Interference identification and suppression method and device based on SAMNet neural network

Abstract

The invention discloses an interference identification and suppression method based on an SAMNet neural network, and the method comprises the steps: carrying out the normalization of an echo signal, and inputting the normalized echo signal into a trained SAMNet identification network and a trained SAMNet suppression network, and obtaining the interference type of the echo signal and an interference suppression signal; wherein the SAMNet recognition network and the SAMNet suppression network are SAMNet neural networks trained by adopting different training samples and training methods; the SAMNet neural network is used for carrying out position embedding coding on the normalized echo signal and then carrying out interference type identification / interference suppression, and the optimization target realized by the SAMNet suppression network based on a loss function during training is to minimize the difference between a predicted interference suppression signal and a real non-interference signal. The method is good in real-time performance and universality.

Description

Technical Field

[0001] This invention belongs to the field of interference identification and suppression technology, specifically relating to an interference identification and suppression method and apparatus based on SAMNet neural network. Background Technology

[0002] In complex electromagnetic environments, multiple types of interference coexist and change rapidly, severely weakening radar's target detection and tracking performance. With the continuous development of deep learning, modern radar is evolving towards intelligence, enabling rapid identification and effective suppression of interference through neural networks. Based on this technical approach, precise identification and effective suppression of interference can be achieved by designing network structures. Currently, the data used for interference identification is mainly images, including time-frequency maps and RD maps. However, this requires converting echo signals into image information and processing them frame by frame, which is time-consuming and difficult to meet the real-time requirements of practical scenarios. Regarding interference suppression, traditional methods generally can only handle single or predefined types of interference. When faced with other types of interference, performance drops significantly, resulting in low reliability in multi-interference scenarios.

[0003] Therefore, current interference identification and suppression methods suffer from low real-time performance, and interference suppression methods, in particular, are characterized by strong specificity but poor versatility. Summary of the Invention

[0004] This invention provides an interference identification and suppression method based on the SAMNet neural network, which can solve the above-mentioned technical problems.

[0005] In a first aspect, embodiments of the present invention provide an interference identification and suppression method based on a SAMNet neural network, the method comprising: After the echo signal is normalized, it is input into the trained SAMNet recognition network and SAMNet suppression network respectively to obtain the interference type and interference suppression signal of the echo signal. The SAMNet recognition network and the SAMNet suppression network are SAMNet neural networks trained with different training samples and training methods. The SAMNet neural network is used to perform interference type recognition / interference suppression after the normalized echo signal is embedded and encoded. The optimization objective of the SAMNet suppression network during training is based on the loss function: to minimize the difference between the predicted interference-suppressed signal and the real interference-free signal.

[0006] Secondly, embodiments of the present invention provide an interference identification and suppression device based on a SAMNet neural network, including an identification module and a suppression module; The recognition module is used to normalize the echo signal and input it into the trained SAMNet recognition network to obtain the interference type of the echo signal; The suppression module is used to normalize the echo signal and input it into the trained SAMNet suppression network to obtain the interference suppression signal; The SAMNet recognition network and the SAMNet suppression network are SAMNet neural networks trained with different training samples and training methods. The SAMNet neural network is used to perform interference type recognition / interference suppression after the normalized echo signal is embedded and encoded. The optimization objective of the SAMNet suppression network during training is based on the loss function: to minimize the difference between the predicted interference-suppressed signal and the real interference-free signal.

[0007] The beneficial effects of this invention compared to the prior art are as follows: This invention uses the SAMNet neural network to identify and suppress interference. Since the processing flow of the SAMNet neural network is based on the signal sequence generated after position encoding of the echo signal, there is no need to convert the echo signal into an image before processing. This saves the time required for signal to image conversion and improves real-time performance. Furthermore, since the SAMNet neural network performing the interference suppression task takes minimizing the difference between the interference-free signal and the predicted interference-suppressed signal as the optimization objective, the network only needs the participation of the interference-free signal and the interference-containing sample echo signal during training, without considering the type and parameters of the interference. This improves the reliability and robustness of the network and can better meet the needs of scenarios with multiple interferences. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the structure of a SAMNet neural network provided in an embodiment of the present invention; Figure 2 A schematic diagram illustrating a scenario where the SAMNet identification network provided by this invention performs interference identification; Figure 3 A schematic diagram of a training scenario for a SAMNet suppression network provided in an embodiment of the present invention; Figure 4 A flowchart illustrating the implementation of an interference identification and suppression method based on a SAMNet neural network, provided in an embodiment of the present invention; Figure 5 A schematic diagram of the structure of an interference identification and suppression device based on a SAMNet neural network provided in an embodiment of the present invention; Figure 6a , 6b This is a schematic diagram illustrating the overall classification and identification accuracy for a single type of interference, provided as an embodiment of the present invention. Figure 7a , 7b This is a schematic diagram of an interference matrix with drying ratios of -18dB and -8dB for a single type of interference, provided in an embodiment of the present invention. Figure 8a , 8b This is a schematic diagram illustrating the overall classification and recognition accuracy against aliasing interference, provided as an embodiment of the present invention. Figure 9a , 9b A schematic diagram of an aliasing interference drying ratio of -14 dB and -2 dB provided for an embodiment of the present invention; Figure 10a , 10b This diagram illustrates the suppression effect of NAM and SMSP type interference provided in an embodiment of the present invention. Detailed Implementation

[0009] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0010] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0011] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0012] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0013] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0014] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0015] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0016] Example 1 Figure 1 The diagram shown illustrates the structure of a SAMNet neural network according to an embodiment of the present invention. As an example and not a limitation, the network may include a position encoding module, an encoder, and a classifier.

[0017] In some embodiments, the position encoding module can divide the normalized echo signal into multiple time segments and add position information to each time segment through position encoding to obtain an encoded signal sequence. The encoder can extract the spatiotemporal feature representation of the echo signal from the encoded signal sequence based on a probabilistic sparse self-attention mechanism. The classifier can then output the interference type based on the spatiotemporal feature representation of the echo signal.

[0018] In one possible implementation, the location embedding module can divide the echo signal into multiple time segments, with each segment serving as a token for the model. Location information is added to each time segment through location encoding, thereby forming a complete time sequence (i.e., the encoded signal sequence).

[0019] For example, the encoded signal sequence can satisfy the following formula:

[0020] in, The encoded signal sequence, It is a matrix composed of time segments. and These represent the number of segments and the number of features within each segment, respectively. This is a location information matrix.

[0021] In one example, the position encoding module can use the Sinusoidal function for position encoding.

[0022] For example, if This represents the actual location of the time segment within the sequence. This is the position vector for that time segment. For the position vector of the first If there are 1 element, then It can be represented as:

[0023] in, For the ()th in the location information matrix t , i The ) element represents the )th element. t The first time segment i Location information, , For the dimension of time segments, .

[0024] In one possible implementation method, see Figure 1 The encoder can be composed of a probabilistic sparse self-attention module, a first residual normalization layer, an extended long short-term memory network, and a second residual normalization layer.

[0025] Specifically, when performing interference identification tasks, the main purpose of the encoder is to extract multi-level, discriminative spatiotemporal feature representations from the echo signals with interference.

[0026] Specifically, when performing the suppression task, the encoder's main task is to filter out interference signals in the echo signal and output the extracted spatiotemporal features as the interference suppression signal.

[0027] For example, the probabilistic sparse self-attention module can perform probabilistic sparse self-attention processing on the encoded signal sequence to obtain self-attention features; the first residual normalization layer can perform residual concatenation between the encoded signal sequence and the self-attention features to obtain a residual vector; the extended long short-term memory network can perform feature extraction and information selection based on the residual vector to obtain context features; the second residual normalization layer can perform residual concatenation and normalization processing on the context features and the residual vector to obtain a spatiotemporal feature representation.

[0028] In one example, compared to traditional self-attention mechanisms, probabilistic sparse self-attention mechanisms employ a specific criterion to identify redundant query vectors. Specifically, a sparsity metric can be calculated for each query vector, and based on the metric results, the top queries with higher metric values ​​are selected. The query vectors form a sparse matrix, and a single-head attention operation is performed based on the sparse matrix.

[0029] For example, the sparsity of a query vector can be measured by the following formula:

[0030] in, Indicates the first i query vectors Sparsity metric, The set of all key-value vectors. For the first A key-value vector, The total number of key-value vectors Dimensions for each attention head.

[0031] For example, when The larger the value, the more it indicates that... The attention distribution is sparse, meaning it contains a lot of information and needs to be calculated precisely; conversely, it contains less information and can be approximated or ignored.

[0032] Specifically, query vectors, key vectors, and value vectors can all be obtained by performing a linear transformation on the encoded signal sequence.

[0033] For example, single-head attention operations can be performed using the following formula:

[0034] in, , , The first The query, key, and value matrix of each attention head; The length of the input feature sequence; For each attention head, there are generally several dimensions. ; For the first Weighted features of each attention head; This is the normalized dot product attention function; For feature splicing operations, This is the activation function used to convert the dot product of the query vector and the key vector into attention weights.

[0035] Therefore, the attention features of the final output can be represented as: ,in , The number of single-head attention; The feature sequence output by multi-head attention (i.e., self-attention features); This is a trainable weight matrix.

[0036] Generally speaking, The values ​​of satisfy:

[0037] in, A constant sampling factor. To query the sequence length.

[0038] Alternatively, zeros can be used to fill the gaps when performing single-head attention operations. This ensures that it maintains the same dimension as the input encoded signal sequence.

[0039] When calculating the attention distribution, the attention weights usually follow a long-tail distribution. Only a few query vectors have attention distributions that are strongly correlated with certain keys, while the attention distributions of most query vectors are close to a uniform distribution and contribute very little to the final result. Therefore, the probabilistic sparse self-attention mechanism uses an approximate KL divergence metric to evaluate the sparsity of each query vector, selects only the u query vectors with the highest sparsity scores, and then only calculates the attention between these selected query vectors and all keys. This can greatly reduce the amount of computation, and at the same time, since it is not necessary to store the complete attention matrix, the space occupied is also greatly reduced.

[0040] In one example, the Extended Long Short-Term Memory (LSTM) network is an improvement on the traditional LSTM network, introducing two new memory structures: Matrix LSTM (mLSTM) and Scalar LSTM (sLSTM). The residual vector is sequentially passed through the first Matrix LSTM network, the Scalar LSTM network, and the second Matrix LSTM network to perform feature extraction and information selection operations.

[0041] For example, mLSTM extends the vector operations in traditional LSTM to matrix operations, enabling it to capture more complex data relationships and patterns in a single time step, greatly enhancing the model's memory and parallel processing capabilities. In addition, it integrates an attention mechanism, allowing it to better focus on global information.

[0042] Specifically, the process of mLSTM performing feature extraction can be represented as follows:

[0043]

[0044]

[0045] At the same time, Projection as , and :

[0046]

[0047]

[0048] in, , , These are the query vector, key vector, and value vector, respectively. , , , For projection weights, , , , For bias terms, The input feature dimension.

[0049] For example, sLSTM retains the scalar memory units of LSTM but introduces normalized states and exponential gating. Normalized states alleviate the vanishing or exploding gradient problem during training, thus ensuring numerical stability over long periods; exponential gating provides more flexible dynamic memory control. Furthermore, sLSTM implements a multi-head parallel processing mechanism through block diagonal linear layers, enabling efficient feature extraction.

[0050] Specifically, the process of feature extraction performed by sLSTM can be represented as follows:

[0051]

[0052]

[0053] in, , , For the input weight vector, , , For loop weights, , , For bias terms, For the input vector, , , These are the input gate, forget gate, and output gate, respectively. It is the sigmoid activation function. for Activation function.

[0054] In one example, see Figure 1The classifier can include cascaded flattening layers, linear layers, and activation layers.

[0055] For example, when performing an interference recognition task, the flattening layer first converts the multidimensional spatiotemporal feature representation into a one-dimensional vector. Subsequently, the linear layer performs an affine transformation to achieve dimensionality reduction. Finally, the probability distribution of classification is obtained through the softmax function after successful activation.

[0056] Specifically, a linear layer can linearly combine one-dimensional spatiotemporal feature representations and then add a bias term to achieve feature dimensionality reduction.

[0057] Specifically, the process of performing an affine transformation in a linear layer can be represented as:

[0058] in, This is the weight matrix. For bias terms, This is the output vector.

[0059] The SAMNet neural network provided by this invention extracts self-attention features through a probabilistic sparse self-attention mechanism, which can greatly reduce the space and time complexity of the processing process and has a greater advantage when processing long sequences. At the same time, the overall processing flow is based on the time series after position encoding, rather than converting the echo signal into an image before processing, which can avoid the conversion from signal to image, further reduce processing time, and improve real-time performance.

[0060] Example 2 Figure 2 The diagram shown illustrates a scenario where the SAMNet identification network provided by this invention performs interference identification.

[0061] As an example, the SAMNet neural network can perform both interference recognition and interference suppression tasks. When performing interference recognition tasks, it is called the SAMNet recognition network; when performing interference suppression tasks, it is called the SAMNet network.

[0062] In one example, when training the SAMNet recognition network, normalized sample echo signals and their true classification labels can be input into the network to obtain the predicted classification of the sample echo signals. Then, multi-class cross-entropy is used as the loss function to calculate the loss of the SAMNet recognition network. The model parameters are then updated in reverse based on the loss value. This training process is repeated until preset conditions are met (e.g., the number of training iterations reaches a threshold, or the classification accuracy meets a standard), resulting in a trained SAMNet recognition network.

[0063] For example, the loss function of the SAMNet recognition network can satisfy the following formula:

[0064] in, and These represent the total number of samples and the total number of categories, respectively. For the sample Category The true label, Predict samples for the model Category The probability of.

[0065] Specifically, the smaller the loss value, the closer the probability distribution predicted by the model is to the true label distribution, and the better the classification performance of the network.

[0066] In one example, see Figure 2 In performing the recognition task, the SAMNet recognition network first performs slicing, embedding, and positional encoding processing through the positional encoding module to obtain the encoded signal sequence. Then, the encoder performs a probabilistic sparse self-attention operation on the signal sequence to obtain a spatiotemporal feature representation. Finally, the spatiotemporal feature representation passes through the flattening layer, linear layer, and activation layer of the classifier to obtain the interference type of the echo signal.

[0067] Example 3 Figure 3 The diagram shown illustrates a training scenario for a SAMNet suppression network provided in an embodiment of the present invention.

[0068] As an example, when training the SAMNet suppression network, both real, interference-free signals and interference-laden sample echo signals can be input into the SAMNet suppression network. The network learns the features of both and obtains the matching result (i.e., the predicted interference-suppressed signal) and the loss function value. The network parameters are then updated in reverse based on the loss value, allowing the network to further learn the difference between the suppression result and the interference-free signal, and continuously feed forward. This process is repeated until a preset condition is met (e.g., the loss function value approaches 0 and tends to stabilize), resulting in a well-trained SAMNet suppression network.

[0069] For example, the loss function of the SAMNet suppression network can satisfy the following formula:

[0070] in, Let SAMNet be the loss function for the suppression network. For the predicted interference suppression signal, The signal is a real, interference-free signal, and N is the total number of samples.

[0071] Specifically, a loss function value close to 0 and tending to stabilize indicates that the difference between the suppression result and the interference-free signal is small, achieving a good interference suppression effect.

[0072] This invention uses minimizing the difference between the interference-suppressed signal and the interference-free signal as the optimization objective of the SAMNet suppression network. This allows the SAMNet network to simply approximate the waveform of the interference-containing signal after receiving it, without considering the type or parameters of the interference. Here, the degree of approximation (i.e., the degree of matching) is measured using the interference suppression loss function.

[0073] Example 4 The interference identification and suppression method based on SAMNet neural network provided in this embodiment of the invention can be applied to electronic devices such as mobile terminals, personal laptops, and supercomputers. This embodiment of the invention does not impose any restrictions on the specific type of electronic device.

[0074] Figure 4 The diagram shown illustrates an implementation flowchart of an interference identification and suppression method based on a SAMNet neural network, provided by an embodiment of the present invention. As an example and not a limitation, the method may include steps S401 and S402, which are described below.

[0075] S401 normalizes the echo signal and inputs it into the trained SAMNet recognition network to obtain the interference type of the echo signal.

[0076] For example, SAMNet identifies the network as the network in Embodiment 2 above.

[0077] In one example, if the radar transmits a linear frequency modulated (LFM) signal, and there are targets and interference signals within the radar's observation range, the echo signal received by the radar can be represented as:

[0078] in, The echo signal contains the target. Interference with the main lobe. It is noise.

[0079] For example, the types of interference signals can include suppression interference and spoofing interference, where suppression interference includes noise amplitude modulation (NAM) and noise frequency modulation (NFM); spoofing interference includes chopping and interleaving (C&I), comb spectrum (CS), interrupted-sampling repeater jamming (ISRJ), and smeared spectrum (SMSP).

[0080] S402 normalizes the echo signal and inputs it into the SAMNet suppression network to obtain the interference suppression signal.

[0081] For example, the SAMNet suppression network can be the network in Embodiment 3 above.

[0082] This invention utilizes the SAMNet neural network to identify and suppress interference. Since the SAMNet neural network's processing flow is based on the signal sequence generated after position encoding of the echo signal, there is no need to convert the echo signal into an image for further processing. This saves the time required for signal-to-image conversion and improves real-time performance. Furthermore, because the SAMNet neural network performing the interference suppression task uses minimizing the difference between the interference-free signal and the predicted interference-suppressed signal as its optimization objective, the network only requires the participation of both interference-free and interference-containing sample echo signals during training, without needing to consider the type or parameters of the interference. This improves the network's reliability and robustness, better meeting the needs of scenarios with multiple interferences coexisting.

[0083] Example 5 Figure 5 The diagram shown illustrates the structure of an interference identification and suppression device based on a SAMNet neural network, provided in an embodiment of the present invention. As an example and not a limitation, the device may include an identification module and a suppression module.

[0084] For example, the recognition module is used to normalize the echo signal and input it into the trained SAMNet recognition network to obtain the interference type of the echo signal; the suppression module is used to normalize the echo signal and input it into the trained SAMNet suppression network to obtain the interference suppression signal; wherein, the SAMNet recognition network and the SAMNet suppression network are SAMNet neural networks trained with different training samples and training methods; the SAMNet neural network is used to perform interference type recognition / interference suppression after performing position embedding encoding on the normalized echo signal, and the optimization objective achieved by the SAMNet suppression network based on the loss function during training is to minimize the difference between the predicted interference suppression signal and the real interference-free signal.

[0085] In one possible implementation, the SAMNet neural network includes a position encoding module, an encoder, and a classifier. The position encoding module is used to divide the normalized echo signal into multiple time segments, and add position information to each time segment through position encoding to obtain the encoded signal sequence. The encoder is used to extract the spatiotemporal feature representation of the echo signal from the encoded signal sequence based on the probabilistic sparse self-attention mechanism; The classifier is used to represent the type of output interference based on spatiotemporal characteristics.

[0086] In one possible implementation, the encoded signal sequence satisfies the following formula:

[0087] in, The encoded signal sequence, It is a matrix composed of time segments. This is a location information matrix; The location information matrix satisfies the following formula:

[0088] in, For the ()th in the location information matrix t , i The ) element represents the )th element. t The first time segment i Location information, , For the dimension of time segments, .

[0089] In one possible implementation, the encoder includes a probabilistic sparse self-attention module, a first residual normalization layer, an extended long short-term memory network, and a second residual normalization layer. The probabilistic sparse self-attention module is used to perform probabilistic sparse self-attention processing on the encoded signal sequence to obtain self-attention features; The first residual normalization layer is used to perform residual concatenation and normalization on the encoded signal sequence and self-attention features to obtain the residual vector; Extended Long Short-Term Memory (LSTM) networks are used for feature extraction and information selection based on residual vectors to obtain contextual features. The second residual normalization layer is used to perform residual concatenation and normalization on the context features and residual vectors to obtain spatiotemporal feature representations.

[0090] In one possible implementation, the loss function of the SAMNet suppression network satisfies the following formula:

[0091] in, Let SAMNet be the loss function for the suppression network. For the predicted interference suppression signal, The signal is a real, interference-free signal, and N is the total number of samples.

[0092] This invention utilizes the SAMNet neural network to identify and suppress interference. Since the SAMNet neural network's processing flow is based on the signal sequence generated after position encoding of the echo signal, there is no need to convert the echo signal into an image for further processing. This saves the time required for signal-to-image conversion and improves real-time performance. Furthermore, because the SAMNet neural network performing the interference suppression task uses minimizing the difference between the interference-free signal and the predicted interference-suppressed signal as its optimization objective, the network only requires the participation of both interference-free and interference-containing sample echo signals during training, without needing to consider the type or parameters of the interference. This improves the network's reliability and robustness, better meeting the needs of scenarios with multiple interferences coexisting.

[0093] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0094] To better illustrate the beneficial effects of the present invention, the following simulation experiments were conducted: Simulation Experiment 1 For example, Experiment 1 tested the overall recognition accuracy of the method provided by the present invention (see...). Figure 6a ) and the recognition accuracy for different single types of interference (see Figure 6b Furthermore, the confusion matrix for interference identification under different drying ratios was tested (see [reference]). Figure 7a and Figure 7bAs can be seen, the method provided by the present invention can achieve an accuracy of over 90% when JNR=-18 dB, an accuracy of 100% when JNR=-8 dB, and an overall accuracy of 96.9% in the range of JNR=-20 dB to 0 dB.

[0095] Simulation Experiment 2 For example, Experiment 2 tested the overall recognition accuracy of the method provided by the present invention (see...). Figure 8a ) and recognition accuracy in different scenarios with two types of aliasing interference (see Figure 8b ), and the confusion matrix for interference identification under different drying ratio conditions (see Figure 9a and Figure 9b As can be seen, when SNR=0 dB, the method provided by this invention can achieve an accuracy of over 95% when JNR=-14 dB, and an accuracy of 100% when JNR is greater than -2 dB.

[0096] Simulation Experiment 3 For example, simulation experiment 3 tested the suppression effect of the present invention on different types of interference (see...). Figure 10a and Figure 10b As can be seen, under the condition of JNR=30 dB, the method provided by this invention exhibits good suppression effects against suppression interference, such as noise amplitude modulation (NAM), and deceptive interference, such as smeared spectrum (SMSP).

[0097] This demonstrates that, compared to other interference identification and suppression methods, this invention can accurately identify both suppressive and deceptive interference using only the time-domain signal of the interference. When suppressing interference, it does not require consideration of the type or parameters of the interference to achieve effective suppression. Compared to existing image-dependent interference identification and single-interference suppression methods, this invention improves the real-time performance of interference identification. Furthermore, the suppression algorithm proposed in this invention achieves end-to-end processing and simultaneous suppression of different types of interference, ensuring stable and reliable interference suppression even in complex and variable interference environments, thus improving the robustness of interference suppression.

Claims

1. A method for interference identification and suppression based on SAMNet neural network, characterized in that, include: After the echo signal is normalized, it is input into the trained SAMNet recognition network and SAMNet suppression network respectively to obtain the interference type and interference suppression signal of the echo signal. The SAMNet recognition network and the SAMNet suppression network are SAMNet neural networks trained with different training samples and training methods. The SAMNet neural network is used to perform interference type recognition / interference suppression after the normalized echo signal is embedded and encoded. The optimization objective of the SAMNet suppression network during training is based on the loss function: to minimize the difference between the predicted interference-suppressed signal and the real interference-free signal.

2. The method according to claim 1, characterized in that, The SAMNet neural network includes a position encoding module, an encoder, and a classifier; The position encoding module is used to divide the normalized echo signal into multiple time segments, and add position information to each time segment through position encoding to obtain the encoded signal sequence. The encoder is used to extract the spatiotemporal feature representation of the echo signal from the encoded signal sequence based on a probabilistic sparse self-attention mechanism; The classifier is used to output the interference type based on the spatiotemporal features.

3. The method according to claim 2, characterized in that, The encoded signal sequence satisfies the following formula: in, The encoded signal sequence, The matrix is ​​composed of the time segments. This is a location information matrix; The location information matrix satisfies the following formula: in, For the ()th position in the location information matrix t , i The )th element represents the )th t The first time segment i Location information, , The dimension of the time segment. .

4. The method according to claim 2, characterized in that, The encoder includes a probabilistic sparse self-attention module, a first residual normalization layer, an extended long short-term memory network, and a second residual normalization layer. The probabilistic sparse self-attention module is used to perform probabilistic sparse self-attention processing on the encoded signal sequence to obtain self-attention features. The first residual normalization layer is used to perform residual concatenation and normalization on the encoded signal sequence and the self-attention feature to obtain a residual vector; The extended long short-term memory network is used to extract features and select information based on the residual vector to obtain context features; The second residual normalization layer is used to perform residual concatenation and normalization processing on the context features and the residual vector to obtain the spatiotemporal feature representation.

5. The method according to claim 1, characterized in that, The loss function of the SAMNet suppression network satisfies the following formula: in, Let SAMNet be the loss function of the suppression network. The predicted interference suppression signal, This refers to the actual, interference-free signal. N The total number of samples.

6. An interference identification and suppression device based on a SAMNet neural network, characterized in that, Includes an identification module and a suppression module; The recognition module is used to normalize the echo signal and input it into the trained SAMNet recognition network to obtain the interference type of the echo signal; The suppression module is used to normalize the echo signal and input it into the trained SAMNet suppression network to obtain the interference suppression signal; The SAMNet recognition network and the SAMNet suppression network are SAMNet neural networks trained with different training samples and training methods. The SAMNet neural network is used to perform interference type recognition / interference suppression after the normalized echo signal is embedded and encoded. The optimization objective of the SAMNet suppression network during training is based on the loss function: to minimize the difference between the predicted interference-suppressed signal and the real interference-free signal.

7. The apparatus according to claim 6, characterized in that, The SAMNet neural network includes a position encoding module, an encoder, and a classifier; The position encoding module is used to divide the normalized echo signal into multiple time segments, and add position information to each time segment through position encoding to obtain the encoded signal sequence. The encoder is used to extract the spatiotemporal feature representation of the echo signal from the encoded signal sequence based on a probabilistic sparse self-attention mechanism; The classifier is used to output the interference type based on the spatiotemporal features.

8. The apparatus according to claim 7, characterized in that, The encoded signal sequence satisfies the following formula: in, The encoded signal sequence, The matrix is ​​composed of the time segments. This is a location information matrix; The location information matrix satisfies the following formula: in, For the ()th position in the location information matrix t , i The )th element represents the )th t The first time segment i Location information, , The dimension of the time segment. .

9. The apparatus according to claim 7, characterized in that, The encoder includes a probabilistic sparse self-attention module, a first residual normalization layer, an extended long short-term memory network, and a second residual normalization layer. The probabilistic sparse self-attention module is used to perform probabilistic sparse self-attention processing on the encoded signal sequence to obtain self-attention features. The first residual normalization layer is used to perform residual concatenation and normalization on the encoded signal sequence and the self-attention feature to obtain a residual vector; The extended long short-term memory network is used to extract features and select information based on the residual vector to obtain context features; The second residual normalization layer is used to perform residual concatenation and normalization processing on the context features and the residual vector to obtain the spatiotemporal feature representation.

10. The apparatus according to claim 6, characterized in that, The loss function of the SAMNet suppression network satisfies the following formula: in, Let SAMNet be the loss function of the suppression network. The predicted interference suppression signal, The true, interference-free signal is represented by N, where N is the total number of samples.

Images