A method for predicting the effective impact range of avalanche human intervention based on deep learning

Abstract

The present invention discloses a method for predicting the effective influence range of artificial intervention in avalanches based on deep learning, which specifically relates to the technical field of avalanche risk control, and comprises the following steps: collecting three-dimensional terrain data and snow body data of snow mountain slopes by means of drones, constructing a slope terrain database and a snow body database, and labeling them in combination with a near-field overpressure database of air blast of explosives; based on a deep learning algorithm, constructing an avalanche probability prediction model using the labeled data set, predicting the avalanche probability of all preset ranges of the snow mountain slopes, and generating a probability cloud map; projecting the cloud map onto a three-dimensional model, and adjusting the display threshold according to different risk management and control requirements, thereby judging whether the effective influence range of artificial intervention in avalanches meets engineering requirements, and realizing precise control; the present invention not only improves the applicability and accuracy of the model, but also can accurately predict the avalanche risks of different grid areas, improve the accuracy of blasting intervention, and avoid unnecessary blasting and waste of resources.

Description

Technical Field

[0001] The present invention relates to the technical field of avalanche risk control, and more specifically, to a method for predicting the effective impact range of artificial intervention in avalanches based on deep learning. Background Art

[0002] In high mountains and snowy mountains, snow accumulation is often accompanied by potential avalanche risks, especially when the snow thickness increases or after human intervention such as blasting, the probability of avalanches increases significantly. In order to prevent avalanches from posing a serious threat to mountain infrastructure, transportation routes, and personnel safety, human intervention measures are usually adopted, such as pre-triggering small-scale avalanches through targeted blasting to avoid larger-scale natural avalanches. However, how to accurately assess and predict the effective impact range of human intervention has always been a difficult point in the field of avalanche control.

[0003] Traditional avalanche intervention plans often rely on experience or simple numerical models, lack accurate predictions of the probability of avalanches, and cannot make effective risk assessments for different terrains and environmental conditions. With the widespread application of deep learning and big data technologies, it has become possible to accurately predict avalanche risks by combining three-dimensional terrain data, snow accumulation data, and blasting parameters. This data-driven prediction model can more accurately determine the possible range of avalanches after blasting, and then optimize blasting intervention plans and achieve precise control. However, how to select the learning algorithm that best suits the current data characteristics from a variety of algorithm models has become the key to further ensuring the effectiveness of artificial intervention in avalanches. Summary of the invention

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A method for predicting the effective impact range of artificial intervention in avalanches based on deep learning, comprising the following steps:

[0006] S1. Collect three-dimensional terrain data and high snow accumulation data of snow mountain slopes through drones, and build a slope terrain database and snow mountain slope snow accumulation database within a preset range;

[0007] S2. Constructing an explosive air blast near-field overpressure database, and labeling the data in the slope terrain database, the snow body database, and the explosive air blast near-field overpressure database;

[0008] S3. Based on the deep learning algorithm, the data set obtained after labeling is input into the avalanche probability prediction model for construction;

[0009] S4. Use the constructed avalanche probability prediction model to perform traversal prediction on all preset ranges of the snow mountain slope to form an avalanche probability cloud map of the entire snow mountain slope;

[0010] S5. Project the probability cloud map of the snow mountain slope obtained in S4 onto the overall three-dimensional model of the snow mountain slope, adjust the probability cloud map display threshold according to different risk management requirements, and judge whether the effective impact range of artificial intervention in avalanche meets the engineering requirements based on the projection display results.

[0011] In a preferred embodiment, the preset range refers to: dividing the entire snow mountain slope into 5m×5m grids, and each grid obtained after the division is the preset range.

[0012] In a preferred embodiment, in the constructed explosive air explosion near-field overpressure database, the air shock wave overpressure data calculation formula of the explosive air explosion is as follows:

[0013] ;

[0014] is the equivalent attenuation distance of air shock wave pressure, It is the overpressure data of air shock wave pressure of explosive air explosion.

[0015] In a preferred embodiment, the air shock wave pressure equivalent attenuation distance The calculation formula is:

[0016] ; R is the shock wave pressure attenuation distance, and W is the mass of the explosive.

[0017] The content of the labeling is: the probability value of avalanche occurring in the snow body on the slope of the snow mountain under the condition of fixed amount of explosives and blasting points.

[0018] In a preferred embodiment, the deep learning algorithm includes fractional regression, DeepAR autoregressive recursion, LSTM, and Transformer models. Feature extraction is performed on the data set obtained after labeling, and a preset learning strategy is used to generate selection coefficients and applicability coefficients respectively. Then, the corresponding unique determination item algorithm model is selected from the fractional regression, DeepAR autoregressive recursion, LSTM, and Transformer models through fuzzy logic to construct an avalanche probability prediction model.

[0019] In a preferred embodiment, the logic for obtaining the selection coefficient is:

[0020] Get the data dimensions of each type of data extracted from the data set obtained after labeling , time series length , the variance of each type of data , and the time series correlation of various types of data , i represents the data type, and then substitute it into the selection coefficient calculation formula:

[0021] ; By calculating the data dimensions of each type of data The product of is, By calculating the variance of each type of data The average value of By calculating the time series correlation of various types of data The average value of is the selection coefficient, , , , All are preset non-zero scale factors.

[0022] In a preferred embodiment, the logic for obtaining the applicable coefficient is:

[0023] Get the learning rate from the preset learning strategy , hidden layer dimension , loss rate , batch value , and then substitute it into the applicable coefficient calculation formula:

[0024] ; is the applicable coefficient, , , , All are preset non-zero influence coefficients.

[0025] In a preferred embodiment, fuzzy logic refers to:

[0026] The selection coefficient and applicability coefficient in the decision-making of building the avalanche probability prediction model are taken as input variables, and the type of learning algorithm used in the construction of the avalanche probability prediction model is taken as the output variable. The input variables are fuzzified and the values ​​of the input variables are converted into fuzzy sets. The output variables are fuzzified and the output variables are converted into fuzzy sets. Fuzzy rules are formulated to describe the adaptability of the learning algorithm under different data type combinations. The fuzzified input variables are inferred through fuzzy rules to obtain the type of learning algorithm used in the construction of the avalanche probability prediction model.

[0027] In a preferred embodiment, judging whether the effective influence range of artificial intervention in avalanche meets engineering requirements means: adjusting the probability cloud map display threshold according to different risk management requirements, and grid points greater than the threshold are regarded as avalanches occurring after blasting.

[0028] Technical effects and advantages of the present invention:

[0029] Based on a deep learning algorithm, the present invention can accurately analyze the probability of avalanche after blasting by introducing three-dimensional terrain data, snow accumulation data and blasting parameters collected by drones. Compared with the traditional empirical method, the present invention can more accurately predict the avalanche risk in different grid areas, improve the accuracy of blasting intervention, and avoid unnecessary blasting and waste of resources.

[0030] The present invention introduces selection coefficient and applicability coefficient, combined with fuzzy logic reasoning, and can automatically screen out the learning algorithm model that best suits the current environment and data characteristics. This method not only improves the applicability and accuracy of the model, but also greatly reduces the time and complexity of manual algorithm selection, ensuring that the most appropriate algorithm model can be selected in different scenarios.

[0031] Through the generation of avalanche probability cloud maps, the present invention can flexibly adjust the display threshold of the probability cloud map according to different risk management requirements, helping engineers accurately determine the effective impact range of manual intervention. Areas greater than the set threshold will be considered high-risk areas, ensuring that manual intervention accurately covers key areas and effectively reduces the occurrence of potential avalanches.

[0032] Accurate risk prediction and optimization of blasting intervention plans can reduce excessive blasting and avoid excessive damage to the environment. At the same time, the present invention can help reduce unnecessary waste of resources and achieve the maximum risk control effect with limited explosives through optimized blasting design. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to facilitate understanding by those skilled in the art, the present invention is further described below in conjunction with the accompanying drawings;

[0034] Figure 1 This is a schematic diagram of a method for predicting the effective impact range of artificial intervention in avalanches based on deep learning in the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments 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.

[0036] Reference Figure 1 The following embodiments are obtained:

[0037] Embodiment 1: A method for predicting the effective impact range of artificial intervention in avalanches based on deep learning, comprising the following steps:

[0038] S1. Collect the three-dimensional terrain data and high snow accumulation data of the snow mountain slopes through drones, and build a slope terrain database and snow accumulation database of the preset range; obtain the three-dimensional terrain and snow accumulation data of the target snow mountain slopes by mounting cameras and infrared equipment on drones. The three-dimensional terrain data includes information such as elevation, slope, and terrain structure, while the snow accumulation data includes snow thickness, humidity, density, etc. The purpose of collecting this data is to build a detailed digital model to provide data support for subsequent analysis. The constructed database can cover the terrain and snow accumulation status of the entire snow mountain slope, forming a detailed basic data set for subsequent avalanche prediction.

[0039] S2. Construct a database of near-field overpressure caused by explosive air blasts, and label the data in the slope terrain database, snow accumulation database, and near-field overpressure database caused by explosive air blasts; this step is to prepare and organize the data. The database of near-field overpressure caused by explosive air blasts records the air shock wave pressure data under conditions such as different amounts of explosives and blasting point locations. It provides important parameters for simulating and analyzing avalanche risks. By labeling the terrain data, snow accumulation data, and blasting data, a training data set is generated. The labeled content may include the probability of avalanches occurring under specific conditions. This step lays the foundation for the subsequent model training, ensuring that the model can "learn" the impact of blasting on avalanches under each environmental condition.

[0040] S3. Based on the deep learning algorithm, the data set obtained after the labeling is input into the avalanche probability prediction model for construction; this step is the core of model construction and training. Based on the deep learning algorithm (such as LSTM, Transformer, etc.), the data set labeled in the previous step is input into the model for training to build a model that can predict the probability of avalanche. The model predicts the probability of avalanche under different conditions by learning the relationship between terrain, snow accumulation and blasting data. Deep learning can handle complex nonlinear data relationships, so it has strong adaptability and accuracy in avalanche risk prediction. Through model construction, a tool that can evaluate the risk of avalanches in future blasting situations can be obtained.

[0041] S4. Use the constructed avalanche probability prediction model to traverse and predict all preset ranges of the snow mountain slope to form an avalanche probability cloud map of the entire snow mountain slope; this is the step of applying the model for global prediction. The constructed avalanche probability prediction model can predict each grid (preset range, such as 5m×5m) of the entire snow mountain slope to form an avalanche probability cloud map. The significance of the cloud map is to visualize the avalanche risk distribution of the entire snow mountain slope and help users identify areas where avalanches may occur. Through traversal prediction, a wide-coverage risk map can be generated for the entire area to support subsequent risk decisions. The predicted value of each grid is the probability of avalanche in the grid after blasting.

[0042] S5. Project the probability cloud map of the snow mountain slope obtained in S4 onto the overall three-dimensional model of the snow mountain slope, adjust the probability cloud map display threshold according to different risk control requirements, and judge whether the effective impact range of artificial intervention in avalanches meets the engineering requirements based on the projection display results. This step is a key step in combining the prediction results with actual engineering needs for decision support. By projecting the avalanche probability cloud map onto the three-dimensional terrain model, the avalanche risk of each area after blasting can be intuitively displayed. By adjusting the display threshold, areas with specific risk levels can be viewed according to different risk control requirements to evaluate whether the blasting has effectively covered high-risk areas. This step helps users judge whether the effective impact range of artificial intervention meets engineering requirements in a visual way, thereby optimizing the blasting plan and improving the accuracy of avalanche control.

[0043] The preset range refers to the division of the entire snow mountain slope into 5m×5m grids, and each grid obtained after division is the preset range. This means that the three-dimensional terrain and snow data of the snow mountain will be divided into 5-meter square areas as the smallest unit for avalanche probability prediction and risk assessment. Prediction within these grids can ensure the precision of the analysis, thereby helping to more accurately assess the possible avalanche impact range after blasting.

[0044] By building a database of near-field overpressure of explosive air explosion, the air shock wave overpressure data at any distance can be called out according to the location of the blasting point, the amount of explosives used and other data; among them, the empirical formula for the air shock wave overpressure of explosive air explosion is as follows:

[0045] ;

[0046] is the equivalent attenuation distance of air shock wave pressure, It is the overpressure data of air shock wave pressure of explosive air explosion.

[0047] Equivalent attenuation distance of air shock wave pressure The calculation formula is:

[0048] ; R is the shock wave pressure attenuation distance, and W is the mass of the explosive.

[0049] The database records the air shock wave overpressure data under different blasting point locations, explosive dosage and other conditions. It is used to simulate the pressure exerted on the environment by the air shock wave after blasting at different distances. The purpose of this database is to provide key data for avalanche risk prediction and help analyze whether the shock wave at the blasting point will lead to an avalanche.

[0050] When calculating the air shock wave overpressure Δp generated by the air explosion of explosives at different distances, different parts of the formula are calculated in sections according to the distance range of the shock wave, including: 0.05≤ <0.3: This is a very close distance, the formula is more complex, and multiple high-order terms are considered to accurately estimate the overpressure. 0.3≤ <1: The formula for the medium distance range is relatively simplified and describes the trend of overpressure change at this distance. 1≤ ≤10: At long distances, the formula is further simplified because the overpressure decays rapidly with distance. Δp represents the air shock wave overpressure, in units of pressure (usually Pascals, Pa). This is the main output we are interested in, used to evaluate the impact force of the explosion. It is the standardized equivalent attenuation distance of air shock wave, which is used to simplify the calculation of the formula. It is defined as: ; R is the actual attenuation distance of the air shock wave, which indicates the distance between the blasting point and the target point, in meters. W is the mass (weight) of the explosive, in kilograms. The larger the mass of the explosive, the greater the overpressure of the shock wave and the wider the impact range. The formula simplifies the calculation of explosion intensity by introducing a standardized equivalent attenuation distance of the air shock wave. Within different distance ranges, the number of terms in the formula is different, and high-order nonlinear relationships are considered to accurately describe the changes in overpressure at different distances. Δp is the pressure value of the air shock wave at the target point after the blast. By calculating Δp at each point, the impact of the blasting on the surrounding environment, especially the impact on the snow mass, can be evaluated. If the shock wave pressure exceeds a certain threshold, an avalanche may occur. Combining the distance R and the explosive mass W into a standardized parameter simplifies the analysis under different blasting conditions. Using standardized distances, the impact of different explosion sizes can be analyzed more systematically. By building a database of near-field overpressures of explosive air blasts and using this empirical formula, the pressure of air shock waves at different distances can be accurately predicted, providing key data support for avalanche prediction. This process not only takes into account the amount of explosives and distance, but also uses a segmented formula to process the changes in overpressure at different distance ranges to ensure the accuracy of the calculation.

[0051] Based on the experience of avalanche blasting, the data in the above slope terrain database, snow accumulation database and explosive air blast near-field overpressure database are labeled with the probability value of avalanche in the snow accumulation on the slope terrain under the condition of fixed explosive amount and blasting point. In the data preparation stage, each data point (including terrain, snow accumulation data, etc.) will be labeled with the possibility of avalanche under specific blasting conditions. These labeled data will be used to train the avalanche probability prediction model so that the model can predict the avalanche risk under different conditions.

[0052] The deep learning algorithm includes fractional regression, DeepAR autoregressive recursion, LSTM, and Transformer models. By extracting features from the data set obtained after labeling, and using the preset learning strategy, the selection coefficient and the applicable coefficient are generated respectively. Then, the corresponding unique determination item algorithm model is selected from the fractional regression, DeepAR autoregressive recursion, LSTM, and Transformer models through fuzzy logic to build an avalanche probability prediction model. Introduction to the avalanche probability prediction model: Input: three-dimensional terrain data of a single point on the snow mountain slope (5m×5m grid range); explosive blasting data: explosive usage, latitude and longitude coordinates of the blasting point, and height of the blasting point from the snow surface; snow accumulation data: snow accumulation humidity, thickness, and density; Output: The probability of avalanche occurring at the slope point after being affected by explosive blasting.

[0053] Fractional regression is a method that extends classical linear regression and is mainly used to estimate different distribution quantiles of the response variable, rather than just the mean. It is particularly suitable for modeling when the data distribution is asymmetric or heteroskedastic. Unlike the least squares method (OLS) which estimates the mean of the response value by minimizing the squared error, fractional regression estimates the value of a specific quantile (such as the 0.1 quantile, 0.5 quantile, etc.) by minimizing the weighted absolute value of the deviation. Application scenario: Fractional regression can be used in avalanche prediction scenarios to predict not only an average probability of occurrence, but also the distribution range of the probability of avalanche occurrence under different conditions (such as extremely high risk in the worst case and more optimistic low risk).

[0054] DeepAR is a time series prediction method based on recurrent neural networks (RNNs), designed to handle simultaneous predictions of multiple related time series. The model autoregressively generates predictions of future values ​​by using historical observations and external features as inputs. The core of DeepAR combines the characteristics of recursive networks and autoregressive models to predict the distribution of possible values ​​in the future by modeling past time series data. Application scenarios: In the prediction of the impact range of avalanche intervention, DeepAR can generate future risk forecasts by combining historical avalanche data from different regions, which is suitable for data with multiple related time series (such as different regions and different weather conditions).

[0055] LSTM is a special recurrent neural network (RNN) that solves the problem of long-term dependencies in traditional RNNs. LSTM introduces "memory cells" and "gate mechanisms" to control the inflow, retention, and discard of information, allowing the network to effectively capture long-term and short-term dependencies. LSTM consists of forget gates, input gates, and output gates, which are used to selectively update and read information in memory cells. Application scenarios: In avalanche prediction, LSTM can capture complex, long-term dependent dynamic patterns related to avalanche occurrence, such as the impact of long-term snow accumulation and temperature changes on the probability of future avalanches.

[0056] Transformer is a neural network architecture based on the self-attention mechanism, which was first applied to natural language processing. Unlike RNNs such as LSTM, Transformer does not rely on the sequential transmission of sequence data, but uses a global self-attention mechanism to directly capture the correlation of any position in the sequence. Its core mechanism is "multi-head self-attention", which can process the entire sequence in parallel and pay attention to other positions in the sequence at each position. The Transformer model is superior to traditional RNN in terms of computational efficiency and modeling long dependencies. Application scenario: In avalanche intervention prediction, Transformer can quickly process a large amount of gridded terrain data, capture the mutual influence between various areas of the snow mountain, and provide a comprehensive avalanche probability forecast for the entire snow mountain slope.

[0057] Fractional regression is suitable for estimating different avalanche risk quantiles and capturing the worst case scenario. DeepAR uses the autoregressive properties of time series and is suitable for multi-region prediction based on historical data. LSTM is suitable for capturing complex, long-term dynamic patterns, such as long-term snow accumulation and environmental changes. Transformer uses a global self-attention mechanism to efficiently capture complex dependencies within the entire area and is suitable for large-scale terrain prediction.

[0058] Although the fractional regression, DeepAR autoregressive recursive, LSTM and Transformer models have different structures and application scenarios, they often involve some common parameter points during the training and prediction process.

[0059] Input data dimensions: All models need to process the dimensions of input data, such as 3D terrain data, snowpack data, blasting point data, etc. for each grid point in avalanche prediction. These input data dimensions determine how much capacity the model needs to process the information.

[0060] Time series length: For DeepAR and LSTM, the length of the time series is an important parameter that directly affects the model's memory capacity and prediction range.

[0061] Learning rate: This is a key training parameter of the deep learning model, which determines the pace at which the model adjusts the weights in each iteration. A larger learning rate may lead to unstable training, while a smaller learning rate may make the convergence speed too slow.

[0062] Dropout: A regularization parameter for deep learning models that prevents overfitting. All models can use dropout, especially when dealing with complex, noisy data.

[0063] Batch size: During training, the batch size affects the stability and convergence speed of the model. Smaller batches usually lead to higher training stability, while larger batches can improve computational efficiency.

[0064] Hidden layer dimension: For LSTM, DeepAR, and Transformer, the hidden layer dimension of the model determines the expressiveness of the model, that is, the ability of the model to capture complex data relationships internally.

[0065] The logic for obtaining the selection coefficient is: obtain the data dimensions of each type of data extracted from the features of the data set obtained after the labeling is completed , time series length , the variance of each type of data , and the time series correlation of various types of data , i represents the data type, and then substitute it into the selection coefficient calculation formula:

[0066] ; By calculating the data dimensions of each type of data The data dimension represents the complexity of each data, such as the representation of terrain data and snow volume data in different dimensions. For example, the dimensions of terrain data are length, width, and height, and the snow volume data has three attributes: humidity, thickness, and density. The more dimensions, the higher the complexity of the data, and the model needs stronger capabilities to process these complex data. The product operation ensures that all data dimensions are taken into account. The larger the dimension, the more effectively the product can reflect the impact of all dimensions, reflecting the importance of high-dimensional data to model selection. Large data dimensions mean that more complex models are needed.

[0067] The time series length is the length of historical data used to train the model (such as the past few months, years, or past event sequences). The square of the time series length amplifies the impact of long time series, indicating that under long time series data, the model needs stronger time dependence capabilities (such as LSTM, Transformer). The square of the time series makes the impact of long sequences greater and the impact of short time series is weakened, emphasizing the importance of long time series for model selection.

[0068] In order to reflect the nonlinearity of the input data, it can be measured by the statistical characteristics of the data set such as standard deviation or variance. By calculating the variance of each type of data The data variance reflects the dispersion of the data. The larger the variance, the stronger the volatility of the data, and the greater the challenge faced by the model in prediction. The variance of the data directly affects the difficulty of model fitting. The larger the variance, the stronger the generalization ability of the model is needed to deal with uncertainty. The effect of variance on model selection is amplified, ensuring that the model can flexibly adapt to different situations under highly volatile data.

[0069] The autocorrelation coefficient ACF measures the internal correlation of the time series. A series with strong correlation is easier to predict, while a series with weak correlation requires a more complex model. Time correlation measures the trend strength in the time dimension of the data and reflects the degree of applicability to the time series model. By calculating the time series correlation of various types of data The average value of the time series correlation is obtained. It is calculated by the existing autocorrelation function and will not be described here. Data with strong autocorrelation indicates that the model can more easily capture the regularity of the data, while data with weak autocorrelation requires a more complex model. The sigmoid function is used to normalize the value of K to ensure that when the autocorrelation is low, the model selection tends to be more complex. is the selection coefficient, , , , All are preset non-zero proportional coefficients that control the importance of each parameter in the selection coefficient calculation. The proportional coefficient indicates that the parameter has a greater impact on the model selection. A larger selection coefficient indicates that the data dimension is high, the time span is large, the variance of the data is large, or the dependence between the data is weak, and a more complex model is needed to process the data. When the selection coefficient is large, it means that a more complex model or a model with more nonlinear features may need to be selected.

[0070] The logic for obtaining the applicable coefficient is: obtain the learning rate from the preset learning strategy , hidden layer dimension , loss rate , batch value , and then substitute it into the applicable coefficient calculation formula:

[0071] ; is the applicable coefficient, , , , All are preset non-zero influence coefficients.

[0072] The learning rate controls the pace of the model weight update during training. The larger the learning rate, the faster the model updates, but it may cause unstable model training; if the learning rate is too small, the model training speed is slow and may not converge effectively. The effect of the learning rate is smoothed so that extreme values ​​of the learning rate do not cause excessive fluctuations in the choice of the model. By using this function, when the learning rate is large, the corresponding fitness coefficient growth rate gradually slows down, while for smaller learning rates, the change is more sensitive.

[0073] The hidden dimension represents the number of hidden neurons in the neural network, which directly affects the expressiveness and complexity of the model. The larger the hidden dimension, the more data features the model can learn, but it may also lead to overfitting. The square root of the hidden dimension balances the impact of model complexity on the fitness coefficient. The importance of large hidden dimension to model fitness is magnified, but by taking the square root, the growth rate is not linear, but gradually slows down, preventing the dimension from having too much impact on model selection when it is too large.

[0074] The dropout rate controls the random blocking of a portion of neurons in the neural network to prevent overfitting. A higher dropout rate can prevent the model from overfitting, but it may also cause underfitting. This fraction maps the loss rate to between 0 and 1, making the impact of the loss rate smoother and ensuring that when the loss rate is very high or very low, it will not have an extreme impact on the model's fitness. As the loss rate increases, the growth rate of the fraction gradually slows down, ensuring that changes in the loss rate have a moderate impact on fitness.

[0075] Batch size is the batch value Determines the number of data samples used in each training. Larger batches generally speed up training, but when they are too large, they affect the generalization ability of the model; smaller batches help capture more subtle data changes, but may lead to long training times or unstable gradients. The square of the batch size amplifies the effect of the batch size on fitness. Larger batches may mean that the model requires higher computing resources, while more adaptable models are more likely to benefit from larger batches. The squared approach ensures that large batch sizes have a more significant impact on the fitness of the model.

[0076] The larger the applicability coefficient, the main requirements of the preset learning strategy for the model include: More precise learning rate control: ensure that the model can work effectively at different learning rates. More complex hidden layer dimensions: support larger neural network capacity to process complex data features. Stronger regularization ability: prevent overfitting through higher loss rate, but still maintain the learning effect of the model. Larger batch size processing capability: support large batch training to improve training efficiency and maintain gradient stability. In the fractional regression model, the learning rate needs to remain relatively stable, because the optimization goal of fractional regression is different from that of classical regression, and the learning rate needs to be moderate to avoid too fast convergence. DeepAR has higher requirements for learning rate because the time dependency in the autoregressive structure needs to be adjusted through appropriate pace. The larger the applicability coefficient, the larger the model capacity is needed to capture complex time dependencies. The capabilities of LSTM are suitable for this scenario, and the batch value The greater the impact, the greater the applicability coefficient, which means that the data volume is large and the complexity is high, requiring stronger parallel processing capabilities. This is exactly the strength of Transformer. The applicability coefficient can, to a certain extent, reflect the degree of adaptability of the deep learning algorithm set, namely fractional regression, DeepAR autoregressive recursion, LSTM, and each model in the Transformer model, to the preset learning strategy.

[0077] Fuzzy logic refers to: taking the selection coefficient and applicability coefficient when building the avalanche probability prediction model as input variables, taking the type of learning algorithm used when building the avalanche probability prediction model as the output variable, fuzzifying the input variables, converting the values ​​of the input variables into fuzzy sets, fuzzifying the output variables, converting the output variables into fuzzy sets, formulating fuzzy rules, describing the degree of adaptation of the learning algorithm under different data type combinations, and reasoning the fuzzified input variables through fuzzy rules to obtain the type of learning algorithm used when building the avalanche probability prediction model.

[0078] The input variables are the selection coefficient and the applicability coefficient. The selection coefficient represents the complexity of the data, such as dimension, time series length, data variance, etc. The applicability coefficient represents the adaptability of the model parameters, such as learning rate, hidden layer dimension, loss rate, and batch size. The output variable is the type of learning algorithm. The model may choose different types of learning algorithms, such as fractional regression, DeepAR, LSTM, Transformer, etc. Each algorithm adapts to different task requirements and data types.

[0079] Fuzzify input variables: First, fuzzify the selection coefficient and the applicability coefficient. The fuzzification process is to convert specific numerical values ​​into fuzzy sets. For example, the value of the selection coefficient can be fuzzified into three levels: "low", "medium", and "high"; the applicability coefficient can also be divided into similar levels. For example, when the value of the selection coefficient is small, it is fuzzified to "low". When the value of the applicability coefficient is large, it is fuzzified to "high".

[0080] Fuzzy output variables: The output variable is the type of learning algorithm, and its adaptability can also be expressed through fuzzification. For example, different learning algorithms can be divided into three categories: "suitable", "relatively suitable", and "unsuitable", describing the adaptability of different algorithms in the current data scenario.

[0081] Formulate fuzzy rules: The formulation of fuzzy rules is the core part of fuzzy logic reasoning. It describes the relationship between input variables and output variables. For example: If the selection coefficient is "high" and the applicability coefficient is "high", then select the Transformer model, "suitable". If the selection coefficient is "low" and the applicability coefficient is "medium", then select the Fractional Regression Model, "relatively suitable". If the selection coefficient is "medium" and the applicability coefficient is "low", then select the DeepAR model, "suitable".

[0082] After the input variables are fuzzified, reasoning is performed based on fuzzy rules. The result of reasoning is the fuzzified output variables, which are defuzzified to obtain the specific learning algorithm type. For example: Assume there is a set of input variables: The selection coefficient is 6, indicating that the data complexity is medium. The applicability coefficient is 8, indicating that the applicability of the current model is strong. Through fuzzification: The selection coefficient is fuzzified to "medium". The applicability coefficient is fuzzified to "high". According to the fuzzy rules, if the selection coefficient is "medium" and the applicability coefficient is "high", the LSTM model is selected, "suitable". After reasoning, the final result is that the LSTM model is considered to be the most suitable learning algorithm type for the current data and conditions.

[0083] Judging whether the effective impact range of artificial intervention for avalanche meets the engineering requirements means: adjusting the probability cloud map display threshold according to different risk control requirements, and grid points greater than the threshold are considered to have avalanches after blasting. In artificial intervention for avalanches, different engineering requirements have different tolerances for risks. By adjusting the display threshold of the probability cloud map, it is possible to decide which areas require higher attention and prevention according to actual engineering requirements. For example, for high-risk areas, a lower threshold may need to be set so that even a smaller avalanche risk will be listed as a key area that needs to be handled. The adjustment of the threshold is based on specific risk control requirements. Engineers can flexibly adjust the threshold according to the safety standards or special needs of different projects to ensure that the intervention effect can meet the safety requirements of a specific environment. A higher threshold can reduce unnecessary blasting, while a lower threshold can increase the blasting range to ensure that all potential avalanche risk areas are covered. By adjusting the threshold, engineers can ensure that the effect of blasting is neither excessive (resulting in waste of resources) nor insufficient (resulting in potential risks not being handled). Grid points greater than the set threshold mean that the grid has a higher probability of avalanche and requires manual intervention to reduce the risk. This method can help optimize the design of blasting plans, so that manual intervention can effectively and economically control risks. Grid points with a risk greater than the threshold will be considered high-risk areas, which can be displayed through a visual probability cloud map, allowing engineering decision makers to intuitively understand the potential areas for avalanches and take reasonable countermeasures. This data-based and prediction-based approach ensures the scientificity and effectiveness of blasting.

[0084] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0085] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0086] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0087] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0088] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for predicting the effective impact range of artificial intervention in avalanches based on deep learning, characterized in that: The following steps are involved: S1. Collect the three-dimensional terrain data and high snow volume data of the snow mountain slope through the UAV, and build the slope terrain database and the snow volume database of the preset range; S2. Constructing an explosive air blast near-field overpressure database, and labeling the data in the slope terrain database, the snow body database, and the explosive air blast near-field overpressure database; S3. Based on the deep learning algorithm, the data set obtained after labeling is input into the avalanche probability prediction model for construction; S4. Use the constructed avalanche probability prediction model to perform traversal prediction on all preset ranges of the snow mountain slope to form an avalanche probability cloud map of the entire snow mountain slope; S5, projecting the probability cloud map of the snow mountain slope obtained in S4 onto the overall three-dimensional model of the snow mountain slope, adjusting the probability cloud map display threshold according to different risk management and control requirements, and judging whether the effective influence range of avalanche artificial intervention meets the engineering requirements according to the projection display results; In the constructed explosive air explosion near-field overpressure database, the calculation formula for the air shock wave overpressure data of explosive air explosion is as follows: ； is the equivalent attenuation distance of air shock wave pressure, It is the overpressure data of air shock wave pressure of explosive air explosion; Equivalent attenuation distance of air shock wave pressure The calculation formula is: ; R is the shock wave pressure attenuation distance, W is the mass of explosive; The deep learning algorithm includes fractional regression, DeepAR autoregressive recursion, LSTM, and Transformer models. By extracting features from the data set obtained after labeling, and using the preset learning strategy, the selection coefficient and the applicability coefficient are generated respectively. Then, the corresponding unique deterministic algorithm model is selected from the fractional regression, DeepAR autoregressive recursion, LSTM, and Transformer models through fuzzy logic to construct the avalanche probability prediction model. The logic for obtaining the selection coefficient is: Get the data dimensions of each type of data extracted from the data set obtained after labeling , time series length , the variance of each type of data , and the time series correlation of various types of data , i represents the data type, and then substitute it into the selection coefficient calculation formula: ; By calculating the data dimensions of each type of data The product of is, By calculating the variance of each type of data The average value of By calculating the time series correlation of various types of data The average value of is the selection coefficient, , , , All are preset non-zero scale factors; The logic for obtaining the applicable coefficient is: Get the learning rate from the preset learning strategy , hidden layer dimension , loss rate , batch value , and then substitute it into the applicable coefficient calculation formula: ; is the applicable coefficient, , , , All are preset non-zero influence coefficients.

2. The method for predicting the effective impact range of artificial intervention in avalanches based on deep learning according to claim 1 is characterized in that: The preset range refers to: dividing the entire snow mountain slope into 5m×5m grids, and each grid obtained after division is the preset range.

3. The method for predicting the effective impact range of artificial intervention in avalanches based on deep learning according to claim 2 is characterized in that: The content of the labeling is: the probability value of avalanche occurring in the snow body on the slope of the snow mountain under the condition of fixed amount of explosives and blasting points.

4. The method for predicting the effective impact range of artificial intervention in avalanches based on deep learning according to claim 3 is characterized in that: Fuzzy logic refers to: The selection coefficient and applicability coefficient in the decision-making of building the avalanche probability prediction model are taken as input variables, and the type of learning algorithm used in the construction of the avalanche probability prediction model is taken as the output variable. The input variables are fuzzified and the values ​​of the input variables are converted into fuzzy sets. The output variables are fuzzified and the output variables are converted into fuzzy sets. Fuzzy rules are formulated to describe the adaptability of the learning algorithm under different data type combinations. The fuzzified input variables are inferred through fuzzy rules to obtain the type of learning algorithm used in the construction of the avalanche probability prediction model.

5. The method for predicting the effective impact range of artificial intervention in avalanches based on deep learning according to claim 4 is characterized in that: Judging whether the effective impact range of artificial intervention in avalanches meets engineering requirements means: adjusting the probability cloud map display threshold according to different risk management requirements, and grid points greater than the threshold are considered to have occurred avalanches after blasting.

Images