Valley debris flow disaster intelligent early warning method based on multi-index fusion and dynamic optimization

Abstract

The invention discloses a valley-type debris flow disaster intelligent early warning method based on multi-index fusion and dynamic optimization, which comprises the steps of data preprocessing, multivariable joint probability modeling, dynamic confidence optimization and threshold updating, and real-time monitoring and model updating, and aims to realize accurate early warning and quick response of valley-type debris flow disasters. According to the method, rainfall, muddy water level height, soil moisture content and other multi-source data are fused, and mathematical modeling and intelligent early warning of a debris flow disaster-causing mechanism are achieved through the steps of data preprocessing, multivariable joint probability modeling, dynamic confidence optimization, real-time monitoring, model updating and the like. According to the method, the Gumbel Copula function suitable for tail joint extreme event modeling is used for constructing the joint distribution model, the method is particularly suitable for valley debris flow disaster early warning scenes with complex terrains and diversified induction mechanisms, and disaster recognition and risk response capabilities in complex geological environments can be remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of geological disaster monitoring and early warning, and particularly relates to a multi-index fusion and dynamic optimization intelligent early warning method for valley-type debris flow disasters. BACKGROUND

[0002] As a sudden, destructive and widely influential geological disaster form, valley-type debris flow mainly occurs in steep valleys, river valleys and other narrow topographies, is often induced by heavy rainfall, snowmelt or sudden geological disturbance, and has obvious stage evolution characteristics of "pregnancy-disaster-occurrence".

[0003] The traditional debris flow early warning method mainly relies on historical monitoring data and expert experience, such as the early rainfall and triggering rainfall criterion method or the unit runoff triggering model, which has the following limitations: strong dependence on a single factor, lack of fusion and utilization of multi-source geophysical indexes (such as soil water content characteristics), weak regional adaptability, high dependence on historical disaster samples, and no cross-regional migration ability; unable to dynamically update, and unable to adapt to the rapid dynamic evolution of debris flow disaster conditions.

[0004] In recent years, with the development of Internet of Things, remote sensing and other technologies, debris flow early warning has begun to introduce various automatic monitoring schemes such as mud water level height monitoring and soil water content monitoring, and has initially realized informationization and data-driven early warning process. However, the existing methods are mostly static experience thresholds or single-factor models, lacking in-depth mining of the correlation mechanism between multi-source data, making it difficult to effectively identify the risk evolution under the joint triggering condition. At the same time, the existing methods lack dynamic updating ability, and cannot optimize the model structure or early warning threshold through real-time data, which cannot guarantee the accuracy and timeliness of the early warning. Therefore, there is an urgent need for a debris flow disaster early warning system that integrates multiple indexes, has real-time updating ability and self-adaptive decision mechanism. SUMMARY

[0005] To solve the above technical problems, the purpose of the present application is to provide a multi-index fusion and dynamic optimization intelligent early warning method for valley-type debris flow disasters, which integrates multi-source data such as rainfall, mud water level height and soil water content, constructs a joint distribution model through Copula function, and adjusts the dynamic threshold value by using ROC curve and exceeding probability combination. Finally, the model and threshold value are dynamically updated by combining real-time data acquisition, and the blue, yellow, orange and red four-level warning levels are output, thereby significantly improving the precision, timeliness and engineering applicability of the early warning system.

[0006] The purpose of the present application is achieved by the following steps: S100, data preprocessing, including time series alignment, data interpolation and noise reduction, to ensure the quality and time consistency of multi-source monitoring data: S101, timing alignment: using a uniform step length to resample rainfall, silt level, and soil moisture, so that multi-index monitoring data at each observation time point is comparable and fusible; S102, data interpolation and noise reduction: data interpolation uses piecewise Hermite interpolation (PCHIP) to fill in missing values, maintaining the smoothness and physical consistency of the time series; noise reduction uses median filtering to suppress measurement noise; S200, multivariate joint probability modeling, including edge distribution fitting, Copula function modeling, and joint exceedance probability calculation, to build a multivariate joint probability model of debris flow disasters, to characterize the joint extreme event probability between multiple indicators: S201, edge distribution fitting: use maximum likelihood estimation to fit the edge distribution of each variable, and select the distribution according to the Akaike information criterion (AIC); Akaike information criterion (AIC) is an index for evaluating the goodness of fit of a statistical model. It introduces a penalty term based on maximum likelihood to prevent model overfitting: where k is the number of model estimation parameters, is the likelihood function value under maximum likelihood estimation; the smaller the AIC value, the better the model; S202, Copula function modeling: select Gumbel Copula model for modeling, construct joint distribution through Sklar theorem, and estimate dependence intensity using Inversion of Kendall's τ method; S203, joint exceedance probability calculation: calculate the multivariate joint trigger probability P c , the formula is: where X i represents each single indicator, is the single indicator warning threshold, is the cumulative distribution value of the edge distribution function, and C(⋅) is the Gumbel Copula function; S300, dynamic confidence optimization and threshold update: based on ROC curve and Youden index analysis to evaluate the discrimination performance, and based on joint trigger probability construction and multi-level risk division, to realize dynamic threshold adjustment and risk level division: S301, discriminant performance evaluation based on ROC curve and Youden index analysis: the Receiver Operating Characteristic (ROC) curve is constructed for joint probability or single index, and the optimal discriminant threshold is calculated to achieve the optimal balance between maximizing the identification rate TPR and minimizing the false positive rate FPR; Specifically, the ROC curve is a function image plotted with the false positive rate FPR as the abscissa and the identification rate TPR as the ordinate to evaluate the advantages and disadvantages of different thresholds; its mathematical expression is: Where TP(t), FP(t), TN(t), FN(t) represent the number of true cases, false positives, true negatives and false negatives at the discriminant threshold t of joint probability, respectively; The expression of Youden index is: Where t is the discriminant threshold of joint probability; S302, joint trigger probability construction: the dynamic fusion probability is defined by combining the marginal exceedance probability and the Copula joint probability, and the formula is: Where P c The Copula function models the joint exceedance probability, d is the dimension of the variable, p j is the marginal exceedance probability of the jth index, F j (x j ) represents the cumulative distribution of the current value, and the weight a is dynamically adjusted through Bayesian optimization, and the formula is: Where f(a) represents the prediction accuracy when the current fusion weight is a, and f(a + ) represents the optimal accuracy in the entire Bayesian optimization process; S303, multi-level risk classification: according to P final The initial warning level (blue, yellow, orange, red four levels) is divided, and the threshold is adjusted through S301 to construct the ROC curve and Youden index, and the optimal threshold of the optimal balance of TPR and FPR is obtained; S400, real-time monitoring and model updating, through real-time monitoring data acquisition, the real-time updating of model parameters and threshold is realized, and then the intelligent early warning of debris flow disaster is realized: S401, real-time data acquisition: real-time acquisition of rainfall, mud water level and soil moisture content monitoring data; S402, model updating and parameter optimization: after performing S100 data preprocessing on real-time monitoring data, merge with existing data, re-execute S200 multivariate joint probability modeling and S300 dynamic confidence optimization and threshold updating steps to realize dynamic updating of multivariate joint probability model, optimize model confidence parameters through real-time adjustment of joint probability distribution, and simultaneously update risk thresholds of each early warning index; S403, real-time early warning level determination: based on the updated model, substitute the current monitoring data to calculate the new joint trigger probability, and accordingly obtain the early warning level (blue, yellow, orange, red four levels) at the current time, realizing dynamic early warning output of disaster risk.

[0007] Preferably, the S102 step of segmented Hermite interpolation is specifically to set the current interpolation point as (x, y), and use the following interpolation formula for calculation: ; Wherein, y k is the known data point, d k is the derivative approximation value, PCHIP will maintain monotonicity to avoid oscillation phenomenon, b k and c k are polynomial coefficients determined by interpolation conditions, .

[0008] Preferably, the formula of maximum likelihood estimation method in S201 step is: ; Wherein, is the parameter estimation value of the jth index edge distribution, n is the sample number, θ is the parameter vector to be estimated, is the probability density function of the jth variable under the parameter θ.

[0009] Preferably, the S202 step of Copula function modeling is specifically: Build joint distribution F(x1,...,x d ), use Sklar theorem: ; Select Gumbel Copula function suitable for tail joint extreme event modeling: ; Wherein, u j represents the cumulative distribution function of single-index edge distribution, natural logarithm ln is used to strengthen tail dependence modeling, parameter θ reflects the dependence strength between variables, and parameter θ is estimated using Inversion of Kendall's τ method, and the formula is: .

[0010] Compared with the prior art, the present application has the following technical effects: 1. Based on Gumbel Copula function joint modeling, fully reflects the correlation between rainfall, mud level, soil moisture content and other multiple indexes, and overcomes the problem of dependence on single threshold value in traditional methods; 2. Dynamic threshold optimization mechanism adapts to complex environmental changes and enhances model robustness; 3. Can be combined with Internet of Things and edge computing to realize low delay and high reliability real-time early warning; 4. Suitable for debris flow high-incidence areas with complex terrain and dramatic data changes, with good engineering practicability and promotion prospect. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 is a flowchart of the method of the present application; Figure 2 is an original data distribution diagram of mud level, soil moisture content and rainfall in Example 1; Figure 3 is a data preprocessing schematic diagram of mud level, soil moisture content and rainfall in Example 1; Figure 4 is a two-dimensional schematic diagram of Gumbel Copula joint modeling in Example 1; Figure 5 is a three-dimensional schematic diagram of Gumbel Copula joint modeling in Example 1; Figure 6 is a ROC curve schematic diagram based on joint probability in Example 1; Figure 7 is a schematic diagram of the relationship between Youden index and discriminant threshold in Example 1. DETAILED DESCRIPTION

[0012] The present application will be further described below in conjunction with the embodiments and drawings, but in no way limits the present application, any transformation or replacement based on the teaching of the present application belongs to the protection scope of the present application.

[0013] Example 1 To verify the effectiveness and applicability of the present application, the following is a typical valley-type debris flow gully in Yunnan area as an example, the specific implementation process is given; The gully is located in the upper reaches of Lancang River in Hengduan Mountains, with large terrain elevation difference, long and narrow channel, and annual average rainfall of about 900 mm, with monthly average rainfall of 121.8 mm in flood season (June to September), which is prone to sudden debris flow; The present application deploys multi-source monitoring nodes in the flood season, including: rain gauge (resolution 0.2 mm, sampling period 10 minutes), mud water level gauge (range 0~15 m, accuracy ±5 mm); The overall collection period is from June to October, and 14813 groups of effective data records are obtained; This embodiment only uses rainfall, mud water level and soil moisture content in June 2025 as an example for description, and the original data distribution diagram is shown in the accompanying Figure 2 ; As shown in the accompanying Figure 1 , the multi-index fusion and dynamic optimization valley-type debris flow disaster intelligent early warning method of the present embodiment comprises the following steps: S100, data preprocessing, specifically including: S101, time series alignment: using a uniform step to resample rainfall, mud water level and soil moisture content, so that multi-index monitoring data have comparability and fusion at each observation time point; S102, data interpolation and noise reduction: data interpolation uses piecewise Hermite interpolation (PCHIP) to fill in missing values, maintaining the smoothness and physical consistency of the time series; Noise reduction uses median filtering to suppress measurement noise; Wherein, the piecewise Hermite interpolation is to set the current interpolation point as (x, y), and the following interpolation formula is used for calculation: ; Wherein, y k is the known data point, d k is the derivative approximation value, PCHIP maintains monotonicity to avoid oscillation phenomenon, b k and c k are polynomial coefficients determined by interpolation conditions, ; The data after preprocessing is shown in the accompanying Figure 3 ; S200, multi-variable joint probability modeling: S201, edge distribution fitting: the maximum likelihood estimation method is used to fit the edge distribution of each variable respectively, and the formula of the maximum likelihood estimation method is: ; Wherein, is the parameter estimation value of the edge distribution of the jth index, n is the sample number, and θ is the estimated parameter vector, The probability density function of the jth variable under the parameter θ; Distribution selection is based on Akaike information criterion (AIC), which is an index for evaluating the goodness of fit of a statistical model. It introduces a penalty term based on maximizing the likelihood to prevent model overfitting: Where k is the number of estimated parameters of the model, The smaller the AIC value, the better the model; S202, Copula function modeling: Gumbel Copula model is selected for modeling, joint distribution is constructed through Sklar theorem, and dependence intensity is estimated by Inversion of Kendall's τ method, the results are shown in Figs. 1 and 2 of the accompanying drawings; Copula function modeling is specifically: Figure 4 , and Figs. 1 and 2 of the accompanying drawings; Figure 5 Construct a joint distribution F(x1,...,x d ) using Sklar theorem: ; Valley-type debris flow is a typical extreme induced geological disaster, and Gumbel Copula function suitable for tail joint extreme event modeling is selected: ; Where u j represents the cumulative distribution function of the single-index marginal distribution, the natural logarithm ln is used to strengthen the tail dependence modeling, the parameter θ reflects the dependence intensity between variables, and the parameter θ is estimated by Inversion of Kendall's τ method, the formula is: ; S203, joint exceedance probability calculation: set the threshold value as x *} = ( x 1 * ,..., x d * ), then the multivariate joint triggering probability P c , the formula is: Where X i represents each single index, is the single-index warning threshold, is the cumulative distribution value of the marginal distribution function, and C(⋅) is the Gumbel Copula function; S300, dynamic confidence optimization and threshold updating, as shown in Figs. 3 and 4 of the accompanying drawings​Figure 6 , attached Figure 7 as shown, specifically comprising: S301, discriminant performance evaluation based on ROC curve and Youden index analysis: construct Receiver Operating Characteristic (ROC) curve for joint probability or single index, calculate the optimal discriminant threshold, to achieve the optimal balance between maximizing the identification rate TPR and minimizing the false positive rate FPR; ROC curve is a function image with false positive rate FPR as abscissa and identification rate TPR as ordinate to evaluate the advantages and disadvantages of different thresholds; its mathematical expression is: Where TP(t), FP(t), TN(t), FN(t) represent the number of true cases, false positives, true negatives and false negatives at the joint probability discriminant threshold t, respectively; The expression of Youden index is: Where t is the discriminant threshold of joint probability; S302, joint trigger probability construction: define dynamic fusion probability combining marginal exceedance probability and Copula joint probability, the formula is: Where P c , Copula function modeling joint exceedance probability, d is the dimension of the variable, p j is the marginal exceedance probability of the jth index, F j (x j ) represents the cumulative distribution of the current value, and the weight α is dynamically adjusted through Bayesian optimization, the formula is: Where f(α) represents the prediction accuracy rate when the current fusion weight is α, f(α + ) represents the optimal accuracy rate in the entire Bayesian optimization process; S303, multi-level risk classification: first, according to the joint trigger probability P final , the warning level is divided into blue (60%), yellow (80%), orange (90%) and red (95%); by adjusting the threshold value in S301, construct ROC curve and Youden index, get the optimal threshold value of TPR and FPR optimal balance: blue (59%), yellow (78.7%), orange (88.5%), red (93.4%); S400, real-time monitoring and model updating: S401, real-time data acquisition: real-time acquisition of rainfall, silt level and soil moisture monitoring data; for example, the current rainfall size (unit: mm / h), the current silt level (unit: m) and the current soil moisture content (%) are collected every ten minutes; Table 1 Partial monitoring data interception reference S402, model updating and parameter optimization: after the real-time monitoring data is preprocessed by S100, it is combined with the existing data, and the steps of S200 multivariate joint probability modeling and S300 dynamic confidence optimization and threshold updating are re-executed to realize dynamic updating of the multivariate joint probability model. Through real-time adjustment of the joint probability distribution, the model confidence parameters are optimized, and the risk thresholds of each early warning index are updated synchronously; S403, real-time early warning level determination: based on the updated model, the new joint trigger probability is calculated by substituting the current monitoring data, and the early warning level (blue, yellow, orange, red four levels) at the current time is obtained accordingly, realizing the dynamic early warning output of disaster risk; For example, under the condition of monitoring the real-time rainfall of 6.3mm / h, the silt level of 0.47m and the soil moisture content of 18.51%, the joint trigger probability P final is 87.2% by model calculation. According to the dynamic early warning level at this time, it meets the yellow warning (78.7%) and does not meet the orange warning (88.5%), so it is determined as yellow warning.

Claims

1. An intelligent early warning method for valley debris flow disasters based on multi-index fusion and dynamic optimization, characterized by The following steps are involved: S100, data preprocessing: S101, Time Series Alignment: Use a unified step size to resample rainfall, mud water level, and soil moisture content to make multi-indicator monitoring data comparable and fusible at each observation time point; S102, Data interpolation and noise reduction: Data interpolation uses piecewise Hermite interpolation to fill missing values ​​and maintain the smoothness and physical consistency of the time series; noise reduction uses median filtering to suppress measurement noise; S200, Multivariate Joint Probability Modeling: S201, marginal distribution fitting: Use maximum likelihood estimation to fit the marginal distribution of each variable, and select the distribution according to the Akaike information criterion; S202, Copula function modeling: The Gumbel Copula model is used for modeling, the joint distribution is constructed through Sklar theorem, and the dependence strength is estimated using the Inversion of Kendall's τ method; S203, joint exceedance probability calculation: Calculate the multivariable joint trigger probability P according to the set threshold c , the formula is: Among them, X i Represents each single indicator, is the single indicator warning threshold, is the cumulative distribution value of the marginal distribution function, C(⋅) is the Gumbel Copula function; S300, dynamic confidence optimization and threshold update: S301. Evaluate the discrimination performance based on ROC curve and Youden index analysis: Construct an ROC curve for the joint probability or single index, and calculate the optimal discrimination threshold to achieve the best balance between maximizing the recognition rate TPR and minimizing the false alarm rate FPR; the Youden index expression is: Among them, t is the discrimination threshold of joint probability; S302, joint trigger probability construction: Combine the edge exceeding probability and the copula joint probability to define the dynamic fusion probability, the formula is: Among them, P c That is, the Copula function models the joint exceedance probability, d is the dimension of the variable, p j is the marginal exceedance probability of the jth indicator, F j (x j ) represents the cumulative distribution of the current value, and the weight α is dynamically adjusted through Bayesian optimization. The formula is: Among them, f(α) represents the prediction accuracy of the current fusion weight α, f(α + ) represents the optimal accuracy of the entire Bayesian optimization process; S303, multi-level risk classification: based on P final The initial warning level is divided, and then the threshold is adjusted through S301 to construct the ROC curve and Youden index to obtain the optimal threshold for the best balance between TPR and FPR; S400, real-time monitoring and model updating: S401, real-time data collection: real-time collection of rainfall, mud water level and soil moisture monitoring data; S402, model update and parameter optimization: After performing S100 data preprocessing on the real-time monitoring data, merge it with the existing data, and re-execute S200 multivariate joint probability modeling and S300 dynamic confidence optimization and threshold update steps to dynamically update the multivariate joint probability model. Through real-time adjustment of the joint probability distribution, optimize the model confidence parameters and synchronously update the risk thresholds of each early warning indicator; S403. Real-time warning level determination: Based on the updated model, the current monitoring data is substituted into the new joint trigger probability to calculate the new joint trigger probability, and the warning level at the current moment is obtained accordingly.

2. The multi-index fusion and dynamic optimization intelligent early warning method for valley-type debris flow disasters according to claim 1 is characterized in that In step S102, the segmented Hermite interpolation is performed by setting the current interpolation point to (x, y) and using the following interpolation formula for calculation: ； Among them, y k is a known data point, d k is the derivative approximation, b k and c k are polynomial coefficients, determined by the interpolation conditions, .

3. The multi-index fusion and dynamic optimization intelligent early warning method for valley-type debris flow disasters according to claim 1 is characterized in that The formula for the maximum likelihood estimation method in step S201 is: ； in, is the parameter estimate of the marginal distribution of the jth indicator, n is the number of samples, θ is the estimated parameter vector, is the probability density function of the jth variable under parameter θ.

4. The multi-index fusion and dynamic optimization intelligent early warning method for valley-type debris flow disasters according to claim 1 is characterized in that The specific modeling of the Copula function in step S202 is: Construct the joint distribution F(x1,...,x d ), using Sklar's theorem: ； The Gumbel Copula function suitable for modeling tail joint extreme events is selected: ； Among them, u j The cumulative distribution function represents the marginal distribution of a single indicator. The natural logarithm ln is used to strengthen the tail dependence modeling. The parameter θ reflects the dependence strength between variables. The parameter θ is estimated using the Inversion of Kendall's τ method, and the formula is: .

Images