Multipath clustering and tracking method for dynamic wireless channel

Abstract

The invention discloses a multipath clustering and tracking method for a dynamic wireless channel, which belongs to the technical field of wireless communication, and comprises the following steps: acquiring time delay, angle and power parameters of multipath components in each time snapshot; running an adaptive neighborhood robust mean shift clustering algorithm, and combining a neighborhood structure quality index to realize static division of a non-uniform density cluster; constructing a directional bounding box of the cluster in a three-dimensional feature space by utilizing principal component analysis, and accurately representing a geometric boundary of the cluster; executing cluster-level tracking based on a directional bounding box and Kalman filtering, and correcting a prediction state; a cluster-level joint similarity is defined based on the shape and density of the bounding box to determine the birth and death of the cluster, and the remaining multipath is re-clustered to identify a new cluster. According to the method, the problems of over-segmentation or under-segmentation and association fuzziness in the prior art are solved. According to the method, the variable density structure of the multipath cluster and the spatio-temporal evolution trajectory of the variable density structure can be adaptively represented, and the precision and robustness of channel modeling in a high-dynamic non-stationary scene are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, specifically to a multipath clustering and tracking method for dynamic wireless channels. Background Technology

[0002] Channel modeling is fundamental to wireless communication systems, and its accuracy directly determines the design boundaries and performance limits of these systems. As wireless technology evolves towards larger bandwidths and massive MIMO antenna arrays, the spatiotemporal resolution of systems has significantly improved. Multipath components exhibit significant "cluster" distribution characteristics in the time delay, angle, and power domains. To address the modeling needs of complex propagation environments, cluster-based channel models have been widely adopted as the current mainstream modeling paradigm. However, designing time-varying channel models suitable for dynamic scenarios requires a deep understanding of the non-stationary evolution mechanism of clusters. Compared to traditional static scenarios, the relative positions of scatterers in dynamic channels constantly change, causing cluster parameters to continuously evolve over time, specifically manifested as the generation, disappearance, and drift of clusters in the parameter space. This dynamic nature renders previously developed static clustering methods inapplicable.

[0003] Specifically, existing multipath clustering and tracking techniques face multiple challenges. In clustering, the traditional K-Power-Means algorithm relies on a preset number of clusters, while the DBSCAN equal-density clustering algorithm depends on fixed neighborhood parameters, making it difficult to adaptively handle uneven density structures and prone to over-segmentation or incorrect merging. In tracking, as cluster density increases, adjacent clusters are easily interleaved in the parameter space. Most existing tracking algorithms rely solely on centroid distance as the association criterion, leading to association ambiguity when cluster spacing is close. Furthermore, existing frameworks often separate clustering and tracking processes, failing to utilize historical trajectory information to assist current clustering, resulting in insufficient recognition capability when facing weak, newly formed clusters. Therefore, developing multipath clustering and tracking methods for dynamic wireless channels has become a current research focus. Summary of the Invention

[0004] The purpose of this invention is to provide a multipath clustering and tracking method for dynamic wireless channels. By combining adaptive neighborhood static clustering with bounding box-based dynamic tracking, it solves the problems of over-segmentation or under-segmentation caused by the reliance on fixed parameters in existing clustering algorithms, and the problem of ambiguous association caused by the reliance on distance metrics in existing tracking algorithms. It achieves significant benefits in many aspects, improves the adaptability of clustering, enhances the robustness of tracking, and significantly improves the ability to identify weak new clusters, thus solving the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A multipath clustering and tracking method for dynamic wireless channels includes:

[0007] S1. Construct the multipath parameter space for dynamic wireless channels:

[0008] Acquire multipath component data obtained through channel measurement or ray tracing simulation. The multipath component data includes the time delay, azimuth angle, elevation angle, and linear power value of each multipath.

[0009] The multipath component data is cleaned, and the isolated forest algorithm is used to identify and remove geometric outliers in the three-dimensional parameter space, retaining the effective multipath components.

[0010] S2. Perform static clustering based on adaptive neighborhood truncation:

[0011] An adaptive neighborhood robust mean drift algorithm is used to perform unsupervised clustering of effective multipath components;

[0012] By constructing a neighborhood distance sequence for sample points and calculating the second-order gradient response of this sequence after multi-scale Gaussian smoothing, the system automatically detects structural boundaries transitioning from high-density regions within a cluster to low-density regions outside the cluster, and adaptively determines the optimal truncation neighborhood size for each sample point. Then, the mean shift is iterated using this neighborhood size until all sample points converge to the corresponding density center, thus obtaining the static cluster partitioning result under the current snapshot;

[0013] S3. Construct a geometric and kinematic description model of the cluster:

[0014] Based on the static cluster partitioning results, principal component analysis is used to determine the principal axis direction of each cluster and construct an oriented bounding box that can tightly enclose the components within the cluster.

[0015] Kalman filtering is used to establish the motion state equation of the cluster in three-dimensional space with time delay, azimuth angle of arrival and elevation angle of arrival. Based on the cluster state at the previous moment, the prior position of the cluster centroid at the current moment is predicted, and the state is corrected a posteriori by combining the observation value at the current moment.

[0016] S4. Perform cluster lifecycle management based on spatiotemporal evolution:

[0017] A serial interference cancellation strategy is used for multipath correlation.

[0018] An observation window is constructed based on the predicted cluster state. The multi-feature matching similarity between the surviving clusters and the candidate observation data at the current time is calculated to determine the survival or death status of the clusters.

[0019] From the raw multipath data at the current moment, the multipath components that fall within the observation window of the surviving clusters are preferentially removed. The remaining residual multipath components are returned as input data and re-clustered. The new clusters obtained by clustering are marked as new clusters, which are used to realize the full life cycle tracking of multipath clusters on continuous time snapshots.

[0020] Preferably, in step S2, the optimal truncation neighborhood size for each sample point is adaptively determined. Perform the following operations:

[0021] S21. Distance sequence construction:

[0022] For each target point to be analyzed in the parameter space Calculate its relationship with all other datasets. The multipath component distances of each sample point are calculated, and these multipath component distances are arranged in ascending order to obtain the neighborhood distance sequence of the target point. ,in ;

[0023] S22, Multi-scale smoothing processing:

[0024] Calculate the cumulative distance curve based on the neighborhood distance sequence. And using the Gaussian kernel function The cumulative distance curve is smoothed by convolution to obtain a smoothed cumulative curve. ,in The scaling parameter for Gaussian smoothing is used to suppress local noise disturbances caused by the randomness of multipath distribution;

[0025] S23, Gradient Feature Extraction:

[0026] For the smoothed cumulative curve Calculate the discrete first-order difference respectively and second-order difference Among them, the first-order difference Approximate representation of the first The radial distance between the nearest neighbor sample and the target point, and the second difference. Representing neighborhood distance with neighbor index The changes in the growth rate;

[0027] S24. Detection of structural mutation points:

[0028] Construct a scale-normalized second-order gradient response function The mathematical expression for quantifying the drastic change in neighborhood density is as follows:

[0029]

[0030] Among them, the numerator term Characterizing the cumulative curve in the th The local bending intensity at a point reflects the acceleration of the change in sample distribution density; the denominator term Used to normalize the effect of the first-order distance slope, preventing the structural bending features from being masked by the large distance value itself;

[0031] S25. Determining the cutoff position:

[0032] Search within the preset neighbor index search range The global maximum point is found, which corresponds to the topological inflection point from the cluster core to the sparse periphery; the index value corresponding to this maximum point is incremented by 1 to determine the adaptive truncation position. , for target point Determine an optimal neighborhood range that can contain effective information within the cluster while eliminating interference between clusters.

[0033] Preferably, in step S25, the adaptive cutoff position is determined. Furthermore, it also includes utilizing neighborhood structure quality indicators. The reliability of the truncation results is evaluated, and the neighborhood structure quality index is determined. Based on global separation Local jump degree and centralized weight The specific calculation formula is as follows:

[0034]

[0035] Among them, global separation The calculation formula is:

[0036]

[0037] In the formula, Indicates the preceding The average distance between the nearest neighbors characterizes the compactness of the cluster; Indicates the first To the The average distance between the neighbors represents the sparsity outside the cluster; To prevent numerically stable terms with a denominator of zero;

[0038] Among them, local jump degree The calculation formula is:

[0039]

[0040] In the formula, Indicates the distance and gap at the boundary; This represents the average increment of the distance between adjacent samples within a cluster; the function Used to map the jump amplitude to interval;

[0041] Among them, centralized weight The calculation formula is:

[0042]

[0043] In the formula, The preset attenuation coefficient, The weight term represents the total number of samples participating in the clustering. It is used to penalize excessively large neighborhood sizes, favoring the formation of compact micro-cluster structures.

[0044] Preferably, in step S3, an oriented bounding box capable of tightly enclosing the components within the cluster is constructed by performing the following operations:

[0045] S31, Basic Bounding Box Construction:

[0046] Calculate the covariance matrix of all multipath components within the cluster in three-dimensional space for time delay, azimuth angle of arrival, and elevation angle of arrival. And perform eigenvalue decomposition on the covariance matrix. The eigenvector matrix is ​​obtained. Using the principal axis direction of the local coordinate system of the cluster as the reference direction, all multipath components within the cluster are projected onto this principal axis direction, and the projection extrema are calculated to determine the basic half-length vector. ;

[0047] S32. Statistical Feature Extension:

[0048] The basic bounding box is elastically expanded based on the root mean square (RMS) value. The power-weighted RMS expansion values ​​of the cluster are calculated in three dimensions: time delay, azimuth angle of arrival, and elevation angle of arrival, respectively, to form the RMS vector. Using a preset expansion factor Calculate the final extended half-length vector This yields an extended bounding box that includes the statistical residual.

[0049] S33. Overlapping Conflict Handling:

[0050] The separation axis theorem is used to detect whether the extended bounding boxes of different clusters geometrically overlap in three-dimensional space; if the first axis is detected... Cluster and the first If clusters overlap, compare their volumes, select the larger cluster, and proceed with the next cluster according to a preset step size. Iteratively reduces its expansion factor, i.e. And recalculate its extended half-length vector until the bounding boxes of all clusters in the space do not overlap, in order to eliminate geometric ambiguity in multipath association.

[0051] Preferably, in step S3, establishing the motion state equations of the cluster using Kalman filtering includes:

[0052] Definition of the first Clusters in State vector at time step It is a six-dimensional vector containing three-dimensional centroid coordinates. and three-dimensional center of mass velocity ,Right now ;

[0053] Based on the assumption of uniform motion, a time update equation for the Kalman filter is established for use in... State prediction at time Prior state at time :

[0054]

[0055] and the prior error covariance matrix :

[0056]

[0057] Wherein, the state transition matrix process noise covariance matrix The specific form is:

[0058]

[0059] In the formula, It is a third-order identity matrix. The time interval between two adjacent snapshots; The acceleration noise variance diagonal matrix, comprising three components—time delay, azimuth angle of arrival, and elevation angle of arrival—is used to characterize the uncertainty introduced by non-uniform drift during cluster motion; the predicted prior state... The position component in the equation is the predicted centroid position.

[0060] Preferably, in step S4, multi-feature matching similarity is calculated. ,include:

[0061] The first Clusters in The geometry of the oriented bounding box at time step is translated to the Kalman filter. Construct an observation window based on the predicted centroid position at each moment;

[0062] Calculate the statistical characteristics of the measurement data at the current time within the observation window, and compare them with... The features at each time point are compared to calculate the shape similarity. and density similarity ;

[0063] shape similarity Based on normalized volume difference Calculate, the expression is ,in, and These are the volumes of the bounding boxes at the current and previous time steps, respectively.

[0064] density similarity Based on normalized density difference Calculate, the expression is ,in, and These represent the multipath point densities within the bounding box at the current and previous time steps, respectively.

[0065] The multi-feature matching similarity is the product of the two features, expressed as follows: ;like Greater than the preset survival threshold If the current observation is obtained, the cluster is determined to be alive at the current moment, and the state of the Kalman filter is updated using the current observation.

[0066] Preferably, in step S4, the execution of a serial interference cancellation strategy includes:

[0067] S41. Based on the predicted bounding box of the surviving cluster, determine the explained region in the current snapshot that belongs to the surviving cluster;

[0068] S42. Traverse all the original multipath components in the current snapshot and check whether they fall within the predicted bounding box of any of the above-mentioned live clusters.

[0069] S43. Treat the multipath components that fall within the range of surviving clusters as old cluster components and remove them from the original dataset to eliminate the shading effect and interference of strong surviving clusters on potential weak new clusters.

[0070] S44. The remaining unremoved multipath components are taken as residual components and input into the adaptive neighborhood robust mean drift algorithm for independent clustering.

[0071] S45. For each new cluster generated by clustering, initialize a new Kalman filter trajectory, initialize the velocity component in its state vector to zero, and mark it as a new cluster, adding it to the tracking list for the next time step.

[0072] Compared with the prior art, the beneficial effects of the present invention are:

[0073] 1. This invention integrates the adaptive neighborhood robust mean drift algorithm and the BoxKF-Cluster tracking algorithm, providing a high-precision multipath parameter extraction scheme for dynamic channel modeling. It also endows the model with the ability to perceive cluster emergence and extinction evolution, achieving significant benefits in multiple aspects. Regarding cluster adaptability, the improved adaptive neighborhood robust mean drift algorithm, by analyzing the second-order gradient response of the neighborhood distance sequence, can automatically detect abrupt changes in density structure, accurately determining the optimal neighborhood size that matches the local density of each sample point. This avoids the over-segmentation or under-segmentation problems caused by fixed parameters in traditional algorithms. Compared with traditional KPM or DBSCAN algorithms, this optimization enables the algorithm to adapt to the non-uniform distribution of dense cores and sparse edges in the channel, significantly improving the accuracy of clustering.

[0074] 2. This invention also achieves excellent performance in tracking robustness. By introducing directional bounding boxes and combining them with Kalman filtering, it no longer relies solely on centroid distance for tracking but integrates the geometric features and motion inertia of the clusters. This mechanism enables the algorithm to maintain trajectory continuity even when clusters are close together or even briefly intersect, effectively solving the correlation ambiguity problem caused by traditional distance metrics. It also has the ability to identify weak targets. By adopting a serial interference cancellation strategy and using tracking prediction information to construct a feedback loop, it prioritizes the removal of strong-power surviving cluster components. This strategy eliminates the masking effect of strong clusters on weak clusters, enabling the effective detection of weak new clusters hidden in the background. This more realistically restores the birth and death process of scatterers in dynamic channels, providing a reliable basis for dynamic cluster channel modeling in 5G / 6G wireless communication systems. Attached Figure Description

[0075] Figure 1 This is a flowchart of the multipath clustering and tracking method for dynamic wireless channels according to the present invention;

[0076] Figure 2 This is a schematic diagram illustrating the principle of determining the optimal neighborhood cutoff position and updating the density center based on the second-order gradient response of the distance sequence in this invention.

[0077] Figure 3 This is a schematic diagram of the oriented bounding box for constructing clusters based on principal component analysis and root mean square statistical extension according to the present invention;

[0078] Figure 4 This is a schematic diagram of the overall process of the multipath cluster tracking, birth and death determination, and serial interference elimination strategy of the present invention;

[0079] Figure 5 This is a schematic diagram showing the decomposition of the state prediction, position correction, and birth-death decision-making processes in the cluster tracking process of the present invention;

[0080] Figure 6 This is a schematic diagram illustrating the effect of multipath component clustering in a simulation scenario according to the present invention;

[0081] Figure 7 This is a schematic diagram illustrating the effect of multipath cluster trajectory tracking in a simulated scenario according to the present invention;

[0082] Figure 8 This is a schematic diagram comparing the clustering effects of the algorithm of this invention with existing comparative algorithms. Detailed Implementation

[0083] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0084] To address the issues of oversegmentation or undersegmentation caused by existing clustering algorithms relying on fixed parameters, and the ambiguity in associations caused by tracking algorithms relying solely on distance metrics, please refer to [link to relevant documentation]. Figures 1-8 This embodiment provides the following technical solution:

[0085] A multipath clustering and tracking method for dynamic wireless channels includes:

[0086] S1. Constructing the multipath parameter space for dynamic wireless channels:

[0087] Multipath component data, including time delay, azimuth angle of arrival, elevation angle of arrival, and linear power value, is acquired through channel measurement or ray tracing simulation. This data is then cleaned, and the isolated forest algorithm is used to identify and remove geometric outliers in the three-dimensional parameter space, retaining only the valid multipath components. Figure 6 As shown in (a), the original multipath distribution contains a certain amount of noise and isolated points. After preprocessing, a clean and effective multipath dataset is obtained.

[0088] S2. Perform static clustering based on adaptive neighborhood truncation:

[0089] An adaptive neighborhood robust mean drift algorithm is used to perform unsupervised clustering of effective multipath components. Static clustering analysis is then performed on the preprocessed multipath components to obtain the static cluster structure at the current time step, such as... Figure 6 As shown in (b), dots of different colors represent different multipath clusters obtained from the division.

[0090] By constructing a neighborhood distance sequence for sample points and calculating the second-order gradient response of this sequence after multi-scale Gaussian smoothing, the system automatically detects structural boundaries transitioning from high-density regions within a cluster to low-density regions outside the cluster, and adaptively determines the optimal truncation neighborhood size for each sample point. The mean shift is iterated using this neighborhood size until all sample points converge to the corresponding density center, thus obtaining the static cluster partitioning result under the current snapshot.

[0091] In this embodiment, in S2, the optimal truncation neighborhood size for each sample point is adaptively determined. Perform the following operations:

[0092] S21. Distance sequence construction:

[0093] For each target point to be analyzed in the parameter space Calculate its relationship with all other datasets. The multipath component distances of each sample point are calculated, and these distances are arranged in ascending order to obtain the neighborhood distance sequence of the target point. ,in ;

[0094] S22, Multi-scale smoothing processing:

[0095] To eliminate the impact of random fluctuations in micro-data, a cumulative distance curve is calculated based on the neighborhood distance sequence. And using the Gaussian kernel function Perform convolution smoothing on the cumulative distance curve to obtain the smoothed cumulative curve. ,in The scaling parameter for Gaussian smoothing is used to suppress local noise disturbances caused by the randomness of multipath distribution;

[0096] S23, Gradient Feature Extraction:

[0097] like Figure 2 As shown in (c), the smoothed cumulative curve Calculate the discrete first-order difference respectively and second-order difference Among them, the first-order difference Approximate representation of the first The radial distance between the nearest neighbor sample and the target point, and the second difference. Representing neighborhood distance with neighbor index The changes in the growth rate;

[0098] S24. Detection of structural mutation points:

[0099] Construct a scale-normalized second-order gradient response function The mathematical expression for quantifying the drastic change in neighborhood density is as follows:

[0100]

[0101] Among them, the numerator term Characterizing the cumulative curve in the th The local bending intensity at a point reflects the acceleration of the change in sample distribution density; the denominator term Used to normalize the effect of the first-order distance slope, preventing the structural bending features from being masked by the large distance value itself;

[0102] S25. Determining the cutoff position:

[0103] like Figure 2 As shown by the curve in (c), within a cluster, the neighborhood distance increases relatively gradually with the index, while at the boundary transitioning from the cluster core to the cluster periphery, the distance growth rate exhibits a significant abrupt change, indicating a shift within the preset neighbor index search range. The global maximum point is found, which corresponds to the topological inflection point from the cluster core to the sparse periphery; the index value corresponding to this maximum point is incremented by 1 to determine the adaptive truncation position. , for target point Determine an optimal neighborhood range that includes both effective intra-cluster information and excludes inter-cluster interference, such as... Figure 2 As shown in (d), the adaptive cutoff position is determined. Using neighborhood parameters, mean drift iteration is performed in parallel on the multipath components. Components that converge to the same density center are marked as the same cluster. Since the neighborhood size of each sample point is adaptively determined according to its local topology, the density center can quickly converge to the high-density region, thus achieving accurate static clustering of the multipath components in the current snapshot.

[0104] In this embodiment, in S25, the adaptive truncation position is determined. Afterwards, as Figure 2 (a) and Figure 2 As shown in (b), the neighborhood structure quality index is used. A reliability assessment of the truncation results shows that the data detected by the second-order gradient... Point-corresponding neighborhood structure quality index The global optimum value indicates that the location can achieve a good balance between intra-cluster compactness and inter-cluster separation, thus determining a suitable local neighborhood range for the target point;

[0105] Neighborhood structure quality index Based on global separation Local jump degree and centralized weight The specific calculation formula is as follows:

[0106]

[0107] Among them, global separation The calculation formula is:

[0108]

[0109] In the formula, Indicates the preceding The average distance between the nearest neighbors characterizes the compactness of the cluster; Indicates the first To the The average distance between the neighbors represents the sparsity outside the cluster; To prevent numerically stable terms with a denominator of zero;

[0110] Among them, local jump degree The calculation formula is:

[0111]

[0112] In the formula, Indicates the distance and gap at the boundary; This represents the average increment of the distance between adjacent samples within a cluster; the function Used to map the jump amplitude to interval;

[0113] Among them, centralized weight The calculation formula is:

[0114]

[0115] In the formula, The preset attenuation coefficient, The weight term represents the total number of samples participating in the clustering. It is used to penalize excessively large neighborhood sizes, favoring the formation of compact micro-cluster structures.

[0116] S3. Construct a geometric and kinematic description model of the cluster:

[0117] Based on the static cluster partitioning results, principal component analysis is used to determine the principal axis direction of each cluster and construct an oriented bounding box that can tightly enclose the components within the cluster.

[0118] Kalman filtering is used to establish the motion state equation of the cluster in three-dimensional space with time delay, azimuth angle of arrival and elevation angle of arrival. Based on the cluster state at the previous moment, the prior position of the cluster centroid at the current moment is predicted, and the state is corrected a posteriori by combining the observation value at the current moment.

[0119] In this embodiment, in S3, an oriented bounding box that can tightly enclose the components within the cluster is constructed, and the following operations are performed:

[0120] S31, Basic Bounding Box Construction:

[0121] like Figure 3 As shown, calculate the covariance matrix of all multipath components within the cluster in three-dimensional space in terms of time delay, azimuth angle of arrival, and elevation angle of arrival. And perform eigenvalue decomposition on the covariance matrix. The eigenvector matrix is ​​obtained. Using the principal axis direction as the local coordinate system of the cluster, all multipath components within the cluster are projected onto the principal axis direction, and the projection extrema are calculated to determine the fundamental half-length vector. ;

[0122] S32. Statistical Feature Extension:

[0123] Perform a flexible expansion of the basic bounding box based on the root mean square, such as... Figure 3 As shown, points of different colors represent different power levels. The power-weighted root mean square spread (RMS) values ​​of the cluster are calculated in three dimensions: time delay, azimuth angle of arrival, and elevation angle of arrival, respectively, forming the RMS vector. Using a preset expansion factor Calculate the final extended half-length vector This yields an extended bounding box containing statistical margins. This approach not only tightly encloses the effective components of the current snapshot but also reserves margins for changes in cluster morphology at the next time step through statistical expansion.

[0124] S33. Overlapping Conflict Handling:

[0125] The separation axis theorem is used to detect whether the extended bounding boxes of different clusters geometrically overlap in three-dimensional space; if the first axis is detected... Cluster and the first If clusters overlap, compare their volumes, select the larger cluster, and proceed with the next cluster according to a preset step size. Iteratively reduces its expansion factor, i.e. And recalculate its extended half-length vector until the bounding boxes of all clusters in the space do not overlap, in order to eliminate geometric ambiguity in multipath association.

[0126] In this embodiment, in S3, the motion state equation of the cluster is established using Kalman filtering, including:

[0127] Definition of the first Clusters in State vector at time step It is a six-dimensional vector containing three-dimensional centroid coordinates. and three-dimensional center of mass velocity ,Right now ;

[0128] Based on the assumption of uniform motion, a time update equation for the Kalman filter is established for use in... State prediction at time Prior state at time :

[0129]

[0130] and the prior error covariance matrix :

[0131]

[0132] Wherein, the state transition matrix process noise covariance matrix The specific form is as follows:

[0133]

[0134] In the formula, It is a third-order identity matrix. The time interval between two adjacent snapshots; The acceleration noise variance diagonal matrix, comprising three components—time delay, azimuth angle of arrival, and elevation angle of arrival—is used to characterize the uncertainty introduced by non-uniform drift during cluster motion; the predicted prior state... The position component in the equation is the predicted centroid position.

[0135] It should be noted that the time update equation of Kalman filtering is used to predict the prior position of the cluster centroid at the next time step, and the extended bounding box of the cluster at the previous time step is translated to the prior position to construct the observation window of the current snapshot, providing a spatiotemporal prior for subsequent birth and death decisions and new cluster detection.

[0136] S4. Perform cluster lifecycle management based on spatiotemporal evolution:

[0137] A serial interference cancellation strategy is used for multipath correlation.

[0138] An observation window is constructed based on the predicted cluster state. The multi-feature matching similarity between the surviving clusters and the candidate observation data at the current time is calculated to determine the survival or death status of the clusters.

[0139] From the raw multipath data at the current moment, the multipath components that fall within the observation window of the surviving clusters are preferentially removed. The remaining residual multipath components are returned as input data and re-clustered. The new clusters obtained by clustering are marked as new clusters, which are used to realize the full life cycle tracking of multipath clusters on continuous time snapshots.

[0140] like Figure 4 and Figure 5 (a)– Figure 5 As shown in (c), multi-feature matching similarity is calculated. ,include:

[0141] The first Clusters in The geometry of the oriented bounding box at time step is translated to the Kalman filter. Construct an observation window based on the predicted centroid position at each moment;

[0142] Calculate the statistical characteristics of the measurement data at the current time within the observation window, and compare them with... The features at each time point are compared to calculate the shape similarity. and density similarity ;

[0143] Shape similarity Based on normalized volume difference Calculate, the expression is ,in, and These are the volumes of the bounding boxes at the current and previous time steps, respectively.

[0144] Density similarity Based on normalized density difference Calculate, the expression is ,in, and These represent the multipath point densities within the bounding box at the current and previous time steps, respectively.

[0145] The similarity of multi-feature matching is the product of the two features, expressed as follows: ;like Greater than the preset survival threshold If the current observation is obtained, the cluster is determined to be alive at the current moment, and the state of the Kalman filter is updated using the current observation.

[0146] In this embodiment, in step S4, the serial interference cancellation strategy is executed, including:

[0147] S41. Based on the predicted bounding box of the surviving cluster, determine the explained regions belonging to the surviving cluster in the current snapshot;

[0148] S42. Traverse all the original multipath components in the current snapshot and check whether they fall within the predicted bounding box of any of the above-mentioned live clusters.

[0149] S43. Treat the multipath components that fall within the range of surviving clusters as old cluster components and remove them from the original dataset to eliminate the shading effect and interference of strong surviving clusters on potential weak new clusters.

[0150] S44. The remaining multipath components that have not been removed are taken as residual components, and the residual components are input into the adaptive neighborhood robust mean drift algorithm for independent clustering.

[0151] S45. For each new cluster generated by clustering, initialize a new Kalman filter trajectory, initialize the velocity component in its state vector to zero, and mark it as a new cluster, adding it to the tracking list for the next time step.

[0152] It should be noted that, since the strong cluster components have been stripped away, the weak new clusters hidden in the background can be clearly identified, and new Kalman filter trajectories are initialized for them and included in the subsequent tracking list.

[0153] like Figure 8 As shown in the figure, the adaptive neighborhood robust mean drift algorithm proposed in this embodiment is compared with the existing typical clustering algorithm in terms of clustering effect. It can be seen that by optimizing the neighborhood selection strategy, this embodiment significantly reduces the phenomena of over-segmentation and under-segmentation.

[0154] Specifically, such as Figure 8 (b) The traditional MP-NN-RMS algorithm, due to its fixed bandwidth, is prone to oversegmentation in regions with drastic density changes (such as cluster edges), incorrectly dividing a potentially complete cluster into multiple smaller clusters; for example... Figure 8 (c) The KPM algorithm requires a pre-defined number of clusters, making it difficult to adapt to dynamic channel environments where the number of clusters is unknown and changes over time. This leads to a rigid clustering structure that deviates from the true physical distribution. Figure 8 The DBSCAN algorithm shown in (d) relies on a fixed density threshold, which can easily misclassify low-density but physically valid multipath components as noise, thus losing some cluster information.

[0155] In comparison, such as Figure 8 (a) shows the clustering results of the adaptive neighborhood robust mean drift clustering algorithm in this embodiment. Figure 2 The second-order gradient response and neighborhood structure quality assessment mechanism shown in the algorithm can adaptively adjust the search range according to the local density, automatically shrink the neighborhood in the cluster core area, and appropriately expand the neighborhood in the cluster edge area, so as to obtain clustering results with clear boundaries and complete structure.

[0156] like Figure 7 As shown, the effect of this embodiment in cluster-level trajectory tracking is as follows: Figure 7 As shown in (b), after isolated discrete paths have been removed in the preprocessing stage, the final cluster trajectory clearly demonstrates the evolution path of the multipath cluster on the time axis. Even in complex situations such as multipath intersections or local occlusion, relying on Figure 3 The directional bounding box modeling capabilities shown and Figure 5The Kalman prediction and observation correction mechanism shown in this embodiment still maintains the continuity and identity consistency of the cluster trajectory, without trajectory swapping or breakage.

[0157] Therefore, this invention integrates the adaptive neighborhood robust mean drift algorithm and the BoxKF-Cluster tracking algorithm to provide a high-precision multipath parameter extraction scheme for dynamic channel modeling, and endows the model with the ability to perceive the evolution of cluster birth and death, achieving significant benefits in many aspects.

[0158] First, regarding clustering adaptability, the improved adaptive neighborhood robust mean drift algorithm can automatically detect abrupt changes in density structure by analyzing the second-order gradient response of the neighborhood distance sequence, accurately determining the optimal neighborhood size that matches the local density of each sample point. This avoids the over-segmentation or under-segmentation problems caused by fixed parameters in traditional algorithms. Compared with traditional KPM or DBSCAN algorithms, this optimization enables the algorithm to adapt to the non-uniform distribution of dense cores and sparse edges in the channel, significantly improving the accuracy of clustering.

[0159] Secondly, in terms of tracking robustness, this invention introduces oriented bounding boxes and combines them with Kalman filtering. Instead of relying solely on the distance between cluster centroids for association, it simultaneously utilizes the geometric features and motion inertia information of the clusters. Even when the cluster spacing is close or there is a brief overlap, the continuity of the trajectory and the consistency of identity can still be maintained, effectively alleviating the problem of association ambiguity in traditional distance-based metrics.

[0160] Furthermore, this invention also possesses a strong ability to identify weak targets. By employing a serial interference cancellation strategy, this invention utilizes tracking and prediction information to construct a feedback loop, preferentially stripping away strong-power surviving cluster components, and explicitly "reclaiming" the residual data space where weak clusters are located from the dataset. This significantly reduces the occlusion effect of strong clusters on weak clusters, enabling the effective detection of weak new clusters hidden in the background, and more realistically reflecting the birth and death process of scatterer clusters in dynamic channels.

[0161] In summary, the multipath clustering and tracking method proposed in this invention demonstrates outstanding performance in improving clustering accuracy, optimizing trajectory continuity, enhancing the ability to identify weak targets, and improving adaptability to complex scenarios, providing strong technical support for the modeling and simulation of dynamic wireless channels.

[0162] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0163] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multipath clustering and tracking method for dynamic wireless channels, characterized in that, include: S1. Construct the multipath parameter space for dynamic wireless channels: Acquire multipath component data obtained through channel measurement or ray tracing simulation. The multipath component data includes the time delay, azimuth angle, elevation angle, and linear power value of each multipath. The multipath component data is cleaned, and the isolated forest algorithm is used to identify and remove geometric outliers in the three-dimensional parameter space, retaining the effective multipath components. S2. Perform static clustering based on adaptive neighborhood truncation: An adaptive neighborhood robust mean drift algorithm is used to perform unsupervised clustering of effective multipath components; By constructing a neighborhood distance sequence for sample points and calculating the second-order gradient response of this sequence after multi-scale Gaussian smoothing, the system automatically detects structural boundaries transitioning from high-density regions within a cluster to low-density regions outside the cluster, and adaptively determines the optimal truncation neighborhood size for each sample point. Then, the mean shift is iterated using this neighborhood size until all sample points converge to the corresponding density center, thus obtaining the static cluster partitioning result under the current snapshot; S3. Construct a geometric and kinematic description model of the cluster: Based on the static cluster partitioning results, principal component analysis is used to determine the principal axis direction of each cluster and construct an oriented bounding box that can tightly enclose the components within the cluster. Kalman filtering is used to establish the motion state equation of the cluster in three-dimensional space with time delay, azimuth angle of arrival and elevation angle of arrival. Based on the cluster state at the previous moment, the prior position of the cluster centroid at the current moment is predicted, and the state is corrected a posteriori by combining the observation value at the current moment. S4. Perform cluster lifecycle management based on spatiotemporal evolution: A serial interference cancellation strategy is used for multipath correlation. An observation window is constructed based on the predicted cluster state. The multi-feature matching similarity between the surviving clusters and the candidate observation data at the current time is calculated to determine the survival or death status of the clusters. From the raw multipath data at the current moment, the multipath components that fall within the observation window of the surviving clusters are preferentially removed. The remaining residual multipath components are returned as input data and re-clustered. The new clusters obtained by clustering are marked as new clusters, which are used to realize the full life cycle tracking of multipath clusters on continuous time snapshots.

2. The multipath clustering and tracking method for dynamic wireless channels according to claim 1, characterized in that, In step S2, the optimal truncation neighborhood size for each sample point is adaptively determined. Perform the following operations: S21. Distance sequence construction: For each target point to be analyzed in the parameter space Calculate its relationship with all other datasets. The multipath component distances of each sample point are calculated, and these multipath component distances are arranged in ascending order to obtain the neighborhood distance sequence of the target point. ,in ; S22, Multi-scale smoothing processing: Calculate the cumulative distance curve based on the neighborhood distance sequence. And using the Gaussian kernel function The cumulative distance curve is smoothed by convolution to obtain a smoothed cumulative curve. ,in The scaling parameter for Gaussian smoothing is used to suppress local noise disturbances caused by the randomness of multipath distribution; S23, Gradient Feature Extraction: For the smoothed cumulative curve Calculate the discrete first-order difference respectively and second-order difference Among them, the first-order difference Approximate representation of the first The radial distance between the nearest neighbor sample and the target point, and the second difference. Representing neighborhood distance with neighbor index The changes in the growth rate; S24. Detection of structural mutation points: Construct a scale-normalized second-order gradient response function The mathematical expression for quantifying the drastic change in neighborhood density is as follows: ； Among them, the numerator term Characterizing the cumulative curve in the th The local bending intensity at a point reflects the acceleration of the change in sample distribution density; the denominator term Used to normalize the effect of the first-order distance slope, preventing the structural bending features from being masked by the large distance value itself; S25. Determining the cutoff position: Search within the preset neighbor index search range The global maximum point is found, which corresponds to the topological inflection point from the cluster core to the sparse periphery; the index value corresponding to this maximum point is incremented by 1 to determine the adaptive truncation position. , for target point Determine an optimal neighborhood range that can contain effective information within the cluster while eliminating interference between clusters.

3. The multipath clustering and tracking method for dynamic wireless channels according to claim 2, characterized in that, In step S25, the adaptive truncation position is determined. Furthermore, it also includes utilizing neighborhood structure quality indicators. The reliability of the truncation results is evaluated, and the neighborhood structure quality index is determined. Based on global separation Local jump degree and centralized weight The specific calculation formula is as follows: ； Among them, global separation The calculation formula is: ； In the formula, Indicates the preceding The average distance between the nearest neighbors characterizes the compactness of the cluster; Indicates the first To the The average distance between the neighbors represents the sparsity outside the cluster; To prevent numerically stable terms with a denominator of zero; Among them, local jump degree The calculation formula is: ； In the formula, Indicates the distance and gap at the boundary; This represents the average increment of the distance between adjacent samples within a cluster; the function Used to map the jump amplitude to interval; Among them, centralized weight The calculation formula is: ； In the formula, The preset attenuation coefficient, The weight term represents the total number of samples participating in the clustering. It is used to penalize excessively large neighborhood sizes, favoring the formation of compact micro-cluster structures.

4. The multipath clustering and tracking method for dynamic wireless channels according to claim 3, characterized in that, In step S3, an oriented bounding box that can tightly enclose the components within the cluster is constructed, and the following operations are performed: S31, Basic Bounding Box Construction: Calculate the covariance matrix of all multipath components within the cluster in three-dimensional space for time delay, azimuth angle of arrival, and elevation angle of arrival. And perform eigenvalue decomposition on the covariance matrix. The eigenvector matrix is ​​obtained. Using the principal axis direction of the local coordinate system of the cluster as the reference direction, all multipath components within the cluster are projected onto this principal axis direction, and the projection extrema are calculated to determine the basic half-length vector. ; S32. Statistical Feature Extension: The basic bounding box is elastically expanded based on the root mean square (RMS) value. The power-weighted RMS expansion values ​​of the cluster are calculated in three dimensions: time delay, azimuth angle of arrival, and elevation angle of arrival, respectively, to form the RMS vector. Using a preset expansion factor Calculate the final extended half-length vector This yields an extended bounding box that includes the statistical residual. S33. Overlapping Conflict Handling: The separation axis theorem is used to detect whether the extended bounding boxes of different clusters geometrically overlap in three-dimensional space; if the first axis is detected... Cluster and the first If clusters overlap, compare their volumes, select the larger cluster, and proceed with the next cluster according to a preset step size. Iteratively reduces its expansion factor, i.e. And recalculate its extended half-length vector until the bounding boxes of all clusters in the space do not overlap, in order to eliminate geometric ambiguity in multipath association.

5. A multipath clustering and tracking method for dynamic wireless channels according to claim 4, characterized in that, In S3, the motion state equations of the cluster are established using Kalman filtering, including: Definition of the first Clusters in State vector at time step It is a six-dimensional vector containing three-dimensional centroid coordinates. and three-dimensional center of mass velocity ,Right now ; Based on the assumption of uniform motion, a time update equation for the Kalman filter is established for use in... State prediction at time Prior state at time : ； and the prior error covariance matrix : ； Wherein, the state transition matrix process noise covariance matrix The specific form is: ； In the formula, It is a third-order identity matrix. The time interval between two adjacent snapshots; The acceleration noise variance diagonal matrix, comprising three components—time delay, azimuth angle of arrival, and elevation angle of arrival—is used to characterize the uncertainty introduced by non-uniform drift during cluster motion; the predicted prior state... The position component in the equation is the predicted centroid position.

6. A multipath clustering and tracking method for dynamic wireless channels according to claim 5, characterized in that, In step S4, multi-feature matching similarity is calculated. ,include: The first Clusters in The geometry of the oriented bounding box at time step is translated to the Kalman filter. Construct an observation window based on the predicted centroid position at each moment; Calculate the statistical characteristics of the measurement data at the current time within the observation window, and compare them with... The features at each time point are compared to calculate the shape similarity. and density similarity ; shape similarity Based on normalized volume difference Calculate, the expression is ,in, and These are the volumes of the bounding boxes at the current and previous time steps, respectively. density similarity Based on normalized density difference Calculate, the expression is ,in, and These represent the multipath point densities within the bounding box at the current and previous time steps, respectively. The multi-feature matching similarity is the product of the two features, expressed as follows: ;like Greater than the preset survival threshold If the current observation is obtained, the cluster is determined to be alive at the current moment, and the state of the Kalman filter is updated using the current observation.

7. A multipath clustering and tracking method for dynamic wireless channels according to claim 6, characterized in that, In step S4, a serial interference cancellation strategy is executed, including: S41. Based on the predicted bounding box of the surviving cluster, determine the explained region in the current snapshot that belongs to the surviving cluster; S42. Traverse all the original multipath components in the current snapshot and check whether they fall within the predicted bounding box of any of the above-mentioned live clusters. S43. Treat the multipath components that fall within the range of surviving clusters as old cluster components and remove them from the original dataset to eliminate the shading effect and interference of strong surviving clusters on potential weak new clusters. S44. The remaining unremoved multipath components are taken as residual components and input into the adaptive neighborhood robust mean drift algorithm for independent clustering. S45. For each new cluster generated by clustering, initialize a new Kalman filter trajectory, initialize the velocity component in its state vector to zero, and mark it as a new cluster, adding it to the tracking list for the next time step.

Images