A method for predicting laboratory earthquakes based on deep convolutional neural networks

Abstract

The present invention discloses a method for predicting laboratory earthquakes based on a deep convolutional neural network. The method uses the extraction capability of a deep convolutional neural network to adaptively learn multi-scale, hierarchical discriminative feature representations from raw acoustic emission signals without the need for manually designed features. Special network structure designs such as causal convolution and dilated convolution are used to achieve accurate signal location and classification while maintaining a large receptive field. The present invention has a natural advantage in parallel computing, can quickly process large-scale time series data, and as the training data increases, the model performance is continuously improved. By directly modeling the acoustic emission time series signal and taking into account its long-distance time dependency and autoregressive characteristics, the dynamic evolution law in the earthquake nucleation process can be better captured.

Description

Technical Field

[0001] The present invention relates to the technical field of earthquake research, and in particular to a method for predicting laboratory earthquakes based on a deep convolutional neural network. Background Art

[0002] In the current research trend, the use of machine learning technology to predict laboratory earthquakes has become an area of ​​great interest. Previous studies have mainly focused on using traditional machine learning methods, such as decision trees, to analyze the mechanisms and behaviors of laboratory earthquakes. These methods usually rely on artificially designed feature sets for specific experimental conditions, such as double direct shear test devices. However, these methods have some significant limitations, including poor dependency and transferability of feature sets, low data utilization capabilities, and failure to fully explore the time-dependent characteristics and physical mechanism information of acoustic emission signals.

[0003] Specifically, the existing technology uses machine learning methods in shear tests on simulated faults to predict the occurrence time of laboratory earthquakes by analyzing continuous acoustic emission signals. However, these methods mainly rely on manually designed feature sets, and their construction process is highly dependent on specific experimental equipment, experimental conditions, and sample settings. This makes it difficult to extend these methods to new experimental conditions and actual engineering applications. In addition, traditional machine learning methods such as decision trees have limited data utilization capabilities, especially for high-frequency sampling, massive, and high-dimensional time series data.

[0004] In addition, the time-dependent characteristics and physical mechanisms of acoustic emission signals are important information for predicting laboratory earthquakes, but traditional methods have failed to fully tap into this information. Traditional methods usually divide acoustic emission signals into independent time windows, ignoring the correlation and dependency between adjacent time windows, making it impossible to accurately model the dynamic evolution law and autoregressive characteristics of the signal, and unable to fully utilize the physical mechanism information contained in the signal. Summary of the invention

[0005] In view of the defects of the prior art, the present invention provides a method for predicting laboratory earthquakes based on a deep convolutional neural network.

[0006] In order to achieve the above invention object, the technical solution adopted by the present invention is as follows:

[0007] A method for predicting laboratory earthquakes based on a deep convolutional neural network comprises the following steps:

[0008] 1) Data acquisition: Collect laboratory earthquake data, including shear stress and acoustic emission data.

[0009] 2) Data preprocessing: Ensure data quality and perform necessary preprocessing, including denoising, detrending, and downsampling.

[0010] 3) Data alignment and time series construction: Align the shear stress and acoustic emission data in time to ensure their temporal consistency. Construct a time series, use shear stress as the target value and acoustic emission data as the feature value to form a training sample.

[0011] 4) Build a deep convolutional neural network model: Design the architecture of the neural network, including the input layer, convolution layer, pooling layer, fully connected layer, as well as activation functions and regularization methods.

[0012] 5) Divide the dataset: Divide the dataset into training set, validation set and test set to ensure the independence and time sequence of the data during the training process.

[0013] 6) Model training: Use the training set to train the deep convolutional neural network model. Use appropriate loss functions, optimization algorithms, and learning rate scheduling strategies to optimize model parameters so that it achieves a better fitting effect on the training set.

[0014] Furthermore, the deep convolutional neural network model also includes:

[0015] Batch Normalization Layer: This layer is used to speed up training by normalizing each small batch of input data to have a mean of 0 and a variance of 1.

[0016] Decoder: It adopts a multi-layer architecture, each layer is a temporal convolutional network module, containing two identical dilated causal convolutions, and a random inactivation layer is added to the decoder module to prevent overfitting.

[0017] Regression layer: It includes a one-dimensional convolutional layer with an output feature of 64 and a kernel size of 2 to further refine the extracted features. Two linear layers are then used to convert the feature representation to the final target value and activated using the rectified linear unit function. The random dropout layer is omitted in the regression layer.

[0018] Dilation rate setting: The multi-layer temporal convolutional network module of the decoder sets different dilation rates. By adjusting the dilation factor, the model's perception of long sequences is controlled to keep the model parameters controllable and flexibly respond to the needs of different tasks. The dilation factor for the i-th layer is set to 2^(i-1), where i=1,2,…,N, indicating the number of layers.

[0019] Furthermore, the deep convolutional neural network model also adds weight normalization;

[0020] Furthermore, the activation function uses the rectified linear unit function, which has the mathematical form of f(x)=max(0,x), that is, all negative inputs are mapped to 0, and positive inputs remain unchanged.

[0021] Furthermore, the deep convolutional neural network model directly adds the input of the module to its output through a residual connection to form a short circuit.

[0022] Furthermore, data preprocessing specifically includes:

[0023] Shear stress data processing, including:

[0024] Denoising: Use Butterworth low-pass filter to filter out high-frequency noise and retain low-frequency signals.

[0025] Detrending: Fit a linear model and subtract the linear predictors to remove the trend from the data.

[0026] Acoustic emission data processing, including:

[0027] Downsampling: The variance of the acoustic emission data is calculated using a moving window, and then the logarithm of the variance is taken.

[0028] Data Alignment: Adjust the time window length to align the shear stress and acoustic emission data in time.

[0029] Constructing time series: Designing input sequences, including the current time point and its recent historical data, for extracting time-dependent features.

[0030] Event-based continuous segmentation: The data set is divided into training set, validation set and test set to maintain the continuity of the time series and ensure that each block of data contains complete earthquake events.

[0031] Normalization: The entire dataset is normalized using the MinMax normalization method to scale the input feature values ​​to the range of 0 to 1.

[0032] Furthermore, the training strategy of the deep convolutional neural network model includes:

[0033] Loss Function: Choose the loss function MAE or MSE to measure the difference between the model prediction and the actual data.

[0034] Back-propagation algorithm: The back-propagation algorithm is used to calculate the gradient of the loss with respect to the model parameters so that the parameters can be adjusted to minimize the loss.

[0035] Learning rate and optimizer: AdamW optimizer is used to dynamically adjust the learning rate, and the change of learning rate is adjusted by cosine annealing schedule.

[0036] Gradient clipping and early stopping mechanism: Gradient clipping technology is used to prevent gradient explosion, and early stopping mechanism is used to avoid overfitting.

[0037] Model parameter initialization: Use Xavier uniform distribution to initialize model parameters.

[0038] Batch size: Choose an appropriate batch size to balance training speed and memory usage.

[0039] Mixed Precision Training: Use mixed precision training to use both single precision and half precision during model training to improve training speed and memory usage efficiency.

[0040] GPU training and distributed data parallel processing: Use GPU to train the model and adopt distributed data parallel (DDP) strategy to balance the load, improve resource utilization and optimize training speed.

[0041] Compared with the prior art, the advantages of the present invention are:

[0042] 1. Solve the problems of insufficient feature representation ability and poor model generalization of existing methods. The present invention utilizes the powerful automatic feature extraction capability of deep convolutional neural networks to adaptively learn multi-scale, hierarchical discriminative feature representations from the original acoustic emission signals, without the need to manually design features for different experimental conditions and settings. The constructed analysis model can be easily migrated and applied to different experimental equipment, experimental materials, stress states, etc., greatly improving the versatility and practicality of the method. At the same time, this method adopts special network structure designs such as causal convolution and dilated convolution, which can achieve accurate signal positioning and classification while maintaining a large receptive field, and establish an acoustic emission signal analysis model that is more accurate, reliable, and transparent than traditional machine learning methods.

[0043] 2. Solve the problems of low data utilization efficiency and small performance improvement space of existing methods. Compared with traditional machine learning methods, the present invention has a natural advantage of parallel computing. Convolution operations can be efficiently implemented through GPU acceleration, so that the model can quickly process large-scale time series data, and the computational complexity will not increase sharply with the increase in data volume. This means that even if a large amount of data is collected during the acoustic emission monitoring process, the present invention can complete the model training and laboratory earthquake prediction tasks within an acceptable time. Therefore, the present invention can more fully mine and utilize the discriminant information contained in the acoustic emission time series, and the model performance can be continuously improved with the increase of training data, while also having good computational efficiency and scalability. Therefore, the present invention can more fully mine and utilize the discriminant information contained in the acoustic emission, and the model performance can be continuously improved with the increase of training data.

[0044] 3. Solve the problem that the existing methods ignore the time series characteristics and physical mechanisms of acoustic emission signals. The present invention directly models the acoustic emission time series signals, and through the causal convolution structure and the dilated convolution mechanism, it can automatically learn the long-distance time dependency in the signal and capture the autoregressive characteristics and dynamic evolution law of the signal. The superposition of multiple layers of causal convolution gives the present invention a certain memory ability, and can model the time accumulation effect of the stress-strain state in the earthquake nucleation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a framework diagram of a deep convolutional neural network model according to an embodiment of the present invention;

[0046] Figure 2 It is a diagram of the temporal convolutional network architecture of an embodiment of the present invention;

[0047] Figure 3 is a flow chart of constructing a regression layer according to an embodiment of the present invention;

[0048] Figure 4 It is a data preprocessing flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples.

[0050] like Figure 1 As shown, the architecture of the deep convolutional neural network model of the present invention is a pure decoder (Decoder-only) architecture, which is divided into three parts: batch normalization layer (BatchNorm1d), decoder (Decoder) and regression layer (RegressionHead).

[0051] The batch normalization layer (BatchNorm1d) at the beginning of the model is used to speed up training. This is a technology widely used in deep neural networks. The basic idea is to normalize the input data of each mini-batch to make its mean 0 and variance 1. The purpose of this is to reduce the internal covariate shift problem in the neural network, making the network training process more stable and converging faster.

[0052] The decoder uses a multi-layer architecture, each of which is a temporal convolutional network (TCN) module (Bai, Kolter and Koltun 2018) ( Figure 2). Each temporal convolutional network module contains two identical dilated causal convolutions (Dilated Causal Conv1D), which are the core components. Causality means that the convolution operation can only see the information of the current time step and before, and cannot use future information, which ensures the temporal structure of the model. Dilation means that during the convolution operation, a certain number of values ​​are skipped between two adjacent input values, which expands the perception field of the convolution kernel and enables the module to capture long-distance dependencies in sequence data more efficiently. Weight normalization (WeightNorm) is also added, which is a parameter initialization and normalization technique. It divides the weight vector of the convolution kernel by its L2 norm so that the length of the weight vector is 1. This normalization method can also accelerate the convergence of the model and improve the training stability. The activation function uses the common rectified linear unit function (ReLU, Rectified Linear Unit), which performs nonlinear transformation on the input and enhances the expression ability of the network. The mathematical form of the ReLU function is f(x) = max(0, x), that is, all negative inputs are mapped to 0, and positive inputs remain unchanged. A random dropout layer is also added, which prevents the model from overfitting and improves the generalization ability of the model by randomly shutting down some neurons during the training process. Finally, the input of the module is directly added to its output through the residual connection to form a short circuit. This structure can help the gradient propagate more smoothly in the network, alleviate the gradient vanishing problem in deep networks, and thus stabilize the model training process.

[0053] The entire decoder is stacked up by multiple layers of temporal convolutional networks (TCN). These TCN modules in different layers are set with different dilation factors. The dilation factor for the i-th layer is set to 2^(i-1), where i=1,2,…,N, indicating the number of layers. Figure 3 The seven-layer decoder is shown, and the corresponding maximum expansion rate is 2^6,64. Therefore, as the number of decoder layers increases, the expansion rate increases exponentially, allowing the decoder's perception of long sequences to grow rapidly while keeping the total number of model parameters under control. In addition, such a setting will also allow the model to be flexibly adjusted when processing different tasks. When processing long time series, increase the number of decoder layers. If processing relatively short time series, reduce the number of decoder layers.

[0054] like Figure 3As shown, after the decoder is the regression layer, which is mainly a one-dimensional convolution layer with an output feature of 64 and a kernel size of 2, which further refines the extracted features. Subsequently, two linear layers (with output dimensions of 128 and 1, respectively) are used to transform the feature representation and finally map it to the target value. The rectified linear unit function (ReLU) is also used. In the decoder module, a random deactivation layer (Dropout) is intentionally placed in this embodiment, but it is intentionally omitted in the regression layer in this embodiment to avoid negative impact on the regression task.

[0055] like Figure 4 As shown, data preprocessing:

[0056] All acoustic emission and shear stress data input to the deep convolutional neural network (whether for training, validation or testing) must go through the following standardized preprocessing process:

[0057] Shear Stress Data: Denoising and Detrending

[0058] Denoising: The raw shear force data recorded in the experiment usually contains high-frequency noise, which comes from the system and the surrounding environment. To reduce this noise, a Butterworth low-pass filter is usually used. This method can effectively filter out high-frequency noise while retaining the low-frequency components of the original signal.

[0059] Detrending: After the denoising step, the target data may still contain trends resulting from non-stationary experimental conditions or systematic drift over time. If not processed, these trends may mask the actual earthquake patterns of interest in this embodiment. Therefore, in time series analysis, this embodiment needs to remove these trends from the data. This method adopts a simple and commonly used detrending method, fitting a linear model to the entire target data, and then subtracting the linear prediction value from the original data.

[0060] Acoustic Emission Data: Downsampling

[0061] Downsampling: This embodiment uses acoustic emission signals as input. Directly inputting the original acoustic emission data into the model may be inefficient, resulting in an overly large initial layer of the model. Therefore, this embodiment downsamples the acoustic emission signal data to reduce the input amount while maintaining its representative features. A moving window is usually used to calculate the various statistical features of the acoustic emission data in each window. Among these features, signal energy is considered to be the most important feature. Since the variance is proportional to the acoustic emission energy release, this embodiment only calculates the variance within each window and ignores other statistical features, which can also significantly reduce the computational requirements. However, this method of calculating only the variance also has problems, because the significant increase in acoustic emission events near the instability stage will cause the variance of this period to far exceed other times, reaching several orders of magnitude. This imbalance may hinder the deep learning model from learning the representation between events. To solve this problem, this embodiment proposes to calculate the logarithm of the variance, which helps to reduce the impact of extreme values ​​while retaining the time evolution of the acoustic emission variance during the event.

[0062] Data alignment: Since the sampling frequency of shear stress and acoustic emission data is often very different, this embodiment will align the two types of data in time by changing the length of the time window. For example, if the shear stress is recorded once every 0.01 seconds and the acoustic emission data is recorded 2 million times per second, then this embodiment will set the time window to 0.1 seconds and the interval of the time window to 0.01 seconds, so that each time window contains 200,000 acoustic emission data points.

[0063] Constructing time series: The designed deep learning model focuses on extracting potential feature values ​​from time dependence, or detecting the pattern of feature changes over time from the time series. In addition, it is also necessary to ensure that the model can only obtain the acoustic emission signal value before the corresponding shear stress timestamp to ensure that the model uses the current and past information of the acoustic emission signal to predict the current value of the shear stress. Therefore, the input consists of the variable x_t and its recent history x_(t-k+1) to x_(t-1), forming a sequence of length k, and the output corresponds to the target y_t. Therefore, the sequence length k is also a hyperparameter of the model.

[0064] Event-based continuous segmentation: Deep learning methods usually divide the data set into training set, validation set and test set. The training set is used for model learning, that is, adjusting the internal parameters of the model and optimizing the predictive ability; the validation set is used to provide an unbiased evaluation of the model performance during and after the training process, help adjust the hyperparameters of the model, and determine the stopping point to avoid overfitting; the test set is strictly used to test the predictive performance of the model after the training phase is completed. Given the time series nature of the data set in this embodiment, it is crucial to maintain the time order during the data segmentation process. Therefore, this embodiment adopts a continuous segmentation method to ensure this. In addition, the data is divided based on earthquake events rather than percentages to ensure that the earthquake events contained in each block are complete during model training, validation and testing.

[0065] Normalization: The MinMax normalization method is used to normalize the entire dataset to ensure that the input feature values ​​are between 0 and 1. To prevent data leakage, only the training set is used to adapt the normalizer.

[0066] Training strategy of the model: Loss function: The goal of training is to optimize the model parameters to minimize the loss function. This minimization means that the difference between the model prediction and the actual data in the training set is minimized. Choosing an appropriate loss function is crucial in this process. For the task of this embodiment, this embodiment considers two standard options: mean absolute error (MAE) and mean squared error (MSE). MAE is a calculation of the mean absolute difference between the observed value and the predicted value, while MSE is the average of the squared differences. The former measures the average size of the error regardless of the direction of the error, while the latter gives more weight to larger errors and is more sensitive to outliers.

[0067] Backpropagation algorithm: Use the backpropagation algorithm to propagate this loss back through the neural network and calculate the gradient of the loss with respect to the model weights, which provides information on how to adjust these weights.

[0068] Learning Rate and Optimizer: Adjustments to model parameters are guided by the learning rate, which is a hyperparameter that determines the magnitude of these adjustments. A higher learning rate may allow the model to converge quickly, but there is also a risk of exceeding the minimum, while a lower learning rate ensures more accurate convergence at the expense of slower training. Therefore, optimizers are often used to efficiently and dynamically adjust the learning rate to minimize losses. This embodiment trains the model using the AdamW optimizer, a variant of the Adam optimizer. AdamW improves model performance and computational efficiency by decoupling weight decay from the optimization step, resulting in more accurate gradient estimates.

[0069] Learning rate adjustment strategy: In addition, this embodiment also uses cosine annealing scheduling to adjust the learning rate. Cosine annealing scheduling adjusts the learning rate during training according to the cosine period, starting from a higher initial learning rate, decreasing over time, and then increasing regularly after a certain training period. This approach prevents the model from falling into a suboptimal solution, thereby improving training efficiency and results.

[0070] Gradient clipping and early stopping mechanism: During the training process, in order to avoid the gradient explosion problem, this embodiment uses the gradient clipping technique to ensure that the gradient remains within the defined range. In addition, this embodiment also uses the early stopping mechanism to prevent overfitting, that is, if the loss on the validation set does not decrease within 20 consecutive cycles, the training is terminated. This method ensures that the model will not overfit.

[0071] Model parameter initialization: Model parameter initialization has an important impact on convergence efficiency and overall performance. In this embodiment, the weights of the regression layer are initialized using Xavier uniform distribution, and the bias term is set to zero.

[0072] Batch size: The batch size determines the number of samples that are forward propagated through the network at a time. A larger batch size allows the model to process more data at the same time, thereby speeding up training, but it also requires more computing memory. In the case of this embodiment, this embodiment tried batch sizes of 32, 64, and 128. Finally, it was found that the largest batch size that can be handled by the GPU memory is used, because this can maximize the efficiency of the training process.

[0073] Mixed Precision Training: Mixed Precision Training is a method that uses both single precision (float32) and half precision (float16) during deep learning training. This method aims to optimize the training speed and memory usage of the model without sacrificing the accuracy or stability of the model. Specifically, during the forward propagation process, half precision (float16) is used for calculations. Half-precision calculations can significantly speed up operations and reduce memory usage, because float16 type data occupies half the memory space of float32 type. However, when calculating loss functions and performing backpropagation, higher numerical precision is often required to avoid numerical instability and gradient disappearance problems. Therefore, in these steps, it is usually switched back to single precision (float32). After the gradient calculation is completed, the gradients are converted from half precision back to single precision, and then these gradients are used to update the network parameters. In short, by intelligently switching different data precisions at different stages of the model, the training speed is improved and the memory usage is optimized, while maintaining the accuracy and stability of model training.

[0074] GPU training and distributed data parallel processing: Finally, this embodiment uses two RTX 4090 GPUs to train the model and adopts the distributed data parallel (DDP) strategy in the Pytorch framework. In DDP, each copy of the model is placed on a different GPU. Each GPU processes a different part of the data set, i.e., data parallelism. In each training iteration, each GPU performs forward calculations and back propagation independently, and then synchronously updates the model weights of all copies through the network. This ensures that the model copies on all GPUs remain consistent throughout the training process. The advantage is that it can balance the load on multiple GPUs, improve resource utilization, and optimize training speed by reducing communication overhead. In addition, DDP also supports flexible partitioning of models and data, so that even large models can be efficiently trained on limited hardware resources.

[0075] The method according to the present invention described above can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD ROM, RAM, floppy disk, hard disk or magneto-optical disk), or as computer code that is originally stored in a remote recording medium or a non-temporary machine-readable medium downloaded through a network and will be stored in a local recording medium, so that the method described herein can be stored in such software processing on a recording medium using a general-purpose computer, a dedicated processor or programmable or dedicated hardware (such as an ASIC or FPGA). It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code, and when the software or computer code is accessed and executed by a computer, a processor or hardware, the method described herein for predicting laboratory earthquakes based on a deep convolutional neural network is implemented. In addition, when a general-purpose computer accesses the code for implementing the processing shown herein, the execution of the code converts the general-purpose computer into a dedicated computer for executing the processing shown herein.

[0076] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the implementation methods of the present invention, and should be understood that the protection scope of the present invention is not limited to such special statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for predicting laboratory earthquakes based on deep convolutional neural networks, characterized in that: The following steps are involved: 1) Data acquisition: Collect laboratory earthquake data, including shear stress and acoustic emission data; 2) Data preprocessing: Ensure data quality and perform necessary preprocessing, including denoising, detrending, and downsampling; 3) Data alignment and time series construction: Align the shear stress and acoustic emission data in time to ensure their temporal consistency; construct a time series, using shear stress as the target value and acoustic emission data as the feature value to form a training sample; Build a deep convolutional neural network model: Design the architecture of the neural network, including the input layer, convolution layer, pooling layer, fully connected layer, as well as activation functions and regularization methods; The deep convolutional neural network model also includes: Batch Normalization Layer: This layer is used to accelerate training by normalizing each small batch of input data to make its mean 0 and variance 1; Decoder: It adopts a multi-layer architecture, each layer is a temporal convolutional network module, containing two identical dilated causal convolutions, and a random dropout layer is added to the decoder module to prevent overfitting; Regression layer: includes a one-dimensional convolutional layer with an output feature of 64 and a kernel size of 2 to further refine the extracted features; two linear layers are then used to convert the feature representation to the final target value, and the rectified linear unit function is used for activation; the random dropout layer is omitted in the regression layer; Dilation rate setting: The multi-layer temporal convolutional network module of the decoder sets different dilation rates. By adjusting the dilation factor, the model's perception of long sequences can be controlled to keep the model parameters controllable and flexibly respond to the needs of different tasks. The dilation factor for the i-th layer is set to 2^(i-1), where i=1,2,…,N, indicating the number of layers; 4) Divide the data set: Divide the data set into training set, validation set and test set; ensure the independence and time sequence of data during training; 5) Model training: Use the training set to train the deep convolutional neural network model; use the loss function, optimization algorithm and learning rate scheduling strategy to optimize the model parameters so that it can achieve a better fitting effect on the training set.

2. The method for predicting laboratory earthquakes based on a deep convolutional neural network according to claim 1, characterized in that: The deep convolutional neural network model also incorporates weight normalization.

3. The method for predicting laboratory earthquakes based on a deep convolutional neural network according to claim 1, characterized in that: The activation function uses the rectified linear unit function, which has the mathematical form of f(x)=max(0,x), that is, all negative inputs are mapped to 0, and positive inputs remain unchanged.

4. The method for predicting laboratory earthquakes based on a deep convolutional neural network according to claim 1, characterized in that: The deep convolutional neural network model directly adds the input of the module to its output through a residual connection, forming a short circuit.

5. The method for predicting laboratory earthquakes based on a deep convolutional neural network according to claim 1, characterized in that: Data preprocessing specifically includes: Shear stress data processing, including: Denoising: Use Butterworth low-pass filter to filter out high-frequency noise and retain low-frequency signals; Detrending: Fitting a linear model and subtracting the linear predictions removes the trend from the data; Acoustic emission data processing, including: Downsampling: Use a moving window to calculate the variance of the acoustic emission data, and then take the logarithm of the variance; Data alignment: Adjust the time window length to align the shear stress and acoustic emission data in time; Constructing time series: Designing input sequences, including the current time point and its recent historical data, to extract time-dependent features; Event-based continuous segmentation: The data set is divided into training set, validation set and test set to maintain the continuity of the time series and ensure that each block of data contains complete earthquake events; Normalization: The entire dataset is normalized using the MinMax normalization method to scale the input feature values ​​to the range of 0 to 1.

6. The method for predicting laboratory earthquakes based on a deep convolutional neural network according to claim 1, characterized in that: Training strategies for deep convolutional neural network models, including: Loss function: Choose the loss function MAE or MSE to measure the difference between the model prediction and the actual data; Back-propagation algorithm: The back-propagation algorithm is used to calculate the gradient of the loss with respect to the model parameters so that the parameters can be adjusted to minimize the loss; Learning rate and optimizer: Use AdamW optimizer to dynamically adjust the learning rate, and adjust the change of learning rate through cosine annealing scheduling; Gradient clipping and early stopping mechanism: Gradient clipping technology is used to prevent gradient explosion, and early stopping mechanism is used to avoid overfitting; Model parameter initialization: Use Xavier uniform distribution to initialize model parameters; Batch size: Choose the batch size to balance training speed and memory usage; Mixed precision training: Use mixed precision training to use both single precision and half precision during model training to improve training speed and memory efficiency; GPU training and distributed data parallel processing: Use GPU to train the model and adopt distributed data parallel strategy to balance the load, improve resource utilization and optimize training speed.

Images