Underwater channel simulation method based on deep learning
By adopting a deep learning-based underwater channel simulation method in the underwater optical communication system, combining the AMSAR module and a variety of deep learning networks, the problem of insufficient channel modeling and poor performance of the model in the existing technology is solved, and higher generalization capabilities and system reliability are achieved.
Patent Information
- Application Number
- CN202510047527.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-04-24
- Filing Date
- 2025-01-13
- Publication Date
- 2025-05-30
AI Technical Summary
When the existing underwater optical communication system faces complex underwater environments, channel modeling is not comprehensive enough, which affects signal transmission stability and reliability. The deep learning model has the problem of forgetting the original channel state on the new water body data set.
A underwater channel simulation method based on deep learning is proposed. AMSAR module is combined with CNN, MLP or GAN network, and a data feature extraction module is constructed through multi-head attention mechanism and adaptive normalization processing, and a residual module is added to the last layer to extract more feature information.
It improves the generalization capability of underwater optical communication systems, reduces the risk of overfitting, can be applicable to different types of water bodies, improves the performance and reliability of the system, and the generated signals are more realistic and reliable.
Smart Images
Figure CN120074713A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of underwater optical communication, and particularly relates to an underwater channel simulation method based on deep learning. Background Art
[0002] In the technical field of underwater optical communication, the current problems mainly include the dynamics, complexity, and variability of the underwater communication environment, resulting in problems such as distance limitation and signal quality degradation when optical signals propagate underwater, which have an important impact on the stability and reliability of underwater communication systems. A real and reliable underwater wireless optical communication channel model is needed to fully capture the influence of each part of the communication system on signal transmission. The traditional Monte Carlo simulation method simulates the propagation process of light beams in water by simulating the movement trajectories of a large number of photons. The modeling method mainly focuses on the physical attenuation of optical signals by the water environment, but ignores the influence of communication devices on signal transmission, which leads to an incomplete modeling of the channel characteristics of underwater wireless optical communication systems. In recent years, deep learning networks have made significant progress in channel modeling, but there are problems such as forgetting the original channel state on new water body datasets, and poor performance resulting in unsatisfactory model effects.
[0003] Therefore, in order to improve the generalization ability of underwater wireless optical communication systems, a new channel modeling method is needed to solve the shortcomings of the existing technology, improve the generalization performance of underwater optical communication systems, reduce the risk of overfitting, and be applicable to different types of water bodies, thereby enhancing the performance and reliability of the system. Summary of the Invention
[0004] In view of the above problems, the present invention provides an underwater channel simulation method based on deep learning, which is characterized in that it includes the following processes: Obtain the data of the real-time underwater transmission signal and the water body attenuation coefficient as the standard input data of the underwater channel simulation model; Input the obtained underwater transmission signal data and the water body attenuation coefficient into the trained underwater channel simulation model, and output the simulated received signal under this water body; The backbone network of the underwater channel simulation model adopts any one of CNN, MLP, or GAN, and an AMSAR module is embedded therein; the AMSAR module is based on the attention mechanism, maps the attention output of each head to the same fully connected layer through parallel connection, and through adaptive normalization processing, weights and sums the information of different heads according to different weights to construct a data feature extraction module. At the same time, based on the Res-Net residual module, a residual unit is added to the last layer to extract more feature information.
[0005] Preferably, for the dataset used to train the underwater channel simulation model, based on the established underwater laser digital communication system, the received signals, the corresponding transmitted signals, and the corresponding water quality attenuation coefficients under various water quality conditions are collected as the dataset for underwater optical communication; and the dataset is preprocessed, including the following processes: Perform operations of data encoding and reconstruction on the digital sequence. According to the value of each element in the digital sequence, zero or non-zero, generate an array of all zeros or all ones, then add a new dimension to the generated array and store it in a list; in each frame of signal, divide every 100 symbols into a sample unit, select the sliding window method, and extract 100 consecutive symbols every time a symbol slides to generate a new sample; standardize the water body attenuation degree of the training data and map it to the [0,1] interval for label encoding processing; the standardization formula is:
[0006] where is a regulation factor used to control the amplitude exceeding the maximum value; represents the attenuation value in tap water; the attenuation value in the test set; finally, splice the water quality attenuation label with the transmitted signal sample and adjust the shape of the new array to a dimension of batch size × 100 × 20 as the input data of the model; adjust the received signal to a dimension of batch size × 100 × 20 for use as a comparison for calculating the loss.
[0007] Preferably, the specific structure of the AMSAR module is: It consists of self-attention mechanisms with multiple heads, and each head calculates the attention weights separately , and the formula is as follows:
[0008]
[0009]
[0010]
[0011] where , and are the weight matrices for training and learning; X is the input data containing the transmitted signal and water body attenuation information; Q, K, and V are respectively abbreviated as the query vector, key vector, and value vector; is the dimension of K; Then, multiply the of each head by the corresponding numerical feature to obtain the weight output :
[0012] Connect the output of each head at this time in parallel to a fully connected layer, and the resulting linear transformation is as follows:
[0013] where is the bias parameter; represents the weight matrix of the linear transformation, and the weight size is adaptively adjusted according to the total number of heads. Its calculation formula is:
[0014] is the number of attention heads.
[0015] Preferably, the number of heads of the multi-head attention of the AMSAR module is set to 8 based on the number of heads of the backbone CNN network, and the number of heads is set to 4 based on the backbone MLP or GAN network.
[0016] Preferably, the last layer in the AMSAR module is a residual module; the residual module is connected to a rectified linear unit LR; the number of input channels of the residual block is 64 for the backbone CNN network, 2000 for the backbone MLP channel, and 20 for the number of channels of the generator in the backbone GAN; The formula for the result output by the residual module is as follows:
[0017] is the final output result; is the output after passing through the last fully connected layer in the AMSAR module; is the input data.
[0018] Preferably, the specific structure of embedding the AMSAR module into the CNN backbone network model is: including the initial AMSAR module, two one-dimensional convolutional layers, and one fully connected layer; the convolutional layer uses the method of max pooling for feature extraction; the activation function of the convolutional layer is ReLU; the final fully connected layer maps the feature vector to the dimension of the required output signal. Preferably, the specific structure of embedding the AMSAR module into the MLP backbone network model is: including four fully connected layers; the AMSAR module is placed at the beginning of the multi-layer perceptron, which can effectively capture the important features and relationships in the input data; the input data is used as the Q, K, and V values of the AMSAR module by calculating with the weight matrix, and the attention weights are calculated , obtain the final AMSAR output and then pass it through the MLP network; the activation function of the four fully connected layers is the ReLU activation function.
[0019] Preferably, the specific structure of embedding the AMSAR module into the GAN backbone network model is as follows: it includes a generator part and a discriminator part; the AMSAR module is added to the generator part; the input of the generator part includes x input data and z constraint conditions; the is a noise sample that satisfies the Gaussian distribution; the generator part includes an AMSAR module, a one-dimensional convolutional layer, a fully connected layer, ReLU and Tanh activation functions; the AMSAR module is located at the beginning of the generator to more effectively capture important features and relationships in the input data; after being input into the one-dimensional convolutional layer, it is used to further capture feature information and position information in the digital sequence; the discriminator part includes a one-dimensional convolutional layer and a ReLU activation function; the discriminator is used to distinguish whether the signal is generated by the generator or is a real signal.
[0020] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes an underwater channel simulation method based on deep learning and its model building method, and proposes an AMSAR module structure for improving generalization ability and embeds it into CNN, MLP and GAN networks respectively. The AMSAR module proposed by the present invention, based on the multi-head attention mechanism, parallelly connects the output of each head to a fully connected layer, and then performs adaptive normalization processing on the output of the fully connected layer. It realizes the weighted summation of different head information and constructs a data feature extraction module. The added residual module is used to solve problems such as gradient disappearance or gradient explosion in training. Compared with the traditional underwater modeling method, the present invention overcomes the limitation of only modeling the water body physical environment. Compared with the currently best underwater channel model, it extracts features more fully and captures the associations between different positions more effectively. At the same time, the present invention can be applied to a variety of new water bodies, expanding the application scope of the network model and having higher generalization performance.
[0021] The experimental results show that when tested in a new water body without training, the network model based on the AMSAR-Net structure shows excellent performance, and the generated signal is more real and reliable. The present invention can provide more effective guidance in the design and parameter selection of underwater optical communication systems, and promote the research in the field of underwater optical communication. Description of the Drawings
[0022] Figure 1 It is a schematic flow chart of the method for building an underwater channel simulation model proposed by the present invention.
[0023] Figure 2 It is the specific structure diagram of the AMSAR module in Example 1 of the present invention.
[0024] Figure 3 It is the structure diagram of the AMSAR-CNN network in Example 1 of the present invention.
[0025] Figure 4 It is the structure diagram of the AMSAR-MLP network in Example 1 of the present invention.
[0026] Figure 5 It is the structure diagram of the AMSAR-GAN network in Example 1 of the present invention.
[0027] Figure 6 In Example 1 of the present invention, the water attenuation is The true received signal diagram below.
[0028] Figure 7 In Example 1 of the present invention, using the water attenuation as The model trained with the dataset simulates and generates The true received signal diagram.
[0029] Figure 8 For the ordinary CNN network, using the water attenuation as The model trained with the dataset simulates and generates The true received signal diagram. Specific implementation manner
[0030] The following further illustrates the present invention in conjunction with specific embodiments.
[0031] The present invention proposes an underwater channel simulation method based on deep learning. Based on the constructed underwater channel simulator model, which is based on the method of multi-head attention mechanism, the output of each head is connected in parallel to a fully connected layer, and then the output of the fully connected layer is subjected to adaptive normalization processing. The added residual module is used to solve problems such as gradient disappearance or gradient explosion in training, and at the same time, it also realizes the effective extraction and learning of target features. The AMSAR network module is respectively embedded into the CNN, MLP, and GAN networks. The signals generated by the network are more real and reliable, and show excellent performance when tested in new waters without training, significantly improving the performance and generalization ability of the network. The construction process of the model is as Figure 1 shown, mainly including the following steps: Step 1: Build an underwater laser digital communication system, and collect the received signals, corresponding transmitted signals, and corresponding water quality attenuation coefficients under various water quality conditions as the dataset for underwater optical communication; Step 2: Preprocess the dataset collected in Step 1. For each water quality, the dataset is divided into a training set and a test set, and the ratio is 8:2; Step 3: Based on the multi-head attention mechanism structure, build an AMSAR model structure that improves the network generalization ability; the attention output of each head of the AMSAR network is mapped to the same fully connected layer through parallel connection, and through adaptive normalization processing, the information of different heads is weighted and summed according to different weights to construct a data feature extraction module; based on the Res-Net residual module, add a residual unit to the last layer to extract more feature information. Embed the AMSAR module into the CNN, MLP, and GAN backbone networks respectively, and adjust and optimize each network structure; use the transmitted and corresponding received signals and the water body attenuation coefficient as inputs, and the final output of the model is the received signal simulated under this water body.
[0032] Step 4: Use the preprocessed training set in Step 2 to train the AMSAR module built in Step 4 combined with the backbone CNN, MLP, and GAN network structures respectively. Step 5: Use the test set in Step 2 to test the trained network model, and select the model with more stable verified generated signals as the final model.
[0033] 1. Obtain the dataset In an underwater laser digital communication system, use the already modulated non-return-to-zero code-on-off keying (NRZ-OOK) signal as the transmitted signal, and use the synchronized signal after offline processing at the receiving end after passing through different water body environments as the received signal to make a dataset. At the same time, use the water body attenuation degree of this dataset as an additional feature and input it into the network model together with the data. Use the dataset with a low water body attenuation coefficient as the training set and test set of the network model, select the model with more stable verified generated signals as the final model, and then use the dataset with a high water body attenuation coefficient to evaluate the performance of the network model.
[0034] 2. Preprocessing of data: Preprocess the obtained transmitted and received data; perform operations of data encoding and reconstruction on the digital sequence. According to the value (zero or non-zero) of each element in the digital sequence, generate a corresponding all-zero or all-one array, and then add a new dimension to the generated array and store it in a list. In each frame of the signal, divide every 100 symbols into a sample unit, and use the sliding window method. Slide one symbol each time to extract 100 consecutive symbols to generate a new sample; standardize the water body attenuation degree of the training data and map it to the [0,1] interval for label encoding processing; the standardization formula is:
[0035] where is the adjustment factor used to control the amplitude exceeding the maximum value; represents the attenuation value in tap water; The attenuation values in the test set; finally, splice the water quality attenuation labels with the transmitted signal samples and adjust the shape of the new array to a dimension of batch size × 100 × 20 as the input data of the model; adjust the received signal to a dimension of batch size × 100 × 20 for use as a comparison for calculating the loss; 3. Model construction In the present invention, based on the network structure of AMSAR, it is respectively embedded into the CNN, MLP, and GAN backbone networks. Adjust and optimize each network structure, and the present invention aims to improve the performance and generalization ability of the model. The specific structure of the AMSAR module applied in the present invention is as Figure 2 shown.
[0036] It consists of the self-attention mechanism of multiple heads, and each head calculates the attention weights separately , and the formula is as follows:
[0037]
[0038]
[0039]
[0040] Among them, , and are the weight matrices for training and learning; X is the input data containing the transmitted signal and water body attenuation information; Q, K, and V are respectively abbreviated as the query vector, key vector, and value vector; is the dimension of K; Then, multiply the of each head by the corresponding numerical feature to obtain the weight output :
[0041] Connect the output of each head at this time in a parallel manner to a fully connected layer, and the resulting linear transformation is as follows:
[0042] Among them is the bias parameter; represents the weight matrix of the linear transformation, and adaptively adjusts the weight size according to the total number of heads. Its calculation formula is:
[0043] is the number of attention heads.
[0044] Then add a residual module to the last layer, and connect a rectified linear unit LR.
[0045] The formula for the result output by the residual module is as follows:
[0046] is the final output result; is the output after the last fully connected layer in the AMSAR module; is the input data.
[0047] The backbone network is based on CNN and includes an AMSAR module, two one-dimensional convolutional layers, and one fully connected layer; the convolutional layer uses the method of max pooling for feature extraction; the activation function of the convolutional layer is ReLU; the last fully connected layer maps the feature vector to the dimension of the required output signal. In the AMSAR module, the attention mechanism is used to pay more attention to the correlation between each data in the input information, and then each head is weighted and fused, so that the attention representations of different heads contribute differently to the final output. A residual module is added to the last layer. Then two convolutional layers perform feature extraction, and the AMSAR output is projected to the dimension of the output signal through the fully connected layer. The specific structure is as Figure 3 shown.
[0048] The backbone network is based on MLP and includes four fully connected layers; the specific structure of the AMSAR module embedded in the MLP backbone network model is: including four fully connected layers; the AMSAR module is placed at the beginning of the multi-layer perceptron, which can effectively capture the important features and relationships in the input data; the input data is used as the Q, K, and V values of the AMSAR module by calculating with the weight matrix, and the attention weights are calculated , and the final AMSAR output is obtained and then passed through a four-layer MLP network; the activation function of the four fully connected layers is the ReLU activation function. The specific structure is as Figure 4 shown.
[0049] The backbone network is based on GAN and includes a generator and a discriminator part. The AMSA module is added to the generator part. The input of the generator part includes the x input data and the z constraint condition; the Noise samples that follow a Gaussian distribution; the generator part includes an AMSAR module, a one-dimensional convolutional layer, a fully connected layer, ReLU and Tanh activation functions; the AMSAR module is located at the beginning of the generator to more effectively capture important features and relationships in the input data; after being input into the one-dimensional convolutional layer, it is used to further capture the feature information and position information in the digital sequence. The discriminator includes four layers of one-dimensional convolutional layers, and the last two layers are fully connected layers. The activation function of the first three layers is the ReLU function, and the other layers are all linear layer outputs. The specific structure is as Figure 5 shown.
[0050] 4. Model Training In this embodiment, the implementation of a channel modeling method for improving the generalization performance of an underwater optical communication system is based on the Linux operating system, the programming language is Python 3.6.10, the deep learning framework is TensorFlow 1.15.0, the CUDA version is 10.0, Adam is used as the optimizer, the learning rate of the backbone CNN is 0.001, the batch size is 15, the number of heads is set to 8, and a total of 200 rounds of training are performed. At the 100th round of training, the learning rate is changed to 0.0001.
[0051] The learning rate of the backbone GAN is 0.001, the batch size is 15, the number of heads is set to 4, and a total of 500 rounds of training are performed. The generator and the discriminator are alternately trained. For every 1 time the generator is trained, the discriminator is updated 5 times. At the 200th round, the learning rate becomes 0.0001.
[0052] The learning rate of the backbone MLP is 0.001, the batch size is 15, the number of heads is set to 8, and a total of 200 rounds of training are performed. At the 120th round of training, the learning rate is changed to 0.0001.
[0053] During the training process of the three models built, the Pearson correlation coefficient and the bit error rate error (BER) between the network-simulated generated signal and the real received signal are used as the evaluation indicators for validating the model. The model with a correlation coefficient close to 1 and a low bit error rate is selected as the optimal model.
[0054] The calculation formula of the Pearson correlation coefficient (Corr) is as follows:[[]]
[0055] where and represent the real received signal and its simulated generated signal respectively,[[]] represents and Covariance. The range of the correlation coefficient is between [-1, 1]. The closer it is to 1, the stronger the correlation. 0 indicates no correlation between the data.
[0056] The calculation formula of the bit error rate (BER) is as follows:
[0057] where represents the absolute value of * 5. Experimental Results In this embodiment, a new untrained water body is used as the test set to compare the performance of the proposed method with that of the original network, and the effectiveness of the proposed solution of the present invention is verified. The training set is the dataset with the attenuation coefficient at a wavelength of of 488 nm, and the test sets are the datasets with the attenuation coefficients at 488 nm of and respectively. The CNN, MLP, and GAN networks based on the AMSAR-Net module are respectively compared with the original network. CORR and BER are used as the evaluation indicators of the model. As shown, the following is the true received signal diagram with the water body attenuation of Figure 6 in the example of the present invention, the following is the true received signal diagram simulated and generated by the model trained with the dataset with the water body attenuation of Figure 7 in the example of the present invention, and the following is the true received signal diagram simulated and generated by the ordinary CNN network in the example of the present invention. Figure 8 The following is the true received signal diagram simulated and generated by the ordinary CNN network in the example of the present invention.
[0058] Table 1 Bit Error Rate Experimental Results
[0059] Table 1 shows the bit error rate experimental results of the model of the present invention trained with the water body with the attenuation coefficient of and tested with the water bodies of and respectively. AMSAR-Net is the network embedded with the AMSAR structure. Analyzing from the aspect of the bit error rate, the bit error rates of AMSAR-Net are all lower than those of the networks without the embedded AMSAR structure.
[0060] Table 2 Correlation Coefficient Experimental Results
[0061] Table 2 shows the correlation coefficient experimental results of the water bodies trained with the attenuation coefficient of and tested with the water bodies of and respectively. In terms of the correlation coefficient, the correlation coefficients of AMSAR-Net are all closer to 1 than those of the networks without the embedded AMSAR structure.
[0062] By comparing the results in Table 1 and Table 2, it can be concluded that the network model based on the AMSAR-Net structure has shown good adaptability and generalization ability in the water body tests under different attenuation coefficients, which proves the effectiveness and superiority of this network structure.
[0063] The above are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.
[0064] Although the specific implementation manners of the present invention have been described above, it is not a limitation to the protection scope of the present invention. Those skilled in the art should understand that on the basis of the technical solution of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.
Claims
1. An underwater channel simulation method based on deep learning, characterized in that: The process includes: Obtain real-time underwater signal transmission data and water attenuation coefficient as standard input data for underwater channel simulation model; The obtained underwater transmission signal data and water body attenuation coefficient are input into the trained underwater channel simulation model, and the simulated receiving signal under the water body is output; The backbone network of the underwater channel simulation model adopts any one of CNN, MLP or GAN, in which an AMSAR module is embedded; The AMSAR module is based on the attention mechanism. It maps the attention output of each head to the same fully connected layer through parallel connections. Through adaptive normalization processing, the information of different heads is weighted and summed according to different weights to build a data feature extraction module. At the same time, based on the Res-Net residual module, a residual unit is added to the last layer to extract more feature information.
2. The underwater channel simulation method based on deep learning according to claim 1, characterized in that: The dataset used to train the underwater channel simulation model is based on the underwater laser digital communication system. The received signals, corresponding transmitted signals and corresponding water quality attenuation coefficients under various water quality conditions are collected as the dataset for underwater optical communication. And preprocess the data set, including the following processes: The operation of encoding and reconstructing the digital sequence, generating an array of all zeros or all ones according to the value of each element in the digital sequence, zero or non-zero, and then adding a new dimension to the generated array and storing it in a list; In each frame signal, every 100 code elements are divided into a sample unit, and a sliding window method is used to extract 100 consecutive code elements every time a code element is slid to generate a new sample; the water body attenuation of the training data is standardized and mapped to the [0,1] interval for label encoding; the standardization formula is: in is the adjustment factor used to control the amplitude beyond the maximum value; Represents the attenuation value in tap water; The attenuation value in the test set; finally, the water quality attenuation label and the transmitted signal sample are concatenated to adjust the shape of the new array to the batch size × 100 × 20 dimensions as the input data of the model; the received signal is adjusted to the batch size × 100 × 20 dimensions for comparison of the calculated loss.
3. The underwater channel simulation method based on deep learning according to claim 1, characterized in that: The specific structure of the AMSAR module is: It consists of a self-attention mechanism with multiple heads, each of which calculates its own attention weight. , the formula is as follows: in, , and is the weight matrix for training and learning; X is the input data containing the transmitted signal and water attenuation information; Q, K and V are referred to as query vector, key vector and value vector respectively; is the dimension of K; Then, each head The corresponding numerical features Multiply to get the weight output : The output of each head at this time Connected in parallel to a fully connected layer, the resulting linear transformation is as follows: in is the bias parameter; The weight matrix representing the linear transformation adjusts the weight size adaptively according to the total number of heads. The calculation formula is: is the number of attention heads.
4. The underwater channel simulation method based on deep learning according to claim 1, characterized in that: The number of multi-heads of the AMSAR module's multi-head attention is set to 8 based on the backbone CNN network, and is set to 4 based on the backbone MLP or GAN network.
5. The underwater channel simulation method based on deep learning according to claim 3, characterized in that: The last layer in the AMSAR module is a residual module; the residual module is connected to a linear rectifier unit LR; the input channel of the residual block has 64 channels in the backbone CNN network, 2000 channels in the backbone MLP, and 20 channels in the generator of the backbone GAN; The result formula of the residual module output is as follows: is the final output result; is the output after the last fully connected layer in the AMSAR module; For input data.
6. The underwater channel simulation method based on deep learning as claimed in claim 3, characterized in that: The specific structure of embedding the AMSAR module into the CNN backbone network model is: It includes the initial AMSAR module, two one-dimensional convolutional layers, and a fully connected layer; the convolutional layer uses the maximum pooling method to extract features; the convolutional layer activation function is ReLU; the final fully connected layer maps the feature vector to the dimension of the required output signal.
7. The underwater channel simulation method based on deep learning as claimed in claim 3, characterized in that: The specific structure of embedding the AMSAR module into the MLP backbone network model is as follows: it includes four fully connected layers; the AMSAR module is placed at the beginning of the multi-layer perceptron, which can effectively capture the important features and relationships in the input data; the input data is calculated with the weight matrix as the Q, K, and V values of the AMSAR module, and the attention weight is calculated , the final AMSAR output is obtained and then passes through the MLP network; the activation function of the four fully connected layers is the ReLU activation function.
8. The underwater channel simulation method based on deep learning as claimed in claim 3, characterized in that: The specific structure of embedding the AMSAR module into the GAN backbone network model is as follows: including a generator part and a discriminator part; the AMSAR module is added to the generator part; the input of the generator part includes x input data and z constraint conditions; The noise sample satisfies the Gaussian distribution; the generator part includes an AMSAR module, a one-dimensional convolution layer, a fully connected layer, a ReLU and a Tanh activation function; the AMSAR module is located at the beginning of the generator to more effectively capture the important features and relationships in the input data; The input to the one-dimensional convolution layer is used to better capture the feature information and position information in the digital sequence; the discriminator part includes a one-dimensional convolution layer and a ReLU activation function; the discriminator is used to determine whether the signal is generated by the generator or a real signal.