Method for modeling behavior of underwater acoustic wideband power amplifier based on LSTM and transfer learning
By using LSTM and transfer learning, a power amplifier model applicable to various underwater acoustic signals is constructed, solving the problem of conventional power amplifier models' dependence on specific signal types and achieving efficient adaptability and low-cost modeling for different signals.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTHEAST UNIV
- Filing Date
- 2023-06-21
- Publication Date
- 2026-07-24
AI Technical Summary
Conventional power amplifier models are limited by the signal dataset used in modeling, and are usually only suitable for specific types of signals, resulting in low efficiency and high cost under different types of signals.
By employing an LSTM-based and transfer learning approach, a power amplifier model applicable to various underwater acoustic signals is established by constructing a basic LSTM model for transfer learning and fixing some LSTM layer parameters during training.
It achieves adaptability to different types of broadband underwater acoustic signals, reduces the model's dependence on signal type, and improves the versatility and efficiency of the power amplifier model.
Smart Images
Figure CN116720470B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power amplifier behavior modeling technology, specifically relating to a behavior modeling method for underwater acoustic broadband power amplifiers based on LSTM and transfer learning. Background Technology
[0002] The power amplifier is a crucial component of an active sonar transmitter, amplifying the modulated signal to the required power level. Identifying the sonar transmitter hinges on recognizing its internal power amplifier; therefore, establishing a power amplifier model that accurately reflects the amplifier's nature is extremely important. Power amplifier behavior modeling aims to describe the system's response characteristics using the amplifier's input and output data in a black-box manner.
[0003] In current research on power amplifier behavior models, traditional modeling methods based on polynomials and other functions require a trade-off between complexity and performance. However, power amplifier models based on neural networks, with their powerful learning and nonlinear fitting capabilities, are a powerful tool for building power amplifier models. Furthermore, since underwater acoustic signals are often broadband, and broadband power amplifiers exhibit strong dynamic characteristics due to their nonlinearity, a nonlinear model with memory effects is necessary. Therefore, applying Long Short-Term Memory (LSTM) neural networks to power amplifier modeling has significant advantages. The neuron structure within the LSTM network corresponds well to the power amplifier's behavioral characteristics, exhibiting good temporal continuity in the network input feature vectors and incorporating temporal dependencies within the network processing units. However, due to the different data distributions of different signal types, which do not meet the same distribution condition, models obtained using specific signal types only learn the power amplifier's behavioral characteristics under that specific signal type, rather than the complete power amplifier characteristics. Therefore, conventional power amplifier behavior models are typically only suitable for the signal type consistent with the modeling process and cannot be effectively applied to other signal types. If conventional methods are used to model sonar transmitter power amplifiers, it is necessary to model all commonly used sonar signals from scratch, constructing separate models for each of these different signals. This would significantly reduce efficiency and be too costly. To establish a power amplifier model that is adaptable to commonly used underwater acoustic signals, this invention proposes a method for modeling the behavior of broadband underwater acoustic power amplifiers based on LSTM and transfer learning. Summary of the Invention
[0004] The purpose of this invention is to address the problem that conventional power amplifier models are limited by the signal dataset during modeling, and the resulting models are usually only suitable for specific types of signals. To improve the adaptability of power amplifier models to different types of broadband underwater acoustic signals, this invention provides a behavior modeling method for underwater acoustic broadband power amplifiers based on LSTM and transfer learning.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for modeling the behavior of underwater acoustic broadband power amplifiers based on LSTM and transfer learning, comprising:
[0006] Step 1: Acquire the input and output signals of the power amplifier, and divide the source domain dataset and target domain dataset according to the signal type;
[0007] Step 2: Preprocess the data and divide it into training and test sets according to the specified ratio;
[0008] Step 3: Construct a basic LSTM model for power amplifier transfer learning. The main structure of the model includes two fully connected layers and four stacked LSTM layers.
[0009] Step 4: Use the training set portion of the source dataset from Step 2 to train the basic LSTM model for power amplifier transfer learning, and save the trained basic LSTM model for power amplifier transfer learning and its parameters.
[0010] Step 5: Load the trained power amplifier transfer learning basic LSTM model and train it. Use the parameters of the trained power amplifier transfer learning basic LSTM model as the initial values for retraining. At the same time, keep the parameters of the first two LSTM layers in the trained power amplifier transfer learning basic LSTM model unchanged during training. Use the target domain training set to optimize the training of the remaining two LSTM layers in the trained power amplifier transfer learning basic LSTM model to obtain the power amplifier model after transfer learning.
[0011] Furthermore, step 1 specifically includes:
[0012] The acquisition system acquires the signal source output as the input V of the power amplifier. in The signal is simultaneously acquired, along with the output V of the power amplifier. out The signal is the input V of the acquired power amplifier. in and output V out Signals are labeled as T by type i Let i = 1, 2, ..., M. Remove incomplete abnormal pulse signals so that the number of pulses collected for each of the M types of signals is N.
[0013] Furthermore, step 2 specifically includes:
[0014] The kth signal from the M signals collected in step 1 constitutes the transfer learning source domain dataset {T}. k The source domain data is used for initial power amplifier modeling, and then the dataset {T} is processed. k The number of pulses in the dataset is divided into a training set and a test set according to a ratio P; then the remaining signal types are used to construct the target domain dataset {T}. i The training set and the test set are divided into groups P according to the same ratio |i=1,2,…,M,i≠k}.
[0015] Furthermore, in step 3, the basic LSTM model for power amplifier transfer learning includes an input fully connected layer, four sets of LSTM layers, and an output fully connected layer stacked sequentially.
[0016] Each LSTM layer consists of a front Batch-Normalization layer, an LSTM feature extraction layer, and a rear Dropout layer stacked in sequence, with the LSTM feature extraction layer containing S neuron nodes.
[0017] Furthermore, step 4 specifically includes:
[0018] Step 4.1: Set the training parameters, which include optimizer type, maximum number of iterations, gradient threshold, initial learning rate, and learning rate descent factor;
[0019] Step 4.2: Input the training set data from the source dataset in Step 2 into the power amplifier transfer learning basic LSTM model to train the training parameters of the power amplifier transfer learning basic LSTM model. During training, the mean absolute error is used as the loss function for the LSTM. and The weights and bias parameters are updated in the direction of gradient descent to make the loss function of the LSTM get closer and closer to the minimum value to train the training parameters of the basic LSTM model for power amplifier transfer learning. The training parameters of the trained basic LSTM model for power amplifier transfer learning are saved to obtain the trained basic LSTM model for power amplifier transfer learning. Here, η is the learning rate of the basic LSTM model for power amplifier transfer learning, W is the linear fully connected weight matrix of the basic LSTM model for power amplifier transfer learning, and b is the bias parameter. During the training process, the weight matrix and bias parameters of the basic LSTM model for power amplifier transfer learning are continuously updated.
[0020] Furthermore, step 5 specifically includes:
[0021] Step 5.1: Load the trained power amplifier transfer learning basic LSTM model, use the parameters of the trained power amplifier transfer learning basic LSTM model as the initial values of the trained power amplifier transfer learning basic LSTM model during retraining, and set the learning rate of the weights and bias parameters of the first two LSTM layers of the trained power amplifier transfer learning basic LSTM model to 0 to keep the parameters of the first two LSTM layers in the trained power amplifier transfer learning basic LSTM model unchanged.
[0022] Step 5.2: Using the training set data from the target domain data in Step 2, input the trained power amplifier transfer learning basic LSTM model to train the parameters of the remaining two LSTM layers in the trained power amplifier transfer learning basic LSTM model. During training, the mean absolute error is used as the loss function for the LSTM. and The weights and bias parameters are updated in the direction of gradient descent to make the loss function of the LSTM get closer and closer to the minimum value to train the training parameters of the basic LSTM model of the power amplifier after transfer learning. After training, the power amplifier model after transfer learning is obtained.
[0023] Furthermore, the loss function of the LSTM is:
[0024]
[0025] Where L is the loss function, N batch This represents the batch size during network training, where n is the index of the summation count, n = 1, 2, ..., N. batch y n It is the actual value of the nth sample in a certain batch during the training process. It is the predicted value of the nth sample in the same batch.
[0026] Furthermore, it also includes: inputting the test sets of the source domain and the target domain into the trained power amplifier transfer learning basic LSTM model and the power amplifier model after transfer learning, respectively, and calculating and comparing the normalized mean square error of the predicted values and actual values of the two models.
[0027] Beneficial effects: The method of this invention breaks down the isolation of power amplifier characteristics caused by differences in signal type, and establishes a power amplifier model that is adaptable to common underwater acoustic signals. Compared with the traditional LSTM power amplifier model, it reduces the model's dependence on signal type. The method of this invention can also be used for sonar transmitter power amplifier modeling. Attached Figure Description
[0028] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention;
[0029] Figure 2 This is the basic network model structure for transfer learning in the embodiment;
[0030] Figure 3 This is a truncation of the LSTM-TL model prediction segment in the embodiment;
[0031] Figure 4 The example shows the prediction errors of the two power amplifier models for three types of signals. Detailed Implementation
[0032] The invention will now be further explained with reference to the accompanying drawings.
[0033] This invention addresses the problem that conventional power amplifier models are limited by the signal dataset used in modeling, and the resulting power amplifier models are usually only suitable for specific types of signals. It provides a method for modeling the behavior of underwater acoustic broadband power amplifiers based on LSTM and transfer learning.
[0034] like Figure 1 As shown, a method for modeling the behavior of underwater acoustic broadband power amplifiers based on LSTM and transfer learning includes:
[0035] Step 1: Acquire the input and output signals of the power amplifier, and divide the source domain dataset and target domain dataset according to the signal type;
[0036] Step 2: Preprocess the data and divide it into training and test sets according to the specified ratio;
[0037] Step 3: Construct a basic LSTM model for power amplifier transfer learning. The main structure of the model includes two fully connected layers and four stacked LSTM layers.
[0038] Step 4: Use the training set portion of the source dataset from Step 2 to train the basic LSTM model for power amplifier transfer learning, and save the trained basic LSTM model for power amplifier transfer learning and its parameters.
[0039] Step 5: Load the trained power amplifier transfer learning basic LSTM model and train it. Use the parameters of the trained power amplifier transfer learning basic LSTM model as the initial values for retraining. At the same time, keep the parameters of the first two LSTM layers in the trained power amplifier transfer learning basic LSTM model unchanged during training. Use the target domain training set to optimize the training of the remaining two LSTM layers in the trained power amplifier transfer learning basic LSTM model to obtain the power amplifier model after transfer learning.
[0040] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:
[0041] In step 1, a broadband pulse signal is generated by a signal source. The signal uses three commonly used underwater acoustic types (LFM, HFM, and CW), with each type having the same pulse width of 0.5s, the same duty cycle of 0.416, and the same frequency range of 500–700Hz. The signal is output through an L2A broadband power amplifier. The acquisition system uses the signal source output as the input V of the power amplifier. in The signal is simultaneously acquired, along with the output V of the power amplifier. out The signal was sampled at a frequency of 22.05 kSa / s. The acquired signals were labeled as T1, T2, and T3. Among the three types of signals acquired, T1 (LFM) constituted the source domain dataset {T1} for transfer learning. The source domain data was used for initial modeling of the power amplifier. The remaining signal types were then used to construct the target domain datasets {T2, T3}.
[0042] In step 2, incomplete or abnormal pulse signals are removed, ensuring that the number of pulses for each type of signal is 1200. Then, the pulse count of the dataset {T1} from step 1 is divided into training and test sets in a 5:1 ratio. Similarly, the pulse count of the target domain datasets {T2, T3} from step 1 is divided into training and test sets in a 5:1 ratio.
[0043] like Figure 2 As shown, in step 3, the basic LSTM model for power amplifier transfer learning includes an input fully connected layer, four sets of LSTM layers, and an output fully connected layer stacked sequentially. Each set of LSTM layers includes a pre-Batch-Normalization layer, an LSTM feature extraction layer, and a post-Dropout layer stacked sequentially, wherein the LSTM feature extraction layer contains 50 neuron nodes.
[0044] Step 4 specifically includes the following steps:
[0045] Step 4.1: Set the training parameters, which include optimizer type, maximum number of iterations, gradient threshold, initial learning rate, and learning rate descent factor. In this embodiment, the optimizer type is adam, the maximum number of iterations is 1000, the gradient threshold is 1, the initial learning rate is 0.005, the learning rate descent factor is 0.2, and the batch size is 128.
[0046] Step 4.2: Input the training set data of the source dataset {T1} from Step 2 into the power amplifier transfer learning basic LSTM model to train the training parameters of the power amplifier transfer learning basic LSTM model. The mean absolute error is used as the loss function of the LSTM during training. and The weights and bias parameters are updated in the direction of gradient descent to bring the loss function of the LSTM closer and closer to its minimum, thus training the parameters of the basic LSTM model for power amplifier transfer learning. The trained parameters of the basic LSTM model are then saved, resulting in the fully trained basic LSTM model for power amplifier transfer learning. The initial learning rate η = 0.005, and after 125 training iterations, the learning rate is decreased with a learning rate decrease factor of 0.2 and a batch size of 128. During training, the weight matrix and bias parameters of the basic LSTM model for power amplifier transfer learning are continuously updated.
[0047] In step 5, the trained power amplifier transfer learning basic LSTM model is loaded, and its parameters are used as the initial values for retraining. The learning rates of the weights and bias parameters of the first two LSTM layers in the trained power amplifier transfer learning basic LSTM model are set to 0 to keep the parameters of the first two LSTM layers unchanged. Then, the training set data of the target domain data {T2, T3} from step 2 is used as input to train the parameters of the remaining two LSTM layers in the trained power amplifier transfer learning basic LSTM model. The mean absolute error is used as the loss function for the LSTM during training. and The weights and bias parameters are updated in the direction of gradient descent, bringing the loss function of the LSTM closer and closer to its minimum. This process trains the training parameters of the basic LSTM model for power amplifier transfer learning, resulting in the power amplifier model after transfer learning. In this embodiment, the power amplifier model after transfer learning is called the LSTM-TL model.
[0048] The training parameters in steps 4 and 5 both include optimizer type, maximum number of iterations, gradient threshold, initial learning rate, and learning rate descent factor.
[0049] Furthermore, the loss function of LSTM in steps 4 and 5 is:
[0050]
[0051] Where L is the loss function, N batch This represents the batch size during network training, where n is the index of the summation count, n = 1, 2, ..., N. batch y n It is the actual value of the nth sample in a certain batch during the training process. It is the predicted value of the nth sample in the same batch.
[0052] When the training set data of the source dataset in step 4 is input into the basic LSTM model for transfer learning of the power amplifier, and the training set data of the target domain data in step 5 is input into the trained basic LSTM model for transfer learning of the power amplifier, the input form is a matrix, and its form is as follows:
[0053]
[0054] When the training set data of the source dataset or the training set data of the target domain data are input into its corresponding model, the current time step and the next time step input are combined to form [V] in the model's temporal input. in (t)V in[t+1] serves as a time node input for the LSTM network's cyclic computation. Multiple input arrays form a time sequence vertically, resulting in the input matrix of the LSTM network's input layer as shown in equation (1). The network output is determined by V. in V corresponding to (t) out The matrix formed by (t).
[0055] A method for modeling the behavior of underwater acoustic broadband power amplifiers based on LSTM and transfer learning further includes: inputting test sets from the source domain and the target domain into the trained power amplifier transfer learning basic LSTM model and the power amplifier model after transfer learning, respectively, and calculating and comparing the normalized mean square error between the predicted values and the actual values of the two models.
[0056] The specific steps for calculating and comparing the normalized mean square error between the predicted and actual values of the two models are as follows:
[0057] Step A1: First, for the trained power amplifier transfer learning basic LSTM model, the power amplifier input signals from the source and target domain test sets in Step 2 are respectively used as inputs according to Equation (1) and applied to the trained power amplifier transfer learning basic LSTM model. The fully connected layer of the trained power amplifier transfer learning basic LSTM model maps the extracted feature vector h to the signal output. The predicted output V of the power amplifier is obtained by the fully connected weight matrix W and the bias parameter b. pre :
[0058] V pre =hW T +b
[0059] Among them W T This indicates the transpose of the linear fully connected weight matrix W.
[0060] Step A2: Calculate the predicted value V pre The measured power amplifier output signal V of the test set out The normalized mean square error (NMSE) is defined as follows:
[0061]
[0062] The prediction performance of the trained power amplifier transfer learning basic LSTM model on the source and target domain test sets was compared based on the NMSE values.
[0063] Step A3: For the power amplifier model after transfer learning, the same two processes as above are used to compare the prediction effect of the power amplifier model after transfer learning on the test set data of the source domain and the target domain, and to verify the wide adaptability of the proposed power amplifier modeling method of transfer learning to signal types.
[0064] like Figure 3As shown, the CW test signal segment was predicted using the power amplifier model after transfer learning. The instantaneous error amplitude in the figure is less than 0.01, and the error is -30.32dB.
[0065] like Figure 4 As shown, the prediction errors of the trained power amplifier transfer learning base LSTM model and the power amplifier model after transfer learning on the test set of three signal types: LFM, HFM, and CW. The trained power amplifier transfer learning base LSTM model has a prediction error of less than -31dB for LFM and HFM signals, showing good prediction performance. The prediction error for CW signals is significantly lower than that for LFM and HFM, with a fluctuation of about 10dB. The power amplifier model after transfer learning performs well in predicting all three signal types, with normalized mean square errors of less than -30dB. The prediction errors for the three signal types are also relatively close, with fluctuations of less than 1dB, showing higher stability than LSTM-Base.
[0066] The examples show that the power amplifier model after transfer learning retains the effective characteristics of the source domain signal and adapts to the new characteristics of the target domain signal during modeling. It can establish a power amplifier model that is suitable for common underwater acoustic signals and reduces the model's dependence on signal type.
[0067] The method of this invention breaks down the isolation of power amplifier characteristics caused by differences in signal type, and establishes a power amplifier model that is adaptable to common underwater acoustic signals. Compared with traditional LSTM power amplifier models, it reduces the model's dependence on signal type. The method of this invention can also be used for sonar transmitter power amplifier modeling.
[0068] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for modeling the behavior of underwater acoustic broadband power amplifiers based on LSTM and transfer learning, characterized in that, include: Step 1: Acquire the input and output signals of the power amplifier, and divide the source domain dataset and target domain dataset according to the signal type; Step 1 specifically includes: The acquisition system acquires the signal source output as the input V of the power amplifier. in The signal is simultaneously acquired, along with the output V of the power amplifier. out The signal is the input V of the acquired power amplifier. in and output V out Signals are labeled as T by type i , i=1,2,…,M, remove incomplete abnormal pulse signals so that the number of pulses collected for each of the M types of signals is N; Step 2: Preprocess the data and divide it into training and test sets according to a set ratio; Step 2 specifically includes: The kth signal from the M signals collected in step 1 constitutes the transfer learning source domain dataset {T}. k The source domain data is used for initial power amplifier modeling, and then the dataset {T} is processed. k The number of pulses in the dataset is divided into a training set and a test set according to a ratio P; then the remaining signal types are used to construct the target domain dataset {T}. i | i=1,2,…,M, i≠k}, and the training set and test set are also divided according to the proportion P; Step 3: Construct a basic LSTM model for power amplifier transfer learning. The main structure of the model includes two fully connected layers and four stacked LSTM layers. Step 4: Use the training set portion of the source dataset from Step 2 to train the basic LSTM model for power amplifier transfer learning, and save the trained basic LSTM model for power amplifier transfer learning and its parameters. Step 5: Load the trained power amplifier transfer learning basic LSTM model and train it. Use the parameters of the trained power amplifier transfer learning basic LSTM model as the initial values for retraining. At the same time, keep the parameters of the first two LSTM layers in the trained power amplifier transfer learning basic LSTM model unchanged during training. Use the target domain training set to optimize the training of the remaining two LSTM layers in the trained power amplifier transfer learning basic LSTM model to obtain the power amplifier model after transfer learning.
2. The underwater acoustic broadband power amplifier behavior modeling method based on LSTM and transfer learning according to claim 1, characterized in that, In step 3, the basic LSTM model for power amplifier transfer learning includes an input fully connected layer, four sets of LSTM layers, and an output fully connected layer stacked sequentially. Each LSTM layer consists of a front Batch-Normalization layer, an LSTM feature extraction layer, and a rear Dropout layer stacked in sequence, with the LSTM feature extraction layer containing S neuron nodes.
3. The underwater acoustic broadband power amplifier behavior modeling method based on LSTM and transfer learning according to claim 1, characterized in that, Step 4 specifically includes: Step 4.1: Set the training parameters, which include optimizer type, maximum number of iterations, gradient threshold, initial learning rate, and learning rate descent factor; Step 4.2: Input the training set data from the source dataset in Step 2 into the power amplifier transfer learning basic LSTM model to train the training parameters of the power amplifier transfer learning basic LSTM model. During training, the mean absolute error is used as the loss function L of the LSTM. and The weights and bias parameters are updated in the direction of gradient descent to make the loss function L of the LSTM get closer and closer to the minimum value to train the training parameters of the basic LSTM model for power amplifier transfer learning. The training parameters of the trained basic LSTM model for power amplifier transfer learning are saved to obtain the trained basic LSTM model for power amplifier transfer learning. Here, η is the learning rate of the basic LSTM model for power amplifier transfer learning, W is the linear fully connected weight matrix of the basic LSTM model for power amplifier transfer learning, and b is the bias parameter. During the training process, the weight matrix and bias parameters of the basic LSTM model for power amplifier transfer learning are continuously updated.
4. The underwater acoustic broadband power amplifier behavior modeling method based on LSTM and transfer learning according to claim 3, characterized in that, Step 5 specifically includes: Step 5.1: Load the trained power amplifier transfer learning basic LSTM model, use the parameters of the trained power amplifier transfer learning basic LSTM model as the initial values of the trained power amplifier transfer learning basic LSTM model during retraining, and set the learning rate of the weights and bias parameters of the first two LSTM layers of the trained power amplifier transfer learning basic LSTM model to 0 to keep the parameters of the first two LSTM layers in the trained power amplifier transfer learning basic LSTM model unchanged. Step 5.2: Using the training set data from the target domain data in Step 2, input the trained power amplifier transfer learning basic LSTM model to train the parameters of the remaining two LSTM layers in the trained power amplifier transfer learning basic LSTM model. During training, the mean absolute error is used as the loss function for the LSTM. and The weights and bias parameters are updated in the direction of gradient descent to make the loss function of the LSTM get closer and closer to the minimum value to train the training parameters of the basic LSTM model of the power amplifier after transfer learning. After training, the power amplifier model after transfer learning is obtained.
5. The underwater acoustic broadband power amplifier behavior modeling method based on LSTM and transfer learning according to claim 3 or 4, characterized in that, The loss function of the LSTM is: Where L is the loss function, N batch This represents the batch size during network training, where n is the index of the summation count, n = 1, 2, ..., N. batch y n It is the actual value of the nth sample in a certain batch during the training process. It is the predicted value of the nth sample in the same batch.
6. The underwater acoustic broadband power amplifier behavior modeling method based on LSTM and transfer learning according to claim 1, characterized in that, Also includes: Input the test sets of the source domain and the target domain into the trained power amplifier transfer learning basic LSTM model and the power amplifier model after transfer learning, respectively, and calculate and compare the normalized mean square error of the predicted values and actual values of the two models.