Parameter estimation method, apparatus, device, and storage medium

Abstract

This application proposes a parameter estimation method, apparatus, device, and storage medium, comprising: converting received multi-target echo signals into a multi-target tensor conforming to a unified tensor model of multiple dimensions; denoising the multi-target tensor to obtain a denoised multi-target tensor; inputting the denoised multi-target tensor and VanderMonde constraint information into a trained tensor decomposition model to obtain factor matrices corresponding to each of the multiple dimensions; and estimating the parameters of the multiple factor matrices to obtain parameter sets corresponding to each of the multiple target objects. The embodiments of this application simplify the decomposition method and improve the accuracy of the decomposition by combining VanderMonde constraint information to decompose the tensor.

Description

Technical Field

[0001] This application belongs to the technical field of mobile communications, specifically relating to a parameter estimation method, apparatus, device, and storage medium. Background Technology

[0002] With the development of next-generation wireless communication technologies, Integrated Sensing and Communication (ISAC) technology, through the deep integration of communication and sensing functions, achieves efficient reuse of spectrum resources and has become a core supporting technology for network sensing. Currently, most ISAC systems adopt a MIMO-OFDM architecture. This architecture can reuse existing communication base station hardware, spectrum resources, and signal processing algorithms, and has been initially applied in scenarios such as multi-target monitoring and logistics scheduling. Its core requirement is to accurately acquire the four-dimensional parameters of multiple targets to provide data support for task scheduling and security management.

[0003] Currently, after acquiring the echo signals of multiple targets, it is necessary to decompose the echo signals to obtain the decomposition results corresponding to each dimension and domain, and further obtain the corresponding four-dimensional parameters.

[0004] However, the current decomposition method is cumbersome and the decomposition results are inaccurate, resulting in inaccurate four-dimensional parameters. Summary of the Invention

[0005] This application proposes a parameter estimation method, apparatus, device, and storage medium that can solve the technical problem that current decomposition methods are cumbersome and the decomposition results are inaccurate, resulting in inaccurate four-dimensional parameters.

[0006] The first aspect of this application proposes a parameter estimation method, including: The received multi-target echo signals are converted into a multi-target tensor that conforms to a unified tensor model with multiple dimensions. The multi-object tensor is denoised to obtain a denoised multi-object tensor; The denoised multi-objective tensor and Vandermonde constraint information are input into the trained tensor decomposition model to obtain the factor matrices corresponding to each of the multiple dimensions. Parameter estimation is performed on multiple factor matrices to obtain parameter sets corresponding to each of the multiple target objects.

[0007] An embodiment of the second aspect of this application provides a parameter estimation apparatus, comprising: The conversion module is used to convert the received multi-target echo signals into multi-target tensors that conform to a unified tensor model with multiple dimensions. The denoising module is used to denoise the multi-target tensor to obtain a denoised multi-target tensor. The input module is used to input the denoised multi-objective tensor and Vandermonde constraint information into the trained tensor decomposition model to obtain the factor matrices corresponding to each of the multiple dimensions. The estimation module is used to estimate the parameters of multiple factor matrices to obtain the parameter sets corresponding to each of the multiple target objects.

[0008] An embodiment of the third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect above.

[0009] An embodiment of the fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the method described in the first aspect above.

[0010] The technical solutions provided in this application embodiment have at least the following technical effects or advantages: This application proposes a parameter estimation method, apparatus, device, and storage medium, comprising: converting received multi-target echo signals into a multi-target tensor conforming to a unified tensor model with multiple dimensions; denoising the multi-target tensor to obtain a denoised multi-target tensor; inputting the denoised multi-target tensor and VanderMonde constraint information into a trained tensor decomposition model to obtain factor matrices corresponding to each of the multiple dimensions; and estimating the parameters of the multiple factor matrices to obtain parameter sets corresponding to each of the multiple target objects. The embodiments of this application simplify the decomposition method and improve the accuracy of the decomposition by combining VanderMonde constraint information to decompose the tensor.

[0011] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0012] 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: Figure 1 A flowchart of a parameter estimation method provided in an embodiment of this application is shown; Figure 2 A flowchart of a parameter estimation method provided in an embodiment of this application is shown; Figure 3 This paper shows a schematic diagram of the structure of a parameter estimation device provided in an embodiment of the present application; Figure 4 This illustration shows a schematic diagram of the structure of an electronic device according to an embodiment of this application; Figure 5 A schematic diagram of a storage medium provided in one embodiment of this application is shown. Detailed Implementation

[0013] Exemplary embodiments of this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0014] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains.

[0015] The parameter estimation method of this application can be executed by a computing device that utilizes cloud computing and virtualization technologies. The computing device can be a server, such as a single server, multiple servers, a server cluster, a cloud computing platform, etc. Optionally, the computing device can also be a terminal device, such as a mobile phone, tablet computer, game console, portable computer, desktop computer, digital signage, all-in-one computer, etc. This application does not limit the type or number of computing devices. Building upon the aforementioned background technology, in existing ISAC systems, multi-target four-dimensional parameter estimation mainly relies on traditional signal processing algorithms and early deep learning methods. Traditional algorithms, such as 3D-FFT, are limited by grid resolution, making it difficult to achieve high-precision estimation; subspace algorithms such as MUSIC and ESPRIT, while improving accuracy, have high computational complexity and cannot meet the needs of real-time sensing; tensor decomposition algorithms such as CP-ALS are sensitive to initialization and are prone to getting trapped in local optima; early deep learning methods are mostly "black box" network structures that do not incorporate prior knowledge of wireless signal structure, resulting in insufficient generalization ability in complex scenarios.

[0016] In practical applications, multi-target parameter estimation faces three key challenges: First, complex channel interference leads to severe signal distortion. Low-altitude static clutter energy easily masks weak target signals, and the superposition of Doppler frequency shifts and spectral overlap between multiple targets further confuses signal characteristics, increasing the difficulty of parameter decoupling. Second, it lacks adaptability to low signal-to-noise ratio environments. Traditional denoising methods are difficult to adapt to the multi-domain structural characteristics of high-dimensional tensor signals. Existing deep learning denoising networks lack dynamic noise adaptation capabilities, resulting in drastic performance fluctuations in scenarios with varying signal-to-noise ratios, and are unable to stably output high-purity signals. Third, resource constraints and inherent algorithm defects lead to accuracy degradation. When the number of subcarriers and antennas is reduced or the coherent processing interval is shortened, the estimation accuracy of existing algorithms will drop significantly. At the same time, traditional tensor decomposition algorithms (such as CP-ALS) are sensitive to initialization and prone to getting trapped in local optima. Early deep learning "black box" models lacked generalization ability and interpretability due to the lack of integration of structured priors of wireless signals. These problems lead to large errors in the estimation of 4D parameters for multiple targets, becoming a core bottleneck restricting the application of the ISAC system in the field of multi-target surveillance. There is an urgent need for a parameter estimation algorithm that balances accuracy, robustness, and efficiency.

[0017] To address the aforementioned problems, this application proposes a parameter estimation method, apparatus, device, and storage medium, comprising: converting received multi-target echo signals into a multi-target tensor conforming to a unified tensor model with multiple dimensions; denoising the multi-target tensor to obtain a denoised multi-target tensor; inputting the denoised multi-target tensor and VanderMonde constraint information into a trained tensor decomposition model to obtain factor matrices corresponding to each of the multiple dimensions; and estimating the parameters of the multiple factor matrices to obtain parameter sets corresponding to each of the multiple target objects. The embodiments of this application simplify the decomposition method and improve the accuracy of the decomposition by combining VanderMonde constraint information to decompose the tensor.

[0018] The following describes a parameter estimation method proposed according to an embodiment of this application, with reference to the accompanying drawings.

[0019] See Figure 1 The method specifically includes the following steps: S101. Convert the received multi-target echo signals into a multi-target tensor that conforms to a unified tensor model with multiple dimensions.

[0020] In some embodiments, during the parameter estimation process, the base station first transmits a signal. After the transmitted signal touches the target object, an echo signal is generated. The base station estimates the parameters of multiple target objects by collecting and analyzing the echo signal.

[0021] In some embodiments, the process of generating the transmitted signal can be implemented as follows: Configure core parameters for a single-site ISAC system, including Subcarriers, There are one coherent processing interval (CPI), and the quadrature modulation order is... Generate communication data of length to be sent. A serial bit stream.

[0022] The serial bit stream is ordered according to the modulation order. Mapped to Each constellation symbol, after serial-to-parallel conversion, is assigned to a time-frequency grid to obtain the frequency domain transmitted signal matrix. Matrix elements Indicates the first The subcarrier and the first One CPI transmission symbol, of which , .

[0023] Perform an N-point Inverse Discrete Fourier Transform (IDFT) on each column of the transmitted signal matrix X to obtain discrete time-domain samples; copy the last N_cp data points of each time-domain sample as a cyclic prefix (CP) to obtain the time-domain signal with CP added, in order to eliminate inter-symbol interference (ISI) caused by multipath delay.

[0024] The discrete time-domain signal is converted into a continuous time-domain signal by a pulse shaping filter, and then the transmitted signal is obtained after beamforming, as shown in equation (1):

[0025] in, It is the transmitted beamforming vector. It is the total number of elements in the Uniform Planar Array (UPA) at the transmitting end, of which and These are the number of array elements in the horizontal direction and the number of array elements in the vertical direction, respectively, with an element spacing of... , For the signal wavelength, For subcarrier spacing, For the transmit pulse shaping function, The duration of a symbol in Orthogonal Frequency Division Multiplexing (OFDM) is [not specified]. ,in For OFDM symbol duration, This represents the duration of the cyclic prefix in the OFDM symbol.

[0026] Furthermore, it can receive the corresponding echo signal.

[0027] Generally, echo signals typically include multiple target objects, static clutter, and noise. Therefore, a joint echo signal model containing targets, static clutter, and noise can be constructed to receive the echo signal.

[0028] In some embodiments, the process of receiving the echo signal can be implemented by initializing the sensing scene parameters and setting the number of targets within the sensing area to be [value missing]. , No. One goal ( The distance, radial velocity, azimuth, elevation angle, and channel gain are denoted as follows: , , , and The number of static scattering clusters is Each cluster contains The scattering point, the th scattering point The th cluster scattering points ( and The propagation delay, azimuth angle, elevation angle, and channel gain are respectively , , and .

[0029] Define the UPA steering vector. Both the transmitting and receiving ends use uniform planar arrays. The number of UPA array elements at the receiving end is... ,in and These are the number of array elements in the horizontal and vertical directions at the receiver, and the number of UPA array elements at the transmitter. ,in and These represent the number of array elements in the horizontal direction and the number of array elements in the vertical direction, respectively, with the spacing between all array elements being... , The signal wavelength. The receiver UPA has an angle... The guiding vector is ,in Indicates the Kronecker product. This is the guiding vector for a horizontally uniform linear array. , This is the guiding vector for a uniform linear array in the vertical direction. Considering a single-site ISAC system, and given that the distance between the receiver UPA and the target is much greater than the distance between the transmitter and receiver UPA, it is assumed that the two UPAs are co-located and that the transmitter and receiver use the same array configuration. Simply replace the corresponding array element parameters.

[0030] Construct a joint channel model where the target channel response is the sum of the scattering contributions from all targets. The subcarrier, the Target channel for each CPI As shown in equation (2):

[0031] in, Indicates the first Doppler frequency shift caused by a single target Indicates the first The propagation delay of each target At the speed of light, Represents conjugate transpose, static clutter channel As shown in equation (3):

[0032] The total channel response is .

[0033] A noisy echo signal is generated. After the transmitted signal passes through the main channel, the echo signal received by the receiver is... As shown in equation (4):

[0034] in It is additive white Gaussian noise. The values ​​are calculated based on the preset SNR and antenna domain normalized energy.

[0035] The receiving end preprocesses the echo signal, and... Perform a cyclic prefix removal operation to obtain a time-domain signal free of ISI, with sampling intervals... Sample it and execute The point-based Discrete Fourier Transform (DFT) is used to transform the signal into the frequency domain, ultimately yielding the received signal for subsequent algorithmic processing. As shown in equation (5):

[0036] in, This means that the cyclic prefix is ​​extracted from the echo signal, and the DFT transform realizes the time-domain to frequency-domain signal conversion.

[0037] In fields such as radar, communication, and imaging, signals typically contain features in the time domain (time information), frequency domain (frequency information), and spatial domain (spatial information). By establishing a unified tensor model, these multi-domain features can be fused together to form a high-order tensor structure. This fusion method can more comprehensively represent the characteristics of the signal and avoid information loss caused by single-domain processing. Therefore, after receiving the echo signal, it is necessary to convert the multi-target echo signal into a multi-target tensor that conforms to a unified tensor model across multiple dimensions.

[0038] In some embodiments, converting multi-target echo signals into a multi-target tensor conforming to a unified tensor model of multiple dimensions can be achieved as follows: Construct multi-dimensional correlations of echo signals to link multi-target echo signals. Integrating at the antenna dimension forms the first... Multi-antenna-multi-symbol echo matrix for each subcarrier Matrix elements Indicates the first The receiving antenna, the first The subcarrier, the first The signal value of each CPI ( As shown in equation (6):

[0039] In the formula, This is the noise matrix.

[0040] Stack echo matrices along the subcarrier dimension, and all echo matrix of subcarriers Stacking is performed along the subcarrier dimension to construct a three-dimensional echo signal tensor. Their dimensions correspond to the receiving antenna domain (size). ), CPI symbol field (size) ), subcarrier frequency domain (size) The stacking operation can be expressed as equation (7), which fully preserves the multi-dimensional structural information of the signal in the spatial, time, and frequency domains:

[0041] In the formula, This indicates a stacking operation along the subcarrier dimension.

[0042] Tensor component decomposition, echo signal tensor It can be decomposed into static environment component tensors. Moving target component tensor and noise component tensor ,Right now:

[0043] in, The echo contribution corresponding to a static scattering cluster has tensor elements that satisfy the no-Doppler frequency shift characteristic. The echo contribution corresponding to the target has tensor elements containing coupled information of Doppler frequency shift and propagation delay. It is the additive white Gaussian noise tensor.

[0044] S102. Denoise the multi-objective tensor to obtain a denoised multi-objective tensor.

[0045] As mentioned above, the multi-target tensor includes multiple target objects, static clutter, and noise. Therefore, it is necessary to remove the static clutter and noise to denoise the multi-target tensor and obtain a denoised multi-target tensor.

[0046] In some embodiments, the multi-target tensor is denoised to obtain a denoised multi-target tensor, including: filtering out static clutter from the multi-target tensor by Doppler domain filtering; and inputting the multi-target tensor after filtering out static clutter into a trained denoising model to obtain a denoised multi-target tensor.

[0047] In some embodiments, static clutter in multi-target tensors can be filtered out using Doppler domain filtering, specifically as follows: For multi-objective tensors CPI symbol field execution Point Discrete Fourier Transform ( To avoid spectral aliasing, the Doppler spectral tensor is obtained. The transformation formula is as shown in equation (9):

[0048] in, To index the Doppler frequency points, the characteristic that static clutter has no radial velocity (Doppler frequency shift is 0) is utilized to... Mid-zero Doppler frequency The energy is set to zero, that is... Perform an inverse DFT on the processed spectral tensor to reconstruct the multi-target tensor after filtering out static clutter. As in equation (10):

[0049] Furthermore, the multi-target tensor after filtering out static clutter is input into the trained denoising model to obtain the predicted noise; based on the multi-target tensor after static clutter and the predicted noise, the denoised multi-target tensor is obtained.

[0050] In some embodiments, the training process of the denoising model includes: obtaining each sample tensor and the estimated noise value corresponding to each sample tensor; inputting the first sample tensor into the denoising model to obtain the predicted noise value, wherein the first sample tensor is any sample tensor among the sample tensors; calculating the first loss function value based on the estimated noise value and the predicted noise value corresponding to the first sample tensor; adjusting the model parameters of the denoising model based on the first loss function value, and continuing training until the first preset training completion condition is met to obtain the trained denoising model.

[0051] In some embodiments, during the training process, the relevant denoising model is generally input with a sample tensor and a noise-free target sample tensor. The sample tensor is then input into the denoising model to obtain a predicted denoised sample tensor. The denoising model is trained based on the difference between the target sample tensor and the predicted denoised sample tensor. However, in practical applications, there is generally no noise-free target sample tensor, making the above method difficult to implement in practice.

[0052] In this embodiment of the application, during the training process of the denoising model, the tensors of each sample are first obtained and the estimated noise values ​​corresponding to each sample tensor are determined.

[0053] In some embodiments, coarse noise estimation can be achieved through eigenvalue decomposition: The multi-object tensor after filtering out static clutter Expanded into a two-dimensional matrix along the CPI symbol domain and subcarrier frequency domain The expansion operation satisfies equation (11):

[0054] In the formula, Let represent the expansion operation along the first dimension, and calculate the Hermitian sample covariance matrix of the expanded matrix, as shown in equation (12):

[0055] In the formula Perform eigenvalue decomposition to obtain eigenvalues, and sort them in descending order. Since the target quantity is The signal energy is mainly concentrated in the largest Of the eigenvalues, the remaining eigenvalues ​​are dominated by noise, therefore the estimated value of the noise variance is given by equation (13):

[0056] After determining the noise variance, the multi-object tensor after filtering out static clutter is... Add artificial disturbance noise Constructing noisy training samples ;Will The real and imaginary parts are separated and flattened as the network input, and the network output is the predicted value of the artificial disturbance noise. The network architecture consists of an encoder and a decoder: the encoder contains multiple consecutive modules, each consisting of a 3D convolutional layer, a batch normalization (BN) layer, and a ReLU activation function, to achieve feature extraction and downsampling, and finally outputs a bottleneck layer of high-dimensional abstract features. The decoder reverses the downsampling process through a 3D deconvolutional layer, and works with the BN layer and ReLU activation function to achieve noise prediction and upsampling. Finally, a 3D convolutional layer ensures that the number of output channels matches the number of inputs. During the training phase, a composite loss function is used to optimize the network parameters. The loss function is defined as shown in equation (14).

[0057] in, Predicting loss for core noise ( ), Loss due to time continuity constraints For spectral smoothness constraint loss, and Let be the constraint strength coefficient, where ( (This is the prediction noise after stacking the real and imaginary parts).

[0058] No artificial noise or disturbance needs to be added during the inference phase. The trained network is input, and the predicted noise in the network output is removed to obtain the final denoised target signal tensor. . In summary, during the training process of the denoising model, the estimated noise value of a sample tensor is first determined. This sample tensor is then input into the denoising model to obtain the predicted noise value. Based on the estimated and predicted noise values, a loss function is constructed to adjust the model parameters of the denoising model until the first preset training completion condition is met, resulting in a well-trained denoising model.

[0059] In the application of the denoising model, the multi-object tensor after filtering out static clutter is input into the trained denoising model. The trained denoising model will output the predicted noise corresponding to the multi-object tensor after filtering out static clutter. The predicted noise can be removed from the multi-object tensor after filtering out static clutter to obtain the denoised multi-object tensor.

[0060] S103. Input the denoised multi-objective tensor and Vandermonde constraint information into the trained tensor decomposition model to obtain the factor matrices corresponding to each of the multiple dimensions.

[0061] Among them, multiple dimensions include the time domain, spatial domain, and frequency domain.

[0062] In some embodiments, the denoised multi-object tensor is obtained by tensor multiplication of three matrices, namely the time-domain factor matrix, the frequency-domain factor matrix, and the spatial-domain factor matrix.

[0063] In some embodiments, the Vandermonde constraint information is typically a Vandermonde matrix, a special type of matrix whose elements consist of powers of a set of parameters. For antenna arrays, the Vandermonde matrix is ​​often used to represent the antenna array's response vector.

[0064] In the time domain, the time delay response vector can be determined by measuring the time difference of arrival (TDOA) of the signal, and the Vandermonde matrix can be constructed.

[0065] In the frequency domain, the frequency response vector can be determined by measuring the frequency response of the signal, and the Vandermonde matrix can be constructed.

[0066] When performing denoising multi-objective tensor decomposition, related technologies randomly initialize a factor matrix, and then obtain two other factor matrices through this factor matrix. The loss function value is then calculated by multiplying the tensors of the three factor matrices with the denoising multi-objective tensor. If the loss function value meets the preset conditions, the three factor matrices are obtained.

[0067] However, the above method cannot guarantee convergence to the global optimum, and usually only converges to the local optimum. Furthermore, the convergence result of the above method is very sensitive to the initial value. Different initial values ​​may lead to different decomposition results. Since the initial value is random, it may lead to inaccurate final decomposition results.

[0068] In this embodiment, a tensor decomposition model is first trained to improve the accuracy of initial value prediction. Then, the factor matrices in the time and frequency domains are constrained by Vandermonde constraint information so that the output time-domain factor matrix conforms to the time-domain characteristics and the output frequency-domain factor matrix conforms to the frequency-domain characteristics, thereby improving the accuracy of the output factor matrix.

[0069] S104. Perform parameter estimation on multiple factor matrices to obtain the parameter sets corresponding to each of the multiple target objects.

[0070] In some embodiments, each column of data in the factor matrix belongs to the same target object, and the corresponding parameters can be obtained by calculating each column of data in the factor matrix.

[0071] Furthermore, by combining the parameters obtained from multiple factor matrices, a parameter set is obtained.

[0072] This application proposes a parameter estimation method, apparatus, device, and storage medium, comprising: converting received multi-target echo signals into a multi-target tensor conforming to a unified tensor model of multiple dimensions; denoising the multi-target tensor to obtain a denoised multi-target tensor; inputting the denoised multi-target tensor and VanderMonde constraint information into a trained tensor decomposition model to obtain factor matrices corresponding to each of the multiple dimensions; and estimating the parameters of the multiple factor matrices to obtain parameter sets corresponding to each of the multiple target objects. The embodiments of this application simplify the decomposition method and improve the accuracy of the decomposition by combining VanderMonde constraint information to decompose the tensor.

[0073] In some embodiments, the denoised multi-objective tensor and Vandermonde constraint information are input into a trained tensor decomposition model to obtain factor matrices corresponding to multiple dimensions, including: inputting the denoised multi-objective tensor into the trained tensor decomposition model to obtain factors corresponding to the time domain and frequency domain; expanding the factors corresponding to the time domain and frequency domain through Vandermonde constraint information to obtain factor matrices corresponding to the time domain and frequency domain; and performing inverse operations based on the factor matrices corresponding to the time domain and frequency domain and the denoised multi-objective tensor to obtain the factor matrix corresponding to the spatial domain.

[0074] In some embodiments, since the factor matrices in the time and frequency domains of the denoised multi-objective tensor conform to the Vandermonde constraint information, and in order to improve the convergence performance of the model, the trained tensor decomposition model generally outputs the corresponding factors in the time and frequency domains respectively. The factors in the time and frequency domains are expanded by the Vandermonde constraint information to obtain the factor matrices in the time and frequency domains respectively. Based on the factor matrices in the time and frequency domains and the denoised multi-objective tensor, the inverse operation is performed to obtain the factor matrix in the spatial domain.

[0075] In some embodiments, denoised multi-object tensors It is separated into real and imaginary dual-channel features, retaining the three-dimensional structure of receiving antenna-CPI symbol-subcarrier, ensuring the complete transmission of spatial, temporal, and frequency domain information.

[0076] Construct a channel-space dual attention mechanism to address both the channel and spatial dimensions of the input features. By assigning dynamic weights, the feature representation of the energy concentration region of the target signal is enhanced, residual noise interference is suppressed, and an enhanced feature tensor is output.

[0077] A Vandermonde-constrained deep unfolded neural network with a dual channel-space attention module is constructed to extract joint features from the spatial, temporal, and frequency domains. The features first pass through the attention module, followed by a batch normalization layer, then multiple basic units. Each basic unit consists of a linear layer, a batch normalization layer, a ReLU function, and a Dropout layer. Finally, a fully connected layer serves as the output layer, outputting a low-dimensional abstract feature vector associated with the target motion. Doppler shift of the target and latency Then, based on the Vandermonde structure, it is extended into a complete matrix: the Doppler factor matrix. The Column expansion to ( (for CPI sign interval), to achieve from 1 parameter to Extension of dimensional vectors; time delay factor matrix The Column expansion to ( (for subcarrier spacing), to achieve from 1 parameter to The expansion of the dimensional vector; finally, based on the multilinear constraints of tensor decomposition, utilizing... , and The spatial factor matrix is ​​derived through pseudo-inverse operation. Its column vector matches the guide vector of the receiving antenna array, corresponding to the target angle information.

[0078] In some embodiments, the training process of the tensor decomposition model includes two stages: a first stage and a second stage. The training process in the first stage includes: acquiring each denoised multi-objective sample tensor; inputting the first denoised multi-objective sample tensor into the tensor decomposition model to obtain the first sample factor matrix corresponding to the time domain and frequency domain respectively, wherein the first denoised multi-objective sample tensor is any denoised multi-objective sample tensor of each denoised multi-objective sample tensor; estimating the first sample factor matrix corresponding to the spatial domain based on the first sample factor matrix corresponding to the time domain and frequency domain respectively and the first denoised multi-objective sample tensor; calculating the predicted denoised multi-objective sample tensor based on the first denoised multi-objective sample tensor and the predicted denoised multi-objective sample tensor; adjusting the model parameters of the tensor decomposition model based on the second loss function value, and continuing training until the second preset training completion condition is met to obtain the first tensor decomposition model.

[0079] In some embodiments, the second-stage training process includes: modifying the model structure of the first tensor decomposition model to obtain a second tensor decomposition model, the second tensor decomposition model outputting sample factors corresponding to the sample factor matrix; inputting the second denoised multi-objective sample tensor into the second tensor decomposition model to obtain the sample factors corresponding to the time domain and frequency domain respectively, the second denoised multi-objective sample tensor being any denoised multi-objective sample tensor of each denoised multi-objective sample tensor; expanding the sample factors corresponding to the time domain and frequency domain respectively through Vandermonde constraint information to obtain the second sample factor matrix corresponding to the time domain and frequency domain respectively; based on The second sample factor matrix corresponding to the time domain and frequency domain, and the second denoised multi-objective sample tensor are estimated to obtain the second sample factor matrix corresponding to the spatial domain. The second predicted denoised multi-objective sample tensor is calculated based on the second sample factor matrix corresponding to the time domain and frequency domain, and the second sample factor matrix corresponding to the spatial domain. The third loss function value is calculated based on the second denoised multi-objective sample tensor and the second predicted denoised multi-objective sample tensor. The model parameters of the second tensor decomposition model are adjusted based on the third loss function value, and training continues until the third preset training completion condition is met to obtain the trained tensor decomposition model.

[0080] In some embodiments, the second preset training completion condition and the third preset completion condition can be flexibly set based on the actual situation.

[0081] As mentioned above, initial values ​​are crucial for tensor decomposition. Therefore, in the first stage of the tensor decomposition model, this application first trains the model to output relatively accurate initial values. Obtain the tensors of each denoised multi-objective sample; input the first denoised multi-objective sample tensor into the tensor decomposition model to obtain the first sample factor matrix corresponding to the time domain and frequency domain respectively; based on the first sample factor matrix corresponding to the time domain and frequency domain respectively and the first denoised multi-objective sample tensor, estimate the first sample factor matrix corresponding to the spatial domain.

[0082] Since the derivation of the pseudo-inverse operation is inaccurate if the two factor matrices are not accurate, the final product of the first sample factor matrix in the spatial domain and the first sample factor matrix in the time and frequency domains will have a large difference from the original first denoised multi-objective sample tensor. Therefore, the second loss function value can be calculated based on the first denoised multi-objective sample tensor and the predicted denoised multi-objective sample tensor. The model parameters of the tensor decomposition model can be adjusted based on the second loss function value, and training can continue until the second preset training completion condition is met to obtain the first tensor decomposition model.

[0083] In some embodiments, the value of the second loss function can be obtained by equation (15):

[0084] In the formula, It is a simplified representation of the CP model. This is the value of the second loss function. This is the first sample factor matrix corresponding to the spatial domain. This is the first sample factor matrix corresponding to the frequency domain. This is the first sample factor matrix corresponding to the time domain.

[0085] After obtaining the first tensor decomposition model, a portion of the network can be frozen to ensure that the initial values ​​output by the first tensor decomposition model are not affected.

[0086] Since the factor matrix needs to be tree-formed using Vandermonde constraint information, the output of the tensor decomposition model needs to be adjusted from the factor matrix to the factors themselves. Furthermore, the model structure of the first tensor decomposition model can be modified to obtain the second tensor decomposition model, which outputs the sample factors corresponding to the sample factor matrix.

[0087] Furthermore, the second denoised multi-objective sample tensor is input into the second tensor decomposition model to obtain the corresponding sample factors in the time and frequency domains. These sample factors are then expanded using Vandermonde constraint information to obtain the corresponding second sample factor matrices in the time and frequency domains. Based on these second sample factor matrices and the second denoised multi-objective sample tensor, the corresponding second sample factor matrix in the spatial domain is estimated. The second predicted denoised multi-objective sample tensor is calculated based on these matrices. A third loss function value is calculated based on the second denoised multi-objective sample tensor and the second predicted denoised multi-objective sample tensor. The model parameters of the second tensor decomposition model are adjusted based on the third loss function value, and training continues until the third preset training completion condition is met, resulting in a trained tensor decomposition model.

[0088] The value of the third loss function can be obtained through equation (16):

[0089] in, The value of the third loss function. For the second denoised multi-objective sample tensor, This is the second sample factor matrix corresponding to the spatial domain. This is the second sample factor matrix corresponding to the frequency domain. This is the second sample factor matrix corresponding to the time domain.

[0090] In this embodiment, during the inference process of the trained tensor decomposition model, the model first generates an initialized time-domain factor matrix and an initialized frequency-domain factor matrix corresponding to the denoised multi-objective tensor. These initialized time-domain and frequency-domain factor matrices are then transformed into time-domain and frequency-domain factors for output. The time-domain and frequency-domain factors are extended using Vandermonde constraint information to obtain corresponding factor matrices in the time and frequency domains. Since the model's initial output and the Vandermonde constraint information have been applied, the corresponding factor matrices in the time and frequency domains are relatively accurate. The spatial domain factor matrix obtained by inverse operation based on the corresponding time-domain and frequency-domain factor matrices and the denoised multi-objective tensor is also relatively accurate. Therefore, this embodiment can utilize the initialized factor matrix output by the network. , and The replacement of random initialization and Vandermonde constraint information extends the time-domain and frequency-domain factors, improving the accuracy of tensor decomposition.

[0091] In some embodiments, each column of data in the factor matrix belongs to the same target object. Parameter estimation is performed on multiple factor matrices to obtain parameter sets corresponding to multiple target objects, including: estimating the parameters of the corresponding target object based on the data in each column of each factor matrix.

[0092] In some implementations, each column of data in the factor matrix belongs to the same target object. Therefore, the parameters of the domain can be estimated by calculating a certain column of data. For example, the target angle can be estimated by calculating the spatial factor matrix, the target velocity can be estimated by calculating the frequency spatial factor matrix, and the target distance can be estimated by calculating the time factor matrix.

[0093] In some embodiments, based on the decomposed spatial factor matrix Solve for the target angle using a uniformly meshed guide vector dictionary. In this process, the target angle parameter is estimated by maximizing the correlation. For the th... One goal, one angle Satisfying equation (17):

[0094] In the formula, For dictionary The array steering vector in the array.

[0095] A two-stage grid search strategy is employed to estimate the target velocity. First, a coarse velocity grid is generated. (scope resolution ), for the first For each objective, calculate the normalized correlation coefficient as shown in equation (18):

[0096] In the formula, To generate a velocity vector, select one that makes... The largest As a rough estimate; subsequently in Perform fine-grained iterative search in the vicinity (resolution) Search range Final speed Satisfying equation (19):

[0097] In the formula This is the final, refined search interval.

[0098] Step F3: Estimate the distance using a two-stage grid search strategy within the distance search interval. Within, distance is estimated by maximizing correlation. As shown in equation (20):

[0099] In the formula Generate a distance vector.

[0100] Furthermore, the angle of each target ,speed ,distance By associating parameters by index, a complete 4D parameter set is formed, and parameter estimation is completed.

[0101] In some embodiments, to describe the above parameter estimation method in detail, this application also provides a schematic diagram of the parameter estimation process, such as... Figure 2 As shown, the method includes the following steps: S201. Convert the received multi-target echo signals into a multi-target tensor that conforms to a unified tensor model with multiple dimensions.

[0102] S202. Static clutter in multi-target tensors is filtered out by Doppler domain filtering.

[0103] S203. Input the multi-objective tensor after filtering out static clutter into the trained denoising model to obtain the predicted noise.

[0104] S204. Based on the multi-target tensor after static clutter and the predicted noise, a denoised multi-target tensor is obtained.

[0105] S205. Input the denoised multi-objective tensor into the trained tensor decomposition model to obtain the corresponding factors in the time domain and frequency domain.

[0106] S206. Extend the corresponding factors in the time domain and frequency domain using the Vandermonde constraint information to obtain the corresponding factor matrices in the time domain and frequency domain.

[0107] S207. Based on the factor matrices corresponding to the time domain and frequency domain, and the denoised multi-objective tensor, perform inverse operations to obtain the factor matrix corresponding to the spatial domain.

[0108] S208. Perform parameter estimation on multiple factor matrices to obtain the parameter sets corresponding to each of the multiple target objects.

[0109] This application also provides a parameter estimation apparatus for performing the parameter estimation method provided in any of the above embodiments. For example... Figure 3 As shown, the device includes a conversion module 301, a noise reduction module 302, an input module 303, and an estimation module 304.

[0110] The conversion module 301 is used to convert the received multi-target echo signal into a multi-target tensor that conforms to a unified tensor model with multiple dimensions. Denoising module 302 is used to denoise the multi-target tensor to obtain a denoised multi-target tensor; Input module 303 is used to input the denoised multi-objective tensor and Vandermonde constraint information into the trained tensor decomposition model to obtain the factor matrix corresponding to each of the multiple dimensions. The estimation module 304 is used to perform parameter estimation on the multiple factor matrices to obtain the parameter sets corresponding to each of the multiple target objects.

[0111] This application proposes a parameter estimation method, apparatus, device, and storage medium, comprising: converting received multi-target echo signals into a multi-target tensor conforming to a unified tensor model of multiple dimensions; denoising the multi-target tensor to obtain a denoised multi-target tensor; inputting the denoised multi-target tensor and VanderMonde constraint information into a trained tensor decomposition model to obtain factor matrices corresponding to each of the multiple dimensions; and estimating the parameters of the multiple factor matrices to obtain parameter sets corresponding to each of the multiple target objects. The embodiments of this application simplify the decomposition method and improve the accuracy of the decomposition by combining VanderMonde constraint information to decompose the tensor.

[0112] In some embodiments, the noise reduction module 302 is specifically used for: Static clutter in the multi-target tensor is filtered out by Doppler domain filtering; The multi-objective tensor after filtering out the static clutter is input into the trained denoising model to obtain the predicted noise; Based on the multi-target tensor after static clutter and the predicted noise, a denoised multi-target tensor is obtained.

[0113] In some embodiments, the training process of the denoising model includes: Obtain the tensor of each sample and the estimated noise value corresponding to each sample tensor; The first sample tensor is input into the denoising model to obtain the predicted noise value. The first sample tensor is any one of the sample tensors. The first loss function value is calculated based on the estimated noise value corresponding to the first sample tensor and the predicted noise value; The model parameters of the denoising model are adjusted based on the first loss function value, and training continues until the first preset training completion condition is met, resulting in a well-trained denoising model.

[0114] In some embodiments, the input module 303 is specifically used for: The denoised multi-objective tensor is input into the trained tensor decomposition model to obtain the corresponding factors in the time domain and frequency domain. By extending the factors in the time and frequency domains using Vandermonde constraint information, we obtain the factor matrices in the time and frequency domains respectively. Based on the factor matrices corresponding to the time domain and frequency domain, and the denoised multi-objective tensor, the inverse operation is performed to obtain the factor matrix corresponding to the spatial domain.

[0115] In some embodiments, the training process of the tensor decomposition model includes two phases: a first phase and a second phase. The training process in the first phase includes: Obtain the tensors of each denoised multi-objective sample; Input the first denoised multi-objective sample tensor into the tensor decomposition model to obtain the first sample factor matrix corresponding to the time domain and frequency domain respectively. The first denoised multi-objective sample tensor is any denoised multi-objective sample tensor of each of the denoised multi-objective sample tensors. The first sample factor matrix corresponding to the spatial domain is obtained based on the first sample factor matrix corresponding to the time domain and frequency domain, respectively, and the first denoised multi-objective sample tensor estimation. The predicted denoised multi-object sample tensor is calculated based on the first sample factor matrix corresponding to the time domain and frequency domain, and the first sample factor matrix corresponding to the spatial domain. The second loss function value is calculated based on the first denoised multi-objective sample tensor and the predicted denoised multi-objective sample tensor; The model parameters of the tensor decomposition model are adjusted based on the second loss function value, and training continues until the second preset training completion condition is met, thus obtaining the first tensor decomposition model.

[0116] In some embodiments, the training process in the first phase includes: Modify the model structure of the first tensor decomposition model to obtain the second tensor decomposition model. The second tensor decomposition model outputs the sample factors corresponding to the sample factor matrix. The second denoised multi-objective sample tensor is input into the second tensor decomposition model to obtain the sample factors corresponding to the time domain and frequency domain, and the second denoised multi-objective sample tensor is any denoised multi-objective sample tensor of each of the denoised multi-objective sample tensors. By extending the sample factors corresponding to the time domain and frequency domain respectively through the Vandermonde constraint information, the second sample factor matrix corresponding to the time domain and frequency domain respectively is obtained. The second sample factor matrix corresponding to the spatial domain is obtained based on the second sample factor matrix corresponding to the time domain and frequency domain, respectively, and the second denoised multi-objective sample tensor estimation. The second prediction denoising multi-object sample tensor is calculated based on the second sample factor matrix corresponding to the time domain and frequency domain, and the second sample factor matrix corresponding to the spatial domain. The third loss function value is calculated based on the second denoised multi-objective sample tensor and the second predicted denoised multi-objective sample tensor; Based on the third loss function value, adjust the model parameters of the second tensor decomposition model and continue training until the third preset training completion condition is met, thus obtaining a trained tensor decomposition model.

[0117] In some embodiments, each column of data in the factor matrix belongs to the same target object. The estimation module is specifically used for: The parameters of the corresponding target object are estimated based on the data in each column of each factor matrix.

[0118] This application also provides an electronic device for performing the above-described parameter estimation method. Please refer to... Figure 4 It illustrates a schematic diagram of an electronic device provided by some embodiments of this application. For example... Figure 4 As shown, the electronic device 7 includes: a processor 700, a memory 701, a bus 702 and a communication interface 703. The processor 700, the communication interface 703 and the memory 701 are connected through the bus 702. The memory 701 stores a computer program that can run on the processor 700. When the processor 700 runs the computer program, it executes the parameter estimation method provided in any of the foregoing embodiments of this application.

[0119] The memory 701 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between the device network element and at least one other network element is achieved through at least one communication interface 703 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc.

[0120] Bus 702 can be an ISA bus, PCI bus, or EISA bus, etc. Buses can be divided into address buses, data buses, control buses, etc. Memory 701 is used to store programs. After receiving execution instructions, processor 700 executes the program. The parameter estimation method disclosed in any of the aforementioned embodiments of this application can be applied to processor 700, or implemented by processor 700.

[0121] The processor 700 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 700 or by instructions in software form. The processor 700 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application-Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 701, and processor 700 reads the information from memory 701 and, in conjunction with its hardware, completes the steps of the above method.

[0122] The electronic device provided in this application embodiment and the parameter estimation method provided in this application embodiment are based on the same inventive concept and have the same beneficial effects as the methods they adopt, operate or implement.

[0123] This application also provides a computer-readable storage medium corresponding to the parameter estimation method provided in the foregoing embodiments. Please refer to... Figure 5 The computer-readable storage medium shown is an optical disc 30, on which a computer program (i.e., a program product) is stored. When the computer program is run by a processor, it executes the parameter estimation method provided in any of the aforementioned embodiments.

[0124] It should be noted that examples of computer-readable storage media may also include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory, or other optical and magnetic storage media, which will not be elaborated here.

[0125] The computer-readable storage medium provided in the above embodiments of this application and the parameter estimation method provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0126] It should be noted that: Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known structures and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0127] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of this application and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination.

[0128] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A parameter estimation method, characterized in that, include: The received multi-target echo signals are converted into a multi-target tensor that conforms to a unified tensor model with multiple dimensions. The multi-object tensor is denoised to obtain a denoised multi-object tensor; The denoised multi-objective tensor and Vandermonde constraint information are input into the trained tensor decomposition model to obtain the factor matrices corresponding to each of the multiple dimensions. Parameter estimation is performed on multiple factor matrices to obtain parameter sets corresponding to each of the multiple target objects.

2. The method according to claim 1, characterized in that, The denoising process of the multi-object tensor to obtain a denoised multi-object tensor includes: Static clutter in the multi-target tensor is filtered out by Doppler domain filtering; The multi-objective tensor after filtering out the static clutter is input into the trained denoising model to obtain the predicted noise; Based on the multi-target tensor after static clutter and the predicted noise, a denoised multi-target tensor is obtained.

3. The method according to claim 2, characterized in that, The training process of the denoising model includes: Obtain the tensor of each sample and the estimated noise value corresponding to each sample tensor; The first sample tensor is input into the denoising model to obtain the predicted noise value. The first sample tensor is any one of the sample tensors. The first loss function value is calculated based on the estimated noise value corresponding to the first sample tensor and the predicted noise value; The model parameters of the denoising model are adjusted based on the first loss function value, and training continues until the first preset training completion condition is met, resulting in a well-trained denoising model.

4. The method according to claim 1, characterized in that, The step involves inputting the denoised multi-objective tensor and Vandermonde constraint information into a trained tensor decomposition model to obtain factor matrices corresponding to multiple dimensions, including: The denoised multi-objective tensor is input into the trained tensor decomposition model to obtain the corresponding factors in the time domain and frequency domain. By extending the factors in the time and frequency domains using Vandermonde constraint information, we obtain the factor matrices in the time and frequency domains respectively. Based on the factor matrices corresponding to the time domain and frequency domain, and the denoised multi-objective tensor, the inverse operation is performed to obtain the factor matrix corresponding to the spatial domain.

5. The method according to claim 1, characterized in that, The training process of a tensor decomposition model consists of two stages: a first stage and a second stage. The first stage of the training process includes: Obtain the tensors of each denoised multi-objective sample; Input the first denoised multi-objective sample tensor into the tensor decomposition model to obtain the first sample factor matrix corresponding to the time domain and frequency domain respectively. The first denoised multi-objective sample tensor is any denoised multi-objective sample tensor of each of the denoised multi-objective sample tensors. The first sample factor matrix corresponding to the spatial domain is obtained based on the first sample factor matrix corresponding to the time domain and frequency domain, respectively, and the first denoised multi-objective sample tensor estimation. The predicted denoised multi-object sample tensor is calculated based on the first sample factor matrix corresponding to the time domain and frequency domain, and the first sample factor matrix corresponding to the spatial domain. The second loss function value is calculated based on the first denoised multi-objective sample tensor and the predicted denoised multi-objective sample tensor; The model parameters of the tensor decomposition model are adjusted based on the second loss function value, and training continues until the second preset training completion condition is met, thus obtaining the first tensor decomposition model.

6. The method according to claim 5, characterized in that, The first phase of the training process includes: Modify the model structure of the first tensor decomposition model to obtain the second tensor decomposition model. The second tensor decomposition model outputs the sample factors corresponding to the sample factor matrix. The second denoised multi-objective sample tensor is input into the second tensor decomposition model to obtain the sample factors corresponding to the time domain and frequency domain, and the second denoised multi-objective sample tensor is any denoised multi-objective sample tensor of each of the denoised multi-objective sample tensors. By extending the sample factors corresponding to the time domain and frequency domain respectively through the Vandermonde constraint information, the second sample factor matrix corresponding to the time domain and frequency domain respectively is obtained. The second sample factor matrix corresponding to the spatial domain is obtained based on the second sample factor matrix corresponding to the time domain and frequency domain, respectively, and the second denoised multi-objective sample tensor estimation. The second prediction denoising multi-object sample tensor is calculated based on the second sample factor matrix corresponding to the time domain and frequency domain, and the second sample factor matrix corresponding to the spatial domain. The third loss function value is calculated based on the second denoised multi-objective sample tensor and the second predicted denoised multi-objective sample tensor; Based on the third loss function value, adjust the model parameters of the second tensor decomposition model and continue training until the third preset training completion condition is met, thus obtaining a trained tensor decomposition model.

7. The method according to claim 1, characterized in that, Each column of data in the factor matrix belongs to the same target object. The parameter estimation of multiple factor matrices yields parameter sets corresponding to each of the multiple target objects, including: The parameters of the corresponding target object are estimated based on the data in each column of each factor matrix.

8. A parameter estimation device, characterized in that, include: The conversion module is used to convert the received multi-target echo signals into multi-target tensors that conform to a unified tensor model with multiple dimensions. The denoising module is used to denoise the multi-target tensor to obtain a denoised multi-target tensor. The input module is used to input the denoised multi-objective tensor and Vandermonde constraint information into the trained tensor decomposition model to obtain the factor matrices corresponding to each of the multiple dimensions. The estimation module is used to estimate the parameters of multiple factor matrices to obtain the parameter sets corresponding to each of the multiple target objects.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by a processor to implement the method as described in any one of claims 1-7.

Images