Geological disaster risk assessment method and system based on multi-source data analysis

Abstract

The invention discloses a geological disaster risk assessment method and system based on multi-source data analysis, relates to the technical field of geological disaster prevention and control, quantifies the uncertainty of multi-source data according to an evidence theory, updates the credibility weight in a Bayesian manner, and adapts the data quality through an attention dynamic neural network. The method has the advantages that uncertainty characteristics of multi-source data are quantified through the evidence theory, the credibility weight of the data source is corrected in combination with a Bayesian dynamic updating mechanism, the reliability of the data source is improved, the reliability of the data source is improved, and the reliability of the data source is improved. Then, a dynamic neural network based on an attention mechanism is utilized to adaptively adjust a structure and a decision boundary according to data uncertainty, robust optimization is introduced to suppress low-credibility data error conduction, and meanwhile key influence factors are calibrated and marked through credibility intervals; the defects that in the prior art, data uncertainty is ignored, an evaluation model cannot adapt to data quality fluctuation, only a single result is output, and no reliable basis exists are effectively overcome.

Description

Technical Field

[0001] This invention relates to the field of geological disaster prevention and control technology, specifically to a geological disaster risk assessment method and system based on multi-source data analysis. Background Technology

[0002] Geological disasters are major natural disasters that threaten the lives and property of residents in mountainous areas and damage infrastructure such as transportation and water conservancy. With the intensification of global climate change and the increase in human engineering activities, the frequency, scope of impact and degree of damage of geological disasters are on the rise. Conducting scientific and efficient geological disaster risk assessment has become a core part of disaster prevention and mitigation work, and it has irreplaceable key significance for identifying disaster hazards in advance, formulating accurate risk avoidance strategies and reducing disaster losses. Existing geological hazard risk assessment technologies mainly rely on single or limited data sources combined with fixed model analysis to determine risks, which has certain shortcomings. First, they ignore the uncertainty characteristics of multi-source data, the assessment model cannot adapt to data quality fluctuations, and only outputs a single risk result, lacking reliable evidence. Second, it is difficult to trace the source of assessment data, making it impossible to continuously optimize the assessment accuracy based on historical results, and it lacks efficient decision-making and collaborative capabilities supported by data. To address these shortcomings, we propose a geological hazard risk assessment method and system based on multi-source data analysis. Summary of the Invention

[0003] The purpose of this invention is to provide a geological hazard risk assessment method and system based on multi-source data analysis.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a geological hazard risk assessment system based on multi-source data analysis, comprising a risk assessment system, wherein the risk assessment system includes the following modules: Data uncertainty quantification module: Based on evidence theory, a multi-source data credibility assessment framework is constructed to quantify the uncertainty characteristics of each data source, and the credibility weight of each data source is corrected by combining a Bayesian dynamic update mechanism, outputting the quantification results and dynamic weights. Robust Model Adaptive Reconstruction Module: Constructs a dynamic neural network based on an attention mechanism, adjusts the structure and decision boundary according to the uncertainty of multi-source data, introduces robust optimization to suppress errors, and outputs basic results for risk assessment; Risk Result Credibility Calibration Module: Combining data uncertainty and model parameters, calculates the credibility range of risk results, identifies key influencing factors, and outputs the final disaster risk level and credibility information; Disaster risk level output module: adopts - Level V graded output, simultaneously displaying risk level, credibility range and key influencing factors, providing intuitive reference for decision-making; Data storage visualization module: store the whole process data and evaluation results, and display through multi-dimensional visualization interface, support data backtracking and trend analysis; Model parameter feedback optimization module: analyze historical evaluation errors, divide optimization priorities, dynamically optimize quantitative parameters and model initial parameters, and improve system evaluation accuracy and stability.

[0005] As a further scheme of the present application: the risk assessment system further comprises a multi-source data acquisition module and a data preprocessing module, the original data acquired by the multi-source data acquisition module specifically includes topography, meteorology, geological structure, hydrology, remote sensing image, sensor monitoring and historical record type data, and the acquisition mode is divided into real-time acquisition and timing acquisition, wherein the sensor monitoring type data adopts a real-time acquisition frequency of 1-5 min / time, the remote sensing image type data adopts a timing acquisition frequency of 1-3 days / time, and the historical data is imported offline.

[0006] As a further scheme of the present application: the data preprocessing module uniformly converts all original data into WGS-84 coordinates, international units and JSON format, performs missing value completion processing, based on K nearest neighbor algorithm, performs redundant data elimination processing, eliminates redundant data with a similarity of more than 95%, performs noise filtering processing, adopts db4 wavelet denoising, and outputs regular data.

[0007] As a further scheme of the present application: the data uncertainty quantification module constructs an identification framework based on the credibility evaluation framework of evidence theory {credible, incredible, uncertain}, identification framework Contains 3 single-element subsets, 3 double-element subsets, 1 full-set subset and 1 empty-set subset, wherein the basic probability assignment value of the empty-set subset is 0, based on calculating the basic probability assignment function of all non-empty subsets All non-empty subsets , quantize the trust degree of each data source to each proposition, and the basic probability assignment function Satisfies: ; The Bayesian dynamic updating mechanism is based on combining the prior credibility weight of the previous moment and the real-time observation data likelihood of the current moment, and iteratively correcting to obtain the posterior credibility weight, and the environmental parameters associated in real time in the updating process include temperature change value, humidity change value, electromagnetic interference intensity and sensor working time length; Based on the posterior credibility weight obtained by the above iterative correction, the credibility level of the data source is divided: when the posterior credibility weight of a certain data source is lower than a preset threshold, the data set corresponding to the data source is defined as low credibility data.

[0008] As a further scheme of the present application: the dynamic neural network of the robust model adaptive reconstruction module comprises an input layer, an attention feature extraction layer, a dynamic hidden layer, a robust optimization layer and an output layer, the attention feature extraction layer is positively correlated with the reliability weight of the data source, the number of neurons of the dynamic hidden layer is adaptively adjusted according to the average level of uncertainty of the multi-source data, the average level of uncertainty is calculated by the average of the uncertainty entropy values of all data sources, and the uncertainty entropy value calculation formula is as follows: ; Among them, is the uncertainty entropy value of the data source, the value range is , , the data source is completely reliable, only = 1, and the rest , is any non-empty subset of the identification framework ; Based on the basic probability assignment function of each data source, the uncertainty entropy value of the data source is calculated, and the uncertainty level of the data source is divided according to the uncertainty entropy value; when the uncertainty entropy value of a certain data source is higher than a preset threshold, the data source is defined as a high-uncertainty data source When the average is less than 0.3, the number of hidden layer neurons is set to 128, when the average is between 0.3 and 0.7, the number of neurons is set to 256, and when the average is greater than 0.7, the number of neurons is set to 512, and the number of neurons is always not more than 512. The robust optimization layer is based on the introduction of a robust optimization regularization term to impose a penalty on the model parameters corresponding to low-reliability data, suppress error propagation, and the output layer uses activation function, outputting the probability distribution of each risk level as the basic result of risk assessment.

[0009] As a further scheme of the present application: the reliability interval calculation of the risk result reliability calibration module is based on the normal distribution confidence interval estimation method, the confidence level is set to 90%-99%, the total variance is calculated based on the variance caused by data uncertainty and the variance caused by model robustness, and the upper and lower limits of the interval are determined in combination with the standard normal distribution quantile; The key influencing factors include the type and corresponding uncertainty entropy value of the high-uncertainty data source, the neuron adjustment amplitude of the model dynamic hidden layer, the weight coefficient of the robust optimization regularization term, and the type and intensity of real-time environmental interference factors, wherein the real-time environmental interference factors include the influence of temperature change on the sensor, the influence of electromagnetic interference on data transmission and the shielding influence of rainfall on remote sensing images.

[0010] As a further scheme of the present application: the data storage visualization module adopts The distributed database stores data, including a raw data set, a pretreated data set, an uncertainty quantization result set, a model reconstruction parameter set, a risk assessment basis result set, a final assessment result set and a historical error data set. The visualization interface includes a data trend curve graph, an uncertainty heat map, a risk level spatial distribution graph, a confidence interval column chart and a model parameter change line chart, and supports a user to query corresponding data through a time range, a space range, a data type and a risk level filtering condition.

[0011] As a further scheme of the present application: the model parameter feedback optimization module calculates the deviation of the historical assessment result and the actual disaster occurrence condition by using a mean square error index, and the calculation formula is as follows: ; Among them, is the quantized value of the actual disaster level, level 1 = 1, level 2 = 2, level 3 = 3, level 6 = 4, level 5 = 5, is the quantized value of the historical assessment result, is the number of historical assessment samples; The optimization priority division standard is: when > 0.8, triggering high-priority optimization, and the optimization iteration number is set to 500, when 0.4-0.8, triggering medium-priority optimization, and the iteration number is set to 300, when < 0.4, triggering low-priority optimization, and the iteration number is set to 100, and the convergence condition of optimization iteration is that the parameter change amount of adjacent two iterations < 10 -5 , and the updated parameter is used to replace the original parameter to realize system update, and the update frequency is consistent with the data timing acquisition frequency.

[0012] In addition, the present application also provides a geological disaster risk assessment method based on multi-source data analysis, and the risk assessment method comprises the following methods: Step 1: Collect multi-dimensional raw data required for geological disaster assessment, covering four categories of geological environment, meteorological conditions, monitoring and sensing, and historical records; Step 2: Perform format standardization, missing value completion, redundancy elimination and noise filtering processing on the raw data, eliminate heterogeneity interference, and output high-quality regular data; Step 3: Construct an identification framework based on evidence theory, quantify the uncertainty characteristics of each data source, correct the confidence weight by combining the Bayesian dynamic updating mechanism, and output the quantization result and dynamic weight; Step 4: Construct a dynamic neural network based on an attention mechanism, adaptively adjust the network structure and decision boundary according to the data uncertainty, introduce robust optimization to suppress errors, and output the risk assessment basis result; Step five, combining data uncertainty and model parameters, calculating risk result confidence interval, marking key influencing factors, obtaining final disaster risk level and confidence information; Step six, adopting I-V grade classification method to display final risk level, confidence interval and key influencing factors; Step seven, storing full-process data and evaluation results, displaying and supporting backtracking analysis through a visual interface, dividing priorities based on historical evaluation errors, and dynamically optimizing quantitative parameters and model initial parameters.

[0013] Compared with the prior art, the beneficial effects of the present application are as follows: 1. The present application quantifies the uncertainty characteristics of multi-source data by evidence theory, combines the Bayesian dynamic updating mechanism to correct the data source confidence weight, and then uses the dynamic neural network based on the attention mechanism to adaptively adjust the structure and decision boundary according to the data uncertainty, introduces robust optimization to suppress the error propagation of low-confidence data, and at the same time, calibrates the key influencing factors through the confidence interval, effectively solves the defects of the prior art that ignores data uncertainty, the evaluation model cannot adapt to data quality fluctuations, only outputs a single result and has no reliable basis, and finally realizes the precision and robustness of geological disaster risk assessment, providing comprehensive reference for disaster prevention and control decision-making with risk level and confidence support; 2. The present application classifies and stores full-process evaluation data through the data storage visualization module and intuitively displays it through a multi-dimensional interface, analyzes historical evaluation errors through the model parameter feedback optimization module, dynamically adjusts quantitative parameters and model initial parameters according to the error level, effectively solves the defects of the prior art that it is difficult to trace the source of evaluation data, cannot continuously optimize the evaluation accuracy according to historical results, and lacks efficient decision-making collaboration ability supported by data, and finally realizes the long-term dynamic improvement of the accuracy of geological disaster risk assessment, and provides complete data support for decision-makers that can be traced and analyzed. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 The figure is a system flowchart in the embodiment of the present application; Figure 2 The figure is a method step diagram in the embodiment of the present application. DETAILED DESCRIPTION

[0015] The specific embodiments of the present application will be further described below in conjunction with the drawings, and it should be noted that the description of these embodiments is used to help understand the present application, but does not constitute a limitation on the present application.

[0016] In addition, the technical features involved in each of the embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0017] Please refer to the attached Figure 1 -Appendix Figure 2 The geological disaster risk assessment system of multi-source data analysis of the application comprises a risk assessment system, which comprises the following modules: Data uncertainty quantification module: based on evidence theory, a multi-source data credibility evaluation framework is constructed to quantify the uncertainty characteristics of each data source, and the credibility weight of each data source is corrected by combining the Bayesian dynamic updating mechanism, and the quantification result and dynamic weight are outputted; Robust model adaptive reconstruction module: a dynamic neural network based on attention mechanism is constructed, the structure and decision boundary are adjusted according to the uncertainty of multi-source data, and robust optimization is introduced to suppress errors, and the basic results of risk assessment are outputted; Risk result credibility calibration module: combining data uncertainty and model parameters, the risk result credibility interval is calculated, the key influencing factors are marked, and the final disaster risk grade and credibility information are outputted; Disaster risk grade output module: adopts -Ⅴ grade classification output, synchronously displays risk grade, credibility interval and key influencing factors, and provides intuitive reference for decision-making; Data storage visualization module: store the whole process data and evaluation results, and display through multi-dimensional visualization interface, support data backtracking and trend analysis; Model parameter feedback optimization module: analyze historical evaluation errors, divide optimization priority, dynamically optimize quantization parameters and model initial parameters, and improve system evaluation accuracy and stability.

[0018] Embodiment one, please refer to the attached Figure 1 -Appendix Figure 2 The multi-source data acquisition module distinguishes real-time and timing acquisition tasks according to preset rules, wherein the acquisition frequency of sensor monitoring type data is set to 1-5 minutes / time, and the sensor working time, environmental temperature and humidity data are recorded synchronously during acquisition, which are used as the basis for subsequent uncertainty quantification. Remote sensing image type data is collected at 1-3 days / time, and after collection, the image cloud cover ratio and shooting time stamp are automatically marked. Historical record type data is supplemented by offline import, and the data format and field integrity need to be verified during import to ensure that the core fields such as disaster occurrence time, grade, cause analysis are included.

[0019] After receiving the data, the data preprocessing module first performs format standardization, converting all data to the WGS-84 coordinate system, with length units standardized to meters, temperature units to degrees Celsius, and rainfall units to millimeters, and the data format standardized to JSON. Next, it performs missing value completion, selecting 10-20 samples with the highest feature similarity to the missing data samples based on the K-nearest neighbor algorithm, and calculating a weighted average to complete the missing values ​​according to the similarity weight. Then, redundant data with feature similarity greater than 95% is removed, and records updated at the acquisition time are retained. Finally, the data is decomposed into three levels of wavelet using the db4 wavelet basis function, and the high-frequency wavelet coefficients are processed using an improved threshold function to filter random noise before outputting regular data to ensure the consistency and high quality of the input data for subsequent modules.

[0020] Example 2, please refer to the appendix. Figure 1 -Appendix Figure 2 The data uncertainty quantification module constructs an identification framework based on evidence theory. {Reliable, Unreliable, Uncertain}, first calculate the data sources in... Basic probability assignment function on all nonempty subsets The initial allocation rules are set in conjunction with the data source type during calculation; Subsequently, the credibility weights are corrected through a Bayesian dynamic update mechanism. During the update, the prior weights of the previous time step are combined with the likelihood of the real-time observed data at the current time step. The likelihood is calculated using a Gaussian distribution fitted to historical credibility data. The resulting dynamic credibility weights are then transmitted to the robust model adaptive reconstruction module. The model's attention feature extraction layer allocates feature extraction priorities according to the rule that "the higher the credibility weight, the greater the attention weight." The dynamic hidden layer adjusts the number of neurons based on the mean uncertainty entropy of all data sources—128 neurons when the mean is <0.3, 256 neurons when the mean is between 0.3 and 0.7, and 512 neurons when the mean is >0.7. At the same time, the robust optimization layer introduces a regularization coefficient to apply L2 regularization penalty to the model parameters corresponding to low credibility data, thereby suppressing error propagation.

[0021] Example 3, please refer to the appendix. Figure 1 -Appendix Figure 2 When the data storage visualization module uses the Hadoop distributed database to store data, it partitions the data according to "time slice + data type". One time slice is generated every 24 hours. Within each slice, the data is stored in categories of "raw data - preprocessed data - evaluation results". It supports fast retrieval by "region + disaster type" - for example, when querying data related to "landslides in the southwestern mountainous areas", the system automatically locates the landslide-specific data partition of the corresponding time slice, and the retrieval response time is controlled within 2 seconds.

[0022] After the model parameter feedback optimization module extracts historical data, it calculates the mean square error by first aligning the historical evaluation results with the actual disaster data in time, ensuring that the evaluation time and the disaster occurrence time differ by no more than 24 hours, then eliminating outliers to avoid interference with error calculation, and then dividing the optimization priority. When performing gradient descent optimization, the "adaptive learning rate" strategy is used: the initial learning rate for high-priority optimization is set to 0.01, and the learning rate is halved every 50 iterations. The initial learning rate for medium-priority optimization is set to 0.005, and the learning rate is halved every 100 iterations. The initial learning rate for low-priority optimization is set to 0.002, and it remains unchanged throughout the process. During optimization, the parameter change is monitored in real time. If the parameter change is less than 10 -5 for three consecutive iterations, it is determined to be converged and the iteration is stopped. The optimized parameters are first verified on the "test data set". If the evaluation accuracy improves by ≥15%, the original parameters are replaced. Otherwise, the parameters are rolled back to the pre-optimization parameters to ensure the reliability of parameter updates.

[0023] Specifically, the multi-source data acquisition module sets data source acquisition priority according to different geological disaster assessment needs: for landslide risk assessment, the priority from high to low is "displacement sensor data > pore water pressure sensor data > rainfall data > remote sensing image data". For debris flow risk assessment, the priority is adjusted to "soil moisture data > river flow data > rainfall data > terrain data". For barrier lake risk assessment, the priority is "remote sensing image data > river flow data > terrain data > historical disaster data". High-priority data sources are given priority in terms of acquisition frequency and data integrity. If there is a conflict in data acquisition (such as overlapping sensor and remote sensing data acquisition times), high-priority data sources are given priority in data acquisition, and low-priority data sources are delayed by no more than 5 minutes in data acquisition to ensure that core data is not missing.

[0024] Specifically, in the data uncertainty quantification module, the calculation of uncertainty entropy H needs to first determine the non-empty subset of the identification framework , which has 7 subsets ({trusted}, {untrusted}, {uncertain}, {trusted, untrusted}, {trusted, uncertain}, {untrusted, uncertain}, {trusted, untrusted, uncertain}). Only subsets with ≠ 0 are involved in the calculation. The calculated value is used in two scenarios: one is to convert it into a data source priori confidence weight, and the other is as a basis for adjusting the number of neurons in the robust model dynamic hidden layer. The mean value of all data source values is calculated. A mean value of <0.3 corresponds to 128 neurons, a mean value of 0.3-0.7 corresponds to 256 neurons, and a mean value of >0.7 corresponds to 512 neurons. The upper limit of the number of neurons is set to 512 to avoid excessive model complexity causing operational delays.

[0025] Specifically, the specific verification logic of "dynamic hidden layer neuron number adjustment" in the robust model self-adaptive reconstruction module is supplemented as follows: after the neuron number adjustment, the system automatically selects the historical data of nearly one month as the verification set, tests the evaluation accuracy of the model under different neuron numbers, if the accuracy after adjustment is improved by >=10%, the new number is retained, if the accuracy decreases or does not change, it is rolled back to the number before adjustment, and the "weight distribution" of the attention feature extraction layer also needs to be optimized in combination with the disaster type: when evaluating landslide, the attention weight of displacement sensor data is additionally increased by 0.1, when evaluating debris flow, the attention weight of soil moisture data is additionally increased by 0.15, so as to ensure that the model is more sensitive to key data of different disasters, and the regularization coefficient of the robust optimization layer is dynamically adjusted according to the "data uncertainty mean value" - when the mean value < 0.3 = 0.2, when 0.3-0.7 = 0.5, when > 0.7 = 0.8, to avoid excessive or insufficient regularization.

[0026] Working principle: Firstly, the multi-source data acquisition module acquires multi-dimensional original data such as geological environment and meteorology, the data preprocessing module outputs regular data after standardization, missing value completion, redundancy removal and denoising, the data uncertainty quantification module quantifies the uncertainty of the data source based on evidence theory, corrects the confidence weight in combination with the Bayesian mechanism, and transmits to the robust model self-adaptive reconstruction module, adjusts the neural network structure according to the data uncertainty and suppresses the error, outputs the basic result of risk assessment, the risk result confidence calibration module calculates the confidence interval and outputs the final risk grade, the disaster risk grade output module displays, the data storage visualization module stores data and visualizes, the model parameter feedback optimization module optimizes the parameters according to the historical error, forms a closed loop, realizes accurate evaluation, and the whole working process ends.

[0027] Although the present application is disclosed in the preferred embodiments as above, it is not intended to limit the present application, and any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present application, therefore, any modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application, which does not deviate from the technical scheme of the present application, all fall within the protection scope defined by the claims of the present application.

Claims

1. A geological hazard risk assessment system based on multi-source data analysis, comprising a risk assessment system, characterized in that, The risk assessment system includes the following modules: Data uncertainty quantification module: Based on evidence theory, a multi-source data credibility assessment framework is constructed to quantify the uncertainty characteristics of each data source, and the credibility weight of each data source is corrected by combining a Bayesian dynamic update mechanism, outputting the quantification results and dynamic weights. Robust Model Adaptive Reconstruction Module: Constructs a dynamic neural network based on an attention mechanism, adjusts the structure and decision boundary according to the uncertainty of multi-source data, introduces robust optimization to suppress errors, and outputs basic results for risk assessment; Risk Result Credibility Calibration Module: Combining data uncertainty and model parameters, calculates the credibility range of risk results, identifies key influencing factors, and outputs the final disaster risk level and credibility information; Disaster risk level output module: adopts - Level V graded output, simultaneously displaying risk level, credibility range and key influencing factors; Data storage visualization module: Stores full-process data and evaluation results, and displays them through a multi-dimensional visualization interface, supporting data backtracking and trend analysis; Model parameter feedback optimization module: Analyzes historical evaluation errors, prioritizes optimizations, dynamically optimizes quantification parameters and initial model parameters, and improves the system's evaluation accuracy and stability.

2. The geological hazard risk assessment system based on multi-source data analysis according to claim 1, characterized in that: The risk assessment system also includes a multi-source data acquisition module and a data preprocessing module. The raw data collected by the multi-source data acquisition module specifically includes topography, meteorology, geological structure, hydrology, remote sensing images, sensor monitoring and historical data. The acquisition methods are divided into real-time acquisition and timed acquisition.

3. The geological hazard risk assessment system based on multi-source data analysis according to claim 2, characterized in that: The data preprocessing module converts all raw data into WGS-84 coordinates, SI units, and JSON format. It also performs missing value completion using the K-nearest neighbor algorithm, removes redundant data with a similarity greater than 95%, filters noise using db4 wavelet denoising, and outputs well-organized data.

4. The geological hazard risk assessment system based on multi-source data analysis according to claim 3, characterized in that: The data uncertainty quantification module constructs an identification framework based on the credibility assessment framework of evidence theory. {Trustworthy, Untrustworthy, Uncertain} Identification Framework It includes 3 single-element subsets, 3 double-element subsets, 1 universal subset, and 1 empty subset, where the basic probability assignment value of the empty subset is 0. This is based on the calculation of each data source within the recognition framework. Basic probability assignment function on all nonempty subsets The basic probability assignment function quantifies the degree of trust that the data source has in each proposition. satisfy: ； The Bayesian dynamic update mechanism is based on combining the prior confidence weight of the previous time step with the likelihood of the real-time observation data at the current time step, and iteratively corrects it to obtain the posterior confidence weight. The environmental parameters associated in real time during the update process include temperature change value, humidity change value, electromagnetic interference intensity and sensor working time.

5. A geological hazard risk assessment system based on multi-source data analysis according to claim 4, characterized in that: The dynamic neural network of the robust model adaptive reconstruction module includes an input layer, an attention feature extraction layer, a dynamic hidden layer, a robust optimization layer, and an output layer. The attention feature extraction layer is positively correlated with the confidence weight of the data source. The number of neurons in the dynamic hidden layer is adaptively adjusted according to the average uncertainty level of the multi-source data. The average uncertainty level is calculated by the mean of the uncertainty entropy values ​​of all data sources. The formula for calculating the uncertainty entropy value is as follows: ； in, The uncertainty entropy value of the data source, with a range of values. , At that time, the data source is completely trustworthy, including only =1, the rest , For identification framework Any non-empty subset; When the mean is <0.3, the number of hidden layer neurons is set to 128; when the mean is between 0.3 and 0.7, the number of neurons is set to 256; and when the mean is greater than 0.7, the number of neurons is set to 512, and the number of neurons never exceeds 512. The robust optimization layer uses a robust optimization regularization term to penalize the model parameters corresponding to low-confidence data, suppressing error propagation. The output layer uses... The activation function outputs the probability distribution of each risk level, which serves as the basis for risk assessment.

6. The geological hazard risk assessment system based on multi-source data analysis according to claim 5, characterized in that: The confidence interval calculation of the risk result confidence calibration module is based on the normal distribution confidence interval estimation method, with the confidence level set at 90%-99%. The total variance is calculated by superimposing the variance caused by data uncertainty and the variance caused by model robustness, and the upper and lower limits of the interval are determined by combining the standard normal distribution quantiles. Key influencing factors include the type of high-uncertainty data source and its corresponding uncertainty entropy value, the adjustment range of neurons in the model's dynamic hidden layer, the weight coefficient of the robust optimization regularization term, and the type and intensity of real-time environmental interference factors.

7. A geological hazard risk assessment system based on multi-source data analysis according to claim 6, characterized in that: The data storage visualization module adopts The distributed database stores data, including the original data set, the preprocessed data set, the uncertainty quantification result set, the model reconstruction parameter set, the risk assessment basic result set, the final assessment result set, and the historical error data set. The visualization interface includes data trend curves, uncertainty heatmaps, risk level spatial distribution maps, confidence interval bar charts, and model parameter change line charts. Users can query corresponding data by filtering conditions such as time range, spatial range, data type, and risk level.

8. A geological hazard risk assessment system based on multi-source data analysis according to claim 7, characterized in that: The model parameter feedback optimization module uses the mean square error index to calculate the deviation between historical assessment results and actual disaster occurrences. The calculation formula is as follows: ； in, This is a quantified value representing the actual disaster level. Level = 1 Level 2, Level III = 3, Level VI = 4, Level V = 5 This is a quantitative value of historical evaluation results. This refers to the number of historical evaluation samples; The optimization priority division criteria are as follows: When the value is greater than 0.8, a high-priority optimization is triggered, and the number of optimization iterations is set to 500. When the performance threshold is 0.4-0.8, a medium-priority optimization is triggered, with 300 iterations. When the parameter value is less than 0.4, a low-priority optimization is triggered. The number of iterations is set to 100. The convergence condition for the optimization iteration is that the parameter change between two consecutive iterations is less than 10. -5 The optimized parameters are used to update the system by replacing the original parameters, and the update frequency is consistent with the data collection frequency.

9. A geological hazard risk assessment method applicable to multi-source data analysis of the risk assessment system according to any one of claims 1-8, characterized in that, The risk assessment methods include the following: Step 1: Collect multi-dimensional raw data required for geological hazard assessment, covering four major categories: geological environment, meteorological conditions, monitoring and sensing, and historical records; Step 2: Standardize the format of the raw data, fill in missing values, remove redundancy and filter noise to eliminate heterogeneity interference and output high-quality, well-organized data. Step 3: Construct an identification framework based on evidence theory, quantify the uncertainty characteristics of each data source, and correct the credibility weights by combining the Bayesian dynamic update mechanism, outputting the quantification results and dynamic weights. Step 4: Construct a dynamic neural network based on the attention mechanism, adaptively adjust the network structure and decision boundary according to the data uncertainty, introduce robust optimization to suppress errors, and output basic results for risk assessment. Step 5: Combining data uncertainty and model parameters, calculate the confidence interval of the risk results, label key influencing factors, and obtain the final disaster risk level and confidence information; Step Six: Present the final risk level, credibility range, and key influencing factors using a tiered classification system of I-V; Step 7: Store the entire process data and evaluation results, display them through a visual interface and support backtracking analysis, prioritize based on historical evaluation errors, and dynamically optimize quantitative parameters and initial model parameters.

Images