Driving broadband spectrum data floor noise estimation method based on improved multi-layer perceptron network

Abstract

The invention discloses a driving broadband spectrum data floor noise estimation method based on an improved multi-layer perceptron network. The network comprises a variable step convolution module, a cavity convolution module and a multi-layer perceptron module. Electromagnetic environment data with different frequency band widths are compressed to be 2-3 times of the output size of a multi-layer perceptron module by using a variable step convolution module, so that the receptive field of a feature vector is improved; the cavity convolution module is used for further expanding the receptive field input by the multi-layer perceptron module, so that the condition that the receptive field of the whole pixel is in a large-bandwidth signal due to the existence of the large-bandwidth signal in the environment is avoided, and meanwhile, the exponential increase of the calculation complexity caused by the increase of the size of a convolution kernel is avoided; the characteristic that electromagnetic environment data is flat in a narrow bandwidth is utilized, the maximum bandwidth capable of being collected by collecting equipment is combined, the size of an output layer of the multi-layer perceptron module is reduced, the limitation that the multi-layer perceptron module must be fitted to the original input size is avoided, and the operation time of a network is shortened; and the accuracy of bottom noise estimation is obviously improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing and deep learning, and in particular relates to a method for estimating background noise of driving broadband spectrum data based on an improved multi-layer perceptron network. Background Art

[0002] Noise estimation is a common signal processing operation, the main purpose of which is to improve the accuracy of the detection algorithm. Currently, energy detection is mostly used in engineering. This method usually designs a threshold value based on the estimated noise floor, compares the collected power spectrum, and obtains the starting frequency and ending frequency of the signal. This detection method is simple in principle and highly flexible, but the accuracy of the noise floor estimation result directly affects the accuracy of the detection result. In a broadband environment, noise fluctuations usually have a certain dependence on frequency. How to accurately estimate environmental noise fluctuations is an engineering problem that needs to be solved urgently.

[0003] The patent application (application number: CN202311694662.9) discloses a burst signal detection method based on background noise estimation. According to the preset minimum pulse interval, the start and end times of the background noise estimation are obtained, and then the environmental background noise distribution is calculated according to the noise duration. This method mainly processes burst signals. When there are frequent signals in the electromagnetic environment, the estimation result will be the superposition of environmental noise and frequent signals, and accurate estimation of environmental noise cannot be achieved. The patent application (application number: CN202411114490.8) discloses a radar background noise estimation method, device, terminal and storage medium. This method pre-stores the background noise model sequence corresponding to the radar at all range gates, and then measures the background noise of all range gates to obtain the background noise sequence as the background noise estimation. This method mainly targets radar signals and solves the main problem that when receiving radar echo signals, the receiver receives the leakage signal component generated by the transmitter, which is not suitable for the scenario in this article.

[0004] Although there are some methods for background noise estimation, these methods are limited in application scenarios and cannot meet the requirements of background noise estimation in wide-band data or unknown electromagnetic environments. Their design methods need to be redesigned. Summary of the invention

[0005] To overcome the deficiencies of the above-mentioned existing technologies, the purpose of the present invention is to provide a method for estimating the background noise of driving broadband spectrum data based on an improved multi-layer perceptron network. Through deep learning, the noise power distribution in a complex electromagnetic environment can be accurately estimated. By using a variable-step convolution module, the process of first compressing and then interpolating for upsampling in mainstream deep learning methods is avoided. By utilizing the characteristic that electromagnetic environment data is flat within a narrow bandwidth, the output layer of the multi-layer perceptron network is further reduced by calculating the relationship between the maximum-size input data and the frequency, avoiding the limitation that the multi-layer perceptron must fit to the original input size, and realizing background noise estimation. It can significantly improve the accuracy of background noise estimation and is applicable to the background noise estimation scenario of wide-frequency and large-bandwidth signals in a complex electromagnetic environment.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions:

[0007] An improved multi-layer perceptron network specifically includes:

[0008] A variable-step convolution module that selects convolution kernels with different step sizes for electromagnetic environment data of different output sizes, performs feature extraction, and compresses to obtain a feature vector size that is 2-3 times the output size of the multi-layer perceptron.

[0009] A dilated convolution module that uses dilated convolution to continue feature extraction on the feature vector processed by the variable-step convolution module, enabling the input data of the multi-layer perceptron to expand the model's receptive field without increasing the computational complexity, and further increasing the frequency receptive range of the feature vector.

[0010] A multi-layer perceptron module that uses the input layer to receive the feature vector extracted by the dilated convolution module and transmits it to the hidden layer; uses multiple hidden layers to perform weighted summation and non-linear transformation on the feature data through neurons, layer by layer extracting and combining features; uses the output layer to fit and transform the final hidden layer into the noise fluctuation in a fluctuating electromagnetic environment.

[0011] The variable-step convolution module includes a convolution layer and an activation layer arranged in sequence;

[0012] The dilated convolution module includes dilated convolutions with different dilation rates;

[0013] The multi-layer perceptron module includes an input layer, multiple hidden layers, and an output layer arranged in sequence.

[0014] A method for estimating the background noise of driving broadband spectrum data based on an improved multi-layer perceptron network specifically includes the following steps:

[0015] Step 1: Use equipment to collect or simulate and generate wide-band electromagnetic environment data of different sizes as the dataset for training the improved multi-layer perceptron network and perform preprocessing;

[0016] Step 2: Use the variable-step convolution module in the improved multi-layer perceptron network to extract features from the wide-band electromagnetic environment data of different sizes collected in Step 1. At the same time, compress the data to 2-3 times the output size of the multi-layer perceptron module;

[0017] Step 3: Use the dilated convolution module in the improved multi-layer perceptron network to continue extracting features from the compressed data obtained in Step 2;

[0018] Step 4: Use an adaptive pooling layer to compress the feature vector obtained by feature extraction in Step 3 to the input layer size of the multi-layer perceptron module; then use the multi-layer perceptron module to fit the feature vector to output noise data, and obtain an estimate of the environmental noise fluctuation;

[0019] Step 5: Backpropagation: Calculate the gradient of the error between the noise data output in Step 4 and the real environmental noise data according to the loss function; update the weights and biases of all layers in the improved multi-layer perceptron network according to the gradient of the error function;

[0020] Step 6: Repeat Step 2 to Step 5 until the performance of the improved multi-layer perceptron network reaches the preset threshold or the set number of training epochs is completed, and complete the noise fluctuation estimation.

[0021] The specific method of Step 2 is as follows:

[0022] The step calculation method of the variable-step convolution module is:

[0023]

[0024] where S is the convolution step, N input is the length of the input data, n is the number of convolution blocks, L MLPoutput is the size of the output layer of the multi-layer perceptron, and the calculation formula can be calculated by the following formula:

[0025] L MLPoutput = ceil(N input * F th / F max )

[0026] where F th is the set hyperparameter, which represents the frequency range of a single noise estimate output by the multi-layer perceptron module, and F max is the maximum frequency range that the receiving device can collect;

[0027] Let the convolution kernel size K inside the variable-step convolution module be at least 2 times the maximum convolution step to ensure that all eigenvalues are involved in the calculation, that is

[0028] K ≥ 2×S

[0029] After the data passes through the variable-stride convolution module, a feature vector of the wideband electromagnetic environment data is obtained. Since the stride of the variable-stride convolution module is adaptively adjusted according to the size of the electromagnetic environment data, the size of the obtained feature vector is 2-3 times the size of the output layer of the multi-layer perceptron module.

[0030] The specific method of step 3 is as follows:

[0031] For the dilated convolution module, the dilated convolution kernel size is custom-set, and the dilation rates are 1, 2, 4, and 6. The feature vectors obtained in step 2 are subjected to dilated convolution in parallel using the above-set dilated convolution module.

[0032] The specific method of step 4 is as follows:

[0033] An adaptive pooling layer is used to compress the feature vectors obtained in step 3 to the input size of the multi-layer perceptron module, which is received by the input layer of the multi-layer perceptron module and then passed to the hidden layer of the multi-layer perceptron module. Through multiple hidden layers with non-linear activation functions, weighted summation is performed in each layer and non-linear transformation is performed through the activation function. The formula is as follows:

[0034] Z (l) = W (l) X (l-1) + b (l)

[0035] where Z (l) is the weighted summation result of the l-th layer, W (l) is the weight of this layer, X (l-1) is the output of the previous layer, and b (l) is the bias vector of this layer; then, after being processed by the non-linear activation function, the final output of this layer is obtained:

[0036] X (l) = σ(Z (l) )

[0037] where σ is the general form of the activation function;

[0038] Finally, the output layer maps it to the form of noise fluctuation in the electromagnetic environment by performing weighted summation on the results of the hidden layer and applying the activation function.

[0039] The specific method of step 5 is as follows:

[0040] According to the loss function, obtain the error between the noise data output in step 4 and the real environmental noise data. Then calculate the gradient of the loss function, and update the weights and biases of each layer of the variable-step convolution module, dilated convolution module, and multi-layer perceptron module in the improved multi-layer perceptron network in the direction of minimizing the error to complete the optimization of network parameters.

[0041] A device for estimating the background noise of driving broadband spectrum data based on an improved multi-layer perceptron network specifically includes:

[0042] A memory for storing computer programs;

[0043] A processor for implementing the method for estimating the background noise of driving broadband spectrum data based on the improved multi-layer perceptron network described in steps 1 to 4 when executing the computer program.

[0044] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can perform background noise estimation based on the method for estimating the background noise of driving broadband spectrum data based on the improved multi-layer perceptron network described in any one of steps 1 to 4.

[0045] Compared with the prior art, the present invention has the following advantages:

[0046] 1. The present invention adopts a deep learning method to realize end-to-end wide-band electromagnetic environment noise fluctuation estimation, without the need to have a certain understanding of the environment, without setting the hyperparameters of the network according to the electromagnetic environment, and has higher applicability. When there are frequently occurring signals in the environment, it can also achieve a good background noise estimation effect.

[0047] 2. The present invention adopts a variable-step convolution module to compress electromagnetic environment data with different frequency band widths to 2-3 times the output size of the multi-layer perceptron module, enhancing the receptive field of the feature vector; ensuring the processing rate of input data of different sizes.

[0048] 3. In step 4 of the present invention, by restricting the number of output layers of the multi-layer perceptron module, the noise in the narrow frequency band range is directly approximated as a value, without the need to restore it to the original input data size, reducing the running time of the network.

[0049] 4. Using dilated convolution to further expand the receptive field of the input of the multi-layer perceptron module, avoiding the situation where the receptive field of the entire pixel is within a large bandwidth signal due to the presence of large bandwidth signals in the environment, and at the same time avoiding the exponential growth of computational complexity caused by an increase in the convolution kernel size.

[0050] 5. By taking advantage of the characteristic that the electromagnetic environment data is flat within a narrow bandwidth and combining it with the maximum bandwidth that the acquisition device can collect, the size of the output layer of the multi-layer perceptron module is reduced, avoiding the limitation that the multi-layer perceptron module must be fitted to the original input size, and reducing the running time of the network.

[0051] In summary, the present invention can be applied to the broadband noise power spectrum estimation under all complex electromagnetic environments. By using the ordinary convolution module with variable step size, while ensuring that the model has a sufficiently large receptive field, it also has a high processing rate for large-size input data. The present invention further expands the receptive field through the method of dilated convolution, ensuring that each data point has a sufficient frequency receptive range before being fed into the multi-layer perceptron module for inference. At the same time, according to the characteristic that the noise fluctuates little within the narrow band range, the size of the output layer of the multi-layer perceptron is adjusted, further improving the inference speed of the model. The model structure is simple and the number of parameters is small. The present invention is applicable to the non-cooperative signal scenarios in the electromagnetic environment where the power spectrum sizes collected by the receiver vary greatly and the electromagnetic environment is unknown, and can ensure the same operation efficiency for different input sizes while accurately estimating the electromagnetic environment noise power distribution; it is applicable to any system platform that can carry a deep model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a schematic diagram of the improved multi-layer perceptron network framework of the present invention.

[0053] Figure 2 is a schematic diagram of the annotation result of the dataset in the embodiment of the present invention.

[0054] Figure 3 is a specific parameter diagram of each module in the embodiment of the present invention.

[0055] Figure 4 is an intuitive diagram of the background noise estimation result of the present invention at 20 MHz - 1020 MHz.

[0056] Figure 5 is an intuitive diagram of the background noise estimation result of the present invention in the broadcast frequency band.

[0057] Figure 6 is the diagram of the operator's downlink signal frequency band of the present invention.

[0058] Table 1 is the specific parameters for issuing instructions to the acquisition device in the embodiment of the present invention.

[0059] Table 2 is the performance comparison between the embodiment of the present invention and the traditional method. DETAILED DESCRIPTION OF THE INVENTION

[0060] The present invention will be further described in detail below with reference to the accompanying drawings.

[0061] Embodiment

[0062] As Figure 1 shown, in the embodiments of this article, BnCm is used to label the improved multi-layer perceptron networks of different scales. B is Block, which means a convolutional block, n is the number of convolutional modules, and C is the number of convolutions within each convolutional module. In this embodiment, B2C1 is taken as an example. As Figure 3 shown, there are a total of two variable-stride convolutional modules, and each module has 1 convolutional layer. After the convolutional layer, an activation layer is used to introduce non-linearity. The activation layer contains an activation function. In this embodiment, the most basic ReLU activation function is selected.

[0063] A method for estimating the background noise of driving broadband spectrum data based on an improved multi-layer perceptron network, the specific steps include:

[0064] Step 1: Use the device to collect or simulate and generate wide-band electromagnetic environment data as the dataset for training the improved multi-layer perceptron network, and complete the dataset annotation task. In the present invention, the actual electromagnetic environment data of a certain city in the southwest is used for training, and the process of constructing the dataset is as follows

[0065] ①. Place the electromagnetic environment spectrum acquisition device in an open area and issue an electromagnetic environment background scanning instruction. The background scanning instruction is to rotate the directional antenna of the acquisition device to scan the air. After the antenna rotates one week, save the collected electromagnetic environment power spectrum. The instruction includes the start and end frequencies of 20 - 1020 MHz, the rotation angle of 1 or 3; the frequency resolution is 6250 Hz, and the power spectrum unit is dBμV, as shown in Table 1;

[0066]

[0067]

[0068] Table 1

[0069] ②. Perform mean processing on the power spectrum data at each angle, plot the noise change trend after mean processing in the simulation software, and select some points in the figure as the noise estimation within a certain range as the rough annotation;

[0070] ③. Manually fine-tune the roughly annotated data at each angle, and record the horizontal and vertical coordinates of the annotation points, as Figure 2 shown;

[0071] ④. Linearly interpolate the annotation results to the corresponding data size, and then randomly crop the data and annotations between the minimum and maximum acquisition bandwidths of the acquisition device;

[0072] ⑤. According to the output layer size of the multi-layer perceptron module, equally spaced select L MLPoutput points as the annotation;

[0073] ⑥. Before training, the data needs to be pre - processed. Here, only the input data and labels are normalized, and the normalization scale of the labels should be the same as that of the data.

[0074] Step 2: The step - by - step calculation method of the variable - step convolution module based on the improved multi - layer perceptron network is as follows:

[0075]

[0076] Among them, N input is the length of the input data, n is the number of convolution blocks, and L MLPoutput is the size of the output layer of the multi - layer perceptron module. The calculation formula can be calculated by the following formula:

[0077] L MLPoutput = ceil(N input * F th / F max )

[0078] Among them, F th is the set hyper - parameter, which means the frequency range represented by a single noise estimate output by the multi - layer perceptron module. F max is the maximum frequency range that the receiving device can collect. To ensure that all feature values are involved in the calculation, the convolution kernel size inside the variable - step convolution module is at least twice the maximum convolution step size, that is

[0079] K ≥ 2×S

[0080] Among them, K is the convolution kernel size inside the variable - step convolution module. After the data passes through the variable - step convolution module, the feature vector of the wide - band electromagnetic environment data is obtained. Since the step size of the variable - step convolution module will be adaptively adjusted according to the size of the electromagnetic environment data, the size of the obtained feature vector will be 2 - 3 times the size of the output layer of the multi - layer perceptron module.

[0081] Step 3: For the dilated convolution module based on the improved multi - layer perceptron network, to ensure that the pixels have a large enough receptive field and avoid the generation of network effects, the convolution kernel size is preferably set to 3 or 5. In the present invention, the dilated convolution kernel size is set to 3, and the dilation rates are 1, 2, 4, and 6. (Using a higher dilation rate may cause the dilated convolution to approximate a 1*1 convolution due to the irrelevance of distant pixels, resulting in a large deviation between the final fitting result and the actual situation). Using the above - set dilated convolution module, the feature vectors obtained in Step 2 are processed by dilated convolution in parallel. The improved multi - layer perceptron network can capture features in a wider frequency range, improving the accuracy of noise floor estimation.

[0082] Step 4: Since the multi-layer perceptron module can only accept data with a fixed input size, before the feature vector obtained in Step 3 enters the multi-layer perceptron module, an adaptive pooling layer is used to compress the feature vector obtained in Step 3 to the input size of the multi-layer perceptron module, which is received by the input layer of the multi-layer perceptron module and then passed to the hidden layer of the multi-layer perceptron module. Through multiple hidden layers with non-linear activation functions, the feature vector is weighted and non-linearly transformed layer by layer. The formula is as follows:

[0083] Z (l) = W (l) X (l-1) + b (l)

[0084] where Z (l) is the weighted summation result of the l-th layer, W (l) is the weight of this layer, X (l-1) is the output of the previous layer, and b (l) is the bias vector of this layer; then, after being processed by the non-linear activation function, the final output of this layer is obtained:

[0085] X (l) = σ(Z (l) )

[0086] where σ is the general form of the activation function and can be modified according to needs;

[0087] Finally, the output layer maps it to the form of noise fluctuation in the electromagnetic environment by weighted summation of the results of the hidden layer and applying the activation function.

[0088] Step 5: Backpropagation: The present invention uses the SmoothL1 loss function to calculate the error between the noise data output in Step 4 and the real environmental noise data. The calculation formula is as follows:

[0089]

[0090] where x is the predicted value of the model and y is the actual label value;

[0091] After that, the gradient of the loss function is calculated, and the weights and biases of each layer of the variable-step convolution module, dilated convolution module, and multi-layer perceptron module in the improved multi-layer perceptron network are updated in the direction of minimizing the error to complete the optimization of the network parameters; here, the most basic gradient descent formula is used for updating, and the calculation formula is as follows:

[0092]

[0093] where α is the learning rate, They are the gradients of the loss function L with respect to the weights and biases respectively. The gradient of the error function is updated by 1e-4 each time, and 10 data are used for each training.

[0094] Step 6: Repeat Step 2 to Step 5 until the performance of the improved multi-layer perceptron network reaches a preset threshold or the set number of training epochs is completed. In the present invention, the number of training epochs from Step 2 to Step 5 is set to 100, and the noise fluctuation estimation is completed.

[0095] Simulation experiment:

[0096] The effect of the present invention can be further demonstrated by the following simulation experiment:

[0097] Here, the power spectrum performances of the existing NLR algorithm, median estimation algorithm, and the dilated convolution-multi-layer perceptron-driven wideband noise floor estimation method of the present invention will be analyzed, and the intuitive estimation performance is as Figures 4 - 6 shown. The red line in the figure represents the manually marked result, the blue dashed line represents the estimated result of the algorithm proposed in this paper, the green dashed line represents the estimated result of the NLR algorithm, and the yellow solid line represents the estimated result of the median estimation algorithm. As can be seen from Figure 5 it, when the signal bandwidth in the data is relatively narrow, the estimation performances of the three algorithms are basically the same. However, when there are large-bandwidth signals in the data ( Figure 6 ), the performance of the median estimation algorithm deteriorates rapidly, and the estimated result of the NLR algorithm also has a large deviation from the actual value. Moreover, when the signal-to-noise ratio is low, the estimation effect is not ideal. On the contrary, the dilated convolution-multi-layer perceptron-driven wideband noise floor estimation method can still maintain a high estimation effect.

[0098] Table 2 shows the comparison of the three algorithms in terms of mean square error and time cost. The mean square error uses the manually annotated noise floor power spectrum data as the actual noise floor, and the mean square errors of the three algorithms are calculated through the MSE formula. The MSE calculation formula is shown as follows.

[0099]

[0100] where n is the sample value, y true 、y predThey are the true value and the predicted value respectively. By calculating the square of the error between all predicted values and the corresponding true values, and then averaging the squared errors of all samples, a value representing the magnitude of the error is obtained. The smaller the MSE, the closer the predicted result of the model is to the true value, and the better the fitting effect. The time cost is evaluated by comparing the total time consumed by the three methods to complete a validation set in the Python environment. It can be seen from Table 2 that since the algorithm in this paper is deployed on the GPU for operation, the algorithm efficiency is higher than that of the NLR algorithm, but the time cost is greater than that of the median estimation algorithm; the traditional method is superior to the new algorithm in terms of computational efficiency and can complete the calculation in a shorter time, but its accuracy is lower; although the algorithm proposed in this paper has a certain time cost in calculation, its noise floor estimation accuracy is significantly higher and can provide a more accurate power spectrum estimation result.

[0101] Algorithm name Running time (ms) Mean squared error NLR filtering algorithm 412.7 9.424e-4 Median estimation algorithm 2.8 1.074e-3 B2C1 algorithm used in the implementation scheme 23.1 5.198e-4

[0102] Table 2

Claims

1. An improved multi-layer perceptron network, characterized in that, Specifically include: A variable stride convolution module that selects convolution kernels with different strides for electromagnetic environment data of different output sizes, extracts features, and compresses to obtain a feature vector size that is 2-3 times the output size of the multi-layer perceptron; A dilated convolution module that uses dilated convolution to continue extracting features from the feature vector processed by the variable stride convolution module, enabling the input data of the multi-layer perceptron to expand the model's receptive field without increasing the computational complexity, and further increasing the frequency receptive range of the feature vector; A multi-layer perceptron module that uses the input layer to receive the feature vector extracted by the dilated convolution module and transmits it to the hidden layer; uses multiple hidden layers to perform weighted summation and non-linear transformation on the feature data through neurons, layer by layer extracting and combining features; uses the output layer to fit and transform the final hidden layer into the noise fluctuation in the fluctuating electromagnetic environment.

2. The improved multi-layer perceptron network according to claim 1, characterized in that The variable stride convolution module includes a convolution layer and an activation layer arranged in sequence; The dilated convolution module includes dilated convolutions with different dilation rates; The multi-layer perceptron module includes an input layer, multiple hidden layers, and an output layer arranged in sequence.

3. A method for estimating the background noise of driving broadband spectrum data based on an improved multi-layer perceptron network, characterized in that, Specifically include the following steps: Step 1: Use equipment to collect or simulate and generate wide-band electromagnetic environment data of different sizes as the dataset for training the improved multi-layer perceptron network and preprocess it; Step 2: Use the variable stride convolution module in the improved multi-layer perceptron network to extract features from the wide-band electromagnetic environment data of different sizes collected in Step 1, and at the same time, compress the data to 2-3 times the output size of the multi-layer perceptron module; Step 3: Use the dilated convolution module in the improved multi-layer perceptron network to continue extracting features from the compressed data obtained in Step 2; Step 4: Use an adaptive pooling layer to compress the feature vector obtained by feature extraction in Step 3 to the input layer size of the multi-layer perceptron module; then use the multi-layer perceptron module to fit and output noise data to obtain an estimate of the environmental noise fluctuation; Step 5: Backpropagation: Calculate the gradient of the error between the noise data output in Step 4 and the real environmental noise data according to the loss function; update the weights and biases of all layers in the improved multi-layer perceptron network according to the gradient of the error function; Step 6: Repeat Step 2 to Step 5 until the performance of the improved multi-layer perceptron network reaches a preset threshold or completes the set number of training epochs, and complete the noise fluctuation estimation.

4. The method for estimating the background noise of drive broadband spectrum data based on the improved multi-layer perceptron network according to claim 3, wherein The specific method of Step 2 is: The step calculation method of the variable stride convolution module is: Among them, S is the convolution stride, N input is the length of the input data, n is the number of convolution blocks, L MLPoutput is the size of the output layer of the multi-layer perceptron, and the calculation formula can be calculated by the following formula: L MLPoutput = ceil(N input * F th / F max ) Among them, F th is a set hyperparameter, representing the frequency range of a single noise estimation value output by the multi-layer perceptron module, and F max is the maximum frequency range that the receiving device can collect; Let the convolution kernel size K inside the variable stride convolution module be at least 2 times the maximum convolution stride to ensure that all eigenvalues are involved in the calculation, that is K≥2×S After the data passes through the variable stride convolution module, a feature vector of the wide-band electromagnetic environment data is obtained; since the stride of the variable stride convolution module will be adaptively adjusted according to the size of the electromagnetic environment data, the obtained feature vector size is 2-3 times the output layer size of the multi-layer perceptron module.

5. The method for estimating the background noise of drive broadband spectrum data based on the improved multi-layer perceptron network according to claim 3, wherein The specific method of Step 3 is: For the dilated convolution module, the size of the dilated convolution kernel is custom - set, and the dilation rates are 1, 2, 4, and 6. The feature vectors obtained in step 2 are subjected to dilated convolution in parallel using the dilated convolution module set as above.

6. The method for estimating the background noise of drive broadband spectrum data based on the improved multi-layer perceptron network according to claim 3, wherein The specific method of step 4 is as follows: Use an adaptive pooling layer to compress the feature vectors obtained in step 3 to the input size of the multi - layer perceptron module, which is received by the input layer of the multi - layer perceptron module and then passed to the hidden layer of the multi - layer perceptron module. Through multiple hidden layers with non - linear activation functions, weighted summation is performed in each layer and non - linear transformation is carried out through the activation function. The formula is as follows: Z (l) = W (l) X (l-1) + b (l) Among them, Z (l) is the weighted sum result of the l-th layer, W (l) is the weight of this layer, X (l-1) is the output of the previous layer, b (l) is the bias vector of this layer; then, after being processed by a non-linear activation function, the final output of this layer is obtained: X (l) = σ(Z (l) ) where σ is the general form of the activation function; Finally, the output layer maps it to the form of noise fluctuation in the electromagnetic environment by performing weighted summation on the results of the hidden layer and applying the activation function.

7. The method for estimating the background noise of driving broadband spectrum data based on the improved multi-layer perceptron network according to claim 3, characterized in that The specific method of step 5 is as follows: According to the loss function, obtain the error between the noise data output in step 4 and the real - environment noise data, and then calculate the gradient of the loss function. Update the weights and biases of each layer of the variable - step convolution module, dilated convolution module, and multi - layer perceptron module in the improved multi - layer perceptron network in the direction of minimizing the error to complete the optimization of network parameters.

8. A device for estimating the background noise of drive broadband spectrum data based on an improved multi-layer perceptron network, characterized in that Specifically, it includes: A memory for storing a computer program; A processor for implementing the method for estimating the background noise of drive broadband spectrum data based on the improved multi - layer perceptron network according to claims 1 to 7 when executing the computer program.

9. A computer - readable storage medium storing a computer program, which can perform background noise estimation based on the method for estimating the background noise of drive broadband spectrum data based on the improved multi - layer perceptron network according to any one of claims 1 to 7 when the computer program is executed by a processor.

Images