A method and system for real-time identification of natural offshore earthquakes using small sample and lightweight deep learning

Abstract

The present application provides a method and system for real-time identification of natural earthquakes at sea using small sample lightweight deep learning, the method comprising: inputting the sound pressure signal recorded by a snorkeling marine seismograph into multiple trained lightweight deep learning networks, and outputting whether a natural earthquake at sea has occurred; for the output results of multiple networks, a voting mechanism is adopted, and the voting result is used as the final recognition result; the training of the lightweight deep learning network comprises: preprocessing the sound pressure signal recorded by the snorkeling marine seismograph; using a supervised learning method to perform multiple model training; the preprocessing comprises: uniformly sampling the data to a fixed frequency and a set length, and performing a 0.05-3Hz bandpass filtering process on the data. The advantages of the present application are: the ability to accurately identify natural earthquake signals in complex background noise; and the ability to be deployed on resource-constrained embedded devices and mobile devices with high requirements for real-time processing.

Description

Technical Field

[0001] The present application belongs to the field of deep seabed detection and natural earthquake identification, and specifically relates to a method and system for real-time identification of natural offshore earthquakes using small sample lightweight deep learning. Background Art

[0002] The ocean area covers 70% of the earth's surface. However, due to the level of observation instruments and economic costs, it is difficult for the seismic network in the ocean area to achieve long-term continuous observation, and the seismic observation data obtained are relatively sparse, which limits the detection of deep mantle structures in the ocean area by seismic tomography. In recent years, a new type of marine earthquake monitoring equipment has been developed: the snorkeling marine seismometer (MERMAID), which can float with the ocean current at a specified depth underwater. It is designed to actively float up and send back seismic waveform signals after recording major global earthquake signals. Due to the limited battery capacity carried by the instrument, misidentification will cause the instrument to float frequently, thus shortening its life and affecting its long-term monitoring function. Being too "sluggish" will affect its reporting of natural earthquakes that can be used for tomography worldwide. Compared with the earthquake monitoring system on land, the snorkeling marine seismometer has put forward more stringent algorithm requirements for improving the recognition rate of earthquake signals and reducing the false alarm rate.

[0003] In theory, the identification of marine earthquakes can refer to the commonly used earthquake cataloging methods on land. Commonly used methods include the short-to-long period average energy ratio (STA / LTA) method and the cross-correlation method. The STA / LTA method is to count the energy of the signal in a sliding time window and compare it with the energy in a longer sliding window to find the situation of abnormal energy increase. The STA / LTA method has high requirements for the signal-to-noise ratio of the signal, and manual confirmation is required after picking up the seismic signal to reduce the probability of false picking. The cross-correlation method is to perform a cross-correlation operation on the intercepted signal with the existing template to detect the similarity between the data and the existing seismic template to determine whether there is a seismic signal in the intercepted signal. This method requires manual selection of templates, and the templates have regional characteristics. Different templates need to be selected for earthquakes in different regions. This method requires high experience of technicians, huge manual workload, and low efficiency.

[0004] The deep learning methods developed in recent years have been introduced into the field of earthquakes, mainly used to distinguish natural earthquakes and explosion signals on land, and some of them are used to detect natural earthquakes. This type of method inputs manually labeled data into the deep learning network to train the model. The source mechanism of natural earthquakes is complex, and the near-surface geological conditions are complex and diverse, which leads to the diversity of natural earthquake signals received in different areas on land. In order to improve the fitting ability of the model, people have designed more complex models. Complex models require larger training samples. In order to prevent overfitting, such models often require massive earthquake data accumulated by the seismic network over the years to train. Insufficient training data will greatly reduce the accuracy of such models. Land earthquakes and marine earthquakes are both caused by the fracture and displacement of the earth's internal plates, but due to the different propagation media of the two, the signals of the two have obvious differences. Monitoring stations on land can receive longitudinal waves, transverse waves and interface waves propagating on the surface of the earth caused by earthquakes. However, since transverse waves cannot propagate in water, pressure detectors in seawater cannot receive transverse waves. Snorkeling marine seismometers are generally placed about 1,500 meters below the sea surface, far away from the seabed, and can hardly receive surface waves propagating along the seabed interface. People have established dense earthquake monitoring stations on land, using broadband velocity or acceleration detectors to monitor signals, and can obtain three-component vector signals with a frequency band range of 120s-150Hz. Land detectors have low self-noise levels and are generally placed in quiet areas and buried 10 meters underground. They are well coupled with the ground and have high signal quality. Compared with the situation on land, the monitoring means of offshore earthquakes are limited, the accuracy of instrument recording is low, and real-time transmission is difficult. Seismic data is recorded by hydrophones, which can only obtain a single scalar pressure signal, generally with a low frequency of several hertz, and as low as 0.1 hertz. The hydrophone floats in the water, and its signal quality is affected by flow noise and interference noise from surrounding ships. In summary, the onshore seismic network is dense, the data is sufficient, the received wave field components are complex, the low-frequency information is rich, and the signal-to-noise ratio is high. However, offshore seismic data collection is difficult, the data is scarce, the received wave field components are single, and the frequency band is narrow, and the signal-to-noise ratio is low. At present, there is a lack of deep learning models designed specifically for the identification of natural offshore earthquakes. The existing earthquake identification models are designed based on the diversity of onshore earthquakes and the large amount of historical earthquake data available. Directly using them in offshore earthquake identification has the problem of poor generalization ability, and retraining models also have the problem that complex models are prone to overfitting on small sample data at sea. In addition, the existing earthquake identification models have a large amount of calculation and are not suitable for deployment in low-power snorkeling marine seismometers. Summary of the invention

[0005] The purpose of this application is to solve the problem that the existing algorithms have low online recognition rate and high missed detection rate for natural earthquakes at sea, resulting in a limited amount of data transmitted back by submersible marine seismographs, to improve the correct recognition rate of natural earthquake signals at sea, to meet the urgent needs of submersible marine seismographs for deploying low-power algorithms, and to realize real-time online high-quality autonomous detection of natural earthquakes by mobile marine equipment.

[0006] In order to achieve the above objectives, this application proposes a small sample lightweight deep learning method for real-time identification of natural earthquakes at sea, including:

[0007] The acoustic pressure signals recorded by the snorkeling marine seismometer are input into multiple trained lightweight deep learning networks to output whether a natural offshore earthquake has occurred.

[0008] For the output results of multiple networks, a voting mechanism is adopted, and the voting results are used as the final recognition results;

[0009] The training of the lightweight deep learning network includes:

[0010] Preprocessing of acoustic pressure signals recorded by snorkeling marine seismographs;

[0011] The data is randomly arranged and then divided into n non-overlapping equal parts. One of the data is selected as the validation set, and the remaining n-1 data are selected as the training set. The supervised learning method is used to perform n model training;

[0012] The preprocessing includes: uniformly sampling the data to a fixed frequency and a set length, and performing 0.05-3 Hz bandpass filtering on the data.

[0013] As an improvement of the above method, the lightweight deep learning network includes 1 convolution layer, 1 activation layer, 1 one-dimensional batch normalization layer, 1 adaptive average pooling layer, 1 flattening layer and 1 fully connected layer;

[0014] Batch normalization layers are added after each layer.

[0015] As an improvement of the above method, the lightweight deep learning network has a convolution kernel size of 33*1, a kernel stride of 1, and a number of channels of 32; a linear rectification function is used for the input layer and the intermediate layer.

[0016] As an improvement of the above method, during the training process of the lightweight deep learning network, a cross entropy loss function is selected to calculate the loss, a small batch stochastic gradient descent algorithm is used to update the model parameters, an Adam optimizer that can adaptively adjust the learning rate is used, and an early stopping strategy is used to prevent overfitting.

[0017] As an improvement of the above method, the set length is 400 seconds.

[0018] The present application also provides a small sample lightweight deep learning real-time identification system for natural earthquakes at sea, which is implemented based on the above method, and the system includes:

[0019] The earthquake recognition module is used to input the sound pressure signals recorded by the snorkeling marine seismometer into multiple trained lightweight deep learning networks to output whether a natural offshore earthquake has occurred; a voting mechanism is used for the output results of multiple networks, and the voting results are used as the final judgment results;

[0020] Training module, used to train lightweight deep learning networks.

[0021] Compared with the prior art, the advantages of this application are:

[0022] 1. Different from the quiet monitoring environment of land seismic networks, hydrophones floating in the ocean face a complex ocean environment of flow noise. Since the network is designed specifically for the characteristics of natural marine earthquake data, and many strategies such as normalization and regularization are used to improve the robustness of the parameter solution optimization process, the model of the present invention can accurately identify natural earthquake signals in complex background noise.

[0023] 2. Due to the lightweight deep learning network design and the training strategy for small sample data sets, the network parameters of the model of the present invention are small and the power consumption is low. The present invention can be deployed on resource-constrained embedded devices and mobile devices with high requirements for real-time processing, and the snorkeling marine seismometer is a typical application scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 The figure shows an original sound pressure signal recorded by a mobile marine seismograph, a sound pressure signal after 0.05-3Hz filtering, a spectrum of the signal, and a time-frequency spectrum obtained after the short-time Fourier transform of the signal. The signal is marked as a positive sample.

[0025] Figure 2 The figure shows an original sound pressure signal recorded by a mobile marine seismograph, a sound pressure signal after 0.05-3Hz filtering, a spectrum of the signal, and a time-frequency spectrum obtained after the short-time Fourier transform of the signal. The signal is marked as a negative sample.

[0026] Figure 3 The figure shows the framework of a method for real-time identification of natural offshore earthquakes using small sample lightweight deep learning;

[0027] Figure 4The figure shows the change of model accuracy (Accuracy) and loss (Loss) with the number of training times (the first group in the four-fold data); Iteraton: iteration; Training Dataset: training data set; validation Dataset: validation data set;

[0028] Figure 5 The figure shows the change of model accuracy (Accuracy) and loss (Loss) with the number of training times (the second group in the four-fold data);

[0029] Figure 6 The figure shows the change of model accuracy (Accuracy) and loss (Loss) with the number of training times (the third group in the four-fold data);

[0030] Figure 7 The figure shows the change of model accuracy (Accuracy) and loss (Loss) with the number of training times (the fourth group in the four-fold data);

[0031] Figure 8 The figure shows the confusion matrices of the four models obtained by four-fold cross validation on the validation data; Predicted: predicted value; Actual: actual value; Recall: recall rate; Precision: accuracy;

[0032] Fig. 9 Shown is the average of the confusion matrices obtained by the four models on the validation data;

[0033] Fig.10 Shown are the model structure and parameters of the small sample lightweight deep learning method for real-time identification of natural offshore earthquakes. DETAILED DESCRIPTION

[0034] The technical solution of the present application is described in detail below with reference to the accompanying drawings.

[0035] Based on the characteristics of offshore earthquakes, the present invention provides a lightweight deep learning model that can be trained on a small sample set of marine seismic data while avoiding overfitting and can be deployed on low-power devices. At the same time, it improves the picking accuracy and realizes real-time autonomous detection and recognition of hydroacoustic event signals based on data-driven in complex marine environments and with limited data samples.

[0036] Example 1

[0037] like Figure 3 As shown, the present invention provides a small sample lightweight deep learning method for identifying natural earthquakes at sea, comprising the following steps:

[0038] 1. Obtain the sound pressure signal recorded by the snorkeling marine seismograph, compare it with the global earthquake catalog, intercept the natural earthquake signal, and avoid intercepting the signal at the time of the earthquake as background noise, resample the data to a fixed frequency and the same length to form a sample library, and perform 0.05-3Hz bandpass filtering on the data in the sample library according to the characteristics of receiving natural earthquake signals at sea. According to the signal's spectrum diagram and time-frequency diagram, comprehensively judge whether the signal is a high-quality earthquake signal or background noise based on multi-scale characteristics.

[0039] 2. Randomly arrange the data, fully "shuffle", and then divide it into n non-overlapping equal parts. Select one of the data as the validation set and the remaining n-1 data as the training set. Perform n model training to ensure that each data has served as validation data to reduce the impact of uneven sample distribution in small sample data sets on model convergence. Implement an early stopping strategy by checking the accuracy of the model on the validation data to avoid the problem of overfitting of deep learning networks on small sample data.

[0040] 3. Construct a lightweight deep learning network, including 1 convolutional layer Conv (Convolutional Layer), 1 activation layer (Activation Function), 1 one-dimensional batch normalization layer (BatchNorm1d), 1 adaptive average pooling layer, 1 flattening layer and 1 fully connected layer FC (Fully-connected layer). The network has a simple hierarchical structure and a small number of parameters. It can be trained on a small sample data set and meet the task requirements of offshore natural earthquake identification. In order to alleviate the possible instability of modern neural networks, a batch normalization (BN) layer is added after each layer of the deep learning network. The size of the network's convolution kernel is designed according to the main frequency band of the sound pressure signal recorded by the hydrophone to fully learn the spectral characteristics of natural earthquakes recorded by snorkeling marine seismographs.

[0041] 4. The model training process uses supervised learning, and both the training set and the validation set data are labeled. The initial value of the model is randomly generated, and the model training is completed by using the forward propagation of data and the back propagation of errors. The model training process is the process of solving or optimizing the model parameters. The cross entropy loss function is selected to calculate the loss. In order to prevent the model from overfitting and improve the generalization ability, the small batch stochastic gradient descent algorithm is used to update the model parameters during the iteration process. In order to make the training process easier to converge, the Adam (adaptive Moment Estimation) optimizer that can adaptively adjust the learning rate is used. The early stopping strategy is used to prevent overfitting. During the training process, when the loss of the validation data does not decrease within several iterations, the training phase will be stopped in advance. According to the complexity of the marine natural earthquake identification task and the sample size, the appropriate batch size is selected. It is ensured that the large batch calculation gradient obtains accurate gradients to avoid the influence of individual wrong labels on the model, and the small batch calculation gradient is taken into account to avoid the model falling into the local optimum.

[0042] 5. After training, n models are obtained. The cross-validation method allows all data to serve as training data and verification data. The data set can be fully utilized for model evaluation, making the obtained model stable and reliable. The test data to be identified is input into the n trained models to obtain n outputs. A voting mechanism is used, and the voting result is used as the final judgment result. The voting method of multiple sub-models can alleviate the disturbance of the model training results caused by a small amount of abnormal data that may be faced by the small sample data set of natural earthquakes at sea.

[0043] The present invention will be further described below in conjunction with embodiments and drawings:

[0044] The snorkeling marine seismometer deployed at a depth of 1500 meters in the deep sea, its hydrophone continuously records the broadband sound pressure signal p(n) radiated by the underwater sound source. By comparing the global earthquake catalog of the Incorporated Research Institutions for Seismology (IRIS) and the China Earthquake Network Center (CENC), the theoretical arrival time of the earthquake is calculated, the signal of the arrival time of the seismic longitudinal wave is intercepted, and the signal is trend filtered, 0.05-3Hz bandpass filtered, spectrum analyzed, and time-spectrum analyzed. The data with a certain signal-to-noise ratio is marked as a natural earthquake signal as a positive sample, such as Figure 1 At the same time, we select the data when there is no natural earthquake, and perform trend filtering, 0.05-3Hz bandpass filtering, spectrum analysis, and time-spectrum analysis. The data when there is no pulse signal is regarded as background noise and marked as negative samples, as shown in Figure 1. Figure 2 shown.

[0045] The data is preprocessed. Considering the composition characteristics of ocean background noise, hydrophones are mainly affected by tides and waves, microseismic, ocean turbulence, distant ships, wind, etc. when working in the deep sea. Tides and waves will form long-period low-frequency interference due to changes in the static pressure of seawater, and its frequency is much lower than the lowest frequency of 0.1 Hz of conventional hydrophones. The wind on the sea surface can generate gravity waves. The nonlinear interaction of surface gravity waves with opposite or nearly opposite propagation directions will generate a hydroacoustic signal with twice the frequency of the sea surface wave. This signal will become a microseismic signal when it reaches the seabed and effectively propagate inside the earth. These signals will generate a peak value around 0.1-0.2Hz. The irregular water flow in the ocean forms turbulence, which will cause the hydrophone to vibrate and generate self-noise. The frequency of turbulence is between 1-20Hz. Ship traffic noise can be transmitted over long distances in water, and its frequency is between tens to hundreds of hertz. Seismographs at sea are sparsely distributed, and most of the signals they receive are natural earthquakes at long distances. The high-frequency attenuation of the signals after long-distance transmission is serious. The frequency of natural earthquakes recorded by hydrophones is generally below 10 Hz, and its main frequency is around 0.5 Hz. Based on the principle of suppressing noise and leaving as many effective signals as possible, the data is bandpass filtered at 0.05-3 Hz. The sampling rates of different devices are different, and the data is resampled to 20 Hz before filtering. The duration of natural earthquake signals at sea is around 180 seconds. In order to obtain as many earthquake signal features as possible, the sample length is fixed at 400 seconds. For samples less than 400 seconds, zero padding is adopted. The prepared data forms a sample library. After fully "shuffling" the data, divide it into four equal parts. The ratio of positive and negative samples in each data is consistent, and there is no overlap between each data. Select three of the data as training data D, and the remaining one as verification data D'. Make each data set serve as verification set D' in turn, and a total of 4 sets of training-verification data divisions can be obtained. 4 models are trained on these training-verification sets. The process of real-time identification of natural offshore earthquakes by small sample lightweight deep learning is as follows Figure 3 shown.

[0046] Construct lightweight convolutional neural networks such as Fig.10 As shown in the figure, it includes 1 convolutional layer Conv (ConvolutionalLayer), 1 activation layer (Activation Function), 1 one-dimensional batch normalization layer (Batch Normalization, BN), 1 adaptive average pooling layer, 1 flattening layer and 1 fully connected layer FC (Fully-connected layer). This lightweight deep learning network contains a total of 1794 parameters. The operation rules of CNN convolution operation are as follows:

[0047] h n =σ n (W n-1 *h n-1 ) (1)

[0048] Among them, h n The data dimension is m n , * represents one-dimensional convolution, W n-1 is the corresponding convolution kernel, and the weight coefficients it contains are the parameters to be solved.

[0049] like The convolution kernel The one-dimensional convolution represented by * is defined as:

[0050]

[0051] Convolution operation is the core component of CNN to extract features from spatially structured data. The main frequency of long-distance natural earthquakes at sea is below 3Hz. In order to obtain sufficient seismic signal features, the size of the convolution kernel is designed to be 33*1, the kernel stride is set to 1, and the number of channels is set to 32. The activation function performs element-by-element nonlinear operations on the output of the previous layer, and uses this result as the output of this layer. The activation function helps the CNN network to represent abstract feature patterns. In order to alleviate the gradient disappearance caused by multi-layer transmission during network parameter optimization, in this network architecture, the input layer and the middle layer use the linear rectified function (Rectifed Linear Unit, ReLU). If x is before activation, then after ReLU:

[0052] σ(x)=max(0,x) (3)

[0053] The ReLU function can automatically delete some neurons, reduce the model capacity, and speed up convergence.

[0054] By adding a batch normalization layer (BatchNorm1d) after each layer of the deep neural network, data from different batches can maintain the same scale at each layer of the network, reducing the impact of the values ​​of the parameters of each layer of the network on the scale of the output data and accelerating the network training process.

[0055] The labeled data is input into the model, and the training process adopts supervised learning. The natural earthquake identification problem can be regarded as a binary classification problem of earthquake and non-earthquake. This network selects the softmax function as the classifier. In the classification problem, the cross entropy loss function is used as the optimization target:

[0056]

[0057] Among them, K represents the training data The number of samples in, M represents the number of categories, represents the prediction of the network model. Θ is the parameter to be optimized for the convolutional network, L CE is the function L of the parameter Θ to be solved CE (Θ), the training of convolutional neural network mainly updates the weights Θ in the network through the back-propagation algorithm. The sample size of natural earthquake data at sea is small, which is prone to model overfitting. To avoid this problem, the mini-batch stochastic gradient descent algorithm (SGD) is used in the iterative cycle to update the model parameters to improve the generalization ability of the model. The learning rate is an important parameter of deep learning. The Adam (adaptive Moment Estimation) optimizer that can adaptively adjust the learning rate is used to update the parameters, and the initial learning rate is set to 0.001. The early stopping strategy is further adopted to prevent the overfitting problem that is easy to occur in small sample set data. When the loss of the validation set does not decrease within 10 training rounds, the training phase will be stopped in advance. The batch size is also a key parameter. According to the complexity of the marine natural earthquake identification task and the sample size, when optimizing the model parameters, the back-propagation gradient of the loss function is calculated on a data batch of size 32. This can not only "smooth" the disturbance of individual wrong label data to the model, but also improve the training efficiency of the model and prevent the model from falling into the local optimum. The accuracy and loss of the network change with the number of iterations on the four training-validation datasets. Figure 4 , Figure 5 , Figure 6 and Figure 7 The solid red line represents the accuracy or loss of the network on the training data, and the dotted black line represents the accuracy or loss of the network on the validation data. As the number of iterations (Epoch) increases, the accuracy of the network on the training data and validation data increases simultaneously, and the loss decreases simultaneously, which shows that the training process of the network is convergent and there is no overfitting problem.

[0058] The confusion matrix of the deep neural network obtained by four-fold cross validation on the validation data is shown in Figure 8 , the average value of the confusion matrix is ​​shown in Fig. 9. The network's recall rate (Recall) for positive samples is above 80%. The recall rate is also called sensitivity, which reflects the detection rate of positive samples. Its definition is R = TP / (TP+FN), which reflects how many positive samples have been successfully identified. Among them, TP represents true positive samples, and FN represents unrecognized positive samples. The precision (Precision) of the four models is above 80%. The definition of precision is P = TP / (TP+FP), which reflects how many data are truly positive samples among the data identified as positive samples, and FP represents false positive samples. Accuracy (Accuracy), Acc = (TP+TN) / (TP+TN+FP+FN), reflects the proportion of correctly identified samples. In this example, the accuracy rate reached more than 90%, where TN represents a true negative sample. The F1 index is a comprehensive reflection of the precision rate and the recall rate, F1 = 2*P*R / (P+R). In this example, F1 is above 80%. The above indicators show that the trained model is effective.

[0059] The data to be judged are input into four trained deep neural networks respectively to obtain four outputs. A voting mechanism is adopted and the voting results are output as the final judgment results. If the votes are 2:2, the output of the model with the highest accuracy during training shall prevail. When the label is [0], it is judged as background noise. When the output label is [1], it is judged as valid natural earthquake data. Finally, the purpose of automatically identifying natural earthquake signals is achieved.

[0060] Example 2

[0061] The present application also provides a small sample lightweight deep learning real-time identification system for natural earthquakes at sea, which is implemented based on the above method, and the system includes:

[0062] The earthquake recognition module is used to input the sound pressure signals recorded by the snorkeling marine seismograph into multiple trained lightweight deep learning networks to output whether a natural offshore earthquake has occurred; a voting mechanism is used for the output results of multiple networks, and the voting results are used as the final judgment results.

[0063] Training module, used to train lightweight deep learning networks.

[0064] The present application may also provide a computer device, comprising: at least one processor, a memory, at least one network interface and a user interface. The various components in the device are coupled together through a bus system. It is understood that the bus system is used to achieve connection and communication between these components. In addition to the data bus, the bus system also includes a power bus, a control bus and a status signal bus.

[0065] The user interface may include a display, a keyboard or a pointing device, such as a mouse, a trackball, a touch pad or a touch screen.

[0066] It is understood that the memory in the embodiments disclosed in the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRRAM). The memories described herein are intended to include, but are not limited to, these and any other suitable types of memories.

[0067] In some embodiments, the memory stores the following elements, executable modules or data structures, or a subset thereof, or an extended set thereof: an operating system and applications.

[0068] The operating system includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., which are used to implement various basic services and process hardware-based tasks. The application includes various application programs, such as a media player (Media Player), a browser (Browser), etc., which are used to implement various application services. The program for implementing the method of the embodiment of the present disclosure can be included in the application.

[0069] In the above embodiment, the processor may also call a program or instruction stored in the memory, specifically, a program or instruction stored in an application program, and is used to:

[0070] Execute the steps of the above method.

[0071] The above method can be applied to a processor or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in the processor or an instruction in the form of software. The above processor may be a general processor, a digital signal processor (Digital Signal Processor, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The above-disclosed methods, steps and logic block diagrams can be implemented or executed. The general processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the above-disclosed method can be directly embodied as a hardware decoding processor to execute, or the hardware and software modules in the decoding processor are combined to execute. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0072] It is understood that the embodiments described in the present application can be implemented by hardware, software, firmware, middleware, microcode or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in the present application or a combination thereof.

[0073] For software implementation, the technology of the present application can be implemented by executing the functional modules (such as procedures, functions, etc.) of the present application. The software code can be stored in a memory and executed by a processor. The memory can be implemented in the processor or outside the processor.

[0074] The present application may also provide a non-volatile storage medium for storing a computer program. When the computer program is executed by a processor, each step in the above method embodiment can be implemented.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present application and are not intended to limit it. Although the present application is described in detail with reference to the embodiments, a person skilled in the art should understand that any modification or equivalent replacement of the technical solution of the present application does not depart from the spirit and scope of the technical solution of the present application and should be included in the scope of the claims of the present application.

Claims

1. A small sample lightweight deep learning method for real-time identification of natural earthquakes at sea, comprising: The acoustic pressure signals recorded by the snorkeling marine seismometer are input into multiple trained lightweight deep learning networks to output whether a natural offshore earthquake has occurred. For the output results of multiple networks, a voting mechanism is adopted, and the voting results are used as the final recognition results; The training of the lightweight deep learning network includes: Preprocessing of acoustic pressure signals recorded by snorkeling marine seismographs; The data is randomly arranged and then divided into n non-overlapping equal parts. One of the data is selected as the validation set, and the remaining n-1 data are selected as the training set. The supervised learning method is used to perform n model training; The preprocessing includes: uniformly sampling the data to a fixed frequency and a set length, and performing a 0.05-3 Hz bandpass filtering process on the data; The lightweight deep learning network includes 1 convolution layer, 1 activation layer, 1 one-dimensional batch normalization layer, 1 adaptive average pooling layer, 1 flattening layer and 1 fully connected layer; Add batch normalization layers after each layer; During the training process of the lightweight deep learning network, a cross entropy loss function is selected to calculate the loss, a small batch stochastic gradient descent algorithm is used to update the model parameters, an Adam optimizer capable of adaptively adjusting the learning rate is used, and an early stopping strategy is used to prevent overfitting; The voting mechanism is to count the output results of multiple networks and adopt the principle of minority obeys majority as the final result. If the number of two results is the same, the output of the network with the highest accuracy during training shall prevail.

2. The method for real-time identification of natural earthquakes at sea by small sample and lightweight deep learning according to claim 1 is characterized in that: The lightweight deep learning network has a convolution kernel size of 33*1, a kernel stride of 1, and a channel number of 32; the input layer and the intermediate layer use linear rectification functions.

3. The method for real-time identification of natural earthquakes at sea by small sample and lightweight deep learning according to claim 1 is characterized in that: The set length is 400 seconds.

4. A small sample lightweight deep learning method for real-time identification of natural earthquake systems at sea, implemented based on any of the methods described in claims 1-3, characterized in that: The system comprises: An earthquake recognition module is used to input the sound pressure signals recorded by the snorkeling marine seismometer into multiple trained lightweight deep learning networks to output whether a natural offshore earthquake has occurred; a voting mechanism is used for the output results of multiple networks, and the voting results are used as the final judgment results; and Training module, used to train lightweight deep learning networks.

Images