Rock burst data generation method and device based on generative adversarial network and product

Abstract

The invention discloses a generative adversarial network-based rockburst data generation method, device and product. The method comprises the following steps of: establishing a rockburst data generation model which takes rockburst geological condition parameters and random noise as input and takes rockburst data corresponding to the rockburst geological condition parameters as output; therefore, in actual use, different types of rockburst data can be obtained only by inputting different rockburst geological condition parameters added with random noise into the rockburst data generation model; therefore, diversified rockburst data can be obtained by inputting rockburst geological condition parameters under different conditions, so that the problems of data scarcity and sample imbalance in the traditional technology are solved, meanwhile, the acquisition of the rockburst data does not need experimental simulation and field acquisition, and the cost and the risk are further reduced; therefore, the invention provides a brand-new rockburst data acquisition mode, so that the method is very suitable for large-scale application and popularization.

Description

Technical Field

[0001] The present invention relates to the technical field of rockburst data generation based on neural networks, and in particular to a rockburst data generation method, device and product based on a generative adversarial network. Background Art

[0002] At present, the commonly used rockburst data types mainly include microseismic waveforms, acoustic emission signals, stress-strain curves, infrared thermal imaging images, etc. These data usually depend on environmental parameters such as geological structure, rock type, and groundwater pressure. Among them, traditional rockburst data collection methods mostly rely on field experiments and real-time monitoring, which have the following shortcomings: (1) Data scarcity. Rockburst events are extremely rare, especially rockburst data under extreme situations. As a result, traditional technologies cannot fully cover the changes in various situations, resulting in poor representativeness and integrity of the data; (2) Sample imbalance. Since rockburst events usually have strong nonlinearity, extremeness and randomness, traditional data sets tend to be biased towards normal or mild rockburst events. Extreme events such as rare high-energy rockbursts can hardly be reflected in real data; (3) High cost and high risk. Experimental simulation and data collection of rockbursts usually need to be carried out in a highly dangerous environment, which has great safety risks and cost investment. Therefore, based on the above shortcomings, how to provide a low-cost, low-risk, and generative adversarial network-based rockburst data generation method that can make up for data scarcity and imbalance has become an urgent problem to be solved. Summary of the Invention

[0003] The technical problem to be solved by the present invention is the problem of rockburst data generation. The purpose is to provide a rockburst data generation method, device and product based on a generative adversarial network, which solves the problems of data scarcity, data imbalance, high cost and high risk caused by traditional technologies relying on field experiments and real-time monitoring.

[0004] The present invention is achieved through the following technical solutions: In a first aspect, a rockburst data generation method based on a generative adversarial network is provided, comprising: Obtain actual rockburst geological condition parameters; adding random noise to the actual rockburst geological condition parameters to obtain noisy actual geological condition parameters, and performing data slicing processing on the noisy actual geological condition parameters to obtain noisy actual geological condition segment parameters of a plurality of continuous time segments; Obtaining a rockburst data generation model, wherein the rockburst data generation model is trained using sample data of several historical rockburst events as input, historical rockburst data of each historical rockburst event as label data, and rockburst generation data of each historical rockburst event as output, and the sample data of any historical rockburst event includes noisy historical geological condition segment parameters corresponding to each time segment within a time period of the occurrence of any historical rockburst event; A number of noisy actual geological condition segment parameters are input into the rockburst data generation model to obtain rockburst generation data corresponding to the actual rockburst geological condition parameters.

[0005] Based on the above-disclosed content, the present invention pre-constructs a rockburst data generation model for rockburst data, wherein the rockburst data generation model is trained by taking sample data corresponding to several historical rockburst events as input, historical rockburst data of each historical rockburst event as labels, and rockburst generation data of each historical rockburst event as output, and the sample data of any historical rockburst event includes noisy historical geological condition segment parameters corresponding to each time segment within the time period of any historical rockburst event; in this way, the present invention is equivalent to constructing a neural network model with rockburst geological condition parameters and random noise as input, and rockburst data corresponding to the rockburst geological condition parameters as output; based on this, in actual application, it is only necessary to obtain the actual rockburst geological condition parameters, and then perform noise processing on them to obtain the noisy actual geological condition parameters; then, the noisy actual geological condition parameters are data sliced ​​to obtain the noisy actual geological condition segment parameters of several continuous time segments; finally, the several noisy actual geological condition segment parameters are input into the aforementioned rockburst data generation model to obtain the rockburst generation data corresponding to the actual geological condition parameters.

[0006] Through the above design, the present invention addresses the problems of scarcity of rockburst data, sample imbalance, and high-cost acquisition, and establishes a rockburst data generation model with rockburst geological condition parameters and random noise as input and rockburst data corresponding to the rockburst geological condition parameters as output; thus, in actual use, different types of rockburst data can be obtained by simply inputting different rockburst geological condition parameters with random noise added into the aforementioned rockburst data generation model; thus, diversified rockburst data can be obtained by inputting rockburst geological condition parameters of different situations, thereby solving the problems of data scarcity and sample imbalance in traditional technologies. At the same time, the present invention also makes it possible to obtain rockburst data without experimental simulation and on-site acquisition, thereby reducing costs and risks; thus, the present invention can generate diversified and highly authentic rockburst data without relying on a large amount of real data, thereby being very suitable for large-scale application and promotion.

[0007] In one possible design, the rockburst data generation model is a trained bimodal GAN ​​model, wherein the bimodal GAN ​​model includes an LSTM network layer, a generator, a rockburst event classifier, a multimodal fusion device, and a discriminator connected in sequence, and any historical rockburst data includes multiple types of rockburst physical quantities; During training, the LSTM network layer is used to generate time-series compressed geological parameter feature data based on the parameters of each noisy historical geological condition segment in the input sample data; A generator for generating target rockburst data based on time-series compressed geological parameter characteristic data, wherein the target rockburst data is rockburst generation data of a target rockburst event, and the target rockburst event is a historical rockburst event corresponding to the input sample data; a rockburst event classifier, configured to perform feature mapping processing on target rockburst data to obtain a latent variable distribution of the target rockburst data, and calculate the KL divergence of the target rockburst data based on the latent variable distribution, so as to determine the rockburst event type of the target rockburst data based on the KL divergence; a multimodal fusion device for performing feature fusion processing on multiple types of rock burst physical quantities in the target rock burst data using a cross-attention mechanism to obtain fused target rock burst data, and performing feature fusion processing on multiple types of rock burst physical quantities in label data corresponding to the input sample data to obtain fused label data; The discriminator is used to perform discriminative processing on the fused target rockburst data based on the fused label data to obtain the probability that the target rockburst data is real data, so as to calculate the loss function value of the generator according to the rockburst event type of the target rockburst data and the probability that the target rockburst data is real data, and adjust the model parameters of the generator according to the loss function value until the training is completed, thereby obtaining the trained bimodal GAN ​​model.

[0008] In one possible design, the rockburst event classifier uses an encoder in a pre-trained variational autoencoder, wherein the variational autoencoder is trained with historical rockburst data of each historical rockburst event as input and reconstructed rockburst data corresponding to each historical rockburst event as output, so that after the training is completed, the pre-trained variational autoencoder is obtained. During the training process, the encoder in the variational autoencoder performs feature mapping on the input historical rockburst data to obtain the latent variable distribution of the input historical rockburst data, and the decoder in the variational autoencoder samples data from the latent variable distribution of the input historical rockburst data to generate reconstructed rockburst data of the input historical rockburst data.

[0009] In one possible design, the loss function of the rockburst event classifier is: (1) In formula (1), represents the loss function of the rockburst event classifier, represents the error between each input historical rockburst data and the corresponding reconstructed rockburst data, represents the hyperparameter, Represents the KL divergence of the input historical rockburst data.

[0010] In one possible design, the error between each input historical rockburst data and the corresponding reconstructed rockburst data is calculated using the following formula (2); (2) In formula (2), is the input i-th historical rockburst data, is the reconstructed rockburst data corresponding to the i-th historical rockburst data, and n represents the total number of historical rockburst data; Correspondingly, the KL divergence of the input historical rockburst data is calculated using the following formula (3): (3) In formula (3), represents the latent variable distribution of the input historical rockburst data, represents the standard normal distribution.

[0011] In a possible design, the rockburst event type of the target rockburst data is a common rockburst event or a rare rockburst event, and the common rockburst event is used to characterize the rockburst event with a rockburst energy less than or equal to an energy threshold, and the rare rockburst event is used to characterize the rockburst event with a rockburst energy greater than the energy threshold; Among them, the loss function of the generator is: (4) In formula (4), represents the loss function of the generator, represents the first weight coefficient, represents the loss item of common rock burst events, represents the second weight coefficient, represents the loss item for rare rock burst events; Accordingly, , where Represents the target rockburst data output by the discriminator is the probability of true data, Represents realistic expectations of common rockburst events; , where represents the true expectation of rare rockburst events, and when the rockburst event type of the target rockburst data is a common rockburst event, is 0, when the rockburst event type of the target rockburst data is a rare rockburst event, is 0.

[0012] In one possible design, after obtaining the trained bimodal GAN ​​model, the method further includes: Obtaining a test data set, and inputting test data in the test data set into the trained bimodal GAN ​​model to obtain rockburst generation data corresponding to each test data; Based on the historical rockburst data and rockburst generation data of each test data, the distribution similarity between the historical rockburst data of all test data and the corresponding rockburst generation data is calculated; Determining whether the distribution similarity meets a preset condition; If not, the model parameters of the generator of the bimodal GAN ​​model are adjusted to obtain an updated bimodal GAN ​​model, so as to retrain the updated bimodal GAN ​​model to obtain a trained bimodal GAN ​​model again.

[0013] In a second aspect, a rockburst data generation device based on a generative adversarial network is provided, comprising: An acquisition unit, used to obtain actual rockburst geological condition parameters; a noise adding unit, configured to add random noise to the actual rockburst geological condition parameter to obtain a noisy actual geological condition parameter, and perform data slicing processing on the noisy actual geological condition parameter to obtain noisy actual geological condition segment parameters of a plurality of continuous time segments; a data generation unit, configured to obtain a rockburst data generation model, wherein the rockburst data generation model is trained using sample data of a plurality of historical rockburst events as input, the historical rockburst data of each historical rockburst event as label data, and rockburst generation data of each historical rockburst event as output, and the sample data of any historical rockburst event includes noisy historical geological condition segment parameters corresponding to each time segment within a time period in which the historical rockburst event occurs; The data generation unit is used to input a number of noise-added actual geological condition segment parameters into the rockburst data generation model to obtain rockburst generation data corresponding to the actual rockburst geological condition parameters.

[0014] In the third aspect, another rockburst data generation device based on a generative adversarial network is provided. Taking the device as an electronic device as an example, the device includes a memory, a processor and a transceiver that are communicatively connected in sequence, wherein the memory is used to store computer programs, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the rockburst data generation method based on a generative adversarial network as described in the first aspect or any possible design of the first aspect.

[0015] In a fourth aspect, a storage medium is provided, on which instructions are stored. When the instructions are run on a computer, the rockburst data generation method based on a generative adversarial network as described in the first aspect or any possible design of the first aspect is executed.

[0016] In a fifth aspect, a computer program product comprising instructions is provided, which, when executed on a computer, causes the computer to execute the rockburst data generation method based on a generative adversarial network as described in the first aspect or any possible design of the first aspect.

[0017] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) In response to the problems of scarcity of rockburst data, imbalanced samples, and high cost of rockburst data collection, the present invention establishes a rockburst data generation model that takes rockburst geological condition parameters and random noise as input and outputs rockburst data corresponding to the rockburst geological condition parameters. Thus, in actual use, different types of rockburst data can be obtained by simply inputting different rockburst geological condition parameters with random noise added into the aforementioned rockburst data generation model. Thus, by inputting rockburst geological condition parameters of different situations, diversified rockburst data can be obtained, thereby solving the problems of data scarcity and imbalanced samples in traditional technologies. At the same time, the present invention also makes it possible to obtain rockburst data without experimental simulation and on-site collection, thereby reducing costs and risks. Thus, the present invention can generate diversified and highly authentic rockburst data without relying on a large amount of real data, making it very suitable for large-scale application and promotion.

[0018] (2) The rockburst data generation model provided by the present invention introduces an LSTM+GAN model based on conditional parameters, in which LSTM can ensure that the generated data maintains consistency and coherence in time evolution, while the input geological condition parameters can ensure that the generated data conforms to specific geological conditions or rockburst types; in this way, the authenticity and reliability of the generated data are guaranteed.

[0019] (3) The rockburst data generation model provided by the present invention introduces a VAE encoder to generate and classify rockburst data. At the same time, the loss function of the generator is divided into a regular event loss term and a rare event loss term with added weight coefficients. During training, the aforementioned loss function is calculated in combination with the classification of rockburst events to adjust the parameters. In this way, the generator can pay more attention to the learning of rare events, thereby enhancing the model's ability to learn rare rockburst events. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the examples. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. A person of ordinary skill in the art can also derive other relevant drawings based on these drawings without inventive effort. In the drawings: Figure 1 A schematic flow chart of the steps of a rockburst data generation method based on a generative adversarial network provided in an embodiment of the present invention; Figure 2 A schematic diagram of the structure of a rockburst generation model provided in an embodiment of the present invention; Figure 3 A schematic diagram showing a comparison between generated data output by a rockburst generation model provided in an embodiment of the present invention and real data; Figure 4 A schematic structural diagram of a rockburst data generation device based on a generative adversarial network provided in an embodiment of the present invention; Figure 5 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0021] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the following examples and accompanying drawings. The exemplary embodiments of the present invention and their descriptions are intended only to explain the present invention and are not intended to limit the present invention. It should be understood that although the terms "first," "second," and so on may be used herein to describe various elements, these elements should not be limited by these terms. These terms are merely used to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element without departing from the scope of the exemplary embodiments of the present invention.

[0022] Example: See also Figure 1As shown, the rockburst data generation method based on the generative adversarial network provided in this embodiment establishes a rockburst data generation model with rockburst geological condition parameters and random noise as inputs and rockburst data corresponding to the rockburst geological condition parameters as outputs; thus, in actual use, different rockburst geological condition parameters with random noise added only need to be input into the aforementioned rockburst data generation model to obtain different types of rockburst data; thus, by inputting rockburst geological condition parameters of different situations, diversified rockburst data can be obtained, thereby solving the problems of data scarcity and sample imbalance in traditional technologies. At the same time, it also makes it possible to obtain rockburst data without experimental simulation and on-site collection, thereby reducing costs and risks; thus, this method provides a new rockburst data acquisition method, which is very suitable for large-scale application and promotion; wherein, for example, this method can be, but is not limited to, running on the rockburst data generation end side. Optionally, the rockburst data generation end can be, but is not limited to, a personal computer or server. It can be understood that the aforementioned execution subject does not constitute a limitation on the embodiments of the present application. Accordingly, the operation steps of this method can be, but are not limited to, as shown in the following steps S1 to S4.

[0023] S1. Obtaining actual rockburst geological condition parameters. In a specific application, for example, the actual rockburst geological condition parameters are rockburst geological condition parameters input during actual use, which are used to generate rockburst data. For example, they may include, but are not limited to, stress mutation intensity within a preset time period (indicating the degree of stress concentration within the rock; a larger value indicates a higher degree of rockburst damage), energy release rate (used to simulate the energy release rate during rockburst occurrence; a higher energy release rate generally indicates a more severe rockburst), fracture scale (used to control the range of rock fracture and can be used to simulate local small rockbursts or large-scale rockbursts), vibration frequency (which affects the vibration signal characteristics during rockburst; different vibration frequencies reflect different types of rockburst events), and geological environment parameters (including burial depth, surrounding rock type, stratum structure, etc., which affect the probability and morphology of rockburst occurrence). In this way, after obtaining the actual rockburst geological condition parameters used to control the occurrence of rockburst events, they can be noised so that corresponding rockburst data can be subsequently generated based on the noised actual rockburst geological condition parameters. The noise addition process of the actual rockburst geological condition parameters is shown in the following step S2.

[0024] S2. Add random noise to the actual rockburst geological condition parameters to obtain noisy actual geological condition parameters, and perform data slicing on the noisy actual geological condition parameters to obtain noisy actual geological condition segment parameters of several continuous time segments; in specific applications, the above-mentioned actual rockburst geological condition parameters are used as conditional inputs of the rockburst data generation model described below, and then the above-mentioned random noise is mixed to obtain the noisy actual geological condition parameters; at the same time, for example, but not limited to, the noisy actual geological condition parameters can be divided into multiple windows, each window containing data of a fixed length, such as the length of each window is 1 minute, that is, the noisy actual geological condition parameters within a preset time period are data sliced ​​at intervals of 1 minute, thereby obtaining several noisy actual geological condition segment parameters; in addition, when performing noise addition, noise data such as Gaussian noise, additive noise or salt and pepper noise can be added. Of course, data noise addition is a commonly used technology for data processing, and its principle will not be repeated here.

[0025] After obtaining the noisy actual geological condition segment parameters, they can be input into a pre-built rockburst data generation model to generate corresponding rockburst generation data; wherein, the acquisition process of the rockburst data generation model is shown in the following step S3.

[0026] S3. Obtain a rockburst data generation model, wherein the rockburst data generation model is trained using sample data of several historical rockburst events as input, historical rockburst data of each historical rockburst event as label data, and rockburst generation data of each historical rockburst event as output, and the sample data of any historical rockburst event includes noisy historical geological condition segment parameters corresponding to each time segment within the time period of any historical rockburst event.

[0027] In specific applications, this embodiment provides a dual-modal GAN ​​model for model training to generate a rockburst data generation model. The training process of the dual-modal GAN ​​model is as follows: S31. Obtain historical rockburst data corresponding to several historical rockburst events and historical rockburst geological condition parameters corresponding to each historical rockburst data, wherein any historical rockburst data includes multiple types of rockburst physical quantities. In this embodiment, the parameter types of the historical rockburst geological condition parameters are the same as the aforementioned actual rockburst geological condition parameters and are not further described here. At the same time, the rockburst physical quantities included in any historical rockburst data may include, but are not limited to, microseismic signals, acoustic emission signals, stress-strain curve signals, and infrared thermal imaging images. In this way, the historical rockburst data corresponding to each historical rockburst event are multimodal data.

[0028] After obtaining the aforementioned raw data, noise addition and data slicing processing may be performed, and the process is shown in the following step S32.

[0029] S32. Add random noise to the historical rockburst geological condition parameters corresponding to each historical rockburst data to obtain a number of noisy historical rockburst geological condition parameters, and perform data slicing processing on each noisy historical rockburst geological condition parameter to obtain noisy historical geological condition segment parameters corresponding to each time segment within the time period of each historical rockburst event. In this embodiment, the noise addition and slicing process of the historical rockburst geological condition parameters can be referred to the aforementioned step S2 and will not be repeated here.

[0030] At the same time, for example, before adding noise, the aforementioned historical rockburst geological condition parameters can also be normalized so that the data can be compared on the same scale; among them, for example, but not limited to, the Z-score normalization method can be used to normalize the aforementioned data; of course, Z-score is a commonly used data normalization method, and its principle will not be repeated here.

[0031] After the noise addition and slicing of the historical rockburst geological condition parameters are completed, a training set can be generated based on the noise addition and slicing, and the process is shown in the following step S33.

[0032] S33. Utilize the noisy historical geological condition segment parameters corresponding to each time segment within the time period of each historical rockburst event to construct sample data for each historical rockburst event. Use the historical rockburst data corresponding to each historical rockburst event as label data, and utilize the sample data and label data corresponding to each historical rockburst event to construct a training set.

[0033] After generating the training set based on the aforementioned step S33, the training set can be used to train the bimodal GAN ​​model, and the process is shown in the following step S34.

[0034] S34. A bimodal GAN ​​model is trained using the sample data corresponding to each historical rockburst event in the training set as input and the rockburst generation data corresponding to each historical rockburst event as output, so as to obtain a rockburst data generation model after training. In this embodiment, the training set is divided into several training subsets, each corresponding to a training batch. The bimodal GAN ​​model is then trained according to the training batches. After each training batch is completed, the model loss function is calculated to adjust the model parameters. Finally, after all training batches are trained, the rockburst data generation model is obtained.

[0035] Furthermore, this embodiment combines the structure of the aforementioned dual-modal GAN ​​model to specifically explain the aforementioned training process: See also Figure 2 As shown, the bimodal GAN ​​model described as an example may include, but is not limited to, sequentially connected LSTM network layers, generators, rockburst event classifiers, multimodal fusers, and discriminators.

[0036] Among them, when training a training batch, the LSTM network layer is used to generate time-series compressed geological parameter feature data based on the parameters of each noisy historical geological condition segment in the input sample data; Specifically, the occurrence of rock burst events has certain temporal evolution characteristics, and the fully connected or convolutional structure used by traditional GAN ​​is difficult to capture this dynamic evolution law. Therefore, this embodiment adds an LSTM network layer to capture the temporal characteristics of rock burst data; Among them, see Figure 2 As shown, the input sample data is essentially the noise-added historical geological condition fragment parameters at different time steps (i.e. Figure 2 v1, v2, ..., vT in ), and the LSTM network layer contains multiple LSTM units connected in sequence. Each LSTM unit corresponds to a noisy historical geological condition segment parameter in chronological order. In this way, the first LSTM unit corresponds to the noisy historical geological condition segment parameter in the first time segment, the second LSTM unit corresponds to the noisy historical geological condition segment parameter in the second time segment, and so on. Each LSTM unit corresponds to the noisy historical geological condition segment parameter of a time segment. Based on this, each LSTM unit will receive the data output by the previous LSTM unit and combine it with the noisy historical geological condition segment parameter in its own corresponding time segment to generate its own output data. Therefore, the data output by the last LSTM unit contains the time series compression feature hT of the entire input sample data, that is, the time series data is modeled, that is, the aforementioned time series compression geological parameter feature data; of course, the LSTM network is a commonly used model for time series modeling, and its principle will not be repeated here.

[0037] In this way, through the aforementioned LSTM network layer, a coherent time series signal that conforms to physical laws can be generated, thereby improving the model's ability to generate time series data. After obtaining the time series compressed geological parameter characteristic data, it can be input into the generator to generate the corresponding rockburst generation data. The process is as follows: A generator is used to generate target rockburst data based on time-series compressed geological parameter characteristic data, wherein the target rockburst data is rockburst generation data of a target rockburst event, and the target rockburst event is a historical rockburst event corresponding to the input sample data.

[0038] After using the generator to generate rock burst generation data of historical rock burst events corresponding to the input sample data, this embodiment further provides a classifier for rock burst data in order to enhance rare events, namely: a rock burst event classifier, which is used to perform feature mapping processing on the target rock burst data to obtain the latent variable distribution of the target rock burst data, and calculate the KL divergence of the target rock burst data based on the latent variable distribution, so as to determine the rock burst event type of the target rock burst data based on the KL divergence; in this embodiment, the rock burst event type of the target rock burst data is taken as an example of a common rock burst event or a rare rock burst event, and the common rock burst event is used to characterize the rock burst event whose rock burst energy is less than or equal to the energy threshold, while the rare rock burst event is used to characterize the rock burst event whose rock burst energy is greater than the energy threshold; in this way, by classifying the rock burst generation data, it is convenient to subsequently combine its classification results to calculate the loss function of the generator; of course, the calculation of the loss function is explained in detail below.

[0039] Among them, this embodiment introduces a pre-trained VAE encoder to perform feature encoding on the generated samples (i.e., the aforementioned target rockburst data) and perform "regular / rare" division in the latent space; specifically, the VAE encoder is trained on real rockburst data and can capture the mainstream feature distribution of rockburst events, thereby identifying whether the generated samples fall in regular areas (high-density distribution) or rare areas (low-density or edge distribution), and then deriving the rockburst event type of the generated samples.

[0040] Furthermore, for example, a rock burst event classifier uses an encoder in a pre-trained variational autoencoder, and the training process of the pre-trained variational autoencoder is: using the historical rock burst data of each historical rock burst event as input and the reconstructed rock burst data corresponding to each historical rock burst event as output to train the variational autoencoder, so as to obtain the pre-trained variational autoencoder after the training is completed, wherein, during the training process, the encoder in the variational autoencoder performs feature mapping on the input historical rock burst data (i.e., maps it to a low-dimensional latent space) to obtain the latent variable distribution of the input historical rock burst data, and the decoder in the variational autoencoder samples data from the latent variable distribution of the input historical rock burst data (i.e., samples data in a low-dimensional latent space) to generate reconstructed rock burst data corresponding to the input historical rock burst data.

[0041] In this embodiment, a variational autoencoder (VAE) is introduced to learn the latent features of rockburst data. It is trained using existing rockburst data. The key to VAE lies in variational inference, that is, optimizing the model by maximizing the lower bound of evidence so that the latent variables obey the standard normal distribution p(z). Therefore, this embodiment discloses the loss function of the aforementioned rockburst event classifier, which can be, but is not limited to, as shown in the following formula (1).

[0042] (1) In formula (1), represents the loss function of the rockburst event classifier, represents the error between each input historical rockburst data and the corresponding reconstructed rockburst data, represents a hyperparameter used to balance the reconstruction loss and KL divergence, Represents the KL divergence of the input historical rockburst data, which is used to measure the difference between the variational distribution and the standard normal distribution.

[0043] Furthermore, for example, the error between each input historical rockburst data and the corresponding reconstructed rockburst data can be calculated using, but not limited to, the following formula (2).

[0044] (2) In formula (2), is the input i-th historical rockburst data, is the reconstructed rockburst data corresponding to the i-th historical rockburst data, and n represents the total number of historical rockburst data.

[0045] Similarly, the KL divergence of the input historical rockburst data can be calculated using, but not limited to, the following formula (3).

[0046] (3) In formula (3), represents the latent variable distribution of the input historical rockburst data, represents the standard normal distribution, and z represents the possible values ​​of the latent variable distribution.

[0047] In this way, by using the loss function of the aforementioned rockburst event classifier to train the variational autoencoder, it is possible to capture the mainstream feature distribution of rockburst events, that is, to obtain the latent variable distribution of the input rockburst data, thereby further decoupling and distinguishing the common features from the rare features in the data; among them, the relationship between the latent variable distribution and common rockburst events and rare rockburst events is: for common rockburst events, in the latent space, conventional samples are concentratedly distributed, which conforms to the backbone structure of the model, that is, the latent variable distribution falls in the "high-density area" of the latent space; while for rare rockburst events, these data exhibit unconventional behavior, are in the marginal area or sparse area, and the distribution of the latent variable deviates from the standard normal distribution; at this time, the KL divergence can be used to measure the degree of deviation.

[0048] In this way, after completing the training of the variational autoencoder, its internal encoder can be used as a rockburst event classifier to classify the rockburst generated data output by the generator into rockburst events, that is, to obtain the latent variable distribution of the target rockburst data output by the generator. Then, based on the above formula (3), the KL divergence of the target rockburst data is calculated. Then, according to the KL divergence, it is determined whether the rockburst event type of the target rockburst data is a common rockburst event or a rare rockburst event; among them, if the KL divergence is less than the divergence threshold, it can be determined that the rockburst event type of the target rockburst data is a common rockburst event, otherwise, it is determined to be a rare rockburst event.

[0049] After completing the classification of the rock burst event type of the target rock burst data output by the generator, multimodal fusion can be performed, that is, a multimodal fusion device is used to adopt a cross-attention mechanism to perform feature fusion processing on multiple types of rock burst physical quantities in the target rock burst data to obtain fused target rock burst data, and to perform feature fusion processing on multiple types of rock burst physical quantities in the label data corresponding to the input sample data to obtain fused label data.

[0050] In specific implementation, rockburst events are usually accompanied by a variety of physical phenomena (such as microseismic, acoustic emission, temperature changes, etc.), that is, the synthetic data output by the generator (i.e., target rockburst data) not only contains time series signals (microseismic waveforms, etc.), but also includes spatial information (infrared thermal imaging, etc.). These data jointly reflect different aspects of the same rockburst event and are highly correlated. Therefore, in order to more effectively judge the authenticity, physical consistency and credibility of the synthetic data, the present invention extracts and fuses features from real rockburst samples and the multimodal rockburst data synthesized by the generator, and uniformly inputs them into the discriminator for classification and judgment.

[0051] Among them, the example multimodal fusion device is a cross-attention network, which uses the cross-attention mechanism to perform weighted fusion of the temporal signal and spatial information in the target rockburst data. Similarly, the temporal signal and spatial information in the label data corresponding to the input sample data are also weightedly fused to obtain the aforementioned fused target rockburst data and fused label data; of course, the cross-attention mechanism is a commonly used technology for feature fusion, and its principle will not be repeated here.

[0052] Therefore, the cross-attention mechanism can ensure the correlation between time series data and spatial data when generating data, thereby improving the overall consistency of rockburst generation data, and then achieving the goal that when generating the microseismic waveform at a certain moment, the model can refer to the infrared thermal imaging information at the corresponding time point to generate data that conforms to physical laws.

[0053] After completing multimodal fusion, it can be input into the discriminator for discrimination, that is: The discriminator is used to perform discriminative processing on the fused target rockburst data based on the fused label data to obtain the probability that the target rockburst data is real data, so as to calculate the loss function value of the generator according to the rockburst event type of the target rockburst data and the probability that the target rockburst data is real data, and adjust the model parameters of the generator according to the loss function value until the training is completed, thereby obtaining the trained bimodal GAN ​​model.

[0054] In this embodiment, it has been explained above that the loss function is calculated once for one training batch. Therefore, the loss function of the generator is explained below using one training batch as an example, as shown in the following formula (4).

[0055] (4) In formula (4), represents the loss function of the generator, represents the first weight coefficient, represents the loss item of common rock burst events, represents the second weight coefficient, Represents the loss term for rare rockburst events.

[0056] in, , where Represents the target rockburst data output by the discriminator is the probability of true data, represents a realistic expectation of common rockburst events, and It represents the mean probability of the discriminator outputting the target rockburst data belonging to common rockburst events among the target rockburst data generated by the generator in a training batch. That is to say, assuming that a training batch corresponds to 10 input sample data, among which, among the 10 target rockburst data generated by the generator, 5 target rockburst data have the rockburst event type of common rockburst events, then, It is the mean of the probabilities that the five target rockburst data output by the discriminator are real data.

[0057] Similarly, , where represents the true expectation of rare rockburst events, and when the rockburst event type of the target rockburst data is a common rockburst event, is 0, when the rockburst event type of the target rockburst data is a rare rockburst event, is 0.

[0058] Furthermore, in order to strengthen the generation of rare events, the second weight coefficient is set greater than the first weight coefficient, so that the generator is more inclined to generate rare events during training and ensure that the discriminator can effectively identify the authenticity of these events; at the same time, during the training process, as the generator is optimized, the weight coefficient is dynamically adjusted, that is, in the early stage, a higher second weight coefficient is set to speed up the learning of rare events. As the model converges, the focus on rare events is gradually reduced (that is, the second weight coefficient is reduced) to avoid overfitting.

[0059] In this way, by dividing the generator's loss function into a regular event loss term and a rare event loss term with added weight coefficients, and combining the classification of rock burst events during training to calculate the aforementioned loss function for parameter adjustment; in this way, the generator can pay more attention to learning rare events, thereby enhancing the model's ability to learn rare rock burst events.

[0060] The above explanation clearly illustrates the training process of sample data in the bimodal GAN ​​model. Based on this, through training of multiple training batches, the trained bimodal GAN ​​model can be obtained, that is, the rockburst data generation model can be obtained.

[0061] Furthermore, in order to ensure the quality, authenticity and physical rationality of the data, this embodiment is also provided with a model verification process, namely: first, a test data set is obtained, and the test data in the test data set is input into the trained bimodal GAN ​​model to obtain the rockburst generation data corresponding to each test data; then, based on the historical rockburst data and rockburst generation data of each test data, the distribution similarity between the historical rockburst data of all test data and all corresponding rockburst generation data is calculated; then, it is judged whether the distribution similarity meets the preset conditions; if not, the model parameters of the generator of the bimodal GAN ​​model are adjusted to obtain an updated bimodal GAN ​​model, so as to retrain the updated bimodal GAN ​​model to obtain the trained bimodal GAN ​​model again.

[0062] In this embodiment, the test data in the test data set is also the sample data corresponding to the historical rockburst event. During training, the training set can be divided into 8:2 ratios, so that the remaining 20% ​​of the training data is used as the test data. At the same time, the distribution similarity between the historical rockburst data of all test data and all rockburst generation data can be measured using KL divergence. That is, when the distribution similarity is greater than the similarity threshold, it is determined that it meets the preset conditions. Otherwise, it is necessary to adjust the parameters and re-train. Of course, KL divergence is a commonly used technology for measuring data distribution similarity, and its principle will not be repeated here.

[0063] In addition, in this embodiment, in order to further improve the quality of generated data, an adversarial retraining optimization method can be used to add generated samples that are misjudged by experts to the training set, further improving the performance of the generator, allowing the generator to continuously evolve and improve the realism of the data.

[0064] Thus, through the aforementioned step S3, the training of the bimodal GAN ​​model can be completed to obtain a rockburst data generation model; then, the rockburst data generation model can be used to generate rockburst data, and the process is shown in the following step S4.

[0065] S4. Inputting a number of noisy actual geological condition segment parameters into the rockburst data generation model to obtain rockburst generation data corresponding to the actual rockburst geological condition parameters; in this embodiment, after obtaining different types of rockburst generation data, the generated rockburst generation data can be used to train a rockburst prediction model, thereby obtaining a rockburst prediction model, and then using the rockburst prediction model to predict the possibility of rockburst occurrence, so as to provide a rockburst risk warning.

[0066] Also, see Figure 3 As shown, this embodiment provides a schematic diagram of the comparison between the generated data output by the rock burst data generation model and the real rock burst data. Figure 3 It can be seen that the distribution trajectory of generated data and real data has a high degree of overlap, indicating that the generated data has high authenticity.

[0067] Therefore, through the rockburst data generation method based on the generative adversarial network described in detail in the aforementioned steps S1 to S4, the present invention establishes a rockburst data generation model that takes rockburst geological condition parameters and random noise as input and outputs rockburst data corresponding to the rockburst geological condition parameters; thus, in actual use, different types of rockburst data can be obtained by simply inputting different rockburst geological condition parameters with random noise added into the aforementioned rockburst data generation model; thus, by inputting rockburst geological condition parameters of different situations, diversified rockburst data can be obtained, thereby solving the problems of data scarcity and sample imbalance in traditional technologies, and at the same time, it also makes it possible to obtain rockburst data without experimental simulation and on-site collection, thereby reducing costs and risks; thus, the present invention provides a new rockburst data acquisition method, which is very suitable for large-scale application and promotion.

[0068] like Figure 4 As shown, the second aspect of this embodiment provides a hardware device for implementing the rockburst data generation method based on a generative adversarial network described in the first aspect of the embodiment, including: The acquisition unit is used to obtain actual rockburst geological condition parameters.

[0069] The noise adding unit is used to add random noise to the actual rockburst geological condition parameters to obtain noisy actual geological condition parameters, and perform data slicing processing on the noisy actual geological condition parameters to obtain noisy actual geological condition segment parameters of several continuous time segments.

[0070] A data generation unit is used to obtain a rockburst data generation model, wherein the rockburst data generation model is trained using sample data of several historical rockburst events as input, historical rockburst data of each historical rockburst event as label data, and rockburst generation data of each historical rockburst event as output, and the sample data of any historical rockburst event includes noisy historical geological condition segment parameters corresponding to each time segment within the time period of any historical rockburst event.

[0071] The data generation unit is used to input a number of noise-added actual geological condition segment parameters into the rockburst data generation model to obtain rockburst generation data corresponding to the actual rockburst geological condition parameters.

[0072] The working process, working details and technical effects of the device provided in this embodiment can be found in the first aspect of the embodiment and will not be described in detail here.

[0073] like Figure 5 As shown, the third aspect of this embodiment provides another rockburst data generation device based on a generative adversarial network. Taking the device as an electronic device as an example, it includes: a memory, a processor and a transceiver that are communicatively connected in sequence, wherein the memory is used to store computer programs, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the rockburst data generation method based on a generative adversarial network as described in the first aspect of the embodiment.

[0074] For example, the memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, first-in first-out memory (FIFO), and / or first-in last-out memory (FILO). Specifically, the processor may include one or more processing cores, such as a quad-core processor or an octal-core processor. The processor may be implemented in at least one of the following hardware forms: a DSP (Digital Signal Processing), an FPGA (Field-Programmable Gate Array), or a PLA (Programmable Logic Array). Furthermore, the processor may include a main processor and a coprocessor. The main processor is a processor for processing data in an awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in a standby state.

[0075] In some embodiments, the processor may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content required to be displayed on the display screen. For example, the processor may be, but is not limited to, a microprocessor of the STM32F105 series, a reduced instruction set computer (RISC) microprocessor, an X86 architecture processor, or a processor with an integrated embedded neural network processing unit (NPU). The transceiver may be, but is not limited to, a Wireless Fidelity (WIFI) wireless transceiver, a Bluetooth wireless transceiver, a General Packet Radio Service (GPRS) wireless transceiver, a ZigBee protocol (a low-power local area network protocol based on the IEEE802.15.4 standard, ZigBee) wireless transceiver, a 3G transceiver, a 4G transceiver, and / or a 5G transceiver. In addition, the device may also include, but is not limited to, a power module, a display screen, and other necessary components.

[0076] The working process, working details and technical effects of the electronic device provided in this embodiment can be found in the first aspect of the embodiment and will not be described in detail here.

[0077] The fourth aspect of this embodiment provides a storage medium that stores instructions for the rockburst data generation method based on a generative adversarial network as described in the first aspect of the embodiment, that is, the storage medium stores instructions, and when the instructions are run on a computer, the rockburst data generation method based on a generative adversarial network as described in the first aspect of the embodiment is executed.

[0078] The storage medium refers to a carrier for storing data, which may include but is not limited to a floppy disk, an optical disk, a hard disk, a flash memory, a USB flash drive and / or a memory stick, and the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0079] The working process, working details and technical effects of the storage medium provided in this embodiment can be found in the first aspect of the embodiment and will not be described in detail here.

[0080] A fifth aspect of this embodiment provides a computer program product comprising instructions, which, when executed on a computer, causes the computer to execute the rockburst data generation method based on a generative adversarial network as described in the first aspect of the embodiment, wherein the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A rockburst data generation method based on generative adversarial networks, characterized in that: include: Obtain actual rockburst geological condition parameters; adding random noise to the actual rockburst geological condition parameters to obtain noisy actual geological condition parameters, and performing data slicing processing on the noisy actual geological condition parameters to obtain noisy actual geological condition segment parameters of a plurality of continuous time segments; Obtaining a rockburst data generation model, wherein the rockburst data generation model is trained using sample data of several historical rockburst events as input, historical rockburst data of each historical rockburst event as label data, and rockburst generation data of each historical rockburst event as output, and the sample data of any historical rockburst event includes noisy historical geological condition segment parameters corresponding to each time segment within a time period of the occurrence of any historical rockburst event; A number of noisy actual geological condition segment parameters are input into the rockburst data generation model to obtain rockburst generation data corresponding to the actual rockburst geological condition parameters.

2. The method according to claim 1, characterized in that The rockburst data generation model is a trained bimodal GAN ​​model, wherein the bimodal GAN ​​model includes an LSTM network layer, a generator, a rockburst event classifier, a multimodal fusion device, and a discriminator connected in sequence, and any historical rockburst data includes multiple types of rockburst physical quantities; During training, the LSTM network layer is used to generate time-series compressed geological parameter feature data based on the parameters of each noisy historical geological condition segment in the input sample data; A generator for generating target rockburst data based on time-series compressed geological parameter characteristic data, wherein the target rockburst data is rockburst generation data of a target rockburst event, and the target rockburst event is a historical rockburst event corresponding to the input sample data; a rockburst event classifier, configured to perform feature mapping processing on target rockburst data to obtain a latent variable distribution of the target rockburst data, and calculate the KL divergence of the target rockburst data based on the latent variable distribution, so as to determine the rockburst event type of the target rockburst data based on the KL divergence; a multimodal fusion device for performing feature fusion processing on multiple types of rock burst physical quantities in the target rock burst data using a cross-attention mechanism to obtain fused target rock burst data, and performing feature fusion processing on multiple types of rock burst physical quantities in label data corresponding to the input sample data to obtain fused label data; The discriminator is used to perform discriminative processing on the fused target rockburst data based on the fused label data to obtain the probability that the target rockburst data is real data, so as to calculate the loss function value of the generator according to the rockburst event type of the target rockburst data and the probability that the target rockburst data is real data, and adjust the model parameters of the generator according to the loss function value until the training is completed, thereby obtaining the trained bimodal GAN ​​model.

3. The method according to claim 2, characterized in that The rockburst event classifier adopts the encoder in the pre-trained variational autoencoder, wherein the historical rockburst data of each historical rockburst event is used as input and the reconstructed rockburst data corresponding to each historical rockburst event is used as output to train the variational autoencoder, so that after the training is completed, the pre-trained variational autoencoder is obtained. During the training process, the encoder in the variational autoencoder performs feature mapping on the input historical rockburst data to obtain the latent variable distribution of the input historical rockburst data, and the decoder in the variational autoencoder samples data from the latent variable distribution of the input historical rockburst data to generate reconstructed rockburst data of the input historical rockburst data.

4. The method according to claim 3, characterized in that The loss function of the rockburst event classifier is: （1） In formula (1), represents the loss function of the rockburst event classifier, represents the error between each input historical rockburst data and the corresponding reconstructed rockburst data, represents the hyperparameter, Represents the KL divergence of the input historical rockburst data.

5. The method according to claim 4, characterized in that The error between each input historical rockburst data and the corresponding reconstructed rockburst data is calculated using the following formula (2); （2） In formula (2), is the input i-th historical rockburst data, is the reconstructed rockburst data corresponding to the i-th historical rockburst data, and n represents the total number of historical rockburst data; Correspondingly, the KL divergence of the input historical rockburst data is calculated using the following formula (3): （3） In formula (3), represents the latent variable distribution of the input historical rockburst data, represents the standard normal distribution.

6. The method according to claim 2, characterized in that The rockburst event type of the target rockburst data is a common rockburst event or a rare rockburst event, wherein the common rockburst event is used to characterize a rockburst event with a rockburst energy less than or equal to an energy threshold, and the rare rockburst event is used to characterize a rockburst event with a rockburst energy greater than an energy threshold; Among them, the loss function of the generator is: （4） In formula (4), represents the loss function of the generator, represents the first weight coefficient, represents the loss item of common rock burst events, represents the second weight coefficient, represents the loss item for rare rock burst events; Accordingly, , where Represents the target rockburst data output by the discriminator is the probability of true data, Represents realistic expectations of common rockburst events; , where represents the true expectation of rare rockburst events, and when the rockburst event type of the target rockburst data is a common rockburst event, is 0, when the rockburst event type of the target rockburst data is a rare rockburst event, is 0.

7. The method according to claim 2, characterized in that After obtaining the trained bimodal GAN ​​model, the method further includes: Obtaining a test data set, and inputting test data in the test data set into the trained bimodal GAN ​​model to obtain rockburst generation data corresponding to each test data; Based on the historical rockburst data and rockburst generation data of each test data, the distribution similarity between the historical rockburst data of all test data and the corresponding rockburst generation data is calculated; Determining whether the distribution similarity meets a preset condition; If not, the model parameters of the generator of the bimodal GAN ​​model are adjusted to obtain an updated bimodal GAN ​​model, so as to retrain the updated bimodal GAN ​​model to obtain a trained bimodal GAN ​​model again.

8. A rockburst data generation device based on generative adversarial network, characterized in that: include: An acquisition unit, used to obtain actual rockburst geological condition parameters; a noise adding unit, configured to add random noise to the actual rockburst geological condition parameter to obtain a noisy actual geological condition parameter, and perform data slicing processing on the noisy actual geological condition parameter to obtain noisy actual geological condition segment parameters of a plurality of continuous time segments; a data generation unit, configured to obtain a rockburst data generation model, wherein the rockburst data generation model is trained using sample data of a plurality of historical rockburst events as input, the historical rockburst data of each historical rockburst event as label data, and rockburst generation data of each historical rockburst event as output, and the sample data of any historical rockburst event includes noisy historical geological condition segment parameters corresponding to each time segment within a time period in which the historical rockburst event occurs; The data generation unit is used to input a number of noise-added actual geological condition segment parameters into the rockburst data generation model to obtain rockburst generation data corresponding to the actual rockburst geological condition parameters.

9. A rockburst data generation device based on generative adversarial network, characterized in that: include: A memory, a processor, and a transceiver that are sequentially communicatively connected, wherein the memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program, and execute the rockburst data generation method based on a generative adversarial network according to any one of claims 1 to 7.

10. A computer program product comprising instructions, characterized in that When the instructions are executed on a computer, the computer is caused to execute the rockburst data generation method based on a generative adversarial network according to any one of claims 1 to 7.

Images