Radio frequency device output prediction method based on autoregressive network

Abstract

The invention discloses an autoregressive network-based radio frequency device output prediction method, which comprises the following steps of: generating a random noise signal by using an arbitrary waveform generator, and recording an input signal set and an output signal set by combining an oscilloscope and a radio frequency device; preprocessing the collected data to obtain an input feature vector and a label for model training; and then constructing and training a full-connection autoregressive neural network based on an automatic encoder architecture, and predicting the output of the radio frequency device by using the trained full-connection autoregressive neural network.

Description

Technical Field

[0001] The present invention belongs to the technical field of radio frequency devices, and more specifically, relates to a radio frequency device output prediction method based on an autoregressive network. Background Art

[0002] As the characteristic size of RF devices decreases and the scale of integrated circuits expands, in order to model and explore these complex structures, physics-based models or electrical models are usually used. Both types of models require the internal circuits of RF devices to be analyzed to create the corresponding models, which will generate a very large set of equations. There is no doubt that from the perspective of computational cost and time cost, this is a very time-consuming and extremely cost-effective method.

[0003] Computer-assisted methods have achieved remarkable results in the macroscopic modeling of RF nonlinear circuits and systems. Neural networks have been proven to be well-suited for device and RF component modeling, significantly improving efficiency in active and passive device modeling, device performance optimization, modeling, and simulation. Two approaches to modeling nonlinear circuits are behavioral modeling and transistor modeling. Behavioral modeling focuses on input and output signals without addressing internal circuit details. Transistor modeling, such as SPICE models, can produce more accurate models but is also more time-consuming. Neural networks, due to their powerful ability to simulate the relationship between input and output, are becoming the preferred modeling approach due to their convenience and speed. Feedforward neural networks are inefficient for modeling nonlinear circuits. Researchers have developed various time-domain neural networks, such as recurrent neural networks (RNNs), dynamic neural networks (DNNs), time-delay neural networks (TDNNs), and state-space dynamic neural networks (SSDNNs), all of which can effectively capture dependencies in long time series.

[0004] For RF device modeling, Neuro-SM and dynamic Neuro-SM have also been proposed for modeling nonlinear devices related to DC, small signals, and large signals. Neuro-SM models use static neural networks such as multilayer perceptrons (MLPs) to map a coarse model of device behavior to a fine model, enabling more accurate matching of device behavior. While modeling is typically performed at the transistor level, most researchers first connect the mapping network to typical transistor-like data (such as DC, S-parameters, and HB data) rather than directly using the device's input and output data as neural network training data. Similarly, in dynamic Neuro-SM models, TDNNs are used instead of MLPs to map the coarse model to a fine model, but dynamic Neuro-SM formulas related to DC, S-parameters, and HB data still need to be developed. Recurrent neural networks, such as long short-term memory (LSTM) networks, can use device input and output data directly as neural network training data, enabling representation of device dynamic behavior in the time domain. However, the numerous trainable parameters in these models make training very slow, and predictions of long sequences are prone to vanishing gradients.

[0005] Most researchers use standard waveforms as data drivers, such as sine waves and square waves of varying amplitudes and frequencies. The corresponding test signals also use the same waveforms. The resulting models can effectively simulate the input-output relationships of RF devices, but the robustness of these data-driven models is limited. However, most models focus on transistor-level modeling rather than device-level modeling. The waveform data is derived from simulation data, free of other noise influences, which significantly differs from the diverse noise environments found in actual RF device operating environments. The real world has a wide and diverse noise spectrum, encompassing the entire frequency range from low to high frequencies. Furthermore, the noise in each frequency band can significantly impact the final prediction performance. Using noise data as a data driver to build a detailed macroscopic behavioral model of RF devices offers greater convenience, enabling output prediction for any waveform within the operating frequency band. This robust model can better assist engineers in addressing complex technical challenges. By designing a network structure, we can learn lower-dimensional features in noise data without the need for manual feature extraction, thereby fully constructing the input-output mapping relationship of the device. Summary of the Invention

[0006] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a method for predicting the output of a radio frequency device based on an autoregressive network, which can predict the time domain output of the radio frequency device in a real working environment by relying only on noise as data drive.

[0007] To achieve the above-mentioned object of the invention, the present invention provides a method for predicting radio frequency device output based on an autoregressive network, characterized by comprising the following steps:

[0008] (1) Data collection;

[0009] (1.1) The arbitrary waveform generator randomly generates a random noise signal with a fixed bandwidth, uniform spectrum and time domain amplitude of a. The random noise signal obeys a uniform distribution in the interval [-a / 2, a / 2] in the time domain;

[0010] (1.2) Input the random noise signal into the oscilloscope and record the input signal x of the RF device through the oscilloscope. i ; Input the random noise signal into the RF device first, then input the output signal of the RF device into the oscilloscope, and record the output signal of the RF device through the oscilloscope i ;

[0011] (1.3) Repeat steps (1.1) and (1.2) S times to obtain the input signal set X = {x1, x2, ..., x i ,…,x S}, x i ={x i,1 ,x i,2 ,…,x i,j ,…,x i,m}, x i,j is the jth sample value in the i-th group of input signals; the output signal set Y={y1,y2,…,y i ,…,y S},y i ={y i,1 ,y i,2 ,…,y i,j ,…,y i,m},y i,j The jth sample value in the i-th group of output signals; m represents the number of samples;

[0012] (2) Preprocessing of collected data;

[0013] (2.1), in the input signal set X={x1,x2,…,x i ,…,x S}, calculate each group of input signals x in turn i ={x i,1 ,x i,2 ,…,x i,j ,…,x i,m The mean μ i , standard deviation σ i ;

[0014] Set the mean threshold μ and standard deviation threshold σ; then determine the mean μi , standard deviation σ i Is it satisfied: μ i ≥μ or σ i ≤σ, if satisfied, retain the set of input signals x i And the corresponding output signal y i ; Otherwise, delete the input signal x i And the corresponding output signal y i , and finally get the input signal set And the corresponding output signal set

[0015] (2.2), delay calibration;

[0016] Calculate each set of input signals x i And the corresponding output signal y i The cross-correlation function R xiyi (k):

[0017]

[0018] Where k represents the number of delay points;

[0019] Find the number of delay points k corresponding to the maximum value of the cross-correlation function max :

[0020]

[0021] According to the number of delay points k max For the output signal y i Perform translation, and fill the missing data with 0 to obtain the output signal

[0022]

[0023] in, The length is m, and the number of 0s filled is k max indivual;

[0024] Finally, the new output signal set is obtained

[0025] (3) Construct training data;

[0026] (3.1), data normalization;

[0027] For the input signal set and output signal set Each group of signals in is normalized to obtain the normalized input signal set and output signal set

[0028] (3.2) Extract feature vectors and label values ​​from the normalized data;

[0029] According to the fixed step, the sliding window is used to process each set of input and output signals. Assume that n data points are sampled each time, and a total of K groups of input feature vectors are obtained, among which, for X norm The set of input feature vectors obtained after the i-th group of input signals slides is recorded as Y norm After the i-th group of input signals slides, take the corresponding n-th value As the output label, Among them, “” indicates normalization;

[0030] (4) Build a fully connected autoregressive neural network based on the autoencoder architecture;

[0031] The fully connected autoregressive neural network based on the autoencoder architecture consists of an input layer, an encoder, a decoder, and an output layer.

[0032] The input layer is used to input the feature vector Input to encoder;

[0033] The encoder consists of L hidden layers, and the number of neurons in each hidden layer gradually decreases from n to N. L , the output of a single neuron in each hidden layer is connected to each neuron in the next hidden layer through an activation function;

[0034] The decoder consists of L hidden layers, where the number of neurons in each hidden layer is N L Gradually increase to n, each hidden layer is connected to each hidden layer of the next layer through a fully connected layer and activation function;

[0035] The output layer is used to output the decoding result;

[0036] (5) Training a fully connected autoregressive neural network based on the autoencoder architecture;

[0037] (5.1), the input feature vector With the corresponding label The input is fed into the encoder through the input layer of a fully connected autoregressive neural network;

[0038] (5.2) In the encoder, the first hidden layer processes the input feature vector Proceed as follows:

[0039] (5.2.1) Using the neurons in the first hidden layer to input feature vector Perform a linear transformation, where the output of the lth neuron is:

[0040]

[0041] in, is the input feature vector The lth element in, l=1,2,…,n,w l and b i are the weight and bias of the lth neuron respectively;

[0042] (5.2.2), the output Y of each neuron l 1 Activate through the activation function to get the output of each neuron after activation

[0043]

[0044] Among them, tanh(·) represents the activation function;

[0045] (5.2.3) Output after all neurons are activated Splice and get the output A of the first hidden layer 1 ;

[0046]

[0047] Among them, Concat(·) represents concatenation on the feature dimension;

[0048] (5.2.4) For the 2nd to Lth layers of the encoder, the output of the previous layer is used as the input of the next layer, and then the operations of steps (5.2.1)-(5.2.3) are repeated, and finally the output A of the Lth hidden layer is obtained. L , and serves as the final output A of the encoder;

[0049]

[0050] (5.3) Decoding operation is performed by the decoder;

[0051] (5.3.1) Input the encoder output A to the decoder;

[0052] (5.3.2) Use the neurons in the Lth hidden layer to perform a linear transformation on the output A to obtain the output of each neuron;

[0053]

[0054] in, and are the weights and biases of the Lth hidden layer of the decoder, l=1,2,…,N L ;

[0055] (5.3.3), the output of each neuron Activate through the activation function to get the output of each neuron after activation

[0056]

[0057] (5.3.4) Output after all neurons are activated Splice and get the output H of the Lth hidden layer L ;

[0058]

[0059] (5.3.5) For the L-1th layer to the 1st layer of the decoder, the output of the previous layer is used as the input of the next layer, and then the operations of steps (5.3.2)-(5.3.4) are repeated to obtain the output of the 1st hidden layer;

[0060]

[0061] (5.3.6) The output of the first hidden layer is output through the output layer to obtain the output prediction value of the RF device

[0062] (5.4), X norm After the i-th group of input signals slides, the K groups of input feature vectors are sequentially input into the fully connected autoregressive neural network. According to steps (5.1)-(5.3), the K groups of output prediction values ​​are obtained, which are recorded as

[0063] (5.5) Calculate the loss value Loss of the fully connected autoregressive neural network after this round of training;

[0064]

[0065] (5.6) Based on the loss value Loss after this round of training, the parameters of the fully connected autoregressive neural network are updated by the gradient descent method, and then the training is repeated until the fully connected autoregressive neural network converges to obtain the trained fully connected autoregressive neural network;

[0066] (6) Predict the output of RF devices;

[0067] Generate a random noise signal according to the method of step (1.1), then input the random noise signal into the oscilloscope, and obtain the input signal x of the RF device through sampling by the oscilloscope; process the input signal x according to steps (2.1), (3.1), and (3.2) to obtain the input feature vector, and input it into the trained fully connected autoregressive neural network to obtain the output prediction value of the RF device

[0068] The object of the invention of the present invention is achieved like this:

[0069] The present invention is based on an autoregressive network-based RF device output prediction method. First, an arbitrary waveform generator is used to generate a random noise signal. Then, an oscilloscope and RF device are used to record the input and output signal sets. The collected data is then preprocessed to obtain input feature vectors and labels for model training. Then, a fully connected autoregressive neural network based on an autoencoder architecture is constructed and trained, and the trained fully connected autoregressive neural network is used to predict the RF device output.

[0070] At the same time, the RF device output prediction method based on the autoregressive network of the present invention also has the following beneficial effects:

[0071] (1) By using noise data to drive the training network, it is possible to predict sine waves of different frequency bands that are different from noise signals, and the model has strong generalization performance;

[0072] (2) Its ability to capture long sequences is better than that of RNN, LSTM and GRU time series prediction models. Its training complexity is much lower than that of RNN and LSTM, and its training time is shorter. It consumes less memory, has better generalization ability, and can better process data in parallel, resulting in higher computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a flow chart of the method for predicting radio frequency device output based on an autoregressive network according to the present invention;

[0074] Figure 2 It is a schematic diagram of the waveform acquisition platform;

[0075] Figure 3 It is the time domain and frequency domain distribution diagram of random noise signal;

[0076] Figure 4 This is a diagram of the fully connected autoregressive neural network architecture based on the autoencoder architecture;

[0077] Figure 5 is the sine wave prediction result;

[0078] Figure 6 It is the triangle wave prediction result. DETAILED DESCRIPTION

[0079] The following describes the specific embodiments of the present invention in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, such descriptions will be omitted here.

[0080] Example

[0081] Figure 1 This is a flow chart of the radio frequency device output prediction method based on the autoregressive network of the present invention.

[0082] In this embodiment, if Figure 1 As shown, the present invention provides a method for predicting radio frequency device output based on an autoregressive network, comprising the following steps:

[0083] S1, data collection;

[0084] In this embodiment, we first need to build a waveform acquisition platform for data acquisition, such as Figure 2 As shown, the waveform acquisition platform includes

[0085] S1.1. The arbitrary waveform generator randomly generates a random noise signal with a bandwidth of 3G, a uniform spectrum distribution, and a time domain amplitude of 1Vpp. The random noise signal obeys a uniform distribution in the interval [-0.5, 0.5] in the time domain. The specific time domain and frequency domain distribution is as follows Figure 3 As shown, Figure 3 (a) is the time domain amplitude distribution histogram, Figure 3 (b) is the frequency domain amplitude distribution diagram.

[0086] S1.2. Input the random noise signal into the oscilloscope and record the input signal x of the RF device through the oscilloscope. i ; Input the random noise signal into the RF device first, then input the output signal of the RF device into the oscilloscope, and record the output signal of the RF device through the oscilloscope i ;

[0087] S1.3. Repeat steps S1.1 and S1.2 for a total of S = 100 times to obtain the input signal set X = {x1, x2, ..., x i ,…,x S}, x i ={x i,1 ,x i,2 ,…,x i,j ,…,x i,m}, x i,j is the jth sample value in the i-th group of input signals; the output signal set Y={y1,y2,…,y i ,…,y S},yi ={y i,1 ,y i,2 ,…,y i,j ,…,y i,m},y i,j The jth sample value in the i-th group of output signals; m represents the number of samples, which is 15000;

[0088] S2, preprocessing of collected data;

[0089] S2.1. In this embodiment, when the arbitrary waveform generator opens a channel and is ready to start, if the oscilloscope is in single trigger mode, an unexpected trigger phenomenon may occur. When the AWG channel is turned on but the waveform has not yet been actually output, the oscilloscope may trigger and collect data in advance, causing the oscilloscope to capture non-preset waveform data. Therefore, the input signal set X = {x1, x2, ..., x i ,…,x S Each input signal x in i ={x i,1 ,x i,2 ,…,x i,j ,…,x i,m The mean μ i , standard deviation σ i ;

[0090] Set the mean threshold μ=0.1 and the standard deviation threshold σ=0.1; then determine the mean μ i , standard deviation σ i Is it satisfied: μ i ≥μ or σ i ≤σ, if satisfied, retain the set of input signals x i And the corresponding output signal y i ; Otherwise, delete the input signal x i And the corresponding output signal y i , and finally get the input signal set And the corresponding output signal set In this embodiment, S' is also 100, indicating that all 100 sets of sampled data can be used for model training;

[0091] S2.2, delay calibration;

[0092] In this embodiment, since the recording of the input signal and the recording of the output signal may be delayed due to different transmission lengths, the delay is obtained by calculating the correlation and calibrated. The specific operation process is as follows:

[0093] Calculate each set of input signals x i And the corresponding output signal y i The cross-correlation function R xiyi (k):

[0094]

[0095] Where k represents the number of delay points;

[0096] Find the number of delay points k corresponding to the maximum value of the cross-correlation function max :

[0097]

[0098] According to the number of delay points k max For the output signal y i Perform translation, and fill the missing data with 0 to obtain the output signal

[0099]

[0100] in, The length is m, and the number of 0s filled is k max indivual;

[0101] Finally, the new output signal set is obtained

[0102] S3, build training data;

[0103] S3.1, data normalization;

[0104] For the input signal set and output signal set Each group of signals in is normalized to obtain the normalized input signal set and output signal set

[0105] In this embodiment, the normalization method is: first find each group of input signals x i The maximum value among the S' maximum values ​​is selected, and the absolute value of the selected maximum value is recorded as x max , and finally divide each element in the input signal set by the maximum value x max Similarly, select the maximum value of the total output signal set, take the absolute value and record it as y max , then divide each element in the output signal set by the maximum value y max ;

[0106] S3.2. Extract feature vectors and label values ​​from the normalized data;

[0107] According to a fixed step, a sliding window is used to perform sliding window processing on each set of input and output signals. Assume that n data points are sampled each time, and a total of K = 2048 sets of input feature vectors are obtained. In this embodiment, sampling is performed with a step size of 1, and n = 4096 data points are sampled each time.

[0108] Among them, for X norm The set of input feature vectors obtained after the i-th group of input signals slides is recorded as Y norm After the i-th group of input signals slides, take the corresponding n-th value As the output label, Among them, "" indicates normalization; so each set of input feature vectors The corresponding output label Composition of training data;

[0109] S4. Build a fully connected autoregressive neural network based on the autoencoder architecture;

[0110] like Figure 4 As shown in Figure 1, the fully connected autoregressive neural network based on the autoencoder architecture includes an input layer, an encoder, a decoder, and an output layer;

[0111] The input layer is used to input the feature vector Input to encoder;

[0112] The encoder consists of 9 hidden layers, where the number of neurons in each hidden layer gradually decreases from 4096 to 16, and the output of a single neuron in each hidden layer is connected to each neuron in the next hidden layer through an activation function;

[0113] The decoder consists of 9 hidden layers, where the number of neurons in each hidden layer gradually increases from 16 to 4096, and each hidden layer is connected to each hidden layer of the next layer through a fully connected layer and an activation function; Figure 4 As shown, in this embodiment, n=N1=4096, N L =16

[0114] The output layer is used to output the decoding result;

[0115] S5. Train a fully connected autoregressive neural network based on the autoencoder architecture;

[0116] S5.1. Input feature vector With the corresponding label The input is fed into the encoder through the input layer of a fully connected autoregressive neural network;

[0117] S5.2. In the encoder, the first hidden layer processes the input feature vector Proceed as follows:

[0118] S5.2.1. Use the neurons in the first hidden layer to input the feature vector Perform a linear transformation, where the output of the lth neuron is:

[0119]

[0120] in, is the input feature vector The lth element in, l=1,2,…,n,w l and b i are the weight and bias of the lth neuron respectively;

[0121] S5.2.2. The output Y of each neuron l 1 Activate through the activation function to get the output of each neuron after activation

[0122]

[0123] Among them, tanh(·) represents the activation function;

[0124] S5.2.3. Output after activating all neurons Splice and get the output A of the first hidden layer 1 ;

[0125]

[0126] Among them, Concat(·) represents concatenation on the feature dimension;

[0127] S5.2.4. For the 2nd to 9th layers of the encoder, the output of the previous layer is used as the input of the next layer, and then the operations of steps S5.2.1-S5.2.3 are repeated to finally obtain the output A of the Lth hidden layer. L , and serves as the final output A of the encoder;

[0128]

[0129] S5.3. Perform decoding operation by the decoder;

[0130] S5.3.1. Input the encoder output A to the decoder;

[0131] S5.3.2. Use the neurons in the ninth hidden layer to perform a linear transformation on the output A to obtain the output of each neuron.

[0132]

[0133] in, and are the weights and biases of the Lth hidden layer of the decoder, l=1,2,…,N L ;

[0134] S5.3.3. The output of each neuron Activate through the activation function to get the output of each neuron after activation

[0135]

[0136] S5.3.4. Output after activating all neurons Splice and get the output H of the Lth hidden layer L ;

[0137]

[0138] S5.3.5. For the decoder layers 8 to 1, use the output of the previous layer as the input of the next layer, and then repeat steps S5.3.2-S5.3.4 to obtain the output of the first hidden layer.

[0139]

[0140] S5.3.6. The output of the first hidden layer is output through the output layer to obtain the output prediction value of the RF device.

[0141] S5.4, X norm After the i-th group of input signals slide, the K groups of input feature vectors are sequentially input into the fully connected autoregressive neural network, and the K groups of output prediction values ​​are obtained according to steps S5.1-S5.3, which are recorded as

[0142] S5.5. Calculate the loss value Loss of the fully connected autoregressive neural network after this round of training;

[0143]

[0144] S5.6. Update the parameters of the fully connected autoregressive neural network using the gradient descent method based on the loss value after this round of training. Then, repeatedly iterate the training until the fully connected autoregressive neural network converges, thereby obtaining a trained fully connected autoregressive neural network.

[0145] S6. Predict the output of the RF device;

[0146] Generate a random noise signal according to the method in step S1.1, then input the random noise signal into the oscilloscope, and obtain the input signal x of the RF device through sampling by the oscilloscope; process the input signal x according to steps S2.1, S3.1, and S3.2 to obtain the input feature vector, and input it into the trained fully connected autoregressive neural network to obtain the output prediction value of the RF device

[0147] Simulation Verification

[0148] To verify the effectiveness of the present invention, predictions were performed for sine waves and triangle waves. An arbitrary waveform generator was used to generate a sine wave with a frequency of 1 GHz and an amplitude of 0.4 Vpp. The input and output signals were collected by an oscilloscope with S = 1 in step S1.3. After normalization in step S2.2, the signals were directly input into the network. The predicted output was compared with the actual output, and the error was calculated.

[0149] Figure 5 The results of this method's time-domain prediction of a sine wave with a frequency of 1 GHz and an amplitude of 0.4 Vpp are shown. As can be seen, the error between the actual and predicted outputs is only 0.000093, reaching the e-5 level. The two curves have a very high degree of overlap, indicating a very accurate prediction. Figure 6 The results of time-domain prediction of a triangle wave with a frequency of 1 GHz and an amplitude of 0.4 Vpp are shown. The error between the actual output and the predicted output is only 0.000530, reaching the e-4 level. The prediction is also almost accurate, indicating that when only driven by noise data, the present invention can still achieve good prediction effects on other waveforms.

[0150] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

Claims

1. A method for predicting RF device output based on autoregressive network. It is characterized in that The following steps are involved: (1) Data collection; (1.1) The arbitrary waveform generator randomly generates a random noise signal with a fixed bandwidth, uniform spectrum and time domain amplitude of a. The random noise signal obeys a uniform distribution in the interval [-a / 2, a / 2] in the time domain; (1.2) Input the random noise signal into the oscilloscope and record the input signal x of the RF device through the oscilloscope. i ; Input the random noise signal to the RF device first, then input the output signal of the RF device to the oscilloscope, and record the output signal of the RF device through the oscilloscope i ; (1.3) Repeat steps (1.1) and (1.2) S times to obtain the input signal set X = {x 1 ,x 2 ,…,x i ,…,x S }, x i ={x i,1 ,x i,2 ,…,x i,j ,…,x i,m }, x i,j is the jth sample value in the i-th group of input signals; the output signal set Y = {y 1 ,y 2 ,…,y i ,…,y S },y i ={y i,1 ,y i,2 ,…,y i,j ,…,y i,m },y i,j The jth sample value in the i-th group of output signals; m represents the number of samples; (2) Preprocessing of collected data; (2.1), in the input signal set X = {x 1 ,x 2 ,…,x i ,…,x S }, calculate each group of input signals x in turn i ={x i,1 ,x i,2 ,…,x i,j ,…,x i,m The mean μ i , standard deviation σ i ; Set the mean threshold μ and standard deviation threshold σ; then determine the mean μ i , standard deviation σ i Is it satisfied: μ i ≥μ or σ i ≤σ, if it is satisfied, keep the group of input signals x i And the corresponding output signal y i ;otherwise , delete the input signal x i And the corresponding output signal y i , and finally get the input signal set And the corresponding output signal set S'≤S; (2.2), delay calibration; Calculate each set of input signals x i And the corresponding output signal y i The cross-correlation function Where k represents the number of delay points; Find the number of delay points k corresponding to the maximum value of the cross-correlation function max : According to the delay point k max For the output signal y i After translation, the missing data is filled with 0 to obtain the output signal. in, The length is m, and the number of filled 0s is k max indivual; Finally, we get a new output signal set (3) Constructing a data set; (3.1) Data normalization; For the input signal set And the output signal set Each group of signals in is normalized to obtain the normalized input signal set And the output signal set (3.2) Extract feature vectors and label values ​​from the normalized data; According to the fixed step, the sliding window is used to process each group of input and output signals. Assume that n data points are sampled each time, and a total of K groups of input feature vectors are obtained, among which, for X norm A set of input feature vectors obtained after the i-th group of input signals slides is recorded as Y norm After the i-th group of input signals slides, take the corresponding n-th value As the output label, Among them, "" indicates normalization; (4) Build a fully connected autoregressive neural network based on the autoencoder architecture; The fully connected autoregressive neural network based on the autoencoder architecture includes an input layer, an encoder, a decoder, and an output layer; The input layer is used to transform the input feature vector Input to encoder; The encoder consists of L hidden layers, and the number of neurons in each hidden layer gradually decreases from n to N. L , the output of a single neuron in each hidden layer is connected to each neuron in the next hidden layer through an activation function; The decoder consists of L hidden layers, where the number of neurons in the first L hidden layers is N L Gradually increase to n, each hidden layer is connected to each hidden layer of the next layer through a fully connected layer and an activation function; The output layer is used to output the decoding result; (5) Train a fully connected autoregressive neural network based on an autoencoder architecture; (5.1), the input feature vector With the corresponding label The input is fed to the encoder through the input layer of a fully connected autoregressive neural network; (5.2) In the encoder, the first hidden layer processes the input feature vector Do the following: (5.2.1) Using the neurons in the first hidden layer to input the feature vector Perform a linear transformation, where the output of the lth neuron is: in, is the input feature vector The lth element in, l=1,2,…,n,w l and b i are the weight and bias of the lth neuron respectively; (5.2.2), the output Y of each neuron l 1 Activate through the activation function to get the output of each neuron after activation Among them, tanh(·) represents the activation function; (5.2.3) Output after all neurons are activated Concatenate and get the output A of the first hidden layer 1 ; Among them, Concat(·) means concatenation on the feature dimension; (5.2.4) For the 2nd to Lth layers of the encoder, the output of the previous layer is used as the input of the next layer, and then the operations of steps (5.2.1)-(5.2.3) are repeated, and finally the output A of the Lth hidden layer is obtained. L , and used as the final output A of the encoder; (5.3) Perform decoding operation through a decoder; (5.3.1) Input the encoder output A to the decoder; (5.3.2) Use the neurons in the Lth hidden layer to perform a linear transformation on the output A to obtain the output of each neuron; in, and are the weights and biases of the Lth hidden layer of the decoder, l=1,2,…,N L ; (5.3.3) The output of each neuron Activate through the activation function to get the output of each neuron after activation (5.3.4) Output after all neurons are activated Concatenate and get the output H of the Lth hidden layer L ; (5.3.5) For the L-1th layer to the 1st layer of the decoder, the output of the previous layer is used as the input of the next layer, and then the operations of steps (5.3.2)-(5.3.4) are repeated to obtain the output of the 1st hidden layer; (5.3.6) The output of the first hidden layer is output through the output layer to obtain the output prediction value of the RF device. (5.4), X norm After the i-th group of input signals slide, K groups of input feature vectors are obtained and input into the fully connected autoregressive neural network in sequence. According to steps (5.1)-(5.3), K groups of output prediction values ​​are obtained, which are recorded as (5.5) Calculate the loss value Loss of the fully connected autoregressive neural network after this round of training; (5.6) According to the loss value Loss after this round of training, the parameters of the fully connected autoregressive neural network are updated by the gradient descent method, and then the fully connected autoregressive neural network is trained repeatedly until the fully connected autoregressive neural network converges to obtain a trained fully connected autoregressive neural network; (6) Predict the output of RF devices; Generate a random noise signal according to the method of step (1.1), then input the random noise signal into an oscilloscope, and obtain the input signal x of the RF device through sampling by the oscilloscope; process the input signal x according to steps (2.1), (3.1), and (3.2) to obtain an input feature vector, and input it into the trained fully connected autoregressive neural network to obtain the output prediction value of the RF device

Images