Event type identification method and device based on waveform identification

Abstract

The invention provides an event type recognition method and device based on waveform recognition, and the method comprises the steps: training an improved Alexnet convolutional neural network through employing a multi-class event waveform sample data set of a specific region in advance, and obtaining an event type recognition model, the event type identification model can be adopted to carry out event type detection on to-be-identified waveform data in a specific area, and events such as natural earthquakes, artificial blasting, collapse and mine earthquakes are determined. According to the method, the original event waveform is used as input data for training and use of the model instead of extracting a plurality of criteria or partial frequency domain data in advance for dimensionality reduction use, so that the preprocessing process of the data can be simplified, the time required for real-time judgment of the event is shortened, and the recognition efficiency is improved; and more original information in the event can be kept in the model training and using process, so that the recognition accuracy of the event can be further improved.

Description

Technical Field

[0001] The present invention relates to the field of earthquake monitoring and early warning technology, and in particular to an event type recognition method based on waveform recognition, an event type recognition device based on waveform recognition, an event type recognition model training method, an event type recognition model training device, an electronic device and a computer-readable medium. Background Art

[0002] With the rapid development of digital seismic observation technology in my country, the density of stations has gradually increased, the earthquake monitoring capacity and level have been greatly improved, and the seismic network can capture more event signals. On the one hand, it can record natural earthquake events with smaller magnitudes, and on the other hand, it can also record a large number of non-natural earthquake events, such as artificial explosions, which can lead to false triggering of earthquake warnings. Timely and accurate removal of artificial explosion events from the natural earthquake catalog to ensure the completeness and accuracy of the catalog is crucial for regional strong earthquake prediction and earthquake hazard assessment; rapid, effective and automatic distinction between natural earthquakes and artificial explosions is of great significance for emergency response in earthquake prevention and disaster relief, and has become one of the important issues facing earthquake monitoring work at present, and it is also an issue that earthquake rescue departments and seismologists have attached great importance to in recent years.

[0003] Since the 1960s, many scholars at home and abroad have conducted extensive research and in-depth discussions on natural earthquakes and non-natural earthquakes. The main methods include the following:

[0004] 1. Use multiple eigenvalues ​​to identify natural and unnatural earthquake waveforms, which can be summarized as follows: P-wave initial motion, focal depth, ratio of body wave magnitude to surface wave magnitude, ratio of P-wave initial motion amplitude to P-wave maximum amplitude, ratio of P-wave spectrum amplitude to Love wave spectrum amplitude, and criteria extracted from the frequency domain;

[0005] 2. Using multivariate statistical analysis methods, a multivariate identification function was proposed. Donat Fah et al. applied it to earthquake and industrial blasting identification in the Alps and central Scotland. By calculating the S / P spectral ratio, the identification rate reached 83%;

[0006] 3. Learning Vector Quantization (LVQ) algorithm, Bian Yinju et al. identified 73 earthquakes and 9 explosions in Beijing and its surrounding areas. During the processing, the logarithmic spectrum amplitude values ​​of 21 sampling points of P and S waves were used as feature inputs. The results showed that the recognition effect was good and the recognition rate reached 92%;

[0007] 4. The genetic algorithm and back propagation algorithm are combined to establish a genetic BP neural network, and the recognition rate reaches 94.7%;

[0008] 5. Wavelet transform energy linearity method;

[0009] The above studies have their own characteristics and have achieved good results. However, due to the complexity of the problem itself, the above studies did not use all the information of the event record for event judgment, but first extracted certain features from the event record, that is, the data was reduced in dimension, which not only increased the complexity of the processing process, but also lost a lot of original event information. At the same time, most of the above studies are based on theoretical exploration of events in specific regions, and have not systematically found universal identification criteria; and often require basic parameters such as event time and space to be determined before further calculations can be obtained, and the timely responsiveness is poor. Therefore, in order to explore the universal identification criteria for natural earthquakes and non-natural earthquakes, and to shorten the time required from the occurrence of an event to the confirmation of the event type as much as possible, it is necessary to directly extract the characteristics that can identify natural earthquakes and non-natural earthquakes from all the information in the event record; systematically study the universal identification criteria for natural earthquakes and non-natural earthquakes, and put them into practical use to serve the specific earthquake rapid reporting work.

[0010] As AlphaGo and ChatGPT become popular around the world, humans are rethinking the development possibilities of artificial intelligence. For human physical labor and non-creative mental labor, the replacement of artificial intelligence can greatly improve efficiency, save labor costs and reduce errors. For example, in the application of seismology, researchers use artificial intelligence to reduce seismic wave noise, estimate earthquake source mechanisms, identify and locate microseismic waves (Gao, Zhang, 2019; Kuang et al, 2021; Zhang et al, 2022). In recent years, some scholars have tried to use artificial intelligence technology to classify seismic waveforms. For example, Chen et al. (2019), Zhao Ming et al. (2019), Men-Andrin et al. (2019), Zhang et al. (2020a) and Zhang et al. (2020b) applied neural network models to classify earthquake and noise waveforms, and the accuracy generally reached more than 95%. Among them, Men-Andrin et al. (2019) applied five neural network models to distinguish real-time earthquake waveforms and noise. The convolutional neural network performed the best, with the highest precision and recall. Many scholars also applied convolutional neural network models to identify natural earthquakes and artificial explosions. For example, Chen Runhang et al. (2018) extracted Mel frequencies from source waveforms. After the cepstral coefficient graph, convolutional neural network was used to classify and identify natural earthquakes and blasting events, with an accuracy of more than 95%; Linville et al. (2019) used convolutional neural network model to distinguish natural earthquake signals and quarry blasting signals in Utah, USA, and the distinction result can reach 99%; Zhou Shaohui et al. (2021) used convolutional neural network model in deep learning technology to learn and train natural earthquakes and non-natural earthquakes in Shandong area, and the training and verification recognition accuracy reached more than 98%; Zheng Zhou et al. (2023) used convolutional neural network to distinguish earthquakes, noise, blasting and abnormal waveforms in Fujian and neighboring provinces, with recognition rates of 97.9%, 99%, 99.2% and 99.3% respectively. The above work has proved that deep learning has certain development potential in seismology, and convolutional neural network is more prominent in waveform distinction.

[0011] However, these previous deep learning studies basically require more complex data processing processes to extract certain features. The data preprocessing process is complicated, which increases the event judgment time and loses a large amount of original event information, reducing the accuracy of recognition. At the same time, most of the above studies are based on theoretical exploration of events in specific regions, and have not systematically found universal recognition criteria, and have not been extended to other regions. Summary of the invention

[0012] In view of the above problems, the present invention is proposed to provide an event type identification method based on waveform recognition, an event type identification device based on waveform recognition, an event type identification model training method, an event type identification model training device, an electronic device and a computer-readable medium that overcome the above problems or at least partially solve the above problems.

[0013] The present invention discloses a method for identifying event types based on waveform recognition, the method comprising:

[0014] Collect waveform data of the area to be identified;

[0015] The waveform data to be identified in the area is input into an event type recognition model to obtain an event type; the event types include natural earthquakes, artificial blasting, collapse, and mine tremors; the event type recognition model is an improved Alexnet convolutional neural network model trained with a regional multi-class waveform sample data set.

[0016] Optionally, the event type recognition model training process is as follows:

[0017] Collect various types of waveform data in the region and mark the type of each waveform data to construct a regional multi-class waveform sample data set;

[0018] The improved Alexnet convolutional neural network model is trained using the regional multi-class waveform sample data set to obtain the event type recognition model.

[0019] Optionally, the improved Alexnet convolutional neural network model includes: 5 convolutional layers, a fully connected layer, a softmax layer and a classification output layer, and the improved Alexnet convolutional neural network model is trained using the regional multi-class waveform sample data set to obtain the event type recognition model, including:

[0020] The multi-class waveform sample data set of the region is used as the input of the improved Alexnet convolutional neural network model, and the waveform feature depth extraction is performed through the five convolutional layers respectively configured with convolution kernels of different sizes and numbers to obtain a waveform feature map and pass it to the fully connected layer;

[0021] The fully connected layer is used to set the number of event classifications and increase the learning rate factor, and uses the ReLU activation function to process the waveform feature map to obtain a waveform feature vector;

[0022] The softmax layer processes the waveform feature vector using a softmax activation function to calculate the probability value of each event category;

[0023] The number of neurons in the classification output layer corresponds to the number of event classifications. The classification output layer outputs the probability value of each event category. The error between the output probability value of each event category and the true value of each event category is calculated by using cross entropy as the cost function and back-propagated to each layer to update the weights of each layer until the cost function curve tends to be stable, thereby obtaining the event type recognition model.

[0024] Optionally, the method further comprises:

[0025] Exception processing is performed on various types of waveform data collected in the area; the exception processing includes: removing calibration, square wave, sudden jump, interference and instrument failure data.

[0026] The present invention also discloses a method for training an event type recognition model, the method comprising:

[0027] Collect various types of waveform data in the region and mark the type of each waveform data to construct a regional multi-class waveform sample data set;

[0028] The improved Alexnet convolutional neural network model is trained using the multi-class waveform sample data set in the region to obtain an event type recognition model.

[0029] Optionally, the improved Alexnet convolutional neural network model includes: 5 convolutional layers, a fully connected layer, a softmax layer and a classification output layer, and the improved Alexnet convolutional neural network model is trained using the regional multi-class waveform sample data set to obtain an event type recognition model, including:

[0030] The multi-class waveform sample data set of the region is used as the input of the improved Alexnet convolutional neural network model, and the waveform feature depth extraction is performed through the five convolutional layers respectively configured with convolution kernels of different sizes and numbers to obtain a waveform feature map and pass it to the fully connected layer;

[0031] The fully connected layer is used to set the number of event classifications and increase the learning rate factor, and uses the ReLU activation function to process the waveform feature map to obtain a waveform feature vector;

[0032] The softmax layer processes the waveform feature vector using a softmax activation function to calculate the probability value of each event category;

[0033] The number of neurons in the classification output layer corresponds to the number of event classifications. The classification output layer outputs the probability value of each event category. The error between the output probability value of each event category and the true value of each event category is calculated by using cross entropy as the cost function and back-propagated to each layer to update the weights of each layer until the cost function curve tends to be stable, thereby obtaining the event type recognition model.

[0034] The present invention also discloses an event type identification device based on waveform recognition, the device comprising:

[0035] The waveform data collection module to be identified is used to collect the waveform data to be identified in the area;

[0036] The event type recognition module is used to input the waveform data to be identified in the area into the event type recognition model to obtain the event type; the event types include natural earthquakes, artificial blasting, collapse, and mine tremors; the event type recognition model is an improved Alexnet convolutional neural network model trained with a regional multi-class waveform sample data set.

[0037] Optionally, the device further comprises:

[0038] A waveform sample set construction module is used to collect various types of waveform data in the region and mark the type of each waveform data to construct a regional multi-type waveform sample data set;

[0039] The event type recognition model training module is used to train the improved Alexnet convolutional neural network model using the regional multi-class waveform sample data set to obtain the event type recognition model.

[0040] Optionally, the improved Alexnet convolutional neural network model includes: 5 convolutional layers, a fully connected layer, a softmax layer and a classification output layer, and the event type recognition model training module includes:

[0041] A feature depth extraction submodule, used to use the regional multi-class waveform sample data set as the input of the improved Alexnet convolutional neural network model, perform waveform feature depth extraction through the five convolutional layers respectively configured with convolution kernels of different sizes and numbers to obtain a waveform feature map and pass it to the fully connected layer;

[0042] A waveform feature vector generation submodule, used in the fully connected layer to set the number of event classifications and increase the learning rate factor, and uses a ReLU activation function to process the waveform feature map to obtain a waveform feature vector;

[0043] A probability value calculation submodule, used for the softmax layer to process the waveform feature vector using a softmax activation function to calculate the probability value of each event category;

[0044] The model iteration optimization module is used in that the number of neurons in the classification output layer corresponds to the number of event classifications, and the classification output layer outputs the probability value of each event category. The error between the output probability value of each event category and the true value of each event category is calculated by using cross entropy as the cost function and back-propagated to each layer to update the weights of each layer until the cost function curve tends to be stable, thereby obtaining the event type recognition model.

[0045] Optionally, the device further comprises:

[0046] The waveform data exception processing module is used to perform exception processing on various types of waveform data collected in the area; the exception processing includes: removing calibration, square wave, sudden jump, interference and instrument failure data.

[0047] The present invention also discloses an event type recognition model training device, the device comprising:

[0048] A waveform sample set construction module is used to collect various types of waveform data in the region and mark the type of each waveform data to construct a regional multi-type waveform sample data set;

[0049] The event type recognition model training module is used to train an improved Alexnet convolutional neural network model using the multi-class waveform sample data set in the region to obtain an event type recognition model.

[0050] Optionally, the improved Alexnet convolutional neural network model includes: 5 convolutional layers, a fully connected layer, a softmax layer and a classification output layer, and the event type recognition model training module includes:

[0051] A feature depth extraction submodule, used to use the regional multi-class waveform sample data set as the input of the improved Alexnet convolutional neural network model, perform waveform feature depth extraction through the five convolutional layers respectively configured with convolution kernels of different sizes and numbers to obtain a waveform feature map and pass it to the fully connected layer;

[0052] A waveform feature vector generation submodule, used in the fully connected layer to set the number of event classifications and increase the learning rate factor, and uses a ReLU activation function to process the waveform feature map to obtain a waveform feature vector;

[0053] A probability value calculation submodule, used for the softmax layer to process the waveform feature vector using a softmax activation function to calculate the probability value of each event category;

[0054] The model iteration optimization module is used in that the number of neurons in the classification output layer corresponds to the number of event classifications, and the classification output layer outputs the probability value of each event category. The error between the output probability value of each event category and the true value of each event category is calculated by using cross entropy as the cost function and back-propagated to each layer to update the weights of each layer until the cost function curve tends to be stable, thereby obtaining the event type recognition model.

[0055] The present invention also discloses an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0056] The memory is used to store computer programs;

[0057] The processor is used to implement the method described in the present invention when executing the program stored in the memory.

[0058] The present invention also discloses one or more computer-readable media having instructions stored thereon, which, when executed by one or more processors, enable the processors to perform the method according to the present invention.

[0059] The present invention includes the following advantages:

[0060] The event type recognition method based on waveform recognition of the present invention uses a multi-class waveform sample data set of a specific area to train an improved Alexnet convolutional neural network model to obtain an event type recognition model in advance, and the event type recognition model can be used to detect the event type of the waveform data to be recognized in the specific area to determine which type of event occurred, such as a natural earthquake, artificial blasting, collapse, mining earthquake, etc. The method of the present invention uses the features containing all information of the event record, such as all waveform records or all power spectra of the original waveform, as input data for model training and use, instead of pre-extracting several criteria or part of the frequency domain data for dimensionality reduction. This can not only simplify the data preprocessing process, shorten the time required for real-time event determination, and improve recognition efficiency, but also ensure that more original information in the event is retained during model training and use, which is conducive to further improving the recognition accuracy of the event. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 is a flowchart of a method for identifying an event type based on waveform recognition provided by an embodiment of the present invention;

[0062] Figure 2 It is a schematic diagram of the structure of an improved Alexnet convolutional neural network model provided by an embodiment of the present invention;

[0063] Figure 3 It is a waveform sample diagram of natural earthquake and artificial explosion provided by an embodiment of the present invention;

[0064] Figure 4 is a workflow diagram of an automatic identifier of an event type identification model provided by an embodiment of the present invention;

[0065] Figure 5 It is the corresponding relationship between the change of the verification set ratio and the number of training samples and the accuracy rate provided by the embodiment of the present invention;

[0066] Figure 6 It is a schematic diagram of the change of the accuracy and cost function of the training set and the validation set with the number of training iterations provided by an embodiment of the present invention;

[0067] Figure 7 It is a flowchart of the steps of an event type recognition model training method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0069] Reference Figure 1 , shows a flowchart of a method for identifying an event type based on waveform recognition provided in an embodiment of the present invention, which may specifically include the following steps:

[0070] Step 101, collecting waveform data of the area to be identified;

[0071] Step 102, input the waveform data to be identified in the area into an event type recognition model to obtain event types; the event types include natural earthquakes, artificial blasting, collapse, and mine tremors; the event type recognition model is an improved Alexnet convolutional neural network model trained with a regional multi-class waveform sample data set.

[0072] The present invention pre-uses multiple types of waveform sample data sets of a specific area to train an improved Alexnet convolutional neural network model to obtain an event type recognition model. The event type recognition model can then be used to perform event type detection on the waveform data to be identified in the specific area to determine which type of event has occurred, such as a natural earthquake, artificial blasting, collapse, or mine tremor.

[0073] Specifically, when you want to identify what events have occurred in a region, you can collect the waveform data to be identified in the region, and intercept it according to the P wave arrival information of the event: intercept the 5 seconds of noise before the P wave arrives, the waveform with a total length of 60 seconds, and can completely include the P and S waves. No special processing is required, and the unified and normalized waveform data to be identified is directly provided to the trained event type recognition model, which can then judge the event type, calculate the probability value of each event type, and determine the main event type that occurs.

[0074] The original waveform image contains effective distinguishing features between different event records. The method of the present invention uses the features of all waveform records or the entire power spectrum of the original waveform, which contain all the information of the event record, as input data for model training and use, instead of pre-extracting several criteria or part of the frequency domain data for dimensionality reduction. This can not only simplify the data preprocessing process, shorten the time required for real-time event judgment, and improve recognition efficiency, but also ensure that more original information in the event is retained during model training and use, which is conducive to further improving the recognition accuracy of the event.

[0075] In one embodiment of the present invention, the event type recognition model training process is as follows:

[0076] Collect various types of waveform data in the region and mark the type of each waveform data to construct a regional multi-class waveform sample data set;

[0077] The improved Alexnet convolutional neural network model is trained using the regional multi-class waveform sample data set to obtain the event type recognition model.

[0078] In one embodiment of the present invention, the improved Alexnet convolutional neural network model includes: 5 convolutional layers, a fully connected layer, a softmax layer and a classification output layer, and the improved Alexnet convolutional neural network model is trained using the regional multi-class waveform sample data set to obtain the event type recognition model, including:

[0079] The multi-class waveform sample data set of the region is used as the input of the improved Alexnet convolutional neural network model, and the waveform feature depth extraction is performed through the five convolutional layers respectively configured with convolution kernels of different sizes and numbers to obtain a waveform feature map and pass it to the fully connected layer;

[0080] The fully connected layer is used to set the number of event classifications and increase the learning rate factor, and uses the ReLU activation function to process the waveform feature map to obtain a waveform feature vector;

[0081] The softmax layer processes the waveform feature vector using a softmax activation function to calculate the probability value of each event category;

[0082] The number of neurons in the classification output layer corresponds to the number of event classifications. The classification output layer outputs the probability value of each event category. The error between the output probability value of each event category and the true value of each event category is calculated by using cross entropy as the cost function and back-propagated to each layer to update the weights of each layer until the cost function curve tends to be stable, thereby obtaining the event type recognition model.

[0083] In one embodiment of the present invention, the method further comprises:

[0084] Exception processing is performed on various types of waveform data collected in the area; the exception processing includes: removing calibration, square wave, sudden jump, interference and instrument failure data.

[0085] The AlexNet convolutional neural network model was designed earlier and is still widely used. This network model is the model used by Krizhevsky to win the ILSVRC competition in 2012. It is mainly composed of 5 convolutional layers and 3 fully connected layers, containing a total of about 650,000 neurons and 60 million trainable parameters. In order to adapt to small batches of training and verification data sets, and to better adapt to the classification of earthquake event characteristics in specific regions, the present invention replaces the last three layers of the AlexNet model with fully connected layers, softmax layers and classification output layers, that is, the 8-layer structure includes: 5 convolutional layers, fully connected layers, softmax layers and classification output layers, and sets the number of classifications in the fully connected layers and increases its learning rate factor to achieve the purpose of accelerating the learning speed of the final layer. The Softmax function normalizes the probability rows in the classification output, and the formula is:

[0086]

[0087] In the formula, i = 0, 1, 2, 3, etc. represent natural earthquake events, artificial blasting events, collapse events, and mining earthquake events, respectively. i (x) is the unnormalized probability of class i output by the last fully connected layer. The activation function used between the two convolutional layers is ReLU, which is a nonlinear operation that replaces all negative pixel values ​​in the feature map with zero, greatly reducing the amount of calculation and accelerating the convergence speed; the pooling layer uses the maximum pooling function, and the calculation formula for each output feature map of the pooling layer is:

[0088]

[0089] In the formula, down() represents the pooling function, and each output feature map corresponds to a weight coefficient β and a bias b. The local response normalization is retained to suppress small neurons and enhance the generalization ability of the model, which is conducive to the application of the model to more areas. Figure 2 As shown, it is a schematic diagram of the convolutional neural network structure used in the present invention. The neural network is operated on dual GPUs, and the input image size is 227×227×3; the convolution kernel sizes used in each convolution layer are 11×11, 5×5, 3×3, 3×3, and 3×3 respectively; the maximum pooling operation used in each pooling layer is 3×3, that is, the window size is scaled with 3×3 to obtain features of different roughness. The purpose of the above improvements is to speed up the speed of model learning and training, improve the final discrimination accuracy, and adapt to the second-level earthquake warning speed.

[0090] The input and output details of each layer of the network used in the present invention are as follows:

[0091] The first layer: This layer is a convolution layer. It uses 96 convolution kernels to perform convolution operations on the input 227×227×3 image. The convolution kernel size is 11×11, the stride is 4, the padding is 0, and the output feature map size is 55×55. This layer uses the ReLU activation function to increase nonlinearity, and uses LRN to further optimize the model. After this layer, the maximum pooling function is used, that is, the input image is downsampled using a 3×3 maximum pooling operation with a stride of 2, which halves the feature image size and increases model stability.

[0092] The second layer: This layer is a convolution layer with 256 convolution kernels, a size of 5×5, a stride of 1, a padding of 2, and an output feature map size of 27×27. This layer also uses the ReLU activation function to increase nonlinearity, and uses LRN to further optimize the model. This layer also uses the maximum pooling function for downsampling with a stride of 2 to halve the feature image size and increase model stability.

[0093] The third layer: This layer is a convolutional layer with 384 convolution kernels of size 3×3, a step size of 1, a padding of 1, an output feature map size of 13×13, and a ReLU activation function to increase nonlinearity.

[0094] The fourth layer is a convolutional layer with 384 convolution kernels of size 3×3, a step size of 1, a padding of 1, an output feature map size of 13×13, and a ReLU activation function to increase nonlinearity.

[0095] The fifth layer is a convolutional layer with 256 convolution kernels, a size of 3×3, a stride of 1, a padding of 1, an output feature map size of 13×13, and a ReLU activation function to increase nonlinearity. This layer also uses the maximum pooling function for downsampling with a stride of 2 to halve the feature image size and increase model stability.

[0096] The sixth layer is a fully connected layer containing 4096 neurons. Its input is the output of the previous convolutional layer after pooling, that is, 13×13×256=43264 features. This layer also uses the ReLU activation function to increase nonlinearity.

[0097] The seventh layer: This layer is the softmax layer, which uses the softmax activation function to calculate the probability value of the category.

[0098] The eighth layer: This layer is the output layer. The number of neurons in this layer corresponds to the final number of classifications, and the probability value of each category is output.

[0099] In the model training process, in order to make the actual network output more accurate, it is necessary to debug the weights of each layer. The present invention uses cross entropy as the cost function to calculate the error between the predicted value and the true value and back-propagates it to each layer to update the weights of each layer. The calculation formula is:

[0100] H(p,q)=-∑ x p(x)logq(x)

[0101] In the formula, p is the true probability value, q is the predicted value, and q is the output value of the last layer of CNN calculated by the Softmax function.

[0102] The following uses the waveform distinction between natural earthquakes and artificial explosions in Guangdong as an example to introduce the training process of the event type recognition model of the present invention:

[0103] 1. Training sample preprocessing and formation

[0104] The research samples were intercepted according to the arrival information of the event P wave: the noise of 5 seconds before the arrival of the P wave was intercepted, the total length of the waveform was 60 seconds, and it could completely contain the P and S waves. In the case of sufficient event data, in order to meet the requirements of rapid identification, minimize the identification time and achieve better identification results, a single unidirectional (vertical) waveform was finally selected as the training data. After removing the data of abnormal information such as calibration, square wave, sudden jump, interference and instrument failure, a total of 1840 valid waveforms were obtained, including 1446 natural earthquake events and 394 artificial blasting waveforms, of which 1000 natural earthquake event waveforms and 300 blasting waveforms were used for training and learning, forming the training data set; the rest were used to test the trained model, forming the test data set. Figure 3 (a)-(c) are waveform samples of natural earthquakes. Figure 3 (d)-(e) are waveform sample diagrams of artificial blasting.

[0105] 2. Sample label establishment

[0106] By using the strong classification ability of neural networks, this study is transformed into a binary classification problem of natural earthquakes and artificial explosions. Natural earthquakes are the first category, and the output label result is 0; artificial explosions are the second category, and the output label result is 1.

[0107] 3. Build a convolutional neural network model

[0108] For the waveform classification problem, a single-channel 60s waveform is used as input, and the discrimination results 0 (natural earthquake) and 1 (artificial explosion) are used as output. The convolutional neural network extracts waveform features through convolution operations, and finally outputs the probability of the corresponding waveform type in a voting-like form after a series of calculations and processing. The recognition method for fast and efficient identification of natural earthquakes and artificial explosions based on the AlexNet convolutional neural network model mainly consists of two modules: training module and testing module, which can be simplified as follows: Figure 4 shown.

[0109] The core steps of the training module include:

[0110] (1) Before building the model, data with high signal-to-noise ratio, clear records, obvious waveform phases, and prominent characteristics of the two types of events were selected as training materials. This will help the model to quickly capture event characteristics and quickly generate a parameter set suitable for distinguishing Guangdong event types.

[0111] (2) Add corresponding event type labels to the training dataset.

[0112] (3) The waveforms of various events used for training are used as input, and the AlexNet convolutional neural network is used to perform event recognition training and verification. The setting of the probability threshold has little effect on the accuracy of event discrimination. CNN has high accuracy in event judgment: the judgment probability value of an image is either close to 1 or close to 0, and there are very few values ​​in the middle. Therefore, for the binary classification of this study, the probability threshold is set to 0.5. In a single training cycle, the validation set cost function and accuracy are randomly selected according to the number of iterations. When the validation set cost function curve tends to a straight line and the recognition accuracy is almost unchanged, the training is stopped. Subsequently, performance evaluation and model optimization are performed based on the results presented on the validation set, including changing the initial learning rate, adjusting the number of iterations, changing the gradient descent speed, etc., so that the model can better adapt to the non-natural earthquake recognition task in Guangdong and achieve the best recognition results.

[0113] Test module (application module):

[0114] The event waveform used for testing is randomly shuffled and input into the convolutional neural model with optimized parameters. According to the event probability output by the model, the waveform is identified as a natural earthquake event or artificial explosion, and an automatic identifier is constructed.

[0115] 4. Evaluation indicators

[0116] Although the accuracy rate can determine the overall correctness rate, it cannot be used as a good indicator to measure the results when the samples are unbalanced. In the case of unbalanced samples of natural earthquake events and artificial blasting events in this article, precision, recall and F1-score are introduced to comprehensively evaluate the model results. The precision rate indicates the number of samples correctly predicted as positive by the model; the recall rate indicates the number of positive samples predicted; and the F1-score is a harmonic average of the precision rate and the recall rate. The calculation formula is as follows:

[0117]

[0118] Wherein: TP and FP represent true positive (i.e., natural earthquake waveform is correctly predicted as natural earthquake waveform) and false positive (i.e., artificial blasting waveform is incorrectly predicted as natural earthquake waveform), respectively, while TN and FN represent true negative (i.e., artificial blasting waveform is correctly predicted as artificial blasting waveform) and false negative (i.e., natural earthquake waveform is incorrectly predicted as artificial blasting waveform).

[0119] 5. Impact of validation set ratio and number of training samples on accuracy

[0120] In the training module of the model, the input training data set is divided into a training set and a validation set, with a common ratio of 7:3 or 8:2. In order to obtain the best training effect, the training data is divided into different ratios, and then tested to see the accuracy of the validation set after training to evaluate the effect. Similarly, in order to test the impact of different numbers of training samples on the training effect, the number of training samples of natural earthquake events with a relatively large amount of data is changed to see the change in the accuracy of the validation set.

[0121] like Figure 5 As shown in (a), when the ratio of training set to validation set is greater than 3:7: the calculated total accuracy increases slightly with the ratio, the accuracy of artificial blasting fluctuates but the overall trend is to increase with the increase of the ratio, and the accuracy of natural earthquake events is higher than 95%; when the ratio is 8:2, the total accuracy reaches a maximum of 97.31%; when the ratio is greater than 8:2 (9:1), the total accuracy shows a downward trend. It can be seen that the number of samples provided to the model for learning during the training process should be more than the number used to verify the model parameters and evaluate the model effect. In this model, the accuracy is best when the ratio of training set to validation set is 8:2. Figure 5As shown in (b), as the number of natural earthquake events used for training increases, its accuracy increases. When the number is greater than 600, the accuracy is higher than 95%. Although the accuracy of artificial blasting decreases with the increase in the number of earthquakes, its accuracy is always higher than 90%. If there are enough blasting materials and they are equivalent to earthquakes, it is expected that its accuracy will reach more than 95%, and the model learning effect will be better.

[0122] 6. Neural network training results

[0123] The 1000 natural earthquake waveforms and 300 blast waveforms used for training and learning were randomly divided into training set and validation set in a ratio of 8:2 and then imported into the AlexNet convolutional neural network model. The detailed training process is as follows Figure 6 As shown in the figure, the training results of the AlexNet convolutional neural network model: the accuracy of the training set and the validation set (a) and the cost function (b) change with the number of training iterations. The entire training process took 7 minutes and 4 seconds. As can be seen from the figure, with the increase in the number of training times, the accuracy of the training set gradually increased and exceeded 90%, and finally stabilized near a certain value; at the same time, the cost function curve dropped rapidly and finally stabilized near a relatively small value and no longer changed; it should be emphasized that the model did not overfit. In the end, the accuracy of the training set was as high as 100%, and the validation set was 97.31%.

[0124] 7. Test results

[0125] This example tests the trained model, randomly shuffles the remaining 446 natural earthquake and 94 artificial blasting event waveforms and imports them into the model for event type recognition. The recognition takes less than 2 seconds, and 526 samples are accurately recognized, with an accuracy rate of 97.41%. This paper evaluates the robustness of the network model by the percentage of correctly classified waveform types, as shown in Table 1. The accuracy of natural earthquakes is over 98%, and the accuracy of blasting is over 90%. It is concluded that the model can effectively identify natural earthquakes and artificial blasting waveforms in Guangdong.

[0126] Table 1 Precision, recall and F1-score of natural earthquakes and artificial explosions obtained from the test

[0127]

[0128] As shown in the confusion matrix in Table 2 (the sum of each row represents the number of true samples of this category, and the sum of each column represents the number of samples predicted to be of this category), there were 9 classification errors for seismic waveforms and 5 classification errors for artificial blasting. From the proportion, it can be seen that the recognition rate of the model for seismic waveforms is better than that for artificial blasting.

[0129] Table 2 Confusion matrix of test recognition results

[0130]

[0131] It can be seen from the Guangdong example that the present invention applies the improved Alexnet convolutional neural network model to the field of earthquake monitoring and early warning, which can solve the problem of manual event classification, which is time-consuming and tests the professional knowledge of the staff; and according to the waveform characteristics of events in different regions, the initial learning rate, the number of iterations, the gradient descent rate, etc. are adjusted to make the method better adapt to regional characteristics and optimize the recognition effect. This example has adjusted the above parameters according to the characteristics of earthquake waveforms in Guangdong to adapt to the distinction between artificial blasting and natural earthquake waveforms in Guangdong. If it is to be applied to other regions and other event waveform distinctions, it only needs to fine-tune the above parameters according to the training process.

[0132] Reference Figure 7 , shows a flowchart of a method for training an event type recognition model provided in an embodiment of the present invention, which may specifically include the following steps:

[0133] Step 701, collecting multiple types of waveform data in a region and marking the type to which each waveform data belongs, to construct a regional multi-class waveform sample data set;

[0134] Step 702: Use the regional multi-class waveform sample data set to train an improved Alexnet convolutional neural network model to obtain an event type recognition model.

[0135] In one embodiment of the present invention, the improved Alexnet convolutional neural network model includes: 5 convolutional layers, a fully connected layer, a softmax layer and a classification output layer, and the improved Alexnet convolutional neural network model is trained using the regional multi-class waveform sample data set to obtain the event type recognition model, including:

[0136] The multi-class waveform sample data set of the region is used as the input of the improved Alexnet convolutional neural network model, and the waveform feature depth extraction is performed through the five convolutional layers respectively configured with convolution kernels of different sizes and numbers to obtain a waveform feature map and pass it to the fully connected layer;

[0137] The fully connected layer is used to set the number of event classifications and increase the learning rate factor, and uses the ReLU activation function to process the waveform feature map to obtain a waveform feature vector;

[0138] The softmax layer processes the waveform feature vector using a softmax activation function to calculate the probability value of each event category;

[0139] The number of neurons in the classification output layer corresponds to the number of event classifications. The classification output layer outputs the probability value of each event category. The error between the output probability value of each event category and the true value of each event category is calculated by using cross entropy as the cost function and back-propagated to each layer to update the weights of each layer until the cost function curve tends to be stable, thereby obtaining the event type recognition model.

[0140] In one embodiment of the present invention, the method further comprises:

[0141] Exception processing is performed on various types of waveform data collected in the area; the exception processing includes: removing calibration, square wave, sudden jump, interference and instrument failure data.

[0142] This method embodiment is basically similar to the embodiment "A method for identifying event types based on waveform recognition", so the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0143] It should be noted that, for the sake of simplicity, the method embodiments are described as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because according to the embodiments of the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0144] An event type identification device based on waveform recognition provided in an embodiment of the present invention may specifically include the following modules:

[0145] The waveform data collection module to be identified is used to collect the waveform data to be identified in the area;

[0146] The event type recognition module is used to input the waveform data to be identified in the area into the event type recognition model to obtain the event type; the event types include natural earthquakes, artificial blasting, collapse, and mine tremors; the event type recognition model is an improved Alexnet convolutional neural network model trained with a regional multi-class waveform sample data set.

[0147] In one embodiment of the present invention, the device further comprises:

[0148] A waveform sample set construction module is used to collect various types of waveform data in the region and mark the type of each waveform data to construct a regional multi-type waveform sample data set;

[0149] The event type recognition model training module is used to train the improved Alexnet convolutional neural network model using the regional multi-class waveform sample data set to obtain the event type recognition model.

[0150] In one embodiment of the present invention, the improved Alexnet convolutional neural network model includes: 5 convolutional layers, a fully connected layer, a softmax layer and a classification output layer, and the event type recognition model training module includes:

[0151] A feature depth extraction submodule, used to use the regional multi-class waveform sample data set as the input of the improved Alexnet convolutional neural network model, perform waveform feature depth extraction through the five convolutional layers respectively configured with convolution kernels of different sizes and numbers to obtain a waveform feature map and pass it to the fully connected layer;

[0152] A waveform feature vector generation submodule, used in the fully connected layer to set the number of event classifications and increase the learning rate factor, and uses a ReLU activation function to process the waveform feature map to obtain a waveform feature vector;

[0153] A probability value calculation submodule, used for the softmax layer to process the waveform feature vector using a softmax activation function to calculate the probability value of each event category;

[0154] The model iteration optimization module is used in that the number of neurons in the classification output layer corresponds to the number of event classifications, and the classification output layer outputs the probability value of each event category. The error between the output probability value of each event category and the true value of each event category is calculated by using cross entropy as the cost function and back-propagated to each layer to update the weights of each layer until the cost function curve tends to be stable, thereby obtaining the event type recognition model.

[0155] In one embodiment of the present invention, the device further comprises:

[0156] The waveform data exception processing module is used to perform exception processing on various types of waveform data collected in the area; the exception processing includes: removing calibration, square wave, sudden jump, interference and instrument failure data.

[0157] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0158] An event type recognition model training device provided in an embodiment of the present invention may specifically include the following modules:

[0159] A waveform sample set construction module is used to collect various types of waveform data in the region and mark the type of each waveform data to construct a regional multi-type waveform sample data set;

[0160] The event type recognition model training module is used to train an improved Alexnet convolutional neural network model using the multi-class waveform sample data set in the region to obtain an event type recognition model.

[0161] In one embodiment of the present invention, the improved Alexnet convolutional neural network model includes: 5 convolutional layers, a fully connected layer, a softmax layer and a classification output layer, and the event type recognition model training module includes:

[0162] A feature depth extraction submodule, used to use the regional multi-class waveform sample data set as the input of the improved Alexnet convolutional neural network model, perform waveform feature depth extraction through the five convolutional layers respectively configured with convolution kernels of different sizes and numbers to obtain a waveform feature map and pass it to the fully connected layer;

[0163] A waveform feature vector generation submodule, used in the fully connected layer to set the number of event classifications and increase the learning rate factor, and uses a ReLU activation function to process the waveform feature map to obtain a waveform feature vector;

[0164] A probability value calculation submodule, used for the softmax layer to process the waveform feature vector using a softmax activation function to calculate the probability value of each event category;

[0165] The model iteration optimization module is used in that the number of neurons in the classification output layer corresponds to the number of event classifications, and the classification output layer outputs the probability value of each event category. The error between the output probability value of each event category and the true value of each event category is calculated by using cross entropy as the cost function and back-propagated to each layer to update the weights of each layer until the cost function curve tends to be stable, thereby obtaining the event type recognition model.

[0166] In one embodiment of the present invention, the device further comprises:

[0167] The waveform data exception processing module is used to perform exception processing on various types of waveform data collected in the area; the exception processing includes: removing calibration, square wave, sudden jump, interference and instrument failure data.

[0168] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0169] In addition, an embodiment of the present invention further provides an electronic device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus.

[0170] Memory, used to store computer programs;

[0171] The processor is used to implement the method described in the above embodiment when executing the program stored in the memory.

[0172] The communication bus mentioned in the above terminal can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0173] The communication interface is used for communication between the above terminal and other devices.

[0174] The memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0175] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0176] In another embodiment of the present invention, a computer-readable storage medium is provided. The computer-readable storage medium stores instructions. When the instructions are executed on a computer, the computer executes the method described in the above embodiment.

[0177] In another embodiment of the present invention, a computer program product including instructions is provided. When the computer program product is run on a computer, the computer executes the method described in the above embodiment.

[0178] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk Solid State Disk (SSD)), etc.

[0179] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0180] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0181] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A method for event type recognition based on waveform recognition, characterized in that: The method comprises: Collect waveform data of the area to be identified; The waveform data to be identified in the area is input into an event type recognition model to obtain an event type; the event types include natural earthquakes, artificial blasting, collapse, and mine tremors; the event type recognition model is an improved Alexnet convolutional neural network model trained with a regional multi-class waveform sample data set.

2. The method according to claim 1, characterized in that: The event type recognition model training process is as follows: Collect various types of waveform data in the region and mark the type of each waveform data to construct a regional multi-class waveform sample data set; The improved Alexnet convolutional neural network model is trained using the regional multi-class waveform sample data set to obtain the event type recognition model.

3. The method according to claim 2, characterized in that The improved Alexnet convolutional neural network model includes: 5 convolutional layers, a fully connected layer, a softmax layer and a classification output layer. The improved Alexnet convolutional neural network model is trained using the regional multi-class waveform sample data set to obtain the event type recognition model, including: The multi-class waveform sample data set of the region is used as the input of the improved Alexnet convolutional neural network model, and the waveform feature depth extraction is performed through the five convolutional layers respectively configured with convolution kernels of different sizes and numbers to obtain a waveform feature map and pass it to the fully connected layer; The fully connected layer is used to set the number of event classifications and increase the learning rate factor, and uses the ReLU activation function to process the waveform feature map to obtain a waveform feature vector; The softmax layer processes the waveform feature vector using a softmax activation function to calculate the probability value of each event category; The number of neurons in the classification output layer corresponds to the number of event classifications. The classification output layer outputs the probability value of each event category. The error between the output probability value of each event category and the true value of each event category is calculated by using cross entropy as the cost function and back-propagated to each layer to update the weights of each layer until the cost function curve tends to be stable, thereby obtaining the event type recognition model.

4. The method according to claim 2, characterized in that: The method further comprises: Exception processing is performed on various types of waveform data collected in the area; the exception processing includes: removing calibration, square wave, sudden jump, interference and instrument failure data.

5. A method for training an event type recognition model, characterized in that: The method comprises: Collect various types of waveform data in the region and mark the type of each waveform data to construct a regional multi-class waveform sample data set; The improved Alexnet convolutional neural network model is trained using the multi-class waveform sample data set in the region to obtain an event type recognition model.

6. The method according to claim 5, characterized in that The improved Alexnet convolutional neural network model includes: 5 convolutional layers, a fully connected layer, a softmax layer and a classification output layer. The improved Alexnet convolutional neural network model is trained using the regional multi-class waveform sample data set to obtain an event type recognition model, including: The multi-class waveform sample data set of the region is used as the input of the improved Alexnet convolutional neural network model, and the waveform feature depth extraction is performed through the five convolutional layers respectively configured with convolution kernels of different sizes and numbers to obtain a waveform feature map and pass it to the fully connected layer; The fully connected layer is used to set the number of event classifications and increase the learning rate factor, and uses the ReLU activation function to process the waveform feature map to obtain a waveform feature vector; The softmax layer processes the waveform feature vector using a softmax activation function to calculate the probability value of each event category; The number of neurons in the classification output layer corresponds to the number of event classifications. The classification output layer outputs the probability value of each event category. The error between the output probability value of each event category and the true value of each event category is calculated by using cross entropy as the cost function and back-propagated to each layer to update the weights of each layer until the cost function curve tends to be stable, thereby obtaining the event type recognition model.

7. An event type identification device based on waveform recognition, characterized in that: The device comprises: The waveform data collection module to be identified is used to collect the waveform data to be identified in the area; The event type recognition module is used to input the waveform data to be identified in the area into the event type recognition model to obtain the event type; the event types include natural earthquakes, artificial blasting, collapse, and mine tremors; the event type recognition model is an improved Alexnet convolutional neural network model trained with a regional multi-class waveform sample data set.

8. An event type recognition model training device, characterized in that: The device comprises: A waveform sample set construction module is used to collect various types of waveform data in the region and mark the type of each waveform data to construct a regional multi-type waveform sample data set; The event type recognition model training module is used to train an improved Alexnet convolutional neural network model using the multi-class waveform sample data set in the region to obtain an event type recognition model.

9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; The memory is used to store computer programs; The processor, when used to execute a program stored in the memory, implements the method described in any one of claims 1-4 or 5-6.

10. One or more computer-readable media having instructions stored thereon, which, when executed by one or more processors, cause the processors to perform the method of any one of claims 1-4 or 5-6.

Images