Intelligent improvement method for electromagnetic radiation monitoring of new energy automobile

Abstract

The invention discloses an intelligent improvement method for electromagnetic radiation monitoring of a new energy automobile. The intelligent improvement method comprises the following steps: firstly, constructing a gating dynamic neural network consisting of two input layers, two gating networks, three sharing units, two weighted summation layers, two feature separation layers and two output layers; training the gating dynamic neural network by using an input electric field signal of electromagnetic radiation of the new energy automobile generated by simulation and an output electric field signal acquired by an electric field sensor to obtain a nonlinear correction model; and then deploying the nonlinear correction model at the output end of the electric field sensor, and correcting and compensating the electric field signal of the electromagnetic radiation of the to-be-detected new energy automobile collected by the electric field sensor through the nonlinear correction model, thereby completing intelligent improvement of the electromagnetic radiation monitoring of the new energy automobile. According to the invention, a large dynamic range signal in new energy automobile electromagnetic radiation monitoring can be intelligently processed, a measurement error caused by nonlinear distortion of the sensor can be effectively identified and compensated, and the accuracy of a measurement result can be significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electromagnetic radiation monitoring, and in particular to an intelligent improvement method for electromagnetic radiation monitoring of new energy vehicles. Background Art

[0002] In the electromagnetic radiation monitoring of new energy vehicles, signals with a large dynamic range will be affected by various nonlinear distortions when collected by electric field sensors. This nonlinear distortion may lead to inaccurate signal measurement results (see the sensor input and output power curves in 1), which in turn affects the evaluation of system performance. In order to reduce the impact of nonlinear distortion on electric field sensors during acquisition, the linearization technologies currently used in related fields mainly include: First, from the hardware perspective, by optimizing the materials and processes of electric field sensors, or reasonably designing the electrode structure of electric field sensors, to optimize the electric field distribution, try to make the electric field evenly distributed inside the electric field sensor, and reduce the impact of nonlinear effects, but this way of improving hardware has the problem of greater difficulty in improvement and limited improvement effect. Second, using the direct linearization method, by selecting a straight line with a smaller fitting error to replace the actual nonlinear curve, if the actual curve is divided into several segments, each segment is replaced by a straight line, then a set of broken lines can be formed that are closer to the characteristic curve of the sensor, but this direct linearization method is relatively complex and has a large error. Summary of the invention

[0003] The present invention aims to solve the problem of nonlinear distortion caused by the electric field sensor itself and the large dynamic range signal characteristics it measures in the electromagnetic radiation monitoring of new energy vehicles, and provides an intelligent improvement method for electromagnetic radiation monitoring of new energy vehicles.

[0004] To solve the above problems, the present invention is achieved through the following technical solutions:

[0005] An intelligent improvement method for electromagnetic radiation monitoring of new energy vehicles includes the following steps:

[0006] Step 1, construct a gated dynamic neural network that describes the input and output relationship of the electric field sensor; the gated dynamic neural network consists of 2 input layers, 2 gated networks, 3 shared units, 2 weighted summation layers, 2 feature separation layers and 2 output layers; the input of the first input layer is used as the input of the first input vector, and the output of the first input layer is simultaneously connected to the input of the two gated networks; the input of the second input layer is used as the output of the second input vector, and the output of the second input layer is simultaneously connected to the input of the three shared units; the output of the first gated network and the output of the three shared units are connected to the input of the first feature separation layer, the output of the first feature separation layer is connected to the input of the first output layer, and the output of the first output layer is used as the input of the first output vector; the output of the second gated network and the output of the three shared units are connected to the input of the second feature separation layer, the output of the second feature separation layer is connected to the input of the second output layer, and the output of the second output layer is used as the input of the second output vector;

[0007] Step 2: First, simulate and generate the electric field signal of the electromagnetic radiation of the new energy vehicle as the input electric field signal of the electric field sensor, then use the electric field sensor to collect the input electric field signal to obtain the output electric field signal of the electric field sensor, and then construct an electric field signal sample data set based on these input electric field signals and output electric field signals;

[0008] Step 3: Process each group of input electric field signals and output signals in the electric field signal sample data set obtained in step 2 to obtain input vectors and output vectors, and construct a training sample data set based on these input vectors and output vectors;

[0009] Step 4: Use the training sample data set obtained in step 3 to perform deep learning on the gated dynamic neural network constructed in step 1 to obtain a learned gated dynamic neural network, and use it as a nonlinear correction model for the electric field sensor;

[0010] Step 5, deploy the nonlinear correction model obtained in step 4 at the output end of the electric field sensor; then, use the electric field sensor to collect the electric field signal of the electromagnetic radiation of the new energy vehicle to be tested, and obtain the input vector after processing the collected electric field signal, and send the input vector to the nonlinear correction model to obtain the output vector, and use the nonlinear correction model to correct and compensate the electric field signal collected by the electric field sensor, thereby completing the intelligent improvement of electromagnetic radiation monitoring of new energy vehicles.

[0011] In step 1, each shared unit is composed of a convolutional layer and a Tanh activation layer. The input of the convolutional layer forms the input of the shared unit, the output of the convolutional layer is connected to the input of the Tanh activation layer, and the output of the Tanh activation layer forms the output of the shared unit.

[0012] In step 1, the gating network is composed of multi-layer perceptrons.

[0013] In step 3, the input vector is constructed based on the input electric field signal and includes a first input vector and a second input vector; the first input vector is composed of the normalized power level PL, bandwidth BW, peak-to-average power ratio PAPR and adjacent channel power ratio ACPR of the input electric field signal; the second input vector is composed of the in-phase component I in (n), orthogonal component Q in (n), magnitude |x(n)|, magnitude squared |x(n) 2 and the magnitude cubed |x(n) 3 The output vector is constructed based on the output electric field signal and includes a first output vector and a second output vector; the first output vector is composed of a probability label y of the offset of the power curve of the output electric field signal true The second output vector is composed of the in-phase component I of the output electric field signal out (n) and the orthogonal component Q out (n) Composition.

[0014] Compared with the prior art, the present invention has the following characteristics:

[0015] 1. Intelligently improve measurement accuracy: Through the nonlinear correction model of the electric field sensor based on the gated dynamic neural network, it can intelligently process the large dynamic range signal in the electromagnetic radiation monitoring of new energy vehicles, effectively identify and compensate for the measurement error caused by the nonlinear distortion of the sensor, and significantly improve the accuracy of the measurement results.

[0016] 2. Multi-task processing capability: The nonlinear correction model of the electric field sensor based on the gated dynamic neural network is a multi-task model, which can perform two key tasks simultaneously: one is to identify the position of the input signal on the power cosine curve, and the other is to compensate and calibrate the distorted signal. This multi-task processing capability improves the overall efficiency and flexibility of the system.

[0017] 3. Adaptation and dynamic adjustment: The nonlinear correction model of the electric field sensor based on the gated dynamic neural network can dynamically adjust the processing strategy according to the characteristics of different input signals, achieve adaptive response to changes in complex electromagnetic environment, and enhance the robustness and reliability of the system.

[0018] 4. Improve the ability to process large dynamic range signals: It provides an effective solution for the large dynamic range signals commonly seen in electromagnetic radiation monitoring of new energy vehicles, which can accurately distinguish and process strong and weak signals, avoid spectrum distortion, and ensure the comprehensiveness and accuracy of the measurement results. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is the input and output power curve of the electric field sensor.

[0020] Figure 2 This is the principle structure diagram of the gated dynamic neural network.

[0021] Figure 3 This is the principle structure diagram of the shared unit.

[0022] Figure 4 This is the principle structure diagram of the gating network. DETAILED DESCRIPTION

[0023] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in combination with specific examples and with reference to the accompanying drawings.

[0024] In order to solve the problem of nonlinear distortion when the electric field sensor collects signals with a large dynamic range in the electromagnetic radiation monitoring of new energy vehicles, the present invention intends to add a nonlinear correction algorithm (model) to the output end of the electric field sensor, and use digital signal processing technology to correct and compensate the signal output by the electric field sensor to obtain a real electric field signal, so as to reduce the influence of nonlinear distortion on the measurement results and improve the accuracy of electromagnetic radiation monitoring of new energy vehicles.

[0025] To this end, the present invention proposes an intelligent improvement method for electromagnetic radiation monitoring of new energy vehicles, comprising the following steps:

[0026] Step 1: Construct a gated dynamic neural network that describes the input and output relationship of the electric field sensor.

[0027] The key to nonlinear correction of electric field sensors is to establish a multi-task model that describes the input and output relationship of the electric field sensor. The gated dynamic neural network multi-task model proposed in the present invention is a gated dynamic neural network, such as Figure 2As shown, the gated dynamic neural network consists of 2 input layers, 2 gated networks, 3 shared units, 2 weighted summation layers, 2 feature separation layers and 2 output layers. The input of the first input layer is used as the input of the first input vector, and the output of the first input layer is simultaneously connected to the input of the two gated networks. The input of the second input layer is used as the output of the second input vector, and the output of the second input layer is simultaneously connected to the input of the three shared units. The output of the first gated network and the output of the three shared units are connected to the input of the first feature separation layer, the output of the first feature separation layer is connected to the input of the first output layer, and the output of the first output layer is used as the input of the first output vector. The output of the second gated network and the output of the three shared units are connected to the input of the second feature separation layer, the output of the second feature separation layer is connected to the input of the second output layer, and the output of the second output layer is used as the input of the second output vector. The first input of the gated dynamic neural network is the characteristic information of the input electric field signal, and the second input is the IQ component and amplitude information of the electric field signal. The first output result of the gated dynamic neural network is used to identify the offset probability of the power curve of the electric field signal collected by the electric field sensor (the left and right offset position with the power peak point of the electric field signal as the axis), and the second output result compensates and calibrates the electric field signal collected by the electric field sensor to obtain a real electric field signal.

[0028] 1) The shared main network (3 shared units) is used to extract the features of the input data. The connection mode of the 3 shared units is parallel connection, and its output can be expressed as

[0029] Each shared unit consists of a convolutional layer and a Tanh activation layer, such as Figure 3 As shown, the input of the convolutional layer forms the input of the shared unit, the output of the convolutional layer is connected to the input of the Tanh activation layer, and the output of the Tanh activation layer forms the output of the shared unit.

[0030] The convolutional layer takes the input data (the second input vector X 2 ) is convolved with multiple local convolution kernels and bias parameters are added to generate corresponding local features. The convolution operation is expressed as:

[0031]

[0032] In the formula, h l represents the convolution output of the lth convolution kernel, X 2 represents the input data, i.e. the second input vector, represents the lth convolution kernel coefficient; l = 1, 2, ..., L, L represents the number of convolution kernels, Represents the convolution operation.

[0033] The output of the activation function of the convolution kernel is:

[0034]

[0035] In the formula, u l represents the output feature map of the lth convolution kernel, represents the bias of the lth convolution kernel, f c (·) is the activation function of the convolution kernel.

[0036] Through the convolution layer composed of convolution kernels, the basis function features required for modeling are generated, and the basis function features are arranged into feature vectors. The feature vector m is written as:

[0037] m=[m 1 ,m 2 ...,m L×B×C ] T =[u 111 ,u 112 ,…,u 211 ,…,u LBC ] T

[0038] Where m is a vector of L×B×C and the dimension of the feature map is B×C.

[0039] The Tanh activation layer maps the features of the convolutional layer output to the output result. The output of the Tanh activation layer can be expressed as:

[0040] h FC =Tanh(w 1 m+b 1 )

[0041] In the formula, w 1 represents the weight coefficient matrix of the Tanh activation layer, with dimensions K×L×B×C, where K is the number of neurons in the Tanh activation layer; b 1 represents the bias coefficient vector of the Tanh activation layer, with a dimension of K×1; m represents the feature vector output by the convolutional layer, and Tanh(·) represents the Tanh function.

[0042] 2) The gating network is used to understand the changing patterns of the behavior of the electric field sensors and further adjust the shared units according to the transmission configuration during the network operation phase.

[0043] Each gating network is composed of a multi-layer perceptron (MLP), such as Figure 4 As shown in Figure 1, the multilayer perceptron consists of an input layer, two hidden layers, and an output layer.

[0044] The input of both gating networks is the first input vector X 1 , the two gating networks each output a set of weights, and these two sets of weight vectors can be used to control different activation levels. The weight g output by the gating network is expressed as:

[0045] g=Sigmoid(W g ·x t +b g )

[0046] Where W g represents the weight, b g represents the bias, x t represents the first input vector, and sigmoid(·) represents the Sigmoid function.

[0047] The weights output by the first gating network select and activate different modules or paths of the shared unit, and can be used to decide which sub-parts of the shared unit (such as specific layers or branches) should be activated. Determine which shared unit modules will be activated, that is, when the weight corresponding to the shared unit The modules that exceed a certain threshold value will be activated and used, and through normalization, the sum of the weights of all shared units is made 1, which is finally used to adjust the output weight of the shared unit. The weight of the i-th shared unit output by the first gating network is for:

[0048]

[0049] In the formula, represents the weighted value of the i-th shared unit, Output of the i-th shared unit, i=1,2,3.

[0050] The weights of the second gating network output are further adjusted or weighted within the activated module, for example, to control the activation level of a specific neuron, or to perform a weighted combination between different modules. Acting on the inside of the activated shared unit, further adjusting or weighting the output. This set of weights can control the neuron activation and output weighting inside each activated shared unit at a finer granularity, ensuring optimal feature extraction and processing. The weight of the jth neuron of the ith shared unit output by the first gating network is for:

[0051]

[0052] In the formula, represents the weighted value of the jth neuron in the ith shared unit, represents the output of the jth neuron of the ith shared unit.

[0053] The gated network selects the signal ratio of each shared unit so that each shared unit has its own learning direction, and finally works together on multiple tasks. During the model training process, the gated network needs to learn how to adjust the main network (shared unit) according to the transmitted information. The more complex the dynamic scene, the stronger the learning ability required by the gated network. In addition, because shared units can enhance parameter sharing, models of multiple targets can be trained jointly, thus reducing the parameter scale of the model and preventing model overfitting.

[0054] The weighted sum layer combines the features of the shared unit outputs under the control of the weights output by the gating network. The output of the weighted sum layer is expressed as:

[0055] y final =Σg·h FC

[0056] In the formula, h FC is the output of the shared unit, g is the output weight of the gating network, y final It is the control output of the gating network and the shared unit.

[0057] The feature separation layer is used to process different features separately. The input features are divided into multiple subsets through different gating networks, and each subset is processed by the feature separation layer, so that different types of features can be processed and modeled in different ways. Each feature separation layer consists of 2 fully connected layers.

[0058] The first feature separation layer is used to output the area recognition of the power curve, and the position of the cosine area where the current input signal is located can be output through the control of the first gating network. The first feature separation layer outputs a single neuron for area judgment of the power curve, and the output is the curve offset probability.

[0059] The first feature separation layer first passes the result of linear transformation through the sigmoid activation function to obtain the probability of each category, and then uses the nonlinear activation function as the output of region recognition. The output of region recognition can be obtained as follows:

[0060]

[0061] In the formula, y final,1 represents the input of the first feature separation layer, W represents the weight of the first feature separation layer, b represents the bias of the first feature separation layer, y task-1 is the output of the first feature separation layer. task-1It is a 1×2 vector, which indicates the probability that the input signal power belongs to the left and right of the cosine curve (with the peak value of the cosine curve as the boundary), corresponding to the two categories of the left and right of the cosine curve. The first element (probability value) indicates the probability that the input signal belongs to the left side of the cosine curve. The second element (probability value) indicates the probability that the input signal belongs to the right side of the cosine curve. Based on this probability vector, a judgment can be made: if the left value is greater than the right value, the input signal belongs to the left side of the cosine curve. Otherwise, the input signal belongs to the right side of the cosine curve.

[0062] The second feature separation layer calibrates the input signal, and the calibrated signal reduces the distortion of amplitude and phase to a certain extent. The dimensions of the in-phase component and orthogonal component output by the second feature separation layer are the same as the dimensions of the in-phase component and orthogonal component of the input, and the final output is a calibrated electric field signal.

[0063] The second feature separation layer obtains nonlinear fitting ability through nonlinear activation function to correct nonlinear cosine. In order to ensure the continuous value of output data, we use the activation function f o Set to linear function y=x to adjust.

[0064]

[0065] In the formula, I out (n) and Q out (n) represents the in-phase component and the orthogonal component of the in-phase component output by the second feature separation layer, y final,2 represents the input of the second feature separation layer, represents the weight of the second feature separation layer, Represents the bias of the second characteristic separation layer.

[0066] In view of the fact that common multi-task network joint training leads to the phenomenon of negative transfer between irrelevant or even conflicting tasks, the final effect is affected. The gated dynamic neural network constructed by the present invention sets a gated network for each task separately, retains the output layer of each specific task, and shares the hidden layer (shared unit) between all tasks. Such a model can obtain different weights of the shared main network for different tasks, so as to realize the selective use of the shared main network, and the gated networks corresponding to different tasks can learn different shared main network combination modes, so the model is more likely to capture the correlation and difference between subtasks. By sharing the network at the bottom layer, it can play a role of complementing each other's learning. The higher the correlation between the tasks in the upper layer, the more sufficient the learning of the model will be. In addition, the more tasks are learned at the same time, the model can capture the same representation of more tasks, resulting in a smaller risk of overfitting on the original task.

[0067] Step 2: First simulate and generate the electric field signal of the electromagnetic radiation of the new energy vehicle as the input electric field signal of the electric field sensor, then use the electric field sensor to collect the input electric field signal to obtain the output electric field signal of the electric field sensor, and then construct an electric field signal sample data set based on these input electric field signals and output electric field signals.

[0068] The electric field signal of the electromagnetic radiation of new energy vehicles is simulated by simulation software or signal generator. The simulated electric field signal not only needs to include the electric field signal of the electromagnetic radiation of new energy vehicles in different states and times, to ensure that the electric field signal sample data contains the distortion caused by sensor nonlinearity and environmental factors; but also needs to ensure that its power covers the entire dynamic range of the electromagnetic radiation of new energy vehicles, to ensure that the electric field signal sample data includes the nonlinear distortion part (the power curve is a cosine curve), to ensure data diversity and improve the subsequent model training effect.

[0069] Step 3: Process each group of input electric field signals and output signals in the electric field signal sample data set obtained in step 2 to obtain a group of input vectors and output vectors, wherein the input vectors are constructed based on the input electric field signals, and the output vectors are constructed based on the output electric field signals, and a training sample data set is constructed based on these input vectors and output vectors.

[0070] Each set of input vectors includes a first input vector X 1 and the second input vector X 2 .

[0071] The first input vector X 1 It is composed of the normalized power level PL of the input electric field signal, the bandwidth BW, the peak-to-average power ratio PAPR and the adjacent channel power ratio ACPR, that is:

[0072] X 1 =[PL BW PAPR ACPR]

[0073] The second input vector X 2 The in-phase component I of the input electric field signal in (n), orthogonal component Q in (n), magnitude |x(n)|, magnitude squared |x(n) 2 and the magnitude cubed |x(n) 3 Composition, namely:

[0074] X 1 =[I in (n) Q in (n) |x(n)| |x(n)| 2 |x(n)| 3 ]

[0075] Each group of output vectors includes a first output vector Y1 and the second output vector Y 2 .

[0076] The first output vector Y 1 The offset probability label y of the power curve of the output electric field signal true Composition, namely:

[0077] Y 1 =[y true ]

[0078] The second output vector Y 2 The in-phase component I of the output electric field signal out (n) and the orthogonal component Q out (n) constitute, that is:

[0079] Y 1 =[I out (n)Q out (n)]

[0080] Step 4: Use the training sample data set obtained in step 3 to perform deep learning on the gated dynamic neural network constructed in step 1 to obtain a learned gated dynamic neural network, and use it as a nonlinear correction model for the electric field sensor.

[0081] The training sample data set includes several training samples, each set of samples includes an input vector (X 1 and X 2 ), and the corresponding output vector (Y 1 and Y 2 ). The training samples in the training sample data set are divided into a training set and a validation set according to a certain ratio, for example, 70% is used for training and 30% is used for validation.

[0082] Model training and verification are performed based on the divided training set and verification set to obtain a learned gated dynamic neural network.

[0083] The goal of the first task of the gated dynamic neural network is to minimize the difference between the predicted first output vector y and the actual first output vector y output by the gated dynamic neural network. true In the present invention, the first output vector y is compared with the actual first output vector y true Represents a one-hot encoded vector, for example, [1,0] represents category 1, which is the left half of the cosine, and [0,1] represents category 2, which is the right half of the cosine. The first task uses the cross loss function to calculate the difference between the predicted value and the true value:

[0084]

[0085] The goal of the second task of the gated dynamic neural network is to minimize the predicted second output vector I of the gated dynamic neural network output out (n) and Q out (n) and the actual second output vector I in (n) and Q in (n). The second task uses the mean square error loss function E mse (θ) Calculate the difference between the predicted value and the true value:

[0086]

[0087] The gated dynamic neural network is trained using the Adam optimization algorithm, and the model parameters are updated until the loss function converges to the preset threshold. Finally, the model is verified and the Adam optimization algorithm is used to train and update the model parameters until the loss function converges to the preset threshold.

[0088] Step 5: deploy the nonlinear correction model obtained in step 4 at the output end of the electric field sensor. After the deployment is completed, the electric field sensor is used to collect the electric field signal of the electromagnetic radiation of the new energy vehicle to be tested, and the collected electric field signal is processed to obtain an input vector, and the input vector is sent to the nonlinear correction model to obtain an output vector, and the nonlinear correction model is used to correct and compensate the electric field signal collected by the electric field sensor, thereby completing the intelligent improvement of electromagnetic radiation monitoring of new energy vehicles.

[0089] The input vector of the nonlinear correction model is regarded as the information of the electric field signal before correction, and the output vector of the nonlinear correction model is regarded as the information of the electric field signal after correction. The subsequent processing of the output vector can be analyzed to obtain the data characteristics of the measured electric field signal. The frequency spectrum of the electric field signal can be observed through FFT, and the accurate size of the measured radiation field strength can be calculated to realize the electromagnetic radiation monitoring of new energy vehicles.

[0090] It should be noted that although the embodiments of the present invention described above are illustrative, they are not intended to limit the present invention, and therefore the present invention is not limited to the above specific embodiments. Without departing from the principles of the present invention, any other embodiments obtained by those skilled in the art under the guidance of the present invention are deemed to be within the protection of the present invention.

Claims

1. An intelligent improvement method for electromagnetic radiation monitoring of new energy vehicles, characterized in that: The steps include: Step 1, construct a gated dynamic neural network that describes the input and output relationship of the electric field sensor; the gated dynamic neural network consists of 2 input layers, 2 gated networks, 3 shared units, 2 weighted summation layers, 2 feature separation layers and 2 output layers; the input of the first input layer is used as the input of the first input vector, and the output of the first input layer is simultaneously connected to the input of the two gated networks; the input of the second input layer is used as the output of the second input vector, and the output of the second input layer is simultaneously connected to the input of the three shared units; the output of the first gated network and the output of the three shared units are connected to the input of the first feature separation layer, the output of the first feature separation layer is connected to the input of the first output layer, and the output of the first output layer is used as the input of the first output vector; the output of the second gated network and the output of the three shared units are connected to the input of the second feature separation layer, the output of the second feature separation layer is connected to the input of the second output layer, and the output of the second output layer is used as the input of the second output vector; Step 2: First, simulate and generate the electric field signal of the electromagnetic radiation of the new energy vehicle as the input electric field signal of the electric field sensor, then use the electric field sensor to collect the input electric field signal to obtain the output electric field signal of the electric field sensor, and then construct an electric field signal sample data set based on these input electric field signals and output electric field signals; Step 3: Process each group of input electric field signals and output signals in the electric field signal sample data set obtained in step 2 to obtain input vectors and output vectors, and construct a training sample data set based on these input vectors and output vectors; Step 4: Use the training sample data set obtained in step 3 to perform deep learning on the gated dynamic neural network constructed in step 1 to obtain a learned gated dynamic neural network, and use it as a nonlinear correction model for the electric field sensor; Step 5, deploy the nonlinear correction model obtained in step 4 at the output end of the electric field sensor; then, use the electric field sensor to collect the electric field signal of the electromagnetic radiation of the new energy vehicle to be tested, and obtain the input vector after processing the collected electric field signal, and send the input vector to the nonlinear correction model to obtain the output vector, and use the nonlinear correction model to correct and compensate the electric field signal collected by the electric field sensor, thereby completing the intelligent improvement of electromagnetic radiation monitoring of new energy vehicles.

2. According to claim 1, the intelligent improvement method for electromagnetic radiation monitoring of new energy vehicles is characterized in that: In step 1, each shared unit is composed of a convolutional layer and a Tanh activation layer. The input of the convolutional layer forms the input of the shared unit, the output of the convolutional layer is connected to the input of the Tanh activation layer, and the output of the Tanh activation layer forms the output of the shared unit.

3. According to claim 1, the intelligent improvement method for electromagnetic radiation monitoring of new energy vehicles is characterized in that: In step 1, the gating network is composed of multi-layer perceptrons.

4. According to claim 1, the intelligent improvement method for electromagnetic radiation monitoring of new energy vehicles is characterized in that: In step 3, The input vector is constructed based on the input electric field signal and includes a first input vector and a second input vector; The first input vector is composed of the normalized power level PL, bandwidth BW, peak-to-average power ratio PAPR and adjacent channel power ratio ACPR of the input electric field signal; the second input vector is composed of the in-phase component I in (n), orthogonal component Q in (n), magnitude |x(n)|, magnitude squared |x(n) 2 and the magnitude cubed |x(n) 3 constitute; The output vector is constructed based on the output electric field signal and includes a first output vector and a second output vector; The first output vector consists of the offset probability label y of the power curve of the output electric field signal true The second output vector is composed of the in-phase component I of the output electric field signal out (n) and the orthogonal component Q out (n) Composition.

Images