A UWB Fuze Target Recognition Method Based on Physical Feature Extraction and SIFCDN Model

Abstract

The application discloses a UWB fuze target recognition method based on physical feature extraction and a SIFCDN model, and comprises the following steps: one-dimensional sequence features and two-dimensional Doppler image features are extracted for measured UWB fuze target echoes, and the two features are fused to construct fuze target physical features; a sequence-image fusion coding deep network model is designed, comprising a multi-head attention sequence-image feature fusion module, a multi-head attention coding module and a linear classification module; the sequence-image fusion coding deep network model is trained and verified based on the fuze target physical features extracted from measured data, and the best model is saved; the measured fuze target echoes under complex scene conditions are used to extract physical features, the sequence-image fusion coding deep network model is subjected to identification performance test, and the effectiveness of the physical feature extraction and the model is verified.

Description

Technical Field

[0001] The present invention relates to the field of radio fuze target recognition, and in particular to a UWB fuze target recognition method based on physical feature extraction and a SIFCDN model. Background Art

[0002] When a carrier-free ultrawideband (CF-UWB) fuze identifies a target, background clutter often mixes with the target echo, making it difficult to accurately identify the target based solely on a one-dimensional echo sequence. To reduce the interference of background clutter, frequency domain filtering and other methods can be used to minimize the background clutter. Furthermore, signal processing methods can be used to extract the Doppler information of moving targets. Conventional modulated ultrawideband fuzes obtain Doppler information based on the Doppler frequency shift, estimating the target's velocity by using the phase or frequency difference between the transmitted and received signals.

[0003] However, the waveform of the transmitted signal from a carrierless ultra-wideband fuze is generally a Gaussian function and its derivative. Compared to commonly used velocity-measuring fuzes, the transmitted signal is unmodulated, and there is no phase difference between the transmitted and received signals to exploit. Therefore, the principle of Doppler information extraction is different. Specifically, the echo signal matrix of a carrierless ultra-wideband fuze reflects range information in fast time and is a discrete-time sampling of a Gaussian pulse signal in slow time, with the sampling interval proportional to velocity. The spectrum of the slow-time signal is a scaled transformation of the Gaussian signal spectrum, with the scaling factor related to velocity. Therefore, the target's Doppler information can be extracted by performing a short-time Fourier transform on the echo signal in the slow-time dimension.

[0004] Based on the above research, this patent proposes a UWB fuze target recognition technology based on physical feature extraction and the SIFCDN (Sequence Image Fusion Coding Deep Network, SIFCDN) model. First, the physical features of the fuze target are constructed based on the fusion of the target's one-dimensional echo sequence and two-dimensional Doppler image. Secondly, the fused physical features are feature encoded and deeply fused through the attention coding network. Finally, the target prediction result is output through the classifier. This method integrates the one-dimensional sequence features and two-dimensional Doppler image features of the target, comprehensively considers the target's distance, speed, structural information, etc., and upgrades fuze target recognition from static recognition to dynamic recognition, providing a new technical approach for target recognition based on carrier-free ultra-wideband fuzes in moving target scenarios. Summary of the Invention

[0005] In order to overcome the deficiencies in the prior art, the present invention provides a UWB fuze target recognition method based on physical feature extraction and SIFCDN model, which can comprehensively utilize the one-dimensional sequence features and two-dimensional Doppler image features of the target to improve the model's recognition performance for moving targets.

[0006] In order to achieve the above-mentioned purpose of the invention, the technical solutions adopted to solve the technical problems are as follows:

[0007] A UWB fuze target recognition method based on physical feature extraction and SIFCDN model includes the following steps:

[0008] Step S100: extracting the physical characteristics of the ultra-wideband fuze target;

[0009] Based on the measured echo of ultra-wideband fuze target, one-dimensional sequence features and two-dimensional Doppler image features are extracted, and the two features are fused to construct the physical characteristics of the fuze target.

[0010] Step S200: Design of sequence-image fusion coding deep network SIFCDN model;

[0011] Design the SIFCDN model, including the multi-head attention sequence-image feature fusion (MSAF) module, the multi-head attention encoding (MSAC) module, and the DNN classification module.

[0012] Step S300: training and verification of SIFCDN model;

[0013] Based on the physical characteristics of the fuze target extracted from the measured data, the SIFCDN model is trained and verified, and the best model is saved;

[0014] Step S400: SIFCDN model validity and generalization analysis;

[0015] The measured fuze target echoes under complex scene conditions are used to extract physical features, and the recognition performance of the saved SIFCDN optimal model is tested to verify the effectiveness of the physical feature extraction and model.

[0016] Furthermore, in step S100, the method for extracting the physical characteristics of the fuze target specifically includes:

[0017] Firstly, the fuze target echoes of M consecutive pulse periods are arranged into a fast-slow time matrix according to the pulse period, and the middle sequence of the matrix in the slow time dimension is extracted as the one-dimensional sequence feature;

[0018] Secondly, perform M-point discrete Fourier transform in the slow time dimension and traverse the fast time dimension to obtain the two-dimensional Doppler image characteristics of the target;

[0019] Finally, the physical characteristics of the target are obtained by combining the one-dimensional sequence features and the two-dimensional Doppler image features. When obtaining the two-dimensional Doppler image through range-Doppler processing, the influence of range ambiguity (or velocity ambiguity) caused by the high (or low) pulse repetition frequency of the fuze can be ignored.

[0020] Furthermore, in step S200, the structure of the SIFCDN model includes a multi-head attention sequence-image feature fusion MSAF module, a multi-head attention encoding MSAC module and a DNN classification module. The model specifically includes:

[0021] The MSAF module includes:

[0022] (1) Two-channel feature extraction layer: the first linear layer Linear0-1, the number of input nodes is 1; the second parallel linear layer Linear0-2, the number of input nodes is 64; Linear0-1 and Linear0-2 both use ReLU activation, the number of output nodes is 128; gating function F;

[0023] (2) Feature fusion layer, including three attention linear layers Linear1-1 to Linear1-3 and a multi-head attention layer. The number of heads in the multi-head attention layer is 8. Two of the three input channels are fusion feature representations, and one channel is Doppler feature representation;

[0024] The MSAC module consists of two layers, each of which includes: LayerNorm layer, multi-head attention layer, LayerNorm layer, MLP layer, the multi-head attention layer has 4 heads, and the three input channels all have the same features; the MLP layer includes the first linear layer, ReLU activation, dropout layer, second linear layer, dropout layer, where the dropout layer randomly drops 1 / 10 of the nodes. The number of output nodes of the first linear layer in the MLP is 512, and the number of output nodes of the second linear layer is 128;

[0025] The DNN classification module includes a linear layer with 4 output nodes.

[0026] Furthermore, the method of fusing the physical features of the fuze target through the multi-head attention sequence-image feature fusion MSAF module includes:

[0027] First, preliminary feature extraction is performed on the target echo sequence and Doppler image through two parallel linear layers, namely:

[0028]

[0029] Among them, S is the target echo sequence, I is the Doppler image of the target, f1 and f2 are the mapping functions corresponding to the Linear0-1 and Linear0-2 linear layers respectively, and fea s ,fes i They are sequence feature representation and Doppler feature representation respectively;

[0030] Secondly, the two features are weighted and summed through the gating function to obtain the fusion feature representation, that is:

[0031] fea fus =λfea s +(1-λ)fea i ,

[0032] Among them, λ is the learnable weight parameter in the gating function, initialized to 0.5, fes fus is the fusion feature representation;

[0033] Finally, based on the Doppler feature representation fea i Get the Value vector and represent the fea based on the fusion feature fus Get the Query vector and Key vector, that is:

[0034]

[0035] Among them, W q ,W k ,W v It is the mapping function corresponding to the attention weight layer Linear1-1 to Linear1-3;

[0036] Furthermore, multi-head attention sequence-image feature fusion is performed to obtain the output of the MSAF module:

[0037]

[0038] Among them, d k The dimension size of the Key vector.

[0039] Furthermore, in step S300, the SIFCDN model is trained and verified based on the physical characteristics of the fuze target extracted from the measured data. When the recognition performance is verified to be stable, the training is stopped and the best model is saved.

[0040] Furthermore, in step S400, the SIFCDN model is tested by extracting physical features from the measured fuze target echo under complex scene conditions, and the recognition rate and confusion matrix are statistically analyzed to verify the effectiveness of the model;

[0041] Ablation experiments were conducted, using the target echo sequence or Doppler image as a single feature input to train, verify, and test the model. The recognition performance of different input models was compared and analyzed to verify the effectiveness of the physical features.

[0042] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art:

[0043] The present invention discloses a UWB fuze target recognition technology based on physical feature extraction and SIFCDN model. The technology constructs target physical features based on the target's one-dimensional echo sequence and two-dimensional Doppler image, and performs feature encoding and deep fusion of the target physical features through an attention coding network. The technology comprehensively considers the target's distance, speed, structural information, etc., and improves fuze target recognition from static recognition to dynamic recognition. This technology provides a new technical approach for target recognition based on carrier-free ultra-wideband fuze in moving target scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without inventive work. In the drawings:

[0045] Figure 1 This is a flow chart of a UWB fuze target recognition technology based on physical feature extraction and SIFCDN model of the present invention;

[0046] Figure 2 These are the four measured targets in this invention: (a) person; (b) car; (c) SUV; (d) truck;

[0047] Figure 3 Doppler images of the measured targets in the present invention; (a) a moving person (v = 1 m / s); (b) a moving car (v = 6 m / s); (c) a moving SUV (v = 3 m / s); (d) a moving truck (v = 10 m / s);

[0048] Figure 4 The training / validation loss and training / validation accuracy of the SIFCDN model in the present invention based on the measured target fusion physical features; (a) training / validation loss; (b) training / validation accuracy;

[0049] Figure 5 Confusion matrix of the SIFCDN model in the present invention tested based on different inputs; (a) sequence features only; (b) Doppler image features only; (c) sequence-Doppler image fusion physical features. DETAILED DESCRIPTION

[0050] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0051] like Figure 1 As shown, this embodiment discloses a UWB fuze target recognition method based on physical feature extraction and SIFCDN model, comprising the following steps:

[0052] Step S100: extracting the physical characteristics of the ultra-wideband fuze target;

[0053] Based on the measured echo of ultra-wideband fuze target, one-dimensional sequence features and two-dimensional Doppler image features are extracted, and the two features are fused to construct the physical characteristics of the fuze target.

[0054] Step S200: Sequence Image Fusion Coding Deep Network (SIFCDN) model design;

[0055] Design the SIFCDN model, which includes a Multi-head Attention Fusion (MSAF) module, a Multi-head Attention Coding (MSAC) module, and a DNN classification module.

[0056] Step S300: training and verification of SIFCDN model;

[0057] Based on the physical characteristics of the fuze target extracted from the measured data, the SIFCDN model is trained and verified, and the best model is saved;

[0058] Step S400: SIFCDN model validity and generalization analysis;

[0059] The measured fuze target echoes under complex scene conditions are used to extract physical features, and the recognition performance of the saved SIFCDN optimal model is tested to verify the effectiveness of the physical feature extraction and model.

[0060] The following is a detailed description of each of the above steps:

[0061] In step S100, the method for extracting the physical characteristics of the fuze target specifically includes:

[0062] Firstly, the fuze target echoes of M consecutive pulse periods are arranged into a fast-slow time matrix according to the pulse period, and the middle sequence of the matrix in the slow time dimension is extracted as the one-dimensional sequence feature;

[0063] Secondly, perform M-point discrete Fourier transform in the slow time dimension and traverse the fast time dimension to obtain the two-dimensional Doppler image characteristics of the target;

[0064] Finally, the physical characteristics of the target are obtained by combining the one-dimensional sequence features and the two-dimensional Doppler image features. When obtaining the two-dimensional Doppler image through range-Doppler processing, the influence of range ambiguity (or velocity ambiguity) caused by the high (or low) pulse repetition frequency of the fuze can be ignored.

[0065] In step S100, the physical characteristics are obtained based on the measured ultra-wideband fuse echo.

[0066] The fuze emission signal can be expressed as:

[0067]

[0068] Among them, p(t) is the Gaussian pulse signal, δ(t) is the impulse signal, T r is the pulse repetition period, and M is the number of pulses.

[0069] The fuze receiving signal can be expressed as:

[0070]

[0071] in, t d is the time delay of the received signal relative to the transmitted signal, A is the amplitude attenuation coefficient of the received signal, and the M / 2th echo sequence is extracted as the one-dimensional sequence feature.

[0072] Arrange the M received signals into a fast and slow time matrix, slice the fast and slow time matrix at the fast time τ, and obtain the sliced ​​signal at the time τ:

[0073]

[0074] Perform DFT transformation to obtain the discrete spectrum of the slice signal at time τ:

[0075]

[0076] According to the above formula, the discrete spectra of other fast time slices and DFT transformation are calculated, and the discrete spectra of all moments are spliced ​​in the fast time sequence to obtain the two-dimensional Doppler image characteristics of the target.

[0077] In step S100, the target data types include: people, cars, SUVs, and trucks. The schematic diagrams of the four targets are as follows: Figure 2 , the target categories and labels correspond to Table 1:

[0078] Target Label people 0 sedan 1 SUV 2 Large truck 3

[0079] Table 1 Target echo signal labels

[0080] Figure 3 Doppler images of four targets are listed for comparison.

[0081] In step S200, the structure of the SIFCDN model includes a multi-head attention sequence-image feature fusion MSAF module, a multi-head attention encoding MSAC module and a DNN classification module. The structure of the model is referenced Figure 1 , specifically including:

[0082] The MSAF module includes:

[0083] (1) Two-channel feature extraction layer: the first linear layer Linear0-1, the number of input nodes is 1; the second parallel linear layer Linear0-2, the number of input nodes is 64; Linear0-1 and Linear0-2 both use ReLU activation, the number of output nodes is 128; gating function F;

[0084] (2) Feature fusion layer, including three attention linear layers Linear1-1 to Linear1-3 and a multi-head attention layer. The number of heads in the multi-head attention layer is 8. Two of the three input channels are fusion feature representations, and one channel is Doppler feature representation;

[0085] The MSAC module consists of two layers, each of which includes: LayerNorm layer, multi-head attention layer, LayerNorm layer, MLP layer, the multi-head attention layer has 4 heads, and the three input channels all have the same features; the MLP layer includes the first linear layer, ReLU activation, dropout layer, second linear layer, dropout layer, where the dropout layer randomly drops 1 / 10 of the nodes. The number of output nodes of the first linear layer in the MLP is 512, and the number of output nodes of the second linear layer is 128;

[0086] The DNN classification module includes a linear layer with 4 output nodes.

[0087] The parameter settings of the above layers are shown in Table 2:

[0088]

[0089] Table 2 Parameter settings of each layer of SIFCDN model

[0090] The activation function of the above DNN classification module adopts the Softmax function, and the activation function of the remaining layers adopts the ReLU function; the Dropout layer discards 1 / 10 of the original nodes.

[0091] Furthermore, the method for fusing the physical features of the fuze target through the multi-head attention sequence-image feature fusion MSAF module includes:

[0092] First, preliminary feature extraction is performed on the target echo sequence and Doppler image through two parallel linear layers, namely:

[0093]

[0094] Among them, S is the target echo sequence, I is the Doppler image of the target, f1 and f2 are the mapping functions corresponding to the Linear0-1 and Linear0-2 linear layers respectively, and fea s ,fes i They are sequence feature representation and Doppler feature representation respectively;

[0095] Secondly, the two features are weighted and summed through the gating function to obtain the fusion feature representation, that is:

[0096] fea fus =λfea s +(1-λ)fea i ,

[0097] Among them, λ is the learnable weight parameter in the gating function, initialized to 0.5, fes fus It is the fusion feature representation;

[0098] Finally, based on the Doppler feature representation fea i Get the Value vector and represent the fea based on the fusion feature fus Get the Query vector and Key vector, that is:

[0099]

[0100] Among them, W q ,W k ,W v It is the mapping function corresponding to the attention weight layer Linear1-1 to Linear1-3;

[0101] Furthermore, multi-head attention sequence-image feature fusion is performed to obtain the output of the MSAF module:

[0102]

[0103] Among them, dk The dimension size of the Key vector.

[0104] In step S300, the SIFCDN model is trained and verified based on the physical characteristics of the fuze target extracted from the measured data. When the recognition performance is verified to be stable, the training is stopped and the best model is saved. The training parameters are set as shown in Table 3:

[0105]

[0106]

[0107] Table 3 Model training parameter settings

[0108] The training loss and recognition accuracy of the model on the physical features of the fuze target extracted from the measured data are as follows: Figure 4 As shown in the figure, the model reaches convergence at about 800 iterations, the recognition rate of the validation set is above 95%, and the best training model is saved.

[0109] In step S400, the measured fuze target echoes under complex scene conditions such as grass, shrubs, sand and stones are used to extract physical features to test the SIFCDN model, and the recognition rate and confusion matrix are statistically analyzed. The results are as follows: Figure 5 ,As shown in Table 4, the average recognition rate of the model is 96.7%, and the recall rate of all categories exceeds 90%, ,which verifies the generalization recognition ability of the model in complex ,scenarios.

[0110] To verify the effectiveness of the physical features, this embodiment uses the target echo sequence or Doppler image as a single feature input to train, verify and test the model, and compares and analyzes the recognition performance of different input models. The results are as follows: Figure 5 , as shown in Table 4. As can be seen from the chart, the model recognition performance of the single sequence feature input is the worst, the model recognition performance of the single Doppler image input is slightly better, and the model performance based on the sequence-Doppler image fusion physical feature input is the best, which verifies the effectiveness of the fusion physical feature.

[0111] Input features Average recognition rate Single sequence features 87.1％ Single Doppler signature 93.8％ Fusion of physical features 96.7％

[0112] Table 4 Generalization analysis of observation azimuth

[0113] Therefore, this embodiment provides a UWB fuze target recognition technology based on physical feature extraction and SIFCDN model, which constructs the target physical features based on the target's one-dimensional echo sequence and two-dimensional Doppler image, and performs feature encoding and deep fusion of the target physical features through an attention coding network. It comprehensively considers the target's distance, speed, structural information, etc., and improves the fuze target recognition from static recognition to dynamic recognition, providing a new technical approach for target recognition based on carrier-free ultra-wideband fuze in moving target scenarios.

[0114] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A UWB fuze target recognition method based on physical feature extraction and SIFCDN model, characterized in that: The following steps are involved: Step S100: extracting the physical characteristics of the ultra-wideband fuze target; Based on the measured echo of ultra-wideband fuze target, one-dimensional sequence features and two-dimensional Doppler image features are extracted, and the two features are fused to construct the physical characteristics of the fuze target. In step S100, the data types of the target include: person, car, SUV, truck; Step S200: Design of sequence-image fusion coding deep network SIFCDN model; Design the SIFCDN model, including the multi-head attention sequence-image feature fusion (MSAF) module, the multi-head attention encoding (MSAC) module, and the DNN classification module. In step S200, the structure of the SIFCDN model includes a multi-head attention sequence-image feature fusion MSAF module, a multi-head attention encoding MSAC module and a DNN classification module. The model specifically includes: The MSAF module includes: (1) Two-channel feature extraction layer: the first linear layer Linear0-1, the number of input nodes is 1; the second parallel linear layer Linear0-2, the number of input nodes is 64; Linear0-1 and Linear0-2 both use ReLU activation, the number of output nodes is 128; gating function F; (2) Feature fusion layer, including three attention linear layers Linear1-1 to Linear1-3 and a multi-head attention layer. The number of heads in the multi-head attention layer is 8. Two of the three input channels are fusion feature representations, and one channel is Doppler feature representation; The MSAC module consists of two layers, each of which includes: LayerNorm layer, multi-head attention layer, LayerNorm layer, MLP layer, the multi-head attention layer has 4 heads, and the three input channels all have the same features; the MLP layer includes the first linear layer, ReLU activation, dropout layer, second linear layer, dropout layer, where the dropout layer randomly drops 1 / 10 of the nodes. The number of output nodes of the first linear layer in the MLP is 512, and the number of output nodes of the second linear layer is 128; The DNN classification module includes a linear layer with 4 output nodes; The method of fusing the physical features of the fuze target through the multi-head attention sequence-image feature fusion MSAF module includes: First, preliminary feature extraction is performed on the target echo sequence and Doppler image through two parallel linear layers, namely: Among them, S is the target echo sequence, I is the Doppler image of the target, f1 and f2 are the mapping functions corresponding to the Linear0-1 and Linear0-2 linear layers respectively, and fea s ,fes i They are sequence feature representation and Doppler feature representation respectively; Secondly, the two features are weighted and summed through the gating function to obtain the fusion feature representation, that is: fea fus =λfea s +(1-λ)fea i , Among them, λ is the learnable weight parameter in the gating function, initialized to 0.5, fes fus is the fusion feature representation; Finally, based on the Doppler feature representation fea i Get the Value vector and represent the fea based on the fusion feature fus Get the Query vector and Key vector, that is: Among them, W q ,W k ,W v is the mapping function corresponding to the attention weight layers Linear1-1 to Linear1-3; furthermore, multi-head attention sequence-image feature fusion is performed to obtain the output of the MSAF module: Among them, d k is the dimension size of the Key vector; Step S300: training and verification of SIFCDN model; Based on the physical characteristics of the fuze target extracted from the measured data, the SIFCDN model is trained and verified, and the best model is saved; Step S400: SIFCDN model validity and generalization analysis; The measured fuze target echoes under complex scene conditions are used to extract physical features, and the recognition performance of the saved SIFCDN optimal model is tested to verify the effectiveness of the physical feature extraction and model.

2. The UWB fuze target recognition method based on physical feature extraction and SIFCDN model according to claim 1 is characterized in that: In step S100, the method for extracting the physical characteristics of the fuze target specifically includes: Firstly, the fuze target echoes of M consecutive pulse periods are arranged into a fast-slow time matrix according to the pulse period, and the middle sequence of the matrix in the slow time dimension is extracted as the one-dimensional sequence feature; Secondly, perform M-point discrete Fourier transform in the slow time dimension and traverse the fast time dimension to obtain the two-dimensional Doppler image characteristics of the target; Finally, the physical characteristics of the target are obtained by combining the one-dimensional sequence features and the two-dimensional Doppler image features. When obtaining the two-dimensional Doppler image through range-Doppler processing, the influence of range ambiguity (or velocity ambiguity) caused by the high (or low) pulse repetition frequency of the fuze can be ignored.

3. The UWB fuze target recognition method based on physical feature extraction and SIFCDN model according to claim 1 is characterized in that: In step S300, the SIFCDN model is trained and verified based on the physical characteristics of the fuze target extracted from the measured data. When the recognition performance is verified to be stable, the training is stopped and the best model is saved.

4. The UWB fuze target recognition method based on physical feature extraction and SIFCDN model according to claim 1 is characterized in that: In step S400, the SIFCDN model is tested by extracting physical features from the measured fuze target echo under complex scene conditions, and the recognition rate and confusion matrix are statistically analyzed to verify the effectiveness of the model; Ablation experiments were conducted, using the target echo sequence or Doppler image as a single feature input to train, verify, and test the model. The recognition performance of different input models was compared and analyzed to verify the effectiveness of the physical features.

Images