Device transmission signal prediction method based on deep learning

Abstract

The invention relates to a device transmission signal prediction method based on deep learning. The method comprises the following steps: S0, converting an actually acquired single-path signal into a signal data sequence represented by an orthogonal component; s1, constructing a signal prediction model based on deep learning to predict a device output signal according to the signal data sequence; the signal prediction model comprises a convolutional neural network used for performing feature extraction on the signal data sequence to obtain a data feature sequence; and the LSTM module is used for predicting according to the data feature sequence to obtain a device output signal. According to the method, the complex nonlinear relationship between the input signal and the output signal of the signal transmission device can be mined very accurately, high-precision prediction of the signal is realized, the model has the characteristics of miniaturization and light weight, the prediction speed is greatly improved while the prediction accuracy is ensured, and the prediction efficiency is improved. And an efficient, practical and advanced solution is provided for the field of signal transmission prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence signal processing, and in particular to a device transmission signal prediction method based on deep learning. Background Art

[0002] In modern electronic systems, signal transmission devices play a vital role. They are responsible for transmitting information from one component to another to ensure the normal operation of the system, and their performance directly affects the operating efficiency and stability of the entire system. Traditional signal transmission devices are easily affected by various interference factors during the signal transmission process, such as distortion, noise, etc. These interference factors will reduce the accuracy and stability of signal transmission, thereby affecting the performance of the entire system. With the increasing complexity of application scenarios, the nonlinear relationship between input and output is becoming more and more complex. Therefore, in order to improve the accuracy and stability of signal transmission, and thus improve the performance of the entire system, under the influence of complex application scenarios of transmission devices and various external interference factors, an innovative method is urgently needed to accurately predict the output signal based on the device input in a short time with higher efficiency. However, the existing methods are difficult to achieve the requirements of quickly and accurately predicting the device output signal based on the device input, and there is a lack of a method that can predict the signal results more quickly and accurately. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a device transmission signal prediction method based on deep learning, which can accurately explore the nonlinear relationship between the input signal and the output signal of the signal transmission device, and at the same time, can complete the prediction task in a very short time while making accurate predictions.

[0004] The technical solution adopted by the present invention to solve the technical problem is: to provide a device transmission signal prediction method based on deep learning, comprising the following steps:

[0005] S0 converts the actual single-channel signal into a signal data sequence represented by orthogonal components;

[0006] S1 constructs a signal prediction model based on deep learning to predict the device output signal according to the signal data sequence; the signal prediction model includes:

[0007] A convolutional neural network is used to extract features from the signal data sequence to obtain a data feature sequence;

[0008] The LSTM module is used to predict the device output signal according to the data feature sequence.

[0009] Furthermore, before step S1, the method further includes a step of normalizing the signal data sequence obtained by conversion.

[0010] Furthermore, the normalization formula is as follows:

[0011]

[0012] Among them, min is the minimum value of all data, max is the maximum value of all data, data in and data out They are the signal data before and after normalization.

[0013] Furthermore, the signal prediction model is trained by the following method:

[0014] The length of the signal data sequence in each sample is used as the window width, one signal data is used as the step length, and the normalized signal data sequence is divided into multiple training samples using a sliding window;

[0015] The obtained training samples are divided into training data set, verification data set and test data set;

[0016] The signal prediction model is trained using a training data set, and the model parameters are optimized according to a loss function, and then verified using a validation data set and tested using a test data set.

[0017] Furthermore, the loss function is expressed as

[0018]

[0019] Here, x and y are the output and label of the model respectively.

[0020] Furthermore, when predicting the amplifier transmission signal, the signal data sequence is obtained by the following formula:

[0021]

[0022] Where am is the amplitude of the amplifier input signal and degree is the phase of the amplifier input signal.

[0023] Furthermore, the convolutional neural network includes a first convolutional layer, an activation function, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer and a normalization layer connected in sequence.

[0024] Furthermore, the convolution kernel size of the convolutional neural network is 3, the convolution step size is 1, and the padding is 1.

[0025] Furthermore, the activation function is a RELU function.

[0026] Beneficial Effects

[0027] Due to the adoption of the above-mentioned technical scheme, the present invention has the following advantages and positive effects compared with the prior art: the present invention targets various signal transmission devices in complex circuit systems, constructs a signal prediction model based on a deep learning model to find the nonlinear relationship between the device input signal and the output signal, and takes into account the large signal gap in different time periods and under different environments, as well as the large difference between the amplitude value and the phase value, and performs conversion preprocessing on the input signal to facilitate model training, thereby ultimately improving the accuracy of the output signal prediction result; at the same time, the present invention specifically preprocesses the device input signal, so that the model can present the characteristics of miniaturization and lightweight, while ensuring the prediction accuracy, greatly improving the prediction speed. Compared with the existing similar methods, the method of the present invention can improve the prediction efficiency by 50%, while ensuring the accuracy, it only takes half the time to achieve the same prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is an original signal diagram collected by the device in an embodiment of the present invention;

[0029] Figure 2 is a signal data diagram after preprocessing in an embodiment of the present invention;

[0030] Figure 3 is a diagram of a data segmentation method in a sliding window form in an embodiment of the present invention;

[0031] Figure 4 It is a schematic diagram of a model of an embodiment of the present invention;

[0032] Figure 5 is a model training flow chart of an embodiment of the present invention;

[0033] Figure 6 It is a work flow chart of an embodiment of the present invention;

[0034] Figure 7 is a diagram showing the change in loss according to an embodiment of the present invention;

[0035] Figure 8 It is the MSE variation diagram of the test set in the implementation manner of the present invention;

[0036] Fig. 9 is a test set prediction result diagram of simulation data in an embodiment of the present invention;

[0037] Fig.10 is a prediction result diagram of the actual collected test set in an embodiment of the present invention;

[0038] Fig.11 is a graph of ablation experiment results of a convolutional network in an embodiment of the present invention;

[0039] Fig.12 It is a graph of LSTM network ablation experiment results in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0041] The embodiment of the present invention relates to a device transmission signal prediction method based on deep learning. For various signal transmission devices in a complex circuit system, a small and lightweight deep learning model is designed to find the nonlinear relationship between the device input signal and the output signal, and the input signal or output signal of each device is predicted, including the following steps:

[0042] 1) Considering the large difference between signals in different time periods and under different environments, as well as the large difference between amplitude and phase values, it is necessary to perform conversion preprocessing on the signal to facilitate model training and improve the accuracy of the final result;

[0043] 2) Design a small and lightweight deep learning model;

[0044] 3) Network training and signal prediction based on deep learning models;

[0045] 4) Performance evaluation of signal prediction models based on deep learning models.

[0046] Due to the different working conditions and working environments of different devices, the methods for data prediction are also different. In fact, no matter what the working conditions of the device are or the changes in the working environment, there is a certain nonlinear relationship between the input signal and the output signal of the device. Changes in working conditions and environment will only make this nonlinear relationship more complicated. Therefore, in theory, a deep learning model can be designed to learn the deep features between signals. By training a large number of samples, the change law between the input signal and the output signal can be established, so as to realize the prediction of the device signal. In order to realize the prediction of the input or output signal of the device, the online learning method based on the deep learning model mainly needs to solve two problems: first, the current output signal of the device is not only related to the current input, but also related to the historical input signal of the device. The problem to be solved is to integrate the input signal sequence; second, the problem to be solved is how to build a small and lightweight deep learning model. By sending the signal into the model, the model can accurately learn the association between the input sequence and the nonlinear relationship between the input signal and the output signal, so as to realize the accurate prediction of the signal.

[0047] Taking the prediction of the amplifier output signal as an example, the original signal of the amplifier is converted and preprocessed. Through the offline learning of the deep learning model and the fitting of the nonlinear relationship, the input signal or output signal of the device can be accurately predicted. First, the amplitude and phase data are converted into signal data by establishing a mathematical model, and the signal data is normalized; then, a small and lightweight deep learning model is designed and built; again, offline learning and fitting of the nonlinear relationship between input and output are performed based on the deep learning model to achieve accurate prediction of the device signal. Finally, the signal prediction performance based on the deep neural network is verified and evaluated. The deep learning model is used to train the input and output signals of the amplifier, and then the trained model is used to predict another section of the output signal or input signal.

[0048] The following further describes the various steps in predicting the amplifier output signal in conjunction with the accompanying drawings.

[0049] Step 1: Data transformation and preprocessing

[0050] In order to predict the device signal, the present invention first preprocesses the signal, that is, converts the amplitude and phase of the signal into signal data, normalizes the signal data, and then uses the preprocessed signal as the input of the deep learning model to train and evaluate the model.

[0051] The formula process for converting amplitude and phase data into signal data is as follows:

[0052]

[0053] Among them, am is the amplitude data of the original signal of the amplifier, and degree is the phase data of the original signal of the amplifier. Figure 1 and Figure 2 The numerical range before and after the signal conversion in the present invention is shown. The original collected data is normalized after the signal conversion so that the data are at the same level to facilitate the training of the model. The normalization formula is as follows:

[0054]

[0055] Where min is the minimum value of all data, max is the maximum value of all data, and data in The present invention utilizes the above principle to pre-process the training samples first, and then utilizes these data to train the model.

[0056] Before training, it is necessary to design the loss function of the network to determine the error between the predicted data of the model and the actual data, which is also an important part of the present invention. The loss function used in the present invention is SmoothL1Loss, which is a smooth loss function widely used in target regression problems, and it balances L1Loss and L2Loss to a certain extent. The expression of the loss function is as follows:

[0057]

[0058] Among them, x and y are the output and label of the model respectively, and |xy| represents the difference between them. When |xy| is less than 1, square error is used; otherwise, linear error is used. This makes SmoothL1Loss smoother than MSELoss, that is, the response to outliers is more gentle.

[0059] Step 2: Design a small and lightweight deep learning model

[0060] After determining the loss function, a small and lightweight deep learning model is designed. In order to better allow the model to extract the feature relationship in the data and simplify the network structure of the model, the present invention adopts a convolutional neural network and a long short-term memory (LSTM) method to extract features. In order to extract the features of the data details, the present invention adopts a convolutional neural network with a small convolution kernel, a convolution kernel size of 3, a convolution step of 1, and a padding of 1. In addition, after the first layer of the convolutional network, an activation function is used to increase the nonlinear relationship between the feature parameter transformations of the model convolutional network, and normalization is added after the fifth layer of the convolutional network to increase the fitting convergence ability of the model. Since the present invention divides the input data into the form of a sequence, a special LSTM module for processing the sequence amount is also used to learn the feature relationship between the sequences. The entire model architecture is to connect the convolutional neural network and the LSTM module in series to form the entire signal prediction model. When training the model, after the data is input into the model, the output is first obtained through feature extraction of the convolutional neural network, and then the output encoding is transmitted to the LSTM module for sequence reconstruction to learn the relationship between the sequences. Finally, the error (SmoothL1Loss) is calculated between the output and the actual true value, and the error value is used to optimize the model so that the model can find the nonlinear relationship between the two (the nonlinear relationship between the input signal and the output signal). It should be noted that the input dimension of the convolutional neural network is the channel of the data. In the case of the present invention, the input and final output dimensions of the convolutional neural network are both 2, that is, two orthogonal components of each signal. In addition, the input dimension and final output dimension of the LSTM are both the length of the sequence.

[0061] Step 3: Deep Neural Network Model Training

[0062] After preprocessing the data in step 1, the present invention divides the labeled data set into three parts, which are respectively used as a training data set, a test data set and a verification data set. The training data set contains 50,000 samples, the verification data set contains 10,000 samples, and the test data set contains 200 samples. After preparing the data set, the present invention trains the data based on the model in step 2. The Adam optimizer is used, the learning rate of the model is set to 0.0001, and the batch size is 1024. Since the input and output of the signal are a continuous sequence, the present invention uses a sliding window method to traverse each signal when making the input data set to avoid missing signals, resulting in the lack of relationship between the data, which will cause invisibility to the learning ability and training efficiency of the model. The data segmentation method in the form of a sliding window is as follows: Figure 3 The structure diagram of the model and the model training flow chart are shown in Figure 4 and Figure 5 shown.

[0063] After learning the known samples in the signal sequence library, the model designed by the present invention will automatically find the nonlinear relationship between the device input data and output data, forming a model network with the ability to predict input and output signals. The network can be used to predict the input or output of the device.

[0064] Step 4: Signal prediction performance evaluation method based on deep learning model

[0065] After completing the above data preparation, model design and model training, it is necessary to evaluate the performance parameters of the model of the present invention. In the present invention, the root mean square error (MSE) and the mean absolute percentage error (MAPE) are used as the accuracy of the evaluation signal prediction model. In the validation set, the calculation formula of the root mean square error (MSE) of the model prediction results is as follows:

[0066]

[0067] On the test set, the mean absolute percentage error (MAPE) of the model is as follows:

[0068]

[0069] Among them, Y i represents the predicted value of the model, Represents the true value of the sample, n is the total number of samples, and i represents the sequence number of the sample.

[0070] Next, ablation experiments are performed to verify the impact of each component module of the model on the final result and the performance of the model for signal prediction. The specific method is to generate the input signal and the corresponding output signal of the amplifier device through simulation software, and use the constructed artificial intelligence signal prediction algorithm to predict the test signal to verify the performance of the signal prediction method based on deep learning of the present invention.

[0071] According to the formula listed, the signal preprocessing data is first generated, such as Figure 1 , Figure 2 After the samples are labeled, they are sent to the deep learning network of the present invention for training and signal prediction. The changes in loss and MSE during the training process are shown in Figure 7 , Figure 8 .

[0072] After the network is trained, new samples are predicted online to verify the performance of the model. New signal samples are sent to the deep learning model for prediction. The results are as follows Fig. 9 Finally, in order to more realistically verify the effectiveness of the present invention, the present invention uses the above model to predict a section of the amplifier signal actually collected and judge the effect of the model. The predicted results are as follows: Fig.10In addition, ablation experiments were performed to verify the effect of the model and the efficiency of each module. The ablation experiment results are shown in Table 1. The I signal prediction results of the ablation experiment are shown in Table 1. Fig.11 , Fig.12 The present invention verifies its signal prediction capability in a very short time according to the final test results.

[0073] Table 1 Ablation experiment results

[0074] Model Validation set mse results Test set mape Conv+LSTM 0.000267 1％ Conv 0.002276 8％ LSTM 0.000267 1.6％

Claims

1. A device transmission signal prediction method based on deep learning, characterized in that: The following steps are involved: S0 converts the actual single-channel signal into a signal data sequence represented by orthogonal components; S1 constructs a signal prediction model based on deep learning to predict the device output signal according to the signal data sequence; the signal prediction model includes: A convolutional neural network is used to extract features from the signal data sequence to obtain a data feature sequence; The LSTM module is used to predict the device output signal according to the data feature sequence.

2. The method according to claim 1, characterized in that Before step S1, the method further includes a step of normalizing the signal data sequence obtained by conversion.

3. The method according to claim 2, characterized in that The normalization formula is as follows: Among them, min is the minimum value of all data, max is the maximum value of all data, data in and data out They are the signal data before and after normalization.

4. The method according to claim 2, characterized in that: The signal prediction model is trained by the following method: taking the length of the signal data sequence in each sample as the window width and one signal data as the step length, and using a sliding window to divide the normalized signal data sequence into a plurality of training samples; The obtained training samples are divided into training data set, verification data set and test data set; The signal prediction model is trained using a training data set, and the model parameters are optimized according to a loss function, and then verified using a validation data set and tested using a test data set.

5. The method according to claim 4, characterized in that The loss function is expressed as Here, x and y are the output and label of the model respectively.

6. The method according to claim 1, characterized in that When predicting the amplifier transmission signal, the signal data sequence is obtained by the following formula: Where am is the amplitude of the amplifier input signal and degree is the phase of the amplifier input signal.

7. The method according to claim 1, characterized in that The convolutional neural network includes a first convolutional layer, an activation function, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer and a normalization layer which are connected in sequence.

8. The method according to claim 7, characterized in that The convolution kernel size of the convolutional neural network is 3, the convolution step size is 1, and the padding is 1.

9. The method according to claim 7, characterized in that: The activation function is a RELU function.

Images