CN111832229A - Vibration transfer system based on CycleGAN model and its training method - Google Patents

Abstract

The invention discloses a vibration transmission system based on CycleGAN model and a training method thereof, including two groups of GAN models, each group of GAN models is composed of generators (G, F) and discriminators (DX, DY), the generators ( G, F) are constructed by one-dimensional convolutional networks, and the discriminators (DX, DY) are constructed by one-dimensional convolutional neural networks; the present invention ignores the specific vibration transmission path of the mechanical equipment structure, and only considers the measuring point A and the measuring point Based on the vibration signal transmission relationship between B, a black box vibration signal transmission model is established. Based on the generative adversarial network model CycleGAN model, based on the original data signal of the measurement points, the generator and discriminator in the CycleGAN model are trained to obtain stable The model parameters are established, and the vibration signal transmission program model from measuring point A to measuring point B is established.

Description

Vibration transfer system based on CycleGAN model and its training method

technical field

The invention relates to the technical field of signal research, in particular to a vibration transmission system based on a CycleGAN model and a training method thereof.

Background technique

Mechanical equipment often has vibration conditions. Usually, vibration testing methods are used to monitor the vibration conditions of mechanical equipment and diagnose the health status of the equipment. The vibration of different positions on the equipment responds differently to the vibration conditions of the mechanical equipment. Therefore, vibration During the test, it is necessary to arrange the measuring points at the position that can best reflect the vibration of the equipment. Through the study of vibration transmission, the relationship between each point of the mechanical equipment and the maximum vibration point can be established, so that the mechanical equipment can only be tested at the maximum vibration point. For vibration testing, the conventional research on vibration transmission path is mainly based on the physical structure characteristics of mechanical equipment, establishes a vibration transmission process model, analyzes the vibration transmission process from measurement point A to measurement point B, and obtains the signal characteristics of measurement point B. ;

However, the prior art relies on the physical structure characteristics of mechanical equipment to deduce complex vibration transmission paths and characteristics, which is difficult to extend to the utilization of transmission characteristics of unknown signals. Therefore, the present invention proposes a vibration transmission system based on the CycleGAN model to solve the problem of existing There are technical problems.

SUMMARY OF THE INVENTION

In view of the above problems, the purpose of the present invention is to propose a vibration transfer system based on the CycleGAN model. The system and its training method do not rely on the physical structure characteristics of mechanical equipment to deduce complex vibration transfer paths and characteristics, and can effectively establish any data. The black box model of vibration signal transmission from A to data B can be widely extended to the utilization of the transmission characteristics of unknown signals.

In order to realize the purpose of the present invention, the present invention is realized through the following technical solutions: a vibration transfer system based on CycleGAN model, including two groups of GAN models, each group of described GAN models is composed of generators (G, F) and discriminators (DX, DY). ), the generator (G, F) is constructed with a one-dimensional convolutional network, and the discriminator (DX, DY) is constructed with a one-dimensional convolutional neural network.

A further improvement is that: the CycleGAN model composed of the two groups of the GAN models has a ring structure, and each group of the GAN models has one generator (G, F) and one discriminator (DX, DY).

The training method of the vibration transfer system based on the CycleGAN model includes the following steps:

Step 1: Preprocess the original data (X, Y) according to the set input data sequence length (series_len), where X represents the data in the X domain, and Y represents the data in the Y domain;

Step 2: Pass the data in the X domain through the generator G to obtain the corresponding Y domain data fake-Y, and reconstruct the fake-Y through the generator F to obtain the X domain data reconstr-X; pass the data in the Y domain through the generator F to obtain the X domain The data fake_X, and then reconstructed back to the original data reconstr_Y input by the Y domain through the generator G;

Step 3: First use the discriminators Dx and Dy to discriminate the respective data X, reconstr_X and Y, and reconstr_Y respectively, to determine whether the data is real data or generated data, and then use Dx to calculate the real data and the loss of the generated data D_x_loss_real and D_x_loss_fake, use Dy calculates the losses D_y_loss_real and D_y_loss_fake for real data and generated data and the total loss D_loss_real and D_loss_fake for the model;

Step 4: Update the parameters of the generator according to the change of the loss value, including training the parameters of the generator and the discriminator convolutional network, and finally make the model reach a balance point and end the training.

A further improvement is: in the fourth step, after the training is completed, the parameters of the optimizer and the one-dimensional convolutional neural network at this time are saved.

The beneficial effects of the present invention are as follows: the present invention ignores the specific vibration transmission path of the mechanical equipment structure, only considers the vibration signal transmission relationship between the measuring point A and the measuring point B, establishes a black box type vibration signal transmission model, based on the generative confrontation network model CycleGAN The model, based on the original data signal of the measuring point, trains the generator and the discriminator in the CycleGAN model, obtains stable model parameters, and establishes a vibration signal transmission program model from the measuring point A to the measuring point B. The present invention does not rely on The derivation of complex vibration transmission paths and characteristics based on the physical structure characteristics of mechanical equipment can effectively establish a black box model of vibration signal transmission from any data A to data B, and can be widely extended to the utilization of transmission characteristics of unknown signals.

Description of drawings

Fig. 1 is the CycleGAN model structure diagram of the present invention;

Fig. 2 is a design program model data change process diagram of the present invention;

Fig. 3 is a design program model data change process diagram of the present invention;

Fig. 4 is the flow chart of designing program training data in the verification example of the present invention;

Fig. 5 is the actual data B sequence diagram in the verification example of the present invention;

Fig. 6 is the actual data B spectrogram in the verification example of the present invention;

Fig. 7 is the time sequence diagram of prediction data C in the verification example of the present invention;

FIG. 8 is a spectrogram of predicted data C in a verification example of the present invention.

Detailed ways

In order to deepen the understanding of the present invention, the present invention will be described in further detail below with reference to the embodiments. The embodiments are only used to explain the present invention and do not constitute a limitation on the protection scope of the present invention.

As shown in FIG. 1, this embodiment provides a vibration transfer system based on the CycleGAN model, including two groups of GAN models, each group of the GAN models is composed of a generator (G, F) and a discriminator (DX, DY), so The generators (G, F) are constructed using a one-dimensional convolutional network, and the discriminators (DX, DY) are constructed using a one-dimensional convolutional neural network.

The CycleGAN model composed of the two groups of the GAN models has a ring structure, and each group of the GAN models has one generator (G, F) and one discriminator (DX, DY).

As shown in Figures 2 and 3, this embodiment provides a training method for a vibration transfer system based on the CycleGAN model, including the following steps:

Step 1: Preprocess the original data (X, Y) according to the set input data sequence length (series_len), where X represents the data in the X domain, and Y represents the data in the Y domain;

Step 2: Pass the data in the X domain through the generator G to obtain the corresponding Y domain data fake-Y, and reconstruct the fake-Y through the generator F to obtain the X domain data reconstr-X; pass the data in the Y domain through the generator F to obtain the X domain The data fake_X, and then reconstructed back to the original data reconstr_Y input by the Y domain through the generator G;

Step 3: First use the discriminators Dx and Dy to discriminate the respective data X, reconstr_X and Y, and reconstr_Y respectively, to determine whether the data is real data or generated data, and then use Dx to calculate the real data and the loss of the generated data D_x_loss_real and D_x_loss_fake, use Dy calculates the losses D_y_loss_real and D_y_loss_fake for real data and generated data and the total loss D_loss_real and D_loss_fake for the model;

Step 4: Update the parameters of the generator according to the change of the loss value, including the parameters of training the generator and the discriminator convolutional network, and finally make the model reach a balance point, end the training, and save the optimizer and one-dimensional convolutional neural network at this time. parameter.

Verification example:

Using the vibration data signal A and signal B that have been measured, first preprocess the data to obtain data of the same data sequence length, train the parameters of the CycleGAN model, and obtain the optimized parameters in the model generator and discriminator, such as parameters such as learning_rate, beta1, beta2, and fixed parameter optimization parameters. See Figure 4.

Then input the actual signal data A into the model again, and output the predicted vibration data C after training the model. After comparing the characteristics of the predicted data C and the actual signal data B according to the vibration characteristic parameters of the vibration test data, it is found that the predicted data C of the CycleGAN model is found. It has a good fit with the actual signal data B, and the main signal characteristics are exactly the same. Demonstrate the effectiveness and practicality of the technique.

Test results obtained according to the CycleGAN model: see Figures 5, 6, 7, and 8.

Table 1 Comparison of actual data and forecast

Timing Diagram Amplitude Spectral eigenfrequency actual data B (-0.3, 0.3) 90.00Hz forecast data C (-0.2, 0.3) 90.00Hz

The invention ignores the specific vibration transmission path of the mechanical equipment structure, only considers the vibration signal transmission relationship between the measurement point A and the measurement point B, establishes a black box vibration signal transmission model, based on the generative confrontation network model CycleGAN model, with the original measurement point Based on the data signal, the generator and discriminator in the CycleGAN model are trained to obtain stable model parameters, and the vibration signal transmission program model from measuring point A to measuring point B is established. The present invention does not rely on the physical structure characteristics of mechanical equipment. The derivation of complex vibration transmission paths and characteristics can effectively establish a black box model of vibration signal transmission from any data A to data B, and can be widely extended to the utilization of transmission characteristics of unknown signals.

The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above-mentioned embodiments. The above-mentioned embodiments and descriptions only illustrate the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will also have Various changes and modifications fall within the scope of the claimed invention. The claimed scope of the present invention is defined by the appended claims and their equivalents.

Claims (4)

1. Vibration transmission system based on cycleGAN model, including two sets of GAN models, its characterized in that: each group of GAN models is composed of a generator (G, F) and discriminators (DX and DY), wherein the generator (G, F) is constructed by adopting a one-dimensional convolution network, and the discriminators (DX and DY) are constructed by adopting a one-dimensional convolution neural network.

2. The CycleGAN model-based vibration transmission system of claim 1, wherein: the cycleGAN model formed by the two groups of GAN models is of an annular structure, and a generator (G, F) and a discriminator (DX and DY) in each group of GAN models are respectively provided with one.

3. The training method of the vibration transmission system based on the cycleGAN model is characterized by comprising the following steps of:

the method comprises the following steps: preprocessing original data (X, Y) according to the set length (series _ len) of an input data sequence, wherein X represents data of an X field, and Y represents data of a Y field;

step two: data of the X domain is processed by a generator G to obtain corresponding Y domain data fake-Y, and the fake-Y is reconstructed by a generator F to obtain X domain data reconstr-X; obtaining X domain data fake _ X from the data of the Y domain through a generator F, and reconstructing original data reconstr _ Y input in the Y domain through a generator G;

step three: firstly, respectively distinguishing the respective data X, reconstr _ X and Y, reconstr _ Y by using discriminators Dx and Dy, judging whether the data is real data or generated data, then calculating the loss D _ X _ loss _ real and D _ X _ loss _ fake of the real data and the generated data by using Dx, and calculating the loss D _ Y _ loss _ real and D _ Y _ loss _ fake of the real data and the generated data and the total loss D _ loss _ real and D _ loss _ fake of the model by using Dy;

step four: and updating parameters of the generator according to the change condition of the loss value, wherein the parameters comprise parameters of a training generator and a discriminator convolution network, and finally, enabling the model to reach an equilibrium point and ending the training.

4. The method for training a CycleGAN model-based vibration transmission system according to claim 3, wherein: in the fourth step, after the training is finished, the parameters of the optimizer and the one-dimensional convolutional neural network at the moment are stored.

Images