HRRP unsupervised identification method, device and equipment based on comparative learning fusion model

Abstract

The invention relates to an HRRP unsupervised identification method, device and equipment based on a comparative learning fusion model. The method comprises the following steps: constructing an adaptive dynamic attention model based on a comparative learning framework; and inputting the HRRP signal sample into a dynamic mixed attention network, and performing shallow feature extraction, middle attention processing and fine-grained modeling in the dynamic mixed attention network to output multi-scale features. And designing multi-scale statistics according to the multi-scale features to adaptively adjust the attention weight so as to output the residual error of the dynamic mixed attention network. And inputting the residual error into a self-adaptive comparison loss module for loss calculation, evaluating the calculated loss through a preset evaluation module to obtain a prediction category of the self-adaptive dynamic attention model, and training the model according to the prediction category of the self-adaptive dynamic attention model. And identifying a to-be-identified HRRP signal by using the trained model to obtain a target image. By adopting the method, the target identification precision can be improved during label-free processing.

Description

Technical Field

[0001] The present invention relates to the field of radar target recognition technology, and in particular to a HRRP unsupervised recognition method, device and equipment based on a contrastive learning fusion model. Background Art

[0002] Radar target recognition is a key task in radar systems. It aims to identify target type, attitude, and position by analyzing radar echo signals. High-resolution range profile (HRRP), an important radar target recognition technology, has been widely studied and applied in recent years. HRRP is the sum of the projection vectors of target scattering point sub-echoes along the radar ray, acquired via broadband radar signals. It provides information about the distribution of target scattering points along the range direction and offers advantages such as high resolution, one-dimensional data, and target structural features.

[0003] With the rapid development of deep learning technology, high-resolution range profile (HRRP) recognition has gradually shifted from traditional methods relying on handcrafted features to a data-driven intelligent paradigm. In the deep learning era, researchers not only extract the spatial local features and temporal dependencies of HRRP using convolutional neural networks (CNNs) and recurrent neural networks (RNNs), respectively, but also innovatively introduce multi-scale dilated convolutions, temporal convolutional networks (TCNs), and Transformer architectures to further enhance sensitivity to the distance between scattering centers and the ability to model long-range correlations. For example, the hybrid CNN-TCN model captures global dependencies across range units through dilated convolutions, significantly improving recognition robustness under low signal-to-noise ratio (SNR) conditions. The HRRP-ViT model, based on a self-attention mechanism, dynamically focuses on key scattering regions, effectively mitigating feature shift in cross-polarization scenarios. Although HRRP recognition technology has made significant progress, there is still room for improvement in recognition accuracy when strong supervisory information, such as labels, is difficult to obtain. Summary of the Invention

[0004] Based on this, it is necessary to provide an adaptive dual-attention Siamese network (ADASiam) HRRP recognition method, device and equipment based on a contrastive learning fusion model, which can improve the HRRP target recognition accuracy in unlabeled processing, in order to address the above technical problems.

[0005] An unsupervised HRRP recognition method based on a contrastive learning fusion model, the method comprising:

[0006] Based on the contrastive learning framework, an adaptive dynamic attention model is constructed. The adaptive dynamic attention model includes a dynamic hybrid attention network and an adaptive contrastive loss module.

[0007] The HRRP signal samples are input into the dynamic hybrid attention network, where shallow feature extraction, mid-level attention processing, and fine-grained modeling are performed to output multi-scale features.

[0008] Multi-scale statistics are designed based on multi-scale features to adaptively adjust the attention weights to output the residual of the dynamic hybrid attention network.

[0009] The residual is input into the adaptive contrast loss module for loss calculation, and the calculated loss is evaluated by the pre-set evaluation module to obtain the predicted category of the adaptive dynamic attention model, so as to train the adaptive dynamic attention model.

[0010] The trained adaptive dynamic attention model is used to identify the HRRP signal to obtain the target image.

[0011] A HRRP unsupervised recognition device based on a contrastive learning fusion model, the device comprising:

[0012] The model building module is used to build an adaptive dynamic attention model based on the contrastive learning framework. The adaptive dynamic attention model includes a dynamic hybrid attention network and an adaptive contrastive loss module.

[0013] The multi-scale feature acquisition module is used to input HRRP signal samples into the dynamic hybrid attention network, where shallow feature extraction, mid-level attention processing, and fine-grained modeling are performed to output multi-scale features.

[0014] The residual output module is used to design multi-scale statistics based on multi-scale features to adaptively adjust the attention weights, thereby outputting the residual of the dynamic hybrid attention network.

[0015] The adaptive loss module is used to input the residual into the adaptive contrast loss module for loss calculation, evaluate the calculated loss through the pre-set evaluation module, and obtain the predicted category of the adaptive dynamic attention model, so as to train the adaptive dynamic attention model.

[0016] The target recognition module is used to identify the HRRP signal to be identified using the trained adaptive dynamic attention model to obtain the target image.

[0017] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0018] Based on the contrastive learning framework, an adaptive dynamic attention model is constructed. The adaptive dynamic attention model includes a dynamic hybrid attention network and an adaptive contrastive loss module.

[0019] The HRRP signal samples are input into the dynamic hybrid attention network, where shallow feature extraction, mid-level attention processing, and fine-grained modeling are performed to output multi-scale features.

[0020] Multi-scale statistics are designed based on multi-scale features to adaptively adjust the attention weights to output the residual of the dynamic hybrid attention network.

[0021] The residual is input into the adaptive contrast loss module for loss calculation, and the calculated loss is evaluated by the pre-set evaluation module to obtain the predicted category of the adaptive dynamic attention model, so as to train the adaptive dynamic attention model.

[0022] The trained adaptive dynamic attention model is used to identify the HRRP signal to obtain the target image.

[0023] The aforementioned unsupervised HRRP recognition method, apparatus, and device based on a contrastive learning fusion model employs an adaptive dynamic attention model comprising a dynamic hybrid attention network and an adaptive contrastive loss module. The dynamic hybrid attention network processes HRRP signal samples in a layered manner, first performing shallow feature extraction to capture the signal's basic information; then performing mid-level attention processing to enhance key features; and finally, implementing fine-grained modeling to output multi-scale features for a comprehensive characterization of the signal. Furthermore, multi-scale statistics designed based on these multi-scale features adaptively adjust attention weights and output network residuals, enabling the model to more accurately focus on different features. During training, the residuals are input into the adaptive contrastive loss module to calculate the loss. The evaluation module then evaluates the loss to determine the model's prediction category, completing the training of the adaptive dynamic attention model. This contrastive learning-based training approach eliminates the need for extensive label information and instead learns the target's feature representation by mining the differences and similarities in the data's own features. Furthermore, when presented with HRRP signals to be recognized, the trained adaptive dynamic attention model accurately identifies the target and outputs the target image, leveraging the feature patterns learned during unlabeled training. By combining an adaptive dynamic attention model with a contrastive learning framework, this technical solution avoids reliance on strong supervisory information. Instead, it leverages the intrinsic structure and feature relationships of the data to achieve deep mining and precise matching of target features under unlabeled conditions, significantly improving target recognition accuracy and providing an innovative and effective solution for target recognition in unlabeled scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1Flowchart of an unsupervised HRRP recognition method based on a contrastive learning fusion model in one embodiment;

[0025] Figure 2 Flowchart of the HRRP unsupervised recognition steps based on the contrastive learning fusion model in one embodiment

[0026] Figure 3 1. is a diagram showing the overall structure of an ADASiam model in one embodiment;

[0027] Figure 4 1. A diagram showing the backbone network structure of an ADASiam model in one embodiment;

[0028] Figure 5 1 is a recognition accuracy curve of HRRP when the adaptive contrast loss and dynamic hybrid channel attention mechanism modules are respectively removed by the method in one embodiment;

[0029] Figure 6 FIG1 is a structural block diagram of an unsupervised HRRP recognition device based on a contrastive learning fusion model in one embodiment;

[0030] Figure 7 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0032] In one embodiment, Figure 1 As shown, an unsupervised HRRP recognition method based on a contrastive learning fusion model is provided, which includes the following steps:

[0033] Step 102: construct an adaptive dynamic attention model based on the contrastive learning framework.

[0034] The adaptive dynamic attention model includes: dynamic hybrid attention network and adaptive contrast loss module.

[0035] In step 104, the HRRP signal sample is input into the dynamic hybrid attention network, where shallow feature extraction, mid-level attention processing, and fine-grained modeling are performed to output multi-scale features.

[0036] In step 106, a multi-scale statistic is designed based on the multi-scale features to adaptively adjust the attention weights, thereby outputting the residual of the dynamic hybrid attention network.

[0037] In step 108, the residual is input into the adaptive contrast loss module for loss calculation, and the calculated loss is evaluated by a pre-set evaluation module to obtain the predicted category of the adaptive dynamic attention model, thereby training the adaptive dynamic attention model.

[0038] Step 110: Use the trained adaptive dynamic attention model to identify the HRRP signal to obtain a target image.

[0039] In the aforementioned unsupervised HRRP recognition method based on contrastive learning fusion models, the adaptive dynamic attention model comprises a dynamic hybrid attention network and an adaptive contrastive loss module. The dynamic hybrid attention network processes HRRP signal samples in a layered manner, first performing shallow feature extraction to capture the signal's basic information; then using mid-level attention processing to enhance key features; and finally, fine-grained modeling to output multi-scale features for a comprehensive characterization of the signal. Furthermore, multi-scale statistics designed based on these multi-scale features adaptively adjust attention weights and output network residuals, enabling the model to more accurately focus on different features. During training, the residuals are input into the adaptive contrastive loss module to calculate the loss. The evaluation module evaluates the loss and determines the model's prediction category, completing the training of the adaptive dynamic attention model. This contrastive learning-based training approach eliminates the need for extensive label information and instead learns the target's feature representation by mining the differences and similarities in the data's own features. Furthermore, when presented with HRRP signals to be recognized, the trained adaptive dynamic attention model accurately identifies the target and outputs the target image, leveraging the feature patterns learned during unlabeled training. By combining an adaptive dynamic attention model with a contrastive learning framework, this technical solution avoids reliance on strong supervisory information. Instead, it leverages the intrinsic structure and feature relationships of the data to achieve deep mining and precise matching of target features under unlabeled conditions, significantly improving target recognition accuracy and providing an innovative and effective solution for target recognition in unlabeled scenarios.

[0040] In one embodiment, the adaptive dynamic attention model is:

[0041] z1=f θ (Τ1(x))

[0042] z2=f θ (Τ2(x))

[0043]

[0044] L=1 / 2[D(p1,z2)+D(p1,z2)]

[0045] Where x is the HRRP signal, z1 and z2 are the HRRP signal samples used for contrastive learning, p1 and p2 are the output features of the projection head, and f θ is an encoder containing dynamic attention, Τ1 and Τ2 are data enhancement operations, is the prediction head, D is the adaptive contrast loss function, and L is the loss function of the contrastive learning framework.

[0046] In one embodiment, after the HRRP signal is input into the dynamic hybrid attention network, the convolution kernel extracted by the shallow feature is used to perform typical scattering peak width matching to obtain the first layer prediction feature:

[0047]

[0048] Among them, F1 is the first layer prediction feature, ReLU is the rectified linear unit function, BN is normalization, Conv1D7 is the 7×1 convolution kernel, x is the HRRP signal, The first layer prediction features are input to the convolution kernel of the middle layer attention processing to capture the correlation of the mid-range scattering points, and the second layer prediction features are obtained:

[0049]

[0050] Among them, F2 is the second-layer prediction feature, DynamicAttn is the dynamic attention mechanism function, and Conv1D5 is a 5×1 convolution kernel. The second-layer prediction feature is input into the convolution kernel of the fine-grained model to extract the microscopic differences between adjacent scattering points and obtain multi-scale features:

[0051]

[0052] Among them, F3 is the multi-scale feature, AdaptiveAvgPool is the adaptive average pooling function, and Conv1D3 is the 3×1 convolution kernel.

[0053] In one embodiment, a multi-scale statistic is designed based on the multi-scale features:

[0054]

[0055] The attention weights are adaptively adjusted according to the multi-scale statistics to output the residual of the dynamic hybrid attention network:

[0056]

[0057] in, is the residual of the dynamic hybrid attention network, s c is the attention weight.

[0058] In one embodiment, the contribution weights of the multi-scale statistics are obtained based on feature concatenation and nonlinear mapping of the multi-scale statistics:

[0059]

[0060] in, is the contribution weight, W1 and W2 are weight matrices, Softmax is the activation function, and GELU is the smooth activation function. The channel is recalibrated according to the contribution weight to generate the attention weight:

[0061]

[0062] Among them, s c is the attention weight, are weight coefficients, W3 and W4 are linear transformation matrices, and Sigmoid is the activation function. The attention weights are adaptively adjusted to output the residual of the dynamic hybrid attention network.

[0063] In one embodiment, the pre-set evaluation module is used to perform weighted K-nearest neighbor classification based on the feature library constructed based on the memory data set, and then select labeled sample data for weighted classification.

[0064] In one embodiment, the calculated loss is evaluated by a preset evaluation module, and the score of each category is calculated respectively. The category corresponding to the maximum value of the score is taken as the predicted category of the adaptive dynamic attention model:

[0065]

[0066] in, is the maximum score, Score(c) is the loss evaluation score, Ι is the identity matrix, y i is the predicted category, c is the category label, is the weight coefficient.

[0067] In one embodiment, Figure 2 As shown in FIG, an unsupervised HRRP recognition step based on a contrastive learning fusion model is provided, and the specific contents are as follows:

[0068] S1 performs data preprocessing and data enhancement on the radar echo data:

[0069] S1.1 performs pre-processing on HRRP:

[0070] The original HRRP data contains four polarization modes: HH, HV, VV, and VH. Here, the HH polarization mode is uniformly selected for processing and analysis. First, the complex matrix COHH is extracted from the MAT file. Its dimension is M×N, where M represents the number of samples and N represents the number of range cells. We extract the complex row vector cohh_data for each sample:

[0071] cohh_data=CoHH[i,:]

[0072] Where i represents the i-th sample. Then all HRRP data are length-normalized and amplitude-normalized. For each sample X, its modulus is:

[0073]

[0074] The normalization factor is:

[0075]

[0076] The normalized modulus is:

[0077]

[0078] Then the center of gravity of HRRP is aligned, and its center of gravity position is:

[0079]

[0080] The offset is: shift = center - centroid, where center is half the length of the HRRP data. The aligned signal is:

[0081] x aligned [i]=roll(x norm , shift)

[0082] Finally, it is saved as an npy file. After reading, unifying the length, normalizing, and center-of-gravity alignment, the HRRP data is converted into a standardized feature vector, providing a high-quality data foundation for subsequent analysis and processing.

[0083] S1.2 performs data enhancement processing on the HRRP data obtained in S1.1:

[0084] After processing in S1.1, the HRRP data is divided into training, validation, and test sets. Data augmentation is then performed after the data is read. Based on the physical characteristics of HRRP, four main data augmentation strategies are employed: sequence-preserving time shifting, adaptive noise injection, local amplitude scaling, and continuous region occlusion.

[0085] S1.2.1 Sequence-preserving timing offset: The physical meaning is to simulate the radial motion of the target relative to the radar, keep the order of the range units unchanged, and avoid destroying the physical structure of HRRP. For the HRRP signal x = [x1, x2, ..., x L ], randomly translate s units: x′=roll(x,s), s~U(-s max ,8 max ),s max =50, where s max =50 is the maximum offset limit to ensure the integrity of the target scattering structure.

[0086] S1.2.2 Adaptive noise injection: The physical meaning is to dynamically adjust the noise intensity based on the signal energy to avoid strong signals being masked or weak signals being over-noised. Dynamically adjust the noise intensity based on signal energy:

[0087]

[0088] Generate noise: ∈~U(0.05, 0.15), this design ensures that the noise intensity is positively correlated with the signal energy and avoids signal-to-noise ratio imbalance.

[0089] S1.2.3 Local Amplitude Scaling: The physical meaning is to improve the model's ability to capture key target area features and enhance the model's detail perception by simulating the natural variation of the target's local scattering intensity. First, randomly determine the scaling window, with the starting position a and window length w satisfying x′=ReplacePart[x,Table[i→0,{i,b,b+m}]], w=Floor[0.3·L], and randomly scale the signal amplitude within the selected window:

[0090]

[0091] S1.2.4 Continuous Area Occlusion: The physical meaning is to simulate the scene where the radar signal is partially lost, forcing the model to rely on the features of the non-occluded area for recognition, thereby improving the robustness of the model and the efficiency of feature utilization. Randomly determine the occlusion area:

[0092] m~Uniform(0.1L, 0.3L), b~Uniform(0, Lm)

[0093] To perform a zero operation on the selected area:

[0094]

[0095] S2 builds the ADASiam network structure:

[0096] ADASiam uses a twin network architecture consisting of a symmetric encoder and prediction head. Its core innovation lies in the collaborative design of a dynamic hybrid attention mechanism and an adaptive hierarchical contrast loss. The network processing flow can be formally expressed as:

[0097]

[0098] where f θ is the encoder containing dynamic attention, h φ is the prediction head, is the data enhancement operation, and D is the adaptive contrast loss function.

[0099] It is worth mentioning that Figure 3 The overall structure diagram of its network is as follows: Figure 4 The backbone network structure diagram shows the model taking dual-view HRRP (High Resolution Range Profile) data as input and performing feature extraction via the HRRP_Backbone module. The HRRP_Backbone module consists of multiple convolutional layers (Conv1D), batch normalization (BN), LeakyReLU activation functions, and a max pooling layer (MaxPool1D) to extract features from HRRP data. During feature extraction, the Dynamic Hybrid Channel Attention (DHCA) module adaptively weights features at different scales, enhancing the model's focus on key features.

[0100] S2.1 Dynamic Hybrid Attention Network Architecture:

[0101] S2.1.1 Multi-scale feature extraction: The input signal is fused with multi-scale features through three-level convolutional coding:

[0102] Shallow feature extraction:

[0103]

[0104] A 7×1 convolution kernel is selected to match the typical scattering peak width (3-5 distance units) to enhance the local peak capture capability. The input length L=984 is compressed to 492 by maximum pooling (stride=2) to retain the key waveform structure.

[0105] Mid-level attention:

[0106]

[0107] A 5×1 convolution kernel is selected to capture the correlation of mid-range scattering points. The dynamic attention module adaptively enhances key channels. Then, a pooling operation is performed: the temporal dimension is compressed twice to 246 to reduce computational complexity.

[0108] Fine-grained modeling:

[0109]

[0110] A 3×1 convolution kernel is selected to extract the microscopic differences between adjacent scattering points, and the final output length is unified to 64 to eliminate the target orientation sensitivity.

[0111] S2.1.2 Statistical feature generation: Comprehensively characterize channel characteristics through multi-scale statistics to provide a basis for dynamic weight fusion.

[0112]

[0113] S2.1.3 Dynamic Weight Fusion: Adaptively adjust the contribution weights of multiple statistics based on input characteristics.

[0114] Feature splicing and nonlinear mapping:

[0115] w=Softmax(W2·GELU(W1[μ; m; σ]))

[0116] S2.1.4 Channel recalibration: Enhance discriminative features and suppress redundant responses through attention weights.

[0117] Attention Generation:

[0118] s c =Sigmoid(W4·GELU(W3(w1μ c +w2m c +w3σ c )))

[0119] Residual Connection:

[0120]

[0121] S2.2 Adaptive Contrast Loss:

[0122] S2.1.1 Dynamic temperature regulation:

[0123]

[0124] Where τ0 is the reference temperature, ρ i and z i denote the output features of the projection head and momentum encoder respectively, and Var is the variance of feature similarity within the batch.

[0125] S2.1.2 Local consistency constraint: The eigenvector p i , Divide into K = 4 segments (each segment d = 256 dimensions), and calculate the segment cosine similarity:

[0126]

[0127] where p (k) =p[(k-1)d:kd], z (k) Similarly, this can force the smooth transition of adjacent distance unit features and preserve the temporal dependency of the HRRP signal.

[0128] S2.1.2 Feature redundancy penalty: Calculate the feature covariance matrix and penalize the off-diagonal elements:

[0129]

[0130] Where d = 1024 is the feature dimension and I is the identity matrix. This reduces the redundancy between feature dimensions, reduces the strength of the off-diagonal elements of the covariance matrix, improves the utilization of the feature space, and enhances the discriminative ability of the classifier.

[0131] S2.1.3 Total loss function:

[0132]

[0133] Among them, λ1,λ2 are balance coefficients, is the contrast loss:

[0134]

[0135] S2.3 Projection and prediction head design: The projection head transforms the dimension from 16384 layers to 1024 layers, enhancing the nonlinearity of features, and the prediction head prevents the bilinear collapse of gradients.

[0136] S3 KNN evaluation module.

[0137] S3.1 Feature Library Construction: Memory Dataset Through the encoder f θ Extract L2 normalized features:

[0138]

[0139] Concatenate features to construct a feature library matrix: and label vector

[0140] S3.2 Weighted K-nearest neighbor classification: for test samples Calculate the cosine similarity between it and the feature library:

[0141]

[0142] Select the Top-K similar sample index: Corresponding tags:

[0143] Its similarity weight is:

[0144]

[0145] Where τ = 0.1 is the temperature parameter, which softens the weight distribution.

[0146] Then calculate the score for each category separately:

[0147]

[0148] The final predicted recognition result is the category with the highest score:

[0149]

[0150] It is worth noting that an effective recognition method for fusing HRRP data generated by different polarization modes is proposed. The proposed fusion method is based on the data level, which can alleviate the problem that the final result of radar target recognition based on feature level fusion depends largely on the quality of feature selection; the introduction of meta-learning method into radar small sample recognition can alleviate the problems of low radar target recognition accuracy and poor generalization ability under small sample conditions to a certain extent. Under small sample conditions, fusing HRRP data generated by different polarization modes for target recognition has important engineering application value.

[0151] In one embodiment, Figure 5 As shown in Figure 2, the recognition accuracy change curves obtained by removing the dynamic hybrid channel attention mechanism module and the adaptive contrast loss module respectively; the method of the present invention is trained for recognition of 12 targets, using 5000 training samples and 1982 test samples for each type of target, and analyzing Figure 5As can be seen, the full ADASiam model (solid line) achieves the highest accuracy during training, with a relatively stable increase in accuracy, reaching a peak accuracy of approximately 95%. This demonstrates that the adaptive contrastive loss and the dynamic mixed channel attention mechanism significantly improve model performance. The model with the adaptive contrastive loss ablated (dotted line) also shows an increase in accuracy, but the overall trend is more gradual, and the final accuracy is lower than that of the full model. This indicates that the adaptive contrastive loss is crucial for improving the model's discriminative ability. By dynamically adjusting the strength of contrastive learning, it helps the model better learn the inherent structure of the data. The accuracy curve of the model with the dynamic mixed channel attention ablated (dashed line) also shows an upward trend, but with greater fluctuations, and the final accuracy is also lower than that of the full model. This demonstrates that the dynamic mixed channel attention mechanism plays a key role in feature extraction and representation learning, helping the model more effectively capture and utilize important features in the data. Ablation experiments demonstrate that the adaptive contrastive loss and the dynamic mixed channel attention mechanism are two key components of the ADASiam model, and their combined effect significantly improves model performance. The complete model showed higher accuracy and more stable training process during training, proving the effectiveness and necessity of these modules.

[0152] It should be understood that although Figure 1-2 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1-2 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0153] In one embodiment, Figure 6 As shown, an unsupervised HRRP recognition device based on a contrastive learning fusion model is provided, comprising: a model construction module 602, a multi-scale feature acquisition module 604, a residual output module 606, an adaptive loss module 608 and a target recognition module 610, wherein:

[0154] The model construction module 602 is used to construct an adaptive dynamic attention model based on a contrastive learning framework. The adaptive dynamic attention model includes a dynamic hybrid attention network and an adaptive contrastive loss module.

[0155] The multi-scale feature acquisition module 604 is used to input the HRRP signal samples into the dynamic hybrid attention network, and perform shallow feature extraction, mid-level attention processing and fine-grained modeling in the dynamic hybrid attention network to output multi-scale features.

[0156] The residual output module 606 is used to design multi-scale statistics based on multi-scale features to adaptively adjust the attention weights, thereby outputting the residual of the dynamic hybrid attention network.

[0157] The adaptive loss module 608 is used to input the residual into the adaptive contrast loss module for loss calculation, evaluate the calculated loss through a pre-set evaluation module, and obtain the predicted category of the adaptive dynamic attention model, thereby training the adaptive dynamic attention model.

[0158] The target recognition module 610 is used to use the trained adaptive dynamic attention model to recognize the HRRP signal to obtain a target image.

[0159] Regarding the specific limitations of the HRRP unsupervised recognition device based on the contrastive learning fusion model, please refer to the limitations of the HRRP unsupervised recognition method based on the contrastive learning fusion model above, which will not be repeated here. Each module in the above-mentioned HRRP unsupervised recognition device based on the contrastive learning fusion model can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0160] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, an HRRP unsupervised recognition method based on a contrastive learning fusion model is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.

[0161] Those skilled in the art will understand that Figure 6-7 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention and does not constitute a limitation on the computer device to which the solution of the present invention is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0162] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0163] Based on the contrastive learning framework, an adaptive dynamic attention model is constructed. The adaptive dynamic attention model includes a dynamic hybrid attention network and an adaptive contrastive loss module.

[0164] The HRRP signal samples are input into the dynamic hybrid attention network, where shallow feature extraction, mid-level attention processing, and fine-grained modeling are performed to output multi-scale features.

[0165] Multi-scale statistics are designed based on multi-scale features to adaptively adjust the attention weights to output the residual of the dynamic hybrid attention network.

[0166] The residual is input into the adaptive contrast loss module for loss calculation, and the calculated loss is evaluated by the pre-set evaluation module to obtain the predicted category of the adaptive dynamic attention model, so as to train the adaptive dynamic attention model.

[0167] The trained adaptive dynamic attention model is used to identify the HRRP signal to obtain the target image.

[0168] Those skilled in the art will appreciate that all or part of the processes in the above-described embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described embodiments. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAM bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0169] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0170] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. An unsupervised HRRP recognition method based on contrastive learning fusion model, characterized by: The method comprises: Based on the contrastive learning framework, an adaptive dynamic attention model is constructed; the adaptive dynamic attention model includes: a dynamic hybrid attention network and an adaptive contrast loss module; Inputting the HRRP signal sample into the dynamic hybrid attention network, performing shallow feature extraction, mid-level attention processing, and fine-grained modeling to output multi-scale features in the dynamic hybrid attention network; Designing multi-scale statistics based on the multi-scale features to adaptively adjust the attention weights, thereby outputting the residual of the dynamic hybrid attention network; Inputting the residual into the adaptive contrast loss module to perform loss calculation, evaluating the calculated loss through a pre-set evaluation module to obtain the predicted category of the adaptive dynamic attention model, thereby training the adaptive dynamic attention model; The trained adaptive dynamic attention model is used to identify the HRRP signal to obtain the target image.

2. The method according to claim 1, characterized in that The adaptive dynamic attention model is: z1=f θ (Τ1(x)) z2=f θ (Τ2(x)) L=1 / 2[D(p1,z2)+D(p1,z2)] Where x is the HRRP signal, z1 and z2 are the HRRP signal samples used for contrastive learning, p1 and p2 are the output features of the projection head, and f θ is an encoder containing dynamic attention, Τ1 and Τ2 are data enhancement operations, is the prediction head, D is the adaptive contrast loss function, and L is the loss function of the contrastive learning framework.

3. The method according to claim 2, characterized in that The HRRP signal sample is input into the dynamic hybrid attention network, where shallow feature extraction, mid-level attention processing, and fine-grained modeling are performed to output multi-scale features, including: After the HRRP signal is input into the dynamic hybrid attention network, the convolution kernel extracted by the shallow feature is used to match the typical scattering peak width to obtain the first-layer prediction feature: Among them, F1 is the first layer prediction feature, ReLU is the rectified linear unit function, BN is normalization, Conv1D7 is the 7×1 convolution kernel, and x is the HRRP signal; The first layer prediction features are input into the convolution kernel of the middle layer attention processing to capture the correlation of the mid-range scattering points, and the second layer prediction features are obtained: Among them, F2 is the second-layer prediction feature, DynamicAttn is the dynamic attention mechanism function, and Conv1D5 is a 5×1 convolution kernel; The second layer prediction features are input into the convolution kernel of the fine-grained modeling to extract the microscopic differences between adjacent scattering points to obtain multi-scale features: Among them, F3 is the multi-scale feature, AdaptiveAvgPool is the adaptive average pooling function, and Conv1D3 is the 3×1 convolution kernel.

4. The method according to claim 3, characterized in that Designing multi-scale statistics based on the multi-scale features to adaptively adjust the attention weights, thereby outputting the residual of the dynamic hybrid attention network, including: Design multi-scale statistics based on the multi-scale features: Among them, μ c is the average value of the c-th channel over all time steps, T is the time series dimension, is the eigenvalue of the cth channel in the feature map output by the second convolution block, x1 is a constant 1, m c is the energy peak in the cth channel, σ c is the standard deviation; The attention weights are adaptively adjusted according to the multi-scale statistics to output the residual of the dynamic hybrid attention network: in, is the residual of the dynamic hybrid attention network, s c is the attention weight.

5. The method according to claim 4, characterized in that Adaptively adjusting the attention weights according to the multi-scale statistics to output the residual of the dynamic hybrid attention network, including: According to the feature splicing and nonlinear mapping of the multi-scale statistics, the contribution weight of the multi-scale statistics is obtained: in, is the contribution weight, W1 and W2 are weight matrices, Softmax is the activation function, and GELU is the smooth activation function; Recalibrate the channel generation attention weights according to the contribution weights: Among them, s c is the attention weight, are weight coefficients, W3 and W4 are linear transformation matrices, and Sigmoid is the activation function; The attention weights are adaptively adjusted to output the residual of the dynamic hybrid attention network.

6. The method according to claim 5, characterized in that Inputting the residual into the adaptive contrast loss module to perform loss calculation, including: The residual is input into the adaptive contrast loss module to perform dynamic temperature adjustment on the output characteristics of the projection head and the HRRP signal samples to obtain the standard temperature: Among them, τ is the standard temperature, τ0 is the reference temperature, p i is the output feature of the projection head, z i is the HRRP signal sample, Var is the variance of feature similarity within the batch, and B is the batch size; According to the standard temperature, the contrast loss is reconstructed to obtain the contrast loss function: Among them, L contrast is the contrast loss function; The piecewise cosine similarity between the output feature of the projection head and the feature vector composed of the HRRP signal samples is calculated to obtain the similarity loss: Among them, L local is the similarity loss, p (k) is the output feature of the projection head, z (k) is the output feature of the momentum encoder, K is the number of segments; Calculate the feature redundancy penalty of the feature covariance matrix composed of the output features of the projection head to obtain the feature redundancy penalty loss: Among them, L textcov is the feature redundancy penalty loss, d is the feature dimension, B is the batch size, p b is the eigenvector in the batch, I is the identity matrix; The loss function of the adaptive contrast loss module is generated according to the preset balance coefficient, the contrast loss function, the similarity loss, and the feature redundancy penalty loss: L total =L contrast +λ1L local +λ2L textcov Among them, L total is the loss function of the adaptive contrast loss module, and λ1 and λ2 are different balance coefficients.

7. The method according to any one of claims 1 to 6, characterized in that The pre-set evaluation module is used to perform weighted K-nearest neighbor classification based on the feature library constructed based on the memory data set, and then select labeled sample data for weighted classification.

8. The method according to claim 7, characterized in that The calculated loss is evaluated by a preset evaluation module to obtain the prediction category of the adaptive dynamic attention model, including: The calculated loss is evaluated by a pre-set evaluation module, and the score of each category is calculated respectively. The category corresponding to the maximum value of the score is taken as the predicted category of the adaptive dynamic attention model: in, is the maximum score, Score(c) is the loss evaluation score, Ι is the identity matrix, y i is the predicted category, c is the category, is the weight coefficient.

9. An unsupervised HRRP recognition device based on contrastive learning fusion model, characterized in that: The device comprises: A model building module is used to build an adaptive dynamic attention model based on a contrastive learning framework; the adaptive dynamic attention model includes: a dynamic hybrid attention network and an adaptive contrastive loss module; A multi-scale feature acquisition module is used to input HRRP signal samples into the dynamic hybrid attention network, and perform shallow feature extraction, mid-level attention processing, and fine-grained modeling to output multi-scale features in the dynamic hybrid attention network; A residual output module, configured to adaptively adjust the attention weights based on the multi-scale features and design multi-scale statistics to output the residual of the dynamic hybrid attention network; An adaptive loss module, configured to input the residual into the adaptive contrast loss module for loss calculation, evaluate the calculated loss through a pre-set evaluation module, obtain a predicted category of the adaptive dynamic attention model, and thereby train the adaptive dynamic attention model; The target recognition module is used to identify the HRRP signal to be identified using the trained adaptive dynamic attention model to obtain the target image.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images