Weak transient signal feature learning and signal reconstruction method based on model guidance

Abstract

The invention belongs to the field of signal processing, and relates to a weak transient signal feature learning and signal reconstruction method based on model guidance, which comprises the following steps: inputting a weak transient signal to be detected into a weak transient signal feature learning network based on model guidance to obtain a reconstructed transient signal; the weak transient signal feature learning network based on model guidance comprises two modules: a model guidance module and a coding-decoding feature learning module, the two modules perform parallel processing on the input weak transient signal, perform unified processing on the processing result through a post-processing convolution block, and output a reconstructed transient signal. According to the method, the transient signal mathematical model with the known waveform is constructed, the training data are generated in batches, and the high cost of data annotation is remarkably reduced; by combining the advantages of a model driving method and deep learning, the detection capability and interpretability of the network are effectively improved. The problem that a weak transient signal is difficult to effectively detect and reconstruct with high precision under a strong noise background is effectively solved.

Description

Technical Field

[0001] The invention belongs to the field of signal processing, and in particular relates to a method for learning weak transient signal features and reconstructing signals based on a model. Background Art

[0002] Transient signals usually refer to events or phenomena with unknown occurrence time and short duration. Effective detection of transient signals is of great significance to medical monitoring, earthquake early warning, radar and infrared detection. In these application scenarios, transient signals often show weak and rapidly changing characteristics due to long propagation distances, high environmental noise or limitations of detection equipment. Weak transient signals are often difficult to detect and separate from background noise due to factors such as low signal energy, short duration and unclear spectral characteristics. Therefore, how to effectively detect weak transient signals in a strong noise background and reconstruct them with high precision is a major challenge in current scientific research and engineering applications.

[0003] Existing weak signal detection methods can be roughly divided into model-driven methods based on traditional signal processing and data-driven deep learning methods. Model-driven methods usually rely on prior assumptions about specific signals, such as extracting signal features based on filtering, envelope analysis, spectral estimation, etc. However, in actual signal processing systems, model-driven methods based on methods such as Fourier transform or wavelet transform can decompose signals to a certain extent, but their ability to capture weak transient signals is limited, especially when the signal is extremely weak or the background noise is large, the signal features are easily lost, resulting in increased reconstruction errors.

[0004] In recent years, data-driven methods based on deep learning have made significant progress in the field of signal processing. Deep neural networks have demonstrated powerful feature extraction capabilities in high-dimensional data processing such as images and speech, and can automatically learn the underlying laws in large-scale data, thereby achieving more accurate signal reconstruction. However, for the processing of weak transient signals, traditional deep learning methods still face significant challenges, mainly reflected in the high cost of data annotation, scarcity of training samples, and lack of interpretability of the decision-making process, making it difficult to directly apply them to actual signal processing systems. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and propose a method for learning and reconstructing weak transient signal features based on model guidance. In the data set preparation stage, by constructing a mathematical model of transient signals and generating training data in batches, the cost of data annotation is significantly reduced. In the process of neural network design and training, by introducing a model-driven method, the neural network is guided by a priori models to learn the characteristics of weak transient signals, which improves the interpretability of the network and the accuracy of signal reconstruction.

[0006] In view of this, the present invention proposes a model-guided weak transient signal feature learning and signal reconstruction method, comprising:

[0007] Inputting the weak transient signal to be detected into the model-guided weak transient signal feature learning network to obtain a reconstructed transient signal;

[0008] The model-guided weak transient signal feature learning network includes two modules: a model-guided module and an encoding-decoding feature learning module. The two modules process the input weak transient signals in parallel, and uniformly process the processing results through a post-processing convolution block to output a reconstructed transient signal.

[0009] Preferably, the model guidance module includes two convolution blocks and four residual blocks, wherein the four residual blocks are connected in series between the two convolution blocks, wherein:

[0010] Each of the convolution blocks includes a one-dimensional convolution layer and a point convolution layer;

[0011] Each of the residual blocks includes two serially connected SBR structures, each of which includes a serially connected one-dimensional separable convolution, batch normalization, and nonlinear activation;

[0012] The processing process of the residual block includes:

[0013] The input features are processed by two SBR structures to obtain intermediate features, which are fused with the input features and then undergo a nonlinear activation to obtain output features.

[0014] Preferably, the model guiding module further includes a signal processing module connected in series after the second convolution block, wherein the signal processing module uses a sliding window to slide the input weak transient signal, and adopts a kernel function method to obtain a set of model features that satisfy the following formula:

[0015]

[0016] Where D represents the model characteristic value of the weak transient signal obtained, and Respectively represent the size of the rear window and the front window, i represents the i-th signal point in the rear window, represents the gray value of the i-th signal point in the rear window, j represents the j-th signal point in the front window, represents the gray value of the jth signal point in the front window, and are the grayscale means of the signals in the rear window and the front window respectively, and e is a natural constant.

[0017] Preferably, the encoding-decoding feature learning module includes: four encoders, four decoders and a connector, wherein:

[0018] The input of the first encoder is the weak transient signal to be detected, and the input of the remaining encoders is the output of the previous encoder; each encoder extracts different features of the weak transient signal to be detected, and gradually extracts multi-level feature representations;

[0019] The connector is used to keep the feature dimension of the fourth encoder output consistent with the input feature dimension of the first decoder;

[0020] The outputs of the first three decoders are fused with the outputs of the encoders at the same level, and the fused features are passed to the next decoder. The output of the fourth decoder is fused with the output of the signal processing module, and the fused features are sent to the post-processing convolution block to obtain the reconstructed transient signal.

[0021] The feature fusion includes feature connection and attention mechanism.

[0022] Preferably, the method includes a training step of a model-guided weak transient signal feature learning network, comprising:

[0023] Step 1) Establish a mathematical model based on a transient signal with a known waveform, generate a transient signal, superimpose noise on the transient signal, generate weak transient signal data with different signal-to-noise ratios, and construct a data set;

[0024] Step 2) Set the loss function and training hyperparameters, input the data set, train the model-guided module in the model-guided weak transient signal feature learning network, and obtain the optimal network parameters, that is, the trained model-guided module. Then, based on the trained model-guided module, train the encoding-decoding feature learning module to obtain the trained encoding-decoding feature learning module, and complete the training of the model-guided weak transient signal feature learning network.

[0025] Preferably, the step 1) comprises:

[0026] Step A1: using the scatter point fitting method to establish a transient signal mathematical model for the transient signal with a known waveform;

[0027] Step A2: for the transient signal mathematical model, different transient signals are generated by changing the amplitude, duration and occurrence position parameters of the signal;

[0028] Step A3: superimposing noise on the generated transient signal to obtain a weak transient signal with noise;

[0029] Step A4: setting a signal-to-noise ratio threshold, retaining weak transient signals that meet the signal-to-noise ratio threshold, and retaining transient signals without noise as labels;

[0030] Step A5: The generated weak transient signals and their labels are constructed into a dataset and further divided into a training set, a validation set, and a test set.

[0031] Preferably, the waveform of step A1 includes pulse wave, sine wave, rectangular wave and shock wave; the noise type of step A3 includes Gaussian white noise, colored noise and Laplace noise.

[0032] Preferably, the step 2) comprises:

[0033] Step C1: For the model hyperparameter prediction network in the model guidance module, the L1 loss function is used;

[0034] Step C2: Use the Adam optimizer to update the model hyperparameter prediction network parameters, set the learning rate lr to 1e-4, set the number of samples for one training batchsize to 32, initialize the training parameters, load the training set and validation set, create the model hyperparameter prediction network structure, perform training, and obtain the optimal parameter model of the model hyperparameter prediction network, that is, the trained model guidance module;

[0035] Step C3: Guide the module based on the trained model, using the weighted loss of regression loss and classification loss:

[0036]

[0037] in, is the total loss function of the encoding-decoding feature learning module, is the weighting factor, For regression loss, a weighted mean square error loss with balanced positive and negative samples is used. For classification loss, binary cross entropy loss is used;

[0038] Step C4: Use the Adam optimizer to update the network parameters and weighting factors of the encoding-decoding feature learning module Set it to 0.8, the learning rate lr is set to 1e-4, and the number of samples trained at a time batchsize is 32. Training is performed to obtain the optimal parameter model of the encoding-decoding feature learning module, that is, the trained encoding-decoding feature learning module, and the training of the model-guided weak transient signal feature learning network is completed.

[0039] Preferably,

[0040] The regression loss Satisfy the following formula:

[0041]

[0042] To use the mean square error loss weighted by the difficulty of regression, the following formula is satisfied:

[0043]

[0044] in, is the transient signal mask matrix, The area where the medium transient signal is located is 1, and the background area is 0; T represents the duration of the transient signal, and L represents the length of the input weak transient signal; is the transient signal label, is the transient signal output by the encoding-decoding feature learning module, Represents the regression difficulty weighting coefficient.

[0045] Compared with the prior art, the advantages of the present invention are:

[0046] The present invention provides a method for learning and reconstructing weak transient signal features based on model guidance. By constructing a mathematical model of transient signals with known waveforms, training data is generated in batches, thereby significantly reducing the high cost of data annotation. At the same time, a model-driven method is introduced when designing a neural network to guide the neural network to learn signal features. By combining the advantages of the model-driven method and deep learning, the detection ability and interpretability of the network are effectively improved. Therefore, the present invention can effectively solve the problem that weak transient signals are difficult to effectively detect and reconstruct with high precision in a strong noise background. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the network structure of weak transient signal feature learning based on model guidance in the method of the present invention;

[0048] Figure 2 It is a schematic diagram of the structure of the encoder;

[0049] Figure 3 It is a schematic diagram of the structure of the decoder;

[0050] Figure 4 It is a structural diagram of the residual block. DETAILED DESCRIPTION

[0051] In order to solve the problem of weak transient signal detection and reconstruction, the present invention proposes a model-guided weak transient signal feature learning and signal reconstruction method, comprising the following steps:

[0052] Step A: Establish a mathematical model for transient signals with known waveforms, and use the mathematical model to generate transient signals with different amplitudes, durations, and locations. Superimpose noise on the generated transient signals to generate weak transient signal data with different signal-to-noise ratios and construct a data set.

[0053] Step B: Construct a weak transient signal feature learning network based on model guidance. The network contains two modules: encoding-decoding feature learning module and model guidance module. The encoding-decoding feature learning module can be divided into two stages: feature encoding stage and feature decoding stage, and a connector is set between the two stages for connection. The module guidance module can be divided into two stages:

[0054] Model hyperparameter prediction stage and model-driven signal processing stage. In the process of weak transient signal feature learning and signal reconstruction, the two modules process the input weak transient signals in parallel, and uniformly process the processing results through a post-processing convolution block, and output them as reconstructed transient signals.

[0055] Step C: Set the loss function and training hyperparameters for the encoding-decoding feature learning module and the model guidance module respectively, and use the weak transient signal data set generated in step A to train the two modules to obtain the optimal network parameter model.

[0056] Step D: Load the network parameter model, input the signal to be detected into the weak transient signal feature learning network guided by the model, and obtain the reconstructed transient signal.

[0057] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0058] Example

[0059] An embodiment of the present invention proposes a method for weak transient signal feature learning and signal reconstruction based on model guidance, which consists of four parts: first, a weak transient signal data set is generated; then a weak transient signal feature learning network based on model guidance is constructed; then a loss function and training hyperparameters are set, and the network is trained using the data set to obtain an optimal model; finally, the optimal model is loaded into the network, and the signal to be detected is input into the weak transient signal feature learning network based on model guidance to obtain a reconstructed transient signal.

[0060] Step 1: Create a dataset

[0061] Step A: Establish a mathematical model for the transient signal with known waveform, and generate transient signals with different amplitudes, durations, and locations through the mathematical model. Superimpose noise on the generated transient signal to generate weak transient signal data with different signal-to-noise ratios and construct a data set. The specific process includes:

[0062] Step A1: Use the scatter point fitting method to establish a mathematical model for a transient signal with a known waveform, including but not limited to a pulse wave, a sine wave, a rectangular wave, a shock wave, etc.

[0063] Step A2: According to the transient signal mathematical model, different transient signals are generated by changing the signal's amplitude, duration, and location parameters.

[0064] Step A3: superimposing noise on the generated transient signal to obtain a weak transient signal with noise, where the noise type includes but is not limited to Gaussian white noise, colored noise, Laplace noise, etc.

[0065] Step A4: According to the actual weak transient signal detection requirements, specify the signal-to-noise ratio (SNR) of the transient signal required in the data set, retain the weak transient signals that meet the SNR, and retain the transient signals without noise as labels. The SNR calculation is obtained by the ratio of the target signal power to the noise signal power.

[0066] Step A5: The generated weak transient signals and their labels are constructed into a dataset, which is further divided into a training set, a validation set, and a test set.

[0067] Step 2: Construct a model-guided weak transient signal feature learning network

[0068] Step B: Construct a weak transient signal feature learning network based on model guidance. The network contains two modules: the encoding-decoding feature learning module and the model guidance module. The encoding-decoding feature learning module can be divided into two stages: the feature encoding stage and the feature decoding stage, and a connector is set between the two stages for connection. The module guidance module can be divided into two stages: the model hyperparameter prediction stage and the model-driven signal processing stage. In the process of weak transient signal feature learning and signal reconstruction, the two modules process the input weak transient signal in parallel, and uniformly process the processing results through a post-processing convolution block, and output them as reconstructed transient signals. The network structure diagram is shown in Figure 1 The specific process includes:

[0069] Step B1: The model-guided weak transient signal feature learning network structure consists of two modules, namely, a model-guided module and an encoding-decoding feature learning module.

[0070] Step B2: In the model guidance module, the input signal first goes through the model hyperparameter prediction stage to obtain the model hyperparameters required by the signal processing module. Figure 1 As shown in Figure 1, it contains two convolutional blocks and four residual blocks. The convolutional block consists of a one-dimensional convolutional layer and a point convolutional layer. The structure of the residual block is shown in Figure 1. Figure 4As shown in the figure, it includes two "one-dimensional separable convolution-batch normalization-non-linear activation" (SeparableConv1d-BatchNorm1d-ReLU, SBR) structures. After the input features (Input) are processed by these two structures to obtain intermediate features, the intermediate features are fused with the input features, and then a nonlinear activation is performed to obtain the output features (Output). The processing process is shown in formula (1).

[0071]

[0072] Step B3: After obtaining the model hyperparameters, in the signal processing module, the model uses the model hyperparameters to process the input signal to obtain the model features. The model-driven processing method used in the signal processing module includes but is not limited to the kernel function method in machine learning, such as formula (2).

[0073]

[0074] in, and Indicates the size of the back window and the front window, represents the input weak transient signal, i represents the i-th signal point in the rear window, Represents the gray value of the i-th signal point in the rear window, represents the gray value of the jth signal point in the front window, and is the mean of the back window and the front window, D represents the model feature value of the weak transient signal, and e is a natural constant. Using the sliding window to slide the input weak transient signal will obtain a set of model features. The model hyperparameters predicted in the model hyperparameter prediction stage are and , which is the window size.

[0075] Step B4: The encoding-decoding feature learning module consists of four encoders, four decoders and a connector, such as Figure 1 As shown in the figure. The input of the first encoder is a weak transient signal, and the input of the remaining encoders is the output of the previous encoder. The input weak transient signal passes through multiple levels of encoders (from top to bottom), each encoder extracts different features, and gradually extracts multi-level feature representations. The connector is used to keep the feature dimension of the fourth encoder output consistent with the input feature dimension of the first decoder. The outputs of the first three decoders will be fused with the outputs of the encoders at the same level. The feature fusion steps include feature connection and attention mechanism. The fused features will be passed to the next decoder. The output of the last decoder will be fused with the output of the signal processing module, and then the fused features will be sent to the post-processing convolution block to obtain the reconstructed transient signal.

[0076] Step B5: The structures of the encoder and decoder are as follows: Figure 2 and Figure 3 As shown. The encoder consists of three "one-dimensional deformable convolution-batch normalization-non-linear activation" (DefC1d-BatchNorm1d-ReLU, DBR) structures. The intermediate features obtained after the input features (Input) pass through two DBR structures will be fused with the input features, and then input into the last DBR structure to obtain the output features (Output), as shown in formula (3). The decoder consists of two DBR structures and a "one-dimensional upsampling convolution-batch normalization-non-linear activation" structure, which is used to restore transient signal features from bottom to top.

[0077]

[0078] Step 3: Network training

[0079] Step C: Set the loss function and training hyperparameters for the encoding-decoding feature learning module and the model guidance module respectively, and use the weak transient signal data set generated in step A to train the two modules to obtain the optimal network parameter model. The specific process includes:

[0080] Step C1: For the model hyperparameter prediction network in the model guidance module, the L1 loss function is used, as shown in formula (4).

[0081]

[0082] Among them, W m is the mth transient signal of the input, is the predicted value of the model hyperparameter prediction network for the mth transient signal, and N is the number of weak transient signals.

[0083] Step C2: Use the Adam optimizer to update the model hyperparameter prediction network parameters, set the learning rate lr to 1e-4, and the batchsize to 32. Initialize the training parameters, load the training set and validation set, create the model hyperparameter prediction network structure, train the network, and train to obtain the optimal parameter model of the model hyperparameter prediction network.

[0084] Step C3: Before training the encoding-decoding feature learning module, first load the optimal parameter model of the model hyperparameter prediction network, and then train the encoding-decoding feature learning module.

[0085] Step C4: The loss function used by the encoding-decoding feature learning module is the weighted loss of regression loss and classification loss, as shown in formula (5).

[0086]

[0087] in is the total loss function of the encoding-decoding feature learning module, is the weighting factor, is the regression loss, is the classification loss. The regression loss uses the weighted mean square error loss with balanced positive and negative samples, as shown in formulas (6) and (7). The classification loss uses the binary cross entropy loss.

[0088]

[0089]

[0090] in, is the transient signal mask matrix, The area where the medium transient signal is located is 1, and the background area is 0; T represents the duration of the transient signal, and L represents the length of the input weak transient signal; is the transient signal label, is the transient signal output by the encoding-decoding feature learning module, Represents the regression difficulty weighting coefficient.

[0091] Step C5: Use the Adam optimizer to update the network parameters and weighting factors of the encoding-decoding feature learning module Set to 0.8, the learning rate lr is set to 1e-4, and the batchsize is 32. Initialize the training parameters, load the training set and the validation set, create the encoding-decoding feature learning module network structure, train the network, and train to obtain the optimal parameter model of the encoding-decoding feature learning module.

[0092] Step 4: Reconstruction of weak transient signals

[0093] Step D: Load the network parameter model, input the signal to be detected into the weak transient signal feature learning network guided by the model, and obtain the reconstructed transient signal. The specific process includes:

[0094] Step D1: Create a model-guided weak transient signal feature learning network, and load the optimal parameter model of the model hyperparameter prediction network and the optimal parameter model of the encoding-decoding feature learning module respectively.

[0095] Step D2: input the weak transient signal data to be detected into the model-guided weak transient signal feature learning network to obtain a reconstructed transient signal.

[0096] Innovation:

[0097] 1 The present invention deeply integrates the model-driven method with the data-driven method in aspects such as data set construction and network design, thereby improving the reconstruction accuracy and reliability of weak transient signals.

[0098] 2 In the hyperparameter setting of the model-driven method, the model hyperparameter prediction network is used to automatically obtain the ideal parameters for model adaptation, which makes up for the lack of robustness caused by manual setting of hyperparameters, making the reconstruction process of weak transient signals without manual intervention and repeated parameter adjustment.

[0099] 3 Encoding-decoding feature learning module and model hyperparameter prediction The network adopts a full convolution structure to adapt to variable-length signal input. At the same time, the one-dimensional deformable convolution and local attention mechanism are introduced to enhance the neural network's ability to extract features of weak transient signals of different scales.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. A model-guided weak transient signal feature learning and signal reconstruction method, comprising: Inputting the weak transient signal to be detected into the model-guided weak transient signal feature learning network to obtain a reconstructed transient signal; The model-guided weak transient signal feature learning network includes two modules: a model-guided module and an encoding-decoding feature learning module. The two modules process the input weak transient signals in parallel, and uniformly process the processing results through a post-processing convolution block to output a reconstructed transient signal.

2. The method for learning and reconstructing weak transient signal features based on model guidance according to claim 1, characterized in that: The model guidance module includes two convolution blocks and four residual blocks, wherein the four residual blocks are connected in series between the two convolution blocks, wherein: Each of the convolution blocks includes a one-dimensional convolution layer and a point convolution layer; Each of the residual blocks includes two serially connected SBR structures, each of which includes a serially connected one-dimensional separable convolution, batch normalization, and nonlinear activation; The processing process of the residual block includes: The input features are processed by two SBR structures to obtain intermediate features, which are fused with the input features and then undergo a nonlinear activation to obtain output features.

3. The method for learning and reconstructing weak transient signal features based on model guidance according to claim 2, characterized in that: The model guiding module also includes a signal processing module connected in series after the second convolution block. The signal processing module uses a sliding window to slide the input weak transient signal and uses a kernel function method to obtain a set of model features that satisfy the following formula: ； Where D represents the model characteristic value of the weak transient signal obtained, and Respectively represent the size of the rear window and the front window, i represents the i-th signal point in the rear window, represents the gray value of the i-th signal point in the rear window, j represents the j-th signal point in the front window, represents the gray value of the jth signal point in the front window, and are the grayscale means of the signals in the rear window and the front window respectively, and e is a natural constant.

4. The method for model-guided weak transient signal feature learning and signal reconstruction according to claim 3, characterized in that: The encoding-decoding feature learning module includes: four encoders, four decoders and a connector, wherein: The input of the first encoder is the weak transient signal to be detected, and the input of the remaining encoders is the output of the previous encoder; each encoder extracts different features of the weak transient signal to be detected, and gradually extracts multi-level feature representations; The connector is used to keep the feature dimension of the fourth encoder output consistent with the input feature dimension of the first decoder; The outputs of the first three decoders are fused with the outputs of the encoders at the same level, and the fused features are passed to the next decoder. The output of the fourth decoder is fused with the output of the signal processing module, and the fused features are sent to the post-processing convolution block to obtain the reconstructed transient signal. The feature fusion includes feature connection and attention mechanism.

5. The method for learning and reconstructing weak transient signal features based on model guidance according to claim 1, characterized in that: The method includes a training step of a model-guided weak transient signal feature learning network, including: Step 1: Establish a mathematical model based on a transient signal with a known waveform, generate a transient signal, superimpose noise on the transient signal, generate weak transient signal data with different signal-to-noise ratios, and construct a data set; Step 2: Set the loss function and training hyperparameters, input the data set, train the model-guided module in the model-guided weak transient signal feature learning network, and obtain the optimal network parameters, that is, the trained model-guided module. Then, based on the trained model-guided module, train the encoding-decoding feature learning module to obtain the trained encoding-decoding feature learning module, and complete the training of the model-guided weak transient signal feature learning network.

6. The method for learning and reconstructing weak transient signal features based on model guidance according to claim 5, characterized in that: The step 1 comprises: Step A1: using the scatter point fitting method to establish a transient signal mathematical model for the transient signal with a known waveform; Step A2: for the transient signal mathematical model, different transient signals are generated by changing the amplitude, duration and occurrence position parameters of the signal; Step A3: superimposing noise on the generated transient signal to obtain a weak transient signal with noise; Step A4: setting a signal-to-noise ratio threshold, retaining weak transient signals that meet the signal-to-noise ratio threshold, and retaining transient signals without noise as labels; Step A5: The generated weak transient signals and their labels are constructed into a dataset and further divided into a training set, a validation set, and a test set.

7. The method for learning and reconstructing weak transient signal features based on model guidance according to claim 6, characterized in that: The waveform of step A1 includes pulse wave, sine wave, rectangular wave and shock wave; the noise type of step A3 includes Gaussian white noise, colored noise and Laplace noise.

8. The method for learning and reconstructing weak transient signal features based on model guidance according to claim 5, characterized in that: The step 2 comprises: Step C1: For the model hyperparameter prediction network in the model guidance module, the L1 loss function is used; Step C2: Use the Adam optimizer to update the model hyperparameter prediction network parameters, set the learning rate lr to 1e-4, set the number of samples for one training batchsize to 32, initialize the training parameters, load the training set and validation set, create the model hyperparameter prediction network structure, perform training, and obtain the optimal parameter model of the model hyperparameter prediction network, that is, the trained model guidance module; Step C3: Guide the module based on the trained model, using the weighted loss of regression loss and classification loss: ； in, is the total loss function of the encoding-decoding feature learning module, is the weighting factor, For regression loss, a weighted mean square error loss with balanced positive and negative samples is used. For classification loss, binary cross entropy loss is used; Step C4: Use the Adam optimizer to update the network parameters and weighting factors of the encoding-decoding feature learning module Set it to 0.8, the learning rate lr is set to 1e-4, and the number of samples trained at a time batchsize is 32. Training is performed to obtain the optimal parameter model of the encoding-decoding feature learning module, that is, the trained encoding-decoding feature learning module, and the training of the model-guided weak transient signal feature learning network is completed.

9. The method for learning and reconstructing weak transient signal features based on model guidance according to claim 8, characterized in that: The regression loss Satisfy the following formula: ； in, To use the mean square error loss weighted by the difficulty of regression, the following formula is satisfied: ； in, is the transient signal mask matrix, The area where the medium transient signal is located is 1, and the background area is 0; T represents the duration of the transient signal, and L represents the length of the input weak transient signal; is the transient signal label, is the transient signal output by the encoding-decoding feature learning module, Represents the regression difficulty weighting coefficient.

Images