Target identification method based on adaptive polynomial kernel function and multi-modal fusion

Abstract

The invention discloses a target recognition method based on an adaptive polynomial kernel function and multi-modal fusion, and belongs to the technical field of target recognition, and the target recognition method comprises the steps: obtaining an original signal of a to-be-recognized target, and carrying out the preprocessing of the original signal based on a wavelet threshold denoising algorithm, and generating a to-be-analyzed signal; performing multi-modal time-frequency analysis on the to-be-analyzed signal, and outputting three types of time-frequency characteristic patterns; carrying out dynamic weighted fusion processing on the three types of time-frequency feature maps to generate a fused time-frequency feature tensor, and carrying out convolutional neural network processing to generate a time sequence feature vector; and respectively inputting the time sequence feature vector into a convolutional neural network and a gradient boosting regression tree GBRT model for processing, and then outputting a preliminary recognition result and a feature verification result based on a target category to complete target recognition. According to the method, through a multi-modal analysis technology, the advantages of short-time Fourier transform and wavelet transform are combined, and accurate recognition of the target is achieved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of target recognition, and particularly relates to a target recognition method based on an adaptive polynomial kernel function and multi-modal fusion. BACKGROUND

[0002] With the rapid development and progress of target recognition technology, the demand of modern recognition systems for complex signal processing increases exponentially, and signals often exhibit nonlinear, time-varying and other characteristics. Accurate evaluation of instantaneous frequency is the key to analyzing the nature of signals and ensuring the quality of subsequent feature extraction and recognition decisions.

[0003] Currently, the evaluation methods for instantaneous frequency in target recognition are developing in a diversified trend, mainly including short-time Fourier transform (STFT), continuous wavelet transform (CWT), Wigner-Ville distribution (WVD) and chirplet transform (CT), etc. They play their respective roles in different scenarios, but also have certain limitations.

[0004] The current mainstream time-frequency analysis methods have limitations in processing nonlinear IF signals in target recognition: short-time Fourier transform (STFT) is limited by the Heisenberg-Gabor inequality, and it is difficult to balance the time-frequency resolution; continuous wavelet transform (CWT) has better time-frequency localization ability, but still cannot accurately capture rapidly changing frequency components; Wigner-Ville distribution (WVD) is easily disturbed by noise and cross terms; chirplet transform (CT) is only suitable for linear IF signals. Although polynomial chirplet transform (PCT) provides a way for nonlinear IF signal analysis, its traditional form uses a fixed polynomial kernel function, which is difficult to dynamically match the time-varying characteristics of signals, and still has the problem of insufficient accuracy in processing strong nonlinear and multi-component target recognition signals.

[0005] In summary, although the existing methods based on time-frequency distribution have advantages in processing nonlinear IF signals in the field of target recognition, they all have different problems, and it is difficult to achieve high-precision analysis and estimation. It is still necessary to improve and optimize the existing technology. SUMMARY

[0006] In order to solve the time-frequency resolution contradiction, cross term interference, adaptability problem and precision decline problem in nonlinear instantaneous frequency signal analysis, the application proposes a target recognition method based on adaptive polynomial kernel function and multi-modal fusion, the application takes the improved polynomial chirplet transform (IPCT) as the core, designs the adaptive kernel function to improve the nonlinear fitting ability, and carries out multi-modal fusion on IPCT, WVD and STFT to maximize the role of IPCT, utilizes the multi-modal analysis technology, combines the advantages of short-time Fourier transform and wavelet transform, and realizes accurate recognition of the target.

[0007] The application proposes a target recognition method based on adaptive polynomial kernel function and multi-modal fusion, and specifically includes the following steps: Obtain the original signal of the target to be recognized, and generate the signal to be analyzed by preprocessing the original signal based on the wavelet threshold denoising algorithm; Perform multi-modal time-frequency analysis on the signal to be analyzed, output the first time-frequency feature map based on the adaptive polynomial kernel function, the second time-frequency feature map based on the Wigner-Ville distribution WVD, and the third time-frequency feature map based on the short-time Fourier transform STFT; The first time-frequency feature map, the second time-frequency feature map and the third time-frequency feature map are dynamically weighted and fused to generate a fusion time-frequency feature tensor, and after convolutional neural network processing, a time series feature vector is generated; The time series feature vector is input into the convolutional neural network and the gradient boosting regression tree GBRT model for processing, and the preliminary recognition result and the characteristic verification result based on the target category are output; According to the preliminary recognition result and the characteristic verification result of the target category, the target recognition is performed.

[0008] Further, the step of preprocessing the original signal based on the wavelet threshold denoising algorithm to generate the signal to be analyzed specifically includes: According to the kurtosis of the original signal, the original signal is decomposed by multi-layer wavelet, if the kurtosis is greater than 3, the original signal contains impulse noise; The Fourier transform is performed on the original signal to obtain the signal power spectral density, and the power spectral flatness is calculated according to the signal power spectral density; If the power spectral flatness is approximately equal to 1, the original signal contains white noise; If the power spectral flatness is less than 1, the original signal contains color noise; Respectively set the corresponding noise threshold for white noise, color noise and impulse noise, and generate the signal to be analyzed with high signal-to-noise ratio after wavelet coefficient processing according to the corresponding noise threshold.

[0009] Furthermore, the steps of setting corresponding noise thresholds for white noise, colored noise, and impulse noise, and generating a high signal-to-noise ratio signal after wavelet coefficient processing based on the corresponding noise thresholds, specifically include: Set corresponding noise thresholds for white noise, colored noise, and impulse noise respectively, as follows: White noise using a threshold ; Color noise uses dynamic threshold ; Impulse noise with kurtosis adaptive hard threshold ; Wavelet coefficients smaller than the noise threshold are set to zero or attenuated, while wavelet coefficients larger than the noise threshold are retained. The retained wavelet coefficients are then reconstructed using an inverse transform to obtain the signal to be analyzed with a high signal-to-noise ratio. in This represents the estimated standard deviation of noise. Indicates the length of the original signal. This represents the estimated noise power spectral density. This is the kurtosis coefficient.

[0010] Furthermore, the step of performing multimodal time-frequency analysis on the signal to be analyzed and outputting a first time-frequency feature map based on an adaptive polynomial kernel function specifically includes: After framing and extracting local features from the signal to be analyzed, a local energy distribution is generated. and frequency change rate ; Construct an adaptive polynomial kernel function containing higher-order terms , is represented as: ; Indicates the first The polynomial coefficients of the frame. Denotes the order of a polynomial. Represents time axis variables, This represents the displacement integral variable in the wavelet transform. Energy weighting factor , for Energy distribution at time, For the first The center moment of the frame, express From frame 1 to frame 2 The maximum value of the local energy distribution within the frame; The polynomial coefficients were analyzed using a particle swarm optimization algorithm. Optimization yields the first Optimal polynomial coefficients of the frame, and an optimal adaptive polynomial kernel function; Based on the optimal adaptive polynomial kernel function Obtain a time-frequency distribution adapted to the nonlinear instantaneous frequency characteristics of the signal to be analyzed, and output a first time-frequency feature map.

[0011] Further, the step of performing multimodal time-frequency analysis on the signal to be analyzed and outputting a second time-frequency feature map based on Wigner-Ville distribution (WVD) and a third time-frequency feature map based on short-time Fourier transform (STFT) comprises: The Wigner-Ville distribution (WVD) is used to perform Gaussian window smoothing on the signal to be analyzed to suppress cross terms, and after noise suppression processing, a second time-frequency feature map with enhanced energy concentration area is outputted; The short-time Fourier transform (STFT) is used to perform adaptive window length adjustment on the signal to be analyzed. If it is a slowly varying signal, a long window is used to improve the frequency resolution, and if it is a fast varying signal, a short window is used to guarantee the time resolution, and finally a third time-frequency feature map is generated as a reference.

[0012] Further, the step of dynamically weighting and fusing the first, second and third time-frequency feature maps to generate a fused time-frequency feature tensor comprises: The first, second and third time-frequency feature maps are set to the same time axis and frequency axis, and are adjusted to the same resolution by interpolation; The energy of the first, second and third time-frequency feature maps after adjusting the resolution is normalized to eliminate the difference in energy scale; A dynamic weighting fusion method is used, and the weights are determined based on the entropy values of the three time-frequency feature maps in different frequency intervals; The energy values of the three time-frequency feature maps at each time-frequency point are weighted and summed according to the determined weights to generate a fused time-frequency feature tensor.

[0013] Further, the step of generating a time series feature vector after the fused time-frequency feature tensor is processed by a convolutional neural network comprises: The convolutional neural network (CNN) is used to extract features from the fused time-frequency feature tensor. Different sizes of convolution kernels are used in the convolution layer to capture local time-frequency patterns. After dimension reduction processing by the pooling layer, a local feature vector is outputted; The local feature vector is unfolded into a time series feature sequence along the time axis, and is inputted into a recurrent neural network (RNN) for time series modeling. The time series feature vector is outputted through the gating mechanism of the RNN.

[0014] Further, the step of inputting the time sequence feature vectors into a convolutional neural network and a gradient boosting regression tree (GBRT) model respectively for processing and outputting preliminary identification results and characteristic verification results based on target categories comprises: dividing the time sequence feature vectors into two paths and inputting them into the convolutional neural network and the GBRT model respectively; the first path of time sequence feature vectors passes through the fully connected layer of the convolutional neural network (CNN) and the Softmax function to output target category probabilities, and the category with the highest probability value is taken as the preliminary identification result and the highest probability value is taken as the probability confidence; the second path of time sequence feature vectors passes through the GBRT model to output a medium frequency estimation value, extracts target characteristic verification parameters carried by the target information of the medium frequency estimation value, matches the target characteristic verification parameters with source characteristic verification parameters in a pre-constructed mapping library, and generates a characteristic verification result; wherein the target characteristic verification parameters include a frequency change rate based on an actual value a medium frequency range and a time-frequency ridge line morphology parameter .

[0015] Further, the step of matching the target characteristic verification parameters with source characteristic verification parameters in a pre-constructed mapping library to generate a characteristic verification result comprises: pre-constructing a "target category-theoretical medium frequency characteristic" mapping library, and each category of target in the mapping library is matched with a theoretical value of the mean and threshold range of the medium frequency characteristic under different maneuvering states and electromagnetic environments; matching actual values based on the frequency change rate the medium frequency range and the time-frequency ridge line morphology parameter with the theoretical values in the mapping library to generate characteristic verification results based on the frequency change rate error the medium frequency range overlap degree the ridge line morphology ; wherein, , represents the mean of the preliminary identification category theoretical frequency change rate, the smaller, the higher the matching degree; , represents the medium frequency range theoretical value, the larger, the higher the matching degree; , represents the number of ridge line sampling points, , respectively represent the first point morphology values of the actual and theoretical ridge lines, The smaller, the higher the matching degree.

[0016] Further, the step of target recognition according to the preliminary identification result of the target category and the characteristic verification result specifically comprises: obtaining the probability confidence of the preliminary identification result and the frequency change rate error of the characteristic verification result , the medium frequency range coincidence degree , the ridge line form . If , and the probability confidence of the preliminary identification result , it is determined that the target recognition result is accurate, and the final target category and the recognition confidence are output. If , or the probability confidence , it is determined that the recognition result is invalid. Wherein, the recognition confidence is represented as the weighted average calculation result of the probability confidence, the frequency change rate error , the medium frequency range coincidence degree , and the ridge line form .

[0017] The present application has the following technical effects: (1) The present application proposes an improved adaptive polynomial chirp transform (IPCT), which sets an adaptive polynomial kernel function and uses it for processing of nonlinear IF signals in target recognition, obtains local fragments by frame processing of the signal to be analyzed, then calculates the local energy distribution and the frequency change rate, and then uses the particle swarm optimization algorithm to dynamically adjust the polynomial coefficients to realize accurate fitting of nonlinear IF signals. (2) The present application designs different thresholds for different noises (white noise fixed threshold, color noise frequency-dependent threshold, and impulse noise kurtosis adaptive threshold), and combines neighborhood smoothing to eliminate residual distortion of impulse noise. Compared with traditional fixed threshold denoising, the signal-to-noise ratio is improved by 3-5dB in a mixed noise scene, providing a higher quality signal basis for subsequent time-frequency analysis. (3) The present application integrates the advantages of multi-modal time-frequency features, compensates for the limitations of single methods, effectively improves the estimation accuracy of nonlinear instantaneous frequency signals and intermediate frequencies, and enhances adaptability in complex environments. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The flow chart of the target recognition method based on adaptive polynomial kernel function and multi-modal fusion proposed by the present application; Figure 2 The schematic diagram of the three time-frequency feature maps output after multi-modal time-frequency analysis in the embodiment of the present application; Figure 3A schematic diagram of a fusion time-frequency feature tensor output after fusion processing of three time-frequency feature maps in an embodiment of the present application. DETAILED DESCRIPTION

[0019] The present application will be further described in conjunction with the accompanying drawings, but the scope of protection of the present application is not limited to the specific embodiments.

[0020] As shown in the figure, the present embodiment provides a target recognition method based on adaptive polynomial kernel function and multi-modal fusion, aiming to improve the processing accuracy of nonlinear IF signals, which specifically includes the following steps: Figure 1 Obtain the original signal of the target to be recognized, and pre-process the original signal based on a wavelet threshold denoising algorithm to generate an analysis signal.

[0021] Due to the noise and clutter (and often mixed scenes of white noise, colored noise, and impulse noise) that inevitably exist in the transmission and reception process of signals in target recognition, a wavelet threshold denoising algorithm with adaptive noise classification and dynamic threshold optimization is used to pre-process the original signal. This algorithm uses wavelet transform to decompose the signal into different frequency subbands, sets appropriate thresholds for processing wavelet coefficients based on the characteristic differences between target echoes and noise and clutter in the wavelet domain, effectively suppresses environmental noise, and provides a high-quality signal basis for subsequent time-frequency analysis.

[0022] In the present embodiment, the original signal is subjected to multi-layer wavelet decomposition according to the kurtosis of the original signal, wherein the kurtosis K is a statistical quantity describing the steepness of the probability distribution of the signal, and is used to distinguish impulse noise.

[0023] ; In the above formula, denotes the mathematical expectation, is the sampling value of the original signal, and the formula can quantify the "sharpness" of the signal distribution. It should be noted that if K>3, the noise is identified as impulse noise, and if K<3, the noise is identified as Gaussian noise.

[0024] In the present embodiment, white noise and colored noise are distinguished according to the power spectrum flatness S, which is an index describing the uniformity of the power spectrum of the signal.

[0025] ; In the above formula, is the power spectrum density of the signal, which is calculated by Fourier transform of the original signal, and reflects the power distribution of the signal at different frequencies ; is the frequency axis variable, and the integration interval is limited to ; ​is the signal sampling frequency, which is the basic parameter of signal processing and is determined by the data acquisition device parameter. Among them, The value range is The closer the value is to 1, the more uniform the signal power spectrum is distributed in each frequency interval. When , the original signal contains white noise, , the original signal contains colored noise.

[0026] Through the above process, the noise is finally divided into three categories: "white noise dominant", "colored noise dominant" and "impulse noise dominant".

[0027] In this embodiment, the threshold is adjusted according to different noise types, and the specific process is as follows: The threshold for white noise is ; The dynamic threshold for colored noise is ; The kurtosis adaptive hard threshold for impulse noise is ; The wavelet coefficients less than the noise threshold are set to zero or attenuated, and the wavelet coefficients greater than the noise threshold are reserved; Among them, represents the noise standard deviation estimate, represents the length of the original signal, represents the noise power spectrum density estimate, is the kurtosis coefficient.

[0028] For white noise, the larger L is and the longer L is, the larger the value is, the stronger the inhibition of noise is; for colored noise, , It is obtained by separating the signal power spectrum from the target echo power spectrum; for impulse noise, , it increases linearly with the increase of kurtosis In order to avoid the residual pseudo-peak after removing the impulse noise, the wavelet coefficients after threshold processing are processed by 3x3 neighborhood smoothing. Through the above process, the wavelet coefficients after threshold processing of white noise, colored noise and impulse noise are inverse transformed and reconstructed to obtain the high signal-to-noise ratio signal to be analyzed.

[0029] It should be noted that the noise type identification process of the embodiment is to more accurately preserve signal characteristics and suppress mixed noise. Different types of noise have different characteristics in the wavelet domain, and it is impossible to consider the effects of all types of noise by using a unified denoising method. The noise in the target recognition scene is not of a single type, and different noises have completely different energy distribution and coefficient characteristics in the wavelet domain. If a unified threshold is used without classification, signal distortion and noise residue will occur. In the embodiment, different threshold strategies are designed according to the characteristics of different noises. For example, a smaller threshold is used for low-frequency color noise subband (to avoid distortion of the low-frequency characteristics of the signal), and a larger threshold is used for high-frequency color noise subband (to strengthen noise suppression). For impulse noise, a hard threshold is used for screening, and then a neighborhood smoothing is used to eliminate residual artifacts, thereby improving the fidelity of the denoised signal. If white noise is left, false peaks will appear in the local energy distribution, which will misjudge the core of the signal energy. If color noise is left, it will interfere with the calculation of the second derivative of the frequency change rate, resulting in the misselection of the polynomial order (n = 2 / 3-4 / 5) of the kernel function. If impulse noise is left, a “pseudo ridge line” will appear in the time-frequency distribution, which will affect the subsequent multi-modal fusion (with WVD and STFT) feature alignment.

[0030] In the “nonlinear IF signal processing for target recognition” scenario of the embodiment, “noise classification” and “wavelet threshold denoising” are combined to design a dynamic threshold for mixed noise, thereby solving the problem that a traditional fixed threshold cannot adapt to mixed noise.

[0031] The multi-modal time-frequency analysis is performed on the to-be-analyzed signal to output a first time-frequency feature map based on an adaptive polynomial kernel function, a second time-frequency feature map based on a Wigner-Ville distribution (WVD), and a third time-frequency feature map based on a short-time Fourier transform (STFT).

[0032] In the embodiment, the process of performing multi-modal time-frequency analysis on the to-be-analyzed signal mainly includes three parts, which are respectively a polynomial chirplet transform using an adaptive polynomial function, a signal processing based on a Wigner-Ville distribution (WVD), and a signal processing based on a short-time Fourier transform (STFT). As shown in Figure 2 The multi-modal time-frequency analysis result is shown, and the specific content is as follows: (1) Processing by polynomial chirplet transform using an adaptive polynomial function Traditional polynomial linear frequency modulated wavelet transform (PCT) relies on a fixed-form polynomial kernel function to perform time-frequency analysis. While this fixed kernel function design ensures the determinism of the computation process, it cannot dynamically adjust the kernel function structure to match the time-varying characteristics of highly nonlinear instantaneous frequency (IF) signals widely present in real-world scenarios (such as signals with multi-component coupling, frequency abrupt changes, or high-order nonlinear variations). This can easily lead to problems such as time-frequency ambiguity and energy dispersion. Therefore, an adaptive polynomial kernel function is proposed. It can be understood that the adaptive polynomial kernel function in this embodiment is actually an improvement on polynomial linear frequency modulated wavelet transform (PCT). To clearly describe the content of this embodiment, the improved polynomial linear frequency modulated wavelet transform is defined as IPCT (Improved PCT, IPCT).

[0033] To capture the time-varying characteristics of a signal, the signal to be analyzed... Frame segmentation is performed, using a sliding window function to divide the long signal into several short-time local segments. By segmenting the data into frames, the time-frequency characteristics of each local segment can be made relatively stable over a short period of time, providing a "local benchmark" for subsequent adaptive polynomial kernel function design. For each local segment... Calculate the reflection of the first Local energy distribution of the intensity spatial distribution of the frame signal and frequency change rate , represents the following: ; ; in, This is the intra-frame time offset. For the first Frame signal in Amplitude at time, The peak position corresponds to the energy core of the signal in that frame. Subsequent kernel functions need to enhance adaptation in this region by calculating the rate of frequency change to quantize the first... The degree of nonlinearity of the instantaneous frequency of the frame.

[0034] The first digit extracted by short-time Fourier transform (STFT) Intra-frame instantaneous frequency; second derivative It directly reflects the curvature of the frequency curve; the larger its absolute value, the stronger the frequency nonlinearity of the signal frame. Time (weakly nonlinear), order of kernel function polynomial ;when Time (nonlinear), ;when Time (strong nonlinearity), ,in, The order of the polynomial that determines the kernel function The peak region is used as a weighting factor for kernel function coefficient optimization.

[0035] The kernel function is the core of time-frequency analysis using polynomial linear frequency modulated wavelet transform, and its structure needs to be consistent with that extracted in step one. Precise matching means that "the degree of nonlinearity determines the order, and the energy distribution determines the weight." To adapt to the time-frequency coupling characteristics of nonlinear IF signals, this embodiment constructs an adaptive polynomial kernel function containing higher-order terms, expressed as: ; in, For the first The polynomial coefficients to be optimized in the frame kernel function. Let be the order of the polynomial. Energy weighting factor . This is the global analysis point, a time-axis variable corresponding to the entire signal's time series. In the analysis of the... Frame time fixed as , For the first The center moment of the frame is a nonvariable. For the displacement integral variable in wavelet transform, express From frame 1 to frame 2 The maximum value of the local energy distribution in the frame.

[0036] In this embodiment, the coefficients of the kernel function The key to its adaptability lies in the particle swarm optimization (PSO) algorithm used to solve for the optimal coefficients, aiming at the energy concentration of the time-frequency distribution, so that the kernel function accurately matches the signal features of the m-th frame. The specific process is as follows: Initialize the particle swarm, with each particle corresponding to a set of coefficients. , dimension . Number of particles Inertial weight Learning factor .

[0037] For each particle Its location Corresponding to a set of coefficients Substituting into the adaptive polynomial kernel function formula, we obtain the first... Particle-specific kernel function for frames Based on this kernel function, an analytical wavelet is constructed: ; in It is the first frame the particle-specific kernel function; is the frequency carrier term, which is used to map the "time-frequency structure" of the kernel function to a specific frequency ; is the analysis wavelet of the frame the particle, whose time-frequency focusing property is completely determined by the kernel function .

[0038] Finally, the frame the :

[0039] ; where, denotes the complex conjugate of the wavelet function, is the frequency axis variable , is the sampling frequency; the integral is accelerated by the fast Fourier transform (FFT) (transforming the time-domain integral into a frequency-domain product, reducing the complexity).

[0040] The fitness (i.e., the energy concentration degree) of the particle is calculated according to : ; wherein the numerator is the "fourth power sum of the time-frequency distribution" (strengthening the contribution of the energy concentration area), and the denominator is the "square of the square sum of the time-frequency distribution" (normalization processing, avoiding the influence of the absolute value of the energy); The value range of is , the closer to 1, the better the adaptability of the kernel function corresponding to the particle.

[0041] Particle iteration update is performed, and the initial positions and velocities of particles are randomly generated. The individual optimal position (the maximum corresponding to the particle history) and the global best position (the maximum corresponding to all particle history) of each particle are recorded; then the particle velocity and position are updated (iteration process is limited to avoid velocity divergence) ; ; ; ​​in, This represents the current iteration number. Is A random number that is uniformly distributed within an interval; The inertia weight is used to balance the global and local search capabilities of the algorithm, and its calculation formula is as follows: , For maximum inertia weight, For minimum inertia weight, This represents the maximum number of iterations. and It is the learning factor, usually a constant greater than zero, used to adjust the degree to which a particle learns towards its individual optimal position and the global optimal position.

[0042] When the number of iterations or ( Represents the globally optimal energy concentration, globally optimal When the value tends to stabilize, stop iterating; at this point... Corresponding coefficients That is, the first Find the optimal polynomial coefficients of the frame and obtain the optimal kernel function. Based on the optimal adaptive polynomial kernel function The time-frequency distribution adapted to the nonlinear instantaneous frequency characteristics of the signal to be analyzed is obtained, and the first time-frequency feature map is output.

[0043] (2) The signal to be analyzed is smoothed by Gaussian window using Wigner-Ville distribution (WVD) to suppress cross terms. After noise suppression, a second time-frequency feature map is output to enhance the energy concentration region. The signal to be analyzed is processed by adaptive adjustment of window length using short-time Fourier transform (STFT). If it is a slow-changing signal, a long window is used to improve frequency resolution. If it is a fast-changing signal, a short window is used to ensure time resolution. Finally, a third time-frequency feature map is generated as a reference.

[0044] It should be noted that Gaussian window smoothing effectively suppresses cross-term interference, but may slightly reduce time-frequency resolution; long window improves frequency resolution and presents narrowband spectrum lines; short window improves time resolution and can capture transient changes.

[0045] The first, second, and third time-frequency feature maps are dynamically weighted and fused to generate a fused time-frequency feature tensor, which is then processed by a convolutional neural network to generate a time-series feature vector.

[0046] In this embodiment, the fusion of the above three types of feature maps specifically includes: The first time-frequency feature map, the second time-frequency feature map, and the third time-frequency feature map are arranged on the same time axis and frequency axis, and are adjusted to the same resolution through interpolation; The energies of the first time-frequency feature map, the second time-frequency feature map, and the third time-frequency feature map after the resolution adjustment are normalized to eliminate differences in energy scales; A dynamic weighted fusion method is adopted, and the weights are determined based on the entropy values of the three time-frequency feature maps in different frequency intervals. The smaller the entropy value, the more concentrated the features in the interval, and the higher the weight. As shown in Figure 3 , the energy values of the three time-frequency feature maps at each time-frequency point are weighted and summed according to the determined weights to generate a fused time-frequency feature tensor.

[0047] A convolutional neural network (CNN) is used to extract features from the fused time-frequency feature tensor. Different sizes of convolution kernels are used in the convolution layer to capture local time-frequency patterns. The dimensionality is reduced and key features are strengthened through the pooling layer. The local feature vector is output. Then, the local feature vector output by the CNN is unfolded into a time series feature sequence along the time axis, and input into a recurrent neural network (RNN) for time series modeling. The global time-varying law of the signal is captured through the gating mechanism of the RNN, and the time series feature vector is output.

[0048] The time series feature vector is input into a convolutional neural network and a gradient boosting regression tree (GBRT) model for processing, and the preliminary recognition result and the characteristic verification result based on the target category are output. The target is identified according to the preliminary recognition result and the characteristic verification result of the target category.

[0049] In this embodiment, the time series feature vector is divided into two paths and input into a convolutional neural network and a gradient boosting regression tree (GBRT) model respectively. The first path of the time series feature vector passes through the fully connected layer of the convolutional neural network (CNN) and the Softmax function to output the probability of the target category, and the class with the highest probability value is taken as the preliminary recognition result and the highest probability value is taken as the probability confidence. The second path of the time series feature vector passes through the gradient boosting regression tree (GBRT) model to output the intermediate frequency estimate value, extracts the target characteristic verification parameter of the target information carried by the intermediate frequency estimate value, matches the target characteristic verification parameter with the source characteristic verification parameter in the pre-constructed mapping library, and generates the characteristic verification result. The target characteristic verification parameter includes the frequency change rate , the intermediate frequency range , and the time-frequency ridge line morphology parameter .

[0050] In this embodiment, during target recognition, intermediate frequency (IF) refers to the frequency range of the signal after processing that falls between high frequency and baseband, and carries key target information. During time-frequency analysis, the frequency range of the signal is used to determine the intermediate frequency. Signal-to-noise ratio Dynamically adjust the scale factor of IPCT based on features. The dataset was constructed using this as the target variable. The Gradient Boosting Regression Tree (GBRT) algorithm was employed, with the tree depth set accordingly. 5. Used to control the complexity of each regression tree, assuming the number of nodes in the decision tree is... In the case of an ideal binary tree, the relationship between the number of nodes and the depth satisfies The learning rate is set to 0.1, which controls the step size of the newly generated regression tree in each iteration for updating the model. The model update formula is as follows: ,in For the first Predicted value after the next iteration This is the predicted value from the previous iteration. For learning rate, For the first The predicted values ​​of each regression tree; setting the number of trees to 100 determines the total number of regression trees generated during model training, and the final prediction result of the model is: ; in For the number of trees, For the first The prediction function for each tree was used. Finally, a 5-fold cross-validation was employed, comparing the mean squared error (MSE) across the five validation iterations. The model with the smallest average MSE is selected as the optimal model, and the optimal construction mapping function for that model is obtained: ; Energy concentration reflects the "frequency coverage range" of the signal. The maneuverability of different targets leads to differences in the frequency range of the nonlinear IF signal. Reflects the "degree of noise interference" of the signal. High indicates weak noise. Low indicates high noise; output variable It is a core adjustment parameter of IPCT, which directly controls the "time-frequency resolution adaptation capability" of IPCT wavelet analysis. As the frequency range increases, the frequency coverage of the wavelet expands. When the wavelet is reduced, its temporal resolution is improved.

[0051] The input signal is subjected to a Fast Fourier Transform (FFT) to convert the time-domain signal into a frequency-domain representation, and then the mapping function is applied... The current signal dynamic scale factor is calculated The dynamic scale factor can be expressed as:

[0052] Wherein is the coefficient obtained by using the least square method fitting.

[0053] Then, the calculated dynamic scale factor is substituted into the intermediate frequency estimation algorithm of the polynomial transformation , the formula is as follows: ; Wherein, is the output variable of the IPCT transformation, representing the instantaneous frequency, is the input signal, is the polynomial base function; the expression is: ; Wherein, is the optimal kernel function complex conjugate of the frame, is the frequency carrier term, is the instantaneous frequency of the IPCT output, t is the global analysis time, is the displacement integral variable; is the Gaussian window function, the expression is: ; is the Gaussian window width parameter, is the polynomial parameter vector, expressed as: ; represents the order of the kernel function, represents the optimal coefficient, k=0,…,n.

[0054] After introducing the dynamic scale factor , the base function becomes: ; The peak search is performed to obtain the intermediate frequency estimation value Since the obtained intermediate frequency estimation value carries the frequency range of the target key information, it is used to perform target “characteristic verification”.

[0055] In this embodiment, the output target category probability is represented as the probability value assigned to different targets (such as "aircraft", "ship", "ground armored vehicle", "clutter interference") (the sum of all category probabilities is 1), and the category with the highest probability is taken as the "preliminary identification result". For example, if the probability of the "aircraft" category is 0.88, "ship" is 0.10, and "clutter interference" is 0.02, then the target is initially determined to be "aircraft", and the "probability confidence level" of this preliminary category (i.e., the highest probability value, which is 0.88 here) is recorded.

[0056] In this embodiment, the second-path time-series feature vector is processed by the Gradient Boosting Regression Tree (GBRT) model to output an intermediate frequency estimate, based on... Peak search of time-frequency energy distribution; extraction of three types of target characteristic verification parameters carrying target information from the mid-frequency estimate; matching of the target characteristic verification parameters with source feature verification parameters in a pre-built mapping library to generate characteristic verification results. In this process, the target characteristic verification parameters include the frequency change rate based on the actual value. Mid-frequency range and time-frequency ridge morphology parameters .

[0057] Among them, the rate of change of frequency Represented as: ; Mid-frequency range: It is determined by the frequency range corresponding to the energy peak in the time-frequency distribution, reflecting the frequency coverage characteristics of the target echo.

[0058] Time-frequency ridge morphology parameters :right The ridge lines are fitted to extract morphological features such as ridge smoothness and number of abrupt changes, which are used to distinguish the target from clutter. (Clutter ridge lines usually have irregular abrupt changes, while target ridge lines show continuous dynamic changes).

[0059] Based on the rate of change of frequency Mid-frequency range and time-frequency ridge morphology parameters The actual value is matched with the theoretical value in the mapping library to generate a value based on the frequency change rate error. Mid-frequency range overlap Ridge morphology The results of the characteristic verification; in: ; ; ; Mean value of the frequency rate of change of the preliminary identified category theory, The smaller, the higher the matching degree. Mean value of the mid-frequency range of the theory, The larger, the higher the matching degree. Number of ridge sampling points, 、 Respectively represent the first Point shape value of the actual and theoretical ridges, The smaller, the higher the matching degree.

[0060] According to the preliminary identification result of the target category and the characteristic verification result, the target identification step specifically includes: Obtain the probability confidence of the preliminary identification result and the frequency change rate error of the characteristic verification result , mid-frequency range coincidence degree , ridge shape ; If , and the probability confidence of the preliminary identification result , it is determined that the target identification result is accurate, and the final target category and identification confidence are output; If , or the probability confidence , it is determined that the identification result is invalid; Wherein, the identification confidence is represented as the weighted average calculation result of the probability confidence and the frequency change rate error , mid-frequency range coincidence degree , and ridge shape .

[0061] It should be noted that in the process of constructing the mapping library, the constructed "target category - theoretical mid-frequency characteristic" mapping library is generated by training a large number of sample signals labeled with real target categories. For each target, the mean and threshold range of the mid-frequency characteristics under different maneuvering states and electromagnetic environments are calculated to form a theoretical standard. Through multiple optimizations, a mapping library that meets the threshold range requirements is formed.

[0062] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and are not limited thereto; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent replacements for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A target recognition method based on adaptive polynomial kernel function and multi-modal fusion, characterized in that, The method comprises the following steps: obtaining an original signal of a target to be identified, and pre-processing the original signal based on a wavelet threshold denoising algorithm to generate an analysis signal; performing multi-modal time-frequency analysis on the analysis signal to output a first time-frequency feature map based on an adaptive polynomial kernel function, a second time-frequency feature map based on a Wigner-Ville distribution (WVD), and a third time-frequency feature map based on a short-time Fourier transform (STFT); performing dynamic weighted fusion processing on the first, second, and third time-frequency feature maps to generate a fusion time-frequency feature tensor, and performing convolutional neural network processing to generate a time series feature vector; inputting the time series feature vector into a convolutional neural network and a gradient boosting regression tree (GBRT) model for processing to output a preliminary identification result and a characteristic verification result based on a target category; performing target identification based on the preliminary identification result and the characteristic verification result of the target category.

2. The object recognition method of claim 1, wherein, The step of pre-processing the original signal based on the wavelet threshold denoising algorithm to generate an analysis signal comprises: performing multi-layer wavelet decomposition on the original signal based on kurtosis, and if the kurtosis is greater than 3, the original signal contains impulse noise; performing Fourier transform calculation on the original signal to obtain a signal power spectral density, and calculating a power spectrum flatness based on the signal power spectral density; if the power spectrum flatness is approximately equal to 1, the original signal contains white noise; if the power spectrum flatness is less than 1, the original signal contains colored noise; setting corresponding noise thresholds for white noise, colored noise, and impulse noise, respectively, and generating an analysis signal with high signal-to-noise ratio after wavelet coefficient processing based on the corresponding noise thresholds.

3. The object recognition method of claim 2, wherein, The step of setting corresponding noise thresholds for white noise, colored noise, and impulse noise, respectively, and generating an analysis signal with high signal-to-noise ratio after wavelet coefficient processing based on the corresponding noise thresholds comprises: setting corresponding noise thresholds for white noise, colored noise, and impulse noise, respectively, specifically: White noise employs thresholding ; Color noise employs dynamic threshold ; Impulse noise employs kurtosis adaptive hard threshold ; setting wavelet coefficients less than the noise threshold to zero or attenuating them, and retaining wavelet coefficients greater than the noise threshold; and reconstructing the retained wavelet coefficients by inverse transform to obtain an analysis signal with high signal-to-noise ratio; wherein denotes the noise standard deviation estimate, denotes the original signal length, denotes the noise power spectral density estimate, is the kurtosis coefficient.

4. The object recognition method of claim 1, wherein, The step of performing multi-modal time-frequency analysis on the analysis signal to output a first time-frequency feature map based on an adaptive polynomial kernel function comprises: frame the signal to be analyzed and extract local features to generate a local energy distribution and the rate of change of frequency ; Constructing adaptive polynomial kernel functions with higher order terms is represented as: ； denotes the polynomial coefficients of the frame, denotes the polynomial order, denotes the time axis variable, denotes the shift integral variable in the wavelet transform, is the energy weight factor, , is the energy distribution at time t, is the center time of the frame, denotes the maximum of the local energy distribution of the 1st to the frame at time t. The polynomial coefficients were analyzed using a particle swarm optimization algorithm. Optimization yields the first The optimal polynomial coefficients of the frame, and the optimal adaptive polynomial kernel function; obtaining a time-frequency distribution that adapts to the nonlinear instantaneous frequency characteristics of the analysis signal based on the optimal adaptive polynomial kernel function, and outputting a first time-frequency feature map.

5. The object recognition method of claim 4, wherein, The step of performing multi-modal time-frequency analysis on the analysis signal to output a second time-frequency feature map based on a Wigner-Ville distribution (WVD) and a third time-frequency feature map based on a short-time Fourier transform (STFT) comprises: performing Gaussian window smoothing processing on the analysis signal using the Wigner-Ville distribution (WVD) to suppress cross terms, and outputting a second time-frequency feature map that enhances the energy concentration area after noise suppression processing; Adopt short-time Fourier transform (STFT) to perform adaptive window length processing on the signal to be analyzed, if it is a slowly varying signal, adopt a long window to improve frequency resolution, if it is a fast varying signal, adopt a short window to guarantee time resolution, and finally generate a third time-frequency feature map as a reference.

6. The object recognition method of claim 5, wherein, The step of performing dynamic weighted fusion processing on the first time-frequency feature map, the second time-frequency feature map and the third time-frequency feature map to generate a fused time-frequency feature tensor specifically includes: The first time-frequency feature map, the second time-frequency feature map and the third time-frequency feature map are set to the same time axis and frequency axis, and are adjusted to the same resolution through interpolation; The energy of the first time-frequency feature map, the second time-frequency feature map and the third time-frequency feature map after adjusting the resolution is normalized to eliminate the difference in energy scale; A dynamic weighted fusion method is adopted, and the weights are determined based on the entropy values of the three time-frequency feature maps in different frequency intervals; The energy values of the three time-frequency feature maps at each time-frequency point are weighted and summed according to the determined weights to generate a fused time-frequency feature tensor.

7. The object recognition method of claim 6, wherein, The step of generating a time series feature vector after processing the fused time-frequency feature tensor through a convolutional neural network (CNN) specifically includes: A convolutional neural network (CNN) is used to extract features from the fused time-frequency feature tensor, and different sizes of convolution kernels are used through the convolution layer to capture local time-frequency patterns. After dimension reduction processing through the pooling layer, a local feature vector is output. The local feature vector is unfolded into a time series feature sequence along the time axis, and is input into a recurrent neural network (RNN) for time series modeling. The recurrent neural network (RNN) outputs a time series feature vector through its gating mechanism.

8. The object recognition method of claim 7, wherein, The step of inputting the time series feature vector into a convolutional neural network and a gradient boosting regression tree (GBRT) model for processing to output a preliminary identification result based on the target category and a characteristic verification result specifically includes: The time series feature vector is divided into two paths and input into a convolutional neural network and a gradient boosting regression tree (GBRT) model respectively. The first path of the time series feature vector passes through the fully connected layer of the convolutional neural network (CNN) and the Softmax function to output the probability of the target category, and the category with the highest probability value is taken as the preliminary identification result and the highest probability value is taken as the probability confidence. The second path of the time series feature vector passes through the gradient boosting regression tree (GBRT) model to output a medium frequency estimation value, extracts a target characteristic verification parameter carried by the target information of the medium frequency estimation value, matches the target characteristic verification parameter with a source characteristic verification parameter in a pre-constructed mapping library, and generates a characteristic verification result. The target characteristic verification parameter includes a frequency variation rate based on an actual value mid-frequency range and a time-frequency ridge line morphology parameter .

9. The target identification method according to claim 8, characterized in that: The step of matching the target characteristic verification parameter with a source characteristic verification parameter in a pre-constructed mapping library to generate a characteristic verification result specifically includes: A "target category-theoretical medium frequency characteristic" mapping library is pre-constructed, and each type of target in the mapping library is matched with a theoretical value of the mean and threshold range of the medium frequency characteristic under different maneuvering states and electromagnetic environments; The actual value of the frequency change rate , the intermediate frequency range , and the time-frequency ridge morphology parameter is matched with the theoretical value in the mapping library to generate a frequency change rate error , intermediate frequency range coincidence degree , ridge morphology characteristic verification result; in, , This represents the mean of the theoretical frequency change rate of the initially identified categories. The smaller the value, the higher the match. , This represents the theoretical value for the intermediate frequency range. The larger the value, the higher the match. , Indicates the number of ridge sampling points. , Representing the actual and theoretical ridge lines respectively Point shape value, The smaller the value, the higher the match.

10. The object recognition method of claim 9, wherein, The step of performing target identification according to the preliminary identification result of the target category and the characteristic verification result specifically includes: Probability confidence of the preliminary recognition result and frequency change rate error of the characteristic verification result , degree of coincidence in the intermediate frequency range , ridge line form ; If , and the probability confidence of the preliminary identification result is , then it is determined that the target identification result is accurate, and the final target category and identification confidence are output. If , or the probability confidence , the recognition result is determined to be invalid. The recognition confidence is expressed as a probability confidence and a frequency change rate error The medium frequency range coincidence degree And the ridge line form Weighted average calculation result.

Images