WGAN-based intelligent noise reduction method for signal analyzer

Abstract

The invention discloses a WGAN-based intelligent noise reduction method for a signal analyzer, and belongs to the technical field of testing. The signal analyzer intelligent noise reduction network based on the WGAN can generate noise data almost completely matched with a real noise floor in real time according to the current state of an instrument, and compared with an attenuator shielding mode used at present, the signal analyzer intelligent noise reduction network based on the WGAN has the main advantages that the real-time performance is high, when the state of the instrument is changed, it is not needed to wait for re-measurement of the noise floor, and the cost is low. The measurement efficiency is greatly improved; a penalty mechanism implemented in training enables a generated noise floor result to be almost completely matched with a real noise floor within a threshold range, thereby ensuring that the variance of a difference value between a noise floor generation value and an actual value of each frequency point in the process of processing unit conversion by the signal analyzer is not too large, and true numbers of logarithms are not negative. Therefore, the problems of point falling and difficult mathematical calculation in the measurement result are avoided, the measurement accuracy is effectively improved, and the intelligent noise reduction of the signal analyzer is realized.

Description

Technical Field

[0001] The present invention belongs to the field of testing technology, and in particular relates to a WGAN-based intelligent noise reduction method for a signal analyzer. Background Art

[0002] The noise of the signal analyzer greatly affects the accuracy and sensitivity of the instrument's measurement. At present, the sources of noise are mainly internal and external. The internal noise sources mainly include a) the inherent noise and nonlinear spurious of electronic components such as amplifiers and mixers; b) thermal noise, flicker noise and device noise that affect the phase noise caused by the oscillator; and c) the resolution bandwidth (RBW) that can affect the internal noise display level and noise floor height of the signal analyzer. The external noise sources mainly come from the interference and noise caused by ambient temperature, vibration, etc. on the signal path.

[0003] Traditional solutions mainly rely on optimizing hardware devices and adjusting circuits according to device coupling relationships to reduce noise generation. However, considering the complex mechanism of noise generation, especially the objective existence of thermal noise, it is very difficult to completely eliminate noise by this means. Microwave device modeling with the help of computers is an important aspect of microwave circuit design optimization, and deep learning has received increasing attention due to its powerful fitting ability for nonlinear curves.

[0004] Ideally, we want to have a curve with exactly the same amplitude as the noise, subtract it numerically, and thus achieve the noise reduction function of the signal analyzer. Based on this idea, we chose the generative WGAN network as a simulator of the noise floor. After learning the distribution and changes of noise under various conditions, the WGAN network can accurately generate the noise floor according to the current state of the instrument, cancel the noise, and achieve intelligent noise reduction.

[0005] The existing noise elimination method is achieved by using an attenuator. When the instrument is in a certain state and the attenuator is set to the maximum value, we believe that the noise at this time is to shield all external input noise, that is, the inherent noise of the instrument in the current state, that is, the noise floor of the signal analyzer. When performing test measurements, if the state of the instrument does not change, the pre-measured noise floor is subtracted at the numerical level to obtain the noise suppression result. There are two main problems with this method. The first problem is that if the instrument state changes, such as RBW or bandwidth changes, then the noise floor needs to be re-measured. This greatly increases the time cost and seriously affects the real-time performance of the measurement. The second problem is that when using this method to reduce noise, when performing noise floor elimination calculations, the measurement result unit needs to be converted from watts (W) to dBm. Logarithmic operation, which will further lead to two problems. First, the pre-measured noise floor is not the real-time noise floor, and the difference between the actual value of the noise floor and the pre-measured recorded value often has a large variance. Since the difference itself is small, the difference is further amplified after taking the logarithm, causing the logarithmic result of the frequency points with a small difference to approach negative infinity, which is ultimately reflected in the measurement results as "drop points" at these frequency points, causing serious jitter in the overall measurement results. Second, if the actual value of the noise floor is less than the pre-measured recorded value, the logarithm of a negative number will be taken after subtraction, which is obviously mathematically incalculable and affects the measurement results. Therefore, both of the above problems will lead to inaccurate measurement results and inability to achieve effective noise reduction. Summary of the invention

[0006] In view of the above technical problems existing in the prior art, the present invention proposes a signal analyzer intelligent noise reduction method based on WGAN, which has a reasonable design, overcomes the shortcomings of the prior art, and has good effects.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A signal analyzer intelligent noise reduction method based on WGAN, comprising the following steps:

[0009] Step 1: Design the generator and discriminator;

[0010] Step 2: Data processing;

[0011] Step 3: Initialize network weights.

[0012] Preferably, step 1 specifically includes the following steps:

[0013] Step 1.1: Design a generator; the generator accepts random numbers that follow a Gaussian distribution and maps them to the target data space;

[0014] Step 1.2: Design a discriminator; the discriminator is used to discriminate the data input to it, and its purpose is to discriminate the data from the real data set as true, and the data from the generator as false;

[0015] Step 1.3: Iterative optimization; when optimizing the generator, the goal is to make the samples generated by the generator gradually approach the real data; when optimizing the discriminator, the goal is to enable the discriminator to accurately determine the authenticity of the data; repeated iterations and mutual feedback allow the data finally generated by the network to closely simulate the real data.

[0016] Preferably, step 2 specifically includes the following steps:

[0017] Step 2.1: Before training, each set of data is processed using maximum and minimum normalization;

[0018] Step 2.2: In each iteration of training, WGAN is used to generate 25 sets of noise floor data, and the kernel density estimation method based on Gaussian kernel function is used to obtain a unique result;

[0019] Step 2.3: After training, when the generated results need to be judged, the training data is processed using a kernel density estimation method based on a Gaussian kernel function to obtain a unique judgment criterion.

[0020] Preferably, step 3 specifically includes the following steps:

[0021] Step 3.1: The weight tensor of the one-dimensional convolutional layer is initialized to a Gaussian distribution before training, and the weights are sampled from a high-low distribution with a mean of 0 and a standard deviation of 0.02;

[0022] Step 3.2: The weight tensor of the batch normalization layer is initialized to a Gaussian distribution before training, and the weights are sampled from a high-low distribution with a mean of 1 and a standard deviation of 0.02;

[0023] Step 3.3: The bias term of the batch normalization layer is initialized to 0.

[0024] Beneficial technical effects brought by the present invention:

[0025] This method utilizes the penalty mechanism of the loss function in the intelligent network training to eliminate the situation where the final result takes the logarithm of a negative number, thus solving mathematically uncalculatable problems and improving measurement accuracy. Secondly, the trained intelligent noise reduction network can generate a noise curve that almost completely matches the current noise floor in real time according to the current instrument status, thereby solving the "drop point" problem and real-time problems, thereby realizing intelligent noise reduction. The details are as follows:

[0026] 1) The WGAN-based signal analyzer intelligent noise reduction network can generate noise data that is almost completely consistent with the actual noise floor in real time according to the current state of the instrument. Compared with the currently used attenuator shielding method, its main advantage is strong real-time performance. When the instrument state changes, there is no need to wait for the noise floor to be measured again, which greatly improves the measurement efficiency;

[0027] 2) The penalty mechanism implemented during training enables the generated noise floor results to almost completely match the actual noise floor within the threshold range, thereby ensuring that the variance of the difference between the generated value and the actual value of the noise floor at each frequency point during the unit conversion process of the signal analyzer will not be too large, and the true number of the logarithm will not be negative, thereby ensuring that there will be no "drop points" and mathematical calculation difficulties in the measurement results, effectively improving the accuracy of the measurement and realizing intelligent noise reduction of the signal analyzer. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of the method of the present invention;

[0029] Figure 2 This is the internal structure diagram of the generator and discriminator;

[0030] Figure 2 (a) is the discriminator structure diagram; (b) is the generator structure diagram. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0032] 1. Basic Concepts of the Invention

[0033] The present invention is used to solve the problems of taking logarithms of negative numbers, "dropping points" and real-time performance when the signal analyzer uses the noise suppression function. Based on the WGAN network, through training and parameter optimization, the distribution and changes of noise under various states of the instrument are learned, and the noise floor under various states of the instrument is generated in real time, so that the signal analyzer can obtain the real-time "noise floor" value, and the intelligent noise reduction of the signal analyzer is realized through difference operation.

[0034] 2. Design of the present invention

[0035] 2.1 Overall network design

[0036] The process of WGAN is as follows Figure 1 shown.

[0037] Generator. The generator accepts random numbers that follow a Gaussian distribution and maps them to the target data space;

[0038] Discriminator. The discriminator is used to discriminate the data input to it. Its purpose is to discriminate the data from the real data set as true and the data from the generator as false;

[0039] Iterative optimization. When optimizing the generator, the goal is to make the samples generated by the generator gradually approach the real data; when optimizing the discriminator, the goal is to make the discriminator accurately judge the authenticity of the data. This repeated iteration and mutual feedback make the data finally generated by the network closely simulate the real data.

[0040] 2.2 Internal Structure of Generator and Discriminator

[0041] The internal structure of the generator and discriminator is as follows Figure 2 (a) and 2(b).

[0042] Discriminator. The input layer is an adaptive global pooling layer; the hidden layer contains 4 one-dimensional convolution layers + batch normalization layer + LeakyReLU activation function layer, where the number of channels of the convolution layer is 1-64-128-256-512, the convolution kernel size is 4, the step size is 2, and the padding is 1; the slope of the activation function is 0.2; the output layer is a one-dimensional convolution layer;

[0043] Generator. The input layer is a one-dimensional deconvolution layer + normalization layer + s2ReLU activation function layer; the hidden layer contains 4 one-dimensional deconvolution layers + batch normalization layers + ReLU or s2ReL activation function layers, where the number of channels of the convolution layer is 512-256-128-64-1, the convolution kernel size is 4, the stride is 2, the padding is 1, and there is also an adaptive global pooling layer; the output layer is a Sigmoid function layer.

[0044] The s2ReLU activation function in the generator. Its formula is sin(2πx)×ReLU(x)×ReLU(1-x). The design of this activation function is based on the characteristics of small noise but high-frequency jitter, which greatly enhances the fitting ability of the WGAN network.

[0045] 2.3 Data processing

[0046] Before training, each set of data is processed using maximum and minimum normalization;

[0047] In training, each iteration uses WGAN to generate 25 sets of noise floor data, and uses the kernel density estimation method based on Gaussian kernel function to obtain a unique result;

[0048] After training, when the generated results need to be judged, the training data is processed using a kernel density estimation method based on a Gaussian kernel function to obtain a unique judgment criterion.

[0049] 2.4 Network Weight Initialization

[0050] The weight tensor of the one-dimensional convolutional layer is initialized to a Gaussian distribution before training, and the weights are sampled from a high-low distribution with a mean of 0 and a standard deviation of 0.02;

[0051] The weight tensor of the batch normalization layer is initialized to a Gaussian distribution before training, and the weights are sampled from a high-low distribution with a mean of 1 and a standard deviation of 0.02;

[0052] The bias term of the batch normalization layer is initialized to 0.

[0053] The protection point of the present invention is to protect a signal analyzer intelligent noise reduction method based on WGAN, which has the following key technologies:

[0054] 1) The WGAN-based real-time noise reduction method abandons the previous method of using an attenuator to shield the external input to pre-measure the noise floor. The generative network generates noise in real time according to the instrument status, avoiding redundant operations such as repeated measurements caused by instrument status changes, ensuring real-time performance while improving measurement efficiency.

[0055] 2) The penalty mechanism in WGAN network training enables the generated results to almost completely match the real noise floor within the threshold range, thus ensuring that the variance of the difference between the generated value and the actual value of the noise floor at each frequency point during the unit conversion of the signal analyzer will not be too large, and the true number of the logarithm will not be negative, thus ensuring that there will be no "drop points" in the measurement results and mathematical calculation difficulties;

[0056] 3) The structure of the WGAN network consists of a discriminator with a convolutional layer and a generator with a deconvolutional layer;

[0057] 4) The training of WGAN network is known to be unstable, and the core parameters involved are obtained after a lot of debugging;

[0058] 5) The s2ReLU activation function in the generator is specially designed for the characteristics of small noise but high-frequency jitter.

[0059] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A signal analyzer intelligent noise reduction method based on WGAN, characterized by: The steps include: Step 1: Design the generator and discriminator; Step 2: Data processing; Step 3: Initialize network weights.

2. The intelligent noise reduction method for a signal analyzer based on WGAN according to claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1.1: Design a generator; the generator accepts random numbers that follow a Gaussian distribution and maps them to the target data space; Step 1.2: Design a discriminator; the discriminator is used to discriminate the data input to it, and its purpose is to discriminate the data from the real data set as true, and the data from the generator as false; Step 1.3: Iterative optimization; when optimizing the generator, the goal is to make the samples generated by the generator gradually approach the real data; When optimizing the discriminator, the goal is to enable the discriminator to accurately determine the authenticity of the data; repeated iterations and mutual feedback allow the data ultimately generated by the network to closely simulate the real data.

3. The intelligent noise reduction method for a signal analyzer based on WGAN according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 2.1: Before training, each set of data is processed using maximum and minimum normalization; Step 2.2: In each iteration of training, WGAN is used to generate 25 sets of noise floor data, and the kernel density estimation method based on Gaussian kernel function is used to obtain a unique result; Step 2.3: After training, when the generated results need to be judged, the training data is processed using a kernel density estimation method based on a Gaussian kernel function to obtain a unique judgment criterion.

4. The intelligent noise reduction method for a signal analyzer based on WGAN according to claim 1, characterized in that: Step 3 specifically includes the following steps: Step 3.1: The weight tensor of the one-dimensional convolutional layer is initialized to a Gaussian distribution before training, and the weights are sampled from a high-low distribution with a mean of 0 and a standard deviation of 0.02; Step 3.2: The weight tensor of the batch normalization layer is initialized to a Gaussian distribution before training, and the weights are sampled from a high-low distribution with a mean of 1 and a standard deviation of 0.02; Step 3.3: The bias term of the batch normalization layer is initialized to 0.

Images