Low-altitude radar target detection method based on large language model semantic prompt and relation reasoning

Abstract

The application discloses a low-altitude radar target detection method based on large language model semantic prompt and relation reasoning, which comprises the following steps: collecting original point track and track data of a radar target in a low-altitude scene, and constructing an original data set; performing data standardization and normalization processing to obtain a standardized feature sequence, performing time sequence segmentation based on a sliding window interception method to obtain a time sequence feature sequence of the radar target, constructing a digital text prompt word, performing hierarchical parameter regulation on a pre-trained large language model, and outputting radar high-level semantic features; performing dynamic screening on the radar high-level semantic features to obtain key features strongly related to a target category and generate a fixed-length feature vector, performing correlation reasoning based on a radar target decoding network, and outputting a category recognition result of the radar target, so that accurate determination of the target category is realized.

Description

Technical Field

[0001] This invention relates to the fields of radar signal processing and artificial intelligence technology, specifically to a low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model. Background Technology

[0002] With the rapid development of the low-altitude economy, low-altitude airspace is gradually opening up, and flight activities are becoming increasingly frequent, leading to a continuous increase in pressure on airspace safety management. Low-altitude targets are diverse, encompassing aircraft such as fixed-wing and rotary-wing aircraft, as well as birds, boulders, and other objects. These targets typically exhibit characteristics such as weak electromagnetic scattering and motion Doppler shift ambiguity, posing a significant challenge to target identification by low-altitude radar in complex environments. Currently, low-altitude target surveillance still primarily relies on traditional radar technology, utilizing electromagnetic wave reflection and the Doppler effect for target detection and identification. However, these methods generally suffer from insufficient environmental adaptability and low target classification accuracy, particularly when dealing with "low, slow, and small" targets (such as drones, birds, and boulders), making it difficult to meet the higher requirements of low-altitude airspace safety monitoring and the current demand for high-precision, high-efficiency target identification in airspace safety management. Against this backdrop, the introduction of artificial intelligence technology, especially intelligent target recognition methods based on deep learning, has become an important way to improve the target identification capabilities of low-altitude radar. In recent years, deep learning technology has been driving radar target recognition from traditional rule matching and signal processing towards intelligence and precision. However, existing deep learning methods still face several challenges: radar signal data is scarce and annotation costs are high, especially in target recognition tasks in complex environments, where model training often relies on large amounts of labeled data. At the same time, existing deep learning models have high computational complexity, making them difficult to deploy directly to edge computing devices or low-power hardware platforms, thus limiting their real-time performance and practicality in real-world applications.

[0003] Therefore, how to effectively integrate deep learning with low-altitude radar systems to improve target recognition performance and meet the dual requirements of real-time performance and efficiency for low-altitude safety monitoring has become a pressing technical challenge. Summary of the Invention

[0004] The purpose of this invention is to propose a low-altitude radar target detection method based on semantic prompts and relational reasoning using a large language model to solve the problems mentioned in the background art.

[0005] To achieve the aforementioned objectives, the first technical solution adopted by this invention is: a low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model, comprising the following steps: S1: Collect raw point and track data of radar targets in low-altitude scenarios to construct the raw dataset; S2, the original dataset is standardized and normalized to obtain the standardized feature sequence. The standardized feature sequence is then segmented temporally based on the sliding window truncation method to obtain the temporal feature sequence of the radar target. S3 encodes the temporal feature sequence of radar targets into structured natural language to construct digital textual prompts; S4, input the digital textual prompts into a pre-trained large language model based on the Transformer structure, perform hierarchical parameter adjustment on the pre-trained large language model, and output high-level semantic features of the radar. S5, based on the adaptive aggregation mechanism of radar target discriminative features, dynamically filters the high-level semantic features of radar to obtain key features strongly related to the target category and generate a fixed-length feature vector. S6. Construct a multi-dimensional association strength representation vector, input the fixed-length feature vector into the radar target decoding network for association reasoning, and output the radar target category recognition result.

[0006] Furthermore, the original dataset in step S1 is:

[0007] Where N is the total number of samples, Let be the data of the i-th point, and ti be the corresponding track data.

[0008] Let the data point P of the i-th point be defined as a 9-dimensional vector: ; Where T is a point in time, B is the batch number, and R is the distance. It refers to location. It is looking up and down. is the Doppler velocity, p is the amplitude, SNR is the signal-to-noise ratio, N is the number of raw points, and m represents the number of time points at 3-second intervals. The i-th trajectory data t is defined as a 10-dimensional vector: ; in, These are the filtered range, azimuth, and pitch; Vtotal is the total velocity. These are the velocity components in the X, Y, and Z directions, respectively. It represents the heading, and m represents the number of time points at 3-second intervals.

[0009] Furthermore, the original dataset undergoes data standardization and normalization processing, specifically including: Select a subset of features related to the target's motion state and scattering characteristics to form the track point feature set of the i-th sample. The track point feature set includes full velocity, Z-axis velocity, heading, azimuth, Doppler velocity, amplitude, signal-to-noise ratio, number of original points, and time features. Z-Score normalization was applied to the full velocity, Z-axis velocity, Doppler velocity, azimuth, and heading. Min-Max normalization was applied to the amplitude, signal-to-noise ratio, and number of original points to obtain the normalized and normalized feature representations.

[0010] Furthermore, the standardized formula is: ; in, These are the standardized eigenvalues. The original feature values ​​are μ, the mean of the m corresponding data, and σ is the standard deviation. The data normalization formula is: ; Among them, X min and X max This represents the minimum and maximum values ​​among m data points. A sliding window of size 5 is used to perform temporal segmentation on the standardized feature sequence to obtain the temporal feature sequence of the radar target.

[0011] Furthermore, the sliding window size is set to 5, and data from 5 consecutive time points are selected each time. Let the current time point be Tn, then the time series window selected for the first time is: Window n =[X n ,X n+1 ,X n+2 ,X n+3 ,X n+4 ]; On the next input, the sliding window moves forward one step, and the new sequence window is: Window n+1 =[X n+1 ,X n+2 ,X n+3 ,X n+4 ,X n+5 ].

[0012] Furthermore, step S4 specifically includes: The generated text prompt sequence is serialized to obtain the token sequence; The token sequence is embedded and mapped, and then superimposed with location information to form an initial feature representation; A hierarchical parameter adjustment strategy is adopted. The initial feature representation is input into a pre-trained large language model consisting of an embedding mapping module and 12 layers of coding units. The parameters of the first 4 layers of the network are frozen, and the parameters of the last 8 layers of the network are fine-tuned to output high-level semantic features of the radar.

[0013] Furthermore, step S5 specifically includes: K learnable category guidance vectors are constructed, and two sets of feature representations are obtained by linear mapping of the high-level semantic features of the radar. The correlation weights between the category guidance vectors and each feature unit are calculated. Based on relevance weights, features are adaptively weighted and aggregated, and aggregated features corresponding to all categories are concatenated to form a fixed-length discriminative feature vector.

[0014] Furthermore, step S6 specifically includes: For each target category Generate an embedding vector The formula for calculating the embedding vector is as follows: ; in, It is a learnable class embedding matrix, and onehot(c) is the one-hot encoded vector of class c; Define sample fusion features With category prototype embedding Multidimensional correlation strength representation vector between ; Multidimensional correlation strength representation vector Inputting the data into a multilayer perceptron (MLP) inference network allows for deep mining of category relationships, yielding category strength association scores. ; For all categories Calculate the association strength score for each category. Then, the scores are summarized to obtain the score vectors for all categories. The formula is as follows: ; The association strength scores are transformed into a probability distribution using the softmax function, as shown in the following formula: ; Radar target category identification results This refers to the category with the highest probability, as shown in the following formula: .

[0015] Furthermore, the Multilayer Perceptron (MLP) inference network performs in-depth mining of category relationships, specifically including: the Multilayer Perceptron (MLP) inference network consists of a first layer MLP, a second layer MLP, and a linear layer; The correlation strength characterization vector The input is fed into the fully connected layer of the first MLP layer to obtain the first intermediate result. The formula is as follows: ; The first intermediate result The input is fed into the second fully connected layer to obtain the second intermediate result. The formula is as follows: ; Where W is the learnable class embedding matrix, b is the bias term, and ReLU is the activation function; Based on the linear layer Converted into the final category association strength score The formula is as follows: .

[0016] Furthermore, it also includes: constructing a teacher-student network architecture, with the teacher network containing a 12-layer Transformer encoder and a 3-layer MLP discriminant module, and the student network containing a 6-layer Transformer encoder and a 2-layer MLP discriminant module; constructing a multi-level cognitive transfer mechanism through feature representation transfer, attention pattern transfer, and discriminant behavior transfer, and obtaining a lightweight edge deployment model by combining a model compression strategy of parameter quantization and structural pruning.

[0017] Due to the application of the above technical solution, the present invention has the following advantages compared with the prior art: This application can transform the numerical features of radar into natural language descriptions with physical semantics, and perform domain fine-tuning through a pre-trained large language model, thereby achieving the extraction of deep semantic features.

[0018] By adaptively aggregating category-guided vectors, physical features strongly correlated with the target category are highlighted, improving the accuracy of target recognition in complex low-altitude scenarios. The sliding window ensures the temporal continuity of the input data, effectively capturing changes in the target over time, preserving the continuous characteristics of target motion, and reducing the interference of short-term radar data noise on the recognition results.

[0019] By constructing a multi-dimensional representation of the correlation strength between sample features and category prototypes, and using inference networks to deeply mine these relationships, accurate determination of the target category can be achieved. Attached Figure Description

[0020] Figure 1 The flowchart of the low-altitude radar target detection method based on semantic prompting and relational reasoning of a large language model provided in the embodiments of the present invention is shown. Figure 2 The diagram shows the detection logic block diagram of the low-altitude radar target detection method based on semantic prompts and relational reasoning of a large language model provided in the embodiments of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or system that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or systems.

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] like Figures 1-2 As shown, this embodiment of the invention provides a low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model, comprising the following steps: S1: Collect raw point and track data of radar targets in low-altitude scenarios to construct the raw dataset; S2, the original dataset is standardized and normalized to obtain the standardized feature sequence. The standardized feature sequence is then segmented temporally based on the sliding window truncation method to obtain the temporal feature sequence of the radar target. S3 encodes the temporal feature sequence of radar targets into structured natural language to construct digital textual prompts; It should be noted that digital text-based prompt generation Based on sliding window temporal modeling, a structured natural language prompt word template rich in domain-specific physical semantic knowledge is constructed and defined as a four-segment prompt word structure.

[0025] This invention converts the continuous / discrete numerical features of radar targets into natural language segments that can be processed by large language models, and defines several semantic coding functions. This is used to format the value of a specified field using a preset text template. The "+" symbol indicates a string concatenation operation. This indicates that the numerical value will be converted into a character sequence according to a predetermined format (such as retaining decimal places, unit conversion, and symbol normalization) and filled into the template placeholder; ":" indicates the corresponding physical quantity obtained from radar observation, estimation, or preprocessing.

[0026] (1) Definition of semantic encoding; Temporal semantic encoding in Indicates a timestamp or sampling time identifier; Amplitude semantic coding ,in The observed or estimated value representing the amplitude of the target echo; Signal-to-noise ratio semantic coding ,in Indicates the target signal-to-noise ratio; Speed ​​semantic encoding: in Indicates the total speed. Represents the vertical velocity component. Indicates the Doppler radial velocity component; Heading semantic encoding ,in Indicates the heading angle; Semantic encoding of the number of original points ,in This indicates the number of original points used to generate the observations for this target; Location semantic encoding ,in Indicates the azimuth angle.

[0027] (2) Structured expression of prompt words; Using the context segment C, instruction segment I, statistical information segment S, and prediction result segment Q, a four-segment prompt word structure P=[C,I,S,Q] is formed, where S = The comprehensive semantic encoding function is defined as follows: The formula is as follows: ; This invention combines the multi-field numerical features of samples at the same time or the same target sample into a statistical information segment S in a predetermined order after semantic encoding.

[0028] Let the first The set of statistical features of a target sample is denoted as , Define a comprehensive semantic encoding function Used to output the corresponding statistical information text string, where the symbol " "This indicates sequential text fragment-level concatenation (distinguished from character-level concatenation "+"), thus ensuring the consistency and parsability of the arrangement of each field in the prompt words; in These are the observed or estimated values ​​for the corresponding fields; The S obtained by the above concatenation is a structured text representation with a fixed set of fields and a fixed order, in order to reduce the sensitivity of large language models to differences in input representations and improve inference stability.

[0029] (3) Four-segment prompt word structure; Context Segment C: Radar charts and trajectory data reflect the target's motion characteristics and behavioral patterns. Target amplitude, signal-to-noise ratio, velocity variations, and heading stability can be used to identify target types and assess potential risks. High signal-to-noise ratio and stable heading generally indicate reliable target tracking, while sudden changes in velocity may indicate anomalous behavior.

[0030] Instruction segment I: Predict target type.

[0031] Statistical information segment S: The current data point contains observations from n (n<=m) time points.

[0032] At a point time Tn:HH:MM:SS.ssssss (point time format), its full speed is , The z-axis velocity is The heading is The direction is Doppler velocity is The amplitude is Signal-to-noise ratio is The number of original points is ; At time Tn+1: HH:MM:SS.ssssss, its full velocity is , z-axis velocity The heading is The direction is Doppler velocity is The amplitude is Signal-to-noise ratio is The number of original points is ; Prediction result segment Q: Target types include: light drones / small drones / birds / aircraft, each corresponding to a prediction score [the maximum value corresponds to the category of the prediction result].

[0033] (4) Generation of complete prompt word sequence; Finally, the complete cue word sequence is obtained, and the generated digital textual cue word sequence is as follows: ; On the next input, the sliding window moves forward one step, and the new input sequence is: ; By using a sliding window to ensure the temporal continuity of the input data, the changes of the target over time can be effectively captured. Each new sequence generated by the sliding window is used as a token and input into the pre-trained GPT-2 model, helping the model to better understand the dynamic behavior of the target.

[0034] S4. Input the digital textual prompts into the pre-trained large language model based on the Transformer structure, perform hierarchical parameter adjustment on the pre-trained large language model, and output the high-level semantic features of the radar. S5, based on the adaptive aggregation mechanism of radar target discriminative features, dynamically filters the high-level semantic features of radar to obtain key features strongly related to the target category and generate a fixed-length feature vector. S6 constructs a multi-dimensional association strength representation vector, inputs the fixed-length feature vector into the radar target decoding network for association reasoning, and outputs the radar target category recognition result.

[0035] It should be noted that this method performs feature selection, standardization, and temporal modeling on the original radar point cloud data, selecting nine core features: point time, signal-to-noise ratio, number of original points, total velocity, Z-velocity, amplitude, heading, azimuth, and Doppler velocity, while removing features highly correlated with geographical location. Next, a sliding window mechanism is used for temporal segmentation, and a semantic mapping function is designed to assign physical semantic labels to the numerical features of these nine radar signals through a three-level transformation mechanism of "numerical-semantic-context," transforming them into structured natural language prompts rich in specific domain knowledge. Then, using a pre-trained GPT-2 model, a hierarchical fine-tuning strategy is employed. The shallow network is frozen to retain general language features, while the deep network is fine-tuned to learn radar-domain-specific semantics. An adaptive aggregation mechanism of radar target discriminative features actively filters and focuses on key discriminative features for each category. Finally, a support set-query set framework is constructed. A radar target category association strength decoder is used to build a multi-dimensional relationship representation between sample features and category prototypes. The category association strength is used to deeply mine inherent association patterns, achieving target recognition with high generalization capability. Finally, a multi-layered knowledge transfer compression method for radar edge deployment is adopted. Through three-dimensional distillation techniques involving feature layers, attention distribution, and relational reasoning, the knowledge of the complex teacher model is transferred to a lightweight student network, meeting the real-time requirements of edge computing while maintaining high accuracy. This invention provides high-precision, highly generalizable, and real-time deployable intelligent recognition technology support for radar target intelligent identification in the development of low-altitude economic security.

[0036] According to an embodiment of the present invention, the original dataset in step S1 is:

[0037] Where N is the total number of samples, Let be the data of the i-th point, and ti be the corresponding track data.

[0038] Let the data point P of the i-th point be defined as a 9-dimensional vector: ; Where T is the point in time, B is the batch number, and R is the distance (m). It is the direction (°). It is pitch (°). Where p is the Doppler velocity (m / s), p is the amplitude, SNR is the signal-to-noise ratio, N is the number of raw points, and m represents the number of time points at 3-second intervals. The i-th trajectory data t is defined as a 10-dimensional vector: ; in, These represent the filtered distance (m), azimuth, and pitch, respectively; Vtotal is the total velocity (m / s). These are the velocity components (m / s) in the X, Y, and Z directions, respectively. The heading is in degrees (°), and m represents the number of time points at 3-second intervals.

[0039] According to an embodiment of the present invention, the original dataset is subjected to data standardization and normalization processing, specifically including: Select a subset of features related to the target's motion state and scattering characteristics to form the track point feature set of the i-th sample. The track point feature set includes full velocity, Z-axis velocity, heading, azimuth, Doppler velocity, amplitude, signal-to-noise ratio, number of original points, and time features. Z-Score normalization was applied to the full velocity, Z-axis velocity, Doppler velocity, azimuth, and heading. Min-Max normalization was applied to the amplitude, signal-to-noise ratio, and number of original points to obtain the normalized and normalized feature representations.

[0040] According to an embodiment of the present invention, the standardized formula is: ; in, These are the standardized eigenvalues. The original feature values ​​are μ, the mean of the m corresponding data, and σ is its standard deviation. The data normalization formula is: ; Among them, X min and X max This represents the minimum and maximum values ​​among m data points. A sliding window of size 5 is used to perform temporal segmentation on the standardized feature sequence to obtain the temporal feature sequence of the radar target.

[0041] According to an embodiment of the present invention, the size of the sliding window is set to 5, and data from 5 consecutive time points are selected each time. Let the current time point be Tn, then the time series window selected for the first time is: Window n =[X n ,X n+1 ,X n+2 ,X n+3 ,X n+4 ]; On the next input, the sliding window moves forward one step, and the new sequence window is: Window n+1 =[X n+1 ,X n+2 ,X n+3 ,X n+4 ,X n+5 ].

[0042] It should be noted that by using a sliding window to ensure the temporal continuity of the input data, the changing characteristics of the target on the time axis can be effectively captured, providing temporal context information for subsequent semantic transformation.

[0043] Assume the target is at discrete sampling time. The statistical eigenvector is denoted as (where subscript) Indicates the time sequence number. (It can be composed of observations such as velocity, heading, azimuth, signal-to-noise ratio, number of points, and amplitude at that moment). Setting the sliding window length to 5 and extracting time sequences with an adjacent window step size of 1, the 0th time sequence window is defined as: The first timing window is obtained after the window slides forward one step. , where the symbol This indicates the concatenation of sequences in chronological order.

[0044] According to an embodiment of the present invention, step S4 specifically includes: The generated text prompt sequence is serialized to obtain the token sequence; The token sequence is embedded and mapped, and then superimposed with location information to form an initial feature representation; A hierarchical parameter adjustment strategy is adopted. The initial feature representation is input into a pre-trained large language model consisting of an embedding mapping module and 12 layers of coding units. The parameters of the first 4 layers of the network are frozen, and the parameters of the last 8 layers of the network are fine-tuned to output high-level semantic features of the radar.

[0045] It should be noted that by leveraging the advantages of the Transformer structure in sequence modeling, its semantic representation capabilities are transferred to the radar feature space, enabling efficient characterization of the semantic attributes of radar targets. The pre-trained model used consists of an embedding mapping module and a multi-layer coding structure, and its overall form can be represented as follows: ; in, Indicates the embedded mapping layer, For the first Layers are feature encoding units based on a self-attention mechanism.

[0046] 1. Input mapping and initial feature construction; First, the constructed radar semantic cue information Serialization is performed to obtain a token sequence composed of discrete symbols: ; in, Represents a set of words, The sequence length is given.

[0047] The token sequence is embedded and mapped, and positional information is superimposed to obtain the initial feature representation: ; This forms the input feature matrix: ; This representation provides the basic input for subsequent deep semantic modeling.

[0048] 2. Semantic transfer strategy based on hierarchical parameter control; To balance the model's generalization ability with its adaptability to radar missions, a hierarchical parameter adjustment strategy is adopted during the model training phase. Specifically, the input features are sequentially fed into 12 layers of coding units for layer-by-layer transformation. The parameters of the first four layers are retained and not updated to maintain the model's ability to represent general sequence structures. The last eight layers are then optimized in a targeted manner to better adapt to the semantic features of radar targets.

[0049] The transformation process of each coding unit can be summarized as follows: ; The multi-head attention module captures the correlation between different features through parallel subspace mapping, while the feedforward network is responsible for performing nonlinear transformations on the fused features. Residual connections and normalization operations are introduced during the encoding process to enhance the stability and convergence of model training.

[0050] 3. High-level semantic feature output from radar; After successive mapping through multiple layers of coding units, the model gradually completes the abstract transformation from low-level structural information to high-level semantic information. Finally, the result of the last layer of coding is taken as the radar semantic feature representation. ; Through the above process, involving 12 layers of processing, each layer performs deep learning and transformation of the input features. Among these... As input features, This represents the transformation operation at the i-th level. The final output has a dimension of Where n represents the sequence length, This represents the feature dimension. Wherein, This high-dimensional radar feature representation, after fusing contextual semantic relationships, effectively reflects the semantic attributes and inherent relational structure of radar targets. This feature serves as input to the subsequent discriminative feature adaptive aggregation module, providing high-quality semantic support for radar target recognition.

[0051] According to an embodiment of the present invention, step S5 specifically includes: K learnable category guidance vectors are constructed, and two sets of feature representations are obtained by linear mapping of the high-level semantic features of the radar. The correlation weights between the category guidance vectors and each feature unit are calculated. Based on relevance weights, features are adaptively weighted and aggregated, and aggregated features corresponding to all categories are concatenated to form a fixed-length discriminative feature vector.

[0052] It should be noted that, considering the specific characteristics of radar target identification tasks, an adaptive aggregation mechanism for radar target discriminative features was designed. This mechanism uses learnable vectors to adaptively weight and aggregate variable-length semantic feature sequences, compressing them into fixed-length feature vectors while highlighting the physical features most relevant to the radar target category (such as signal-to-noise ratio, speed, and heading).

[0053] 1. Construction of category-guided vectors: This invention introduces K learnable category guidance vectors to characterize the discriminative feature patterns corresponding to different categories, which are defined as follows:

[0054] in, This represents the feature guidance vector associated with the i-th category, used to guide the model to focus on the information most relevant to that category during feature aggregation; In the formula, Category bootstrap vector pool (stores bootstrap vectors for all categories); :Bundle The vectors are stacked / assembled into a whole in sequence; each A guideline representing a category; : No. Guided vectors (learnable parameters) for each category; : Number of categories (total number of learnable categories); : Matrix space over the real number field, representing It is OK, The real-valued matrix of columns; Dimensions of the model's semantics / latent space.

[0055] 2. Mapping representation of encoded features: The output of the last layer of the encoder is denoted as This output contains high-level semantic features of each temporal unit in the input sequence. To facilitate subsequent attention calculation and feature fusion, we perform a linear mapping on this output to obtain two sets of feature representations: ; in, For learnable parameter matrix, Used for participating in correlation measurement Used to carry the aggregated feature information.

[0056] 3. Feature relevance weight calculation: For each category guiding vector Calculate the correlation strength between it and each feature unit in the input sequence. Let... Indicates the first The category-guided vector for the first... The importance coefficient of each feature unit is calculated as follows: ; This process measures the degree of matching between the class guide vector and the input features using an inner product form, and uses softmax for normalization to highlight the most discriminative feature locations.

[0057] 4. Adaptive aggregation of discriminative features: After obtaining the aforementioned relevance weights, the representation of the feature information is... We perform a weighted summation to form the aggregated features corresponding to each category: ; in, Indicates the first The adaptive pooling features corresponding to each category comprehensively reflect the discriminative information in the input sequence that is most relevant to that category.

[0058] 5. Generation of fixed-length feature representations: Concatenate the pooled features corresponding to all categories to form a unified fixed-length feature representation. ; This feature vector While maintaining the integrity of category discrimination information, the model effectively compresses variable-length input sequences. Through the aforementioned category-guided adaptive feature aggregation mechanism, the model can automatically filter and fuse the most discriminative physical feature information from variable-length radar echo sequences. Each category-guided vector plays the role of "category prior" in the feature aggregation process, enabling the model to focus on key information closely related to the target category. For example, the signal-to-noise ratio reflects the target echo intensity; velocity and its changing trend characterize the target's maneuverability; and heading stability reflects the differences in motion patterns among different target types. The final feature representation Z comprehensively characterizes the discriminative features of radar targets, providing high-quality input for subsequent target classification and recognition tasks, thereby effectively improving the recognition performance of complex low-altitude targets.

[0059] According to an embodiment of the present invention, step S6 specifically includes: For each target category Generate an embedding vector The formula for calculating the embedding vector is as follows: ; in, It is a learnable class embedding matrix, and onehot(c) is the one-hot encoded vector of class c; Indicates the dimension of the category embedding space; Represents a set of categories. This represents the number of categories in the set. Indicate category The embedding vector; Define sample fusion features With category prototype embedding Multidimensional correlation strength representation vector between Among these methods: direct splicing captures the overall correlation; difference reveals the differential features; and element-wise multiplication represents the interactive synergistic effect. Through these operations, the correlation strength representation vector comprehensively depicts the intrinsic connection between radar target features and category prototypes.

[0060] Multidimensional correlation strength representation vector Inputting the data into a multilayer perceptron (MLP) inference network allows for deep mining of category relationships, yielding category strength association scores. ; For all categories Calculate the association strength score for each category. Then, the scores are summarized to obtain the score vectors for all categories. The formula is as follows: ; The association strength scores are transformed into a probability distribution using the softmax function, as shown in the following formula: ; In the formula, : Probability function; : Random variables (category label variables) corresponding to the classification results; A candidate category (belonging to a set of categories) ); : Condition symbol, indicating "under the condition of"; : Fused feature vector, which is the target representation after fusing multi-source / multi-modal information, used for subsequent classification and determination; Given fusion features Under these conditions, the target belongs to category The posterior probability; :category Association strength score (from (corresponding components in the middle). : Represents any one of the categories in the sample and category set. The correlation strength score, Exponential function (natural index) This is used to ensure that the output is positive and to amplify the difference in fractions; : For a set of categories Sum the probabilities of all categories one by one (using normalization / partition function) to ensure that the sum of the probabilities of all categories is 1; It iterates over variables.

[0061] Radar target category identification results This refers to the category with the highest probability, as shown in the following formula: .

[0062] It should be noted that, through the radar target inference decoder, the model can delve into the intrinsic correlation between sample features and category prototypes, learning transferable relational reasoning patterns. This reasoning approach enables the model to achieve accurate classification through more detailed relational analysis when dealing with complex situations involving "low, slow, and small" targets (such as drones, birds, and aerostats) that have similar radar echo characteristics, demonstrating stronger discriminative ability and generalization performance.

[0063] According to an embodiment of the present invention, a multilayer perceptron (MLP) inference network is used for in-depth mining of category relationships, specifically including: the multilayer perceptron (MLP) inference network includes a first layer MLP, a second layer MLP, and a linear layer; The correlation strength characterization vector The input is fed into the fully connected layer of the first MLP layer to obtain the first intermediate result. The formula is as follows: ; The first intermediate result The input is fed into the second fully connected layer to obtain the second intermediate result. The formula is as follows: ; Where W is the learnable class embedding matrix, b is the bias term, and ReLU is the activation function; Based on the linear layer Converted into the final category association strength score The formula is as follows: .

[0064] According to an embodiment of the present invention, the method further includes: constructing a teacher-student network architecture, wherein the teacher network contains a 12-layer Transformer encoder and a 3-layer MLP discriminant module, and the student network contains a 6-layer Transformer encoder and a 2-layer MLP discriminant module; constructing a multi-level cognitive transfer mechanism through feature representation transfer, attention pattern transfer, and discriminant behavior transfer, and obtaining a lightweight edge deployment model by combining a model compression strategy of parameter quantization and structural pruning.

[0065] It should be noted that, for applications where radar target recognition models face limitations in computing power, power sensitivity, and high real-time requirements on edge computing platforms, this invention proposes a multi-scale cognitive transfer and model compression method tailored to the characteristics of radar signal processing. This method, centered on teacher-student network collaborative learning, achieves effective transfer of high-performance radar target recognition capabilities to lightweight models through hierarchical transfer of radar representation structures, attention distribution patterns, and discrimination decision-making behaviors. This supports real-time intelligent perception for edge radar devices, specifically including: 1. Radar-based cognitive-guided teacher-student network design: The teacher network adopts a high-capacity radar feature modeling structure, consisting of... The model consists of a Transformer encoder and a 3-layer Multilayer Perceptron (MLP) discriminator module. This model can fully learn the temporal correlations, spectral structure, and target motion characteristics in radar echo sequences, forming a stable and highly discriminative high-level semantic representation of radar. The student network performs structural pruning and depth compression for edge computing scenarios, including... The 1-layer Transformer encoder and 2-layer MLP decision mode significantly reduce the model parameter size and computational complexity while retaining the ability to model key physical features of radar.

[0066] 2. A multi-level cognitive transfer mechanism oriented towards radar characteristics: To avoid the problem that traditional distillation methods only focus on output consistency and ignore the internal structural features of radar signals, this invention constructs a three-layer progressive knowledge transfer mechanism of feature representation, attention mode, and discrimination behavior.

[0067] (1) Radar feature representation migration; This level constrains the way the student model models radar echo features in the mid-to-high-level representation space, making it approximate the teacher model in terms of feature distribution structure. Its feature representation transfer loss is defined as: ; in, and d represents the radar feature representation of the teacher and student models at the corresponding levels, respectively, where d is the feature dimension.

[0068] (2) Radar attention pattern shift; In radar target recognition, the contributions of different time slices and feature dimensions to the final discrimination result vary significantly. To ensure that the student model maintains consistency with the teacher model in feature attention behavior, thereby continuously focusing on key radar physical features (such as strong echo regions, locations of abrupt changes in motion state, etc.), a constraint term based on the difference in attention distribution is constructed, and its loss function is defined as: ; in, and The teacher model and the student model respectively represent the first two models in the second stage. In the sample, the first Attention weights at each feature location.

[0069] (3) Radar discrimination behavior migration; In the final discrimination stage, by constraining the numerical difference between the outputs of the teacher model and the student model, the student model learns the discrimination boundary shape of the teacher model. Its discrimination behavior transfer loss is defined as: ; in, and These represent the teacher model and the student model in the sample, respectively. The output results on This represents the number of target categories. This constraint helps the student model maintain stable and consistent classification behavior in complex radar target scenarios.

[0070] (4) Joint cognitive transfer objective function; Based on the above multi-level cognitive transfer constraints, the overall knowledge transfer loss function is constructed as follows: ; in, These are weighting coefficients used to balance the transfer contributions of different cognitive levels.

[0071] 3. Lightweight collaborative optimization of models for edge deployment: After completing multi-level knowledge transfer training, a radar edge deployment-friendly model compression strategy is further introduced into the student model. During training, the optimization objective of the student model is defined as: ; in, This represents the parameters of the lightweight model. Based on this, the following two types of compression operations are combined: (1) Parameter quantization: Reduce the numerical precision of weights and activations to reduce storage overhead and computational energy consumption; (2) Structural pruning: Remove redundant computing units that contribute little to radar target discrimination to further improve inference efficiency.

[0072] 4. Final Lightweight Radar Edge Model: Through multi-level cognitive transfer and collaborative compression optimization, the final deployable model is obtained: ; This model maintains radar target recognition accuracy while significantly reducing model size and computational complexity, enabling stable operation on resource-constrained radar edge computing devices. This method provides a practical deployment solution for radar intelligent sensing systems in low-altitude economic and security scenarios, characterized by high precision, low power consumption, and strong robustness.

[0073] In summary, the present invention has the following advantages compared with the prior art: This application can transform the numerical features of radar into natural language descriptions with physical semantics, and perform domain fine-tuning through a pre-trained large language model, thereby achieving the extraction of deep semantic features.

[0074] By adaptively aggregating category-guided vectors, physical features strongly correlated with the target category are highlighted, improving the accuracy of target recognition in complex low-altitude scenarios. The sliding window ensures the temporal continuity of the input data, effectively capturing changes in the target over time, preserving the continuous characteristics of target motion, and reducing the interference of short-term radar data noise on the recognition results.

[0075] By constructing a multi-dimensional representation of the correlation strength between sample features and category prototypes, and using inference networks to deeply mine these relationships, accurate determination of the target category can be achieved.

[0076] Those skilled in the art will understand that, for ease of explanation, the example is provided with one memory and one processor. In actual terminals or servers, multiple processors and memories may exist. Memory can also be referred to as storage medium or storage device, etc., and the embodiments of this application do not limit this.

[0077] It should be understood that in the embodiments of this application, the processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor may also be a general-purpose microprocessor, graphics processing unit (GPU), or one or more integrated circuits to execute relevant programs to achieve the functions required by the embodiments of this application.

[0078] The processor can also be an integrated circuit chip with signal processing capabilities. In implementation, each step of this application can be completed through integrated logic circuits in the processor hardware or instructions in software form. The aforementioned processor can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The steps of the methods disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the functions required by the units included in the methods, systems, and storage media of the embodiments of this application.

[0079] It should also be understood that the memory mentioned in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache.

[0080] By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0081] The memory can also be a Compact Disc Read-Only Memory (CD-ROM) or other optical disc storage, optical disk storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto. The memory can exist independently and be connected to the processor via a bus. The memory can also be integrated with the processor. The memory can store programs, and when the program stored in the memory is executed by the processor, the processor performs the various steps of the method determined in the above embodiments of this application.

[0082] It should be noted that when the processor is a general-purpose processor, DSP, ASIC, FPGA, or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, the memory (storage module) is integrated into the processor. It should be noted that the memory described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0083] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0084] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules within the processor. The software modules can reside in mature storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. Since this storage medium is located in memory, the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method; to avoid repetition, these will not be described in detail here.

[0085] Those skilled in the art will recognize that the various illustrative logical blocks (ILBs) and steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0086] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer-programmed program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a processor, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a computer network, or other programmable device.

[0087] This embodiment also provides a computer-readable storage medium storing a computer program that enables a computer to execute in order to implement the above-described low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model.

[0088] It should be noted that computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic) or wireless (e.g., infrared, wireless, microwave, etc.) means, or from one website, computer, server, or data center to a mobile phone processor via a wired means. A computer-readable storage medium can be any usable medium that a computer can access, or a data storage system such as a server or data center that integrates one or more usable media. Usable media can be magnetic media (e.g., floppy disks, hard disks), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives), etc.

[0089] Finally, it should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model, characterized in that, Includes the following steps: S1: Collect raw point and track data of radar targets in low-altitude scenarios to construct the raw dataset; S2, the original dataset is standardized and normalized to obtain the standardized feature sequence. The standardized feature sequence is then segmented temporally based on the sliding window truncation method to obtain the temporal feature sequence of the radar target. S3 encodes the temporal feature sequence of radar targets into structured natural language to construct digital textual prompts; S4, input the digital textual prompts into a pre-trained large language model based on the Transformer structure, perform hierarchical parameter adjustment on the pre-trained large language model, and output high-level semantic features of the radar. S5, based on the adaptive aggregation mechanism of radar target discriminative features, dynamically filters the high-level semantic features of radar to obtain key features strongly related to the target category and generate a fixed-length feature vector. S6. Construct a multi-dimensional association strength representation vector, input the fixed-length feature vector into the radar target decoding network for association reasoning, and output the radar target category recognition result.

2. The low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model as described in claim 1, characterized in that, The original dataset in step S1 is: ， Where N is the total number of samples, Let ti be the data of the i-th point, and ti be the corresponding track data; Let the data point P of the i-th point be defined as a 9-dimensional vector: ； Where T is a point in time, B is the batch number, and R is the distance. It refers to location. It is looking up and down. is the Doppler velocity, p is the amplitude, SNR is the signal-to-noise ratio, N is the number of raw points, and m represents the number of time points at 3-second intervals. The i-th trajectory data t is defined as a 10-dimensional vector: ； in, These are the filtered range, azimuth, and pitch; Vtotal is the total velocity. These are the velocity components in the X, Y, and Z directions, respectively. It represents the heading, and m represents the number of time points at 3-second intervals.

3. The low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model as described in claim 2, characterized in that, The original dataset undergoes data standardization and normalization processing, specifically including: Select a subset of features related to the target's motion state and scattering characteristics to form the track point feature set of the i-th sample. The track point feature set includes full velocity, Z-axis velocity, heading, azimuth, Doppler velocity, amplitude, signal-to-noise ratio, number of original points, and time features. Z-Score normalization was applied to the full velocity, Z-axis velocity, Doppler velocity, azimuth, and heading. Min-Max normalization was applied to the amplitude, signal-to-noise ratio, and number of original points to obtain the normalized and normalized feature representations.

4. The low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model as described in claim 3, characterized in that, The standardized formula is: ； in, These are the standardized eigenvalues. The original feature values ​​are μ, the mean of the m corresponding data, and σ is the standard deviation. The data normalization formula is: ； Among them, X min and X max This represents the minimum and maximum values ​​among m data points. A sliding window of size 5 is used to perform temporal segmentation on the standardized feature sequence to obtain the temporal feature sequence of the radar target.

5. The low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model as described in claim 1, characterized in that, The sliding window size is set to 5, and data from 5 consecutive time points are selected each time. Let the current time point be... The first selected timing window is: Window n =[X n ,X n+1 ,X n+2 ,X n+3 ,X n+4 ]; On the next input, the sliding window moves forward one step, and the new sequence window is: Window n+1 =[X n+1 ,X n+2 ,X n+3 ,X n+4 ,X n+5 ]。 6. The low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model as described in claim 1, characterized in that, Step S4 specifically includes: The generated text prompt sequence is serialized to obtain the token sequence; The token sequence is embedded and mapped, and then superimposed with location information to form an initial feature representation; A hierarchical parameter adjustment strategy is adopted. The initial feature representation is input into a pre-trained large language model consisting of an embedding mapping module and 12 layers of coding units. The parameters of the first 4 layers of the network are frozen, and the parameters of the last 8 layers of the network are fine-tuned to output high-level semantic features of the radar.

7. The low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model as described in claim 6, characterized in that, Step S5 specifically includes: K learnable category guidance vectors are constructed, and two sets of feature representations are obtained by linear mapping of the high-level semantic features of the radar. The correlation weights between the category guidance vectors and each feature unit are calculated. Based on relevance weights, features are adaptively weighted and aggregated, and aggregated features corresponding to all categories are concatenated to form a fixed-length discriminative feature vector.

8. The low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model as described in claim 7, characterized in that, Step S6 specifically includes: For each target category Generate an embedding vector The formula for calculating the embedding vector is as follows: ； in, It is a learnable class embedding matrix, and onehot(c) is the one-hot encoded vector of class c; Define sample fusion features With category prototype embedding Multidimensional correlation strength representation vector between ; Multidimensional correlation strength representation vector Inputting the data into a multilayer perceptron (MLP) inference network allows for deep mining of category relationships, yielding category strength association scores. ; For all categories Calculate the association strength score for each category. Then, the scores are summarized to obtain the score vectors for all categories. The formula is as follows: ； The association strength scores are transformed into a probability distribution using the softmax function, as shown in the following formula: ； Radar target category identification results This refers to the category with the highest probability, as shown in the following formula: 。 9. The low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model as described in claim 8, characterized in that, The Multilayer Perceptron (MLP) inference network is used for in-depth mining of category relationships. Specifically, the Multilayer Perceptron (MLP) inference network includes a first-layer MLP, a second-layer MLP, and a linear layer. The correlation strength characterization vector The input is fed into the fully connected layer of the first MLP layer to obtain the first intermediate result. The formula is as follows: ； The first intermediate result The input is fed into the second fully connected layer to obtain the second intermediate result. The formula is as follows: ； Where W is the learnable class embedding matrix, b is the bias term, and ReLU is the activation function; Based on the linear layer Converted into the final category association strength score The formula is as follows: 。 10. The low-altitude radar target detection method based on semantic prompting and relational reasoning using a large language model as described in claim 1, characterized in that, Also includes: Construct a teacher-student network architecture. The teacher network contains a 12-layer Transformer encoder and a 3-layer MLP discriminant module, while the student network contains a 6-layer Transformer encoder and a 2-layer MLP discriminant module. A multi-level cognitive transfer mechanism is constructed by feature representation transfer, attention pattern transfer, and discriminative behavior transfer. Combined with the model compression strategy of parameter quantization and structural pruning, a lightweight edge deployment model is obtained.

Images