A method and device for enhancing acoustic emission signals based on generative adversarial networks

Abstract

The present application provides an acoustic emission signal enhancement method and device based on a generative adversarial network, the method comprising: obtaining a random noise signal; inputting the random noise signal into a pre-trained generator of a generative adversarial network to obtain a generated signal similar to a real signal; wherein, during the training process of the generative adversarial network, the ability of the generator to generate acoustic emission signals is improved by adversarial training between the generator and the discriminator; the generator is composed of a bidirectional long short-term memory network layer and a self-attention mechanism layer, and the discriminator is composed of a bidirectional long short-term memory network layer. A bidirectional long short-term memory network is used to replace the fully connected network in the traditional generative adversarial network, and an attention mechanism is added between the layers to improve the quality of the generated signal, so that the signal generated by the generator is closer to the real signal and the effect of data enhancement is improved.

Description

Technical Field

[0001] The present application relates to the technical field of acoustic emission signal processing, and in particular to a method and device for enhancing acoustic emission signals based on a generative adversarial network. Background Art

[0002] In the field of structural health monitoring, acoustic emission (AE) technology, as a cutting-edge non-destructive detection method, has become the focus of research and application. It has shown great potential in the monitoring of complex and critical structures, and its application in concrete structure monitoring has become relatively mature.

[0003] However, although deep learning-based methods have shown great potential in structural health monitoring, their application still faces several key challenges. One of the most important challenges is that acoustic emission data for concrete often faces the problem of limited quantity and imbalanced class distribution. In practical applications, there are relatively few instances of severe damage in acoustic emission events, while instances of no damage or minor damage occupy the majority of the data set. This data imbalance poses a significant challenge to the training and performance optimization of deep learning models. Since the effectiveness of deep learning models depends heavily on large, balanced and diverse training data, class imbalance may lead to the performance degradation of the model when identifying less common damage types, thus affecting the overall detection accuracy and reliability. In view of these challenges, data augmentation technology has emerged as a key means to solve these problems. Traditional data augmentation techniques, such as adding noise and translation, can expand the number of datasets to a certain extent, but they have limitations in large-scale data generation. These methods mainly rely on simple changes to the original data, so they may lead to data duplication when a large number of enhancements are performed, thus limiting the ability of deep learning models to learn complex or unknown data patterns. Therefore, how to improve the quality of data-enhanced acoustic emission signals has become a technical issue that cannot be underestimated. Summary of the invention

[0004] In view of this, the purpose of this application is to provide an acoustic emission signal enhancement method and device based on a generative adversarial network, which adopts a bidirectional long short-term memory network to replace the fully connected network in the traditional generative adversarial network, and adds an attention mechanism between each layer to encourage the generative adversarial network to pay attention to important time points and features of the signal, improve the quality of the generated signal, so that the signal generated by the generator is closer to the real signal, and improve the effect of data enhancement.

[0005] In a first aspect, an embodiment of the present application provides an acoustic emission signal enhancement method based on a generative adversarial network, the acoustic emission signal enhancement method comprising:

[0006] Get a random noise signal;

[0007] The random noise signal is input into the generator of a pre-trained generative adversarial network to obtain a generated signal similar to a real signal; wherein the generative adversarial network includes a generator and a discriminator, and during the training process of the generative adversarial network, the ability of the generator to generate acoustic emission signals is improved by adversarial training between the generator and the discriminator; the generator is composed of a bidirectional long short-term memory network layer and a self-attention mechanism layer, and the discriminator is composed of a bidirectional long short-term memory network layer.

[0008] Furthermore, the generative adversarial network is obtained by the following steps:

[0009] Inputting a first noise signal into the generator to obtain a first generated signal;

[0010] Inputting the real signal and the first generated signal into the discriminator, fixing the weight parameters of the generator, training the discriminator, and continuously optimizing the weight parameters of the discriminator;

[0011] Inputting a second noise signal into the generator to obtain a second generated signal, setting a corresponding label for the second generated signal, and inputting the second generated signal into the discriminator, calculating a loss value based on a discrimination result of the discriminator, and updating a weight parameter of the generator in the form of error back propagation; wherein the label is 1;

[0012] Return to the step of inputting the first noise signal into the generator, repeatedly and alternately train the generator and the discriminator until the generator and the discriminator reach a Nash equilibrium state, and obtain the generative adversarial network.

[0013] Furthermore, the generator includes a multi-layer superimposed network structure and a fully connected layer, and each layer of the superimposed network structure includes a bidirectional long short-term memory network layer and a self-attention mechanism layer; the first noise signal is input into the generator to obtain a first generated signal, including:

[0014] The first noise signal is input into the first layer of the superimposed network structure in the multi-layer superimposed network structure, the output of the last layer of the superimposed network structure in the multi-layer superimposed network structure is then input into the fully connected layer, and the signal output by the fully connected layer is used as the first generated signal.

[0015] Furthermore, the loss value is calculated by the following formula:

[0016]

[0017] Wherein, loss represents the loss value, x represents the real signal, z represents the second noise signal, G represents the generator, D represents the discriminator, P data(x) represents the probability distribution of the real signal, P z(z) represents the probability distribution of the second noise signal, G(z) represents the second generated signal generated by the generator, n is the total number of data points in the signal sequence, G(t) and G(t+1) represent the generated signals generated by the generator at time t and t+1 respectively, λ is the weight coefficient, D(G(z)) represents the discrimination result of the discriminator on the second generated signal, and D(x) represents the discrimination result of the discriminator on the true signal.

[0018] In a second aspect, an embodiment of the present application further provides an acoustic emission signal enhancement device based on a generative adversarial network, the acoustic emission signal enhancement device comprising:

[0019] An acquisition module, used for acquiring a random noise signal;

[0020] An input module is used to input the random noise signal into the generator of a pre-trained generative adversarial network to obtain a generated signal similar to a real signal; wherein the generative adversarial network includes a generator and a discriminator, and during the training process of the generative adversarial network, the ability of the generator to generate acoustic emission signals is improved by adversarial training between the generator and the discriminator; the generator is composed of a bidirectional long short-term memory network layer and a self-attention mechanism layer, and the discriminator is composed of a bidirectional long short-term memory network layer.

[0021] Furthermore, the acoustic emission signal enhancement device also includes a training module, and the training module is used to obtain the generative adversarial network through the following steps:

[0022] Inputting a first noise signal into the generator to obtain a first generated signal;

[0023] Inputting the real signal and the first generated signal into the discriminator, fixing the weight parameters of the generator, training the discriminator, and continuously optimizing the weight parameters of the discriminator;

[0024] Inputting a second noise signal into the generator to obtain a second generated signal, setting a corresponding label for the second generated signal, and inputting the second generated signal into the discriminator, calculating a loss value based on a discrimination result of the discriminator, and updating a weight parameter of the generator in the form of error back propagation; wherein the label is 1;

[0025] Return to the step of inputting the first noise signal into the generator, repeatedly and alternately train the generator and the discriminator until the generator and the discriminator reach a Nash equilibrium state, and obtain the generative adversarial network.

[0026] Furthermore, the generator includes a multi-layer superimposed network structure and a fully connected layer, each of which includes a bidirectional long short-term memory network layer and a self-attention mechanism layer; when the training module is used to input the first noise signal into the generator to obtain a first generated signal, the training module is also used to:

[0027] The first noise signal is input into the first layer of the superimposed network structure in the multi-layer superimposed network structure, the output of the last layer of the superimposed network structure in the multi-layer superimposed network structure is then input into the fully connected layer, and the signal output by the fully connected layer is used as the first generated signal.

[0028] Furthermore, the training module is also used to calculate the loss value by the following formula:

[0029]

[0030] Wherein, loss represents the loss value, x represents the real signal, z represents the second noise signal, G represents the generator, D represents the discriminator, P data(x) represents the probability distribution of the real signal, P z(z) represents the probability distribution of the second noise signal, G(z) represents the second generated signal generated by the generator, n is the total number of data points in the signal sequence, G(t) and G(t+1) represent the generated signals generated by the generator at time t and t+1 respectively, λ is the weight coefficient, D(G(z)) represents the discrimination result of the discriminator on the second generated signal, and D(x) represents the discrimination result of the discriminator on the true signal.

[0031] In a third aspect, an embodiment of the present application further provides an electronic device, comprising: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory through the bus, and when the machine-readable instructions are executed by the processor, the steps of the acoustic emission signal enhancement method based on the generative adversarial network are performed as described above.

[0032] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the acoustic emission signal enhancement method based on a generative adversarial network as described above are executed.

[0033] An embodiment of the present application provides an acoustic emission signal enhancement method and device based on a generative adversarial network. The acoustic emission signal enhancement method comprises: first, obtaining a random noise signal; then, inputting the random noise signal into a pre-trained generator of a generative adversarial network to obtain a generated signal similar to a real signal; wherein the generative adversarial network comprises a generator and a discriminator, and during the training process of the generative adversarial network, the ability of the generator to generate acoustic emission signals is improved by adversarial training between the generator and the discriminator; the generator is composed of a bidirectional long short-term memory network layer and a self-attention mechanism layer, and the discriminator is composed of a bidirectional long short-term memory network layer.

[0034] This application uses a bidirectional long short-term memory network to replace the fully connected network in the traditional generative adversarial network, and adds an attention mechanism between each layer to encourage the generative adversarial network to pay attention to important time points and features of the signal, improve the quality of the generated signal, so that the signal generated by the generator is closer to the real signal and improve the effect of data enhancement.

[0035] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 A flowchart of an acoustic emission signal enhancement method based on a generative adversarial network provided in an embodiment of the present application;

[0038] Figure 2 A schematic diagram of the structure of a generative adversarial network provided in an embodiment of the present application;

[0039] Figure 3 A schematic diagram of the structure of an acoustic emission signal enhancement device based on a generative adversarial network provided in an embodiment of the present application;

[0040] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0041] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application usually described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, each other embodiment obtained by those skilled in the art without making creative work belongs to the scope of protection of the present application.

[0042] First, the application scenarios to which the present application is applicable are introduced. The present application can be applied to the field of acoustic emission signal processing technology.

[0043] In the field of structural health monitoring, acoustic emission (AE) technology, as a cutting-edge non-destructive detection method, has become the focus of research and application. It has shown great potential in the monitoring of complex and critical structures, and its application in concrete structure monitoring has become relatively mature.

[0044] However, although deep learning-based methods have shown great potential in structural health monitoring, their application still faces several key challenges. One of the most important challenges is that acoustic emission data for concrete often faces the problem of limited quantity and imbalanced class distribution. In practical applications, there are relatively few instances of severe damage in acoustic emission events, while instances of no damage or minor damage occupy the majority of the data set. This data imbalance poses a significant challenge to the training and performance optimization of deep learning models. Since the effectiveness of deep learning models depends heavily on large, balanced and diverse training data, class imbalance may lead to the performance degradation of the model when identifying less common damage types, thus affecting the overall detection accuracy and reliability. In view of these challenges, data augmentation technology has emerged as a key means to solve these problems. Traditional data augmentation techniques, such as adding noise and translation, can expand the number of datasets to a certain extent, but they have limitations in large-scale data generation. These methods mainly rely on simple changes to the original data, so they may lead to data duplication when a large number of enhancements are performed, thus limiting the ability of deep learning models to learn complex or unknown data patterns. Therefore, how to improve the quality of data-enhanced acoustic emission signals has become a technical issue that cannot be underestimated.

[0045] Based on this, an embodiment of the present application provides an acoustic emission signal enhancement method based on a generative adversarial network to improve the quality of the generated signal so that the data-enhanced acoustic emission signal is closer to the original acoustic emission signal, thereby improving the effect of data enhancement.

[0046] See also Figure 1 , Figure 1 A flowchart of an acoustic emission signal enhancement method based on a generative adversarial network provided in an embodiment of the present application. Figure 1 As shown in, the acoustic emission signal enhancement method provided in the embodiment of the present application includes:

[0047] S101. Obtain a random noise signal.

[0048] Regarding the above step S101, during specific implementation, a random noise signal is acquired.

[0049] S102, inputting the random noise signal into a generator of a pre-trained generative adversarial network to obtain a generated signal similar to a real signal.

[0050] Here, the purpose of the generative adversarial network is to generate acoustic emission signals that are closer to real signals in order to expand the number of acoustic emission signals and expand the original data set.

[0051] For the above step S102, in the specific implementation, the acquired random noise signal is input into the generator of the pre-trained generative adversarial network to obtain a generated signal similar to the real signal, thereby achieving the purpose of data enhancement. Specifically, the random noise signal is input into the generator part of the generative adversarial network, and data enhancement is performed to obtain a generated signal with similar characteristics to the real signal, so as to achieve the purpose of expanding the concrete acoustic emission data. These data-enhanced acoustic emission signals maintain the characteristics of the real signal to a certain extent and are closer to the real signal. Such a data enhancement method helps to improve the diversity of training data in the data set, thereby expanding the original data set, and when using the training data in the data set to train the concrete damage recognition model, the generalization ability of the concrete damage recognition model is improved, thereby further improving the performance and effect of the concrete damage recognition model.

[0052] See also Figure 2 , Figure 2 This is a schematic diagram of the structure of a generative adversarial network provided in an embodiment of the present application. Figure 2 As shown in , the generative adversarial network includes a generator and a discriminator. During the training process of the generative adversarial network, the ability of the generator to generate acoustic emission signals is improved by adversarial training between the generator and the discriminator. In this application, the generator is composed of a bidirectional long short-term memory network layer and a self-attention mechanism layer, and the discriminator is composed of a bidirectional long short-term memory network layer.

[0053] Specifically, generative adversarial networks are an important research direction of deep learning unsupervised algorithms. They have the ability to fit the original data distribution and can achieve data regeneration. They have been widely used in many fields such as computer vision, natural language processing, and network security, and still have greater development potential. Figure 2 As shown in Figure 1, the generative adversarial network consists of a generator G and a discriminator D. The generator generates data, and the discriminator determines whether the input data is real data or generated data. The two networks improve together in the process of dynamic game and finally reach Nash equilibrium.

[0054] The LSTM network effectively regulates the flow of information between network units by introducing a unique gating mechanism (including input gate, forget gate, and output gate). This structure enables the network to retain long-term dependencies in time series, thereby performing well in various sequence prediction and classification tasks. Based on the LSTM network, the bidirectional LSTM network further expands its capabilities. The bidirectional LSTM network combines two independent LSTM networks to process the forward and reverse information flows of sequence data respectively. This structure enables the network to capture more comprehensive contextual information at each point in the sequence, thereby more accurately understanding and analyzing data with complex time dependencies. Since the bidirectional LSTM network can consider both forward and backward information at the same time, it is particularly suitable for application scenarios that require a comprehensive understanding of the data context. For example, in acoustic emission signal analysis, this feature of the bidirectional LSTM network enables it to more effectively capture the temporal characteristics of the signal, thereby improving the accuracy of damage detection and classification.

[0055] The attention mechanism is an important neural network component that is widely used in natural language processing, computer vision, speech recognition and other fields. Its core idea is to give neural networks the ability to dynamically focus, so that they can focus on the most critical information when processing sequence data. Compared with the static and equal information processing method in traditional methods, the attention mechanism provides a more flexible and efficient information capture strategy.

[0056] This application uses a bidirectional long short-term memory network to replace the fully connected network in the traditional generative adversarial network, and adds an attention mechanism between each layer to encourage the model to pay attention to important time points and features of the signal to improve the quality of the generated signal.

[0057] Specifically, according to the embodiment provided in this application, the generative adversarial network is obtained through the following steps:

[0058] A: Inputting a first noise signal into the generator to obtain a first generated signal.

[0059] Here, for the above step A, in the specific implementation, firstly, a first noise signal is acquired, and then the first noise signal is input into the generator to obtain a first generated signal generated by the generator.

[0060] like Figure 2 As shown in , the generator provided in the embodiment of the present application includes a multi-layer superimposed network structure and a fully connected layer, and each superimposed network structure includes a bidirectional long short-term memory network layer and a self-attention mechanism layer. For the above step A, the first noise signal is input into the generator to obtain a first generated signal, including:

[0061] The first noise signal is input into the first layer of the superimposed network structure in the multi-layer superimposed network structure, the output of the last layer of the superimposed network structure in the multi-layer superimposed network structure is then input into the fully connected layer, and the signal output by the fully connected layer is used as the first generated signal.

[0062] Here, for the above steps, in the specific implementation, since the generator includes a multi-layer superimposed network structure, the first noise signal is first input into the first layer of the superimposed network structure in the multi-layer superimposed network structure. Specifically, the first noise signal is input into the bidirectional long short-term memory network layer in the first layer of the superimposed network structure to extract the timing characteristics of the first noise signal. BiLSTM is an extended version of the LSTM model, that is, sequence information is obtained in the forward and reverse directions, so that the network can have context information at the same time. The LSTM unit consists of a forget gate, an input gate, and an output gate. The forget gate controls how much of the previous memory is retained, the input gate controls the influence of the current input, the output gate controls the output value, and the forget gate f t Determines which information of the previous unit is deleted, o t Decide to output information by filtering the input and the previous hidden state. Assume that the input time series data is x = (x1, x2, x3, ..., x n ). The long short-term memory network mechanism can be expressed by the following formula:

[0063] f t =σ(W f [h t-1 ,x t ]+b f )

[0064] i t =σ(W i [h t-1 ,x t ]+b i )

[0065]

[0066]

[0067] o t =σ(W o [h t-1 ,x t ]+b o )

[0068] h t =o t tanh(C t )

[0069] Where σ represents the Sigmoid function; i t and h t represents the activation vector of the hidden state and input gate; C t and Update state for cell state and time step, W i ,W c ,W f ,W o is the corresponding weight matrix, b i ,b c ,b f ,b o is the corresponding deviation; the symbol (·) represents element-wise multiplication. This feature of the bidirectional long short-term memory network enables it to capture the timing characteristics of the signal more effectively, thereby obtaining the timing characteristics of the first noise signal.

[0070] The temporal features are then fed into the self-attention mechanism layer in the first layer of the stacked network structure. The attention layer is an important part of causing the network to pay attention to the signal. Figure 2The first half of the real signal has a large amplitude, indicating that there is more information and features in this part. The implementation of the attention mechanism in the neural network is mainly completed by assigning a weight to each element in the sequence. These weights represent the relative importance of each element to the current task. First, a scoring function is used to calculate the similarity between the query (generally the hidden state of the previous step) and each key (input element) in the sequence to generate an attention score. These scores are then fed into the softmax function for normalization to obtain the attention weight of each element. These weights are multiplied and summed with the hidden state of the sequence element to generate a context vector, which represents the weighted representation of the sequence in the current step. This context vector is then used to generate the output of the current time step and may be used as the input of the next step. In this way, the attention mechanism enables the network to dynamically focus on the most relevant part of the sequence to the current task when processing each time step, effectively improving the processing ability of long sequences and complex problems. Specifically, the attention mechanism has three main elements: query, key, and value. By calculating the similarity score between query and key, the attention mechanism determines the weight of each key, and then applies the weight to the corresponding value to generate a weighted output. The mathematical expression is shown in the following formula:

[0071]

[0072] Among them, Q is the query vector, which is used to represent the content of interest; K is the key matrix, which is used to represent the input content; V is the value matrix, which is used to represent the content to be obtained; d k is the dimension of the query and key vectors; the softmax function is used to calculate the normalized attention weights, ensuring that their sum is 1. When calculating the attention weights, higher similarity scores will result in higher weights, making the network pay more attention to important information.

[0073] After passing through the first layer of superimposed network structure, it is input into the second layer of superimposed network structure, and so on. After multiple layers of superposition, the output of the last layer of superimposed network structure in the multi-layer superimposed network structure is input into the fully connected layer of the generator, and the signal output by the fully connected layer is used as the first generated signal.

[0074] B: Inputting the real signal and the first generated signal into the discriminator, fixing the weight parameters of the generator, training the discriminator, and continuously optimizing the weight parameters of the discriminator.

[0075] For the above step B, in the specific implementation, the first stage of training is to input the real signal and the first generated signal generated by the generator into the discriminator. This process only updates the weight parameters of the discriminator, which is used to train the discriminator to distinguish between true and false signals. Specifically, the first noise signal is output as the first generated signal after passing through the generator. Then the real signal and the first generated signal are fed to the discriminator as input, and the real signal is labeled 1, and the first generated signal is labeled 0. The weight parameters of the generator are fixed, the discriminator is trained, and the weight parameters of the discriminator are continuously optimized. Since the discriminator can easily distinguish between the two signals, the discriminator only uses the BiLSTM layer. Through the above process, a simple two-class discriminator training is completed. At this time, the discriminator is easier to distinguish between the real signal and the generated signal because the quality of the generated signal is poor in the early stage.

[0076] C: Input the second noise signal into the generator to obtain a second generated signal, set a corresponding label for the second generated signal, and input the second generated signal into the discriminator, calculate the loss value based on the discrimination result of the discriminator, and update the weight parameters of the generator in the form of error back propagation.

[0077] Here, the label set to 1 is for the second generated signal.

[0078] For the above step C, in the specific implementation, the second stage of training fixes the weight of the discriminator and trains the weight of the generator. The second noise signal z is input into the generator to obtain the second generated signal G(z). At this time, the second generated signal is labeled 1. When the error is back-propagated, the generator is prompted to update the weight so that the generator output G(z) is judged as 1 by the discriminator, which means that the signal generated by the generator cannot be distinguished as true or false by the discriminator. The discriminator outputs the judgment result, which is the discriminator judging the true signal x and the second generated signal G(z) to judge whether each sample is true or false, and outputs D(x) and D(G(z)) respectively. Then, the loss value is calculated based on the judgment result of the discriminator, and the weight parameters of the generator are updated in the form of error back-propagation.

[0079] Specifically, according to the embodiment provided in this application, the loss value is calculated by the following formula:

[0080]

[0081] Among them, loss represents the loss value, x represents the real signal, z represents the second noise signal, G represents the generator, D represents the discriminator, P data(x) represents the probability distribution of the real signal, P z(z)represents the probability distribution of the second noise signal, G(z) represents the second generated signal generated by the generator, n is the total number of data points in the signal sequence, G(t) and G(t+1) represent the generated signals generated by the generator at time t and t+1 respectively, λ is the weight coefficient, D(G(z)) represents the discrimination result of the discriminator on the second generated signal, and D(x) represents the discrimination result of the discriminator on the real signal.

[0082] Here, as shown in the above formula, a time difference regularization term is added to the generator loss function, because it is found that the signal generated by the generator does have the trend of the real signal, but the local oscillation is more serious and the overall is not smooth. This regularization term is the mean of the square of the point at the next moment of the signal minus the point at the previous moment. Therefore, when optimizing the loss function, the difference between adjacent points of the signal is taken into account, and the gap between adjacent points is gradually narrowed, making the signal smoother and closer to the real signal.

[0083] D: Return to the step of inputting the first noise signal into the generator, repeatedly and alternately train the generator and the discriminator until the generator and the discriminator reach a Nash equilibrium state, and obtain the generative adversarial network.

[0084] Here, the Nash equilibrium state means that the discriminator cannot determine whether the generated signal generated by the generator is true or false.

[0085] For the above step D, in the specific implementation, the above steps A-C are only a cycle of adversarial training, and the above process will be repeated later, returning to the step of inputting the first noise signal into the generator in the above step A, and repeatedly training the generator and the discriminator alternately. In the training of the generative adversarial network, the generator and the discriminator compete with each other and continuously optimize their respective parameters. When the signal generated by the generator gradually approaches the real signal, and the discriminator finds it difficult to distinguish the generated signal from the real signal, the Nash equilibrium state is reached. In this state, the signal generated by the generator is of high quality, and the discriminator cannot accurately judge the true or false, so that the generated signal is more realistic. In summary, this embodiment can effectively train the generative adversarial network by generating signal input discriminator, Nash equilibrium state and iterative adversarial training and other operations, so that it generates realistic and diverse signals, thereby improving the effect of data enhancement and the quality of the generated signal.

[0086] In the specific implementation, in the initial training stage, when the discriminator is trained, the discriminator can usually distinguish the real signal from the generated signal with a high accuracy, because the generator has not yet learned the data distribution of the real signal. However, when the generator is trained, the recognition ability of the discriminator is weakened, resulting in a decrease in accuracy. This phenomenon reflects that the learning ability of the generator is increasing. Such repeated training leads to repeated oscillations in the early stage. When the generative adversarial network is close to the Nash equilibrium, the generator can imitate the distribution of the real signal, and the quality of the generated signal is enough to confuse the discriminator, so that the discriminator cannot accurately distinguish the real signal from the generated signal. The accuracy of the discriminator gradually approaches 50%, but this is not the attenuation of the discriminator performance, but the improvement of the generator performance, which reflects the success of the model training, that is, the generator has learned a complex distribution that matches the real signal. The introduction of time difference mean square error regularization in the loss function has a significant improvement in the quality of the generated signal, because the mean square error can encourage the generated signal to change little between adjacent points, thereby increasing the smoothness of the signal. The signal generated by the generative adversarial network in this application is smoother and more similar to the real data in details.

[0087] In a traditional generative adversarial network, the generator and the discriminator are usually composed of a fully connected neural network. However, considering the time domain characteristics of the acoustic emission signal, in the acoustic emission signal enhancement method based on the generative adversarial network provided in the embodiment of the present application, a bidirectional long short-term memory network is used to replace the fully connected network in the traditional generative adversarial network, so as to more effectively capture and utilize the temporal information of the signal data. The bidirectional structure of the bidirectional long short-term memory network enables the model to comprehensively consider the forward and backward contextual information of the signal, thereby improving the understanding and generation ability of the temporal data. In addition, the present application introduces an attention mechanism between each bidirectional long short-term memory network layer to further improve the model's attention to the important parts of the signal. The attention mechanism gives the model a higher weight to key information when generating data, so that the model can focus more on the time points and features that have a decisive influence when processing complex signal data. In order to further improve the quality of the generated data and make it closer to the real signal data, the generator's loss function introduces time difference mean square error regularization, which encourages the generator to produce smooth and coherent signals to ensure that the signal is visually similar to the real signal. This improves the quality of the generated signal, making the signal generated by the generator closer to the real signal and improving the effect of data enhancement.

[0088] See also Figure 3 , Figure 3 This is a schematic diagram of the structure of an acoustic emission signal enhancement device based on a generative adversarial network provided in an embodiment of the present application. Figure 3 As shown in , the acoustic emission signal enhancement device 300 includes:

[0089] An acquisition module 301 is used to acquire a random noise signal;

[0090] The input module 302 is used to input the random noise signal into the generator of a pre-trained generative adversarial network to obtain a generated signal similar to a real signal; wherein the generative adversarial network includes a generator and a discriminator, and during the training process of the generative adversarial network, the ability of the generator to generate acoustic emission signals is improved by adversarial training between the generator and the discriminator; the generator is composed of a bidirectional long short-term memory network layer and a self-attention mechanism layer, and the discriminator is composed of a bidirectional long short-term memory network layer.

[0091] Furthermore, the acoustic emission signal enhancement device also includes a training module, and the training module is used to obtain the generative adversarial network through the following steps:

[0092] Inputting a first noise signal into the generator to obtain a first generated signal;

[0093] Inputting the real signal and the first generated signal into the discriminator, fixing the weight parameters of the generator, training the discriminator, and continuously optimizing the weight parameters of the discriminator;

[0094] Inputting a second noise signal into the generator to obtain a second generated signal, setting a corresponding label for the second generated signal, and inputting the second generated signal into the discriminator, calculating a loss value based on a discrimination result of the discriminator, and updating a weight parameter of the generator in the form of error back propagation; wherein the label is 1;

[0095] Return to the step of inputting the first noise signal into the generator, repeatedly and alternately train the generator and the discriminator until the generator and the discriminator reach a Nash equilibrium state, and obtain the generative adversarial network.

[0096] Furthermore, the generator includes a multi-layer superimposed network structure and a fully connected layer, each of which includes a bidirectional long short-term memory network layer and a self-attention mechanism layer; when the training module is used to input the first noise signal into the generator to obtain a first generated signal, the training module is also used to:

[0097] The first noise signal is input into the first layer of the superimposed network structure in the multi-layer superimposed network structure, the output of the last layer of the superimposed network structure in the multi-layer superimposed network structure is then input into the fully connected layer, and the signal output by the fully connected layer is used as the first generated signal.

[0098] Furthermore, the training module is also used to calculate the loss value by the following formula:

[0099]

[0100] Wherein, loss represents the loss value, x represents the real signal, z represents the second noise signal, G represents the generator, D represents the discriminator, P data(x) represents the probability distribution of the real signal, P z(z) represents the probability distribution of the second noise signal, G(z) represents the second generated signal generated by the generator, n is the total number of data points in the signal sequence, G(t) and G(t+1) represent the generated signals generated by the generator at time t and t+1 respectively, λ is the weight coefficient, D(G(z)) represents the discrimination result of the discriminator on the second generated signal, and D(x) represents the discrimination result of the discriminator on the true signal.

[0101] See also Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 4 As shown in , the electronic device 400 includes a processor 410 , a memory 420 and a bus 430 .

[0102] The memory 420 stores machine-readable instructions executable by the processor 410. When the electronic device 400 is running, the processor 410 communicates with the memory 420 via the bus 430. When the machine-readable instructions are executed by the processor 410, the above-mentioned Figure 1 The steps of the acoustic emission signal enhancement method based on the generative adversarial network in the method embodiment shown in the figure can be specifically implemented by referring to the method embodiment, which will not be described in detail here.

[0103] The present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the computer program can execute the above-mentioned Figure 1 The steps of the acoustic emission signal enhancement method based on the generative adversarial network in the method embodiment shown in the figure can be specifically implemented by referring to the method embodiment, which will not be described in detail here.

[0104] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0105] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0106] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0107] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0108] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application can essentially be embodied in the form of a software product, or in other words, the part that contributes to the prior art or the part of the technical solution. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0109] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present application, which are used to illustrate the technical solutions of the present application, rather than to limit them. The protection scope of the present application is not limited thereto. Although the present application is described in detail with reference to the above-mentioned embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-mentioned embodiments within the technical scope disclosed in the present application, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A method for enhancing acoustic emission signals based on a generative adversarial network, characterized in that: The acoustic emission signal enhancement method comprises: Get a random noise signal; The random noise signal is input into the generator of a pre-trained generative adversarial network to obtain a generated signal similar to a real signal; wherein the generative adversarial network includes a generator and a discriminator, and during the training process of the generative adversarial network, the ability of the generator to generate acoustic emission signals is improved by adversarial training between the generator and the discriminator; the generator is composed of a bidirectional long short-term memory network layer and a self-attention mechanism layer, and the discriminator is composed of a bidirectional long short-term memory network layer; The generative adversarial network is obtained by the following steps: Inputting a first noise signal into the generator to obtain a first generated signal; Inputting the real signal and the first generated signal into the discriminator, fixing the weight parameters of the generator, training the discriminator, and continuously optimizing the weight parameters of the discriminator; Inputting a second noise signal into the generator to obtain a second generated signal, setting a corresponding label for the second generated signal, and inputting the second generated signal into the discriminator, calculating a loss value based on a discrimination result of the discriminator, and updating a weight parameter of the generator in the form of error back propagation; wherein the label is 1; Returning to the step of inputting the first noise signal into the generator, repeatedly and alternately training the generator and the discriminator until the generator and the discriminator reach a Nash equilibrium state, thereby obtaining the generative adversarial network; The loss value is calculated by the following formula: Wherein, loss represents the loss value, x represents the real signal, z represents the second noise signal, G represents the generator, D represents the discriminator, and p data(x) represents the probability distribution of the real signal, p z (z) represents the probability distribution of the second noise signal, G(z) represents the second generated signal generated by the generator, n is the total number of data points in the signal sequence, G(t) and G(t+1) represent the generated signals generated by the generator at time t and t+1 respectively, λ is the weight coefficient, D(G(z)) represents the discrimination result of the discriminator on the second generated signal, and D(x) represents the discrimination result of the discriminator on the true signal.

2. The method for enhancing acoustic emission signals according to claim 1, characterized in that: The generator includes a multi-layer superimposed network structure and a fully connected layer, each layer of the superimposed network structure includes a bidirectional long short-term memory network layer and a self-attention mechanism layer; the first noise signal is input into the generator to obtain a first generated signal, including: The first noise signal is input into the first layer of the superimposed network structure in the multi-layer superimposed network structure, the output of the last layer of the superimposed network structure in the multi-layer superimposed network structure is then input into the fully connected layer, and the signal output by the fully connected layer is used as the first generated signal.

3. An acoustic emission signal enhancement device based on a generative adversarial network, characterized in that: The acoustic emission signal enhancement device comprises: An acquisition module, used for acquiring a random noise signal; An input module, used for inputting the random noise signal into the generator of a pre-trained generative adversarial network to obtain a generated signal similar to a real signal; wherein the generative adversarial network includes a generator and a discriminator, and during the training process of the generative adversarial network, the ability of the generator to generate acoustic emission signals is improved by adversarial training between the generator and the discriminator; the generator is composed of a bidirectional long short-term memory network layer and a self-attention mechanism layer, and the discriminator is composed of a bidirectional long short-term memory network layer; The acoustic emission signal enhancement device further includes a training module, which is used to obtain the generative adversarial network through the following steps: Inputting a first noise signal into the generator to obtain a first generated signal; Inputting the real signal and the first generated signal into the discriminator, fixing the weight parameters of the generator, training the discriminator, and continuously optimizing the weight parameters of the discriminator; Inputting a second noise signal into the generator to obtain a second generated signal, setting a corresponding label for the second generated signal, and inputting the second generated signal into the discriminator, calculating a loss value based on a discrimination result of the discriminator, and updating a weight parameter of the generator in the form of error back propagation; wherein the label is 1; Returning to the step of inputting the first noise signal into the generator, repeatedly and alternately training the generator and the discriminator until the generator and the discriminator reach a Nash equilibrium state, thereby obtaining the generative adversarial network; The training module is also used to calculate the loss value by the following formula: Wherein, loss represents the loss value, x represents the real signal, z represents the second noise signal, G represents the generator, D represents the discriminator, and p data(x) represents the probability distribution of the real signal, p z (z) represents the probability distribution of the second noise signal, G(z) represents the second generated signal generated by the generator, n is the total number of data points in the signal sequence, G(t) and G(t+1) represent the generated signals generated by the generator at time t and t+1 respectively, λ is the weight coefficient, D(G(z)) represents the discrimination result of the discriminator on the second generated signal, and D(x) represents the discrimination result of the discriminator on the true signal.

4. The acoustic emission signal enhancement device according to claim 3, characterized in that: The generator includes a multi-layer superimposed network structure and a fully connected layer, each of which includes a bidirectional long short-term memory network layer and a self-attention mechanism layer; when the training module is used to input the first noise signal into the generator to obtain a first generated signal, the training module is also used to: The first noise signal is input into the first layer of the superimposed network structure in the multi-layer superimposed network structure, the output of the last layer of the superimposed network structure in the multi-layer superimposed network structure is then input into the fully connected layer, and the signal output by the fully connected layer is used as the first generated signal.

5. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate through the bus, and the machine-readable instructions are executed by the processor to perform the steps of the acoustic emission signal enhancement method based on the generative adversarial network as described in any one of claims 1 to 2.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the acoustic emission signal enhancement method based on a generative adversarial network as claimed in any one of claims 1 to 2 are executed.

Images