Hrrp signal feature recognition method, device and equipment based on time-frequency domain fusion

Abstract

This application relates to a method, apparatus, and device for HRRP signal feature recognition based on time-frequency domain fusion. A sample set is constructed using HRRP signal data obtained from multiple radar stations detecting the same target. After preprocessing, this sample set is used to train an HRRP signal feature recognition network. This network extracts time-domain, frequency-domain, and time-frequency characteristic primitives from the data through parallel time-domain segmentation coding units, global frequency-domain embedding units, and time-frequency analysis units. These primitives are then fused to obtain single-station fused features for target recognition. During the training phase, the single-station fused features from multiple radar stations, target prediction results, and ground truth labels are combined to calculate cross-station contrastive learning loss and cross-target category contrastive loss. The network's learnable parameters are adjusted until convergence. Finally, the trained network is used to recognize targets from HRRP signals, effectively improving target recognition accuracy.

Description

Technical Field

[0001] This application relates to the field of radar signal processing and pattern recognition technology, and in particular to a method, apparatus and device for HRRP signal feature recognition based on time-frequency domain fusion. Background Technology

[0002] High-resolution range profiles (HRRPs) are target radar characteristic signals acquired through broadband radar systems. Essentially, they are the coherent superposition of the projections of the target's scattering centers along the radar's line-of-sight direction. While maintaining ease of acquisition and efficient storage, these characteristic signals effectively characterize the target's scattering point distribution, geometric structure, and other physical attributes, thus possessing significant application value in the field of radar automatic target recognition (RATR).

[0003] With the development of deep learning technology, data-driven HRRP recognition methods have made significant progress. However, existing methods generally share a common limitation: these strategies primarily treat the HRRP signal as a whole for implicit feature learning, failing to fully incorporate key prior information about the target and environment. This approach leads to two prominent problems: firstly, the models often ignore the differential contributions of local features in different regions of the HRRP, failing to effectively capture local discriminative segments that play a crucial role in recognition; secondly, existing methods underestimate the impact of global discriminative features on the final recognition result, limiting the model's representational capabilities.

[0004] Specifically, current HRRP identification methods have significant shortcomings in feature utilization: First, the localized discriminative segments contained in the time-domain waveform can directly reflect the specific structural features of the target, but existing methods lack explicit modeling of these local features; second, the frequency domain spectrum provides a global feature signature of the target's electromagnetic scattering characteristics, which has better robustness to local perturbations, but existing methods do not fully utilize this; third, time-frequency representation can simultaneously capture the dynamic characteristics of the spectrum and its temporal correspondence, revealing transient phenomena and providing a complementary perspective that pure time-domain or frequency-domain analysis cannot fully provide, but this important characteristic has not been fully explored in current methods. Summary of the Invention

[0005] Therefore, it is necessary to provide a time-frequency domain fusion-based HRRP signal feature recognition method, apparatus, and device that can effectively improve the accuracy and robustness of target recognition in addressing the aforementioned technical problems.

[0006] A method for HRRP signal feature recognition based on time-frequency domain fusion, the method comprising:

[0007] Acquire an HRRP signal sample set, which includes HRRP signal data obtained from multiple radar stations detecting the same target;

[0008] Time-frequency alignment and motion compensation were performed on the HRRP signal data obtained from each radar site to obtain preprocessed data.

[0009] The preprocessed data corresponding to each radar station is input into the corresponding HRRP signal feature recognition network. In the HRRP signal feature recognition network, the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives of the preprocessed data are extracted by parallel time-domain segmentation coding unit, global frequency-domain embedding unit, and time-frequency analysis unit, respectively. The recognition unit then fuses the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives to obtain single-station fused features. Target recognition is then performed based on the single-station fused features to obtain target prediction results.

[0010] By combining the single-station fusion features of multiple radar stations, target prediction results, and corresponding ground truth labels, cross-station contrastive learning loss and cross-target category contrastive loss are calculated. The learnable parameters in each of the HRRP signal feature recognition networks are adjusted until convergence, thus obtaining the trained HRRP signal feature recognition network.

[0011] The HRRP signal received from a certain station is acquired, and the target recognition result is obtained by using the trained HRRP signal feature recognition network to perform target recognition on the HRRP signal.

[0012] In one embodiment, in the time-domain segmented coding unit:

[0013] The preprocessed data is randomly sampled and divided to obtain multiple HRRP time domain segments;

[0014] A scattering feature weighting calculation network is used to weight each HRRP time domain segment by calculating the importance of the target scattering characteristics in each HRRP time domain segment;

[0015] The weighted HRRP time-domain segment is input into the time-domain coding network to obtain the time-domain feature code, i.e., the time-domain feature primitive.

[0016] In one embodiment, in the global frequency domain embedding unit:

[0017] Perform a discrete Fourier transform on the preprocessed data to obtain the corresponding spectral signal;

[0018] An attention mask matrix is ​​calculated based on the amplitude spectrum of the spectrum signal. The amplitude spectrum and phase spectrum of the spectrum signal are then masked using the attention mask matrix to obtain the masked amplitude spectrum and phase spectrum.

[0019] The amplitude spectrum and phase spectrum after the masking process are encoded to obtain amplitude encoding vector and phase encoding vector, and the amplitude encoding vector and phase encoding vector are used as the frequency domain characteristic primitives.

[0020] In one embodiment, in the time-frequency analysis unit:

[0021] Perform a short-time Fourier transform on the preprocessed data to obtain the corresponding time-frequency diagram;

[0022] The time-frequency graph is encoded using a convolutional neural network to obtain a time-frequency graph encoding vector, which is then used as the time-frequency characteristic primitive.

[0023] In one embodiment, in the identification unit, the single-station fusion features are extracted based on the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives using a feature extraction network with a Transformer architecture.

[0024] In one embodiment, the cross-site contrastive learning loss is expressed as:

[0025] ;

[0026] In the above formula, Indicates the number of radar stations. Indicates the extracted first... Individual station integration features Representation and Features Feature sets from different radar stations in the same sequence Indicates the inner product. Indicates the temperature coefficient. Represents the set consisting of all features. Then it means except for the feature The set consisting of all features other than Then they respectively represent from the set With sets Its characteristics.

[0027] In one embodiment, the cross-target class contrast loss is expressed as:

[0028] ;

[0029] In the above formula, Representation and Features A set of features from the same category of targets. Then it represents the cardinality of the set. Indicates that it comes from a set Its characteristics.

[0030] This application also provides an HRRP signal feature recognition device based on time-frequency domain fusion, the device comprising:

[0031] The HRRP signal sample acquisition module is used to acquire an HRRP signal sample set, which includes HRRP signal data obtained by multiple radar stations detecting the same target.

[0032] The data preprocessing module is used to perform time-frequency alignment and motion compensation on the HRRP signal data obtained from each radar site to obtain preprocessed data.

[0033] The time-frequency domain feature extraction module is used to input the preprocessed data corresponding to each radar station into the corresponding HRRP signal feature recognition network. In the HRRP signal feature recognition network, the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives of the preprocessed data are extracted by parallel time-domain segmentation coding unit, global frequency-domain embedding unit, and time-frequency analysis unit, respectively. The recognition unit then fuses the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives to obtain single-station fused features. Target recognition is then performed based on the single-station fused features to obtain target prediction results.

[0034] The multi-site joint network training module is used to combine the single-site fusion features of multiple radar sites, target prediction results and corresponding ground truth labels to calculate cross-site contrastive learning loss and cross-target category contrastive loss, and adjust the learnable parameters in each HRRP signal feature recognition network until convergence, thus obtaining the trained HRRP signal feature recognition network.

[0035] The target recognition module is used to acquire the HRRP signal received from a certain station, and use a trained HRRP signal feature recognition network to perform target recognition on the HRRP signal to obtain the target recognition result.

[0036] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the specific steps in the above-described HRRP signal feature recognition method based on time-frequency domain fusion.

[0037] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the specific steps of the above-described HRRP signal feature recognition method based on time-frequency domain fusion.

[0038] The aforementioned HRRP signal feature recognition method, apparatus, and device based on time-frequency domain fusion acquires an HRRP signal sample set, constructed from HRRP signal data obtained from multiple radar stations detecting the same target. After preprocessing, this sample set is input into a corresponding HRRP signal feature recognition network. Within this network, parallel time-domain segmentation coding units, global frequency-domain embedding units, and time-frequency analysis units extract time-domain characteristic primitives, frequency-domain characteristic primitives, and time-frequency characteristic primitives from the preprocessed data, respectively. A recognition unit then fuses these primitives to obtain single-station fused features. Target recognition is then performed based on these single-station fused features to obtain target prediction results. The cross-station contrastive learning loss and cross-target category contrastive loss are calculated by combining the single-station fused features from multiple radar stations, the target prediction results, and the corresponding ground truth labels. The learnable parameters in each HRRP signal feature recognition network are adjusted until convergence, resulting in a trained HRRP signal feature recognition network. The trained HRRP signal feature recognition network is then used to perform target recognition on HRRP signals to obtain target recognition results. This method can effectively improve the accuracy of target recognition. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating a time-frequency domain fusion-based HRRP signal feature recognition method in one embodiment.

[0040] Figure 2 This is a schematic diagram of the training process of the HRRP signal feature recognition network in one embodiment;

[0041] Figure 3 This is a structural block diagram of an HRRP signal feature recognition device based on time-frequency domain fusion in one embodiment;

[0042] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0044] Current HRRP identification methods have significant shortcomings in feature utilization. First, the localized discriminative segments contained in the time-domain waveform can directly reflect the specific structural features of the target, but existing methods lack explicit modeling of these local features. Second, the frequency domain spectrum provides a global signature of the target's electromagnetic scattering characteristics, exhibiting better robustness to local perturbations, but current methods do not fully utilize this. Third, time-frequency representation can simultaneously capture the dynamic characteristics of the spectrum and their temporal correspondence, revealing transient phenomena and providing a complementary perspective that pure time-domain or frequency-domain analysis cannot fully offer, but this important characteristic has not been fully explored in current methods.

[0045] In response to the above problems, such as Figure 1 As shown, a method for HRRP signal feature recognition based on time-frequency domain fusion is provided, which specifically includes the following steps:

[0046] Step S100: Obtain the HRRP signal sample set, which includes HRRP signal data obtained by multiple radar stations detecting the same target.

[0047] Step S110: Time-frequency alignment and motion compensation are performed on the HRRP signal data obtained from each radar station to obtain preprocessed data.

[0048] Step S120: The preprocessed data corresponding to each radar station is input into the corresponding HRRP signal feature recognition network. In the HRRP signal feature recognition network, the time-domain characteristic primitives, frequency-domain characteristic primitives, and time-frequency characteristic primitives of the preprocessed data are extracted by the parallel time-domain segmented coding unit, global frequency-domain embedding unit, and time-frequency analysis unit, respectively. The recognition unit is used to fuse the time-domain characteristic primitives, frequency-domain characteristic primitives, and time-frequency characteristic primitives to obtain single-station fused features. Then, target recognition is performed based on the single-station fused features to obtain the target prediction result.

[0049] Step S130: Combine the single-station fusion features of multiple radar stations, target prediction results and corresponding ground truth labels to calculate cross-station contrastive learning loss and cross-target category contrastive loss. Adjust the learnable parameters in each HRRP signal feature recognition network until convergence, and then obtain the trained HRRP signal feature recognition network.

[0050] Step S140: Obtain the HRRP signal received from a certain station, and use the trained HRRP signal feature recognition network to perform target recognition on the HRRP signal to obtain the target recognition result.

[0051] In this method, unlike treating HRRP as a single entity, we obtain a more powerful and robust feature representation by decoupling and explicitly modeling its different information dimensions. Specifically, we propose a novel fusion framework that can synergistically utilize temporal local features, frequency domain global features, and complementary temporal and frequency domain features to overcome the performance bottleneck of existing HRRP recognition methods.

[0052] Specifically, this method proposes a feature mining and recognition approach for multi-station HRRP signals based on time-frequency domain fusion. It achieves comprehensive characterization of multi-station HRRP signals by collaboratively mining time-domain, frequency-domain, and time-frequency-domain features. The framework comprises three core modules: a Time-Domain Segmented Coding Unit (TSE), a Global Frequency Embedding Unit (GFE), and a Time-Frequency Analysis Unit (TFA). The TSE extracts local discriminative features, dynamically discovers and weights key time segments in the HRRP signal, focusing the model's attention on the most significant target features. The GFE calculates the global spectrum of the HRRP and adaptively weights the most discriminative frequency components using an energy-guided attention mechanism, effectively capturing the inherent electromagnetic scattering characteristics of the target. The TFA generates a time-frequency representation of the HRRP through short-time Fourier transform and learns to extract joint time-frequency features crucial for distinguishing complex targets. Furthermore, based on the aforementioned triple feature extraction, a multi-contrast learning fusion strategy (MCL) is designed. This strategy enhances the feature consistency of the same target across different radar stations while ensuring high separability of the fused representation across different target categories, thereby achieving effective synergy of complementary information. Specifically, this strategy constructs a dual optimization objective of cross-station consistency and cross-class discriminability, which not only solves the problem of insufficient modeling of cross-station shared features in traditional methods but also significantly improves the discriminative ability of the fused representation.

[0053] In step S100, to build the basic data support for the subsequent training of the HRRP signal feature recognition network, it is necessary to first obtain an HRRP signal sample set. The core of this sample set consists of HRRP signal data collected by multiple radar stations deployed at different spatial locations (such as different longitudes, latitudes, or altitudes) after synchronous or temporal detection of the same target (including the target's morphology under different attitudes, motion states, or environmental interference).

[0054] Specifically, radar stations can employ different operating parameters (such as transmission frequency band, pulse repetition frequency, polarization mode, etc.) to cover the scattering characteristics of targets under diverse observation angles and electromagnetic environments. Furthermore, the detection of the same target needs to include echo signals from multiple scenarios, including static and dynamic conditions, to ensure that the HRRP data in the sample set reflects both the distribution characteristics of scattering points along the radar line of sight and the signal variations caused by changes in target attitude and differences in observations from multiple stations. This provides comprehensive data support for subsequent model learning of robustness characteristics.

[0055] In step S110, the HRRP data collected by each radar station are time-frequency aligned and motion compensated to eliminate phase distortion caused by target micro-motion and solve the position sensitivity and amplitude fluctuation problems in the HRRP sequence.

[0056] Next, feature extraction is performed on the preprocessed data from each radar site. This method employs a three-tiered parallel information flow processing architecture: The time-domain segmentation coding stage uses a network-guided Markov chain Monte Carlo adaptive segmentation mechanism to dynamically discover key time segments in the HRRP signal. Attention-weighted generation of position-sensitive time-domain characteristic primitives effectively captures the discriminative features of the target's local scattering structure. Specifically, this includes three parallel feature extraction stages: a global frequency domain embedding stage, which calculates the global HRRP spectrum, constructs an energy-guided attention mechanism, adaptively selects and emphasizes the most discriminative frequency components, and generates frequency-domain characteristic primitives representing the target's electromagnetic scattering characteristics. The time-frequency analysis stage uses short-time Fourier transform to generate the time-frequency representation of HRRP, extracts joint time-frequency features, captures transient phenomena and local spectral dynamics, and forms complementary time-frequency characteristic primitives.

[0057] Meanwhile, in the feature fusion and optimization stage, a multi-contrast learning mechanism is designed: by constructing a dual optimization objective of cross-station consistency and cross-class discriminability, the feature alignment of the same target across different radar stations is strengthened, while ensuring the feature separability between different target categories. This strategy effectively solves the problem of insufficient cross-station feature sharing modeling and significantly improves the robustness and accuracy of recognition in complex scenarios such as missing data and noise interference.

[0058] A network architecture for HRRP signal feature recognition was designed within the framework of the three-stage feature extraction, feature fusion and optimization of multiple radar sites.

[0059] In this embodiment, in the time-domain segmented coding unit: the preprocessed data is randomly sampled and divided to obtain multiple HRRP time-domain segments. The scattering feature weight calculation network is used to calculate the importance of the target scattering characteristics in each HRRP time-domain segment, and the weighted HRRP time-domain segments are weighted. The weighted HRRP time-domain segments are input into the time-domain coding network to obtain the time-domain feature code, i.e., the time-domain feature primitive.

[0060] Specifically, the length of the radar station is set as follows: The original HRRP signal is ,in, Indicates the first The wavelet is obtained from each range cell. After the basic preprocessing of the radar signal, i.e., step S110, the resulting HRRP time series can be represented as follows: Based on the preprocessed HRRP signal, the time-domain segmented coding unit can be composed of the following steps: random sampling division, scattering characteristic weight calculation, and time-domain characteristic coding.

[0061] Furthermore, the temporal sampling partition of HRRP can be expressed as:

[0062] (1)

[0063] In formula (1), For the first The sub-segments obtained by random partitioning, the starting position of the segment. From uniform distribution The HRRP time-domain segments are obtained through sampling. These segments are then input into the scattering feature weight calculation network to calculate the importance of the target scattering characteristics contained in each segment and to weight each time-domain segment. This process can be represented by the following mathematical process:

[0064] (2)

[0065] (3)

[0066] In formulas (2) and (3), and These are the activation functions, and These are the learnable weights and biases, respectively. Weight The time domain segment obtained after reweighting is In this step, time-domain segments containing more target scattering characteristics will receive greater weight, and the time-varying information in the segments will become more significant after reweighting; conversely, time-domain segments containing less target scattering characteristics will have the opposite effect.

[0067] Furthermore, the weighted time-domain segment obtained will be input into the time-domain coding network to obtain the scattering characteristic primitives of the HRRP sequence in the time domain, which can be specifically represented as:

[0068] (4)

[0069] In formula (4), , For learnable weights, This is a learnable bias. That is the activation function. The time-domain feature encoding obtained after encoding is called the time-domain feature primitive.

[0070] In this embodiment, in the global frequency domain embedding unit: the preprocessed data is subjected to discrete Fourier transform to obtain the corresponding spectral signal; the attention mask matrix is ​​calculated based on the amplitude spectrum of the spectral signal; the amplitude spectrum and phase spectrum of the spectral signal are masked using the attention mask matrix to obtain the masked amplitude spectrum and phase spectrum; the masked amplitude spectrum and phase spectrum are encoded to obtain the amplitude encoding vector and phase encoding vector; and the amplitude encoding vector and phase encoding vector are used as frequency domain characteristic primitives.

[0071] Specifically, the preprocessed HRRP sequence is input into a global frequency domain embedding unit for processing. This unit includes frequency domain transformation, attention mask calculation, and frequency domain feature encoding. Specifically, after discrete Fourier transform, the spectral signal is obtained. The calculation process for its attention mask matrix is ​​as follows:

[0072] (5)

[0073] In formula (5), Indicates amplitude spectrum, For the mask matrix, and These are the learnable weights and biases, respectively. The masking process can be represented as:

[0074] (6)

[0075] (7)

[0076] In formulas (6) and (7), The phase spectrum is represented. Finally, the amplitude and phase spectra after masking are encoded:

[0077] (8)

[0078] (9)

[0079] In formulas (8) and (9), These are the learnable weights for amplitude and phase, respectively. For learnable bias, and The amplitude and phase encoded vectors, respectively.

[0080] In this embodiment, in the time-frequency analysis unit: the preprocessed data is subjected to short-time Fourier transform to obtain the corresponding time-frequency graph, the time-frequency graph is encoded using a convolutional neural network to obtain the time-frequency graph encoding vector, and the time-frequency graph encoding vector is used as the time-frequency characteristic primitive.

[0081] Specifically, the input HRRP signal will undergo a short-time Fourier transform to obtain a time-frequency diagram, as follows:

[0082] (10)

[0083] In formula (10), For the obtained time-frequency diagram, These represent the frame and frequency index, respectively. Next, a convolutional neural network is used to encode the time-frequency graph; this process can be represented as:

[0084] (11)

[0085] In formula (11), This is the time-frequency graph encoding vector. For batch normalization processing, This is a two-dimensional convolutional network.

[0086] In this embodiment, in the recognition unit, a feature extraction network based on the Transformer architecture extracts single-station fusion features based on time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives. The process is represented as follows:

[0087] (12)

[0088] Furthermore, the recognition unit performs identification based on the fusion features of single-station radar stations to obtain the recognition results. During the training process, the recognition results obtained from multiple radar stations are combined to calculate cross-station contrastive learning loss and cross-target category contrastive loss representation.

[0089] Specifically, the cross-site comparison learning loss is expressed as:

[0090] (13)

[0091] In formula (13), Indicates the number of radar stations. Indicates the extracted first... Individual station integration features Representation and Features Feature sets from different radar stations in the same sequence Indicates the inner product. Indicates the temperature coefficient. Represents the set consisting of all features. Then it means except for the feature The set consisting of all features other than , Then they respectively represent from the set With sets Its characteristics.

[0092] In one embodiment, the cross-target class contrast loss is expressed as:

[0093] ;

[0094] In the above formula, Representation and Features A set of features from the same category of targets. Then it represents the cardinality of the set. Indicates that it comes from a set Its characteristics.

[0095] Furthermore, the final multiple contrastive loss function can be expressed as:

[0096] (15)

[0097] In this embodiment, the total loss function of the HRRP signal feature recognition network is expressed as:

[0098] (16)

[0099] In formula (16), This represents the classification loss based on cross-entropy. This represents the corresponding weight coefficient. Based on the above loss function, the update of the learnable parameters in the recognition model can be expressed as:

[0100] (17)

[0101] In formula (17), the parameter , For learning rate, For optimizers.

[0102] like Figure 2 The diagram shown above illustrates the training process.

[0103] In step S140, the trained HRRP signal feature recognition network can be implemented at a single radar site for target recognition, or it can be implemented at multiple radars in a multi-radar site system to identify the same target.

[0104] This paper also demonstrates the effectiveness of the proposed method through experimental results. Tables 1 to 3 below show the performance results of the proposed technique, with the scenario being the identification of twelve types of aircraft targets. Table 1 presents a performance comparison between the proposed technique and existing advanced radar HRRP target identification methods, showing that the proposed technique significantly outperforms existing methods. Table 2 compares the identification performance in noisy scenarios, with the results also indicating that the proposed method has the most stable identification capability. Table 3 compares the computational resource consumption of each method, showing that the proposed method achieves the best identification performance under the same computational consumption level. In Tables 1 and 2, the F1 score represents the harmonic mean of precision and recall.

[0105] Table 1 Comparison of recognition performance of various methods

[0106]

[0107] Table 2 Comparison of recognition performance of various methods in noisy scenes

[0108]

[0109] Table 3 Comparison of resource consumption for each method

[0110]

[0111] Table 4 shows the full English abbreviations of the methods used for comparison in all experiments, as listed in Tables 1 to 3 above:

[0112] Table 4. Full English and Chinese names of each method

[0113]

[0114] The aforementioned HRRP signal feature recognition method based on time-frequency domain fusion innovatively achieves unified modeling of local time-domain features, global frequency-domain features, and complementary time-frequency features through a triple information flow collaborative mining mechanism and a multi-contrast learning fusion strategy. This method effectively solves the core problems of existing methods, such as the lack of time-frequency physical consistency and cross-site feature alignment failure. It significantly improves feature discrimination capabilities in complex environments with noise interference and missing data, achieving breakthrough progress in recognition accuracy under challenging scenarios such as low signal-to-noise ratio and high missing rate.

[0115] This method can extract and learn the target characteristics of HRRP signals in the time and frequency domains. It can obtain separable features with good characterization for radar HRRP signals. The separable features obtained by this method can effectively improve the performance of classification and recognition algorithms and maintain good recognition performance under conditions of missing data and strong interference such as noise.

[0116] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed 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 some of the sub-steps or stages of other steps.

[0117] In one embodiment, such as Figure 3 As shown, a time-frequency domain fusion-based HRRP signal feature recognition device is provided, comprising: an HRRP signal sample acquisition module 200, a data preprocessing module 210, a time-frequency domain feature extraction module 220, a multi-site joint network training module 230, and a target recognition module 240, wherein:

[0118] The HRRP signal sample acquisition module 200 is used to acquire an HRRP signal sample set, which includes HRRP signal data obtained by multiple radar stations detecting the same target.

[0119] The data preprocessing module 210 is used to perform time-frequency alignment and motion compensation on the HRRP signal data obtained from each radar station to obtain preprocessed data.

[0120] The time-frequency domain feature extraction module 220 is used to input the preprocessed data corresponding to each radar station into the corresponding HRRP signal feature recognition network. In the HRRP signal feature recognition network, the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives of the preprocessed data are extracted by parallel time-domain segmentation coding unit, global frequency-domain embedding unit, and time-frequency analysis unit, respectively. The recognition unit is used to fuse the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives to obtain single-station fused features. Then, target recognition is performed based on the single-station fused features to obtain target prediction results.

[0121] The multi-site joint network training module 230 is used to combine the single-site fusion features of multiple radar sites, target prediction results and corresponding ground truth labels to calculate cross-site contrast learning loss and cross-target category contrast loss, and adjust the learnable parameters in each HRRP signal feature recognition network until convergence, thus obtaining the trained HRRP signal feature recognition network.

[0122] The target recognition module 240 is used to acquire the HRRP signal received by a certain station, and use the trained HRRP signal feature recognition network to perform target recognition on the HRRP signal to obtain the target recognition result.

[0123] Specific limitations regarding the HRRP signal feature recognition device based on time-frequency domain fusion can be found in the limitations of the HRRP signal feature recognition method based on time-frequency domain fusion described above, and will not be repeated here. Each module in the aforementioned HRRP signal feature recognition device based on time-frequency domain fusion can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0124] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a time-frequency domain fusion-based HRRP signal feature recognition method. The display screen can be an LCD screen or an e-ink display screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0125] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0126] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0127] Acquire an HRRP signal sample set, which includes HRRP signal data obtained from multiple radar stations detecting the same target;

[0128] Time-frequency alignment and motion compensation were performed on the HRRP signal data obtained from each radar site to obtain preprocessed data.

[0129] The preprocessed data corresponding to each radar station is input into the corresponding HRRP signal feature recognition network. In the HRRP signal feature recognition network, the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives of the preprocessed data are extracted by parallel time-domain segmentation coding unit, global frequency-domain embedding unit, and time-frequency analysis unit, respectively. The recognition unit then fuses the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives to obtain single-station fused features. Target recognition is then performed based on the single-station fused features to obtain target prediction results.

[0130] By combining the single-station fusion features of multiple radar stations, target prediction results, and corresponding ground truth labels, cross-station contrastive learning loss and cross-target category contrastive loss are calculated. The learnable parameters in each of the HRRP signal feature recognition networks are adjusted until convergence, thus obtaining the trained HRRP signal feature recognition network.

[0131] The HRRP signal received from a certain station is acquired, and the target recognition result is obtained by using the trained HRRP signal feature recognition network to perform target recognition on the HRRP signal.

[0132] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0133] Acquire an HRRP signal sample set, which includes HRRP signal data obtained from multiple radar stations detecting the same target;

[0134] Time-frequency alignment and motion compensation were performed on the HRRP signal data obtained from each radar site to obtain preprocessed data.

[0135] The preprocessed data corresponding to each radar station is input into the corresponding HRRP signal feature recognition network. In the HRRP signal feature recognition network, the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives of the preprocessed data are extracted by parallel time-domain segmentation coding unit, global frequency-domain embedding unit, and time-frequency analysis unit, respectively. The recognition unit then fuses the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives to obtain single-station fused features. Target recognition is then performed based on the single-station fused features to obtain target prediction results.

[0136] By combining the single-station fusion features of multiple radar stations, target prediction results, and corresponding ground truth labels, cross-station contrastive learning loss and cross-target category contrastive loss are calculated. The learnable parameters in each of the HRRP signal feature recognition networks are adjusted until convergence, thus obtaining the trained HRRP signal feature recognition network.

[0137] The HRRP signal received from a certain station is acquired, and the target recognition result is obtained by using the trained HRRP signal feature recognition network to perform target recognition on the HRRP signal.

[0138] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. 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 embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0139] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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.

[0140] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for HRRP signal feature recognition based on time-frequency domain fusion, characterized in that, The method includes: Acquire an HRRP signal sample set, which includes HRRP signal data obtained from multiple radar stations detecting the same target; Time-frequency alignment and motion compensation were performed on the HRRP signal data obtained from each radar site to obtain preprocessed data. The preprocessed data corresponding to each radar station is input into the corresponding HRRP signal feature recognition network. In the HRRP signal feature recognition network, parallel time-domain segmented coding units, global frequency-domain embedding units, and time-frequency analysis units are used to extract the time-domain characteristic primitives, frequency-domain characteristic primitives, and time-frequency characteristic primitives of the preprocessed data, respectively. The recognition unit then fuses these time-domain characteristic primitives, frequency-domain characteristic primitives, and time-frequency characteristic primitives to obtain single-station fused features. Target recognition is then performed based on these single-station fused features to obtain target prediction results. Specifically, in the time-domain segmented coding unit: the preprocessed data is randomly sampled and divided into multiple HRRP time-domain segments. A scattering feature weight calculation network is used to calculate the scattering feature weights in each of the HRRP time-domain segments. The importance of target scattering characteristics is considered. Each HRRP time-domain segment is weighted, and the weighted HRRP time-domain segment is input into a time-domain coding network to obtain time-domain feature codes, i.e., the time-domain feature primitives. In the global frequency-domain embedding unit: the preprocessed data undergoes a discrete Fourier transform to obtain the corresponding spectral signal. An attention mask matrix is ​​calculated based on the amplitude spectrum of the spectral signal. The attention mask matrix is ​​then used to mask the amplitude spectrum and phase spectrum of the spectral signal to obtain masked amplitude and phase spectra. The masked amplitude and phase spectra are then encoded to obtain amplitude encoding vectors and phase encoding vectors. These amplitude encoding vectors and phase encoding vectors are used as the frequency-domain feature primitives. By combining the single-site fusion features of multiple radar sites, target prediction results, and corresponding ground truth labels, the cross-site contrastive learning loss and cross-target category contrastive loss are calculated. The learnable parameters in each of the HRRP signal feature recognition networks are then adjusted until convergence, resulting in a trained HRRP signal feature recognition network. The cross-site contrastive learning loss is expressed as: In the above formula, Indicates the number of radar stations. Indicates the extracted first... Individual station integration features Representation and Features Feature sets from different radar stations in the same sequence Indicates the inner product. Indicates the temperature coefficient. Represents the set consisting of all features. Then it means except for the feature The set consisting of all features other than , Then they respectively represent from set With sets Features; The HRRP signal received from a certain station is acquired, and the target recognition result is obtained by using the trained HRRP signal feature recognition network to perform target recognition on the HRRP signal.

2. The HRRP signal feature recognition method based on time-frequency domain fusion according to claim 1, characterized in that, In the time-frequency analysis unit: Perform a short-time Fourier transform on the preprocessed data to obtain the corresponding time-frequency diagram; The time-frequency graph is encoded using a convolutional neural network to obtain a time-frequency graph encoding vector, which is then used as the time-frequency characteristic primitive.

3. The HRRP signal feature recognition method based on time-frequency domain fusion according to claim 1, characterized in that, In the recognition unit, the feature extraction network based on the Transformer architecture extracts the single-station fusion features based on the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives.

4. The HRRP signal feature recognition method based on time-frequency domain fusion according to any one of claims 1-3, characterized in that, The cross-target category contrast loss is expressed as: In the above formula, Representation and Features A set of features from the same category of targets. Then it represents the cardinality of the set. Indicates that it comes from a set Its characteristics.

5. A HRRP signal feature recognition device based on time-frequency domain fusion, characterized in that, The apparatus implements the HRRP signal feature recognition method based on time-frequency domain fusion as described in any one of claims 1-4, and the apparatus comprises: The HRRP signal sample acquisition module is used to acquire an HRRP signal sample set, which includes HRRP signal data obtained by multiple radar stations detecting the same target. The data preprocessing module is used to perform time-frequency alignment and motion compensation on the HRRP signal data obtained from each radar site to obtain preprocessed data. The time-frequency domain feature extraction module is used to input the preprocessed data corresponding to each radar station into the corresponding HRRP signal feature recognition network. In the HRRP signal feature recognition network, the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives of the preprocessed data are extracted by parallel time-domain segmentation coding unit, global frequency-domain embedding unit, and time-frequency analysis unit, respectively. The recognition unit then fuses the time-domain feature primitives, frequency-domain feature primitives, and time-frequency feature primitives to obtain single-station fused features. Target recognition is then performed based on the single-station fused features to obtain target prediction results. The multi-site joint network training module is used to combine the single-site fusion features, target prediction results and corresponding ground truth labels of multiple radar sites to calculate cross-site contrastive learning loss and cross-target category contrastive loss, and adjust the learnable parameters in each HRRP signal feature recognition network until convergence, thus obtaining the trained HRRP signal feature recognition network. The target recognition module is used to acquire the HRRP signal received from a certain station, and use a trained HRRP signal feature recognition network to perform target recognition on the HRRP signal to obtain the target recognition result.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

Images