Sensing data determination methods, devices and network equipment

By constructing an observation tensor and performing tensor decomposition and weighted fusion, the problem of insufficient adaptability of sensing data in the fusion of different frequency bands is solved, and higher accuracy and robustness of target parameter estimation are achieved.

CN123123948APending Publication Date: 2026-11-17BEIJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611334262.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, it is difficult to achieve adaptive fusion of sensing data across different frequency bands, resulting in insufficient preservation of multi-dimensional amplitude and phase structure and joint subspace information in cross-frequency band echoes, which affects the accuracy and robustness of target parameter estimation.

Method used

By constructing an observation tensor, performing tensor decomposition to determine the reference frequency band and teacher reference subspace, updating the decomposed data, and performing weighted fusion, multi-dimensional adaptive weighted fusion is achieved by leveraging the complementary advantages of high and low frequency bands.

Benefits of technology

It improves the accuracy and robustness of joint estimation of distance, velocity and angle, and makes full use of the complementary advantages of high and low frequency bands in different perception dimensions, thereby enhancing the performance of multi-band target parameter estimation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123948A_ABST
    Figure CN123123948A_ABST
Patent Text Reader

Abstract

This application provides a method, apparatus, and network device for determining sensing data. The method includes: constructing an observation tensor based on received sensing echo data of at least two frequency bands for at least one target object; the observation tensor includes observation tensors corresponding to the at least two frequency bands; performing tensor decomposition on the observation tensor to determine decomposed data and a reference frequency band; the reference frequency band is the frequency band with the highest reliability across multiple sensing dimensions; constructing a teacher reference subspace for a guiding frequency band based on the sensing data of multiple dimensions corresponding to the reference frequency band; updating the decomposed data through the teacher reference subspace to obtain target decomposed data corresponding to multiple frequency bands; and performing weighted fusion on the target decomposed data to obtain target sensing data for the at least one target object; the weights of the weighted fusion are related to the reliability of the target decomposed data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of sensor integration technology, specifically to a method, apparatus, and network device for determining sensing data. Background Technology

[0002] In Integrated Sensing and Communication (ISAC) technology, the sensing data typically utilizes multi-band complementary information by first processing the echoes of each frequency band individually, and then fusing the processing results from different frequency bands. In related technologies, the fusion point usually occurs after the individual frequency band processing, or relies on specific parameter ratios between the frequency bands, making it difficult to fully preserve the multidimensional amplitude and phase structure and joint subspace information in the cross-frequency band echoes.

[0003] For example, in data-level fusion, the original observation information across frequency bands has been compressed into a small number of parameters during the single-frequency estimation process. When an erroneous peak or a large estimation deviation occurs in a certain frequency band, subsequent fusion is difficult to recover the lost effective information. In symbol-level fusion, cross-frequency band collaboration usually still occurs after each frequency band extracts features separately, failing to enter the multidimensional decomposition process of joint representation of distance, velocity, and angle. In spectrum-level fusion, the spectral estimation error of a single frequency band will first enter its respective spectrum and then be brought into the fusion process. Furthermore, fixed superposition or fixed weights are difficult to reflect the reliability differences of each frequency band in different parameter dimensions, making it difficult to achieve adaptive fusion for each dimension. Summary of the Invention

[0004] At least one embodiment of this application provides a method, apparatus, and network device for determining sensing data, which addresses the problem in the prior art that it is difficult to achieve adaptive fusion of sensing data for different frequency bands across various dimensions.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows:

[0006] In a first aspect, embodiments of this application provide a method for determining sensing data, including:

[0007] An observation tensor is constructed based on received sensing echo data of at least two frequency bands for at least one target object; the observation tensor includes the observation tensors corresponding to the at least two frequency bands.

[0008] The observed tensor is decomposed to determine the decomposed data and the reference frequency band; the decomposed data includes: feature matrices and core tensors corresponding to data in multiple sensing dimensions; the reference frequency band is the frequency band with the highest reliability in the multiple sensing dimensions.

[0009] Based on the multi-dimensional perception data corresponding to the reference frequency band, a teacher reference subspace for the guidance frequency band is constructed;

[0010] The decomposed data is updated by the teacher reference subspace to obtain target decomposed data corresponding to multiple frequency bands;

[0011] The target decomposition data is weighted and fused to obtain target perception data of the at least one target object; the weight of the weighted fusion is related to the reliability of the target decomposition data.

[0012] Furthermore, based on the received sensing echo data in at least two frequency bands concerning at least one target object, an observation tensor is constructed, including:

[0013] The sensed echo data is subjected to pilot removal and normalization to obtain standard sensed data;

[0014] Feature parameters are extracted from the standard sensing data based on the subcarrier spacing, orthogonal frequency division multiplexing symbol length, and the arrangement of the receiving antennas; the feature parameters include: distance features, velocity features, and angle features;

[0015] The observation tensor is constructed based on the characteristic parameters.

[0016] Further, the observed tensor is decomposed to determine the decomposed data, including:

[0017] The observed tensor is estimated using the minimum description length method to obtain the modulus rank corresponding to multiple feature parameters;

[0018] The principal component analysis model is initialized using a higher-order orthogonal iteration method to obtain the target decomposition model;

[0019] The observed tensor is decomposed using the target decomposition model to obtain the decomposed data.

[0020] Further, the observed tensor is decomposed to determine the decomposed data and the reference frequency band, including:

[0021] Based on the decomposed data, construct the projection matrix of the ideal response subspace corresponding to the feature parameters;

[0022] Based on the projection matrix, determine the subspace of data for each of the perception dimensions;

[0023] Based on the residuals of the subspace, the estimated values ​​of the data for each sensing dimension corresponding to the sensing data of each frequency band are determined;

[0024] The reliability of the estimated value is calculated, and the reference frequency band is determined based on the reliability of the estimated value.

[0025] Further, calculating the reliability of the estimated value includes:

[0026] The root mean square standard deviation of the local uncertainty of the estimated value is calculated based on the covariance of the Jacobian matrix approximation parameter of the fitted residual.

[0027] Perform multiple random samplings with replacement on the subspace and calculate the perturbation dispersion of the median result for different sampling results;

[0028] The estimated value is optimized, and the attraction domain consistency of the estimated value is calculated;

[0029] The reliability of the estimated value is determined based on the root mean square standard deviation, the perturbation dispersion, and the consistency of the attraction domain.

[0030] Furthermore, by updating the decomposed data through the teacher reference subspace, target decomposed data corresponding to multiple frequency bands are obtained, including:

[0031] The teacher reference subspace is added as a guiding term to the objective function to update the feature matrix and the core tensor, thereby obtaining the target decomposition data;

[0032] The objective function includes a data reconstruction term and a reference-guided regularization term.

[0033] Further, the target decomposition data is weighted and fused to obtain target perception data for the at least one target object, including:

[0034] Determine the concentration of the peak value of the target decomposition data corresponding to each frequency band with respect to the peak value distribution of all target decomposition data;

[0035] Determine the degree of fit of the target decomposition data for each frequency band with respect to the corresponding observation tensor;

[0036] Calculate the normalized reconstruction residuals of the fit;

[0037] The weights of the weighted fusion are determined based on the concentration and the normalized reconstruction residuals.

[0038] Based on the weights of the weighted fusion, the target decomposition data is weighted and fused to obtain target perception data of the at least one target object.

[0039] Furthermore, the method also includes:

[0040] Based on the fusion result obtained by weighted fusion of the target decomposition data, candidate sensing data is extracted;

[0041] Determine the set of indices for the candidate sensor data across all sensory dimensions;

[0042] Construct a candidate response matrix for the sensing data corresponding to each frequency band;

[0043] By fusing the frequency band correlation tensors of multiple frequency bands, a joint correlation tensor is obtained; each element in the joint correlation tensor represents the fusion correlation strength corresponding to the index.

[0044] Select a preset number of target indices with the highest fusion correlation strength from the set of indices;

[0045] Based on the target index, the corresponding parameter values ​​are read from the candidate sensing data, and the estimation result of the target sensing data is output.

[0046] Secondly, embodiments of this application provide a sensing data determination device, comprising:

[0047] The first construction module is used to construct an observation tensor based on received sensing echo data of at least two frequency bands for at least one target object; the observation tensor includes the observation tensors corresponding to the at least two frequency bands.

[0048] The decomposition module is used to perform tensor decomposition on the observed tensor to determine the decomposed data and the reference frequency band; the decomposed data includes: feature matrices and core tensors corresponding to data in multiple sensing dimensions; the reference frequency band is the frequency band with the highest reliability in the multiple sensing dimensions.

[0049] The second construction module is used to construct a teacher reference subspace for the guidance frequency band based on the multiple dimensions of perception data corresponding to the reference frequency band.

[0050] The update module is used to update the decomposed data through the teacher reference subspace to obtain target decomposed data corresponding to multiple frequency bands;

[0051] The fusion module is used to perform weighted fusion on the target decomposition data to obtain target perception data of the at least one target object; the weight of the weighted fusion is related to the reliability of the target decomposition data.

[0052] Thirdly, embodiments of this application provide a network device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the method described in the first aspect.

[0053] Compared with existing technologies, the sensing data determination method, apparatus, network device, and storage medium provided in this application introduce a teacher subspace provided by a reference frequency band during the decomposition process of the guided frequency band. This allows the guided frequency band to update itself while interpreting its own observation data, referencing a more reliable subspace structure. This enables cross-band information to participate in processing before the single-frequency decomposition error is fixed, thereby reducing the impact of noise, sidelobes, multi-target coupling, or frequency-selective scattering on the decomposition results and improving the accuracy and robustness of joint estimation of range, velocity, and angle. Simultaneously, it can perform multi-dimensional adaptive weighted fusion based on the actual decomposition quality of the current data, thus making fuller use of the complementary advantages of high and low frequency bands in different sensing dimensions and improving the performance of multi-band target parameter estimation. Attached Figure Description

[0054] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0055] Figure 1 This is a schematic diagram illustrating the steps of the sensing data determination method according to an embodiment of this application;

[0056] Figure 2 This is a schematic diagram of the module of the sensing data determination device according to an embodiment of this application;

[0057] Figure 3 This is a schematic diagram of the structure of a network device according to an embodiment of this application. Detailed Implementation

[0058] The terms "first," "second," etc., used in this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, without limiting the number of objects; for example, the first object can be one or more. Furthermore, "or" in this application indicates at least one of the connected objects. For example, "A or B" covers three scenarios: Scenario 1: including A but not B; Scenario 2: including B but not A; Scenario 3: including both A and B. The character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0059] The term "instruction" in this application can be either a direct instruction (or explicit instruction) or an indirect instruction (or implicit instruction). A direct instruction can be understood as one in which the sender explicitly informs the receiver of specific information, the operation to be performed, or the requested result, etc.; an indirect instruction can be understood as one in which the receiver determines the corresponding information based on the instruction sent by the sender, or makes a judgment and determines the operation to be performed or the requested result, etc., based on the judgment result.

[0060] To enable those skilled in the art to better understand the embodiments of this application, the following description is provided first:

[0061] The basic idea of ​​ISAC is to simultaneously complete communication and environmental awareness within the same wireless system. Taking a multi-input multi-output orthogonal frequency division multiplexing (MIMO-OFDM) signal as an example, the target echo will produce different phase changes in three directions: subcarrier, OFDM symbol, and receiving antenna. These changes contain the target's time delay, Doppler, and angle information, respectively.

[0062] High-frequency bands typically offer higher range, velocity, or angular resolution; low-frequency bands suffer less propagation loss and are often more stable under low signal-to-noise ratio or obstruction conditions. However, the carrier frequency, bandwidth, subcarrier spacing, symbol length, and antenna size of the two bands are usually different, so combining the advantages of different bands cannot be achieved simply by directly splicing the echoes.

[0063] In carrier aggregation or multi-band sensing integrated systems, different frequency bands typically exhibit varying range resolution, velocity sensitivity, angle resolution, and noise immunity due to differences in carrier frequency, bandwidth, subcarrier spacing, symbol duration, and array response. Related technologies generally leverage the complementarity between multiple frequency bands to improve target parameter estimation performance. Common approaches include fusion at the parameter estimation result level, fusion at the complex symbol or eigenvector level, and fusion at the range spectrum, velocity spectrum, or range-velocity spectrum level. While these methods can utilize multi-band observation information to some extent, their fusion location is usually after single-band processing or relies on specific frequency band parameter ratios, making it difficult to fully preserve the multidimensional amplitude and phase structure and joint subspace information in cross-band echoes. Therefore, existing multi-band fusion methods can still be broadly categorized into three types: data-level fusion, symbol-level fusion, and spectrum-level fusion. Specifically:

[0064] 1. Data-level fusion

[0065] Data-level fusion is a relatively straightforward method for multi-band target parameter estimation. This type of method typically involves first inputting echo signals from different frequency bands into their respective parameter estimation modules, independently estimating range, velocity, or angle within each band. Then, the estimated values ​​from multiple frequency bands are weighted and fused at the parameter result level. For example, one scheme processes high-frequency and low-frequency echoes separately, first obtaining the target's range and velocity estimates in both frequency bands, and then calculating the fusion weights based on indicators such as estimation accuracy, received signal-to-noise ratio, or noise variance for each frequency band, thereby obtaining the final target parameter estimate.

[0066] The advantages of this type of scheme are its simplicity; the fusion end only needs to process a small number of physical parameters output from each frequency band, resulting in low computational complexity and ease of integration with existing single-frequency sensing algorithms. However, since data-level fusion occurs after single-frequency band parameter estimation, the original complex phase, amplitude correlation, and multi-dimensional coupling structures between distance, velocity, and angle in the echoes from different frequency bands have already been compressed into a small number of parameter results during the single-frequency estimation process. In other words, the fusion module can only correct the already formed estimation results and cannot reuse the original observation information across frequency bands. When a frequency band produces erroneous peaks or large estimation biases due to noise, sidelobes, multi-target coupling, or insufficient resolution, subsequent weighted averaging often fails to recover the lost effective information and may even further propagate the erroneous estimation to the final result. Therefore, data-level fusion cannot fully leverage the complementary advantages of multi-frequency band echoes in joint parameter estimation.

[0067] 2. Symbol-level fusion

[0068] Symbol-level fusion combines complex symbols, echo features, or intermediate eigenvectors from different frequency bands before physical parameter estimation, and then completes target parameter estimation based on the fused features. Compared with data-level fusion, this type of method retains more echo amplitude and phase information. For example, one existing scheme first performs angle compensation, spatial filtering, and cyclic cross-correlation processing on high- and low-frequency echoes to correct initial phase differences between different frequency bands; then, it extracts time delay eigenvectors and Doppler eigenvectors from the echoes of each frequency band; then, based on the parameter relationships between subcarrier spacing, carrier frequency, and symbol duration in different frequency bands, it rearranges, pads with zeros, averages, or recovers missing elements from the eigenvectors; finally, it achieves cross-frequency band parameter estimation by searching for distance and velocity peaks using an improved Inverse Discrete Fourier Transform / Discrete Fourier Transform (IDFT / DFT).

[0069] Compared to data-level fusion, this type of scheme can retain some complex feature information before parameter estimation, thus generally having better fusion potential in range or velocity estimation. However, this scheme often requires specific parameter ratios between different frequency bands, such as alignable, rearrangeable, or zero-padded structural relationships between subcarrier spacing, symbol periods, or carrier frequencies; otherwise, feature vector reconstruction and joint search become complex. Furthermore, existing symbol-level fusion typically still extracts range or velocity-related features separately within each frequency band before performing cross-frequency combination; cross-frequency band collaboration does not directly enter the multi-dimensional subspace decomposition process of joint representation of range, velocity, and angle. Therefore, although this type of scheme utilizes multi-band information earlier than data-level fusion, it still struggles to fully exploit the joint structure of different frequency bands in the multi-dimensional parameter space.

[0070] 3. Spectral fusion

[0071] Spectral fusion refers to first performing spectral analysis on the echo signals of each frequency band separately to obtain spectral results such as range spectrum, velocity spectrum, angle spectrum, or range-velocity spectrum. Then, the spectra of different frequency bands are mapped to a unified physical coordinate system for superposition, weighting, or joint decision-making. Specifically, if the subcarrier spacing, symbol period, or carrier frequency of different frequency bands meet integer multiples, approximate integer multiples, or other alignable relationships, spectral point alignment can be performed directly. If the frequency band parameters are not perfectly matched, it is usually necessary to transform the spectra of different frequency bands to a unified range, velocity, or angle coordinate system through interpolation, averaging, nearest neighbor mapping, or coordinate transformation before fusion processing.

[0072] Spectral-level fusion is susceptible to errors in single-band spectrum estimation. For example, when a frequency band has strong sidelobes, noise peaks, overlapping peaks of multiple targets, or subspace bias, these errors will first be reflected in the spectrum of that band and then incorporated into the fusion process. If fixed stacking or fixed weights are used, the reliability differences between different frequency bands in the three dimensions of distance, velocity, and angle are difficult to accurately characterize. For instance, high-frequency bands may have advantages in angle or distance resolution, while low-frequency bands may have advantages in anti-occlusion or robustness; the reliability of the same frequency band is not consistent across different parameter dimensions. Existing spectral-level fusion methods typically struggle to adaptively fuse based on actual reliability differences in the multi-dimensional parameter space, thus their adaptability to complex multi-target scenarios and low signal-to-noise ratio scenarios remains limited.

[0073] Although the three schemes mentioned above have different fusion levels, their overall processing order is quite similar. They typically perform preprocessing, feature extraction, parameter estimation, or spectrum construction separately in the high-frequency and low-frequency bands, and then fuse the results obtained from different frequency bands. In other words, in existing schemes, the collaboration between high-frequency and low-frequency bands mostly occurs after single-band processing, without fully introducing cross-band complementary information during the multi-dimensional sensing information decomposition process.

[0074] Data-level fusion schemes typically first estimate physical parameters such as range, velocity, or angle independently within each frequency band, and then perform weighted fusion of the parameter estimation results from different frequency bands. This type of scheme is simple to implement and has low computational complexity. However, because fusion occurs after single-band parameter estimation, the complex phase and amplitude correlations in the original echoes, as well as the multi-dimensional coupling structures between range, velocity, and angle, have been compressed into a small number of parameter results. Therefore, when a frequency band produces erroneous peaks or large estimation deviations due to noise, sidelobes, multi-target coupling, or insufficient resolution, subsequent weighted fusion struggles to recover the lost effective information, thus limiting the utilization efficiency of multi-band complementary information and the accuracy of target parameter estimation.

[0075] Symbol-level fusion schemes typically combine complex symbols, echo features, or intermediate eigenvectors from different frequency bands before parameter estimation, preserving more amplitude and phase information compared to data-level fusion. However, existing symbol-level fusion schemes usually require correction and matching of initial phase, subcarrier spacing, symbol period, or carrier frequency ratios between different frequency bands, relying on eigenvector rearrangement, zero-padding, averaging, or missing element recovery steps. When parameters in different frequency bands do not meet specific ratios, feature alignment and joint processing become challenging. Furthermore, these schemes typically still extract range or velocity-related features separately within each frequency band before cross-frequency combination; cross-frequency band cooperation does not directly enter the multidimensional tensor quantum space decomposition process composed of subcarriers, OFDM symbols, and receiving antennas. Therefore, their utilization of the joint structure of range, velocity, and angle remains insufficient.

[0076] Spectral fusion schemes typically first calculate the range, velocity, and angle spectra or range-velocity spectra for each frequency band separately, and then map the spectra of different frequency bands onto a unified physical coordinate system for superposition, weighting, or joint decision-making. This type of scheme is susceptible to errors in single-band spectrum estimation. For example, when a frequency band has strong sidelobes, noise peaks, overlapping peaks of multiple targets, or subspace bias, these errors will first be reflected in the single-spectrum graph and then carried over into the subsequent fusion process. Furthermore, fixed superposition or fixed weights are insufficient to reflect the reliability differences of the same frequency band across different dimensions of range, velocity, and angle, making it difficult to adaptively adjust the fusion results based on actual perceived quality.

[0077] Based on the problems existing in the prior art, this application mainly addresses the following technical issues: First, how to introduce complementary sensing information from high-frequency and low-frequency bands into the multidimensional tensor decomposition process, so that cross-band collaboration is no longer limited to single-frequency parameter estimation, feature extraction, or spectrum construction, thereby utilizing the joint low-rank structure of multi-frequency echoes at an earlier stage; Second, how to calculate the corresponding fusion weights based on the actual decomposition quality of different frequency bands in distance, velocity, and angle dimensions after tensor decomposition, thereby achieving adaptive weighted fusion for different sensing dimensions.

[0078] This application provides a method and apparatus for determining sensing data. The method and apparatus are based on the same concept, and since they solve problems based on similar principles, their implementations can be mutually referenced; repeated details will not be repeated.

[0079] like Figure 1 As shown in the figure, an embodiment of this application provides a method for determining sensing data, which includes the following steps:

[0080] Step 101: Construct an observation tensor based on the received sensing echo data of at least two frequency bands for at least one target object; the observation tensor includes the observation tensors corresponding to the at least two frequency bands.

[0081] Step 102: Perform tensor decomposition on the observed tensor to determine the decomposed data and reference frequency band; the decomposed data includes: feature matrices and core tensors corresponding to data in multiple sensing dimensions; the reference frequency band is the frequency band with the highest reliability in the multiple sensing dimensions;

[0082] Step 103: Construct a teacher reference subspace for the guidance frequency band based on the multi-dimensional perception data corresponding to the reference frequency band;

[0083] Step 104: Update the decomposed data through the teacher reference subspace to obtain target decomposed data corresponding to multiple frequency bands;

[0084] Step 105: Perform weighted fusion on the target decomposition data to obtain target perception data of the at least one target object; the weight of the weighted fusion is related to the reliability of the target decomposition data.

[0085] Optionally, the at least two frequency bands include a low frequency band (e.g., 2.6 GHz, 3.5 GHz, etc.) and a high frequency band (e.g., 28 GHz, 73 GHz, 140 GHz, etc.).

[0086] The sensing data determination method in this application introduces a teacher subspace provided by a reference frequency band during the decomposition process of the guided frequency band. This allows the guided frequency band to update itself while referencing a more reliable subspace structure, enabling cross-frequency band information to participate in processing before the single-frequency decomposition error is fixed. This reduces the impact of noise, sidelobes, multi-target coupling, or frequency-selective scattering on the decomposition results, improving the accuracy and robustness of joint estimation of range, velocity, and angle. Simultaneously, it can perform multi-dimensional adaptive weighted fusion based on the actual decomposition quality of the current data, thereby making fuller use of the complementary advantages of high and low frequency bands in different sensing dimensions and improving the performance of multi-frequency band target parameter estimation.

[0087] It should be noted that the embodiments of this application are applicable to a sensor-integrated collaborative sensing system, such as a system operating in... Cooperative sensing systems on discontinuous frequency bands. The b-th frequency band contains... Root transmitting antenna, Root receiving antenna, Subcarriers and L OFDM symbols. The system simultaneously observes L targets through beam scanning. The MIMO-OFDM transmission signal in the b-th frequency band is:

[0088] ;

[0089] in, Indicates the transmission power. Represents the nth subcarrier, ... Pilots on OFDM symbols and These represent the carrier frequency and subcarrier spacing, respectively. Indicates the symbol duration. The received echo of the b-th frequency band on the n-th subcarrier and the m-th OFDM symbol is:

[0090] ;

[0091] in, , and These represent time delay, Doppler shift, and angle, respectively. Indicates the initial phase related to the frequency band. and These represent the return complex gain and the transmit beam gain, respectively. Includes noise and other reflection interference. Transmission steering vector. and receiving guide vector for:

[0092] ;

[0093] in , and These represent the element spacing and wavelength, respectively.

[0094] Optionally, based on received sensing echo data in at least two frequency bands concerning at least one target object, an observation tensor is constructed, including:

[0095] The sensed echo data is subjected to pilot removal and normalization to obtain standard sensed data;

[0096] Feature parameters are extracted from the standard sensing data based on the subcarrier spacing, orthogonal frequency division multiplexing symbol length, and the arrangement of the receiving antennas; the feature parameters include: distance features, velocity features, and angle features;

[0097] The observation tensor is constructed based on the characteristic parameters.

[0098] In this embodiment of the application, the pilot removal and normalization processing of the sensed echo data to obtain standard sensed data is specifically as follows:

[0099] In this embodiment of the application, after removing the pilot signals and normalizing, we follow the subcarrier... OFDM symbol and receiving antenna Arrange the data to obtain the third-order observation tensor. Specifically:

[0100] ;

[0101] in, , and These correspond to time delay, Doppler response, and angular response, respectively.

[0102] It should be noted that the carrier frequency, subcarrier spacing, symbol duration, and antenna configuration can differ across frequency bands. For example, the subcarrier spacing in high-frequency bands can be greater than that in low-frequency bands, thus offering an advantage in range resolution; while low-frequency bands have lower propagation loss and higher signal-to-noise ratios under low or obstructed conditions. The construction method of the observation tensor (arranged by subcarrier, OFDM symbol, and receiving antenna) does not depend on the proportional relationship between frequency band parameters, therefore, there is no need to align or map parameters for different frequency bands.

[0103] Optionally, the observed tensor is decomposed to determine the decomposed data, including:

[0104] The observed tensor is estimated using the minimum description length method to obtain the modulus rank corresponding to multiple feature parameters;

[0105] The principal component analysis model is initialized using a higher-order orthogonal iteration method to obtain the target decomposition model;

[0106] The observed tensor is decomposed using the target decomposition model to obtain the decomposed data.

[0107] It should be noted that decomposing the observed tensor requires establishing a decomposition model, which is the Tucker decomposition model.

[0108] For example, Tucker models are established for the high-frequency band and the low-frequency band respectively, specifically:

[0109] ;

[0110] in, It is the core tensor. , and These represent feature matrices for the three perception dimensions (e.g., distance, velocity, and angle subspaces). , and It is the corresponding modulo rank. This represents the inner product. The three feature matrices respectively cover the target's time delay, Doppler response, and angular response. The inclusion relationship between the three feature matrices and the physical time delay, Doppler response, and angular response subspaces is as follows:

[0111] ;

[0112] The baseline optimization problem of band-by-band independent Tucker decomposition is as follows:

[0113] ;

[0114] in This represents the Frobenius operator. Traditional methods minimize the reconstruction errors of the two frequency bands separately, thus reliable information from one band does not participate in the feature update of the other. Therefore, we add a teacher-guided term in addition to the reconstruction error. The specific objective of the high-low frequency complementary Tucker decomposition after introducing the teacher reference subspace is:

[0115] ;

[0116] The set of optimization variables in the complementary Tucker decomposition is represented as follows:

[0117] ;

[0118] The teacher-guided regularization terms for the three dimensions of distance, speed, and angle are represented as follows:

[0119] ;

[0120] and , and It is a three-dimensional teacher reference subspace. Determine the current frequency band in dimension The extent of guidance received and the method for obtaining its values ​​will be described in detail in the next section.

[0121] In this embodiment of the application, firstly... It unfolds along three modes. The result is... .in, It is the row number of the expanded matrix. This is the number of snapshots that can be used to estimate covariance. The row dimension and number of snapshots for each modulus expansion matrix are expressed as:

[0122] ;

[0123] The empirical covariance matrix for each frequency band and each sensing dimension is expressed as follows:

[0124] ;

[0125] Subsequently, the Minimum Description Length (MDL) score is calculated based on the eigenvalues ​​of the covariance matrix, yielding the modulus rank for each frequency band and each dimension. The maximum value among all estimated modulus ranks is taken as the number of distinguishable targets. Then initialize the two frequency bands using HOOI. Specifically:

[0126] 1) For candidate modulo rank The corresponding minimum description length score is:

[0127] ;

[0128] 2) The estimated modulo-rank of each frequency band and each sensing dimension is finally expressed as:

[0129] ;

[0130] 3) The final number of distinguishable targets can be expressed as: ;

[0131] 4) Based on the estimated number of distinguishable targets Using the HOOI algorithm to perform Tucker decomposition on the two frequency bands respectively, the initial core tensor and subspace are obtained, which can be represented as:

[0132] .

[0133] Optionally, the observed tensor is decomposed to determine the decomposed data and the reference frequency band, including:

[0134] Based on the decomposed data, construct the projection matrix of the ideal response subspace corresponding to the feature parameters;

[0135] Based on the projection matrix, determine the subspace of data for each of the perception dimensions;

[0136] Based on the residuals of the subspace, the estimated values ​​of the data for each sensing dimension corresponding to the sensing data of each frequency band are determined;

[0137] The reliability of the estimated value is calculated, and the reference frequency band is determined based on the reliability of the estimated value.

[0138] In this embodiment, to compare the initial subspaces of two frequency bands, initial physical parameters need to be obtained first. The ideal response vectors corresponding to range, velocity, and angle explicitly include the subcarrier spacing, carrier frequency, symbol duration, and array parameters for each frequency band. Therefore, the ideal physical response vectors for the three dimensions of range, velocity, and angle can be expressed as:

[0139] ;

[0140] Bundle After assembling a matrix from the candidate responses, a projection matrix of the ideal response subspace can be constructed. Initial parameter estimates are obtained by minimizing the residuals of the initial Tucker subspace on this projected orthogonal complement. Specifically:

[0141] 1) The ideal response matrix composed of candidate parameters is expressed as:

[0142] ;

[0143] The subspace projection matrix corresponding to the ideal response matrix is ​​expressed as:

[0144] ;

[0145] 2) Initial parameter estimates can be obtained by minimizing the subspace residuals:

[0146] .

[0147] Optionally, calculating the reliability of the estimated value includes:

[0148] The root mean square standard deviation of the local uncertainty of the estimated value is calculated based on the covariance of the Jacobian matrix approximation parameter of the fitted residual.

[0149] Perform multiple random samplings with replacement on the subspace and calculate the perturbation dispersion of the median result for different sampling results;

[0150] The estimated value is optimized, and the attraction domain consistency of the estimated value is calculated;

[0151] The reliability of the estimated value is determined based on the root mean square standard deviation, the perturbation dispersion, and the consistency of the attraction domain.

[0152] In this embodiment of the application, the reliability of the initial estimate is determined from three perspectives, specifically:

[0153] First, the root mean square standard deviation of the local uncertainty is obtained by approximating the parameter covariance using the Jacobian matrix of the fitted residuals. The local covariance of the parameter estimation is approximated as:

[0154] ;

[0155] The root mean square standard deviation of the local uncertainty is:

[0156] ;

[0157] Second, perform multiple Bootstrap resampling operations on the rows of the subspace and repeat parameter fitting. The more concentrated the center of the resampling results relative to the median, the more stable the estimate is to observational perturbations. The row-constrained subspace and response matrix in the Bootstrap experiment are represented as follows:

[0158] ;

[0159] No. The Bootstrap parameter estimates are expressed as follows:

[0160] ;

[0161] The element-wise median and perturbation dispersion estimated by Bootstrap are expressed as follows:

[0162] and ;

[0163] Third, the proportion of optimization results falling into the neighborhood of the master solution is statistically analyzed by repeatedly optimizing from multiple initial values. This index is defined as the consistency of the attraction domain of the multi-initial-value estimation, expressed as:

[0164]

[0165] in, Indicator functions are generally represented as: ;

[0166] The nominal resolution scales for the three dimensions of distance, velocity, and angle are expressed as follows:

[0167] ;

[0168] Finally, local uncertainty, Bootstrap dispersion, and attraction field consistency are combined into... The smaller the value, the stronger the frequency band. In dimensions The more reliable the estimate, the better. Specifically, it is expressed as follows:

[0169] .

[0170] In this embodiment of the application, the teacher reference subspace of the guidance frequency band is constructed based on the multiple dimensions of perception data corresponding to the reference frequency band, which can be:

[0171] Comparing high and low frequency bands in each dimension If the difference between the two is very small, neither frequency band is forced to guide the other; if the difference exceeds a threshold, the frequency band with the lowest reliability index (highest reliability) is selected as the winning frequency band (the reference frequency band). The selection method can be expressed as follows:

[0172] ;

[0173] Once the winning frequency band is determined, the guided frequency band can be represented as:

[0174] ;

[0175] The parameters of the winning frequency band are then substituted into the physical response function of the guided frequency band to obtain the mapped response matrix:

[0176]

[0177] Gram-Schmidt orthogonalization of the mapped response matrix:

[0178] ;

[0179] After orthogonalization, a teacher reference subspace matching the dimension of the guided frequency band is obtained. The more significant the reliability difference between the two frequency bands, the greater the guidance weight, which is specifically expressed as:

[0180] ;

[0181] Ultimately, the fractal dynamic superior frequency band mechanism outputs... .

[0182] Optionally, the decomposed data is updated through a teacher reference subspace to obtain target decomposed data corresponding to multiple frequency bands, including:

[0183] The teacher reference subspace is added as a guiding term to the objective function to update the feature matrix and the core tensor, thereby obtaining the target decomposition data;

[0184] The objective function includes a data reconstruction term and a reference-guided regularization term.

[0185] In this embodiment, after obtaining the teacher reference subspace and guidance weights for each perceptual dimension, the present invention uses an alternating optimization method to update the Tucker decomposition results for high and low frequency bands. To simplify subsequent expressions, frequency bands are defined. In dimensions The effective guiding strength on is For the The next iteration follows these steps:

[0186] (1) Update the distance factor matrix

[0187] In the In this iteration, the velocity feature matrix, angle feature matrix, and core tensor are fixed, and only the distance feature matrix is ​​updated. First, construct the coefficient matrix corresponding to the distance modulus expansion:

[0188] ;

[0189] in, Let Kronecker product be represented. The teacher-guided regularization subproblem of the distance factor is:

[0190] ;

[0191] Its unorthogonalized closed-form solution is:

[0192] ;

[0193] Subsequently Orthogonalization is performed to obtain a new distance factor matrix. .

[0194] (2) Update the velocity factor matrix

[0195] With the updated distance factor matrix, current angle factor matrix, and core tensor fixed, construct the coefficient matrix corresponding to the velocity magnitude expansion:

[0196] ;

[0197] The teacher-guided regularization subproblem of the velocity factor is:

[0198] ;

[0199] Its unorthogonalized closed-form solution is:

[0200] ;

[0201] Subsequently Orthogonalization is performed to obtain a new velocity factor matrix. .

[0202] (3) Update the angle factor matrix

[0203] With the updated distance factor matrix, velocity factor matrix, and current core tensor fixed, construct the coefficient matrix corresponding to the angle magnitude expansion:

[0204] ;

[0205] The teacher-guided regularization subproblem of the angle factor is:

[0206] ;

[0207] Its unorthogonalized closed-form solution is:

[0208] ;

[0209] Subsequently Orthogonalization is performed to obtain a new angle factor matrix. .

[0210] (4) Update the core tensor

[0211] After updating the three factor matrices, the observation tensor is projected into the new distance, velocity, and angle subspaces, and updated in conjunction with the core tensor regularization term. :

[0212] ;

[0213] (5) Convergence judgment

[0214] After completing one round of updates to the distance factor, velocity factor, angle factor, and core tensor, calculate the... The objective function value for the next iteration:

[0215] ;

[0216] If the relative change of the objective function between two consecutive iterations satisfies:

[0217] ;

[0218] in, If a preset convergence threshold is met, iteration will stop; otherwise, the alternating update process described above will continue until the convergence condition is met or the maximum number of iterations is reached.

[0219] Optionally, the target decomposition data is weighted and fused to obtain target perception data of the at least one target object, including:

[0220] Determine the concentration of the peak value of the target decomposition data corresponding to each frequency band with respect to the peak value distribution of all target decomposition data;

[0221] Determine the degree of fit of the target decomposition data for each frequency band with respect to the corresponding observation tensor;

[0222] Calculate the normalized reconstruction residuals of the fit;

[0223] The weights of the weighted fusion are determined based on the concentration and the normalized reconstruction residuals.

[0224] Based on the weights of the weighted fusion, the target decomposition data is weighted and fused to obtain target perception data of the at least one target object.

[0225] In this embodiment, after the decomposition converges, the corresponding subspace projection matrices are constructed using the refined distance, velocity, and angle factor matrices, respectively, and the refined subspace spectrum is calculated within the corresponding parameter search interval. The refined subspace projection matrix after complementary decomposition convergence can be expressed as:

[0226] ;

[0227] Search range for corresponding parameters Inside, candidate parameters The corresponding physical response vector is projected onto the refined subspace to obtain the refined subspace spectral values ​​at the candidate parameter locations:

[0228] ;in, This indicates that frequency band b in dimension q corresponds to the candidate parameters. The corresponding physical response vector. After normalizing the refined subspace spectrum, the largest value is selected. One dominant peak is identified, resulting in a set of candidate peak locations for each frequency band and each sensing dimension:

[0229] ;

[0230] To evaluate the actual quality of the decomposition results for different frequency bands, the spectral fusion weights are composed of two posterior quality indices: normalized reconstruction residuals and spectral concentration. The normalized reconstruction residuals... The residual is used to measure how well the Tucker decomposition results fit the original observation tensor; the smaller the residual, the better the decomposition results can explain the current observation data; spectral concentration. This is used to measure the prominence of the dominant peak relative to the overall spectral background. The greater the concentration, the easier it is to distinguish the target candidate parameters in the corresponding dimension.

[0231] ;

[0232] ;

[0233] Finally, the spectral fusion weights are expressed as:

[0234] ;

[0235] Finally, the normalized subspace spectra of the two frequency bands are weighted and fused according to the fusion weights, and candidate parameters for range, velocity, and angle are extracted from the fused spectrum. The candidate parameters for range, velocity, or angle in the fused spectrum are extracted using spectral fusion weights.

[0236] ;

[0237] in, Let represent the fusion candidate parameter vector along dimension p. At this point, we can obtain... One distance candidate value, One speed candidate value and Candidate values ​​for each angle.

[0238] Optionally, the method further includes:

[0239] Based on the fusion result obtained by weighted fusion of the target decomposition data, candidate sensing data is extracted;

[0240] Determine the set of indices for the candidate sensor data across all sensory dimensions;

[0241] Construct a candidate response matrix for the sensing data corresponding to each frequency band;

[0242] By fusing the frequency band correlation tensors of multiple frequency bands, a joint correlation tensor is obtained; each element in the joint correlation tensor represents the fusion correlation strength corresponding to the index.

[0243] Select a preset number of target indices with the highest fusion correlation strength from the set of indices;

[0244] Based on the target index, the corresponding parameter values ​​are read from the candidate sensing data, and the estimation result of the target sensing data is output.

[0245] Since the candidate sensing data for distance, velocity, and angle are extracted separately from the fused spectrum, it is necessary to further determine which distances, velocities, and angles belong to the same target object. To this end, this application first enumerates all index combinations of the candidate sensing data for the three dimensions, then uses the candidate response matrix to fit the original observation tensor for each frequency band, and evaluates the degree of matching between different candidate sensing data (e.g., feature parameters of the distance, velocity, and angle triples) and the actual observation data based on the fitting results. Specifically:

[0246] (1) The complete index combination of candidate feature parameters for distance, velocity, and angle is specifically written as:

[0247] ;

[0248] (2) For each frequency band, candidate response matrices are constructed based on the fused distance, velocity, and angle candidate parameters, and their pseudo-inverses are used to perform multi-mode projection on the original observation tensor to obtain the frequency band correlation tensor:

[0249]

[0250] in, , and They represent frequency bands respectively. The response matrix is ​​constructed from candidate parameters of distance, velocity, and angle. This represents the Moore-Penrose pseudoinverse. Each parameter of the associated tensor... Depicting the The distance candidate value, the th distance candidate value, The candidate velocity value and the first The ability of multiple candidate angles to jointly interpret the observation data of the current frequency band.

[0251] Then, to ensure consistency across high and low frequencies for each candidate combination, the final candidate combinations need to be selected from the joint high and low frequency band correlation tensor, which is represented as:

[0252] ;

[0253] in, Indicates candidate triples The strength of the fusion correlation is considered. A stronger fusion correlation indicates that the corresponding combination of distance, velocity, and angle can simultaneously explain both high- and low-frequency observational data. Therefore, the triple with the strongest fusion correlation is selected from all candidate triples. Combining these elements yields the target index set:

[0254] ;

[0255] Finally, based on the target index set The corresponding distance, radial velocity, and angle are read from the fusion candidate parameters to obtain the first... Parameter estimation results for each target:

[0256] ;

[0257] Through the above processing, the present invention ultimately associates the extracted candidate values ​​of distance, velocity, and angle as follows: The system generates a triplet of target-level parameters and outputs the estimated distance, radial velocity, and angle for each target.

[0258] The various methods of the embodiments of this application have been described above. Apparatus for implementing the above methods will now be provided.

[0259] like Figure 2 As shown in the figure, this application embodiment also provides a sensing data determination device 200, including:

[0260] The first construction module 201 is used to construct an observation tensor based on received sensing echo data of at least two frequency bands for at least one target object; the observation tensor includes the observation tensors corresponding to the at least two frequency bands.

[0261] Decomposition module 202 is used to perform tensor decomposition on the observation tensor to determine the decomposed data and reference frequency band; the decomposed data includes: feature matrices and core tensors corresponding to data in multiple sensing dimensions; the reference frequency band is the frequency band with the highest reliability in the multiple sensing dimensions;

[0262] The second construction module 203 is used to construct a teacher reference subspace for the guidance frequency band based on the multiple dimensions of perception data corresponding to the reference frequency band.

[0263] The update module 204 is used to update the decomposed data through the teacher reference subspace to obtain target decomposed data corresponding to multiple frequency bands;

[0264] The fusion module 205 is used to perform weighted fusion on the target decomposition data to obtain target perception data of the at least one target object; the weight of the weighted fusion is related to the reliability of the target decomposition data.

[0265] The sensing data determination device provided in this application introduces a teacher subspace provided by a reference frequency band during the decomposition process of the guided frequency band. This allows the guided frequency band to update itself while referencing a more reliable subspace structure, enabling cross-frequency band information to participate in processing before the single-frequency decomposition error is fixed. This reduces the impact of noise, sidelobes, multi-target coupling, or frequency-selective scattering on the decomposition results, improving the accuracy and robustness of joint estimation of range, velocity, and angle. Simultaneously, it can perform multi-dimensional adaptive weighted fusion based on the actual decomposition quality of the current data, thereby making fuller use of the complementary advantages of high and low frequency bands in different sensing dimensions and improving the performance of multi-frequency band target parameter estimation.

[0266] This application also provides a network device, such as... Figure 3 As shown, it includes a transceiver 310, a processor 300, a memory 320, and a program or instructions stored in the memory 320 and executable on the processor 300; when the processor 300 executes the program or instructions, it implements the various processes of the above-described method embodiments on the network device side and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0267] The transceiver 310 is used to receive and send data under the control of the processor 300.

[0268] Among them, Figure 3 In this context, the bus architecture may include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 300) and memory (memory 320). The bus architecture may also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 310 may be multiple elements, including transmitters and receivers, providing a unit for communicating with various other devices over a transmission medium. The processor 300 is responsible for managing the bus architecture and general processing, and the memory 320 may store data used by the processor 300 during operation.

[0269] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the above-described sensing data determination method embodiments and achieves the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0270] This application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, they implement the various processes of the above-described perception data determination method embodiments and achieve the same technical effects. To avoid repetition, they will not be described again here.

[0271] It should be noted that, in this document, 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 a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0272] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0273] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A perception data determination method, comprising: include: Construct an observation tensor based on received sensing echo data in at least two frequency bands for at least one target object; The observation tensor includes the observation tensors corresponding to the at least two frequency bands; The observed tensor is decomposed to determine the decomposed data and the reference frequency band; The decomposed data includes: feature matrices and core tensors corresponding to data from multiple sensing dimensions; the reference frequency band is the frequency band with the highest reliability across the multiple sensing dimensions. Based on the multi-dimensional perception data corresponding to the reference frequency band, a teacher reference subspace for the guidance frequency band is constructed; The decomposed data is updated by the teacher reference subspace to obtain target decomposed data corresponding to multiple frequency bands; The target decomposition data is weighted and fused to obtain target perception data of the at least one target object; the weight of the weighted fusion is related to the reliability of the target decomposition data.

2. The method of claim 1, wherein, Based on received sensing echo data in at least two frequency bands concerning at least one target object, an observation tensor is constructed, including: The sensed echo data is subjected to pilot removal and normalization to obtain standard sensed data; Feature parameters are extracted from the standard sensing data based on the subcarrier spacing, orthogonal frequency division multiplexing symbol length, and the arrangement of the receiving antennas; the feature parameters include: distance features, velocity features, and angle features; The observation tensor is constructed based on the characteristic parameters.

3. The method of claim 2, wherein, Perform tensor decomposition on the observed tensor to determine the decomposed data, including: The observed tensor is estimated using the minimum description length method to obtain the modulus rank corresponding to multiple feature parameters; The principal component analysis model is initialized using a higher-order orthogonal iteration method to obtain the target decomposition model; The observed tensor is decomposed using the target decomposition model to obtain the decomposed data.

4. The method according to claim 3, characterized in that, The observed tensor is decomposed into tensors to determine the decomposed data and reference frequency band, including: Based on the decomposed data, construct the projection matrix of the ideal response subspace corresponding to the feature parameters; Based on the projection matrix, determine the subspace of data for each of the perception dimensions; Based on the residuals of the subspace, the estimated values ​​of the data for each sensing dimension corresponding to the sensing data of each frequency band are determined; The reliability of the estimated value is calculated, and the reference frequency band is determined based on the reliability of the estimated value.

5. The method of claim 4, wherein, Calculating the reliability of the estimated value includes: The root mean square standard deviation of the local uncertainty of the estimated value is calculated based on the covariance of the Jacobian matrix approximation parameter of the fitted residual. Perform multiple random samplings with replacement on the subspace and calculate the perturbation dispersion of the median result for different sampling results; The estimated value is optimized, and the attraction domain consistency of the estimated value is calculated; The reliability of the estimated value is determined based on the root mean square standard deviation, the perturbation dispersion, and the consistency of the attraction domain.

6. The method of claim 4, wherein, The decomposed data is updated by the teacher reference subspace to obtain target decomposed data corresponding to multiple frequency bands, including: The teacher reference subspace is added as a guiding term to the objective function to update the feature matrix and the core tensor, thereby obtaining the target decomposition data; The objective function includes a data reconstruction term and a reference-guided regularization term.

7. The method of claim 6, wherein, The target decomposition data is weighted and fused to obtain target perception data for the at least one target object, including: Determine the concentration of the peak value of the target decomposition data corresponding to each frequency band with respect to the peak value distribution of all target decomposition data; Determine the degree of fit of the target decomposition data for each frequency band with respect to the corresponding observation tensor; Calculate the normalized reconstruction residuals of the fit; The weights of the weighted fusion are determined based on the concentration and the normalized reconstruction residuals. Based on the weights of the weighted fusion, the target decomposition data is weighted and fused to obtain target perception data of the at least one target object.

8. The method of claim 1, wherein, The method further includes: Based on the fusion result obtained by weighted fusion of the target decomposition data, candidate sensing data is extracted; Determine the set of indices for the candidate sensor data across all sensory dimensions; Construct a candidate response matrix for the sensing data corresponding to each frequency band; By fusing the frequency band correlation tensors of multiple frequency bands, a joint correlation tensor is obtained; each element in the joint correlation tensor represents the fusion correlation strength corresponding to the index. Select a preset number of target indices with the highest fusion correlation strength from the set of indices; Based on the target index, the corresponding parameter values ​​are read from the candidate sensing data, and the estimation result of the target sensing data is output.

9. A perception data determination apparatus, characterized by, include: The first construction module is used to construct the observation tensor based on the received sensing echo data of at least two frequency bands about at least one target object; The observation tensor includes the observation tensors corresponding to the at least two frequency bands; The decomposition module is used to perform tensor decomposition on the observed tensor to determine the decomposed data and the reference frequency band. The decomposed data includes: feature matrices and core tensors corresponding to data from multiple sensing dimensions; the reference frequency band is the frequency band with the highest reliability across the multiple sensing dimensions. The second construction module is used to construct a teacher reference subspace for the guidance frequency band based on the multiple dimensions of perception data corresponding to the reference frequency band. The update module is used to update the decomposed data through the teacher reference subspace to obtain target decomposed data corresponding to multiple frequency bands; The fusion module is used to perform weighted fusion on the target decomposition data to obtain target perception data of the at least one target object; the weight of the weighted fusion is related to the reliability of the target decomposition data.

10. A network device, comprising: include: Transceiver, processor, memory, and programs or instructions stored in the memory and executable on the processor; When the processor executes the program or instructions, it implements the steps of the perceptual data determination method as described in any one of claims 1 to 8.