Aviation simulation signal intelligent noise adding method based on information constraint generative adversarial network

Abstract

The invention discloses an aviation simulation signal intelligent noise adding method based on an information constraint generative adversarial network, and relates to the technical field of aviation testing, and the method comprises the steps: forming a sensor noise adding signal and a corresponding controllable noise coding vector training sample set in a real noise environment; creating an aviation noise-adding simulation data intelligent generation network, and carrying out model initialization and optimizer configuration; the training sample sets are subjected to aviation noise-adding simulation data intelligent generation network loop training in batches; and after the training is completed, inputting a controllable noise coding vector to be simulated, and generating network output aviation noise adding simulation data with the same noise distribution in a real environment. According to the method, the information constraint generative adversarial network is adopted on the basis of a traditional generative adversarial network, controllable noise coding is suitable for any defined noise attribute, the time-frequency characteristics and non-stationary distribution of noise can be better captured, more real simulation data can be generated, and the generalization ability and universality of the method are higher.

Description

Technical Field

[0001] The present invention relates to the field of aviation testing technology, and in particular to an intelligent noise addition method for aviation simulation signals based on an information-constrained generative adversarial network. Background Art

[0002] Noise signals (such as electromagnetic interference, mechanical vibration noise, and aerodynamic noise) generated by aircraft in complex electromagnetic environments can severely impact sensor data quality and, in turn, interfere with the performance of back-end test systems. During automated aviation testing, it is necessary to simulate the deterministic physical signals measured by sensors to generate sensor simulation data. Furthermore, to verify the adaptability and robustness of the test object in a real-world noise environment, the generated sensor simulation data must be subjected to noise-addition simulation. The quality of the generated simulation data is directly related to the authenticity of the test results. Current methods for generating aviation noise-addition simulation data fall into two main categories: physics-based modeling and data-driven approaches. However, both have significant drawbacks. Aviation noise is multi-source, non-stationary, and strongly coupled, involving the effects of multiple physical fields, such as electromagnetic, mechanical, and fluidic. Traditional noise generation processes based on electromagnetic field equations and statistical probability models struggle to accurately simulate noise characteristics in dynamic environments. Traditional machine learning data-driven methods, such as support vector machines, random forests, or shallow neural networks, struggle to capture the time-frequency characteristics and non-stationary distribution of noise signals, resulting in unrealistic results. In addition, the traditional generative adversarial network model can only generate simulated data with the same distribution as the real labeled training samples, but cannot capture the noise distribution characteristics of the current signal in the real environment and realize the noise simulation of different noise sample properties. Summary of the Invention

[0003] The present invention provides an intelligent noise addition method for aviation simulation signals based on information-constrained generative adversarial networks, which solves the problems of low generation quality and weak versatility of existing physical modeling-based and data-driven methods, and improves the authenticity and generalization of aviation noise-added simulation data.

[0004] To achieve the above object, the technical solution of the present invention is as follows:

[0005] An intelligent noise addition method for aviation simulation signals based on information-constrained generative adversarial networks includes the following steps:

[0006] Step 1: Under normal working conditions, collect the sensor noise signal x in a real noise environment noise and the corresponding controllable noise coding vector c, forming a training sample set {x noise ,c}.

[0007] Step 2: Model initialization and optimizer configuration. Create an intelligent generation network for aviation noise simulation data, including a generator network G, a discriminator network D, and an information constraint network E. Configure the Adam optimizer for each network. Specifically, the generator network G uses a fully connected layer network structure, while the discriminator network D and the information constraint network E use the same convolutional neural network architecture.

[0008] Step 3: The training sample set is fed into the intelligent generation network of aviation noise simulation data in batches for cyclic training. Each cycle includes the following sub-steps:

[0009] Step 3.1: The Gaussian distribution z with mean 0 and standard deviation 1 is used as the latent vector and the controllable noise encoding vector c to enter the generative network G, and the generated sensor noise simulation data x fake =G(z,c).

[0010] Step 3.2: Real sensor noise signal x noise Generate sensor noise simulation data x fake Together they enter the discriminant network D to discriminate true from false and output the discrimination result D(x noise ) and D(x fake ).

[0011] Step 3.3: Real sensor noise signal x noise Generate sensor noise simulation data x fake Together enter the information constraint network E to perform controllable noise coding prediction, and noise with x fake The prediction results are E(x noise ) and E(x fake ).

[0012] Step 3.4: Update network parameters adversarially.

[0013] Furthermore, we calculate x fake The discriminant loss D loss (x fake ) and the predicted mutual information loss I(c,E(x fake )), I is the mutual information measure.

[0014] Furthermore, according to the generated network G loss G loss =D loss (x fake )-λI(c,E(x fake ))Back propagation updates the network parameters of G.

[0015] Furthermore, the network parameters of the generated network G are fixed and the response to x is calculated. noise The discriminant loss D loss (x noise), continuously generate x through the generation network G fake , calculate x fake The discriminant loss D loss (x fake ) and the predicted loss E loss (x fake ), respectively according to D loss =D loss (x fake )+D loss (x noise ) and E loss =E loss (x fake ) Perform back propagation on the discriminant network D and the information constraint network E to update the network parameters.

[0016] Step 4: After completing the training of all training sample sets, fix the network parameters of the generation network G, the discriminant network D, and the information constraint network E, input the controllable noise code vector c′ to be simulated, and generate the network G to output the aviation noise simulation data x′ with the same noise distribution as the real environment noise .

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] (1) Compared with existing physical modeling-based methods and shallow machine learning-based methods, the present invention adopts a generative adversarial network intelligent algorithm. When faced with the simulation difficulties of multi-source, non-stationary and strongly coupled aviation noise, it can better capture the time-frequency characteristics and non-stationary distribution of noise with a small number of samples, generate more realistic simulation data, and have stronger generalization ability.

[0019] (2) The present invention adopts an information-constrained generative adversarial network. Compared with the traditional generative adversarial network, it adds a controllable noise coding vector and an information-constrained auxiliary network. The mutual information metric is used to realize the noise simulation required for different controllable noise coding vectors. It can be data of different sensor types, different noise types or different noise intensities. The method is more versatile. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 It is an overall block diagram of the technical solution of the present invention.

[0022] Figure 2Generate a network structure diagram for the present invention.

[0023] Figure 3 This is the network structure diagram for the present invention.

[0024] Figure 4 This is the information constraint network structure diagram of the present invention. Specific implementation plan

[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0026] Reference Figure 1 The present invention discloses an intelligent noise addition method for aviation simulation signals based on information-constrained generative adversarial networks, which includes the following specific steps:

[0027] Step 1: Under normal working conditions, collect the sensor noise signal x in a real noise environment noise and the corresponding controllable noise coding vector c, forming a training sample set {x noise ,c}, the controllable noise encoding vector c can be any defined sample noise attribute, which can be but not limited to the same sensor number, different sensor types and different noise amplitudes.

[0028] Furthermore, in the embodiment, the sensor noise signal is an analog signal, and the sensor noise signal x noise The controllable noise encoding vector is the sensor number of the sensor output signal collected and stored under normal field use conditions of aviation equipment.

[0029] Step 2: Model initialization and optimizer configuration. Create an intelligent generation network for aviation noise simulation data, including a generator network G, a discriminator network D, and an information constraint network E. Configure the Adam optimizer for each network. Specifically, the generator network G uses a fully connected layer network structure, while the discriminator network D and the information constraint network E use the same convolutional neural network architecture.

[0030] Further, refer to Figure 2 , the generation network G consists of a four-layer network structure. The first three layers are composed of linear layers and ReLU activation function layers, and the last layer is composed of linear layers and Tanh activation function layers.

[0031] Further, refer to Figure 3,The discriminant network D consists of a four-layer network structure. The first three layers are convolutional layers, pooling and ReLU activation function layers, and the last layer is composed of linear layers and sigmoid activation function layers.

[0032] Further, refer to Figure 4 ,The information constraint network E consists of a four-layer network structure. The first three layers are composed of linear layers and ReLU activation function layers, and the last layer is composed of a linear layer and Tanh activation function layer.

[0033] Step 3: The training sample set is fed into the intelligent generation network of aviation noise simulation data in batches for cyclic training. Each cycle includes the following sub-steps:

[0034] Step 3.1: Submit the Gaussian distribution z with mean 0 and standard deviation 1 as the potential vector and the controllable noise encoding vector c to the generative network G to generate the same vector as x. noise Sensor noise simulation data x with the same dimension fake =G(z,c).

[0035] Step 3.2: Real sensor noise signal x noise Generate sensor noise simulation data x fake Together they enter the discriminant network D to discriminate true from false. The discriminant result D(x noise ) and D(x fake ).

[0036] Furthermore, the discriminant network D outputs the probability dp of whether each sample is a real noise signal, and the training set discrimination result D(x noise )={dp},D(x fake )={1-dp}.

[0037] Step 3.3: Real sensor noise signal x noise Generate sensor noise simulation data x fake Together enter the information constraint network E to predict the sensor number, and noise with x fake The prediction results are the encoding vector E(x noise ) and E(x fake ).

[0038] Step 3.4: Update network parameters adversarially.

[0039] Furthermore, we calculate x fake The discriminant loss D loss (x fake ) and the predicted mutual information loss I(c,E(x fake )), I is the mutual information measure.

[0040] Furthermore, Dloss (x fake )=-mean(D(x fake )), mean(·) represents the averaging operation.

[0041] Furthermore, according to the generated network G loss G loss =D loss (x fake )+λI(c,E(x fake )) Back propagation updates the network parameters of G, I is the mutual information metric, I(c,E(x fake ))=-mean(sum(Log_softmax(E(x fake )*c))), sum(·) represents summation, Log_softmax(·) represents logarithm after softmax function, and λ is a hyperparameter, which is set to 1.0.

[0042] Furthermore, the network parameters of the generated network G are fixed and the response to x is calculated. noise The discriminant loss D loss (x noise ), continuously generate x through the generation network G fake , calculate x fake The discriminant loss D loss (x fake ) and the predicted loss E loss (x fake ), respectively according to D loss =D loss (x fake )+D loss (x noise ) and E loss =E loss (x fake ) Perform back propagation on the discriminant network D and the information constraint network E to update the network parameters.

[0043] Furthermore, D loss (x noise )=-mean(D(x noise )), E loss (x fake )=I(c,E(x fake )).

[0044] Step 4: After completing the training of all training sample sets, the noise distribution law under each noise attribute in the controllable noise coding vector is learned. The network parameters of the generator network G, the discriminator network D, and the information constraint network E are fixed. The sensor number vector c′ to be simulated is input, and the generator network G outputs the aviation noise simulation data x′ with the same noise distribution as the real environment. noise .

[0045] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0046] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. An intelligent noise addition method for aviation simulation signals based on information-constrained generative adversarial networks, characterized by Step 1: Under normal working conditions, collect the sensor noise signal x in a real noise environment noise and the corresponding controllable noise coding vector c, forming a training sample set {x noise ,c}.

2. Step 1 according to claim 1, characterized in that The sensor noise signal x noise The sensor output signals are collected and stored for aviation equipment under normal field use conditions.

3. Step 1 according to claim 1, characterized in that The controllable noise coding vector c can be any defined sample noise attribute, including but not limited to the same sensor number, different sensor types, and different noise amplitudes.

4. An intelligent noise addition method for aviation simulation signals based on information-constrained generative adversarial networks, characterized by This includes step 2: model initialization and optimizer configuration. This involves creating an intelligent generation network for aviation noise simulation data, including a generator network G, a discriminator network D, and an information constraint network E. Adam optimizers are configured for each network.

5. Step 2 according to claim 4, characterized in that The generative network G adopts a fully connected layer network structure, and the discriminative network D and the information constraint network E adopt a convolutional neural network structure.

6. Step 2 according to claim 4, characterized in that The generative network G consists of a four-layer network structure. The first three layers are composed of linear layers and ReLU activation function layers, and the last layer is composed of linear layers and Tanh activation function layers. The discriminant network D consists of a four-layer network structure. The first three layers are composed of convolutional layers, pooling layers, and ReLU activation function layers, and the last layer is composed of linear layers and sigmoid activation function layers. The information constraint network E consists of a four-layer network structure. The first three layers are composed of linear layers and ReLU activation function layers, and the last layer is composed of linear layers and Tanh activation function layers.

7. An intelligent noise addition method for aviation simulation signals based on information-constrained generative adversarial networks, characterized by The method includes step 3: the training sample set enters the aviation noise simulation data intelligent generation network in batches for cyclic training.

8. Step 3 according to claim 6, characterized in that: Each training cycle includes sub-step S1: a Gaussian distribution z with a mean of 0 and a standard deviation of 1 is used as a potential vector and a controllable noise encoding vector c to enter the generator network G, and the generator network G outputs the generated sensor noise simulation data x fake =G(z,c)。 9. Step 3 according to claim 6, characterized in that: Each training cycle includes sub-step S2: the real sensor noise signal x noise Generate sensor noise simulation data x fake Together they enter the discriminant network D to discriminate true from false and output the discrimination result D(x noise ) and D(x fake ).

10. Step 3 according to claim 6, characterized in that: Each training cycle includes sub-step S3: the real sensor noise signal x noise Generate sensor noise simulation data x fake Together enter the information constraint network E to perform controllable noise coding prediction, and noise with x fake The prediction results are E(x noise ) and E(x fake ).

11. Step 3 according to claim 6, characterized in that: Each training cycle includes sub-step S4: adversarial update of network parameters.

12. Sub-step S4 according to claim 11, characterized in that: Calculate x fake The discriminant loss D loss (x fake ) and the predicted mutual information loss I(c,E(x fake )), I is the mutual information metric, I(c,E(x fake ))=-mean(sum(Log_softmax(E(x fake )*c))), sum(·) represents summation, Log_softmax(·) represents logarithm after softmax function, and λ is a hyperparameter, which is set to 1.

0. According to the generated network G loss G loss =D loss (x fake )-λI(c,E(x fake )) Back propagation updates the network parameters of G; Fixed network parameters of the generated network G, calculate the x noise The discriminant loss D loss (x noise ), continuously generate x through the generation network G fake , calculate x fake The discriminant loss D loss (x fake ) and the predicted loss E loss (x fake ), respectively according to D loss =D loss (x fake )+D loss (x noise ) and E loss =E loss (x fake ) Perform back propagation on the discriminant network D and the information constraint network E to update the network parameters.

13. An intelligent noise addition method for aviation simulation signals based on information-constrained generative adversarial networks, characterized by Step 4: After completing the training of all training sample sets, the generative network G learns the noise distribution law under each noise attribute in the controllable noise coding vector. The network parameters of the generative network G are fixed, the controllable noise coding vector c′ to be simulated is input, and the generative network G outputs the aviation noise simulation data x′ with the same noise distribution as the real environment. noise .

Images