A smart geological disaster early warning method and system for 5G high-speed rail
By running deep learning models and synthetic aperture radar technology on edge servers, the problems of insufficient computing power of edge devices and blind spots in template matching detection are solved, enabling efficient and comprehensive early warning of earthquakes and landslides.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING TICOM TECH
- Filing Date
- 2023-12-29
- Publication Date
- 2026-06-02
Smart Images
Figure CN117877229B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geological disaster early warning, specifically a smart geological disaster early warning method and system for 5G high-speed rail. Background Technology
[0002] With the development of 5G and IoT technologies, geological disaster monitoring equipment can enable intelligent monitoring and early warning of disaster processes. However, due to the limited computing power of edge devices, complex machine learning models cannot be run directly on the devices. To solve this problem, an edge-cloud architecture can be adopted, where edge computing servers are placed at the network edge. Edge devices transmit the collected data to the edge servers, which are responsible for running machine learning models for analysis and prediction, and then transmitting the results to the cloud. This architecture reduces the workload of edge devices. Existing edge-cloud architectures mainly focus on the server side, and research on the resource shortage problem of edge devices is insufficient. Further research is needed on how to achieve efficient edge-cloud collaboration under the condition of edge devices with limited computing power to support critical tasks such as geological disaster early warning.
[0003] With the development of earthquake monitoring networks, waveform similarity-based methods such as template matching have become crucial for earthquake detection. These methods significantly improve the sensitivity and quantity of earthquake detection by matching continuous waveform data with waveform templates of existing earthquake events. However, template matching-based detection can only detect known earthquakes similar to existing templates, and is ineffective for entirely new events in unknown regions. To achieve comprehensive earthquake monitoring without blind spots, it is necessary to develop novel, widely detectable earthquake phase recognition methods. In recent years, deep learning technology has made significant progress in fields such as computer vision, and its excellent pattern recognition capabilities have brought new opportunities for earthquake signal analysis. Compared to matching-based detection methods such as template matching, deep neural networks can reliably predict novel and unknown patterns by learning generalized feature representations. Therefore, using deep learning to achieve end-to-end modeling of earthquake waveforms can significantly improve the level of earthquake monitoring. Summary of the Invention
[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0005] A smart geological disaster early warning method for 5G high-speed rail includes the following steps:
[0006] S1: Data Acquisition: Multiple sensor monitoring devices are deployed along the high-speed railway line. The devices collect seismic waveform time-domain signals and upload them to the edge server.
[0007] S2: Data preprocessing: The edge server downloads the pre-trained model from the railway early warning and control center, preprocesses the data collected in real time by the sensing and monitoring equipment, inputs the pre-processed data into the pre-trained model, obtains the prediction results, and transmits them to the railway early warning and control center.
[0008] S3: Early Warning Judgment: The railway early warning control center uses synthetic aperture radar based on interferometry technology to obtain the displacement time series of the landslide area, determine the landslide distribution area, and make early warning judgments based on the received prediction results to control the high-speed rail operation status in real time.
[0009] The sensing and monitoring equipment, edge server, and railway early warning and control center communicate via 5G.
[0010] Furthermore, in step S1, the earthquake waveform time-domain signal includes a waveform time-domain signal with three orthogonal components: vertical component Z(t), east-west component E(t), and north-south component N(t).
[0011] Furthermore, in step S2, data preprocessing involves processing the three component signals: the vertical component Z(t), the east-west component E(t), and the north-south component N(t). The specific steps are as follows:
[0012] A1. Detrending processing to eliminate DC component;
[0013] A2, high-pass filter above 2Hz to filter out micro-vibration noise;
[0014] A3. Resampling, sampling frequency f s =100Hz;
[0015] A4. Obtain the seismic signal matrix. in, The component time-domain sequences processed by steps A1-A2 are normalized to obtain the normalized seismic signal matrix.
[0016] Furthermore, in step S2, the pre-trained model is generated based on the generalized phase detection method, and the specific steps are as follows:
[0017] B1. From the waveform archive of the China Earthquake Networks Center, extract earthquake records that occurred within a 100-kilometer radius and with magnitudes between -0.8 and 5.7 over the past M years. Select P-wave and S-wave waveforms, and extract a total of 1.5 million P-wave samples and 1.5 million S-wave samples. At the same time, randomly select 1.5 million background noise waveforms as negative samples and preprocess the selected waveforms.
[0018] B2. Use the sliding window technique to segment each preprocessed waveform signal. Set the length of the sliding window Δt, and extract a signal segment of length Δt from the beginning of the signal as the first window w(t1)={X(t)}. normalization Given the sequence {t1 ≤ t ≤ t1 + Δt}, a sliding window is used to extract the next signal segment according to a step size Δt, until the window reaches the end of the signal, resulting in the waveform window time series W(t) = {w(t1), w(t2), ..., w(t...}. k )}, where t1 is the starting time point of the time series, t k Let w(t) be the k-th time point in the time series, where k is a positive integer. k ) is at time t k to t k The data within +Δt, after being normalized in the same way as w(t1), are used to repeat step B2 to obtain the time series of each preprocessed signal waveform window.
[0019] B3 and W(t) are passed as input to the neural network model, and the model output is the predicted probability result of each window in W(t) [P]. 纵 P 横 P 噪 ], where P 纵 Let P be the probability value of belonging to the longitudinal wave. 横 Let P be the probability value of belonging to a transverse wave. 噪 Let P0 be the probability value that belongs to noise. If:
[0020] P 纵 >P0, mark the original waveform of the current window as a longitudinal wave;
[0021] P 横 >P0, mark the original waveform of the current window as a transverse wave;
[0022] P 纵 <P0 and P 横 <P0 represents noise and is not marked.
[0023] Repeat step c to evaluate the prediction results for all windows.
[0024] Furthermore, the neural network model is a one-dimensional convolutional neural network based on ConvNet. The model architecture includes four convolutional layers and two fully connected layers. Rectified Linear Units (ReLU) are used as the activation function in each layer. Each convolutional layer is followed by a pooling layer using max pooling to learn and extract temporal features from the signal. The first convolutional layer has an 11×1 kernel size, 24 kernels, a stride of 1, and VALID padding (filling within image boundaries). The first pooling layer uses max pooling with a 2×1 kernel size and a stride of 2. The second convolutional layer has a 5×1 kernel size, 48 kernels, a stride of 1, and VALID padding. The padding is set to VALID; the second pooling layer uses max pooling with a kernel size of 2×1 and a stride of 2; the third convolutional layer has a kernel size of 3×1, 48 kernels, a stride of 1, and padding set to VALID; the third pooling layer uses max pooling with a kernel size of 2×1 and a stride of 2; the fourth convolutional layer has a kernel size of 3×1, 64 kernels, a stride of 1, and padding set to VALID; the fourth pooling layer uses max pooling with a kernel size of 2×1 and a stride of 2; the flattening layer flattens the features into a one-dimensional vector; the first fully connected layer has 100 nodes and uses ReLU activation function; the second fully connected layer has 3 nodes and uses Softmax activation function.
[0025] The training objective is to minimize the cross-entropy loss function:
[0026] Among them, y ij P represents the true label (0 or 1) of sample i for class j. ij Let C represent the probability that sample i is predicted to be class j, C represent the number of classes, and G represent the total number of samples in the training set, which is 75% of the total number of samples in the dataset. When training the model, the Adam optimizer is used, and the learning rate is dynamically adjusted based on the cosine annealing strategy.
[0027] Furthermore, in step S3, the railway early warning control center makes an early warning judgment based on the received prediction results. Multiple sensor monitoring devices collect data in real time and upload it to the edge server. A pre-trained model is used for prediction. When an edge server detects a possible seismic waveform, i.e., P... 纵 >P0 or P 横 When P > 0, the prediction result, including the waveform occurrence time t, will be used. p The waveform type is transmitted to the railway early warning and control center. Based on the prediction results of all edge servers, the early warning and control center performs correlation analysis in the time and spatial domains. If more than two edge servers detect possible seismic waves within the time window Δt, it is determined that an earthquake has occurred.
[0028] Holographic correlation is performed based on the time difference of waveform occurrence at different stations. The optimal point source location, latitude and longitude, and occurrence time t of the seismic event are determined by the L1 norm minimization optimization method. r ;
[0029] Based on the relative position of the earthquake source and the railway and with reference to the theoretical propagation speed, the arrival time of the seismic waves is calculated. Combined with the landslide distribution area, the section requiring early warning is determined, trains running in the warning area are controlled, and early warning information is issued to remind drivers and passengers to take protective measures.
[0030] Furthermore, in step S3, the step of determining the landslide distribution area is as follows:
[0031] C1. Use synthetic aperture radar to acquire data of the same area at time phase h (h≥2), then perform interferometric processing to obtain the interferometric phase. in and These are the synthetic aperture radar image phases for the 0th and 1st time phases, respectively;
[0032] C2. According to the formula The displacement Δd of the surface in this region at time phase o is obtained. o Where λ is the wavelength of the synthetic aperture radar; the displacements of multiple time phases are superimposed to obtain the displacement time series D(t)=∑Δd o ;
[0033] C3. Using the Star-Based Augmentation System (SBAS), a displacement-time series with even better accuracy and reliability after correction is obtained. in, The rate of change of phase over time;
[0034] C4. If D'(t) > D0, where D0 is the displacement threshold, then the area is determined to be a landslide distribution area.
[0035] A smart geological disaster early warning system for 5G high-speed rail includes sensor detection equipment, edge servers, and a railway early warning and control center.
[0036] Multiple of the aforementioned sensing and detection devices are deployed along the high-speed railway line to collect seismic waveform time-domain signals and upload them to the edge server;
[0037] The edge server preprocesses the data collected in real time by the sensing and detection device, and inputs the preprocessed data into the pre-trained model to obtain the prediction result.
[0038] The railway early warning and control center uses synthetic aperture radar based on interferometry technology to acquire the displacement time series of landslide areas, determine the landslide distribution area, and make early warning judgments based on the received prediction results, so as to carry out real-time control of the high-speed rail operation status.
[0039] The sensing and monitoring equipment, edge server, and railway early warning and control center communicate via 5G.
[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0041] This invention involves deploying multiple sensing and detection devices along a high-speed railway line to collect seismic waveform time-domain signals and upload them to an edge server. The edge server preprocesses the data collected in real time by the sensing and detection devices, and the preprocessed data is input into a pre-trained model to obtain prediction results.
[0042] The pre-trained model is generated based on the generalized phase detection method, which can reliably detect P-waves and S-waves in earthquakes ranging from very small to very large. The sliding window technique is used to obtain the time series of each pre-processed signal waveform window, which is then input into the neural network model to achieve reliable prediction of novel and unknown patterns.
[0043] The railway early warning and control center uses synthetic aperture radar based on interferometry technology to acquire the displacement time series of landslide areas, determine the distribution area of landslides, and make early warning judgments based on the received prediction results, so as to carry out real-time control of the high-speed rail operation status.
[0044] This invention can reliably detect P-waves and S-waves of earthquakes ranging from very small to very large, without requiring specific waveform templates, thus achieving comprehensive geological monitoring and timely early warning without blind spots. Attached Figure Description
[0045] Figure 1 This is an architecture diagram of a smart geological disaster early warning system for 5G high-speed rail, as described in this invention.
[0046] Figure 2 A flowchart of an intelligent geological disaster early warning method for 5G high-speed rail, as described in the invention;
[0047] Figure 3 This is a diagram of the neural network architecture based on the generalized phase detection method in step S2 of the present invention;
[0048] Figure 4 A diagram showing the relationship between training loss and training rounds provided in an embodiment of the present invention;
[0049] Figure 5 A graph showing the relationship between test accuracy and training rounds provided in an embodiment of the present invention. Example:
[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0051] like Figure 2As shown, a smart geological disaster early warning method for 5G high-speed rail includes the following steps:
[0052] S1: Data Acquisition: Multiple sensor monitoring devices are deployed along the high-speed railway line. The devices collect seismic waveform time-domain signals and upload them to the edge server.
[0053] S2: Data preprocessing: The edge server downloads the pre-trained model from the railway early warning and control center, preprocesses the data collected in real time by the sensing and monitoring equipment, inputs the pre-processed data into the pre-trained model, obtains the prediction results, and transmits them to the railway early warning and control center.
[0054] S3: Early Warning Judgment: The railway early warning control center uses synthetic aperture radar based on interferometry technology to obtain the displacement time series of the landslide area, determine the landslide distribution area, and make early warning judgments based on the received prediction results to control the high-speed rail operation status in real time.
[0055] The sensing and monitoring equipment, edge servers, and railway early warning and control center communicate via 5G.
[0056] In step S1, the earthquake waveform time-domain signal contains three orthogonal components: vertical component Z(t), east-west component E(t), and north-south component N(t).
[0057] In step S2, data preprocessing involves processing the three component signals: the vertical component Z(t), the east-west component E(t), and the north-south component N(t). The specific steps are as follows:
[0058] A1. Detrending processing to eliminate DC component;
[0059] A2, high-pass filter above 2Hz to filter out micro-vibration noise;
[0060] A3. Resampling, sampling frequency f s =100Hz;
[0061] A4. Obtain the seismic signal matrix. in, The component time-domain sequences processed by steps A1-A2 are normalized to obtain the normalized seismic signal matrix.
[0062] In step S2, the pre-trained model is generated based on the generalized phase detection method. The specific steps are as follows:
[0063] B1. From the waveform archive of the China Earthquake Networks Center, extract earthquake records that occurred within a 100-kilometer radius and with magnitudes between -0.8 and 5.7 over the past M years. Select P-wave and S-wave waveforms, and extract a total of 1.5 million P-wave samples and 1.5 million S-wave samples. At the same time, randomly select 1.5 million background noise waveforms as negative samples and preprocess the selected waveforms.
[0064] B2. Use the sliding window technique to segment each preprocessed waveform signal. Set the length of the sliding window Δt, and extract a signal segment of length Δt from the beginning of the signal as the first window w(t1)={X(t)}. normalization Given the sequence {t1 ≤ t ≤ t1 + Δt}, a sliding window is used to extract the next signal segment according to a step size Δt, until the window reaches the end of the signal, resulting in the waveform window time series W(t) = {w(t1), w(t2), ..., w(t...}. k )}, where is the starting time point of the time series, is the th time point of the time series, is a positive integer, and is the data within the time range , after being normalized by the same method. Repeat step B2 to obtain the time series of each preprocessed signal waveform window.
[0065] B3 and W(t) are passed as input to the neural network model, and the model output is the predicted probability result of each window in W(t) [P]. 纵 P 横 P 噪 ], where P 纵 Let P be the probability value of belonging to the longitudinal wave. 横 Let P be the probability value of belonging to a transverse wave. 噪 Let P0 be the probability value that belongs to noise. If:
[0066] P 纵 >P0, mark the original waveform of the current window as a longitudinal wave;
[0067] P 横 >P0, mark the original waveform of the current window as a transverse wave;
[0068] P 纵 <P0 and P 横 <P0 represents noise and is not marked.
[0069] Repeat step c to evaluate the prediction results for all windows.
[0070] like Figure 3As shown, the neural network model is a one-dimensional convolutional neural network based on ConvNet. The model architecture includes four convolutional layers and two fully connected layers. Rectified Linear Units (ReLU) are used as the activation function in each layer. Each convolutional layer is followed by a pooling layer using max pooling to learn and extract temporal features from the signal. The first convolutional layer has an 11×1 kernel size, 24 kernels, a stride of 1, and VALID padding (filling within image boundaries). The first pooling layer uses max pooling with a 2×1 kernel size and a stride of 2. The second convolutional layer has a 5×1 kernel size, 48 kernels, a stride of 1, and VALID padding. The first fully connected layer uses max pooling with a 2×1 kernel and a stride of 2. The second fully connected layer uses max pooling with a 3×1 kernel, 48 kernels, a stride of 1, and padding set to VALID. The third fully connected layer uses max pooling with a 2×1 kernel and a stride of 2. The fourth fully connected layer uses max pooling with a 2×1 kernel and a stride of 2. The flattening layer flattens the features into a one-dimensional vector. The first fully connected layer has 100 nodes and uses ReLU activation. The second fully connected layer has 3 nodes and uses Softmax activation.
[0071] The training objective is to minimize the cross-entropy loss function:
[0072] Among them, y ij P represents the true label (0 or 1) of sample i for class j. ij Let C represent the probability that sample i is predicted to be class j, C represent the number of classes, and G represent the total number of samples in the training set, which is 75% of the total number of samples in the dataset. When training the model, the Adam optimizer is used, and the learning rate is dynamically adjusted based on the cosine annealing strategy.
[0073] In step S3, the railway early warning control center makes an early warning judgment based on the received prediction results. Multiple sensor monitoring devices collect data in real time and upload it to the edge server. Pre-trained models are used for prediction. When an edge server detects a possible seismic waveform, i.e., P... 纵 >P0 or P 横 When P > 0, the prediction result, including the waveform occurrence time t, will be used. p The waveform type is transmitted to the railway early warning and control center. Based on the prediction results of all edge servers, the early warning and control center performs correlation analysis in the time and spatial domains. If more than two edge servers detect possible seismic waves within the time window Δt, it is determined that an earthquake has occurred.
[0074] Holographic correlation is performed based on the time difference of waveform occurrence at different stations. The optimal point source location, latitude and longitude, and occurrence time t of the seismic event are determined by the L1 norm minimization optimization method. r ;
[0075] Based on the relative position of the earthquake source and the railway and with reference to the theoretical propagation speed, the arrival time of the seismic waves is calculated. Combined with the landslide distribution area, the section requiring early warning is determined, trains running in the warning area are controlled, and early warning information is issued to remind drivers and passengers to take protective measures.
[0076] In step S3, the steps for determining the landslide distribution area are as follows:
[0077] C1. Use synthetic aperture radar to acquire data of the same area at time phase h (h≥2), then perform interferometric processing to obtain the interferometric phase. in and These are the synthetic aperture radar image phases for the 0th and 1st time phases, respectively;
[0078] C2. According to the formula The displacement Δd of the surface in this region at time phase o is obtained. o Where λ is the wavelength of the synthetic aperture radar; the displacements of multiple time phases are superimposed to obtain the displacement time series D(t)=∑Δd o ;
[0079] C3. Using the Star-Based Augmentation System (SBAS), a displacement-time series with even better accuracy and reliability after correction is obtained. in, The rate of change of phase over time;
[0080] C4. If D'(t) > D0, where D0 is the displacement threshold, then the area is determined to be a landslide distribution area.
[0081] like Figure 1 As shown, a smart geological disaster early warning system for 5G high-speed rail includes sensor detection equipment, edge servers, and a railway early warning and control center.
[0082] Multiple sensing and detection devices are deployed along the high-speed railway line to collect seismic waveform time-domain signals and upload them to the edge server;
[0083] The edge server preprocesses the data collected in real time by the sensing and detection devices, and the preprocessed data is input into the pre-trained model to obtain the prediction result.
[0084] The railway early warning and control center uses synthetic aperture radar based on interferometry technology to acquire the displacement time series of landslide areas, determine the distribution area of landslides, and make early warning judgments based on the received prediction results, so as to carry out real-time control of the high-speed rail operation status.
[0085] The sensing and monitoring equipment, edge servers, and railway early warning and control center communicate via 5G.
[0086] The technical solution provided by this invention is further illustrated below through specific embodiments. The experimental environment is a Windows operating system, using the TensorFlow and Keras deep learning development frameworks, and Python as the development language. The CPU used in the experiment is an Intel Core i7-8550U, and the GPU is three NVIDIA GTX 1060s. During training, the batch size is set to 1000*3, the initial learning rate is set to 0.001, and the epoch is 480.
[0087] The algorithm runs in a Python 3.7 environment, as shown in the table below:
[0088]
[0089]
[0090] The necessary dependency libraries were installed in a virtual environment, and the seismic maps were randomly divided into a training set (75%) and a validation set (25%). The model was then trained using the cross-entropy loss function and the ADAM optimization algorithm, with mini-batches of 480 records each, using three NVIDIA GTX 1060 graphics processing units. The training process was programmed to terminate when the validation loss did not decrease over more than five epochs (a full iteration of the dataset), and the epoch with the lowest loss value on the validation set was selected as the best model. Finally, maps of the three component traces, plotted picking, and maps of the P-wave and S-wave probability flows were generated.
[0091] Figure 4 and Figure 5 The trends of training loss and test accuracy with respect to training epochs are presented. We can see that the generalized phase detection method proposed in this invention outperforms benchmark algorithms, demonstrating the effectiveness of our proposed scheme.
[0092] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A smart geological disaster early warning method for 5G high-speed rail, characterized in that, Includes the following steps: S1: Data Acquisition: Multiple sensor monitoring devices are deployed along the high-speed railway line. The devices collect seismic waveform time-domain signals and upload them to the edge server. S2: Data Preprocessing: The edge server downloads the pre-trained model from the railway early warning and control center, preprocesses the data collected in real time by the sensing and monitoring equipment, inputs the pre-processed data into the pre-trained model to obtain prediction results, and transmits them to the railway early warning and control center; the pre-trained model is generated based on the generalized phase detection method, and the specific steps are as follows: B1. From the waveform archive of the China Earthquake Networks Center, extract earthquake records that occurred within a 100-kilometer radius and with magnitudes between -0.8 and 5.7 over the past M years. Select P-wave and S-wave waveforms, and extract a total of 1.5 million P-wave samples and 1.5 million S-wave samples. At the same time, randomly select 1.5 million background noise waveforms as negative samples and preprocess the selected waveforms. B2. Use the sliding window technique to segment each preprocessed waveform signal, and set the length of the sliding window. Capture from the start position of the signal A signal segment of length, used as the first window According to step size The sliding window captures the next signal segment until it reaches the end of the signal, thus obtaining the waveform window time series. ,in, This is the starting time point of the time series. The first time series At a certain point in time, It is a positive integer. In time arrive The data within, after being processed by the same Repeat step B2 for each preprocessed signal waveform window time series, using the same normalized result. B3 The input is passed to the neural network model, and the model output is... Predicted probability results for each window ,in, This represents the probability value of belonging to the longitudinal wave. This represents the probability value of belonging to a transverse wave. Set the detection probability threshold to represent the probability value that belongs to noise. ,if: Mark the original waveform of the current window as a longitudinal wave; Mark the original waveform of the current window as a transverse wave; and This is noise and will not be marked. Repeat step B3 to evaluate the prediction results for all windows; S3: Early Warning Judgment: The railway early warning control center uses synthetic aperture radar based on interferometry technology to obtain the displacement time series of the landslide area, determine the landslide distribution area, and make early warning judgments based on the received prediction results to control the high-speed rail operation status in real time. The sensing and monitoring equipment, edge server, and railway early warning and control center communicate via 5G.
2. The intelligent geological disaster early warning method for 5G high-speed rail according to claim 1, characterized in that, In step S1, the seismic waveform time-domain signal includes a waveform time-domain signal with three orthogonal components: the vertical component... East-west component North-South Components .
3. The intelligent geological disaster early warning method for 5G high-speed rail according to claim 2, characterized in that, In step S2, data preprocessing is performed on the vertical component. East-west component North-South Components The three component signals are processed in the following steps: A1. Detrending processing to eliminate DC component; A2, high-pass filter above 2 Hz to filter out micro-vibration noise; A3. Resampling, sampling frequency =100 Hz; A4. Obtain the seismic signal matrix. ,in, The time-domain sequences of the components processed by A1-A2 are normalized to obtain the normalized seismic signal matrix. .
4. The intelligent geological disaster early warning method for 5G high-speed rail according to claim 1, characterized in that, The neural network model is a one-dimensional convolutional neural network based on ConvNet. The model architecture includes four convolutional layers and two fully connected layers. Rectified linear units (ReLU) are used as the activation function in each layer. Each convolutional layer is followed by a pooling layer using max pooling to learn and extract temporal features from the signal. The first convolutional layer has an 11×1 kernel size, 24 kernels, and a stride of 1. Padding is set to VALID, meaning it fills within the image boundaries; the first pooling layer uses max pooling with a kernel size of 2×1 and a stride of 2; the second convolutional layer has a kernel size of 5×1, 48 kernels, a stride of 1, and padding is set to VALID; the second pooling layer uses max pooling with a kernel size of 2×1 and a stride of 2; the third convolutional layer has a kernel size of 3×1, 48 kernels, a stride of 1, and padding is set to VALID; the third pooling layer uses max pooling with a kernel size of 2×1 and a stride of 2; the fourth convolutional layer has a kernel size of 3×1, 64 kernels, a stride of 1, and padding is set to VALID; the fourth pooling layer uses max pooling with a kernel size of 2×1 and a stride of 2; the flattening layer flattens the features into a one-dimensional vector; the first fully connected layer has 100 nodes and uses ReLU activation; the second fully connected layer has 3 nodes and uses Softmax activation. The training objective is to minimize the cross-entropy loss function: ,in, Indicates sample For category The true label, which is 0 or 1. Indicates sample Predicted as category The probability, Indicates the number of categories. This represents the total number of samples in the training set, which is 75% of the total number of samples in the dataset. When training the model, the Adam optimizer is used, and the learning rate is dynamically adjusted based on the cosine annealing strategy.
5. A smart geological disaster early warning method for 5G high-speed rail according to claim 1, characterized in that, In step S3, the railway early warning control center makes an early warning judgment based on the received prediction results. Multiple sensor monitoring devices collect data in real time and upload it to the edge server. Pre-trained models are used for prediction. When an edge server detects a possible seismic waveform, that is... or At that time, the prediction results, including the waveform occurrence time, will be used. The waveform type is transmitted to the railway early warning and control center. Based on the prediction results from all edge servers, the early warning and control center performs correlation analysis in the time and spatial domains. If the time window... If two or more edge servers detect possible seismic waves, an earthquake is determined to have occurred. Holographic correlation is performed based on the time difference of waveform occurrence at different stations. The optimal point source location, latitude and longitude, and occurrence time of the seismic event are determined by the L1 norm minimization optimization method. ; Based on the relative position of the earthquake source and the railway and with reference to the theoretical propagation speed, the arrival time of the seismic waves is calculated. Combined with the landslide distribution area, the section requiring early warning is determined, trains running in the warning area are controlled, and early warning information is issued to remind drivers and passengers to take protective measures.
6. A smart geological disaster early warning method for 5G high-speed rail according to any one of claims 1-5, characterized in that, In step S3, the step of determining the landslide distribution area is as follows: C1. Using synthetic aperture radar to survey the same area Phase acquisition, Then, interference processing is performed to obtain the interference phase. ,in and The first The phase of the synthetic aperture radar image of the first and second time phases; C2. According to the formula The result is the Displacement of the Earth's surface in this region at a given time phase ,in, The wavelength of the synthetic aperture radar is used; the displacements of multiple time phases are superimposed to obtain the displacement time series. ; C3. Using the Star-Based Augmentation System (SBAS), a displacement-time series with even better accuracy and reliability after correction is obtained. ,in, The rate of change of phase over time; C4. If , The displacement threshold is used to determine that the area is a landslide distribution area.
7. A smart geological disaster early warning system for 5G high-speed rail, characterized in that, The device for performing the method according to any one of claims 1-6 includes a sensing and detection device, an edge server, and a railway early warning and control center. Multiple of the aforementioned sensing and detection devices are deployed along the high-speed railway line to collect seismic waveform time-domain signals and upload them to the edge server; The edge server preprocesses the data collected in real time by the sensing and detection device, and inputs the preprocessed data into the pre-trained model to obtain the prediction result. The railway early warning and control center uses synthetic aperture radar based on interferometry technology to acquire the displacement time series of landslide areas, determine the landslide distribution area, and make early warning judgments based on the received prediction results, so as to carry out real-time control of the high-speed rail operation status. The sensing and detection equipment, edge server, and railway early warning and control center communicate via 5G.