Method for identifying working mode of multifunctional radiation source with distributed hierarchical architecture

Abstract

The invention discloses a method for identifying a working mode of a multifunctional radiation source with a distributed hierarchical architecture, which adopts a distributed hierarchical model architecture, carries out receiver-level hierarchical feature coding through a Conformer encoder, carries out cross-receiver feature fusion through a Transformer encoder, realizes feature coding of a plurality of radiation source signal receivers, and carries out feature fusion on a plurality of radiation source signal receivers. The method comprises the steps that firstly, feature representation is obtained after aggregation and fusion, then, an attention pooling mechanism is used for aggregating the feature representation after fusion according to the learned weight, a single feature vector is obtained, finally, probability distribution mapped to a plurality of predefined working modes is achieved through a linear classification layer and a Softmax function, and the category with the highest probability is the recognition result. According to the method provided by the invention, the limitation of a single-station multifunctional radiation source working mode identification method in the prior art in the aspects of signal agility processing, incomplete observation, complex electromagnetic environment interference, unknown modes and the like is solved, and the accuracy and robustness of identification are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of radiation source working mode recognition, and in particular relates to a distributed hierarchical architecture multifunctional radiation source working mode recognition method. Background Art

[0002] In recent years, advancements in related technologies have led to significant progress in target detection, interference mitigation, and intelligent operations. These advancements have significantly enhanced operational efficiency and intelligent processing capabilities in modern applications. Understanding multi-function emitters is a key component of this advancement. Multi-function emitters, capable of scheduling multiple concurrent functions within their timeline, such as wide-area surveillance and precise target tracking, have been widely deployed on various platforms.

[0003] Effectively identifying the operating modes of multifunctional emitters is crucial for understanding system behavior, but this faces significant technical challenges. First, processing the non-cooperative nature of multifunctional emitter signals is inherently difficult, especially in complex electromagnetic environments. Second, manual signal labeling is impractical due to cost and reliability issues. In addition, the lack of intercepted data for the key operating modes of multifunctional emitters often leads to unbalanced training data sets for identification algorithms. Therefore, the research community has begun to focus on methods such as multifunctional emitter data enhancement, adversarial generative learning, and few-shot learning to alleviate these limitations.

[0004] Furthermore, signals intercepted by receivers during dynamic operation are often subject to errors or distortion caused by non-ideal factors. These issues stem from limitations in signal sorting and deinterleaving capabilities, as well as the fundamental physical constraints of the receiver hardware. Intercepted signals may also contain emissions from completely unknown multi-functional emitter patterns. The resulting discrepancy between the signals used to train recognition models and those encountered in real-world operation leads to a shift in the class distribution. This requires robust mechanisms within the recognition system to effectively sort and detect these unknown or anomalous signals. Consequently, specialized research areas focusing on signal sorting and open-set recognition have emerged.

[0005] To address these issues, current mainstream research methods for multi-functional radiation source operating mode recognition are mainly focused on single-station scenarios. These methods mainly use models based on recurrent neural networks (RNNs), including gated recurrent units (GRUs), long short-term memory networks (LSTMs), and their derivative architectures. In recent years, research on radiation source systems using distributed architectures and multi-sensor fusion technology has steadily increased. This research aims to meet the challenges of modern electromagnetic environments and involves areas such as active detection, pulse sequence segmentation, and pulse stream representation. However, research on the specific application of distributed architectures to the field of multi-functional radiation source operating mode recognition is still insufficient.

[0006] In general, existing methods for recognizing the working modes of multifunctional radiation sources face challenges in dealing with signal agility, incomplete observations, and complex electromagnetic environments. There is an urgent need for a method for recognizing the working modes of multifunctional radiation sources based on a distributed hierarchical architecture. Summary of the Invention

[0007] In order to solve the above technical problems, the present invention provides a distributed hierarchical architecture multi-function radiation source working mode recognition method, which solves the limitations of the existing single-station multi-function radiation source working mode recognition method in processing signal agility, incomplete observation, complex electromagnetic environment interference and unknown modes.

[0008] The technical solution adopted by the present invention is: a distributed hierarchical architecture multifunctional radiation source working mode recognition method, the specific steps are as follows:

[0009] S1, data generation;

[0010] Based on the typical pulse characteristic parameter settings of the multifunctional radiation source, multiple radiation source signal receivers are constructed to intercept the pulse description word PDW sequence data set describing the activity of the same multifunctional radiation source within a certain time window. The data set is divided into training set, validation set, and test set according to the set ratio, and the data of the training set, validation set, and test set are enhanced by simulating site loss and non-ideal factors.

[0011] S2, data preprocessing;

[0012] Based on step S1, the pulse descriptor PDW sequence dataset is obtained, and the pulse parameters in each sequence are normalized at the feature level. Then the sequence is padded to achieve a unified maximum sequence length, and the corresponding attention mask is generated. Finally, the sequences processed by all radiation source signal receivers are transposed and stacked to form the input tensor of the model.

[0013] S3, distributed hierarchical model architecture construction;

[0014] The distributed hierarchical model includes an input embedding layer, a Conformer encoder module, a Transformer encoder module, and an output layer, and is deployed in a system including multiple spatially distributed radiation source signal receivers.

[0015] The distributed hierarchical model first performs receiver-level hierarchical feature encoding through the Conformer encoder and cross-receiver feature fusion through the Transformer encoder to achieve feature encoding of multiple radiation source signal receivers and obtain the aggregated and fused feature representation;

[0016] Then, the attention pooling mechanism is used to aggregate the fused feature representations according to the learned weights to obtain a single feature vector; finally, a linear classification layer and a Softmax function are used to map it to the probability distribution of multiple predefined working modes, and the category with the highest probability is the result of this recognition.

[0017] S4, distributed hierarchical model training;

[0018] The preprocessed training set and validation set obtained in step S2 are input into the distributed hierarchical model architecture in step S3, and the model is trained end-to-end using a composite loss function adapted to the distributed architecture to obtain a model that has converged.

[0019] S5, distributed hierarchical model testing;

[0020] The preprocessed test set obtained in step S2 is input into the trained converged model obtained in step S4 for testing, and the test results are output and the accuracy of model recognition is evaluated.

[0021] Furthermore, the step S1 is specifically as follows:

[0022] First, a simulation scenario including a mobile multifunctional radiation source transmitter and K mobile distributed radiation source signal receivers is set, and the basic parameters of the transmitter and receiver are set.

[0023] Among them, the basic parameters include: radiation source transmission power, radiation source transmission antenna gain, receiver reception sensitivity, receiver system loss, radiation source transmission elevation angle and azimuth angle range, receiver reception gain, movement speed and position information of the radiation source transmitter and radiation source signal receiver.

[0024] Then, the C multi-function radiation source working modes that need to be identified are defined, and the variation range of the pulse parameters, modulation type, typical pulse number PN, and PRI group number GN in each mode are set to generate an ideal multi-function radiation source signal sample set and data labels, and divide them into training set, validation set, and test set according to the preset ratio.

[0025] Each sample in the sample set represents an operating mode instance, and its parameters are randomly generated within a preset range. The pulse parameters in each mode include: PRI, PW, and RF.

[0026] We then simulated site loss and non-ideal factors to perform data augmentation on the training, validation, and test sets. For each ideal example, we generated samples with different receiver combinations, setting the data for failed receivers to zero. We also retained samples where all receivers functioned normally. Finally, we simulated non-ideal factors during signal transmission and reception to obtain the final training, validation, and test sets, along with their corresponding data labels.

[0027] Furthermore, the step S2 is specifically as follows:

[0028] S21, feature normalization;

[0029] Set the sequence of receiver k in the data set obtained in step S1 to P (k) , for the i-th pulse in the sequence The three parameters, PRI, PW, and RF, are normalized to obtain the normalized results. Finally, the normalized sequence is obtained

[0030] Among them, k∈K.

[0031] S22, sequence filling;

[0032] Determine the maximum length of all K receiver sequences in the current batch or the entire dataset Each normalized sequence Pad with zero vector to maximum length L max , update the normalized sequence after filling

[0033] Among them, L k represents the length of the receiver k sequence; p pad Indicates the filling amount, represents the normalized pulse parameters, Indicates size is M f ×L max The vector space of M f Indicates the number of features.

[0034] S23, attention mask generation;

[0035] Based on step S22, the padded normalized sequence for each receiver is Generate a corresponding binary attention mask where the original impulse positions are 1 and the padded positions are 0. Combine the masks of the K receivers into

[0036] S24, distributed tensor construction;

[0037] The K padded normalized sequences Transpose and stack to form the input feature tensor

[0038] in,(·) T represents transpose, Indicates size is K×L max ×M f The vector space of .

[0039] S25, construct model input;

[0040] The input feature tensor P and the attention mask M form the model input (P, M).

[0041] Furthermore, the step S3 is specifically as follows:

[0042] S31, input embedding layer;

[0043] M in the input tensor P f The d-dimensional impulse vector is projected onto the d model Dimension, and d model The positional encoding of the dimension is added element by element to obtain an embedded representation including temporal information

[0044] in, Indicates size is K×L max ×d model The vector space of .

[0045] S32, receiver-level hierarchical feature encoding;

[0046] At each receiver sequence X (k) Add [CLS] identifier before embedding Get sequence

[0047] Among them, concat(·) represents the concatenation operation, dim=0 means that the dimension of the concatenation operation starts from 0, Indicates size (L max +1)×d model The vector space of .

[0048] Then use the N enc An independent Conformer encoder module f of each Conformer block conf_enc Processing Sequence and the corresponding mask Output

[0049] Among them, h k,CLS represents the hidden state corresponding to [CLS], Indicates that the kth receiver is from 1 to the maximum length L max The hidden state of all pulses.

[0050] Finally, take the hidden state h corresponding to [CLS] k,CLS As the characteristic of the receiver k And aggregate the features of multiple receivers to obtain the aggregate feature matrix

[0051] S33, cross-receiver feature fusion;

[0052] Each receiver feature r (k) with a learnable receiver embedding Add element by element to get the modified aggregate feature matrix

[0053] Then use a fus The Transformer encoder module f is a stack of standard Transformer blocks fus Processing the Aggregate Feature Matrix Get the fusion feature matrix

[0054] Among them, f (1) ,...,f (K) represents the fused features from receiver 1 to receiver K.

[0055] S34, output layer;

[0056] The attention pooling mechanism is used to aggregate the fusion feature matrix F. First, each fusion feature u is calculated through a linear transformation with tanh activation k =tanh(W att f (k) +b att ); Then, by calculating the context vector Get importance score

[0057] Among them, u k 、u c 、e kRepresents fusion features, context vectors, and importance scores, respectively. W att and b att represents the learnable parameters in the linear transformation, and Indicates size d att ×d model and The vector space of Represents the size of the learnable parameters.

[0058] Apply the Softmax activation function to all K importance scores to obtain the normalized attention weights Weighted aggregation of F is used to obtain the final aggregated feature vector

[0059] Finally, the aggregate feature vector z is input into a weighted Bias is b c ∈R C The linear classification layer and Softmax activation function output the probability distribution of C predefined working modes The one with the highest probability is the final recognition result.

[0060] Furthermore, the step S4 is specifically as follows:

[0061] Use the preprocessed training set D obtained in step S2 train and validation set D val Data, using a composite loss function Conduct training.

[0062] Among them, the main loss function y c and They represent the probability of the true category c being the label and the recognition category c, γ represents the regularization coefficient, and θ L2 represents the internal variable parameter of L2 regularization, Θ represents the set of all learnable parameters in the model; auxiliary loss function Indicates that it acts on r (k) The auxiliary classifier results, λ represents the auxiliary loss weight coefficient; the contrast loss function represents the projection representation, sim(·,·) represents the cosine similarity calculation, N represents the batch size during training, K represents the number of distributed receivers, τ represents the contrast loss temperature parameter, and μ represents the contrast loss weight coefficient.

[0063] Finally, the optimizer is used according to L composite Update the model parameters θ, set the hyperparameters, train until convergence, and obtain the optimal model parameters θ * .

[0064] The hyperparameters include: number of training rounds, batch size, learning rate and learning rate decay strategy, and early stopping mechanism based on validation set performance.

[0065] Furthermore, the step S5 is specifically as follows:

[0066] The test set D obtained by preprocessing in step S2 test Input the trained DHCT model obtained in step S4 The model outputs a working mode probability distribution for each test sample Determine the predicted category based on the highest probability The predicted category and the true labels of the test set Make a comparison.

[0067] Calculate the recognition accuracy as a performance evaluation indicator, which is defined as the ratio of the number of correctly classified samples to the total number of samples in the test set.

[0068] Among them, N c Indicates the number of samples of category c that are correctly classified, N test Indicates the total number of samples in the test set.

[0069] Beneficial effects of the present invention: The method of the present invention adopts a distributed hierarchical model architecture, performs receiver-level hierarchical feature encoding through a Conformer encoder, and performs cross-receiver feature fusion through a Transformer encoder to realize feature encoding of multiple radiation source signal receivers, and obtains the aggregated and fused feature representation, and then uses the attention pooling mechanism to aggregate the fused feature representation according to the learned weights to obtain a single feature vector, and finally uses a linear classification layer and a Softmax function to map it to the probability distribution of multiple predefined working modes, and the category with the highest probability is the result of this recognition. The method of the present invention adopts a distributed architecture and utilizes information from multiple radiation source signal receivers, overcoming the shortcomings of a single-station observation with limited viewing angle and susceptibility to occlusion and interference, and can more comprehensively capture the signal characteristics of a multifunctional radiation source. A hierarchical processing mechanism is adopted, in which an independent Conformer encoder first performs robust time-domain feature extraction on the data of each receiver to effectively isolate the noise source, and then a Transformer fusion module performs cross-receiver information fusion, with a clear structure and suitable for distributed processing. An attention pooling mechanism is adopted to adaptively aggregate the fused features, and dynamic weighting can be performed based on the relevance of the features to the final classification task. A composite loss function is used for training, in which the auxiliary loss provides intermediate supervision and promotes the effective learning of the receiver-level feature extractor, while the contrast loss enhances the discriminative power of the features, prompting representations from different viewing angles of the same radiation source to aggregate in the feature space, thereby improving the robustness of the representation and its resistance to non-ideal conditions. This solves the limitations of the existing single-station multifunctional radiation source working mode recognition method in processing signal agility, incomplete observations, complex electromagnetic environment interference, and unknown modes, and improves the accuracy and robustness of recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 This is a flow chart of a method for identifying the working mode of a multifunctional radiation source with a distributed hierarchical architecture according to the present invention.

[0071] Figure 2 Schematic diagram of the geometry of the simulation experiment scene in an embodiment of the present invention.

[0072] Figure 3 Schematic diagram of non-ideal factors affecting the intercepted pulse sequence in an embodiment of the present invention.

[0073] Figure 4 Schematic diagram of the core building blocks of the DHCT model architecture in an embodiment of the present invention.

[0074] Figure 5 This is a comparison chart of the average accuracy of the DHCT model and the comparison model in a mixed non-ideal scenario in an embodiment of the present invention.

[0075] Figure 6 Schematic diagram of the accuracy recognition result under a single non-ideal factor in an embodiment of the present invention. DETAILED DESCRIPTION

[0076] like Figure 1 As shown in FIG, a flow chart of a method for identifying a multifunctional radiation source working mode in a distributed hierarchical architecture of the present invention is shown, and the specific steps are as follows:

[0077] The specific steps are as follows:

[0078] S1, data generation;

[0079] Based on the typical pulse characteristic parameter settings of the multifunctional radiation source, multiple radiation source signal receivers are constructed to respectively intercept the pulse description word (PDW) sequence data set describing the activity of the same multifunctional radiation source within a certain time window. The data sets are divided into training set, validation set, and test set according to the set ratio, and the data of the training set, validation set, and test set are enhanced by simulating site loss and non-ideal factors.

[0080] S2, data preprocessing;

[0081] Based on step S1, a pulse descriptor word (PDW) sequence dataset is obtained. The pulse parameters in each sequence are normalized at the feature level, and then the sequence is padded to achieve a unified maximum sequence length. At the same time, the corresponding attention mask is generated. Finally, the sequences processed by all radiation source signal receivers are transposed and stacked to form the input tensor of the model.

[0082] S3, distributed hierarchical model architecture construction;

[0083] The distributed hierarchical Conformer-Transformer (DHCT) model includes an input embedding layer, a Conformer encoder module, a Transformer encoder module, and an output layer, and is deployed in a system including multiple spatially distributed radiation source signal receivers.

[0084] The distributed hierarchical model first performs receiver-level hierarchical feature encoding through the Conformer encoder and cross-receiver feature fusion through the Transformer encoder to achieve feature encoding of multiple radiation source signal receivers and obtain the aggregated and fused feature representation;

[0085] Then, the attention pooling mechanism is used to aggregate the fused feature representations according to the learned weights to obtain a single feature vector; finally, a linear classification layer and a Softmax function are used to map it to the probability distribution of multiple predefined working modes, and the category with the highest probability is the result of this recognition.

[0086] S4, distributed hierarchical model training;

[0087] The preprocessed training set and validation set obtained in step S2 are input into the distributed hierarchical model architecture in step S3, and the model is trained end-to-end using a composite loss function adapted to the distributed architecture to obtain a model that has converged.

[0088] S5, distributed hierarchical model testing;

[0089] The preprocessed test set obtained in step S2 is input into the trained converged model obtained in step S4 for testing, and the test results are output and the accuracy of model recognition is evaluated.

[0090] In this embodiment, step S1 is specifically as follows:

[0091] First, a simulation scenario is set up that includes a mobile multifunctional radiation source transmitter and K mobile distributed radiation source signal receivers, and the basic parameters of the transmitter and receiver are set. The simulation scenario of the red and blue radiation sources and receivers in this embodiment is as follows: Figure 2 The figure shows the positions and trajectories of a multifunctional radiation source transmitter (red dot) and six distributed radiation source signal receivers (blue dot). It includes a mobile multifunctional radiation source transmitter and K = 6 mobile distributed radiation source signal receivers, showing their relative positions and movement trajectories.

[0092] The basic parameters described in this embodiment include: a radiation source transmission power of 20 kW, a radiation source transmission antenna gain of 30 dB, a receiver sensitivity of -80 dBm, a receiver system loss of 10 dB, a radiation source transmission elevation angle range of -25° to 25° and an azimuth angle range of -60° to 60°, a receiver gain of 3 dB, a radiation source transmitter speed of 500 m / s, a radiation source signal receiver speed of 400 m / s, and position information. The radiation source is at the origin of the three-dimensional coordinate axis, and the position of the signal receiver is shown in Table 1.

[0093] Table 1

[0094]

[0095] Among them, R m 、 θ m They respectively represent the distance, azimuth and elevation angle of the signal receiver relative to the radiation source.

[0096] Then, the C = 6 multifunctional radiation source operating modes to be identified are defined, and the pulse parameter variation range, modulation type, typical pulse number (PN), and PRI group number (GN) are set for each mode, as shown in Table 2. Based on Table 2, this embodiment generates 2400 ideal multifunctional radiation source signal sample sets and data labels with a length of 200. The number of samples for each operating mode remains consistent at 400. While ensuring the proportion of each operating mode, the sets are divided into training, validation, and test sets according to a preset ratio. In this embodiment, the preset ratio is 8:1:1.

[0097] Table 2

[0098]

[0099] Each sample in the sample set represents an operating mode instance, and its parameters are randomly generated within a preset range. The pulse parameters in each mode include: pulse repetition interval PRI, pulse width PW, and carrier frequency RF. U(·) represents the average value, (·) α Indicates that when the modulation type is agile, the agile value is used instead of the average value.

[0100] We then simulated site loss (simulating receiver failures that can occur in real applications) and non-ideal factors to augment the training, validation, and test sets. For each ideal example, we generated replicas with different receiver combinations (ranging from 0 to 5 failed receivers), with the data for failed receivers set to zero. We also retained replicas with all receivers functioning properly, expanding the total sample set by a factor of 69. Finally, we simulated non-ideal factors during signal transmission and reception to obtain the final training, validation, and test sets, along with their corresponding data labels.

[0101] Among them, the simulation results of non-ideal factors are as follows Figure 3 As shown, the non-ideal factors include: missed pulses, false pulses, measurement errors, and mixed scenarios. Figure 3 (a) is an ideal pulse, Figure 3 (b) is the leakage pulse, Figure 3 (c) is a false pulse, Figure 3 (d) is the measurement error, Figure 3 (e) is a mixed scenario. The settings of various non-ideal factors are shown in Table 3.

[0102] Table 3

[0103]

[0104] The measurement error is added by adding additive white Gaussian noise and multiplicative noise, σ bias,PRI , σ bias,PW , σ bias,RFrepresents the standard deviation of additive noise, σ rel,PRI , σ rel,PW , σ rel,RF Represents the relative proportion of multiplicative noise; Missing pulses are simulated by randomly removing pulses from the active receiver data according to a specified missing pulse rate; False pulses are randomly added according to a specified false pulse rate; * indicates that each mixing scenario level means that the measurement error, missing pulse rate, and false pulse rate are all set to the same level.

[0105] This embodiment addresses non-ideal factors:

[0106] Training set D train Expanded 5-fold, each original ideal sample is retained, and four additional samples are generated by independently applying a random non-ideal factor type of level 1 strength to each active receiver, resulting in a final total of 662,400 training samples.

[0107] Validation set D val Expanded 9x to cover nine different conditions for all active receivers: ideal conditions, plus four non-ideal conditions tested at intensities 3 and 6, resulting in 149,040 validation samples.

[0108] Test set D test The 45-fold expansion covers ideal conditions, 12 intensity levels of each of three different non-ideal factors, and 8 levels of mixed factors. Each of these 45 conditions is uniformly applied to each sample, forming a different test subset, which ultimately includes 745,200 test samples.

[0109] In this embodiment, step S2 is specifically as follows:

[0110] S21, feature normalization;

[0111] Set the sequence of receiver k in the data set obtained in step S1 to P (k) , for the i-th pulse in the sequence The three parameters, PRI, PW, and RF, are normalized to obtain the normalized results. Finally, the normalized sequence is obtained

[0112] Wherein, k∈K. In this embodiment, Max-Min global normalization is adopted.

[0113] S22, sequence filling;

[0114] Determine the maximum length of all K receiver sequences in the current batch or the entire dataset Each normalized sequence Pad with zero vector to maximum length L max, update the normalized sequence after filling

[0115] Among them, L k represents the length of the receiver k sequence; p pad Indicates the filling amount, represents the normalized pulse parameters, Indicates size is M f ×L max The vector space of M f =3 indicates the number of features.

[0116] S23, attention mask generation;

[0117] Based on step S22, the padded normalized sequence for each receiver is Generate a corresponding binary attention mask where the original spike positions are 1 and the padded positions are 0. Combine the masks of the K receivers into

[0118] S24, distributed tensor construction;

[0119] The K padded normalized sequences Transpose and stack to form the input feature tensor

[0120] in,(·) T represents transpose, Indicates size is K×L max ×M f The vector space of .

[0121] S25, construct model input;

[0122] The input feature tensor P and the attention mask M form the model input (P, M).

[0123] In this embodiment, for the data set generated in step S1, 662,400 preprocessed training set samples, 149,040 preprocessed validation samples, and 745,200 preprocessed test samples are constructed according to steps S21 to S25.

[0124] In this embodiment, step S3 is specifically as follows:

[0125] S31, input embedding layer;

[0126] M in the input tensor P f The d-dimensional impulse vector is projected onto the d model Dimension, and d model The positional encoding of the dimension is added element by element to obtain an embedded representation including temporal information

[0127] in, Indicates size is K×L max ×d model The vector space of .

[0128] S32, receiver-level hierarchical feature encoding;

[0129] At each receiver sequence X (k) Add [CLS] identifier before embedding Get sequence

[0130] Among them, concat(·) represents the connection operation, which splices multiple tensors or sequences together along the specified dimension to form a larger tensor or sequence. dim=0 means that the dimension of the connection operation starts from 0. Indicates size (L max +1)×d model The vector space of .

[0131] Then use the N enc An independent Conformer encoder module f of each Conformer block conf_enc Processing Sequence and the corresponding mask Output

[0132] Among them, h k,CLS represents the hidden state corresponding to [CLS], Indicates that the kth receiver is from 1 to the maximum length L max The hidden state of all pulses.

[0133] Finally, take the hidden state h corresponding to [CLS] k,CLS As the characteristic of the receiver k And aggregate the features of multiple receivers to obtain the aggregate feature matrix

[0134] S33, cross-receiver feature fusion;

[0135] Each receiver feature r (k) with a learnable receiver embedding Add element by element to get the modified aggregate feature matrix

[0136] Then use a fus The Transformer encoder module f is a stack of standard Transformer blocks fus Processing the Aggregate Feature Matrix Get the fusion feature matrix

[0137] Among them, f (1) ,...,f (K) represents the fused features from receiver 1 to receiver K.

[0138] The multi-head self-attention mechanism in each Transformer block is computed between the K receiver representations, so that the representation of each receiver can pay attention to and fuse information from other receivers.

[0139] S34, output layer;

[0140] The attention pooling mechanism is used to aggregate the fusion feature matrix F. First, each fusion feature u is calculated through a linear transformation with tanh activation k =tanh(W att f (k) +b att ); Then, by calculating the context vector Get importance score

[0141] Among them, u k 、u c 、e k Represent fusion features, context vectors, and importance scores respectively, and represents the learnable parameters in the linear transformation, and Indicates size d att ×d model and The vector space of Represents the size of the learnable parameters.

[0142] Apply the Softmax activation function to all K importance scores to obtain the normalized attention weights Weighted aggregation of F is used to obtain the final aggregated feature vector

[0143] Finally, the aggregate feature vector z is input into a weighted Bias is b c ∈R C The linear classification layer and Softmax activation function output the probability distribution of C predefined working modes The one with the highest probability is the final recognition result.

[0144] In this embodiment, the distributed hierarchical architecture model is as follows: Figure 1 As shown in step S3 of the DHCT model, the core building blocks of the DHCT model architecture are as follows: Figure 4 As shown, Figure 4 (a) is the Conformer block structure, Figure 4 (b) is the Transformer block structure, where x represents the input data of a single block. Other parameters in the model are shown in Table 4.

[0145] Table 4

[0146]

[0147] In this embodiment, step S4 is specifically as follows:

[0148] Use the preprocessed training set D obtained in step S2 train and validation set D val Data, using a composite loss function Conduct training.

[0149] Among them, the main loss function y c and They represent the probability of the true category c being the label and the recognition category c, γ represents the regularization coefficient, and θ L2 represents the internal variable parameter of L2 regularization, Θ represents the set of all learnable parameters in the model; auxiliary loss function Indicates that it acts on r (k) The auxiliary classifier results, λ represents the auxiliary loss weight coefficient; the contrast loss function represents the projection representation, sim(·,·) represents the cosine similarity calculation, N represents the batch size during training, K represents the number of distributed receivers, τ represents the contrast loss temperature parameter, and μ represents the contrast loss weight coefficient.

[0150] This example finally uses the AdamW optimizer according to L composite Update the model parameters θ, set the hyperparameters, train until convergence (the training process continues until the maximum number of rounds is reached or the early stopping condition is met), and obtain the optimal model parameters θ * .

[0151] The model parameters θ include the embedding layer, Conformer encoder, Transformer encoder, attention pooling, classification layer, receiver embedding, and the weight and bias of the projection head. The hyperparameters include the number of training rounds, batch size, learning rate and ReduceLROnPlateau learning rate decay strategy, and early stopping mechanism based on validation set performance, as shown in Table 5.

[0152] Table 5

[0153]

[0154] In this embodiment, step S5 is specifically as follows:

[0155] The test set D obtained by preprocessing in step S2 test Input the trained DHCT model obtained in step S4 The model outputs a working mode probability distribution for each test sample Determine the predicted category based on the highest probability The predicted category and the true labels of the test set Make a comparison.

[0156] Calculate the recognition accuracy as the main performance evaluation indicator, which is defined as the ratio of the number of correctly classified samples to the total number of samples in the test set.

[0157] Among them, N c Indicates the number of samples of category c that are correctly classified, N test Indicates the total number of samples in the test set.

[0158] This example further conducted comparative verification. Referring to the hyperparameter settings and loss functions in Table 5, the comparative model settings adopted are as follows:

[0159] a) Distributed hierarchical Transformer-Transformer model, also known as DHTT (Attention Pooling) or DHTT (Attention);

[0160] b) Distributed hierarchical Conformer-Conformer model, namely DHCC (Attention Pooling) or DHCC (Attention);

[0161] c) Replace the static pooling DHCT model, i.e., DHCT(MeanPooling) or DHCT(Mean) and DHCT(MaxPooling) or DHCT(Max) for comparison.

[0162] Then, through the hyperparameter settings in Table 5, the loss function is set to the cross entropy function, and the Conformer, Transformer, GRU, and LSTM models are introduced for comparison.

[0163] After the training of each model converges, each model is trained on the test set D obtained by step S2 preprocessing. test Processing is performed to obtain the corresponding accuracy results in different scenarios. The recognition accuracy comparison results in mixed scenarios are as follows: Figure 5 As shown in the figure, the accuracy recognition results under a single non-ideal factor are as follows Figure 6 shown.

[0164] Depend on Figure 5 and Figure 6 As can be seen from testing and calculated accuracy results, the DHCT model proposed in this paper demonstrates strong robustness and accuracy against typical single non-ideal factors and in mixed error scenarios in the context of distributed full-station identification. Compared to other fusion models, this method demonstrates higher recognition accuracy and robustness under various non-ideal conditions, particularly high false pulse rates and complex mixed error scenarios.

[0165] In summary, the method of the present invention operates through a two-stage process: first, an independent Conformer encoder processes the pulse sequence data of each radiation source signal receiver to robustly capture complex timing dependencies and local waveform patterns; second, a Transformer module uses a cross-receiver attention mechanism to fuse these individual receiver representations to capture the correlation between receivers and form a global understanding of the state of the multi-functional radiation source, thereby identifying the working mode of the multi-functional radiation source. The attention pooling mechanism contained in this method can adaptively weight the fused features, which is better than the static pooling strategy. At the same time, the auxiliary supervision and comparative supervision used in the training process further enhance the discriminative ability and overall robustness of the features. The method of the present invention aims to solve the challenges of signal agility, incomplete observations and working mode recognition of multi-functional radiation sources in complex electromagnetic environments, especially in the presence of high-density false pulses and mixed working mode scenarios, showing good accuracy and robustness.

[0166] Those skilled in the art will appreciate that the embodiments described herein are intended to aid the reader in understanding the principles of the present invention, and it should be understood that the scope of the present invention is not limited to such specific descriptions and embodiments. Various modifications and variations are readily apparent to those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims.

Claims

1. A distributed hierarchical architecture multifunctional radiation source working mode recognition method, the specific steps are as follows: S1, data generation; Based on the typical pulse characteristic parameter settings of the multifunctional radiation source, multiple radiation source signal receivers are constructed to intercept the pulse description word PDW sequence data set describing the activity of the same multifunctional radiation source within a certain time window. The data set is divided into training set, validation set, and test set according to the set ratio. The training set, validation set, and test set are enhanced by simulating site loss and non-ideal factors. S2, data preprocessing; Based on the pulse description word PDW sequence dataset obtained in step S1, the pulse parameters in each sequence are normalized at the feature level, and then the sequence is padded to achieve a unified maximum sequence length. At the same time, the corresponding attention mask is generated. Finally, the sequences processed by all radiation source signal receivers are transposed and stacked to form the input tensor of the model; S3, distributed hierarchical model architecture construction; The distributed hierarchical model includes: an input embedding layer, a Conformer encoder module, a Transformer encoder module, and an output layer; and is deployed in a system including multiple spatially distributed radiation source signal receivers; The distributed hierarchical model first performs receiver-level hierarchical feature encoding through the Conformer encoder and cross-receiver feature fusion through the Transformer encoder to achieve feature encoding of multiple radiation source signal receivers and obtain the aggregated and fused feature representation; The attention pooling mechanism is then used to aggregate the fused feature representations based on the learned weights to obtain a single feature vector. Finally, a linear classification layer and a Softmax function are used to map the feature representations to the probability distributions of multiple predefined working modes. The category with the highest probability is the recognition result. S4, distributed hierarchical model training; The preprocessed training set and validation set obtained in step S2 are input into the distributed hierarchical model architecture in step S3, and the model is trained end-to-end using a composite loss function adapted to the distributed architecture to obtain a model that has converged. S5, distributed hierarchical model testing; The preprocessed test set obtained in step S2 is input into the trained converged model obtained in step S4 for testing, and the test results are output and the accuracy of model recognition is evaluated.

2. The method for identifying the working mode of a multifunctional radiation source with a distributed hierarchical architecture according to claim 1, characterized in that: The step S1 is specifically as follows: First, a simulation scenario is set up, including a mobile multifunctional radiation source transmitter and K mobile distributed radiation source signal receivers, and the basic parameters of the transmitter and receiver are set; The basic parameters include: radiation source transmission power, radiation source transmission antenna gain, receiver reception sensitivity, receiver system loss, radiation source transmission elevation angle and azimuth angle range, receiver reception gain, movement speed and position information of the radiation source transmitter and the radiation source signal receiver; Then, the C multifunctional radiation source operating modes that need to be identified are defined, and the variation range of pulse parameters, modulation type, typical pulse number PN, and PRI group number GN under each mode are set. An ideal multifunctional radiation source signal sample set and data labels are generated, and then divided into training set, validation set, and test set according to the preset ratio. Each sample in the sample set represents an operating mode instance, and its parameters are randomly generated within a preset range; the pulse parameters in each mode include: PRI, PW, and RF; Then, we simulate site loss and non-ideal factors to perform data augmentation on the training set, validation set, and test set. For each ideal sample, we generate samples under different receiver combinations, and set the data of the failed receiver to zero. At the same time, we retain samples in which all receivers are working properly. Finally, we simulate non-ideal factors in the signal transmission and reception process to obtain the final training set, validation set, test set, and corresponding data labels.

3. The method for identifying the working mode of a multifunctional radiation source with a distributed hierarchical architecture according to claim 1, characterized in that: The step S2 is specifically as follows: S21, feature normalization; Set the sequence of receiver k in the data set obtained in step S1 to P (k) , for the i-th pulse in the sequence The three parameters, PRI, PW, and RF, are normalized to obtain the normalized results. Finally, the normalized sequence is obtained Where k∈K; S22, sequence filling; Determine the maximum length of all K receiver sequences in the current batch or the entire dataset Each normalized sequence Pad with zero vector to maximum length L max , update the normalized sequence after filling Among them, L k represents the length of the receiver k sequence; p pad Indicates the filling amount, represents the normalized pulse parameters, Indicates size is M f ×L max The vector space of M f Indicates the number of features; S23, attention mask generation; Based on step S22, the padded normalized sequence for each receiver is Generate a corresponding binary attention mask with 1 at the original pulse position and 0 at the padded position; combine the masks of the K receivers into S24, distributed tensor construction; The K padded normalized sequences Transpose and stack to form the input feature tensor in,(·) T represents transpose, Indicates size is K×L max ×M f The vector space of S25, construct model input; The input feature tensor P and the attention mask M form the model input (P, M).

4. The method for identifying the working mode of a multifunctional radiation source with a distributed hierarchical architecture according to claim 1, characterized in that: The step S3 is specifically as follows: S31, input embedding layer; M in the input tensor P f The d-dimensional impulse vector is projected onto the d model Dimension, and d model The positional encoding of the dimension is added element by element to obtain an embedded representation including temporal information in, Indicates size is K×L max ×d model The vector space of S32, receiver-level hierarchical feature encoding; At each receiver sequence X (k) Add [CLS] identifier before embedding Get sequence Among them, concat(·) represents the concatenation operation, dim=0 means that the dimension of the concatenation operation starts from 0, Indicates size (L max +1)×d model The vector space of Then use the N enc An independent Conformer encoder module f of each Conformer block conf_enc Processing Sequence and the corresponding mask Output Among them, h k,CLS represents the hidden state corresponding to [CLS], Indicates that the kth receiver is from 1 to the maximum length L max The hidden state of all pulses; Finally, take the hidden state h corresponding to [CLS] k,CLS As the characteristic of the receiver k And aggregate the features of multiple receivers to obtain the aggregate feature matrix S33, cross-receiver feature fusion; Each receiver feature r (k) with a learnable receiver embedding Add element by element to get the modified aggregate feature matrix Then use a fus The Transformer encoder module f is a stack of standard Transformer blocks fus Processing the Aggregate Feature Matrix Get the fusion feature matrix Among them, f (1) ,...,f (K) represents the fused features from receiver 1 to receiver K; S34, output layer; The attention pooling mechanism is used to aggregate the fusion feature matrix F; first, each fusion feature u is calculated through a linear transformation with tanh activation k =tanh(W att f (k) +b att ); Then, by calculating the context vector Get importance score Among them, u k 、u c 、e k Represents fusion features, context vectors, and importance scores, respectively. W att and b att represents the learnable parameters in the linear transformation, and Indicates size d att ×d model and The vector space of Indicates the size of the learnable parameters; Apply the Softmax activation function to all K importance scores to obtain the normalized attention weights Weighted aggregation of F is used to obtain the final aggregated feature vector Finally, the aggregate feature vector z is input into a weighted Bias is b c ∈R C The linear classification layer and Softmax activation function output the probability distribution of C predefined working modes The one with the highest probability is the final recognition result.

5. The method for identifying the working mode of a multifunctional radiation source with a distributed hierarchical architecture according to claim 1, characterized in that: The step S4 is specifically as follows: Use the preprocessed training set D obtained in step S2 train and validation set D val Data, using a composite loss function Conduct training; Among them, the main loss function y c and They represent the probability of the true category c being the label and the recognition category c, γ represents the regularization coefficient, and θ L2 represents the internal variable parameter of L2 regularization, Θ represents the set of all learnable parameters in the model; auxiliary loss function Indicates that it acts on r (k) The auxiliary classifier results, λ represents the auxiliary loss weight coefficient; the contrast loss function represents the projection representation, sim(·,·) represents the cosine similarity calculation, N represents the batch size during training, K represents the number of distributed receivers, τ represents the contrast loss temperature parameter, and μ represents the contrast loss weight coefficient; Finally, the optimizer is used according to L composite Update the model parameters θ, set the hyperparameters, train until convergence, and obtain the optimal model parameters θ * ; The hyperparameters include: number of training rounds, batch size, learning rate and learning rate decay strategy, and early stopping mechanism based on validation set performance.

6. The method for identifying the working mode of a multifunctional radiation source with a distributed hierarchical architecture according to claim 1, characterized in that: The step S5 is specifically as follows: The test set D obtained by step S2 preprocessing test Input the trained DHCT model f obtained in step S4 θ* The model outputs a working mode probability distribution for each test sample Determine the predicted category based on the highest probability The predicted category and the true labels of the test set Make comparisons; Calculate the recognition accuracy as a performance evaluation indicator, which is defined as the ratio of the number of correctly classified samples to the total number of samples in the test set. Among them, N c Indicates the number of samples of category c that are correctly classified, N test Indicates the total number of samples in the test set.

Images