Ice rock collapse potential dangerous area identification device in alpine region

Abstract

The invention relates to the field of geological disaster monitoring, and discloses an alpine region ice rock collapse potential dangerous area identification device comprising an acquisition and preprocessing module used for acquiring and preprocessing alpine region ice rock multi-source data to form basic data; the fusion extraction module is used for performing data fusion processing and feature extraction on the basic data to form fused feature data; the identification and prediction module is used for identifying a potential risk area of the ice rock in the alpine region and predicting an ice rock collapse scene according to the fused feature data to form a risk area judgment result and ice rock collapse simulation and prediction information; and an optimization layout module. Basic data are formed by collecting and preprocessing ice rock multi-source data in the alpine region, fused feature data are formed by conducting data fusion processing and feature extraction on the basic data, and therefore spatial features of satellite images, time sequence features of sensor monitoring data, environment features of meteorological data and the like are organically combined.

Description

Technical Field

[0001] The invention relates to the technical field of geological disaster monitoring, and in particular to a device for identifying potential dangerous areas of ice and rock avalanches in high-cold areas. Background Art

[0002] In the earth's high-altitude and cold regions, such as alpine glacier zones and polar regions, due to the influence of special natural conditions such as low temperature, freeze-thaw cycles, and glacier movement, the stability of ice-rock assemblages is poor. Ice and rock avalanches in these areas occur frequently, posing a huge threat to the surrounding ecological environment, infrastructure, and human activities. With global climate change, environmental changes in high-altitude and cold regions have become more complex and drastic, and the frequency and severity of ice and rock avalanches have a tendency to further increase. This makes the monitoring and early warning of ice and rock avalanches in high-altitude and cold regions an important research direction in the field of geological disaster prevention and control.

[0003] Traditional devices for identifying potential dangerous areas of ice and rock avalanches in high-altitude and cold regions rely on strain sensors or displacement sensors arranged on site during monitoring. They judge the stability of the ice and rock bodies by monitoring the local strain or displacement changes. This method can only obtain limited local information and is difficult to fully reflect the ice and rock avalanche risk status of the entire area. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a device for identifying potential dangerous areas of ice and rock avalanches in high-altitude and cold regions, which solves the problem that traditional devices for identifying potential dangerous areas of ice and rock avalanches in high-altitude and cold regions rely on strain sensors or displacement sensors arranged on site during monitoring, and judge the stability of the ice and rock bodies by monitoring the local strain or displacement changes. This method can only obtain limited local information and is difficult to fully reflect the ice and rock avalanche risk status of the entire area.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a device for identifying potential dangerous areas of ice and rock avalanches in high-cold areas, comprising:

[0006] Collection and preprocessing module: used to collect and preprocess multi-source data of ice and rock in alpine areas to form basic data;

[0007] Fusion extraction module: used to perform data fusion processing and feature extraction on basic data to form fusion feature data;

[0008] Identification and prediction module: used to identify potential dangerous areas of ice and rock in high-cold areas and predict ice and rock collapse scenarios based on fused feature data, forming dangerous area determination results and ice and rock collapse simulation prediction information;

[0009] Layout optimization module: used to optimize the monitoring equipment layout of the acquisition preprocessing module based on multi-source data and monitoring equipment status information;

[0010] Early warning interaction module: used to provide early warning prompts and visual displays based on the results of dangerous area determination, ice and rock collapse simulation prediction information, and monitoring equipment status information.

[0011] Preferably, the acquisition and preprocessing module includes a sensor network unit, a satellite and aerial remote sensing unit, a meteorological data unit, a preprocessing unit, and a communication unit. The sensor network unit is used to collect corresponding data in high-cold areas through sensors. The sensors include a magnetoresistive three-axis magnetometer and a temperature and humidity sensor. The corresponding data include data on the intensity and direction changes of the magnetic field inside the mountain, and real-time data on ambient temperature and humidity. The satellite and aerial remote sensing unit is used to regularly obtain satellite images and three-dimensional point cloud data in high-cold areas. The satellite images include images of terrain, glaciers, and rock textures. The meteorological data access unit is used to obtain local meteorological data in high-cold areas. The meteorological data include temperature, precipitation, wind speed and direction, air pressure, and solar radiation intensity data and integrate history. The preprocessing unit is used to perform data cleaning, normalization and standardization, as well as preliminary construction operations of feature engineering to form basic data. The communication unit is used to provide data transmission and communication channels between modules.

[0012] Preferably, the preprocessing unit includes a data cleaning subunit, a normalization subunit, and a feature engineering construction subunit. The data cleaning subunit uses a robust statistical method based on Hampel identifiers to identify sensor time series outliers and repairs them through cubic spline interpolation, uses adaptive median filtering to remove noise from images, fills in missing time series data through seasonal decomposition-trend extrapolation, and repairs missing image pixels based on a regional growing algorithm. The normalization subunit is used to use Min-Max normalization to map data to the [0,1] interval and Z-score standardization to reshape the data normal distribution for different data, so that each type of data has a unified dimension and a suitable data distribution form. The feature engineering construction subunit is used to derive features from the original data, the features include calculating the comprehensive value of magnetic field intensity and direction cosine features from the three-axis data of the magnetometer, and constructing meteorological data features, the meteorological data features include daily temperature difference, precipitation intensity, and wind speed and direction change rate.

[0013] Preferably, the fusion extraction module includes a basic data processing unit, a fusion and bidirectional encoding unit, an attention mechanism unit, and a pre-training and fine-tuning unit. The basic data processing unit is used to construct a 12-layer DenseNet variant network for satellite and aerial images. The first layer uses 7×7 convolution kernels, 64 kernels, a step size of 2 and a padding of 3 for convolution operation followed by a maximum pooling layer. The middle layer strengthens feature transfer through dense connection blocks. The last layer outputs an image space feature map through global average pooling and full connection. The local time series features of the sensor monitoring data are extracted through a one-dimensional convolutional network, and the effective spatial and temporal feature information in the image and monitoring data are extracted respectively to form feature information. The fusion and bidirectional encoding unit is used to splice the feature information in the feature dimension through a bidirectional long short-term memory network and input it according to the time step, and fuse the bidirectional hidden state to output a comprehensive feature vector.

[0014] Preferably, the attention mechanism unit is used to access the multi-head attention mechanism after the output of the bidirectional long short-term memory network, determine the weight of each feature based on complex calculations, focus on the key precursor features of ice and rock avalanches, and highlight important information. The pre-training and fine-tuning unit is used to pre-train the model backbone network based on public geological disaster images and meteorological environment monitoring data sets with multi-type annotations, freeze the parameters of the bottom 8 convolutional layers and then load the local ice and rock avalanche related data, use the AdamW optimizer in combination with a small learning rate and the learning rate is exponentially decayed according to the training rounds for fine-tuning, use the focus loss function FL(p_t)=-\alpha_t(1-p_t)^{\gamma}\log(p_t), optimize the model to make it adapt to the local ice and rock avalanche monitoring task, and output fused feature data. The multi-type annotations include ice and rock avalanches, debris flows, and landslides.

[0015] Preferably, the identification and prediction module includes a simulation and prediction unit and an enhanced identification unit. The simulation and prediction unit is used to simulate the occurrence and development process of ice and rock avalanches and give prediction results through adversarial training between discriminators and generators based on cross entropy loss, and at the same time assist risk decision-making based on the uncertainty of simulation results through Monte Carlo sampling statistical analysis. The enhanced identification unit is used to select CatBoost, Random Forest and 4-layer multilayer perceptron (MLP) as base classifiers, and use fused feature data and derived high-order features as input to train classification models for distinguishing whether a mountain is in potential danger of ice and rock avalanches and dividing danger levels. The soft voting method is used for integration to comprehensively output dangerous area determination results and ice and rock avalanche simulation prediction information.

[0016] Preferably, the generator is a 6-layer transposed convolutional neural network, which inputs a 128-dimensional random noise vector z and an environmental feature vector e encoded by a multi-layer perceptron, the first layer has a transposed convolution kernel of 4×4, a step size of 2, and an output channel of 1024. After batch normalization and ReLU activation, the samples are cascaded layer by layer, and the last layer is activated by Tanh to output a simulated ice and rock collapse evolution image sequence and a corresponding physical quantity field. The discriminator adopts a 5-layer convolutional neural network, the first layer has a convolution kernel of 4×4, a step size of 2, 32 channels, and LeakyReLU is activated. After multiple convolutions and activations, the last layer is fully connected to output the discrimination probability D(x).

[0017] Preferably, the optimization layout module includes an environmental state unit, a reward function unit, and a reinforcement learning unit. The environmental state unit is used to construct a three-dimensional digital twin model of a high-cold mountain and discretize it into grid units. The state vector of each unit includes terrain, rock, meteorology, historical ice and rock collapse frequency, surrounding vegetation coverage, and soil moisture information. The monitoring equipment action vector includes three-dimensional coordinates, coverage radius, sampling frequency, transmission power, and monitoring angle parameters, and the environmental state and equipment action space are confirmed by this. The reward function unit is used to guide the optimization direction of the monitoring equipment layout of the acquisition preprocessing module through the reward function. The reward function is R=w1C υ -w2C cost -w3R penalty , where C υ is the effective monitoring coverage rate of ice and rock collapse precursors, C cost is the comprehensive cost of equipment deployment, R penalty The penalty item targets device signal blocking, redundant deployment, and frequent failures. The reinforcement learning unit is used to optimize the monitoring device layout strategy of the acquisition preprocessing module based on reward feedback through the proximal strategy optimization algorithm, interactive iterative training of the strategy network and the value network.

[0018] Preferably, the early warning interaction module includes an early warning unit and an interaction unit. The early warning unit is used to perform graded early warning according to a preset danger level threshold value compared with the risk probability output by the identification prediction module. When the risk probability is greater than the threshold, early warning information is pushed through multiple channels. The multiple channels include SMS gateways, sound and light alarms, and mobile phone APPs. The early warning information includes precise geographic coordinates of dangerous areas, estimated scale of ice and rock avalanches, and expected time range of occurrence. The interaction unit is used to display real-time monitoring curves of sensors, heat maps of ice and rock avalanche danger areas, and animations of predicted ice and rock avalanche movement trajectories based on WebGIS combined with Cesium three-dimensional earth engine and Echarts visualization library, and provides mouse wheel zooming, translation, rotation of viewing angle, historical data tracing back, and simulation of protection plans, so as to intuitively present monitoring and prediction situations to assist decision-making.

[0019] The method for identifying potential dangerous areas of ice and rock avalanches in high-altitude cold regions includes the following steps:

[0020] S1. Collection and preprocessing: Collect and preprocess multi-source data of ice and rock in high-cold areas to form basic data;

[0021] S2, fusion extraction: perform data fusion processing and feature extraction on the basic data to form fusion feature data;

[0022] S3, Identification and prediction: Identify potential dangerous areas of ice and rock in high-cold areas and predict ice and rock collapse scenarios based on fused feature data, and form dangerous area determination results and ice and rock collapse simulation prediction information;

[0023] S4. Optimize layout: optimize the layout of monitoring equipment based on multi-source data and monitoring equipment status information;

[0024] S5. Early warning interaction: Provide early warning prompts and visual displays based on the results of dangerous area determination, ice and rock collapse simulation prediction information, and monitoring equipment status information.

[0025] The present invention provides a device for identifying potential dangerous areas of ice and rock avalanches in high-cold areas. It has the following beneficial effects:

[0026] 1. The present invention collects multi-source data of ice and rock in high-cold areas and performs preprocessing to form basic data, and forms fused feature data by performing data fusion processing and feature extraction on the basic data, thereby organically combining the spatial characteristics of satellite images, the time series characteristics of sensor monitoring data, and the environmental characteristics of meteorological data, and mining the deep-level correlations and potential patterns between the data, thereby realizing the extraction of fused features with key indicative significance for ice and rock collapse judgment from massive raw data.

[0027] 2. The present invention can accurately identify potential dangerous areas of ice and rock in high-altitude cold regions and accurately predict ice and rock collapse scenarios by integrating multi-source data and conducting in-depth analysis and processing. It obtains rich and comprehensive information through the coordinated collection of high-precision sensors, satellite and aerial remote sensing, and meteorological data. The deep learning model integrates feature extraction and GAN simulation prediction, which overcomes the limitations of traditional single data source monitoring. It can provide early insights into subtle signs of ice and rock collapse, provide accurate basis for disaster prevention, effectively reduce the uncertainty risk of disasters, and ensure the safety of life and property and the stable operation of infrastructure.

[0028] 3. The present invention can dynamically adjust the layout of monitoring equipment according to multi-source data and monitoring equipment status information through the optimization layout module based on reinforcement learning. By constructing a detailed environmental model and designing a reasonable reward function, the PPO algorithm is used to enable the monitoring equipment to adapt to environmental changes, accurately focus on high-risk areas, avoid wasting resources in low-risk or invalid monitoring areas, improve monitoring efficiency and coverage, and solve the problem that traditional fixed monitoring layouts are difficult to cope with complex and changeable high-cold environments and dynamic changes in ice and rock collapse risks, ensuring the optimal monitoring effect with limited resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a system architecture diagram of the device for identifying potential dangerous areas of ice and rock avalanches in high-cold areas proposed by the present invention;

[0030] Figure 2 This is a diagram of the architecture of the acquisition preprocessing module of the device for identifying potential dangerous areas of ice and rock avalanches in high-cold regions proposed by the present invention;

[0031] Figure 3 This is a diagram of the preprocessing unit architecture of the device for identifying potential dangerous areas of ice and rock avalanches in high-cold regions proposed by the present invention;

[0032] Figure 4 This is a fusion extraction module architecture diagram of the device for identifying potential dangerous areas of ice and rock avalanches in high-cold areas proposed by the present invention;

[0033] Figure 5 This is a diagram of the recognition and prediction module architecture of the device for identifying potential dangerous areas of ice and rock avalanches in high-cold regions proposed by the present invention;

[0034] Figure 6 This is a diagram of the optimized layout module architecture of the device for identifying potential dangerous areas of ice and rock avalanches in high-cold regions proposed by the present invention;

[0035] Figure 7 This is a diagram of the early warning interaction module architecture of the device for identifying potential dangerous areas of ice and rock avalanches in high-cold regions proposed by the present invention;

[0036] Figure 8 This is a flow chart of the method for identifying potential dangerous areas of ice and rock avalanches in high-cold regions proposed by the present invention. DETAILED DESCRIPTION

[0037] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] Example:

[0039] Please see attached Figure 1 The embodiment of the present invention provides a device for identifying potential dangerous areas of ice and rock avalanches in high-cold areas, comprising:

[0040] Collection and preprocessing module: used to collect and preprocess multi-source data of ice and rock in alpine areas to form basic data;

[0041] Fusion extraction module: used to perform data fusion processing and feature extraction on basic data to form fusion feature data;

[0042] Identification and prediction module: used to identify potential dangerous areas of ice and rock in high-cold areas and predict ice and rock collapse scenarios based on fused feature data, forming dangerous area determination results and ice and rock collapse simulation prediction information;

[0043] Layout optimization module: used to optimize the monitoring equipment layout of the acquisition preprocessing module based on multi-source data and monitoring equipment status information;

[0044] Early warning interaction module: used to provide early warning prompts and visual displays based on the results of dangerous area determination, ice and rock collapse simulation prediction information, and monitoring equipment status information.

[0045] Specifically, basic data are formed by collecting multi-source data on ice and rocks in high-altitude cold areas and preprocessing them, thereby integrating data resources from different channels and types, including data on changes in the intensity and direction of the internal magnetic field of the mountain obtained by the sensor network, real-time data on environmental temperature and humidity, terrain, glacier, rock texture images and three-dimensional point cloud data obtained by satellite and aerial remote sensing, and meteorological data such as temperature, precipitation, wind speed and direction collected by meteorological data units and integrating historical data. Through the preliminary sorting and purification of these multi-source data, a comprehensive and rich basic data system has been built, providing sufficient data materials for subsequent precise analysis.

[0046] By performing data fusion processing and feature extraction on the basic data, fused feature data is formed, thereby organically combining the spatial characteristics of satellite images, the time series characteristics of sensor monitoring data, and the environmental characteristics of meteorological data, and digging out the deep-level connections and potential patterns between the data. For example, through a specific deep learning network architecture, the effective spatial and temporal feature information in the images and monitoring data are extracted respectively, and these feature information are fused using a bidirectional long short-term memory network, and then the key features are focused on with the help of the attention mechanism, thereby realizing the extraction of fused features with key indicative significance for ice and rock avalanche judgment from massive raw data, and solving the problem of traditional identification devices for potential dangerous areas of ice and rock avalanches in high-altitude cold regions. When monitoring, they rely on strain sensors or displacement sensors arranged on site, and judge the stability of the local strain or displacement changes of the ice and rock body. This method can only obtain limited local information and it is difficult to fully reflect the risk status of ice and rock avalanches in the entire area.

[0047] By identifying potential dangerous areas of ice and rock in high-altitude and cold regions and predicting ice and rock avalanche scenarios based on fused feature data, we can form dangerous area determination results and ice and rock avalanche simulation prediction information, and then use advanced machine learning algorithms to simulate the occurrence and development process of ice and rock avalanches and give prediction results, while considering uncertainty factors to assist risk decision-making; the integrated learning classification model combines the advantages of multiple base classifiers to accurately judge whether the mountain is in potential danger of ice and rock avalanche and the level of danger, thereby achieving precise positioning of potential dangers of ice and rock in high-altitude and cold regions and detailed simulation and prediction of ice and rock avalanche scenarios, improving the accuracy and reliability of hazard identification and prediction.

[0048] By optimizing the monitoring equipment layout of the acquisition preprocessing module based on multi-source data and monitoring equipment status information, a three-dimensional digital twin model of high-cold mountains is constructed, and each unit state vector (covering terrain, rocks, meteorology, historical ice and rock collapse frequency, surrounding vegetation coverage, soil moisture information) and monitoring equipment action vector (including three-dimensional coordinates, coverage radius, sampling frequency, transmission power, monitoring angle parameters) are determined. Based on this, a reasonable reward function is designed to guide the layout optimization direction, and a proximal strategy optimization algorithm is used to dynamically adjust according to reward feedback. The intelligent and dynamic layout of the monitoring equipment is realized, so that it can be flexibly adjusted according to the real-time changes in ice and rock collapse risks and changes in environmental conditions.

[0049] Early warning prompts and visual displays are provided based on the results of dangerous area determination, ice and rock avalanche simulation prediction information, and monitoring equipment status information. The risk probability is compared based on the preset danger level threshold. When the threshold is exceeded, detailed early warning information including precise geographic coordinates of dangerous areas, estimated ice and rock avalanche scale, and expected time range of occurrence is pushed in a timely manner through multiple channels such as SMS gateways, sound and light alarms, and mobile phone APPs. At the same time, based on WebGIS combined with Cesium 3D Earth Engine and Echarts visualization library, the real-time monitoring curve of sensors, heat maps of ice and rock avalanche dangerous areas, and animations of predicted ice and rock avalanche movement trajectories are intuitively displayed. Interactive functions such as mouse wheel zooming, panning, rotating perspective, tracing back historical data, and simulating protection plans are provided, thereby achieving efficient early warning communication and intuitive decision-making assistance.

[0050] Please see attached Figure 2 , Attachment Figure 3The acquisition and preprocessing module includes a sensor network unit, a satellite and aerial remote sensing unit, a meteorological data unit, a preprocessing unit, and a communication unit. The sensor network unit is used to collect corresponding data through sensors in high-cold areas. The sensors include a magnetoresistive three-axis magnetometer and a temperature and humidity sensor. The corresponding data include the internal magnetic field strength and direction change data of the mountain, and the real-time data of environmental temperature and humidity. The satellite and aerial remote sensing unit is used to regularly obtain satellite images and three-dimensional point cloud data in high-cold areas. Satellite images include images of terrain, glaciers, and rock textures. The meteorological data access unit is used to obtain local meteorological data in high-cold areas. Meteorological data includes temperature, precipitation, wind speed and direction, air pressure, and solar radiation intensity data and integrates history. The preprocessing unit is used to perform data cleaning, normalization and standardization, as well as preliminary construction operations of feature engineering to form basic data. The communication unit is used to provide data transmission and communication channels between modules.

[0051] The preprocessing unit includes a data cleaning subunit, a normalization subunit, and a feature engineering construction subunit. The data cleaning subunit uses a robust statistical method based on Hampel identifiers to identify sensor time series anomalies and repairs them through cubic spline interpolation. Adaptive median filtering is used to remove noise from images. The missing data of the time series are filled by seasonal decomposition-trend extrapolation method, and the missing pixels of the image are repaired based on the regional growing algorithm. The normalization subunit is used to map the data to the [0,1] interval and Z-score standardization for different data, respectively, to reshape the normal distribution of the data, so that various types of data have a unified dimension and a suitable data distribution form. The feature engineering construction subunit is used to derive features from the original data. The features include calculating the comprehensive value of magnetic field intensity and direction cosine features from the three-axis data of the magnetometer, and constructing meteorological data features. The meteorological data features include daily temperature difference, precipitation intensity, and wind speed and direction change rate.

[0052] Specifically, a magnetoresistive three-axis magnetometer is used to collect data on the intensity and direction changes of the magnetic field inside the mountain, temperature and humidity sensors are used to collect real-time data on environmental temperature and humidity, satellite and aerial remote sensing units are used to obtain satellite images and three-dimensional point cloud data, and meteorological data units are used to collect meteorological data such as temperature and precipitation and integrate historical data to form multi-source data.

[0053] By using a robust statistical method based on Hampel identifiers to identify sensor time series outliers and repairing them through cubic spline interpolation, adaptive median filtering is used to remove noise from images, seasonal decomposition-trend extrapolation method is used to fill missing data in time series, and regional growing algorithm is used to repair missing pixels in images, thereby ensuring the accuracy and completeness of the data and removing errors and missing interference in the data.

[0054] By using Min-Max normalization to map the data to the [0,1] interval and Z-score standardization to reshape the data normal distribution, various types of data have unified dimensions and appropriate data distribution forms, which is convenient for subsequent fusion processing.

[0055] By deriving features from the original data, such as calculating the comprehensive value of magnetic field intensity and direction cosine features from the three-axis data of the magnetometer, and constructing meteorological data features including daily temperature difference, precipitation intensity, wind speed and direction change rate, etc., we can enrich the data dimensions and mine potential information, providing a basis for a comprehensive analysis of ice and rock avalanche risks.

[0056] Please see attached Figure 4 The fusion extraction module includes a basic data processing unit, a fusion and bidirectional encoding unit, an attention mechanism unit, and a pre-training and fine-tuning unit. The basic data processing unit is used to construct a 12-layer DenseNet variant network for satellite and aerial images. The first layer uses 7×7 convolution kernels, 64, with a step size of 2 and a padding of 3 for convolution operation followed by a maximum pooling layer. The middle layer strengthens feature transmission through dense connection blocks. The last layer outputs the image space feature map through global average pooling and full connection. The local time series features of the sensor monitoring data are extracted through a one-dimensional convolutional network. The effective spatial and temporal feature information in the image and monitoring data are extracted to form feature information. The fusion and bidirectional encoding unit is used to splice the feature information in the feature dimension through a bidirectional long short-term memory network and input it according to the time step. The bidirectional hidden state is fused to output a comprehensive feature vector.

[0057] The attention mechanism unit is used to access the multi-head attention mechanism after the output of the bidirectional long short-term memory network, determine the weight of each feature based on complex calculations, focus on the key precursor features of ice and rock avalanche, and highlight important information. The pre-training and fine-tuning unit is used to pre-train the model backbone network based on public geological disaster images and meteorological environment monitoring data sets with multi-type annotations, freeze the parameters of the bottom 8 convolutional layers and load the local ice and rock avalanche related data, use the AdamW optimizer combined with a small learning rate and the learning rate is exponentially decayed according to the training rounds for fine-tuning, and use the focus loss function FL(p_t)=-\alpha_t(1-p_t)^{\gamma}\log(p_t) to optimize the model to adapt it to the local ice and rock avalanche monitoring task, and output fused feature data. Multi-type annotations include ice and rock avalanche, debris flow, and landslide.

[0058] Specifically, a 12-layer DenseNet variant network is constructed for satellite and aerial images. The first layer uses 7×7 convolution kernels, 64 kernels, stride 2 and padding 3 for convolution operation, followed by a maximum pooling layer. The middle layer strengthens feature transfer through densely connected blocks. The last layer outputs the image spatial feature map through global average pooling and full connection. The local time series features of the sensor monitoring data are extracted through a one-dimensional convolutional network. The effective spatial and temporal feature information in the image and monitoring data are extracted respectively to form feature information, so as to accurately capture the key features of different types of data. The feature information is spliced ​​in the feature dimension through a bidirectional long short-term memory network and input according to the time step. The bidirectional hidden state is fused to output the comprehensive feature vector, which realizes the deep fusion of multi-source data in the spatiotemporal dimension and effectively integrates information.

[0059] By connecting a multi-head attention mechanism after the output of the bidirectional long short-term memory network, the weight of each feature is determined by complex calculation, focusing on the key precursor features of ice and rock avalanches, highlighting important information, and facilitating the model to focus on the feature changes closely related to ice and rock avalanches. The backbone network of the model is pre-trained based on public geological disaster images and meteorological environment monitoring data sets with multi-type annotations (ice and rock avalanches, debris flows, and landslides). After freezing the parameters of the bottom 8 convolutional layers, the local ice and rock avalanche related data are loaded. The AdamW optimizer is combined with a small learning rate and the learning rate is exponentially decayed according to the training rounds for fine-tuning. The focus loss function FL(p_t)=-\alpha_t(1-p_t)^{\gamma}\log(p_t) is used to optimize the model to make it adapt to the local ice and rock avalanche monitoring task, and output fused feature data, thereby realizing the deep fusion and feature extraction of multi-source data.

[0060] Please see attached Figure 5 The recognition and prediction module includes a simulation and prediction unit and an enhanced recognition unit. The simulation and prediction unit is used to simulate the occurrence and development process of ice and rock avalanches and give prediction results through adversarial training between the discriminator and the generator based on cross entropy loss. At the same time, the uncertainty of the simulation results is analyzed based on Monte Carlo sampling statistics to assist risk decision-making. The enhanced recognition unit is used to select CatBoost, Random Forest and 4-layer multi-layer perceptron (MLP) as base classifiers, and use fused feature data and derived high-order features as input to train classification models for distinguishing whether the mountain is in potential danger of ice and rock avalanches and dividing the danger level. The soft voting method is used for integration to comprehensively output the dangerous area determination results and ice and rock avalanche simulation prediction information.

[0061] The generator is a 6-layer transposed convolutional neural network, which inputs a 128-dimensional random noise vector z and an environmental feature vector e encoded by a multi-layer perceptron. The first layer has a transposed convolution kernel of 4×4, a step size of 2, and an output channel of 1024. After batch normalization and ReLU activation, the convolution is cascaded and upsampled layer by layer. The last layer is activated by Tanh to output a simulated ice and rock collapse evolution image sequence and the corresponding physical quantity field. The discriminator uses a 5-layer convolutional neural network. The first layer has a convolution kernel of 4×4, a step size of 2, 32 channels, and LeakyReLU is activated. After multiple convolutions and activations, the last layer is fully connected to output the discrimination probability D(x).

[0062] Specifically, the discriminator and the generator are trained adversarially based on cross entropy loss. The generator is a 6-layer transposed convolutional neural network. It inputs a 128-dimensional random noise vector z and an environmental feature vector e encoded by a multi-layer perceptron. The first layer has a transposed convolution kernel of 4×4, a step size of 2, and an output channel of 1024. After batch normalization and ReLU activation, the convolution is upsampled layer by layer. The last layer is activated by Tanh to output the simulated ice and rock avalanche evolution image sequence and the corresponding physical quantity field. The discriminator adopts a 5-layer convolutional neural network. The first layer has a convolution kernel of 4×4, a step size of 2, and 32 channels. LeakyReLU is activated. After multiple convolutions and activations, the last layer is fully connected to output the discrimination probability D(x), thereby simulating the occurrence and development process of ice and rock avalanches and giving prediction results. The uncertainty of the simulation results is analyzed based on Monte Carlo sampling statistics to assist risk decision-making, which can more comprehensively consider the uncertainty factors of ice and rock avalanches and provide more reliable prediction information.

[0063] By selecting CatBoost, RandomForest and 4-layer multi-layer perceptron (MLP) as base classifiers, and taking fused feature data and derived high-order features as input, classification models are trained to determine whether a mountain is in potential danger of ice and rock avalanche and to divide the danger level. The soft voting method is used for integration, and the dangerous area determination results and ice and rock avalanche simulation prediction information are comprehensively output. The advantages of multiple models are complemented to improve the recognition accuracy, thereby achieving accurate identification of potential ice and rock danger areas in high-altitude cold regions and accurate prediction of ice and rock avalanche scenarios.

[0064] Please see attached Figure 6 The optimization layout module includes an environmental state unit, a reward function unit, and a reinforcement learning unit. The environmental state unit is used to construct a three-dimensional digital twin model of a high-cold mountain and discretize it into grid units. The state vector of each unit includes terrain, rock, meteorology, historical ice and rock collapse frequency, surrounding vegetation coverage, and soil moisture information. The action vector of the monitoring equipment includes three-dimensional coordinates, coverage radius, sampling frequency, transmission power, and monitoring angle parameters, and the environmental state and equipment action space are confirmed by this. The reward function unit is used to guide the optimization direction of the monitoring equipment layout of the acquisition preprocessing module through the reward function. The reward function is R=w1C υ -w2C cost -w3Rpenalty , where C υ is the effective monitoring coverage rate of ice and rock collapse precursors, C cost is the comprehensive cost of equipment deployment, R penalty As penalty items for equipment signal obstruction, redundant deployment, and frequent failures, the reinforcement learning unit is used to optimize the monitoring equipment layout strategy of the acquisition preprocessing module based on reward feedback through the proximal strategy optimization algorithm and interactive iterative training between the strategy network and the value network.

[0065] Specifically, by constructing a three-dimensional digital twin model of the high-cold mountain and discretizing it into grid units, the state vector of each unit (including terrain, rocks, meteorology, historical frequency of ice and rock avalanches, surrounding vegetation coverage, and soil moisture information) and the action vector of the monitoring equipment (including three-dimensional coordinates, coverage radius, sampling frequency, transmission power, and monitoring angle parameters) are determined to confirm the environmental status and equipment action space, thereby fully understanding the monitoring environment and equipment conditions.

[0066] The reward function is used to guide the optimization direction of the monitoring equipment layout of the acquisition preprocessing module, so that the monitoring equipment layout is adjusted in the direction of improving monitoring effects and reducing costs. The proximal strategy optimization algorithm is used, and the strategy network and the value network are interactively iteratively trained. The monitoring equipment layout strategy of the acquisition preprocessing module is optimized based on reward feedback, and the dynamic optimization of the monitoring equipment layout is achieved, thereby realizing the intelligent optimization of the monitoring equipment layout, solving the problem that the traditional monitoring equipment layout is fixed, lacks flexibility and adaptability, and cannot be adjusted according to the dynamic changes of ice and rock avalanche risks.

[0067] Please see attached Figure 7 The early warning interaction module includes an early warning unit and an interactive unit. The early warning unit is used to perform graded early warning based on the risk probability output by the identification and prediction module according to the preset danger level threshold. When the risk probability is greater than the threshold, early warning information is pushed through multiple channels, including SMS gateways, sound and light alarms, and mobile phone APPs. The early warning information includes the precise geographic coordinates of the dangerous area, the estimated scale of ice and rock avalanches, and the expected time range of occurrence. The interactive unit is used to display the real-time monitoring curve of the sensor, the heat map of the ice and rock avalanche danger area, and the animation of the predicted ice and rock avalanche movement trajectory based on WebGIS combined with Cesium three-dimensional earth engine and Echarts visualization library, and provides mouse wheel zooming, translation, rotation of perspective, historical data tracing back, and simulation of protection plans, so as to intuitively present the monitoring and prediction situation to assist decision-making.

[0068] Specifically, a graded warning is issued by comparing the risk probability output by the identification and prediction module according to the preset danger level threshold. When the risk probability is greater than the threshold, warning information is pushed through multiple channels such as SMS gateway, sound and light alarm, and mobile phone APP. The information includes the precise geographic coordinates of the dangerous area, the estimated scale of the ice and rock avalanche, and the expected time range of occurrence, thereby ensuring that the warning information can be conveyed to relevant personnel in a timely and accurate manner, thereby improving the timeliness and pertinence of the warning.

[0069] Based on WebGIS combined with Cesium 3D Earth Engine and Echarts visualization library, it displays real-time monitoring curves of sensors, heat maps of dangerous areas of ice and rock avalanches, and animations of predicted movement trajectories of ice and rock avalanches. It also provides functions such as mouse wheel zooming, panning, rotating perspective, backtracking historical data, and simulating protection plans. It intuitively presents monitoring and prediction situations to assist decision-making, enabling decision makers to fully understand the situation of ice and rock avalanches and respond effectively.

[0070] Please see attached Figure 8 , a method for identifying potential dangerous areas of ice and rock avalanches in high-cold areas, comprising the following steps:

[0071] S1. Collection and preprocessing: Collect and preprocess multi-source data of ice and rock in high-cold areas to form basic data;

[0072] S2, fusion extraction: perform data fusion processing and feature extraction on the basic data to form fusion feature data;

[0073] S3, Identification and prediction: Identify potential dangerous areas of ice and rock in high-cold areas and predict ice and rock collapse scenarios based on fused feature data, and form dangerous area determination results and ice and rock collapse simulation prediction information;

[0074] S4. Optimize layout: optimize the layout of monitoring equipment based on multi-source data and monitoring equipment status information;

[0075] S5. Early warning interaction: Provide early warning prompts and visual displays based on the dangerous area determination results, ice and rock collapse simulation prediction information, and monitoring equipment status information.

[0076] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A device for identifying potential dangerous areas of ice and rock avalanches in high-cold areas, characterized in that: include: Collection and preprocessing module: used to collect and preprocess multi-source data of ice and rock in alpine areas to form basic data; Fusion extraction module: used to perform data fusion processing and feature extraction on basic data to form fusion feature data; Identification and prediction module: used to identify potential dangerous areas of ice and rock in high-cold areas and predict ice and rock collapse scenarios based on fused feature data, forming dangerous area determination results and ice and rock collapse simulation prediction information; Layout optimization module: used to optimize the monitoring equipment layout of the acquisition preprocessing module based on multi-source data and monitoring equipment status information; Early warning interaction module: used to provide early warning prompts and visual displays based on the results of dangerous area determination, ice and rock collapse simulation prediction information, and monitoring equipment status information.

2. The device for identifying potential dangerous areas of ice and rock avalanches in high-cold areas according to claim 1 is characterized by: The acquisition and preprocessing module includes a sensor network unit, a satellite and aerial remote sensing unit, a meteorological data unit, a preprocessing unit, and a communication unit. The sensor network unit is used to collect corresponding data through sensors in high-cold areas. The sensors include a magnetoresistive three-axis magnetometer and a temperature and humidity sensor. The corresponding data include data on the intensity and direction changes of the magnetic field inside the mountain and real-time data on ambient temperature and humidity. The satellite and aerial remote sensing unit is used to regularly obtain satellite images and three-dimensional point cloud data in high-cold areas. The satellite images include images of terrain, glaciers, and rock textures. The meteorological data access unit is used to obtain local meteorological data in high-cold areas. The meteorological data includes temperature, precipitation, wind speed and direction, air pressure, and solar radiation intensity data and integrates history. The preprocessing unit is used to perform data cleaning, normalization and standardization, as well as preliminary construction operations of feature engineering to form basic data. The communication unit is used to provide data transmission and communication channels between modules.

3. The device for identifying potential dangerous areas of ice and rock avalanches in high-cold regions according to claim 2 is characterized in that: The preprocessing unit includes a data cleaning subunit, a normalization subunit, and a feature engineering construction subunit. The data cleaning subunit uses a robust statistical method based on Hampel identifiers to identify sensor time series anomalies and repairs them through cubic spline interpolation, uses adaptive median filtering to remove noise from images, fills in missing time series data through seasonal decomposition-trend extrapolation, and repairs missing image pixels based on regional growing algorithm. The normalization subunit is used to use Min-Max normalization to map data to the [0,1] interval and Z-score standardization for different data to reshape the normal distribution of data, so that various types of data have a unified dimension and a suitable data distribution form. The feature engineering construction subunit is used to derive features from the original data, including calculating the comprehensive value of magnetic field intensity and direction cosine features from the three-axis data of the magnetometer, and constructing meteorological data features, and the meteorological data features include daily temperature difference, precipitation intensity, and wind speed and direction change rate.

4. The device for identifying potential dangerous areas of ice and rock avalanches in high-cold regions according to claim 1 is characterized in that: The fusion extraction module includes a basic data processing unit, a fusion and bidirectional encoding unit, an attention mechanism unit, and a pre-training and fine-tuning unit. The basic data processing unit is used to construct a 12-layer DenseNet variant network for satellite and aerial images. The first layer uses 7×7 convolution kernels, 64 kernels, a step size of 2 and a padding of 3 for convolution operation followed by a maximum pooling layer. The middle layer strengthens feature transmission through dense connection blocks. The last layer outputs an image space feature map through global average pooling and full connection. The local time series features of the sensor monitoring data are extracted through a one-dimensional convolutional network, and the effective spatial and temporal feature information in the image and monitoring data are extracted respectively to form feature information. The fusion and bidirectional encoding unit is used to splice the feature information in the feature dimension through a bidirectional long short-term memory network and input it according to the time step, and fuse the bidirectional hidden state to output a comprehensive feature vector.

5. The device for identifying potential dangerous areas of ice and rock avalanches in high-cold regions according to claim 4 is characterized in that: The attention mechanism unit is used to access the multi-head attention mechanism after the output of the bidirectional long short-term memory network, determine the weight of each feature based on complex calculations, focus on the key precursor features of ice and rock avalanche, and highlight important information. The pre-training and fine-tuning unit is used to pre-train the model backbone network based on public geological disaster images and meteorological environment monitoring data sets with multi-type annotations, freeze the parameters of the bottom 8 convolutional layers and then load the local ice and rock avalanche related data, use the AdamW optimizer in combination with a small learning rate and the learning rate is exponentially decayed according to the training rounds for fine-tuning, use the focus loss function FL(p_t)=-\alpha_t(1-p_t)^{\gamma}\log(p_t), optimize the model to make it adapt to the local ice and rock avalanche monitoring task, and output fused feature data. The multi-type annotations include ice and rock avalanche, debris flow, and landslide.

6. The device for identifying potential dangerous areas of ice and rock avalanches in high-cold regions according to claim 1 is characterized by: The identification and prediction module includes a simulation and prediction unit and an enhanced identification unit. The simulation and prediction unit is used to simulate the occurrence and development process of ice and rock avalanches and give prediction results through adversarial training between discriminators and generators based on cross-entropy loss. At the same time, the uncertainty of simulation results is analyzed based on Monte Carlo sampling statistics to assist risk decision-making. The enhanced identification unit is used to select CatBoost, Random Forest and 4-layer multi-layer perceptron (MLP) as base classifiers, and use fused feature data and derived high-order features as input to train classification models for distinguishing whether a mountain is in potential danger of ice and rock avalanches and dividing danger levels. The soft voting method is used for integration to comprehensively output dangerous area determination results and ice and rock avalanche simulation prediction information.

7. The device for identifying potential dangerous areas of ice and rock avalanches in high-cold regions according to claim 6 is characterized by: The generator is a 6-layer transposed convolutional neural network, which inputs a 128-dimensional random noise vector z and an environmental feature vector e encoded by a multi-layer perceptron. The first layer has a transposed convolution kernel of 4×4, a step size of 2, and an output channel of 1024. After batch normalization and ReLU activation, the samples are cascaded layer by layer. The last layer is activated by Tanh to output a simulated ice and rock collapse evolution image sequence and a corresponding physical quantity field. The discriminator adopts a 5-layer convolutional neural network, with a first layer convolution kernel of 4×4, a step size of 2, 32 channels, and LeakyReLU activated. After multiple convolutions and activations, the last layer is fully connected to output the discrimination probability D(x).

8. The device for identifying potential dangerous areas of ice and rock avalanches in high-cold regions according to claim 1 is characterized by: The optimization layout module includes an environmental state unit, a reward function unit, and a reinforcement learning unit. The environmental state unit is used to construct a three-dimensional digital twin model of a high-cold mountain and discretize it into grid units. Each unit state vector includes terrain, rock, weather, historical ice and rock collapse frequency, surrounding vegetation coverage, and soil moisture information. The monitoring equipment action vector includes three-dimensional coordinates, coverage radius, sampling frequency, transmission power, and monitoring angle parameters, and the environmental state and equipment action space are confirmed by this. The reward function unit is used to guide the optimization direction of the monitoring equipment layout of the acquisition preprocessing module through the reward function. The reward function is R=w1C v -w2C cost -w3R penalty , where C v is the effective monitoring coverage rate of ice and rock collapse precursors, C cost is the comprehensive cost of equipment deployment, R penalty The penalty item targets device signal blocking, redundant deployment, and frequent failures. The reinforcement learning unit is used to optimize the monitoring device layout strategy of the acquisition preprocessing module based on reward feedback through the proximal strategy optimization algorithm, interactive iterative training of the strategy network and the value network.

9. The device for identifying potential dangerous areas of ice and rock avalanches in high-cold regions according to claim 1 is characterized by: The early warning interaction module includes an early warning unit and an interaction unit. The early warning unit is used to perform graded early warning based on the risk probability output by the identification prediction module according to a preset danger level threshold. When the risk probability is greater than the threshold, early warning information is pushed through multiple channels. The multiple channels include SMS gateways, sound and light alarms, and mobile phone APPs. The early warning information includes precise geographic coordinates of dangerous areas, estimated ice and rock avalanche scale, and expected time range of occurrence. The interaction unit is used to display real-time monitoring curves of sensors, heat maps of ice and rock avalanche danger areas, and animations of predicted ice and rock avalanche movement trajectories based on WebGIS combined with Cesium three-dimensional earth engine and Echarts visualization library, and provides mouse wheel zooming, translation, rotation of viewing angle, historical data tracing back, and simulated protection plans, so as to intuitively present monitoring and prediction situations to assist decision-making.

10. A method for identifying potential dangerous areas of ice and rock avalanches in high-cold regions, characterized in that: The following steps are involved: S1. Collection and preprocessing: Collect and preprocess multi-source data of ice and rock in high-cold areas to form basic data; S2, fusion extraction: perform data fusion processing and feature extraction on the basic data to form fusion feature data; S3, Identification and prediction: Identify potential dangerous areas of ice and rock in high-cold areas and predict ice and rock collapse scenarios based on fused feature data, and form dangerous area determination results and ice and rock collapse simulation prediction information; S4. Optimize layout: optimize the layout of monitoring equipment based on multi-source data and monitoring equipment status information; S5. Early warning interaction: Provide early warning prompts and visual displays based on the dangerous area determination results, ice and rock collapse simulation prediction information, and monitoring equipment status information.

Images