Marine riser fault detection method

Abstract

The invention provides a marine riser fault detection method, and relates to the technical field of marine engineering monitoring. Comprising the steps that S1, a marine riser simulation model is established through finite element analysis, and simulation calculation is conducted to obtain simulation data; s2, building a deep learning model based on a self-attention-scalar long short-term memory network, and training the deep learning model by using the simulation data obtained in the step S1; s3, constructing a generative adversarial domain reconstruction migration network, constructing a generator and a discriminator, and performing adversarial training on the generator and the discriminator by using real data and analog data; and S4, taking the generator as a preprocessing part of real data, and performing fault detection by using the deep learning model trained in the step S2. According to the method, collaborative fusion of multiple technologies is achieved, and the problem that an existing method for conducting fault detection on the marine riser has defects is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of marine engineering monitoring, and in particular to a marine riser fault detection method. Background Art

[0002] Marine risers, transmission pipelines connecting underwater production systems and floating structures, are a crucial component of deepwater oil and gas field development systems. Due to the time-varying nature of deepwater marine environmental loads and the complex interactions between risers and the seabed soil, riser safety presents unique challenges and significant uncertainties. Furthermore, marine risers are not only subject to corrosion from the natural environment, such as seawater, but also to long-term weathering, which weakens the protective effects of their external protective layers. Furthermore, the transportation of high-temperature, high-pressure petroleum gas within the pipes accelerates internal wear and tear. Given these combined factors, a riser failure could jeopardize the safety of the entire deepwater oil and gas field system and could even lead to a series of serious accidents.

[0003] Therefore, it is very necessary to perform fault detection on marine risers. However, existing technologies in this field have limitations in practical applications. Specifically, existing riser designs are affected by factors such as differences between assumed environmental loads and actual loads, limitations in riser calculation and analysis models, uncertainty in riser damage, and possible deviations between riser installation and actual loads. Therefore, it is difficult to accurately reflect the actual loads and stresses acting on the risers through preliminary design alone. Currently, common riser fault detection methods include sensor technology, fiber optic sensing technology, sonar detection technology, drone and remote sensing technology, data analysis and pattern recognition algorithms, etc. For example, the patent publication number CN106697209A provides a deepwater tension leg platform device and its riser real-time monitoring method, which adopts the method of deploying sensors. With the rise and development of deep learning technology, data-driven methods have made riser fault detection more intelligent. However, practical verification has shown that these current methods still have many shortcomings, such as poor model generalization ability, and the performance of existing deep learning models significantly degrades when applied across scenarios; high data collection and labeling costs, and the need to re-collect data for each different scenario; the labels to which samples belong need to be marked, which requires a lot of time and effort. In addition, in actual engineering, it is difficult to directly obtain risers with certain faults. In this case, the label judgment is based on theoretical calculations, such as traditional peak analysis, which further challenges the accuracy of the results.

[0004] In summary, the present invention provides a method for detecting marine riser faults. Summary of the Invention

[0005] The object of the present invention is to provide a method for detecting faults of marine risers, so as to solve the problem mentioned in the above background art that the existing methods for detecting faults of marine risers have shortcomings.

[0006] The present invention is achieved by adopting the following technical solutions:

[0007] A method for detecting faults in a marine riser comprises the following steps:

[0008] Step S1: establishing a marine riser simulation model through finite element analysis, and performing simulation calculations to obtain simulation data;

[0009] Step S2: Build a deep learning model based on the self-attention-scalar long short-term memory network, and train the deep learning model using the simulated data obtained in step S1;

[0010] Step S3: Build a generative adversarial domain reconstruction transfer network, build a generator and a discriminator, and use real data and simulated data to perform adversarial training on the generator and the discriminator respectively;

[0011] Step S4: Use the generator as the preprocessing part of the real data and use the deep learning model trained in step S2 for fault detection.

[0012] Furthermore, the step S1 includes the following sub-steps:

[0013] Step S1-1: using finite element analysis to establish a marine riser simulation model based on the parameters of the marine riser;

[0014] Step S1-2: define boundary conditions;

[0015] Step S1-3: Determine the monitoring point and set the fault type;

[0016] Step S1-4: performing simulation calculations on the marine riser simulation model to obtain vibration response signals of multiple designated sampling points of the marine riser under different working conditions;

[0017] Step S1-5: pre-processing the simulation data obtained in step S1-4, dividing the vibration response signal into a (batch, 1, seq) format;

[0018] Step S1-6: perform label processing on the simulation data obtained in step S1-4.

[0019] Furthermore, step S2 includes the following sub-steps:

[0020] Step S2-1: Combine the self-attention mechanism with a neural network module with a scalar long short-term memory network to obtain a deep learning model;

[0021] Step S2-2: Adding a position detection head and a fault detection head to the deep learning model obtained in step S2-1, wherein the position detection head predicts the location of the fault, and the fault detection head predicts the type of fault;

[0022] Step S2-3: Input the simulated data obtained in step S1 into the deep learning model to train the deep learning model.

[0023] Furthermore, step S3 includes the following sub-steps:

[0024] Step S3-1: Define the set of real data as the target domain, define the set of simulated data as the source domain, and use a generative adversarial network to align the target domain data with the source domain data;

[0025] Step S3-2: Based on the target domain data and source domain data, a one-dimensional convolutional neural network is used to build a generator and a discriminator.

[0026] Furthermore, in step S4, the generator parameters are frozen, and real data is input into the generator to obtain data aligned with the feature domain of the simulated data, which is then input into the deep learning model for fault monitoring.

[0027] Furthermore, in step S4, confidence thresholds are set for the detection results of the position detection head and the fault detection head, detection results below the confidence threshold are not processed, and detection results above the confidence threshold are determined as corresponding fault modes.

[0028] Furthermore, in step S1-1, the parameters of the marine riser include size, material and connection method; in step S1-2, the boundary conditions include the installation and constraint method of the marine riser, and the action method and magnitude of the external load on the marine riser.

[0029] Furthermore, in step S1-6, the label corresponding to the vibration response signal has two dimensions, wherein the first dimension is the sampling position, represented by numbers 0 to n-1, where n is the number of sampling points; the second dimension is the fault type, including no fault, corrosion, crack, deformation, and mixed type faults.

[0030] Furthermore, in step S2-3, a dynamic weight method is used to balance the two detection heads during training, and the loss weight of the initial position detection head is set to 1, and the loss weight of the initial fault detection head is set to 0.5. The loss weight of the position detection head decreases linearly to 0.2 with training, and the loss weight of the fault detection head increases linearly to 1 with training.

[0031] Furthermore, in step S3-2, a batch of data is randomly selected from the source domain and input into the discriminator for training, at which time the generator parameters are frozen; then, the discriminator parameters are frozen, the generator parameters are unfrozen, and a batch of data is randomly selected from the target domain and input into the generator, and then the data generated by the generator is input into the discriminator for training; in this step, the above training process needs to be repeated until the preset conditions are met.

[0032] The beneficial effects achieved by the present invention are:

[0033] Provided is a marine riser fault detection method that covers the entire process of simulation modeling, network architecture design, and training strategy, forming a complete technical closed loop. The method achieves a collaborative integration of multiple technologies by acquiring data through finite element simulation, building a deep learning model based on a self-attention-scalar long short-term memory network, and constructing a network with generative adversarial domain reconstruction and migration. Based on this, the present invention addresses the inherent deficiencies of traditional fault detection methods in data acquisition, model recognition, and generalization capabilities for the special scenario of marine risers, effectively improving the accuracy of marine riser fault detection and providing advanced technical support for the safe operation and maintenance of marine risers, thereby ensuring the efficient and reliable operation of marine oil and gas production systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 1 is a reference schematic diagram of the process of the marine riser fault detection method according to an embodiment of the present invention;

[0035] Figure 2 3. It is a reference schematic diagram of the adversarial training process in the marine riser fault detection method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0037] Example 1

[0038] This embodiment provides a method for detecting faults in a marine riser. Figure 1 , including the following steps:

[0039] Step S1: Establish a marine riser simulation model through finite element analysis, and perform simulation calculations to obtain simulation data.

[0040] Specifically, the method includes the following sub-steps:

[0041] Step S1-1: Using finite element analysis, a simulation model of the marine riser is constructed based on the marine riser's parameters, including dimensions, materials, and connection methods. In this embodiment, the specific parameters are the marine riser's wall thickness, diameter, length, and the material's elastic modulus, Poisson's ratio, and yield strength.

[0042] Step S1-2: Define boundary conditions, including the installation and restraint methods of the marine riser, as well as the mode and magnitude of the external loads acting on the marine riser. Marine riser installation methods include rigid connections, flexible connections, and other special connection methods. In this example, flexible connections are selected. External loads include waves and currents. In this embodiment, the JONSWAP spectrum is used to simulate irregular waves. The effective wave height is set to 5 meters, the average period is 8 seconds, and the wave direction is at a 45° angle to the platform axis. The current is simplified to a stratified uniform flow. The surface current velocity is 1.2 m / s, which decreases linearly with depth to 0.5 m / s at the seabed. The current direction is the same as the wave direction.

[0043] Step S1-3: Determine monitoring points and set fault types, where commonly considered fault types include corrosion, cracks, and deformation. In this embodiment, two points are selected at the top, middle, and bottom of the marine riser as sampling points, and three different values are set for corrosion, cracks, and deformation to simulate different degrees of damage. Corrosion is achieved by modifying the geometry of the deep learning model. Specifically, a smaller area is created at the corresponding location of the marine riser model, and the wall thickness of this area is reduced to simulate material loss caused by corrosion. Cracks are simulated using the fracture mechanics module in ABAQUS. Specifically, a crack surface is defined in the model, and the crack starting position, propagation direction, and propagation length are specified. Deformation is simulated by changing the geometry of the model. Specifically, a concave geometry is created at the corresponding location.

[0044] Step S1-4: Perform simulation calculations on the marine riser simulation model using dedicated software to obtain vibration response signals of multiple designated sampling points of the marine riser under different working conditions. In this embodiment, the dedicated software used is ABAQUS.

[0045] Step S1-5: Preprocess the simulated data obtained in step S1-4, dividing the vibration response signal into a (batch, 1, seq) format. The first dimension, batch, represents the batch size, indicating the number of data stacks during one forward propagation during training; the second dimension, 1, indicates that only data from a single sampling point is input each time; and the third dimension, seq, represents the number of samples of the vibration response signal. In this example, the number of data stacks is 64 and the number of samples of the vibration response signal is 2000, so the format is (64, 1, 2000), and the data is randomly packaged.

[0046] Step S1-6: Label the analog data obtained in step S1-4. The labels corresponding to the vibration response signals have two dimensions: the first dimension is the sampling location, represented by numbers 0 to n-1, where n is the number of sampling points; the second dimension is the fault type, which includes no fault, corrosion, crack, deformation, and mixed faults. In this embodiment, the first dimension is represented by numbers 0 to 5; in the second dimension, 0 represents no fault, 1, 2, and 3 represent low, medium, and high corrosion, respectively; 4, 5, and 6 represent low, medium, and high cracks, respectively; 7, 8, and 9 represent low, medium, and high deformation, respectively; 10 represents a mixed corrosion and crack fault; 11 represents a combined corrosion and deformation fault; 12 represents a combined crack and deformation fault; and 13 represents all three faults.

[0047] Step S2: Build a deep learning model based on the self-attention-scalar long short-term memory network, and train the deep learning model using the simulated data obtained in step S1.

[0048] Specifically, the method includes the following sub-steps:

[0049] Step S2-1: Combine the self-attention mechanism with the neural network module with scalar long short-term memory network to obtain a deep learning model.

[0050] The self-attention mechanism can be used to capture long-distance dependencies and can better capture the relationship between elements that are far apart in the sequence. At the same time, different input sequences will produce different attention weights, which will make the deep learning model more powerful. The core formula of the self-attention mechanism is:

[0051]

[0052] Where Q is the query matrix, K is the key matrix, V is the value matrix, d k is the key vector dimension.

[0053] Unlike traditional long short-term memory (LSTM) networks, scalar long short-term memory (sLSTM) networks introduce exponential gating, normalization, and stabilizers. Exponential gating allows the input and forget gates to have exponential activation functions, allowing the model to dynamically adjust the storage and forgetting of information based on the input, effectively solving the problem of traditional long short-term memory networks being unable to flexibly revise storage decisions. Specifically:

[0054] The forward propagation process of the traditional long short-term memory network is as follows:

[0055] c t =f t ⊙c t-1 +it ⊙z t ;

[0056]

[0057] Among them, c t Represents the cell state, h t represents the hidden state, z t represents the cell input, i t represents the input gate, f t represents the forget gate, o t Represents the output gate; "~" represents the calculation intermediate state of the corresponding parameter; W z 、W i 、W f 、W o The input weight vectors corresponding to the input and cell input, input gate, forget gate and output gate are weighted R. z 、R i 、R f 、R o and correspond to the loop weights between the hidden state and the cell input, input gate, forget gate, and output gate, respectively; b z 、b i 、b f 、b o and are the corresponding bias terms; ψ() = tanh(), σ(x) = 1 / (1+exp(-x)).

[0058] In the scalar long short-term memory network, exponential gating, normalization state, and stabilization state are added to the traditional long short-term memory network. Its forward propagation process is as follows:

[0059] c t =f t ⊙c t-1 +i t ⊙z t ;

[0060] n t =f t ⊙n t-1 +i t ;

[0061]

[0062] Add a stabilization state so that:

[0063] m t =max(log(f t )+m t-1 ,log(i t )),

[0064] i′ t=exp(log(i t )-m t ),

[0065] f′ t =exp(log(f t )+m t-1 -m t ),

[0066] where n t is the normalized state, m t It is a stable gating state to prevent the value overflow caused by the exponential activation function. The meaning of other letters is the same as in the formula of the forward propagation process of the traditional long short-term memory network.

[0067] Building on the core mechanism of a scalar long short-term memory network, layer normalization (LN), causal convolution, swish activation function, Gaussian error linear unit (GELU activation function), block diagonal layer, and linear projection layer are added, and connected via a residual structure to build a neural network module with a scalar long short-term memory network. In this embodiment, when combining the self-attention mechanism with a neural network module with a scalar long short-term memory network, training is performed on an RTX4070 graphics card, using one self-attention module and two neural network modules with a scalar long short-term memory network.

[0068] Step S2-2: Add a position detection head and a fault detection head to the deep learning model obtained in step S2-1. Both the position detection head and the fault detection head contain an activation function layer, a random dropout layer, and a linear layer. Except for the output dimension, they are the same. The position detection head is used to predict the location of the fault. The prediction value dimension obtained in this embodiment is (64, 6). 64 means that 64 samples will be stacked in one training, and 6 means that there are 6 positions of the sampling points. The specific value of each represents the confidence in the position prediction; the fault detection head is used to predict the type of fault. The prediction value dimension obtained in this embodiment is (64, 14). 64 means that 64 samples will be stacked in one training, and 14 means that there are 14 positions of the sampling points. The specific value of each represents the confidence in the fault prediction. Step S2-3: Input the simulation data obtained in step S1 into the deep learning model to train the deep learning model. During training, a dynamic weighting approach is used to balance the two detection heads. The initial position detection head's loss weight is set to 1, and the initial fault detection head's loss weight is set to 0.5. The position detection head's loss weight decreases linearly to 0.2 with training, and the fault detection head's loss weight increases linearly to 1 with training. In this embodiment, training is stopped if there is no decrease in accuracy and loss for 10 consecutive rounds after 200 or 100 rounds of training. The training loss function includes position loss and classification loss, and the formula is as follows:

[0069] Ltotal =αL pos +βL cls ;

[0070] Among them, L pos is the position classification loss, L cls is the fault classification loss, and the cross entropy loss function is used; α and β are dynamic weights;

[0071] The further specific formula is as follows:

[0072] L pos =L CE (f pos (x), y pos );

[0073] L cls =L CE (f cls (x), y fal );

[0074] Among them, L CE represents the cross entropy loss function, f pos (·) represents the function from the network input to the position detection head output, y pos is the position label, y fal is the fault type label;

[0075]

[0076] Where t is the current training round and T is the total training rounds.

[0077] Step S3: Build a generative adversarial domain reconstruction transfer network, build the generator and discriminator, and use real data and simulated data to perform adversarial training on the generator and discriminator respectively.

[0078] Specifically, it includes the following sub-steps:

[0079] Step S3-1: Define the set of all real data as the target domain, define the set of all simulated data as the source domain, and use a generative adversarial network to align the target domain data with the source domain data.

[0080] Step S3-2: Based on the target domain data and source domain data, a one-dimensional convolutional neural network is used to build the generator and discriminator. The adversarial training process is as follows Figure 2 As shown, in this embodiment, Resnet18 is selected as the generator and discriminator. The loss function of the discriminator is the binary cross entropy loss function, and the loss function of the generator is the inverse of the cross entropy function. The parameters of the generator and the discriminator are not shared.

[0081] During training, it is agreed that label 0 represents the real data in the source domain and label 1 represents the generated data or target domain data; first, the generator parameters are frozen, and a batch of data is randomly selected from the source domain and input into the discriminator for training. The data format of this batch of data is the same as the simulated data format, which is (64, 1, 2000); then the output dimension of the discriminator is (64, 2), which indicates the probability prediction of each sample belonging to two categories (real data in the source domain / generated data); then the binary cross entropy loss function is used to calculate the error between the discriminator output and the real domain label (source domain data corresponds to label 0); the training goal at this stage is to minimize the discrimination error, that is, to improve the discriminator's ability to accurately distinguish between true and false data.

[0082] Next, freeze the discriminator parameters and unfreeze the generator parameters. A batch of data is randomly selected from the target domain in the format of (64, 1, 2000). This batch of target domain data is fed into the generator, and the generator outputs data in the same format of (64, 1, 2000). The generator's purpose here is to attempt to learn the distribution characteristics of the target domain data, thereby generating samples that are closer to real data. The data generated by the generator is then fed into the discriminator, whose output dimension is (64, 2), indicating the discriminator's probability prediction of each sample belonging to one of two categories (source domain real data / target domain generated data). A binary cross-entropy loss function is then used to calculate the error between the discriminator output and the domain label. The goal of training the generator is to maximize the discrimination error, meaning it is more difficult for the discriminator to distinguish between the generator-generated data and the real target domain data, thereby improving the generator's ability to deceive the discriminator.

[0083] During the training process, the above steps need to be repeated until the loss does not decrease for 10 consecutive rounds after 200 rounds or 100 rounds, and then the training is terminated.

[0084] Step S4: Freeze the generator parameters, save the trained generator parameters, and use the generator as a preprocessor for real data. When using the deep learning model built in step S2 for fault detection, first input the real data into the generator to obtain data aligned with the simulated data feature domain, and then input this data into the deep learning model for fault monitoring. A confidence threshold is set for the position head detection results and the fault detection head results. Results below the threshold are not processed, and results above the confidence threshold are determined to be the corresponding fault mode. This solution completes domain migration, eliminating the need to retrain the deep learning model built in step 2.

[0085] The above-mentioned offshore riser fault detection method achieves a synergistic integration of multiple technologies by using finite element simulation to acquire data, building a deep learning model based on a self-attention-scalar long short-term memory network, and constructing a network for generative adversarial domain reconstruction transfer. This method achieves several beneficial effects, specifically:

[0086] First, using simulation to obtain a large amount of data and using the high-fidelity data generated by finite element simulation as a training basis will help solve the problem of the difficulty and scarcity of actual marine riser failure data.

[0087] Secondly, a deep learning model based on the self-attention-scalar long short-term memory network was built, which has the advantage of high recognition accuracy. Unlike CNN (convolutional neural network) and RNN (recurrent neural network), the self-attention mechanism enables the model to better build global relationships and establish global associations between sampling points. The scalar long short-term memory network has both the ability to train quickly and the ability to reason quickly, and can dynamically adjust the stored information and capture information from multiple angles. That is, the model can dynamically adjust its behavior according to different inputs, which enhances the generalization ability of the model. At the same time, the scalar long short-term memory network also introduces a stable state to avoid the gradient explosion problem during training.

[0088] Third, to address the problem of poor recognition results when directly inputting real data due to domain differences between simulated data and real data, a generative adversarial network domain transfer method is adopted. This method can perform fault detection when real data has no labels and does not require retraining of the recognition model, which meets actual usage needs.

[0089] It should be noted that the parts not described in detail or expanded in the above scheme are all prior art, which do not belong to the improvements made by the present invention over the prior art, nor do they belong to the scope of protection of the technical solution of the present invention, so they will not be described in detail herein. Of course, the above content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of the embodiments of the present invention. The present invention is not limited to the above examples, and equal changes and improvements made by ordinary technicians in this technical field within the essential scope of the present invention should all fall within the scope of the patent coverage of the present invention.

Claims

1. A method for detecting faults in a marine riser, characterized in that: The steps include: Step S1: establishing a marine riser simulation model through finite element analysis, and performing simulation calculations to obtain simulation data; Step S2: Build a deep learning model based on the self-attention-scalar long short-term memory network, and train the deep learning model using the simulated data obtained in step S1; Step S3: Build a generative adversarial domain reconstruction transfer network, build a generator and a discriminator, and use real data and simulated data to perform adversarial training on the generator and the discriminator respectively; Step S4: Use the generator as the preprocessing part of the real data and use the deep learning model trained in step S2 for fault detection.

2. The method for detecting faults in a marine riser according to claim 1, wherein: The step S1 includes the following sub-steps: Step S1-1: using finite element analysis to establish a marine riser simulation model based on the parameters of the marine riser; Step S1-2: define boundary conditions; Step S1-3: Determine the monitoring point and set the fault type; Step S1-4: performing simulation calculations on the marine riser simulation model to obtain vibration response signals of multiple designated sampling points of the marine riser under different working conditions; Step S1-5: pre-processing the simulation data obtained in step S1-4, dividing the vibration response signal into a (batch, 1, seq) format; Step S1-6: perform label processing on the simulation data obtained in step S1-4.

3. The method for detecting faults in a marine riser according to claim 2, wherein: The step S2 includes the following sub-steps: Step S2-1: Combine the self-attention mechanism with a neural network module with a scalar long short-term memory network to obtain a deep learning model; Step S2-2: Adding a position detection head and a fault detection head to the deep learning model obtained in step S2-1, wherein the position detection head predicts the location of the fault, and the fault detection head predicts the type of fault; Step S2-3: Input the simulated data obtained in step S1 into the deep learning model to train the deep learning model.

4. The method for detecting faults in a marine riser according to claim 3, wherein: The step S3 includes the following sub-steps: Step S3-1: Define the set of real data as the target domain, define the set of simulated data as the source domain, and use a generative adversarial network to align the target domain data with the source domain data; Step S3-2: Based on the target domain data and source domain data, a one-dimensional convolutional neural network is used to build a generator and a discriminator.

5. The method for detecting faults in a marine riser according to claim 4, wherein: In step S4, the generator parameters are frozen, and real data is input into the generator to obtain data aligned with the feature domain of the simulated data, which is then input into the deep learning model for fault detection.

6. The method for detecting faults in a marine riser according to claim 5, wherein: In step S4, confidence thresholds are set for the detection results of the position detection head and the fault detection head. Detection results below the confidence threshold are not processed, and detection results above the confidence threshold are determined as corresponding fault modes.

7. The method for detecting faults in a marine riser according to claim 2, wherein: In step S1-1, the parameters of the marine riser include size, material and connection method; in step S1-2, the boundary conditions include the installation and restraint method of the marine riser, and the action method and magnitude of the external load on the marine riser.

8. The method for detecting faults in a marine riser according to claim 2, wherein: In step S1-6, the label corresponding to the vibration response signal has two dimensions, wherein the first dimension is the sampling position, represented by numbers 0 to n-1, where n is the number of sampling points; the second dimension is the fault type, including no fault, corrosion, crack, deformation, and mixed type faults.

9. The method for detecting faults in a marine riser according to claim 3, wherein: In step S2-3, dynamic weights are used to balance the two detection heads during training, and the loss weight of the initial position detection head is set to 1, and the loss weight of the initial fault detection head is set to 0.

5. The loss weight of the position detection head decreases linearly to 0.2 with training, and the loss weight of the fault detection head increases linearly to 1 with training.

10. The method for detecting faults in a marine riser according to claim 4, wherein: In step S3-2, a batch of data is randomly selected from the source domain and input into the discriminator for training. At this time, the generator parameters are frozen; Then, freeze the discriminator parameters, unfreeze the generator parameters, randomly select a batch of data from the target domain and input it into the generator, and then input the data generated by the generator into the discriminator for training; In this step, the above training process needs to be repeated until the preset conditions are met.

Images