A landslide signal identification method, device and electronic equipment

Abstract

The application discloses a landslide signal identification method and device and electronic equipment, and relates to the technical field of landslide signal identification. The method comprises the following steps: initializing a U-shaped network structure; replacing the convolution kernel of each convolution layer in the U-shaped network structure with a connection structure of a layer-by-layer convolution kernel and a point-by-point convolution kernel, and setting the layer-by-layer convolution kernel to one dimension to obtain a light-weight U-shaped network structure; obtaining seismic waveform data, laboratory rock fracture acoustic emission data and landslide waveform data, and performing mixed training on the light-weight U-shaped network structure by using the seismic waveform data, the laboratory rock fracture acoustic emission data and the landslide waveform data to obtain a landslide signal identification model; and identifying whether the to-be-detected signal comprises a landslide signal component by using the pre-trained landslide signal identification model. The technical scheme provided by the application provides a low-power-consumption and high-precision method for identifying landslide signals, which is suitable for low-power-consumption devices such as ARM and STM32.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of landslide disaster detection, and in particular to a landslide signal identification method, device and equipment. BACKGROUND

[0002] In seismology, accurate detection of landslide disasters is crucial. Accurate prediction of landslide signals and timely warning can effectively protect people's lives and property. Landslide signal detection methods include traditional methods and deep learning methods. Traditional methods include long-short window ratio, ARAIC algorithm, etc. Although such detection algorithms are fast, they have low accuracy in low signal-to-noise ratio conditions. Deep neural networks (DNN) can make full use of large amounts of manually annotated data to achieve high accuracy in seismic phase picking. At present, it has been widely applied. At present, there are many seismic phase picking algorithms based on convolutional neural networks. For example, the PhaseNet model can achieve a precision rate of 93.3% and a recall rate of 94.1% in P and S wave picking. The EQTransformer model used for seismic phase detection combines the Transformer model and the RNN model to achieve more excellent results. In addition, the EQTransformer model has achieved good results on most data sets.

[0003] However, although the deep learning method has high accuracy, it requires more computing resources in actual use, and generally needs a computing unit such as a graphics processing unit (GPU) with high power consumption (usually more than 100 W) to complete the calculation. On the seismic motion device, the power consumption is usually less than 5 W, and the computing performance is limited on such devices, so it is difficult to complete the deep learning calculation work, which limits the deep learning algorithm to complete the calculation work on the device, and an algorithm with low power consumption and high precision is needed to complete the work of landslide signal recognition. File CN114611674A provides a lightweight method of U-net model, which can be applied in low-computing edge mobile devices. This method sets a pre-pruning factor layer between every two convolution layers, and the pre-pruning factor layer includes multiple 1*1 convolution kernel kernels, and the number of input and output channels thereof is the same as that of the convolution layer connected thereto. Then the U-net model including the pre-pruning factor layer is trained for the first time using data, and after the training is completed, it is analyzed whether the parameters of each convolution kernel in the pre-pruning factor layer are greater than a preset threshold value. If the parameters of a certain convolution kernel are less than the preset threshold value, it indicates that the current channel is not important, and thus the current channel is pruned. A lightweight U-net model is obtained through this method, and the computational complexity is reduced. However, this method directly reduces the number of input and output channels of the U-net model, and the accuracy of the model completely depends on the U-shaped structure. If the number of feature channels is pruned too much, the training data cannot be supported for the landslide signal which is difficult to obtain, and the model obtained by training cannot guarantee the accuracy of landslide signal recognition. In addition, the U-net model is a fully convolutional neural network, and in general cases, the input and output are images, which is not suitable for direct application to landslide signals. SUMMARY

[0004] Therefore, the embodiments of the present application provide a landslide signal recognition method, device and equipment, thereby providing a low-power and high-precision algorithm suitable for ARM, STM32 and other low-power devices for landslide signal recognition.

[0005] According to a first aspect, the embodiments of the present application provide a landslide signal recognition method, which comprises: initializing a U-shaped network structure; replacing the convolution kernel of each convolution layer in the U-shaped network structure with a connection structure of layer-by-layer convolution kernel and point-by-point convolution kernel, and setting the layer-by-layer convolution kernel to one dimension to obtain a lightweight U-shaped network structure; obtaining seismic waveform data, laboratory rock fracture acoustic emission data and landslide waveform data, and using the seismic waveform data, the laboratory rock fracture acoustic emission data and the landslide waveform data to perform mixed training on the lightweight U-shaped network structure to obtain a landslide signal recognition model; and identifying whether a to-be-detected signal includes a landslide signal component through the pre-trained landslide signal recognition model.

[0006] Optionally, the initialization of the U-shaped network structure comprises: initializing the U-shaped network structure comprising 6 down-sampling layers and 4 up-sampling layers, and removing the skip connection between the corresponding down-sampling layer and the up-sampling layer in the U-shaped network structure; wherein the convolution step length of the first down-sampling layer to the second down-sampling layer is set to 1, the convolution step length of the third down-sampling layer to the sixth down-sampling layer is set to 4, and the deconvolution step length of the first up-sampling layer to the fourth up-sampling layer is set to 4; wherein the output channel number of the first down-sampling layer is defined as F, and the output channel number of the second down-sampling layer to the fourth up-sampling layer is F, 2F, 3F, 4F, 5F, 4F, 3F, 3F, and F in turn.

[0007] Optionally, the hybrid training of the lightweight U-shaped network structure using the seismic waveform data, the laboratory rock fracture acoustic emission data, and the landslide waveform data comprises: inputting the seismic waveform data and the laboratory rock fracture acoustic emission data into the lightweight U-shaped network structure for training to adjust the model parameters of each down-sampling layer and up-sampling layer in the lightweight U-shaped network structure; fixing the model parameters of other layers in the lightweight U-shaped network structure except the last up-sampling layer; inputting the landslide waveform data into the lightweight U-shaped network structure for training to adjust the model parameters of the last up-sampling layer of the lightweight U-shaped network structure.

[0008] Optionally, the method further comprises: converting the adjusted model parameters of each down-sampling layer and up-sampling layer into half-precision floating point or 8-bit format.

[0009] Optionally, replacing the convolution kernel of each convolution layer in the U-shaped network structure with a connection structure of layer-by-layer convolution kernel and point-by-point convolution kernel comprises: replacing the convolution kernel of each convolution layer in the U-shaped network structure according to the order structure of the first point-by-point convolution kernel connected to the layer-by-layer convolution kernel, and the layer-by-layer convolution kernel connected to the second point-by-point convolution kernel; the number of the first point-by-point convolution kernel is greater than the input feature number, which is used to expand the number of input features, the layer-by-layer convolution kernel is used to extract features from the input features to obtain extracted features, and the second point-by-point convolution kernel is used to fuse the extracted features and output features.

[0010] Optionally, the size of the layer-by-layer convolution kernel is 1*7.

[0011] Optionally, the P-wave arrival time and the S-wave arrival time in the seismic waveform data and the landslide waveform data, and the P-wave arrival time in the laboratory rock fracture acoustic emission data are labeled using a triangular function or a normal distribution function.

[0012] According to a second aspect, an embodiment of the present application provides a landslide signal identification device, the device comprising: a model initialization module configured to initialize a U-shaped network structure; a model lightening module configured to replace a convolution kernel of each convolution layer in the U-shaped network structure with a connection structure of a layer-by-layer convolution kernel and a point-by-point convolution kernel, and set the layer-by-layer convolution kernel to one dimension, to obtain a lightened U-shaped network structure; a model training module configured to obtain seismic waveform data, laboratory rock rupture acoustic emission data and landslide waveform data, and perform mixed training on the lightened U-shaped network structure by using the seismic waveform data, the laboratory rock rupture acoustic emission data and the landslide waveform data, to obtain a landslide signal identification model; and an identification module configured to identify whether a to-be-detected signal includes a landslide signal component by using the pre-trained landslide signal identification model.

[0013] According to a third aspect, an embodiment of the present application provides an electronic device, comprising a memory and a processor, which are in communication connection with each other, and the memory stores computer instructions, and the processor executes the computer instructions to perform the method in the first aspect or any one of the optional embodiments of the first aspect.

[0014] According to a fourth aspect, an embodiment of the present application provides a computer readable storage medium, which stores computer instructions, and the computer instructions are used to make a computer execute the method in the first aspect or any one of the optional embodiments of the first aspect.

[0015] The technical scheme provided in the present application has the following advantages:

[0016] The technical solution provided in this application offers a lightweight and high-precision landslide signal recognition model, which improves upon the traditional U-shaped network structure. Specifically, the convolutional kernels of each convolutional layer in the U-shaped network structure are replaced with a connection structure of layer-by-layer convolutional kernels and point-by-point convolutional kernels, significantly reducing the number of model parameters to be trained. Furthermore, setting the layer-by-layer convolutional kernels to 1-dimensional allows for further reduction of model parameters under the condition of further adapting to seismic waveform signals, resulting in a lightweight U-shaped network structure. Then, the lightweight U-shaped network structure is trained using acquired seismic waveform data, laboratory rock fracture acoustic emission data, and landslide waveform data. Considering that there are many samples available for natural earthquake signals and acoustic emission signals, and that they are similar to the micro-fracture signals of landslides, the trained model can be used for processing seismic motion data for landslide monitoring. Transfer learning is then performed on the model using landslide waveform data to improve the accuracy in landslide signal detection, resulting in a landslide signal recognition model with stronger generalization ability. Then, the landslide signal recognition model trained by the above training method is deployed into a low-power earthquake detection device, which can perform high-precision recognition of the signal to be detected under low power conditions and accurately determine whether the signal to be detected is a landslide signal. Attached Figure Description

[0017] The features and advantages of the invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the invention in any way. In the drawings:

[0018] Figure 1 The diagram illustrates the steps of a landslide signal identification method according to one embodiment of the present invention.

[0019] Figure 2 A schematic diagram of a U-shaped network structure according to one embodiment of the present invention is shown;

[0020] Figure 3 A schematic diagram of the landslide signal transfer learning process is shown in one embodiment of the present invention;

[0021] Figure 4 A schematic diagram of a convolutional structure according to one embodiment of the present invention is shown;

[0022] Figure 5 A schematic diagram of the structure of labeled data in one embodiment of the present invention is shown;

[0023] Figure 6 A schematic diagram of a landslide signal identification device according to one embodiment of the present invention is shown;

[0024] Figure 7 A schematic diagram of an electronic device according to one embodiment of the present invention is shown. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figure 1 In one embodiment, a landslide signal identification method specifically includes the following steps:

[0027] Step S1: Initialize the U-shaped network structure.

[0028] Step S2: Replace the convolution kernel of each convolutional layer in the U-shaped network structure with a connection structure of layer-by-layer convolution kernels and point-by-point convolution kernels, and set the layer-by-layer convolution kernels to 1 dimension to obtain a lightweight U-shaped network structure.

[0029] Specifically, this embodiment first initializes a U-shaped network structure, namely the U-net model, and then performs lightweight improvements on the U-shaped network structure. Considering the difficulty in obtaining landslide signals, training data samples are often insufficient. Directly pruning the input and output channels of convolutional layers is difficult to guarantee the training accuracy of the model. This embodiment improves the U-shaped network structure by replacing the ordinary convolutional kernels of each convolutional layer with a layer-by-layer convolutional kernel and a point-by-point convolutional kernel connection structure. An ordinary convolutional kernel is a filter, and each ordinary convolutional kernel includes multiple kernels. The number of ordinary convolutional kernels is equal to the number of output channels, and the number of kernels in each ordinary convolutional kernel is equal to the number of input channels. The convolution operation of an ordinary convolutional kernel involves convolving each kernel with the corresponding input feature, and then merging the convolution results of each kernel to obtain the output feature of the current ordinary convolutional kernel. Multiple ordinary convolutional kernels can output multiple features. This embodiment first processes the input features using layer-by-layer convolutional kernels. Each layer-by-layer convolutional kernel has only one kernel, and the number of kernels equals the number of input channels. The convolution operation of each kernel involves convolving with the corresponding input feature, outputting multiple intermediate features. These intermediate features are then processed separately using pointwise convolutional kernels. A pointwise convolutional kernel consists of multiple 1x1 kernels, the number of which equals the number of output channels. The number of 1x1 kernels within a pointwise convolutional kernel equals the number of input channels. The convolution operation of a pointwise convolutional kernel involves convolving each of its internal 1x1 kernels with the corresponding intermediate feature. The multiple convolved intermediate features are then combined into a single output feature. By using multiple pointwise convolutional kernels, multiple output features can be output. This approach reduces the number of model training parameters without changing the number of input and output channels. Furthermore, considering that the U-net model is a fully convolutional neural network, and its input and output are generally two-dimensional images, it is not suitable for direct application to landslide signals. Therefore, this embodiment sets the dimension of the layer-by-layer convolutional kernel to 1 dimension. On the one hand, this is more suitable for landslide signal data, enabling it to process three-dimensional seismic waveform data; on the other hand, it further reduces the number of training parameters. It should be noted that this embodiment only uses the above-mentioned connection structure of one layer-by-layer convolutional kernel plus one layer-by-point convolutional kernel as an example for explanation. As long as the number of input and output channels in each convolutional layer in the U-shaped network structure remains the same before and after the improvement, other combinations of different numbers of layer-by-layer convolutional kernels and point-by-point convolutional kernels are acceptable, and this embodiment does not impose any special limitations on this.

[0030] In one embodiment, assuming the training signal format is B×C²×T×1, representing the batch size, number of input channels, and feature length of each feature, respectively, and assuming the size of the ordinary convolutional kernel is C²×C×L×1, representing the number of input channels, number of kernels, and kernel length and width, respectively. Assuming the ordinary convolutional kernel size is 256×128×7 and the input image size is 32×128×112, then the number of trainable parameters is 256×128×7 = 229376. Furthermore, the number of floating-point operations is related to the input image size. Addition and multiplication are both floating-point operations (FLOPs). For a pixel in the output feature map, the number of floating-point multiplication operations required is K·C, and the number of floating-point addition operations required after multiplication is K·C-1. Therefore, outputting one feature point requires (2·K·C-1) operations. Each output feature point requires the same number of multiplications and additions. The total number of features is B×C²×T. Taking the assumed size as an example, the total number of floating-point operations is (2·K·C⁻¹)·B·²·T = 1,643,249,664 = 1.6 (GFLOPs). Taking a typical CPU as an example, assuming a core frequency of 3.0GHz, and one floating-point operation (multiplication or addition) is performed per clock cycle, then in a single-core scenario, a normal convolution calculation takes at least 0.54 seconds to complete one convolution calculation (not considering memory access issues and vector instruction set optimization).

[0031] In this improved embodiment, a filter is designed for each input feature map using layer-by-layer convolution, with a trainable number of parameters K·C = 896. Each filter only needs to process its corresponding feature map, requiring (2·K-1)·T·C·B = 5,963,776 = 0.006 GFLOPs for computation. After layer-by-layer convolution, pointwise convolution is added to change the number of channels in the feature map. The kernel size of pointwise convolution is 1×1, and the trainable number of parameters is C·C² = 32,768. Based on previous calculations, the computational complexity is (2·1·C-1)·B·C²·T = 233,963,520 = 0.234 GFLOPs. Therefore, the total computational complexity is 0.06 + 0.234 = 0.240 (FLOPs). This reduces the number of floating-point operations required by traditional convolution by more than 6 times, and completes one inference in only 0.08 seconds.

[0032] Meanwhile, the number of trainable parameters is 32768 + 896 = 33664, which is only 14.7% of the number of parameters in a typical convolutional model. This effectively reduces the number of trainable parameters and the number of operations, significantly improving computational efficiency for low-power, low-performance computing devices. On the one hand, the number of channels in the model remains unchanged; on the other hand, combined with the mixed data provided in subsequent steps for mixed training, research shows that this optimization does not cause a significant decrease in model accuracy.

[0033] Furthermore, considering that the optimized convolutional computational complexity is mainly concentrated in the pointwise convolution part, if the convolutional kernel size is changed from 3 to 5 (with other parameters remaining the same as above), the number of trainable parameters only increases by 8.9%, while the number of floating-point operations only increases by 6.1%. Compared with the improved method of this embodiment, if the convolutional kernel of the traditional convolutional method is changed from 3 to 5, the number of trainable parameters increases by 227.8%, while the number of floating-point operations increases by 178.4%. This makes traditional convolution unsuitable for designing excessively large kernels. However, the improved U-shaped network structure of this embodiment can avoid the problem of excessive increase in computational complexity when using large convolutional kernels. Therefore, setting a 1*7 layerwise convolutional kernel is optimal in this embodiment.

[0034] It should be noted that, in order to meet the needs of different accuracy scenarios, the lightweight U-shaped network structure provided in this embodiment can also be designed with different scales. That is, the number of network layers is the same for different scales, but the number of output features per layer is different. The output channels are pruned to cope with earthquake detection equipment with lower hardware configuration.

[0035] Step S3: Acquire seismic waveform data, laboratory rock fracture acoustic emission data, and landslide waveform data, and use the seismic waveform data, laboratory rock fracture acoustic emission data, and landslide waveform data to perform mixed training on the lightweight U-shaped network structure to obtain the landslide signal recognition model.

[0036] Step S4: Identify whether the signal to be detected includes a landslide signal component using a pre-trained landslide signal recognition model.

[0037] Specifically, for the aforementioned lightweight U-shaped network structure, this embodiment of the invention, in order to further compensate for the inaccuracy caused by the lightweight model, employs three types of data for mixed training: seismic waveform data, laboratory rock fracture acoustic emission data, and landslide waveform data. Since natural seismic signals and acoustic emission signals are abundant and easily obtained, and are similar to landslide micro-fracture signals, this embodiment trains the model using natural seismic signals (including P-waves and S-waves) and acoustic emission signals from rock fractures (mainly P-waves). After training, the model can be used to monitor landslide earthquakes. Subsequently, landslide waveform data (including P-waves and S-waves) is used for transfer learning of the model. This not only allows for multiple uses of a single model but also significantly improves the accuracy of the lightweight U-shaped network structure in landslide signal detection, thus obtaining a landslide signal recognition model. Specifically, the training data comes from manually labeled waveform data. The labeled data requires labeling the P-waves and S-waves in the landslide signals. The natural seismic waveform data comes from the China Earthquake Administration Data Center, and the laboratory rock fracture acoustic emission data comes from continuous acoustic emission data from rock friction fracture tests conducted by the inventor's laboratory.

[0038] After training a landslide signal recognition model that meets the requirements according to the above method, this disclosure provides a landslide signal recognition method. Specifically, the landslide signal recognition model pre-trained in the above steps is deployed in low-power seismic detection equipment such as ARM and STM32. The signal to be detected is input into the seismic detection equipment, and the system identifies whether the signal to be detected includes a landslide signal component and determines the arrival times of P-waves and S-waves. This achieves a low-power, high-precision landslide signal recognition method, improves recognition efficiency, and reduces hardware costs while ensuring recognition accuracy.

[0039] Specifically, in one embodiment, step S1 above includes the following steps:

[0040] Step 1: Initialize a U-shaped network structure consisting of 6 downsampling layers and 4 upsampling layers, and remove the skip connections between the corresponding downsampling and upsampling layers in the U-shaped network structure; wherein, the convolution stride from the first downsampling layer to the second downsampling layer is set to 1, the convolution stride from the third downsampling layer to the sixth downsampling layer is set to 4, and the deconvolution stride from the first upsampling layer to the fourth upsampling layer is set to 4; wherein, the number of output channels of the first downsampling layer is defined as F, then the number of output channels of the second downsampling layer to the fourth upsampling layer are F, 2F, 3F, 4F, 5F, 4F, 3F, 3F, F in sequence.

[0041] Specifically, such as Figure 2As shown, the U-shaped network structure provided in this embodiment of the invention includes 10 layers, specifically 6 downsampling layers and 4 upsampling layers. Considering that the skip connections in the U-shaped network structure are mainly used to fuse shallow and deep features to process data with different receptive fields, which is more effective for complex two-dimensional data, while one-dimensional landslide data is relatively simple and the influence of the receptive field is small, this embodiment of the invention removes the skip connections in the U-shaped network structure. Actual testing has verified that this not only has no impact on the accuracy of the output model but is actually higher, while also reducing the fusion process of skip connections, further lightweighting the U-shaped network structure.

[0042] In addition, the U-shaped network structure provided in the embodiments of the present invention is shown in Table 1.

[0043] Table 1. Schematic diagram of U-shaped network structure

[0044]

[0045] Layers 1 through 6 are downsampling layers used to extract deep features from landslide signal data. The convolution stride from the first to the second downsampling layer is set to 1. If the output channel of the first downsampling layer is defined as F, then the output channel of the second downsampling layer is also defined as F. The convolution stride from the third to the sixth downsampling layers is set to 4, and the output channels are defined as 2F, 3F, 4F, and 5F respectively compared to the first downsampling layer. Layers 7 through 10 are upsampling layers. The deconvolution stride from the first to the fourth upsampling layers is set to 4, interpolating the signal waveform by a factor of 4 to increase the feature map length. The notation 1 / 4 means that the length of the processed feature map becomes 4 times. Compared to the first downsampling layer, the number of output channels in the upsampling layers is 4F, 3F, 3F, and F respectively. This structure allows for the smooth decomposition of the original landslide signal and the analysis of the deep features of each part, up to the sixth downsampling layer. Then, the shorter, more channel-rich 5F features are gradually fused and amplified to recover the output features with the same length F as the input signal. The number of model layers and parameters is kept moderate, balancing accuracy and efficiency. Finally, a fully connected layer outputs three types of signals, representing the recovered original signal, the P-wave extracted from the original signal, and the S-wave extracted from the original signal. The value of F can be 4, 8, or 12, representing small, medium, and large models respectively, to accommodate seismic detection equipment with different hardware configurations.

[0046] Specifically, in one embodiment, step S3 above includes the following steps:

[0047] Step 2: Input the seismic waveform data and laboratory rock fracture acoustic emission data into the lightweight U-shaped network structure for training, in order to adjust the model parameters of each downsampling layer and upsampling layer in the lightweight U-shaped network structure.

[0048] Step 3: Fix the model parameters of all layers in the lightweight U-shaped network structure except for the last upsampling layer.

[0049] Step 4: Input the landslide waveform data into the lightweight U-shaped network structure for training, in order to adjust the model parameters of the last upsampling layer of the lightweight U-shaped network structure.

[0050] Specifically, in this embodiment of the invention, in order to further improve the generalization ability and accuracy of the landslide signal recognition model, a method is proposed to train a U-shaped network structure using seismic waveform data and laboratory rock fracture acoustic emission data, and then perform transfer training using landslide waveform data. For example... Figure 3 As shown, firstly, seismic waveform data and laboratory rock fracture acoustic emission data are input into a lightweight U-shaped network structure for training to adjust the model parameters of each downsampling and upsampling layer in the lightweight U-shaped network structure. This model is already ready for earthquake detection. Then, the model parameters of layers one through nine in the lightweight U-shaped network structure are fixed, and landslide waveform data is input into the lightweight U-shaped network structure for training, adjusting only the model parameters of the tenth layer. Combining the advantages of natural earthquake signals and laboratory rock fracture acoustic emission signals being similar to landslide micro-fracture signals and having a large number of readily available samples, this embodiment trains the model using natural earthquake signals (including P-waves and S-waves) and rock fracture acoustic emission signals (mainly P-waves), and then uses landslide waveform data (including P-waves and S-waves) for transfer learning and fine-tuning of the model. This not only allows for multiple uses of a single model but also significantly improves the accuracy of the lightweight U-shaped network structure in landslide signal detection.

[0051] Specifically, in one embodiment, the adjusted model parameters of each downsampling and upsampling layer are converted to half-precision floating-point or 8-bit format. Specifically, the trainable parameters in the model are converted to half-precision floating-point or 8-bit integer for calculation, further reducing the model size throughout the computation process and accelerating computation on low-power devices. The choice can also be made based on different precision requirements; half-precision floating-point can be used for calculations with higher device performance, while 8-bit integer can be used for calculations with lower device performance. Quantization and model size can be adjusted accordingly, improving the flexibility of model usage.

[0052] Specifically, in one embodiment, step S2 above includes the following steps:

[0053] Step 5: Replace the convolutional kernels of each convolutional layer in the U-shaped network structure according to the following sequence: first pointwise convolutional kernels are connected to layer-by-layer convolutional kernels, and layer-by-layer convolutional kernels are then connected to second pointwise convolutional kernels. The number of first pointwise convolutional kernels is greater than the number of input features, which is used to expand the number of input features. Layer-by-layer convolutional kernels are used to extract features from the input features. The second pointwise convolutional kernels are used to fuse the extracted features and output the features.

[0054] Specifically, the U-shaped network structure needs to ensure the amount of information at each layer of the network, but the addition of activation functions may cause information loss. If the ReLU activation function is used directly to transform between the input and output, it will result in incomplete information.

[0055] Therefore, the optimization approach of using layer-by-layer convolutional kernels and point-by-point convolutional kernels in the embodiments of the present invention specifically includes: such as Figure 4 As shown, a first-level pointwise convolutional kernel is used to enlarge and widen the feature map of the image, embedding it into a high-dimensional space (e.g., expanding the input features by 1.2 times, and the output channel being 1.2 times the input channel). Then, a layer-by-layer convolutional kernel is used to process the features in the high-dimensional space. Finally, a second-level pointwise convolutional kernel is used to compress the number of features. As mentioned earlier, the computational cost of layer-by-layer convolution is small, so the overall computational efficiency of the process can be guaranteed, and the problem of information loss is solved.

[0056] Specifically, in one embodiment, the arrival times of P-waves and S-waves in seismic and landslide waveform data, and the arrival time of P-waves in laboratory rock fracture acoustic emission data, are all labeled using trigonometric or normal distribution functions. Specifically, in practical applications, a certain degree of labeling and recognition error is usually permissible. To address this, this embodiment of the invention uses trigonometric functions or normal distribution functions to label the P-waves and S-waves in the training signal data. For example... Figure 5 As shown, taking trigonometric function annotation as an example, after the model training is completed, the signal to be detected is input into the landslide signal recognition model. If the signal to be detected includes landslide signal components, the output structure includes three types of signals. The first type has the same waveform as the input signal to be detected, representing the original signal. The second and third types both contain trigonometric functions. In the second and third types of signals, the position where the trigonometric function appears indicates the position of the arrival time of the longitudinal wave or transverse wave on the time axis. The highest vertex of the triangle indicates that the probability of the arrival time of the longitudinal wave and transverse wave at that position is the highest, and the lowest vertex positions on both sides of the triangle indicate that the probability of the arrival time of the longitudinal wave and transverse wave at that position is the lowest. Thus, this annotation method allows for a certain amount of annotation error and recognition error, and can also perform other data statistics based on the range of the triangle annotation.

[0057] Through the above steps, the technical solution provided in this application offers a lightweight and high-precision landslide signal recognition model, which improves upon the traditional U-shaped network structure. Specifically, the convolutional kernels of each convolutional layer in the U-shaped network structure are replaced with a connection structure of layer-by-layer convolutional kernels and point-by-point convolutional kernels, significantly reducing the number of model parameters to be trained. Furthermore, setting the layer-by-layer convolutional kernels to 1-dimensional further reduces the number of model parameters to be trained while adapting to seismic waveform signals, resulting in a lightweight U-shaped network structure. Then, the lightweight U-shaped network structure is trained using acquired seismic waveform data, laboratory rock fracture acoustic emission data, and landslide waveform data. Considering that there are many samples available for natural earthquake signals and acoustic emission signals, and that they are similar to the micro-fracture signals of landslides, the trained model can be used for processing seismic motion data for landslide monitoring. The landslide waveform data is then used for transfer learning of the model to improve its accuracy in landslide signal detection, resulting in a landslide signal recognition model with stronger generalization ability. Then, the landslide signal recognition model trained by the above training method is deployed into a low-power earthquake detection device, which can perform high-precision recognition of the signal to be detected under low power conditions and accurately determine whether the signal to be detected is a landslide signal.

[0058] like Figure 6 As shown, this embodiment also provides a landslide signal identification device, which includes:

[0059] Model initialization module 1 is used to initialize the U-shaped network structure. For details, please refer to the relevant description of step S1 in the above method embodiment, which will not be repeated here.

[0060] Model lightweighting module 2 is used to replace the convolutional kernels of each convolutional layer in the U-shaped network structure with a connection structure of layer-by-layer convolutional kernels and point-by-point convolutional kernels, and the layer-by-layer convolutional kernels are set to 1 dimension, resulting in a lightweight U-shaped network structure. For details, please refer to the relevant description of step S2 in the above method embodiment, which will not be repeated here.

[0061] Model training module 3 is used to acquire seismic waveform data, laboratory rock fracture acoustic emission data, and landslide waveform data. It then uses these data to perform hybrid training on a lightweight U-shaped network structure to obtain a landslide signal recognition model. For details, please refer to the relevant description of step S3 in the above method embodiments; it will not be repeated here.

[0062] The identification module 4 is used to identify whether the signal to be detected includes a landslide signal component using a pre-trained landslide signal identification model. For details, please refer to the relevant description of step S4 in the above method embodiment, which will not be repeated here.

[0063] The landslide signal identification device provided in this embodiment of the invention is used to execute the landslide signal identification method provided in the above embodiment. Its implementation method and principle are the same. For details, please refer to the relevant description of the above method embodiment, which will not be repeated here.

[0064] Through the synergistic collaboration of the aforementioned components, the technical solution provided in this application offers a lightweight and high-precision landslide signal recognition model, which improves upon the traditional U-shaped network structure. Specifically, the convolutional kernels of each convolutional layer in the U-shaped network structure are replaced with a connection structure of layer-by-layer convolutional kernels and point-by-point convolutional kernels, significantly reducing the number of model parameters to be trained. Furthermore, setting the layer-by-layer convolutional kernels to 1-dimensional further reduces the number of model parameters to be trained while adapting to seismic waveform signals, resulting in a lightweight U-shaped network structure. Then, the lightweight U-shaped network structure is trained using acquired seismic waveform data, laboratory rock fracture acoustic emission data, and landslide waveform data. Considering that there are many samples available for natural earthquake signals and acoustic emission signals, and that they are similar to the micro-fracture signals of landslides, the trained model can be used for processing seismic motion data for landslide monitoring. The landslide waveform data is then used for transfer learning of the model to improve its accuracy in landslide signal detection, resulting in a landslide signal recognition model with stronger generalization capabilities. Then, the landslide signal recognition model trained by the above training method is deployed into a low-power earthquake detection device, which can perform high-precision recognition of the signal to be detected under low power conditions and accurately determine whether the signal to be detected is a landslide signal.

[0065] Figure 7 An electronic device according to an embodiment of the present invention is shown. The device includes a processor 901 and a memory 902, which can be connected via a bus or other means. Figure 7 Taking the example of a connection between China and Israel via a bus.

[0066] Processor 901 can be a Central Processing Unit (CPU). Processor 901 can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0067] The memory 902, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the above method embodiments. The processor 901 executes various functional applications and data processing of the processor by running the non-transitory software programs, instructions, and modules stored in the memory 902, thereby implementing the methods in the above method embodiments.

[0068] The memory 902 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor 901, etc. Furthermore, the memory 902 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory 902 may optionally include memory remotely located relative to the processor 901, and these remote memories may be connected to the processor 901 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0069] One or more modules are stored in memory 902, and when executed by processor 901, they perform the methods described in the above method embodiments.

[0070] The specific details of the aforementioned electronic device can be understood by referring to the relevant descriptions and effects in the above method embodiments, and will not be repeated here.

[0071] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The implemented program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0072] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A landslide signal identification method, characterized in that, The method includes: Initialize the U-shaped network structure; The convolutional kernels of each convolutional layer in the U-shaped network structure are replaced with a connection structure of layer-by-layer convolutional kernels and point-by-point convolutional kernels, and the layer-by-layer convolutional kernels are set to 1 dimension to obtain a lightweight U-shaped network structure. Seismic waveform data, laboratory rock fracture acoustic emission data, and landslide waveform data are acquired, and a lightweight U-shaped network structure is trained using the seismic waveform data, the laboratory rock fracture acoustic emission data, and the landslide waveform data to obtain a landslide signal recognition model. The pre-trained landslide signal recognition model identifies whether the signal to be detected includes a landslide signal component. The initialization of the U-shaped network structure includes: initializing a U-shaped network structure comprising 6 downsampling layers and 4 upsampling layers, and removing the skip connections between corresponding downsampling and upsampling layers in the U-shaped network structure; wherein, the convolution stride from the first downsampling layer to the second downsampling layer is set to 1, the convolution stride from the third downsampling layer to the sixth downsampling layer is set to 4, and the deconvolution stride from the first upsampling layer to the fourth upsampling layer is set to 4; wherein, the number of output channels of the first downsampling layer is defined as F, then the number of output channels from the second downsampling layer to the fourth upsampling layer are F, 2F, 3F, 4F, 5F, 4F, 3F, 3F, F respectively; The step of replacing the convolutional kernels of each convolutional layer in the U-shaped network structure with a connection structure of layer-by-layer convolutional kernels and point-by-point convolutional kernels includes: replacing the convolutional kernels of each convolutional layer in the U-shaped network structure in the following order: connecting the first point-by-point convolutional kernel to the layer-by-layer convolutional kernels, and then connecting the layer-by-layer convolutional kernels to the second point-by-point convolutional kernels; the number of the first point-by-point convolutional kernels is greater than the number of input features, which is used to expand the number of input features; the layer-by-layer convolutional kernels are used to extract features from the input features to obtain extracted features; and the second point-by-point convolutional kernels are used to fuse the extracted features and output features.

2. The method according to claim 1, characterized in that, The method of using the seismic waveform data, the laboratory rock fracture acoustic emission data, and the landslide waveform data to perform hybrid training on the lightweight U-shaped network structure includes: The seismic waveform data and the acoustic emission data of the laboratory rock fracture are input into the lightweight U-shaped network structure for training, so as to adjust the model parameters of each downsampling layer and upsampling layer in the lightweight U-shaped network structure; The model parameters of all layers in the lightweight U-shaped network structure, except for the last upsampling layer, are fixed. The landslide waveform data is input into the lightweight U-shaped network structure for training, so as to adjust the model parameters of the last upsampling layer of the lightweight U-shaped network structure.

3. The method according to claim 2, characterized in that, The method further includes: Convert the adjusted model parameters of each downsampling and upsampling layer into half-precision floating-point or 8-bit format.

4. The method according to claim 1, characterized in that, The size of the layer-by-layer convolution kernel is 1*7.

5. The method according to claim 1, characterized in that, The arrival times of P-waves and S-waves in the earthquake waveform data and the landslide waveform data, as well as the arrival times of P-waves in the acoustic emission data of the laboratory rock fracture, are labeled using trigonometric functions or normal distribution functions.

6. A landslide signal identification device, characterized in that, The device includes: The model initialization module is used to initialize the U-shaped network structure. Initializing the U-shaped network structure includes: initializing a U-shaped network structure comprising 6 downsampling layers and 4 upsampling layers, and removing skip connections between corresponding downsampling and upsampling layers in the U-shaped network structure; wherein the convolution stride from the first downsampling layer to the second downsampling layer is set to 1, the convolution stride from the third downsampling layer to the sixth downsampling layer is set to 4, and the deconvolution stride from the first upsampling layer to the fourth upsampling layer is set to 4; wherein, the number of output channels of the first downsampling layer is defined as F, then the number of output channels from the second downsampling layer to the fourth upsampling layer are F, 2F, 3F, 4F, 5F, 4F, 3F, 3F, F, respectively. A lightweight model module is used to replace the convolutional kernels of each convolutional layer in the U-shaped network structure with a connection structure of layer-by-layer convolutional kernels and point-by-point convolutional kernels, wherein the layer-by-layer convolutional kernels are set to 1 dimension, resulting in a lightweight U-shaped network structure. Replacing the convolutional kernels of each convolutional layer in the U-shaped network structure with a connection structure of layer-by-layer convolutional kernels and point-by-point convolutional kernels includes: replacing the convolutional kernels of each convolutional layer in the U-shaped network structure with a structure in which a first point-by-point convolutional kernel connects to a layer-by-layer convolutional kernel, and then the layer-by-layer convolutional kernels are connected to a second point-by-point convolutional kernel; the number of the first point-by-point convolutional kernels is greater than the number of input features, used to expand the number of input features; the layer-by-layer convolutional kernels are used to extract features from the input features to obtain extracted features; and the second point-by-point convolutional kernels are used to fuse the extracted features and output features. The model training module is used to acquire seismic waveform data, laboratory rock fracture acoustic emission data and landslide waveform data, and to use the seismic waveform data, the laboratory rock fracture acoustic emission data and the landslide waveform data to perform mixed training on a lightweight U-shaped network structure to obtain a landslide signal recognition model; The identification module is used to identify whether the signal to be detected includes a landslide signal component through the pre-trained landslide signal identification model.

7. An electronic device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method as described in any one of claims 1-5.

Images