adaptively sampled and feature extracted. The local feature entropy of each point is calculated, and the sampling density threshold is set according to the feature entropy for downsampling to obtain a feature point sequence. A multi-scale covariance matrix of the feature point sequence is constructed to extract the geometric feature vector. The geometric feature vector is fused with the depth feature vector to generate a vertex feature vector. The dynamic threshold K nearest neighbor algorithm is used to construct a graph structure network, and the spatial propagation characteristics and geometric similarity of the connecting edges are calculated to obtain the edge feature vector.

[0078] For example, for a point cloud data containing 1 million points, the local feature entropy of each point can be calculated, the sampling density threshold is set to 0.5, and 100,000 feature points are obtained by downsampling. Five scale covariance matrices are constructed for these feature points, and 128-dimensional geometric feature vectors are extracted. The geometric feature vector is concatenated with the 256-dimensional depth feature vector to obtain a 384-dimensional vertex feature vector. The dynamic threshold K-nearest neighbor algorithm with K=8 is used to construct the graph structure, and the 32-dimensional spatial propagation features and 32-dimensional geometric similarity features of the edge are calculated to obtain a 64-dimensional edge feature vector.

[0079] Next, the vertex feature vector and edge feature vector are input into the multi-head cross attention network. The attention weight coefficient is obtained by feature channel recalibration and spatial position encoding. Based on this coefficient, the features in the graph structure network are adaptively aggregated to obtain local fusion features. The local fusion features are input into the residual feature enhancement network to extract discriminative features. The discriminative features are transformed into the spectral domain and then input into the dynamic graph convolution network to finally obtain the scene semantic information.

[0080] Specifically, an 8-head cross attention mechanism can be used to interact with 384-dimensional vertex features and 64-dimensional edge features. The attention weight coefficient is obtained through linear projection and softmax operation, and the features are weighted aggregated to obtain 256-dimensional local fusion features. The local fusion features are enhanced using a 3-layer residual connection fully connected layer to obtain 128-dimensional discriminative features. The discriminative features are Fourier transformed and input into a 3-layer dynamic graph convolutional network to finally obtain the probability distribution of 20 categories of semantic labels for each point.

[0081] Then, a similarity measurement model with semantic consistency constraints is constructed based on scene semantic information. Reliable matching point pairs are screened through bidirectional cross-validation, matching point pairs are sampled probabilistically, and the transformation matrix is ​​calculated in combination with the local geometric structure consistency constraints. The optimal initial transformation matrix is ​​selected using the robustness evaluation criterion.

[0082] Finally, a hierarchical KD tree (K-Dimensional Tree) is used to perform local feature matching on the area covered by the optimal initial transformation matrix to obtain corresponding point pairs. A Mahalanobis distance metric space that considers the semantic similarity between point pairs is constructed, and the transformation matrix is ​​iteratively optimized in this space. The registration accuracy and scene consistency of the optimized transformation matrix are evaluated in real time, and the environment matching data is output when the evaluation index meets the convergence condition.

[0083] In this embodiment, efficient semantic segmentation of environmental feature data is achieved through graph neural networks, which improves the accuracy and robustness of scene understanding. Adaptive sampling and multi-scale feature extraction effectively retain key structural information, the multi-head attention mechanism enhances the expressiveness of features, and the dynamic graph convolutional network adapts to scene changes of different scales. Combining semantic information and geometric structure constraints for point cloud registration improves the accuracy and reliability of matching. Semantic consistency constraints effectively filter out false matches, probabilistic sampling and local geometric structure constraints improve the robustness of transformation matrix estimation, and hierarchical KD trees accelerate the local feature matching process. Iterative optimization is performed using the Mahalanobis distance metric space considering semantic similarity to improve registration accuracy and scene consistency. Real-time evaluation of registration quality and dynamic adjustment of optimization strategies ensure the accuracy and reliability of matching results. The overall method realizes efficient real-time matching of environmental feature data with three-dimensional maps.

[0084] In an optional embodiment,

[0085] A hierarchical KD tree is used to perform local feature matching on the area covered by the optimal initial transformation matrix to obtain corresponding point pairs, and a Mahalanobis distance metric space considering the semantic similarity between point pairs is constructed. The transformation matrix is ​​iteratively optimized in the metric space, and the registration accuracy and scene consistency of the optimized transformation matrix are evaluated in real time. When the evaluation index meets the convergence condition, the output environment matching data includes:

[0086] Performing feature extraction on the point cloud data within the coverage area of ​​the optimal initial transformation matrix, extracting local geometric structure features and semantic label features of the point cloud data, and performing feature fusion on the local geometric structure features and the semantic label features to obtain a multi-dimensional feature vector;

[0087] Calculating feature similarity for the multidimensional feature vector, constructing a feature similarity matrix, performing feature decomposition based on the feature similarity matrix to obtain feature main directions, using the feature main directions to perform spatial division on the point cloud data and construct a hierarchical KD tree;

[0088] Calculate the geometric information entropy and semantic information entropy of the hierarchical KD tree nodes, construct a node splitting cost function by combining the geometric information entropy, semantic information entropy and position constraint, determine the optimal splitting parameter based on the minimization result of the node splitting cost function, organize similar feature points into the same subtree to obtain a hierarchical feature index;

[0089] Calculating the correlation strength of feature points in the source point cloud and the target point cloud based on the hierarchical feature index, establishing a corresponding relationship between the feature points according to the correlation strength, and constructing edge connections according to the corresponding relationship to obtain a matching graph structure;

[0090] Iteratively updating node features and edge weights in the matching graph structure, calculating a reconstruction error of the node features, constructing a metric learning objective function based on the reconstruction error, and optimizing the metric learning objective function to obtain a discriminative metric matrix;

[0091] Calculating similarity scores of feature point pairs in the metric space defined by the discriminant metric matrix, constructing a transformation matrix optimization objective function by combining the similarity scores with local geometric constraints and global semantic constraints, and solving the transformation matrix optimization objective function to obtain an optimal transformation matrix;

[0092] Calculate the registration error of the optimal transformation matrix at multiple feature scales, combine the registration error with the semantic label consistency to obtain an environment matching evaluation score, output the final environment feature matching data when the environment matching evaluation score meets the preset convergence condition, and when the environment matching evaluation score does not meet the preset convergence condition, optimize and update the feature extraction parameters, feature matching parameters and metric calculation parameters according to the environment matching evaluation score, and return to the feature extraction step to continue iterative optimization.

[0093] Exemplarily, first, feature extraction is performed on the point cloud data within the area covered by the optimal initial transformation matrix. Local geometric structure features of the point cloud data, such as normal vectors, curvature, etc., are extracted; semantic label features, such as the semantic category to which the point belongs, are extracted. The local geometric structure features are fused with the semantic label features to obtain a multidimensional feature vector. For example, for a point, its feature vector can be [x, y, z, nx, ny, nz, c, label], where x, y, z are coordinates, nx, ny, nz are normal vectors, c is curvature, and label is a semantic label.

[0094] Then, the feature similarity is calculated for the multi-dimensional feature vectors and a feature similarity matrix is ​​constructed. The similarity between feature vectors can be calculated using metrics such as Euclidean distance or cosine similarity. Feature decomposition is performed based on the feature similarity matrix to obtain the feature main direction. The feature main direction is used to spatially divide the point cloud data and construct a hierarchical KD tree.

[0095] Next, the geometric information entropy and semantic information entropy of the hierarchical KD tree nodes are calculated. The geometric information entropy can characterize the geometric distribution of points within the node, and the semantic information entropy can characterize the semantic distribution of points within the node. The node splitting cost function is constructed by combining the geometric information entropy, semantic information entropy and position constraints. The optimal splitting parameter is determined based on the minimization result of the node splitting cost function, and similar feature points are organized into the same subtree to obtain a hierarchical feature index.

[0096] The correlation strength of feature points in the source point cloud and the target point cloud is calculated based on the hierarchical feature index. The correlation strength can be calculated using metrics such as the Euclidean distance or cosine similarity of feature vectors. The corresponding relationship between feature points is established based on the correlation strength, and the edge connection is constructed based on the corresponding relationship to obtain the matching graph structure.

[0097] Iteratively update node features and edge weights in the matching graph structure. Calculate the reconstruction error of node features and construct a metric learning objective function based on the reconstruction error. Optimize the metric learning objective function to obtain a discriminative metric matrix.

[0098] Calculate the similarity score of feature point pairs in the metric space defined by the discriminative metric matrix. Combine the similarity score with local geometric constraints and global semantic constraints to construct the transformation matrix optimization objective function. Solve the transformation matrix optimization objective function to obtain the optimal transformation matrix.

[0099] Calculate the registration error of the optimal transformation matrix at multiple feature scales. Combine the registration error with the semantic label consistency to obtain the environment matching evaluation score. When the environment matching evaluation score meets the preset convergence condition, output the final environment feature matching data. When the environment matching evaluation score does not meet the preset convergence condition, optimize and update the feature extraction parameters, feature matching parameters and metric calculation parameters according to the environment matching evaluation score, and return to the feature extraction step to continue iterative optimization.

[0100] For example, for an indoor scene point cloud, first extract the feature vector [x, y, z, nx, ny, nz, c, label] of each point. Then calculate the Euclidean distance between the feature vectors and construct the feature similarity matrix. Perform feature decomposition on the similarity matrix to obtain the feature main directions [1, 0, 0], [0, 1, 0], [0, 0, 1]. Construct a KD tree based on the feature main directions and calculate the geometric information entropy and semantic information entropy of each node. Set the node splitting cost function to the weighted sum of the geometric information entropy and the semantic information entropy. Iterate and optimize the splitting parameters to obtain a hierarchical feature index.

[0101] Based on the feature index, the Euclidean distance between the feature points in the source point cloud and the target point cloud is calculated to establish the corresponding relationship. The matching graph structure is constructed, and the node features and edge weights are updated iteratively. The metric learning objective function is optimized to obtain the discriminative metric matrix. The similarity score of the feature point pairs is calculated in the new metric space. The transformation matrix is ​​optimized by combining the similarity score with geometric constraints and semantic constraints.

[0102] Calculate the registration error of the optimized transformation matrix at different feature scales, such as point-to-point distance, normal vector consistency, etc. Combined with the semantic label consistency, the environment matching evaluation score is obtained. If the evaluation score does not meet the convergence condition, update the parameters of feature extraction, matching, and metric calculation, and return to continue iterative optimization.

[0103] In this embodiment, by constructing a hierarchical KD tree for local feature matching, the efficiency and accuracy of feature matching are improved. The hierarchical structure can quickly locate similar features and reduce invalid matching. At the same time, multi-scale features are taken into account to enhance the robustness of matching. The introduction of semantic information for feature fusion and constraint improves the feature expression ability and matching accuracy. Semantic information can distinguish objects of different categories and reduce semantically inconsistent erroneous matching. At the same time, semantic constraints can ensure the semantic consistency of the scene. An iterative optimization strategy is adopted to evaluate the registration accuracy and scene consistency in real time, and the parameters are adaptively adjusted to improve the accuracy and robustness of environment matching. Through multiple iterative optimizations, the matching results can be continuously improved to adapt to changes in different scenes.

[0104] S102. Control the movement of the robot based on the robot motion trajectory planning data, perform broadband parallel acquisition of the test signal through the orthogonal polarization directional antenna array in the phased array antenna system, obtain multi-dimensional signal feature data including signal strength, arrival angle, phase difference and polarization characteristics, input the multi-dimensional signal feature data and the robot's real-time posture data and preset environmental electromagnetic propagation model parameters into the tensor decomposition model, construct a signal feature distribution map, input the signal feature distribution map into a deep neural network for feature extraction to obtain a signal feature vector, and process the signal feature vector based on a multi-source information fusion algorithm to obtain signal source initial position data and signal emission characteristic data.

[0105] The robot is controlled to move based on the generated trajectory planning data. The robot collects the signals to be tested in the environment through the phased array antenna system. The antenna system contains a 16x16 array of orthogonal polarization directional antenna units, which can achieve broadband parallel reception in the 300MHz-3GHz frequency band. Through digital beamforming technology, the system can simultaneously obtain multi-dimensional feature data such as signal strength, arrival angle, phase difference and polarization characteristics.

[0106] The collected multi-dimensional signal feature data is combined with the robot's real-time posture data and the preset environmental electromagnetic propagation model parameters, and then input into the tensor decomposition model for processing. The model uses the Tucker decomposition method to decompose high-dimensional data into the product of a low-dimensional core tensor and a factor matrix, effectively extracting the time, space, and frequency characteristics of the signal. A signal feature distribution map is constructed based on the decomposition results to intuitively reflect the distribution of the signal in space.

[0107] The constructed signal feature distribution map is input into the pre-trained deep convolutional neural network to extract high-dimensional feature vectors. The network consists of multiple layers of convolution, pooling and fully connected layers, which can effectively capture the spatial correlation of the signal. The extracted feature vectors are then processed by a multi-source information fusion algorithm, which combines Bayesian estimation and DS evidence theory to fuse heterogeneous data from different sensors, and finally obtain the initial position estimation and emission characteristic parameters of the signal source.

[0108] In an optional embodiment,

[0109] Inputting the environment matching data into a pre-trained trajectory planning model, fusing the obstacle position prediction information and robot motion constraint parameters in the scene semantic information, and generating robot motion trajectory planning data includes:

[0110] Mapping the environment matching data to the grid map to obtain an initial grid map, calculating the logarithmic value of the obstacle occupancy probability in the initial grid map to construct a static cost map, extracting the dynamic obstacle position information to perform Gaussian kernel density calculation to construct a dynamic cost map, converting the grid point semantic labels based on the semantic cost mapping function to construct a semantic cost map, and performing weighted combination of the static cost map, the dynamic cost map and the semantic cost map to obtain an environmental comprehensive cost map;

[0111] Based on the comprehensive cost map of the environment, a state vector is constructed, wherein the state vector includes the robot position coordinates, motion speed, target position coordinates and local cost information, the state vector is input into the policy neural network to obtain the action parameters of linear velocity and angular velocity, the state-action value function is calculated according to the action parameters, a reward function including a target distance term, an obstacle avoidance term and a trajectory smoothing term is constructed based on the state-action value function, the parameters of the policy neural network are optimized by using a proximal policy algorithm, and the optimized action parameters are subjected to kinematic constraint processing to obtain a planned trajectory;

[0112] The planned trajectory is interpolated with a cubic spline function to obtain a smooth trajectory, an elastic potential field is constructed based on the smooth trajectory, the trajectory curvature is set as an internal potential energy function, the obstacle repulsion force in the environmental comprehensive cost map is set as an external potential energy function, the weighted sum of the internal potential energy function and the external potential energy function is minimized based on the gradient descent method to obtain an optimized trajectory, the optimized trajectory is kinematically verified, and the final robot motion trajectory planning data is output.

[0113] Exemplarily, first, the environment matching data is input into the pre-trained trajectory planning model. The environment matching data includes three-dimensional point cloud data, RGB image data, etc. of the current environment. The pre-trained trajectory planning model is trained based on the deep reinforcement learning method and can generate the robot's motion trajectory according to the input environment data.

[0114] Next, the environment matching data is mapped to the grid map to obtain the initial grid map. Specifically, the 3D point cloud data can be projected onto a 2D plane, and the plane can be divided into several grid cells according to the set grid resolution (such as 0.1m). The state of each grid cell can be occupied, idle or unknown. For example, for a 10m×10m environment, a 100×100 grid map can be obtained.

[0115] Then, the logarithm of the obstacle occupancy probability in the initial grid map is calculated to construct a static cost map. The inverse distance weighted method can be used to calculate the obstacle occupancy probability of each grid, and the logarithm is taken to obtain the cost value. For example, if the occupancy probability of a grid is 0.8, then its cost value is log(0.8)≈-0.223.

[0116] Next, the dynamic obstacle location information is extracted, and the Gaussian kernel density is calculated to construct a dynamic cost map. The target detection algorithm can be used to identify dynamic obstacles (such as pedestrians, vehicles, etc.) from the RGB image and map their locations to the grid map. With the dynamic obstacle location as the center, the two-dimensional Gaussian distribution is calculated to obtain the dynamic cost map. For example, for a pedestrian at (50, 50), its influence range may be the surrounding 5×5 grid area.

[0117] Subsequently, the semantic labels of the grid points are converted based on the semantic cost mapping function to construct a semantic cost map. The RGB image can be processed using a semantic segmentation algorithm to obtain the semantic label of each pixel (such as road, sidewalk, green belt, etc.). The semantic label is then mapped to the grid map, and the corresponding cost value is assigned to each grid according to the predefined semantic cost mapping relationship. For example, the cost value of the road area can be set to 0.1, and the cost value of the sidewalk area can be set to 0.5.

[0118] The static cost map, dynamic cost map and semantic cost map are weighted and combined to obtain the comprehensive cost map of the environment. A linear weighting method can be used, such as 0.5×static cost map+0.3×dynamic cost map+0.2×semantic cost map. In this way, a cost map that comprehensively considers static obstacles, dynamic obstacles and environmental semantic information can be obtained.

[0119] The state vector is constructed based on the comprehensive cost graph of the environment. The state vector contains the robot position coordinates (x, y), motion speed (v, ω), target position coordinates (xg, yg) and local cost information. The local cost information can be the cost value within a certain range around the robot (such as 3m×3m). For example, the state vector can be expressed as (2, 3, 0.5, 0.1, 8, 9, [0.1, 0.2, 0.3, ...]).

[0120] The state vector is input into the policy neural network to obtain the action parameters of linear velocity and angular velocity. The policy neural network can adopt a multi-layer perceptron structure, where the number of nodes in the input layer is the same as the dimension of the state vector, and the output layer has 2 nodes corresponding to the linear velocity and angular velocity respectively. For example, when the above state vector is input, the output may be (0.8, -0.2), indicating that the linear velocity is 0.8m / s and the angular velocity is -0.2rad / s.

[0121] The state-action value function is calculated based on the action parameters. The temporal difference learning method can be used to calculate the Q value using the current state, action, next state and reward. The reward function is constructed based on the state-action value function, which includes the target distance term, obstacle avoidance term and trajectory smoothing term. The target distance term can be the negative value of the Euclidean distance between the robot and the target point; the obstacle avoidance term can be the inverse of the distance between the robot and the nearest obstacle; the trajectory smoothing term can be the square of the change in angular velocity between two adjacent moments. These three terms can be weighted summed to obtain the total reward value.

[0122] The proximal strategy algorithm is used to optimize the parameters of the strategy neural network. Specifically, a batch of trajectory data can be obtained through multiple sampling, the objective function value can be calculated and the gradient can be obtained, and then the neural network parameters can be updated using the gradient descent method. During the optimization process, it is necessary to ensure that the KL divergence of the new and old strategies does not exceed the preset threshold (such as 0.01).

[0123] The optimized motion parameters are processed with kinematic constraints to obtain the planned trajectory. It is necessary to consider the robot's maximum linear velocity, maximum angular velocity, maximum acceleration and other constraints, and cut the motions that exceed the constraint range. For example, if the maximum linear velocity is 1.5m / s, the speed greater than 1.5m / s will be limited to 1.5m / s.

[0124] The planned trajectory is interpolated using a cubic spline function to obtain a smooth trajectory. The key points in the planned trajectory can be selected as control points, and a continuous and smooth trajectory curve can be generated using the cubic spline interpolation algorithm. For example, for 5 control points, a spline curve composed of 4 cubic polynomials can be obtained.

[0125] The elastic potential field is constructed based on the smooth trajectory. The trajectory curve can be regarded as an elastic band, and its internal potential energy function is defined as the square integral of the curvature. The external potential energy function can be defined based on the obstacle repulsion force in the environmental comprehensive cost map, for example, a repulsion force potential field in the form of 1 / r, where r is the distance to the obstacle.

[0126] The optimized trajectory is obtained by minimizing the weighted sum of the internal potential energy function and the external potential energy function based on the gradient descent method. The position of each point on the trajectory can be iteratively adjusted to make the total potential energy reach a local minimum. The optimization process needs to balance the internal potential energy (keeping the trajectory smooth) and the external potential energy (avoiding obstacles).

[0127] Finally, the optimized trajectory is kinematically verified to ensure that it meets the robot's motion constraints. You can check whether the curvature, speed, and acceleration of the trajectory are within the allowable range. If they are out of range, appropriate adjustments need to be made. After verification, the final robot motion trajectory planning data can be output.

[0128] In this embodiment, by fusing multi-source information to construct an environmental comprehensive cost map, factors such as static obstacles, dynamic obstacles, and environmental semantics are fully considered, thereby improving the safety and adaptability of trajectory planning. The deep reinforcement learning-based method is used for trajectory generation and optimization, which can adaptively learn the optimal decision-making strategy in complex environments, improving the generalization ability and robustness of the algorithm. The trajectory is post-processed and optimized in combination with the elastic potential field model, effectively avoiding obstacles while ensuring the smoothness of the trajectory, achieving local fine adjustment of the trajectory, and improving the quality and executability of the planning results.

[0129] In an optional embodiment,

[0130] Acquiring multidimensional signal feature data including signal strength, arrival angle, phase difference and polarization characteristics, inputting the multidimensional signal feature data, the robot's real-time posture data and preset environmental electromagnetic propagation model parameters into a tensor decomposition model, and constructing a signal feature distribution map includes:

[0131] The power spectral density of the received signal is estimated by using a periodogram averaging method and signal strength data is obtained by constant false alarm threshold detection. An array manifold is constructed based on the antenna directional diagram to locate the array response and a multiple signal classification algorithm is used to calculate the spatial spectrum to obtain the angle of arrival data. Phase difference data is extracted by multi-channel coherent sampling and phase unwrapping processing. Polarization matrix is ​​reconstructed based on dual orthogonal polarization channels to obtain polarization characteristic data. The signal strength data, angle of arrival data, phase difference data and polarization characteristic data are combined to form multi-dimensional signal feature data.

[0132] According to the multi-dimensional signal characteristic data, an environmental electromagnetic propagation model including distance attenuation, atmospheric absorption, reflection diffraction and frequency shift is established, terrain shielding and building shielding losses are calculated by ray tracing method, and delay spread characteristics are analyzed based on power delay spectrum to obtain environmental electromagnetic propagation model parameters;

[0133] Rearranging the multidimensional signal feature data and the environmental electromagnetic propagation model parameters according to the signal feature dimension, spatial dimension, time dimension and posture dimension to construct a feature tensor, performing tensor decomposition on the feature tensor to obtain a core tensor and a factor matrix, optimizing the core tensor and the factor matrix using an alternating least squares method, introducing regularization constraints to achieve iterative convergence, and obtaining an optimized core tensor and factor matrix;

[0134] The optimized core tensor and factor matrix are modally projected to obtain compressed feature data, and the compressed feature data are processed by grid resampling and interpolation smoothing to generate an initial feature distribution map. The noise in the initial feature distribution map is suppressed by singular value threshold processing and norm regularization method, and the boundary features of the initial feature distribution map are optimized in combination with variational regularization and gradient preserving filtering method, and the target feature distribution map is output.

[0135] Exemplarily, first, the power spectral density of the received signal is estimated by the periodogram averaging method. Specifically, the received signal is segmented, each segment of the signal is subjected to a fast Fourier transform to obtain a periodogram, and then a plurality of periodograms are averaged to obtain a smooth power spectral density estimate. For example, a 1-second signal can be divided into 10 segments, each of which is 0.1 seconds, and the periodograms are calculated and averaged. Then, the signal strength data is obtained by constant false alarm threshold detection. A fixed false alarm probability is set, such as 0.01, and the corresponding detection threshold is calculated according to the noise power spectral density. The frequency point above the threshold is a valid signal, and its amplitude is recorded as the signal strength data.

[0136] Then, the array response is mapped based on the antenna directional diagram to construct the array manifold. First, the gain of each antenna unit in different directions is measured to obtain the antenna directional diagram. Taking the 8-element linear array as an example, the gain of each antenna unit is measured every 5° in the 0°-360° direction. According to the antenna unit position and directional diagram, the array manifold matrix is ​​constructed. The spatial spectrum is calculated using the multiple signal classification algorithm to obtain the signal arrival angle data. Specifically, the covariance matrix of the received signal is eigen-decomposed, and the orthogonality of the signal subspace and the noise subspace is used to construct the pseudo-spectral function and search for the peak to obtain the signal arrival angle.

[0137] Next, the phase difference data is extracted through multi-channel coherent sampling and phase unwrapping. Multiple receiving channels are sampled simultaneously and the phase difference between channels is calculated. Since the phase measurement range is limited to -π to π, phase unwrapping is required. The phase difference accumulation method can be used to adjust ±2π when the phase difference between adjacent sampling points exceeds π.

[0138] Polarization characteristic data is obtained by reconstructing the polarization matrix based on dual orthogonal polarization channels. Signals in two orthogonal polarization directions are received and a 2×2 polarization matrix is ​​constructed. The polarization state of the signal is obtained by eigenvalue decomposition of the polarization matrix, including characteristics such as polarization type (linear / circular polarization / elliptical polarization) and polarization angle.

[0139] The signal strength data, arrival angle data, phase difference data and polarization characteristic data are combined to form multi-dimensional signal characteristic data. For example, a 5-dimensional data structure can be constructed, corresponding to the signal strength, azimuth and elevation angles of the arrival angle, phase difference, and the major axis direction and ellipticity of the polarization ellipse.

[0140] Based on the acquired multi-dimensional signal characteristic data, an environmental electromagnetic propagation model including distance attenuation, atmospheric absorption, reflection diffraction and frequency shift is established. First, the free space path loss is calculated according to the propagation distance. Then, considering the absorption of oxygen and water vapor in the atmosphere, the atmospheric absorption loss is calculated according to parameters such as frequency, temperature, and air pressure. For reflection and diffraction, the ray tracing method is used to calculate the terrain shielding and building shielding losses. Specifically, multiple rays are emitted from the transmitting point to the receiving point, and the intersection of the rays with the terrain or building is detected, and the additional loss is calculated according to the number of reflections and the diffraction edge. For mobile scenarios, the frequency shift caused by the Doppler effect is considered. The delay spread characteristics are analyzed based on the power delay spectrum to obtain the delay spread parameters of multipath propagation. Finally, a set of model parameters describing the characteristics of environmental electromagnetic propagation are obtained.

[0141] The multi-dimensional signal feature data and the parameters of the environmental electromagnetic propagation model are rearranged according to the signal feature dimension, space dimension, time dimension and posture dimension to construct a feature tensor. For example, a 5th-order tensor can be constructed, with each dimension corresponding to 5 signal features, xyz three-dimensional space coordinates, time and robot 6-DOF posture. The feature tensor is decomposed to obtain a core tensor and a factor matrix. The Tucker decomposition method is used to decompose the high-order tensor into the product of a low-order core tensor and multiple factor matrices.

[0142] The core tensor and factor matrix are optimized using the alternating least squares method. Other factors are fixed, a factor matrix is ​​optimized, and the iteration is repeated until convergence. In order to avoid overfitting, L1 and L2 regularization constraints are introduced. The regularization coefficient is determined by cross-validation, such as L1 coefficient is 0.01 and L2 coefficient is 0.001. After multiple iterations of optimization, the final core tensor and factor matrix are obtained.

[0143] The optimized core tensor and factor matrix are modally projected to obtain compressed feature data. Specifically, important feature modes are selected for projection, such as retaining the principal component that explains 80% of the variance. The grid resampling method is used to interpolate the compressed feature data of irregular sampling points to regular grid points. Then the grid data is smoothed using a Gaussian smoothing kernel to generate an initial feature distribution map.

[0144] The noise in the initial feature distribution map is suppressed by the singular value threshold processing method. The feature distribution map is subjected to singular value decomposition, and the singular values ​​less than the set threshold (such as 1% of the maximum singular value) are set to zero, and then the denoised distribution map is reconstructed. The L1 norm regularization method is combined to further enhance the sparsity.

[0145] Finally, variational regularization and gradient preserving filtering methods are used to optimize the boundary features of the initial feature distribution map. Variational regularization achieves smooth denoising by minimizing the energy functional while preserving the edge. Gradient preserving filtering adaptively adjusts the filter strength according to the local gradient information, strengthens the filter in the smooth area and weakens the filter in the edge area. After boundary optimization, the final target feature distribution map is output.

[0146] In this embodiment, through multi-dimensional signal feature extraction and tensor decomposition modeling, various feature information such as signal strength, arrival angle, phase difference, polarization, etc. are effectively integrated, which improves the comprehensiveness and accuracy of the feature distribution map. The environmental electromagnetic propagation model is combined with the robot posture data to achieve accurate modeling of complex electromagnetic environments and enhance the adaptability of the feature distribution map in different scenarios. The introduction of regularization constraints and boundary optimization technology effectively suppresses noise interference, improves the anti-interference ability and edge retention effect of the feature distribution map, and further improves the quality of the distribution map.

[0147] S103. Based on the initial position data of the signal source and the signal emission characteristic data, a zoning strategy is adopted to divide the measurement area, sampling grid points are constructed in the measurement area, the robot is controlled to perform signal measurement at the sampling grid points, a space-time two-dimensional adaptive beamforming algorithm is used to process the multi-dimensional signal characteristic data, a signal propagation model is constructed in combination with multipath effect compensation, the three-dimensional spatial coordinates of the signal source are calculated through a multi-constraint optimization algorithm, the three-dimensional spatial coordinates of the signal source are corrected by a particle filter algorithm, the corrected three-dimensional spatial coordinates of the signal source are fused and calibrated with the indoor three-dimensional map data based on a deep calibration network, and the final signal source position coordinates are output.

[0148] Based on the preliminary estimated signal source location, the measurement area is divided into multiple sub-areas using a partitioning strategy. Evenly distributed sampling grid points are constructed in each sub-area, and the robot is controlled to move to each grid point in turn for fine measurement. At each measurement point, a space-time two-dimensional adaptive beamforming algorithm is used to process the received signal. The algorithm adaptively adjusts the weights of the antenna array through the minimum mean square error criterion to achieve spatial filtering of the target signal and effectively suppress interference and noise.

[0149] Combined with multipath effect compensation technology, an accurate signal propagation model is constructed. This model considers the superposition effect of direct path and reflection path, and simulates the propagation characteristics of signals in complex indoor environments through ray tracing method. Based on the established propagation model, a multi-constraint optimization algorithm is used to solve the three-dimensional spatial coordinates of the signal source. The algorithm combines multiple constraints such as signal strength and arrival angle to minimize the error between the measured value and the model prediction value through iterative optimization.

[0150] To further improve the positioning accuracy, the particle filter algorithm is used to correct the calculated signal source coordinates. This algorithm approximates the posterior probability distribution of the signal source position through a large number of randomly sampled particles, which can effectively deal with nonlinear and non-Gaussian noise problems. Finally, the corrected signal source coordinates are fused and calibrated with the indoor three-dimensional map data using a deep calibration network. The network learns the mapping relationship between the signal measurement results and the actual position, compensates for the system error, and outputs the final high-precision signal source positioning results.

[0151] In an optional embodiment,

[0152] The space-time two-dimensional adaptive beamforming algorithm is used to process multi-dimensional signal feature data, and the signal propagation model is constructed in combination with multipath effect compensation. The three-dimensional spatial coordinates of the signal source are calculated through a multi-constraint optimization algorithm, including:

[0153] Calculate the time-frequency spectrum of multidimensional signal feature data to obtain a frequency domain feature matrix, perform singular value decomposition on the frequency domain feature matrix to extract principal component feature vectors, construct a two-dimensional covariance matrix containing spatial correlation and time correlation, perform feature decomposition on the two-dimensional covariance matrix using regularization constraints to obtain a signal subspace and a noise subspace, calculate a projection operator based on the orthogonality of the signal subspace and the noise subspace, calculate the expected signal steering vector in combination with a wave direction estimation algorithm, and use a recursive least squares criterion to optimize and obtain an optimal weight vector;

[0154] Establish an indoor environment model including the geometric characteristics of the reflecting surface and the dielectric characteristics of the material, use a layered ray tracing method to identify the direct path and the multipath propagation path, calculate the reflection point coordinate sequence and the diffraction edge position sequence of each propagation path in combination with the optimal weight vector, calculate the path loss based on the reflection coefficient and the unified diffraction theory, and construct a multipath propagation channel model considering frequency selective fading based on the path loss, path delay and phase offset;

[0155] Constructing a multi-constraint optimization objective function based on the multipath propagation channel model, wherein the multi-constraint optimization objective function includes an observation matrix constraint term, a signal power constraint term, a propagation delay constraint term, and a carrier phase constraint term, introducing an adaptive weight coefficient for each constraint term in the multi-constraint optimization objective function, and using a variable splitting strategy to decompose the multi-constraint optimization objective function into a sub-problem sequence;

[0156] The sub-problem sequence is constructed as an augmented Lagrangian function, wherein the augmented Lagrangian function includes a main variable, an auxiliary variable and a dual variable, the main variable is iteratively updated by a proximal gradient descent algorithm, the auxiliary variable is iteratively updated by a soft threshold operation, the dual variable is iteratively updated in combination with a dynamic step size strategy, and a preprocessing matrix is ​​constructed by online estimating curvature to improve the convergence speed;

[0157] The three-dimensional spatial coordinates of the initial signal source are obtained according to the iterative update result of the augmented Lagrangian function, and an error evaluation function including geometric consistency and energy consistency is constructed. When the calculation result of the error evaluation function is less than a preset threshold, the final three-dimensional spatial coordinates of the signal source are output.

[0158] First, the received multi-dimensional signal feature data is subjected to time-frequency analysis to calculate the frequency domain feature matrix. The frequency domain feature matrix is ​​subjected to singular value decomposition to extract the principal component eigenvector. Then a two-dimensional covariance matrix is ​​constructed, which contains both spatial correlation and time correlation information. Regularization constraints are imposed on the two-dimensional covariance matrix and eigendecomposition is performed to obtain the signal subspace and the noise subspace. Based on the orthogonality of the signal subspace and the noise subspace, the projection operator is calculated. Combined with the wave direction estimation algorithm, the steering vector of the expected signal is calculated. Finally, the recursive least squares criterion is used for optimization to obtain the optimal weight vector.

[0159] For example, for a uniform linear array with 8 elements, the received signal sampling length is 1024 points. First, perform short-time Fourier transform on the signal to obtain a 128×128 time-frequency spectrum matrix. Perform singular value decomposition on the matrix and take the first 4 principal component eigenvectors to form an 8×4 matrix U. Then calculate the two-dimensional covariance matrix R=UU^H, where H represents the conjugate transpose. Perform eigendecomposition on R, take the first 2 eigenvectors to form the signal subspace, and the remaining 6 eigenvectors to form the noise subspace. Calculate the projection operator P based on orthogonality. Use the MUSIC algorithm to estimate the signal arrival direction θ and obtain the steering vector a(θ). Finally, the optimal weight vector w is obtained by minimizing |Pa(θ)-w|^2.

[0160] Next, an indoor environment model is established that includes the geometric characteristics of the reflecting surface and the dielectric characteristics of the material. The layered ray tracing method is used to identify the direct path and multipath propagation path. Combined with the optimal weight vector obtained previously, the reflection point coordinate sequence and diffraction edge position sequence of each propagation path are calculated. Based on the reflection coefficient and unified diffraction theory, the path loss of each path is calculated. According to the path loss, path delay and phase offset, a multipath propagation channel model considering frequency selective fading is constructed.

[0161] For example, for a rectangular room of 10m×8m×3m, the wall material is concrete with a dielectric constant of 5-j0.3. The transmitter is located at (2, 1, 1.5) and the receiver is located at (8, 7, 1.5). Using the hierarchical ray tracing method, 1 direct path and 4 reflected paths are identified. For reflection path 1, the coordinates of the reflection point are (5, 1, 1.5). According to the incident angle and material parameters, the reflection coefficient is calculated to be 0.7∠-30°. The total path length is 10.2m, the corresponding path loss is 62dB, and the delay is 34ns. Taking all the paths into consideration, the multipath channel impulse response h(t) in this environment is obtained.

[0162] Then, a multi-constraint optimization objective function is constructed based on the multipath propagation channel model. The objective function contains observation matrix constraint terms, signal power constraint terms, propagation delay constraint terms and carrier phase constraint terms. Adaptive weight coefficients are introduced for each constraint term. The variable splitting strategy is adopted to decompose the multi-constraint optimization objective function into a series of sub-problems.

[0163] The above sub-problem sequence is constructed as an augmented Lagrangian function, which includes the main variable, auxiliary variable and dual variable. The proximal gradient descent algorithm is used to iteratively update the main variable. The soft threshold operation is used to iteratively update the auxiliary variable. The dual variable is iteratively updated in combination with the dynamic step size strategy. The preprocessing matrix is ​​constructed by online estimation of curvature to improve the convergence speed of the algorithm.

[0164] Finally, the initial signal source three-dimensional spatial coordinates are obtained according to the iterative update results of the augmented Lagrangian function. An error evaluation function including geometric consistency and energy consistency is constructed. When the calculation result of the error evaluation function is less than the preset threshold, the final signal source three-dimensional spatial coordinates are output.

[0165] In an optional embodiment,

[0166] The calculation formula of the augmented Lagrangian function is as follows:

[0167]

[0168] Among them, L rd (·) represents the augmented Lagrangian function, x represents the main variable, y represents the auxiliary variable, λ represents the dual variable, β represents the step size adjustment factor, f(x) represents the objective function, g i (x) represents the i-th constraint, and ρ represents the augmentation factor.

[0169] In this embodiment, space-time two-dimensional adaptive beamforming is adopted to make full use of the space-time characteristics of the signal, improve the anti-interference ability and spatial resolution. The influence of multipath effect is taken into account, a more accurate signal propagation model is established, and the positioning accuracy in complex environments is improved. A multi-constraint optimization algorithm is used to comprehensively utilize a variety of prior information to improve the robustness and reliability of positioning. It can effectively handle complex nonlinear optimization problems, especially those with multiple constraints. By incorporating constraints into the objective function, the constrained optimization problem can be converted into an unconstrained optimization problem, simplifying the solution process. It has good convergence and stability. By introducing penalty terms and gradually adjusting the dual variables, it is possible to gradually approach the optimal solution while ensuring that the constraints are met. Compared with the traditional Lagrange multiplier method, this method is less sensitive to the selection of initial values ​​and has stronger robustness.

[0170] Figure 2 FIG. 1 is a schematic diagram of the structure of the automatic walking and positioning system of the signal determination robot under complex environment according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0171] The first unit is used to obtain indoor three-dimensional map data, obtain environmental multimodal perception data through binocular vision sensors and three-dimensional laser radar scanning, perform feature extraction on the multimodal perception data to obtain environmental feature data, use graph neural network to perform semantic segmentation on the environmental feature data to obtain scene semantic information, combine point cloud registration technology to perform real-time matching of the scene semantic information with the three-dimensional map data to obtain environmental matching data, input the environmental matching data into a pre-trained trajectory planning model, integrate obstacle position prediction information and robot motion constraint parameters in the scene semantic information, and generate robot motion trajectory planning data;

[0172] The second unit is used to control the movement of the robot based on the robot motion trajectory planning data, perform broadband parallel acquisition of the measured signal through the orthogonal polarization directional antenna array in the phased array antenna system, obtain multi-dimensional signal feature data including signal strength, arrival angle, phase difference and polarization characteristics, input the multi-dimensional signal feature data and the robot's real-time posture data and preset environmental electromagnetic propagation model parameters into a tensor decomposition model, construct a signal feature distribution map, input the signal feature distribution map into a deep neural network for feature extraction to obtain a signal feature vector, and process the signal feature vector based on a multi-source information fusion algorithm to obtain signal source initial position data and signal emission characteristic data;

[0173] The third unit is used to divide the measurement area by adopting a partitioning strategy according to the initial position data of the signal source and the signal emission characteristic data, construct sampling grid points in the measurement area, control the robot to perform signal measurement at the sampling grid points, process the multi-dimensional signal feature data by adopting a space-time two-dimensional adaptive beamforming algorithm, construct a signal propagation model in combination with multipath effect compensation, calculate the three-dimensional spatial coordinates of the signal source by a multi-constraint optimization algorithm, correct the three-dimensional spatial coordinates of the signal source by adopting a particle filter algorithm, fuse and calibrate the corrected three-dimensional spatial coordinates of the signal source with the indoor three-dimensional map data based on a deep calibration network, and output the final signal source position coordinates.

[0174] According to a third aspect of the embodiments of the present invention,

[0175] An electronic device is provided, comprising:

[0176] processor;

[0177] a memory for storing processor-executable instructions;

[0178] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0179] According to a fourth aspect of the embodiments of the present invention,

[0180] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0181] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0182] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatic movement and positioning of a signal measurement robot in a complex environment, characterized in that: include: Acquire indoor three-dimensional map data, acquire environmental multimodal perception data through binocular vision sensors and three-dimensional laser radar scanning, perform feature extraction on the multimodal perception data to obtain environmental feature data, use graph neural network to perform semantic segmentation on the environmental feature data to obtain scene semantic information, combine point cloud registration technology to perform real-time matching of the scene semantic information with the three-dimensional map data to obtain environmental matching data, input the environmental matching data into a pre-trained trajectory planning model, integrate obstacle position prediction information and robot motion constraint parameters in the scene semantic information, and generate robot motion trajectory planning data; The robot movement is controlled based on the robot motion trajectory planning data, broadband parallel acquisition of the measured signal is performed through the orthogonal polarization directional antenna array in the phased array antenna system to obtain multi-dimensional signal feature data including signal strength, arrival angle, phase difference and polarization characteristics, the multi-dimensional signal feature data and the robot real-time posture data and the preset environmental electromagnetic propagation model parameters are input into the tensor decomposition model to construct a signal feature distribution map, the signal feature distribution map is input into the deep neural network for feature extraction to obtain a signal feature vector, and the signal feature vector is processed based on the multi-source information fusion algorithm to obtain the signal source initial position data and signal emission characteristic data; According to the initial position data of the signal source and the signal emission characteristic data, a zoning strategy is adopted to divide the measurement area, sampling grid points are constructed in the measurement area, the robot is controlled to perform signal measurement at the sampling grid points, a space-time two-dimensional adaptive beamforming algorithm is used to process the multi-dimensional signal feature data, a signal propagation model is constructed in combination with multipath effect compensation, the three-dimensional spatial coordinates of the signal source are calculated by a multi-constraint optimization algorithm, the three-dimensional spatial coordinates of the signal source are corrected by a particle filter algorithm, the corrected three-dimensional spatial coordinates of the signal source are fused and calibrated with the indoor three-dimensional map data based on a deep calibration network, and the final signal source position coordinates are output.

2. The method according to claim 1, characterized in that The environment feature data is semantically segmented using a graph neural network to obtain scene semantic information, and the scene semantic information is matched with the three-dimensional map data in real time using point cloud registration technology to obtain environment matching data, including: Adaptive sampling is used to calculate the local feature entropy of each point in the environmental feature data, a sampling density threshold is set based on the local feature entropy to perform downsampling to obtain a feature point sequence, a multi-scale covariance matrix of the feature point sequence is constructed to extract a geometric feature vector, the geometric feature vector is fused with a depth feature vector to generate a vertex feature vector, a dynamic threshold K nearest neighbor algorithm is used to construct a graph structure network, and the spatial propagation characteristics and geometric similarity of the connecting edges in the graph structure network are calculated to obtain an edge feature vector; Input the vertex feature vector and the edge feature vector into a multi-head cross attention network, obtain the attention weight coefficient through feature channel recalibration and spatial position encoding, adaptively aggregate the features in the graph structure network based on the attention weight coefficient to obtain local fusion features, input the local fusion features into the residual feature enhancement network to extract discriminative features, perform spectral domain transformation on the discriminative features and input them into the dynamic graph convolution network to obtain scene semantic information; Based on the scene semantic information, a similarity measurement model with semantic consistency constraints is constructed, reliable matching point pairs are screened through bidirectional cross-validation, the matching point pairs are probabilistically sampled and the transformation matrix is ​​calculated in combination with the local geometric structure consistency constraints, and the optimal initial transformation matrix is ​​selected using a robustness evaluation criterion; A hierarchical KD tree is used to perform local feature matching on the area covered by the optimal initial transformation matrix to obtain corresponding point pairs, and a Mahalanobis distance metric space that considers the semantic similarity between point pairs is constructed. The transformation matrix is ​​iteratively optimized in the metric space, and the registration accuracy and scene consistency of the optimized transformation matrix are evaluated in real time. When the evaluation index meets the convergence condition, the environment matching data is output.

3. The method according to claim 2, characterized in that A hierarchical KD tree is used to perform local feature matching on the area covered by the optimal initial transformation matrix to obtain corresponding point pairs, and a Mahalanobis distance metric space considering the semantic similarity between point pairs is constructed. The transformation matrix is ​​iteratively optimized in the metric space, and the registration accuracy and scene consistency of the optimized transformation matrix are evaluated in real time. When the evaluation index meets the convergence condition, the output environment matching data includes: Performing feature extraction on the point cloud data within the coverage area of ​​the optimal initial transformation matrix, extracting local geometric structure features and semantic label features of the point cloud data, and performing feature fusion on the local geometric structure features and the semantic label features to obtain a multi-dimensional feature vector; Calculating feature similarity for the multidimensional feature vector, constructing a feature similarity matrix, performing feature decomposition based on the feature similarity matrix to obtain feature main directions, using the feature main directions to perform spatial division on the point cloud data and construct a hierarchical KD tree; Calculate the geometric information entropy and semantic information entropy of the hierarchical KD tree nodes, construct a node splitting cost function by combining the geometric information entropy, semantic information entropy and position constraint, determine the optimal splitting parameter based on the minimization result of the node splitting cost function, organize similar feature points into the same subtree to obtain a hierarchical feature index; Calculating the correlation strength of feature points in the source point cloud and the target point cloud based on the hierarchical feature index, establishing a corresponding relationship between the feature points according to the correlation strength, and constructing edge connections according to the corresponding relationship to obtain a matching graph structure; Iteratively updating node features and edge weights in the matching graph structure, calculating a reconstruction error of the node features, constructing a metric learning objective function based on the reconstruction error, and optimizing the metric learning objective function to obtain a discriminative metric matrix; Calculating similarity scores of feature point pairs in the metric space defined by the discriminant metric matrix, constructing a transformation matrix optimization objective function by combining the similarity scores with local geometric constraints and global semantic constraints, and solving the transformation matrix optimization objective function to obtain an optimal transformation matrix; Calculate the registration error of the optimal transformation matrix at multiple feature scales, combine the registration error with the semantic label consistency to obtain an environment matching evaluation score, output the final environment feature matching data when the environment matching evaluation score meets the preset convergence condition, and when the environment matching evaluation score does not meet the preset convergence condition, optimize and update the feature extraction parameters, feature matching parameters and metric calculation parameters according to the environment matching evaluation score, and return to the feature extraction step to continue iterative optimization.

4. The method according to claim 1, characterized in that: Inputting the environment matching data into a pre-trained trajectory planning model, fusing the obstacle position prediction information and robot motion constraint parameters in the scene semantic information, and generating robot motion trajectory planning data includes: Mapping the environment matching data to the grid map to obtain an initial grid map, calculating the logarithmic value of the obstacle occupancy probability in the initial grid map to construct a static cost map, extracting the dynamic obstacle position information to perform Gaussian kernel density calculation to construct a dynamic cost map, converting the grid point semantic labels based on the semantic cost mapping function to construct a semantic cost map, and performing weighted combination of the static cost map, the dynamic cost map and the semantic cost map to obtain an environmental comprehensive cost map; Based on the comprehensive cost map of the environment, a state vector is constructed, wherein the state vector includes the robot position coordinates, motion speed, target position coordinates and local cost information, the state vector is input into the policy neural network to obtain the action parameters of linear velocity and angular velocity, the state-action value function is calculated according to the action parameters, a reward function including a target distance term, an obstacle avoidance term and a trajectory smoothing term is constructed based on the state-action value function, the parameters of the policy neural network are optimized by using a proximal policy algorithm, and the optimized action parameters are subjected to kinematic constraint processing to obtain a planned trajectory; The planned trajectory is interpolated with a cubic spline function to obtain a smooth trajectory, an elastic potential field is constructed based on the smooth trajectory, the trajectory curvature is set as an internal potential energy function, the obstacle repulsion force in the environmental comprehensive cost map is set as an external potential energy function, the weighted sum of the internal potential energy function and the external potential energy function is minimized based on the gradient descent method to obtain an optimized trajectory, the optimized trajectory is kinematically verified, and the final robot motion trajectory planning data is output.

5. The method according to claim 4, characterized in that Acquiring multidimensional signal feature data including signal strength, arrival angle, phase difference and polarization characteristics, inputting the multidimensional signal feature data, the robot's real-time posture data and preset environmental electromagnetic propagation model parameters into a tensor decomposition model, and constructing a signal feature distribution map includes: The power spectral density of the received signal is estimated by using a periodogram averaging method and signal strength data is obtained by constant false alarm threshold detection. An array manifold is constructed based on the antenna directional diagram to locate the array response and a multiple signal classification algorithm is used to calculate the spatial spectrum to obtain the angle of arrival data. Phase difference data is extracted by multi-channel coherent sampling and phase unwrapping processing. Polarization matrix is ​​reconstructed based on dual orthogonal polarization channels to obtain polarization characteristic data. The signal strength data, angle of arrival data, phase difference data and polarization characteristic data are combined to form multi-dimensional signal feature data. According to the multi-dimensional signal characteristic data, an environmental electromagnetic propagation model including distance attenuation, atmospheric absorption, reflection diffraction and frequency shift is established, terrain shielding and building shielding losses are calculated by ray tracing method, and delay spread characteristics are analyzed based on power delay spectrum to obtain environmental electromagnetic propagation model parameters; Rearranging the multidimensional signal feature data and the environmental electromagnetic propagation model parameters according to the signal feature dimension, spatial dimension, time dimension and posture dimension to construct a feature tensor, performing tensor decomposition on the feature tensor to obtain a core tensor and a factor matrix, optimizing the core tensor and the factor matrix using an alternating least squares method, introducing regularization constraints to achieve iterative convergence, and obtaining an optimized core tensor and factor matrix; The optimized core tensor and factor matrix are modally projected to obtain compressed feature data, and the compressed feature data are processed by grid resampling and interpolation smoothing to generate an initial feature distribution map. The noise in the initial feature distribution map is suppressed by singular value threshold processing and norm regularization method, and the boundary features of the initial feature distribution map are optimized in combination with variational regularization and gradient preserving filtering method, and the target feature distribution map is output.

6. The method according to claim 1, characterized in that The space-time two-dimensional adaptive beamforming algorithm is used to process multi-dimensional signal feature data, and the signal propagation model is constructed in combination with multipath effect compensation. The three-dimensional spatial coordinates of the signal source are calculated through a multi-constraint optimization algorithm, including: Calculate the time-frequency spectrum of multidimensional signal feature data to obtain a frequency domain feature matrix, perform singular value decomposition on the frequency domain feature matrix to extract principal component feature vectors, construct a two-dimensional covariance matrix containing spatial correlation and time correlation, perform feature decomposition on the two-dimensional covariance matrix using regularization constraints to obtain a signal subspace and a noise subspace, calculate a projection operator based on the orthogonality of the signal subspace and the noise subspace, calculate the expected signal steering vector in combination with a wave direction estimation algorithm, and use a recursive least squares criterion to optimize and obtain an optimal weight vector; Establish an indoor environment model including the geometric characteristics of the reflecting surface and the dielectric characteristics of the material, use a layered ray tracing method to identify the direct path and the multipath propagation path, calculate the reflection point coordinate sequence and the diffraction edge position sequence of each propagation path in combination with the optimal weight vector, calculate the path loss based on the reflection coefficient and the unified diffraction theory, and construct a multipath propagation channel model considering frequency selective fading based on the path loss, path delay and phase offset; Constructing a multi-constraint optimization objective function based on the multipath propagation channel model, wherein the multi-constraint optimization objective function includes an observation matrix constraint term, a signal power constraint term, a propagation delay constraint term, and a carrier phase constraint term, introducing an adaptive weight coefficient for each constraint term in the multi-constraint optimization objective function, and using a variable splitting strategy to decompose the multi-constraint optimization objective function into a sub-problem sequence; The sub-problem sequence is constructed as an augmented Lagrangian function, wherein the augmented Lagrangian function includes a main variable, an auxiliary variable and a dual variable, the main variable is iteratively updated by a proximal gradient descent algorithm, the auxiliary variable is iteratively updated by a soft threshold operation, the dual variable is iteratively updated in combination with a dynamic step size strategy, and a preprocessing matrix is ​​constructed by online estimating curvature to improve the convergence speed; The three-dimensional spatial coordinates of the initial signal source are obtained according to the iterative update result of the augmented Lagrangian function, and an error evaluation function including geometric consistency and energy consistency is constructed. When the calculation result of the error evaluation function is less than a preset threshold, the final three-dimensional spatial coordinates of the signal source are output.

7. The method according to claim 6, characterized in that The calculation formula of the augmented Lagrangian function is as follows: Among them, L rd (·) represents the augmented Lagrangian function, x represents the main variable, y represents the auxiliary variable, λ represents the dual variable, β represents the step size adjustment factor, f(x) represents the objective function, g i (x) represents the i-th constraint, and ρ represents the augmentation factor.

8. An automatic walking and positioning system for a signal measurement robot in a complex environment, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain indoor three-dimensional map data, obtain environmental multimodal perception data through binocular vision sensors and three-dimensional laser radar scanning, perform feature extraction on the multimodal perception data to obtain environmental feature data, use graph neural network to perform semantic segmentation on the environmental feature data to obtain scene semantic information, combine point cloud registration technology to perform real-time matching of the scene semantic information with the three-dimensional map data to obtain environmental matching data, input the environmental matching data into a pre-trained trajectory planning model, integrate obstacle position prediction information and robot motion constraint parameters in the scene semantic information, and generate robot motion trajectory planning data; The second unit is used to control the movement of the robot based on the robot motion trajectory planning data, perform broadband parallel acquisition of the measured signal through the orthogonal polarization directional antenna array in the phased array antenna system, obtain multi-dimensional signal feature data including signal strength, arrival angle, phase difference and polarization characteristics, input the multi-dimensional signal feature data and the robot's real-time posture data and preset environmental electromagnetic propagation model parameters into a tensor decomposition model, construct a signal feature distribution map, input the signal feature distribution map into a deep neural network for feature extraction to obtain a signal feature vector, and process the signal feature vector based on a multi-source information fusion algorithm to obtain signal source initial position data and signal emission characteristic data; The third unit is used to divide the measurement area by adopting a partitioning strategy according to the initial position data of the signal source and the signal emission characteristic data, construct sampling grid points in the measurement area, control the robot to perform signal measurement at the sampling grid points, process the multi-dimensional signal feature data by adopting a space-time two-dimensional adaptive beamforming algorithm, construct a signal propagation model in combination with multipath effect compensation, calculate the three-dimensional spatial coordinates of the signal source by a multi-constraint optimization algorithm, correct the three-dimensional spatial coordinates of the signal source by adopting a particle filter algorithm, fuse and calibrate the corrected three-dimensional spatial coordinates of the signal source with the indoor three-dimensional map data based on a deep calibration network, and output the final signal source position coordinates.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images