A geological disaster risk assessment system based on cloud computing

Abstract

The application discloses a geological disaster risk assessment system based on cloud computing and relates to the technical field of geological information processing, which comprises a data processing module, a feature extraction module, an intelligent decision module and a scheme decision module. First, geological disaster monitoring data are acquired and subjected to data block coding processing to obtain a historical disaster sequence matrix and a regional space feature matrix, and then a parameterized unitary orthogonal matrix is constructed to orthogonally code and reversibly compress the historical disaster sequence to obtain a first disaster matrix. The first disaster matrix is subjected to disaster occurrence probability calculation to obtain an exploration action, the exploration action is optimized through a greedy strategy to generate an optimized monitoring sequence. The optimized monitoring sequence and the historical disaster sequence matrix are subjected to structural analysis through orthogonal inverse transformation to obtain a disaster monitoring point layout, and the adaptive improvement of exploration efficiency is realized, and an intelligent exploration task that is continuously optimized and can be referred to is formed.

Description

Technical Field

[0001] This invention relates to the field of geological information processing technology, and in particular to a cloud computing-based geological hazard risk assessment system. Background Technology

[0002] As the core of disaster prevention and mitigation, geological disaster monitoring often faces severe challenges such as massive data volume, complex decision-making, and uncertain outcomes. The monitoring process generates massive amounts of heterogeneous data from multiple sources, including topography, meteorology, hydrology, geotechnical mechanics, and historical disasters, which are highly correlated in time and space. Traditional data management and decision-making methods struggle to deeply integrate and efficiently mine this data, leading to suboptimal resource allocation in monitoring scheme development and low early warning efficiency.

[0003] With the development of cloud computing and artificial intelligence technologies, existing methods for processing geological disaster monitoring information generally suffer from the following problems: First, the high dimensionality and heterogeneity of geological disaster data make feature extraction difficult, and directly using machine learning for modeling can easily fall into the curse of dimensionality, and the interpretability of the model is poor; Second, monitoring decisions are constrained by the existing state, with significant delayed returns and state-dependent characteristics, and conventional static optimization or simple supervised learning models are difficult to characterize the dynamic evolution and long-term risks of geological disasters; Finally, the output of most intelligent algorithms is still an indicative predictive probability intermediate result, which cannot form a complete and executable plan that can directly guide monitoring deployment and emergency response, resulting in a breakpoint between decision analysis and engineering implementation. Summary of the Invention

[0004] The technical problem solved by this invention is that existing technologies fail to efficiently compress geological disaster monitoring data and extract key features with low complexity, make it difficult to dynamically adjust the monitoring process and combine regions while taking into account multi-objective constraints, and lack the ability to transform abstract strategies into a sequence of executable monitoring schemes.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a cloud computing-based geological hazard risk assessment system, comprising a data processing module, a feature extraction module, an intelligent decision-making module, and a scheme decision-making module;

[0006] The data processing module is used to acquire geological disaster monitoring data and perform data block encoding processing to obtain a historical disaster sequence matrix and a regional spatial feature matrix;

[0007] The feature extraction module is used to construct a parameterized unitary orthogonal matrix, perform orthogonal encoding and invertible transformation on the historical disaster sequence matrix, and obtain the first disaster matrix.

[0008] The intelligent decision-making module is used to calculate the probability of disaster occurrence from the first disaster matrix, obtain monitoring actions, optimize the monitoring actions through a greedy strategy, and generate an optimized monitoring sequence.

[0009] The scheme decision module is used to perform structured analysis on the optimized monitoring sequence and historical disaster sequence matrix through orthogonal inverse transformation to obtain the target monitoring task.

[0010] Preferably, the data processing module is used to acquire geological disaster monitoring data and perform data block encoding processing to obtain a historical disaster sequence matrix and a regional spatial feature matrix;

[0011] The target monitoring area is divided into M two-dimensional grid units, and the monitoring process is discretized into T monitoring rounds. Geological disaster monitoring is carried out in each grid unit according to the round number. Topographic data, meteorological and hydrological data, soil and rock mechanics data, slope stability monitoring data and historical disaster record data of the target area in each round are collected and saved as a geological disaster monitoring sequence. The geological disaster monitoring data is subjected to coordinate unification and standardized coding to obtain the regional spatial feature matrix (M*K) of the current round, where K is the dimension of the geological disaster monitoring sequence.

[0012] Each completed geological disaster monitoring task is recorded as a historical disaster sequence. The historical disaster sequence includes the task time, technology used, target area, cost, abnormal indicators, and result evaluation. The historical disaster sequence is arranged in chronological order to obtain a historical disaster sequence matrix.

[0013] The topographic data includes slope, aspect, elevation, surface curvature, and lithology.

[0014] The meteorological and hydrological data include rainfall intensity, rainfall duration, evaporation, runoff coefficient, and groundwater level;

[0015] The soil and geotechnical data include soil type, pore water pressure, internal friction angle, cohesion, and permeability coefficient.

[0016] The slope stability monitoring data includes displacement monitoring values, stress and strain data, anchor cable load, and crack width;

[0017] The historical disaster record data includes disaster type, occurrence time, scope of impact, loss assessment, and recurrence index.

[0018] Preferably, the feature extraction module includes a parameter integration unit, an orthogonal matrix construction unit, and an orthogonal feature encoding unit;

[0019] The parameter integration unit is used to construct a monitoring environmental parameter vector using the regional spatial feature matrix and the historical disaster sequence matrix;

[0020] Regional complexity parameters are statistically analyzed from the regional spatial feature matrix, including geological complexity and data noise complexity.

[0021] The logic for handling the geological complexity is as follows:

[0022] The variance of each column in the regional spatial feature matrix is ​​calculated to obtain K feature variances. The mean values ​​of the slope and internal friction angle are calculated to obtain the mean slope and mean internal friction angle. The feature variances, the mean slope and the mean internal friction angle are added together and averaged, and then normalized to obtain the geological complexity.

[0023] The logic for handling data noise complexity is as follows:

[0024] Meteorological and hydrological data and slope stability monitoring data are extracted from the regional spatial feature matrix and the regional spatial signal-to-noise ratio is obtained by moving average filtering. The reciprocal of the regional spatial signal-to-noise ratio is then taken and normalized to obtain the data noise complexity.

[0025] Monitoring and evaluation parameters are statistically analyzed from the historical disaster sequence matrix, including monitoring cost parameters and monitoring success rate parameters.

[0026] The processing logic for the monitoring cost parameters is as follows:

[0027] Costs are extracted from the historical disaster sequence matrix to obtain cost sequences. The average value of the cost sequences is calculated and normalized to obtain monitoring cost parameters.

[0028] The processing logic for the monitoring success rate parameter is as follows:

[0029] The results evaluation sequence is extracted from the historical disaster sequence matrix. The results evaluation includes successful early warning and missed reporting. Successful early warning and missed reporting are mapped to 1 and 0, respectively. The success rate of the results evaluation sequence is calculated and normalized as a monitoring success parameter.

[0030] The regional complexity parameter and the monitoring and evaluation parameter are concatenated into a monitoring environment parameter vector.

[0031] Preferably, the orthogonal matrix construction unit is used to obtain an orthogonal matrix by performing a unitary orthogonal transformation on the regional spatial feature matrix and the historical disaster sequence matrix;

[0032] The monitoring environment parameter vector received by the parameter integration unit is used to calculate the control parameters and initial values ​​of the Logistic chaotic mapping. The processing logic for the control parameters and initial values ​​is as follows:

[0033] A comprehensive index is calculated by monitoring environmental parameter vectors. This comprehensive index is the average of the regional complexity parameter and the monitoring and evaluation parameter. Control parameters are then calculated based on this comprehensive index, and the expression for the control parameters is as follows:

[0034] ;

[0035] in, For control parameters, As a comprehensive indicator;

[0036] The mean and standard deviation of each sample feature are calculated based on the regional spatial feature matrix to obtain the first mean vector and the first standard deviation vector. The mean of the sample features is calculated based on the historical disaster sequence matrix to obtain the second mean vector. The first mean vector, the first standard deviation vector and the second mean vector are concatenated according to the time series dimension and then projected and normalized to generate the initial value.

[0037] Using control parameters and initial values ​​as control factors, and setting the number of iterations to Q, where Q is four times the dimension of the historical disaster sequence, a chaotic sequence is generated iteratively using the control parameters and initial values ​​according to the Logistic chaotic mapping. The iterative calculation expression for the chaotic sequence is as follows:

[0038] ;

[0039] in, For the k-th chaotic sequence, For control parameters, For the (k+1)th chaotic sequence, obtain the chaotic sequence. The chaotic sequence is mapped to a rotation angle sequence, and the calculation expression for the rotation angle sequence mapping is as follows:

[0040] ;

[0041] in, Let k be the k-th rotation angle after mapping. It is a mathematical natural constant;

[0042] The covariance matrices of the regional spatial feature matrix and the historical disaster sequence matrix are calculated separately to obtain the first covariance matrix and the second covariance matrix. For each sample feature, the Pearson correlation coefficient of the first covariance matrix and the second covariance matrix is ​​calculated to obtain the first Pearson coefficient and the second Pearson coefficient. The sum of the absolute values ​​of the first Pearson coefficient and the second Pearson coefficient is calculated to obtain the comprehensive correlation strength. For each sample feature dimension i, the feature dimension j with the largest comprehensive correlation strength is selected. The index coordinates corresponding to the sample feature and the feature with the strongest correlation are added to the rotation plane candidate set. Through the rotation angle sequence and the rotation plane candidate set, a rotation matrix sequence is constructed. Each matrix in the rotation matrix sequence is multiplied to obtain an orthogonal matrix. The row dimension and column dimension of the orthogonal matrix are equal to the feature dimensions of the historical disaster sequence matrix.

[0043] Preferably, the orthogonal feature encoding unit is used to encode the historical disaster sequence matrix according to the orthogonal matrix to obtain the first disaster matrix;

[0044] The orthogonal matrix is ​​transposed to obtain the transposed orthogonal matrix. The historical disaster sequence matrix is ​​then transformed and encoded using the transposed orthogonal matrix to obtain the transformed historical disaster sequence matrix, which serves as the first disaster matrix. The transformation encoding expression for the first disaster matrix is ​​as follows:

[0045] ;

[0046] in, This is the first disaster matrix. This is a matrix of historical disaster sequences. It is a transpose orthogonal matrix.

[0047] Preferably, the intelligent decision-making module includes a state construction unit, an action generation unit, and a probability evaluation unit;

[0048] The state construction unit is used to construct an extended monitoring state vector based on the first disaster matrix and the current monitoring round;

[0049] Based on the regional spatial feature matrix of the current round, calculate the mean vector, standard deviation vector and corresponding outlier index vector of each feature dimension to obtain statistics. Then, concatenate the statistics with the remaining budget, remaining time window and environmental constraint indicators to obtain the current state components.

[0050] The monitoring status component is obtained by calculating the anomaly index based on the historical disaster sequence matrix. The monitoring status component is used to represent the coverage and regional priority distribution of the currently monitored area. The monitoring status component includes the monitoring hit rate and the monitoring occupancy rate. The monitoring hit rate is the ratio of the number of abnormal areas with detected anomalies to the number of monitored areas. The monitoring occupancy rate is the ratio of the area of ​​the monitored area to the total area of ​​the target monitoring area.

[0051] Extract the row vectors corresponding to the most recent L monitoring records from the first disaster matrix as historical state components;

[0052] Save the current state component, the monitoring state component, and the historical state component as an extended monitoring state vector.

[0053] Preferably, the action generation unit is used to perform monitoring probability calculation on the extended monitoring state vector through a lightweight decision network to obtain the monitoring action and target space unit set for the current round;

[0054] The extended monitoring state vector is input into a lightweight decision network, which outputs the monitoring technology probability and spatial unit priority probability. The processing logic of the lightweight decision network is as follows:

[0055] The extended monitoring state vector is input into a fully connected layer, where feature fusion is performed to obtain an intermediate feature vector. This intermediate feature vector is then input into the monitoring technology head, which includes a fully connected layer and a softmax activation function. The head outputs the probability of each monitoring technology. The feature vector of each grid cell is extracted from the regional spatial feature matrix of the current round. The feature vector of each grid cell has a dimension of K. The intermediate layer feature vector is projected onto the same dimension K as the feature vector of the grid cell through a fully connected layer to obtain a query vector. The dot product of the query vector and the feature vector of each grid cell in the regional spatial feature matrix is ​​calculated to obtain the priority score of each grid cell. The priority score is then subjected to a sigmoid transformation to obtain the spatial cell priority probability and the monitoring technology probability.

[0056] The monitoring action with the highest probability is selected as the monitoring action for the current round. The spatial unit priority scores are sorted in descending order to obtain the target spatial unit set. The monitoring action for the current round and the target spatial unit set are saved as monitoring actions.

[0057] Preferably, the probability evaluation unit is used to optimize the monitoring action using a greedy strategy to obtain an optimized monitoring sequence;

[0058] Based on the remaining budget and iteration step size, iterate and optimize starting from the first monitoring round:

[0059] Based on the state of the previous moment and the selected monitoring actions, the state construction unit is invoked, the current state component is recalculated, the monitoring action with the highest probability is selected, and spatial units are selected according to the priority score of spatial units in the regional spatial feature matrix based on the current monitoring action and the score, until the cumulative estimated cost of the spatial units exceeds the preset remaining budget, then the iteration is terminated, and the actions selected in each iteration step are added to the optimized monitoring sequence in sequence to obtain the optimized monitoring sequence.

[0060] Preferably, the scheme decision module includes an orthogonal inverse transformation unit and an analysis unit;

[0061] The orthogonal inverse transformation unit is used to restore the optimized monitoring sequence through orthogonal inverse transformation to obtain the optimized historical disaster sequence matrix;

[0062] Encode each monitoring action in the optimized monitoring sequence by encoding the monitoring technology type, target spatial unit set, estimated cost vector, and time arrangement vector to generate an action coding vector. Then, concatenate the action codes of each round in chronological order to form an action coding sequence.

[0063] The action coding sequence and the most recent L codes of the first disaster matrix are input into the encoder network. The latent variables are sampled using a reparameterized Gaussian distribution. An incremental feature matrix is ​​generated in the coding space through the decoder network. The incremental feature matrix is ​​inversely transformed by the orthogonal matrix output by the orthogonal matrix construction unit to calculate the scheme increment matrix of the original task space. The scheme increment matrix is ​​superimposed with the historical monitoring sequence matrix to obtain the optimized historical disaster sequence matrix.

[0064] Preferably, the parsing unit is used to perform structured parsing on the optimized historical disaster sequence matrix to generate target monitoring tasks;

[0065] The last row vector of the optimized historical disaster sequence matrix is ​​parsed to map the monitoring scheme record corresponding to the last row vector to the target monitoring task. The target monitoring task includes the monitoring round, the type of technology used, the set of target spatial units, the suggested sampling density, the budget allocation, the time arrangement, and the evaluation of the prediction results.

[0066] The beneficial effects of this invention are as follows: Traditional methods employ fixed transformations or simple dimensionality reduction, while this invention dynamically drives Logistic chaotic mapping by monitoring environmental parameters to generate a parameterized sparse orthogonal matrix. This matrix combines orthogonality, sparsity, and environmental adaptability, ensuring lossless and reversible information while reducing computational overhead. The parameterized sparse orthogonal matrix is ​​used to orthogonally encode historical disaster sequences, resulting in a decorrelated first disaster matrix, laying a high-quality data foundation for subsequent intelligent analysis. This invention also solves the problems of non-Markovianness and long-term dependency in monitoring sequence decision-making. The current monitoring state is used as a cached state, and an extended state space with fixed-window historical encoding is adopted, enabling the system to comprehensively consider the current environment, the cumulative effects of executed actions, and historical experience context during decision-making. Combined with a lightweight dual-head decision network, the system outputs technical probabilities and regional priorities, and designs a rolling greedy strategy considering budget constraints. This allows the system to efficiently generate globally coordinated optimized monitoring action sequences in the encoding space, achieving multi-round scientific planning. Finally, this invention learns the pattern increment guided by the optimized sequence in the coding space through a conditional variational autoencoder, and decodes the increment back to the original task space using the inverse transformation of the same orthogonal matrix. This increment is then fused with the historical sequence to form a structured new scheme, generating an executable task for geological disaster monitoring. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the basic process of a cloud computing-based geological hazard risk assessment system provided in one embodiment of the present invention. Detailed Implementation

[0068] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0069] Example, refer to Figure 1 This paper presents a cloud computing-based geological hazard risk assessment system, which includes a data processing module, a feature extraction module, an intelligent decision-making module, and a scheme decision-making module.

[0070] The data processing module is used to acquire geological disaster monitoring data and perform data block encoding to obtain a historical disaster sequence matrix and a regional spatial feature matrix.

[0071] The feature extraction module is used to construct a parameterized unitary orthogonal matrix, perform orthogonal encoding and invertible transformation on the historical disaster sequence matrix, and obtain the first disaster matrix.

[0072] The intelligent decision-making module is used to calculate the probability of disaster occurrence from the first disaster matrix, obtain monitoring actions, optimize the monitoring actions through a greedy strategy, and generate an optimized monitoring sequence.

[0073] The scheme decision module is used to perform structured analysis on the optimized monitoring sequence and historical disaster sequence matrix through orthogonal inverse transformation to obtain the target monitoring task.

[0074] In this embodiment, the data processing module receives the multi-source geological disaster monitoring data stream and regularizes it into a historical disaster sequence matrix (T task records × D-dimensional features) and a regional spatial feature matrix (M grids × K-dimensional features) with spatiotemporal labels. The feature extraction module dynamically generates a parameterized unitary orthogonal matrix based on the statistical characteristics of the current monitoring environment and performs orthogonal transformation on the historical sequence matrix to obtain a compressed and decorrelated first disaster matrix Z. The intelligent decision-making module analyzes the historical patterns encoded by matrix Z and the current regional state, calculates the monitoring technology probability and regional priority through a lightweight decision network, and uses a greedy strategy to generate a multi-round optimized monitoring sequence under budget constraints. Finally, the scheme decision-making module uses the inverse transformation property of the orthogonal matrix to decode the scheme increment generated in the encoding space of the abstract sequence back to the original task space, superimposes it with the historical disaster sequence matrix, and parses it to obtain the target monitoring task containing specific technology, region, budget, and time.

[0075] The data processing module is used to acquire geological disaster monitoring data and perform data block encoding to obtain historical disaster sequence matrix and regional spatial feature matrix;

[0076] The target monitoring area is divided into M two-dimensional grid units, and the monitoring process is discretized into T monitoring rounds. Geological disaster monitoring is carried out in each grid unit according to the round number. Topographic data, meteorological and hydrological data, soil and rock mechanics data, slope stability monitoring data and historical disaster record data of the target area in each round are collected and saved as a geological disaster monitoring sequence. The geological disaster monitoring data is subjected to coordinate unification and standardized coding to obtain the regional spatial feature matrix (M*K) of the current round, where K is the dimension of the geological disaster monitoring sequence.

[0077] Each completed geological disaster monitoring task is recorded as a historical disaster sequence. The historical disaster sequence includes the task time, technology used, target area, cost, abnormal indicators, and results evaluation. The historical disaster sequences are arranged in chronological order to obtain a historical disaster sequence matrix.

[0078] Topographic data includes slope, aspect, elevation, surface curvature, and lithology.

[0079] Meteorological and hydrological data include rainfall intensity, rainfall duration, evaporation, runoff coefficient, and groundwater level;

[0080] Soil and geotechnical data include soil type, pore water pressure, internal friction angle, cohesion, and permeability coefficient;

[0081] Slope stability monitoring data includes displacement monitoring values, stress and strain data, anchor cable loads, and crack widths;

[0082] Historical disaster records include disaster type, time of occurrence, scope of impact, loss assessment, and recurrence indicators.

[0083] In this embodiment, a monitoring area is divided into M grids (M=10000 units), and a complete monitoring project is discretized into T rounds. One round is randomly selected, and the slope, rainfall intensity, soil internal friction angle, displacement monitoring values, and historical disaster counts for each grid are collected. These data are then spatially interpolated to grid coordinates and standardized to form a regional spatial feature matrix with grids and K columns of features. Simultaneously, N past monitoring tasks, each record including task time, technology used, target area, cost, anomaly indicators, and result evaluation, are arranged chronologically into an N-row × D-column historical disaster sequence matrix. This embodiment obtains gridded data that directly characterizes geological disasters by spatiotemporally gridding and standardizing geological disaster information. Geological disaster monitoring activities are abstracted into a regional spatial feature matrix representing spatial states and a historical disaster matrix representing historical monitoring experience, providing a unified input for subsequent matrix operations.

[0084] The feature extraction module includes a parameter integration unit, an orthogonal matrix construction unit, and an orthogonal feature encoding unit;

[0085] The parameter integration unit is used to construct a monitoring environmental parameter vector using the regional spatial feature matrix and the historical disaster sequence matrix;

[0086] Regional complexity parameters are statistically analyzed from the regional spatial feature matrix. These parameters include geological complexity and data noise complexity.

[0087] The logic for handling geological complexity is as follows:

[0088] The variance of each column in the regional spatial feature matrix is ​​calculated to obtain K feature variances. The mean values ​​of slope and internal friction angle are calculated to obtain the mean slope and mean internal friction angle. The feature variances, mean slope and mean internal friction angle are added together and averaged, and then normalized to obtain the geological complexity.

[0089] The logic for handling data noise complexity is as follows:

[0090] Meteorological and hydrological data and slope stability monitoring data are extracted from the regional spatial feature matrix and the regional spatial signal-to-noise ratio is obtained by moving average filtering. The reciprocal of the regional spatial signal-to-noise ratio is taken and normalized to obtain the data noise complexity.

[0091] Monitoring and evaluation parameters are statistically analyzed from the historical disaster sequence matrix. These parameters include monitoring cost parameters and monitoring success rate parameters.

[0092] The processing logic for monitoring cost parameters is as follows:

[0093] Costs are extracted from the historical disaster sequence matrix to obtain cost sequences. The average value of the cost sequences is calculated and normalized to obtain monitoring cost parameters.

[0094] The processing logic for the success rate monitoring parameter is as follows:

[0095] The results evaluation sequence is extracted from the historical disaster sequence matrix. The results evaluation includes successful early warning and missed reporting. Successful early warning and missed reporting are mapped to 1 and 0 respectively. The success rate of the results evaluation sequence is calculated and normalized as the monitoring success parameter.

[0096] The regional complexity parameter and the monitoring and evaluation parameter are concatenated into a monitoring environment parameter vector.

[0097] In this embodiment, the historical disaster sequence matrix and regional spatial feature matrix are received from the data processing module. The variance of the K feature columns of the regional spatial feature matrix is ​​calculated, and the average values ​​of the slope and internal friction angle of the entire region are calculated. The variance of the K feature columns is added to the two average values, and then the average is normalized to [0,1] to obtain the geological complexity. Next, meteorological and hydrological data and slope stability monitoring data are extracted from the regional spatial feature matrix. The signal and noise are separated by sliding window filtering. The average signal-to-noise ratio is calculated, the reciprocal is taken, and normalized to obtain the data noise complexity. The historical average cost is calculated from the cost column of the historical disaster sequence matrix and normalized to obtain the monitoring cost parameter. The success rate (number of successful warnings / total number) is calculated from the achievement evaluation column and normalized to obtain the monitoring success rate parameter. Finally, the geological complexity, data noise complexity, monitoring cost parameter, and monitoring success rate parameter are concatenated into a monitoring environmental parameter vector.

[0098] By quantifying the parameters of the monitoring environment, massive low-order features are transformed into high-order environmental parameters (monitoring environment parameter vectors), achieving high compression and effective representation of information. This enables the system to understand whether the current monitoring environment is a complex or simple region, the level of data noise, and the tightness of the budget, and adjust the orthogonal matrix accordingly, thus achieving intelligent processing that adapts to the environment.

[0099] The orthogonal matrix construction unit is used to obtain an orthogonal matrix by performing a unitary orthogonal transformation on the regional spatial feature matrix and the historical disaster sequence matrix;

[0100] The monitoring environment parameter vector received by the parameter integration unit is used to calculate the control parameters and initial values ​​of the Logistic chaotic mapping. The processing logic for the control parameters and initial values ​​is as follows:

[0101] A comprehensive index is calculated by monitoring environmental parameter vectors. The comprehensive index is the average of the regional complexity parameter and the monitoring and evaluation parameter. Control parameters are then calculated based on the comprehensive index. The expression for the control parameters is as follows:

[0102] ;

[0103] in, For control parameters, As a comprehensive indicator;

[0104] The mean and standard deviation of each sample feature are calculated based on the regional spatial feature matrix to obtain the first mean vector and the first standard deviation vector. The mean of the sample features is calculated based on the historical disaster sequence matrix to obtain the second mean vector. The first mean vector, the first standard deviation vector and the second mean vector are concatenated according to the time series dimension and then projected and normalized to generate the initial value.

[0105] Using control parameters and initial values ​​as control factors, and setting the number of iterations to Q, where Q is four times the dimension of the historical disaster sequence, a chaotic sequence is generated iteratively using the control parameters and initial values ​​according to the Logistic chaotic mapping. The iterative calculation expression for the chaotic sequence is as follows:

[0106] ;

[0107] in, For the k-th chaotic sequence, For control parameters, For the (k+1)th chaotic sequence, obtain the chaotic sequence. The chaotic sequence is mapped to a rotation angle sequence. The calculation expression for the rotation angle sequence mapping is as follows:

[0108] ;

[0109] in, Let k be the k-th rotation angle after mapping. It is a mathematical natural constant;

[0110] The covariance matrices of the regional spatial feature matrix and the historical disaster sequence matrix are calculated separately to obtain the first covariance matrix and the second covariance matrix. For each sample feature, the Pearson correlation coefficient of the first covariance matrix and the second covariance matrix is ​​calculated to obtain the first Pearson coefficient and the second Pearson coefficient. The sum of the absolute values ​​of the first Pearson coefficient and the second Pearson coefficient is calculated to obtain the comprehensive correlation strength. For each sample feature dimension i, the feature dimension j with the largest comprehensive correlation strength is selected. The index coordinates corresponding to the sample feature and the feature with the strongest correlation are added to the rotation plane candidate set. Through the rotation angle sequence and the rotation plane candidate set, a rotation matrix sequence is constructed. Each matrix in the rotation matrix sequence is multiplied to obtain an orthogonal matrix. The row dimension and column dimension of the orthogonal matrix are equal to the feature dimensions of the historical disaster sequence matrix.

[0111] In this embodiment, by receiving the monitoring environmental parameter vector, the comprehensive index $s$ and control parameter $\mu$ are calculated. The mean and standard deviation of the regional spatial feature matrix and the mean of the historical disaster sequence matrix are calculated. After concatenation, normalization and projection are performed to obtain initial values. Using the control parameter and initial values ​​as control factors, the Logistic mapping is iterated Q times to generate a chaotic sequence, which is then mapped to a rotation angle sequence. The covariance matrix of the regional spatial feature matrix and the historical disaster sequence matrix is ​​then calculated. For each feature i, the most relevant feature j is found (considering the correlation between the two matrices), forming a candidate set of rotation planes {(i, j)}. For each rotation angle... Construct a plane with non-zero elements (cos 1 and 2) only in the corresponding row and column of the (i,j) plane. , -sin sin cos The Givens rotation matrix is ​​obtained by multiplying the rotation matrix to obtain an orthogonal matrix.

[0112] Chaotic mapping is used to generate angles to ensure the strong randomness and unpredictability of the matrix. Meanwhile, the control factor is determined by the exploration environment parameters. The most relevant feature pair is selected as the rotation plane, which is the feature with the strongest correlation in the future orthogonal transformation, realizing the removal of correlation and information aliasing. The final generated orthogonal matrix can ensure the reversibility of the transformation, reduce the computational complexity, and allow the rotation matrix to adapt to the monitoring environment parameters. This effectively reduces the computational complexity of the matrix, reduces the loss of information in the transformation, and facilitates the effective implementation of subsequent inverse transformation.

[0113] The orthogonal feature coding unit is used to encode the historical disaster sequence matrix according to the orthogonal matrix to obtain the first disaster matrix;

[0114] The orthogonal matrix is ​​transposed to obtain the transposed orthogonal matrix. The historical disaster sequence matrix is ​​then transformed and encoded using the transposed orthogonal matrix to obtain the transformed historical disaster sequence matrix, which serves as the first disaster matrix. The transformation encoding expression for the first disaster matrix is ​​as follows:

[0115] ;

[0116] in, This is the first disaster matrix. This is a matrix of historical disaster sequences. It is a transpose orthogonal matrix.

[0117] In this embodiment, the historical disaster sequence matrix is ​​multiplied with the transposed orthogonal matrix to obtain the first disaster matrix. The purpose is to project the initially correlated historical disaster sequence matrix onto a new orthogonal coordinate system spanned by the column vectors of the orthogonal matrix. Under the orthogonal coordinate system, the correlation between the data dimensions is eliminated to the greatest extent, effectively reducing the learning difficulty of subsequent intelligent decision-making and improving the generalization ability.

[0118] The state construction unit is used to construct an extended monitoring state vector based on the first disaster matrix and the current monitoring round;

[0119] Based on the regional spatial feature matrix of the current round, calculate the mean vector, standard deviation vector and corresponding outlier index vector of each feature dimension to obtain statistics. Then, concatenate the statistics with the remaining budget, remaining time window and environmental constraint indicators to obtain the current state components.

[0120] The monitoring status component is obtained by calculating the anomaly indicators based on the historical disaster sequence matrix. The monitoring status component is used to represent the coverage and regional priority distribution of the currently monitored area. The monitoring status component includes the monitoring hit rate and the monitoring occupancy rate. The monitoring hit rate is the ratio of the number of abnormal areas where anomalies were found to the number of monitored areas. The monitoring occupancy rate is the ratio of the area of ​​the monitored area to the total area of ​​the target monitoring area.

[0121] Extract the row vectors corresponding to the most recent L monitoring records from the first disaster matrix as historical state components;

[0122] Save the current state component, the monitoring state component, and the historical state component as an extended monitoring state vector.

[0123] In this embodiment, during the t-th round of operation, based on the regional spatial feature matrix of the current round, the mean, standard deviation, and anomaly ratio of each feature are calculated, and then the budget and time constraints are concatenated to form the current state component. Secondly, based on the anomaly detection situation of the monitored area, the monitoring hit rate (when 5 anomalies have been detected / 100 grids have been monitored = 5%) and the monitoring occupancy rate (50 square kilometers have been monitored / total area of ​​100 square kilometers = 50%) are calculated to form the monitoring state component. Finally, the row vectors of the most recent L records are extracted from the first disaster matrix to form the historical state component. The three are concatenated to obtain the extended monitoring state vector.

[0124] The current state component describes the current objective environment, the monitoring state component describes the cumulative effect of actions taken (similar to a cached state), and the historical state component provides historical decision-making context. By obtaining the extended monitoring state vector through these three state components, the observability and non-Markovianness issues in reinforcement learning monitoring scenarios are addressed. Traditional MDPs assume the current state contains all historical information, but current grid data in monitoring cannot reflect the global monitoring history. By explicitly incorporating the first hazard matrix and monitoring results (monitoring hit rate and monitoring occupancy rate) into the state, the state space is expanded, making the extended monitoring state vector approximately satisfy the Markov property and avoiding ineffective duplicate monitoring.

[0125] The action generation unit is used to calculate the monitoring probability of the extended monitoring state vector through a lightweight decision network to obtain the monitoring action and target space unit set for the current round.

[0126] The extended monitoring state vector is input into the lightweight decision network, which outputs the monitoring technology probability and spatial unit priority probability. The processing logic of the lightweight decision network is as follows:

[0127] The extended monitoring state vector is input into a fully connected layer, where feature fusion is performed to obtain an intermediate feature vector. This intermediate feature vector is then input into the monitoring technology head, which includes a fully connected layer and a softmax activation function. The head outputs the probability of each monitoring technology. The feature vector of each grid cell is extracted from the regional spatial feature matrix of the current round. The feature vector of each grid cell has a dimension of K. The intermediate layer feature vector is projected onto the same dimension K as the feature vector of the grid cell through a fully connected layer to obtain the query vector. The dot product of the query vector and the feature vector of each grid cell in the regional spatial feature matrix is ​​calculated to obtain the priority score of each grid cell. The priority score is then subjected to a sigmoid transformation to obtain the spatial cell priority probability and the monitoring technology probability.

[0128] The monitoring action with the highest probability is selected as the monitoring action for the current round. The spatial unit priority scores are sorted in descending order to obtain the target spatial unit set. The monitoring action for the current round and the target spatial unit set are saved as monitoring actions.

[0129] In this embodiment, the extended state vector is input into a lightweight decision network. The lightweight decision network first performs feature fusion through a fully connected layer, then splits into two attention heads: one outputs the probability distribution of six monitoring technologies through a fully connected layer and Softmax, and the other projects the intermediate features into a query vector and calculates the dot product similarity with the feature vectors of all grids in the regional spatial feature matrix. The spatial unit priority probability of each grid is then obtained through a sigmoid function. During decision-making, the technology with the highest monitoring technology probability (e.g., displacement monitoring) is selected, along with several grids with the highest spatial unit priority probabilities, forming a target spatial unit set within the budget. The combination of these two factors constitutes the monitoring action for the current round.

[0130] In this embodiment, a dual-head architecture is used to handle discrete technology selection and continuous spatial selection problems respectively. By using the dot product attention mechanism of query vectors and grid features, the monitoring value of each region is dynamically evaluated based on the current comprehensive status, which effectively improves the efficiency of monitoring region selection.

[0131] The probability evaluation unit is used to optimize the monitoring actions using a greedy strategy to obtain an optimized monitoring sequence;

[0132] Based on the remaining budget and iteration step size, iterate and optimize starting from the first monitoring round:

[0133] Based on the state of the previous moment and the selected monitoring actions, the state construction unit is invoked, the current state component is recalculated, the monitoring action with the highest probability is selected, and spatial units are selected according to the priority score of spatial units in the regional spatial feature matrix based on the current monitoring action and the score, until the cumulative estimated cost of the spatial units exceeds the preset remaining budget, then the iteration is terminated, and the actions selected in each iteration step are added to the optimized monitoring sequence in sequence to obtain the optimized monitoring sequence.

[0134] In this embodiment, a greedy algorithm is used for multi-round sequence optimization. It iterates from the initial state according to the total budget. In the h-th step, the state construction unit is called to update the state based on the executed action sequence, and the action generation unit is called to obtain the current optimal action. The cost of executing the current optimal action is checked. When the accumulated cost is less than the remaining budget, the current optimal action is added to the optimization monitoring sequence and the corresponding budget is deducted. The simulated state is transferred to the next round. This process is repeated until the budget is exhausted or the maximum number of rounds is reached, and the optimization monitoring sequence is output.

[0135] The greedy strategy in this embodiment means that in each round, the best action that seems best at the moment is selected. By sequential decision-making and state simulation transition, the multi-step decision problem is decomposed into a series of single-step problems. By simulating the decision-making process of the entire monitoring cycle online, a globally coordinated optimized monitoring sequence is obtained, which effectively balances the computational complexity of global optimization and the optimal monitoring action.

[0136] The scheme decision module includes an orthogonal inverse transformation unit and an analytical unit;

[0137] The orthogonal inverse transform unit is used to restore the optimized monitoring sequence through orthogonal inverse transform to obtain the optimized historical disaster sequence matrix;

[0138] Encode each monitoring action in the optimized monitoring sequence by encoding the monitoring technology type, target spatial unit set, estimated cost vector, and time arrangement vector to generate an action coding vector. Then, concatenate the action codes of each round in chronological order to form an action coding sequence.

[0139] The action coding sequence and the most recent L codes of the first disaster matrix are input into the encoder network. The latent variables are sampled using a reparameterized Gaussian distribution. The incremental feature matrix is ​​generated in the coding space through the decoder network. The orthogonal matrix output by the orthogonal matrix construction unit is used to perform an inverse transformation on the incremental feature matrix to calculate the scheme increment matrix of the original task space. The scheme increment matrix is ​​superimposed with the historical monitoring sequence matrix to obtain the optimized historical disaster sequence matrix.

[0140] The optimized monitoring sequence is encoded, for example, the magnetic exploration is converted into a one-hot encoded vector, the selected grid list is converted into a multi-hot vector, and the one-hot encoded vector and multi-hot vector are concatenated with the estimated cost and time vectors to obtain the action encoded vector. The action encoded vector and L rows are input into the conditional variational autoencoder, which outputs the distribution of latent variables. After sampling, the decoder generates an incremental feature matrix, where the dimension of the incremental feature matrix is ​​the same as that of the first disaster matrix. An inverse transformation is performed using an orthogonal matrix, and the matrix is ​​aligned and added with the historical disaster sequence matrix to obtain the optimized historical disaster sequence matrix.

[0141] The resulting optimized historical disaster sequence matrix is ​​a structured data object that can be directly used to guide the next steps of data collection and updating.

[0142] The parsing unit is used to perform structured parsing on the optimized historical disaster sequence matrix to generate target monitoring tasks;

[0143] The last row vector of the optimized historical disaster sequence matrix is ​​parsed to map the monitoring scheme record corresponding to the last row vector to the target monitoring task. The target monitoring task includes the monitoring round, the type of technology used, the set of target spatial units, the suggested sampling density, the budget allocation, the time arrangement, and the evaluation of the prediction results.

[0144] This embodiment reads the last row of the optimized historical disaster sequence matrix and parses it according to predetermined field mapping rules. For example, it decodes the technology type as displacement monitoring from a specific column, decodes the target grid number list from another set of columns, and parses the suggested sampling density, budget allocation, time arrangement, and prediction result evaluation from other columns. This information is then filled into the target monitoring task template, improving the practicality and ease of use of the entire system. This allows the intelligent results to be seamlessly integrated into the existing monitoring workflow and drive monitoring actions.

[0145] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0146] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A cloud computing-based geological hazard risk assessment system, characterized in that, It includes a data processing module, a feature extraction module, an intelligent decision-making module, and a solution decision-making module; The data processing module is used to acquire geological disaster monitoring data and perform data block encoding processing to obtain a historical disaster sequence matrix and a regional spatial feature matrix; The feature extraction module is used to construct a parameterized unitary orthogonal matrix, perform orthogonal encoding and invertible transformation on the historical disaster sequence matrix, and obtain the first disaster matrix. The intelligent decision-making module is used to calculate the probability of disaster occurrence from the first disaster matrix, obtain monitoring actions, optimize the monitoring actions through a greedy strategy, and generate an optimized monitoring sequence. The scheme decision module is used to perform structured analysis on the optimized monitoring sequence and historical disaster sequence matrix through orthogonal inverse transformation to obtain the target monitoring task; The feature extraction module includes a parameter integration unit, an orthogonal matrix construction unit, and an orthogonal feature encoding unit; The orthogonal matrix construction unit is used to perform unitary orthogonal transformation on the regional spatial feature matrix and the historical disaster sequence matrix to obtain an orthogonal matrix; The monitoring environment parameter vector received by the parameter integration unit is used to calculate the control parameters and initial values ​​of the Logistic chaotic mapping. The processing logic for the control parameters and initial values ​​is as follows: A comprehensive index is calculated by monitoring environmental parameter vectors. This comprehensive index is the average of the regional complexity parameter and the monitoring and evaluation parameter. Control parameters are then calculated based on this comprehensive index, and the expression for the control parameters is as follows: ； in, For control parameters, As a comprehensive indicator; The mean and standard deviation of each sample feature are calculated based on the regional spatial feature matrix to obtain the first mean vector and the first standard deviation vector. The mean of the sample features is calculated based on the historical disaster sequence matrix to obtain the second mean vector. The first mean vector, the first standard deviation vector and the second mean vector are concatenated according to the time series dimension and then projected and normalized to generate the initial value. Using control parameters and initial values ​​as control factors, and setting the number of iterations to Q, where Q is four times the dimension of the historical disaster sequence matrix, a chaotic sequence is generated iteratively using the control parameters and initial values ​​according to the Logistic chaotic mapping. The iterative calculation expression for the chaotic sequence is as follows: ； in, For the k-th chaotic sequence, For control parameters, For the (k+1)th chaotic sequence, obtain the chaotic sequence. The chaotic sequence is mapped to a rotation angle sequence, and the calculation expression for the rotation angle sequence mapping is as follows: ； in, Let k be the k-th rotation angle after mapping. It is a mathematical natural constant; The covariance matrices of the regional spatial feature matrix and the historical disaster sequence matrix are calculated separately to obtain the first covariance matrix and the second covariance matrix. For each sample feature, the Pearson correlation coefficient of the first covariance matrix and the second covariance matrix is ​​calculated to obtain the first Pearson coefficient and the second Pearson coefficient. The sum of the absolute values ​​of the first Pearson coefficient and the second Pearson coefficient is calculated to obtain the comprehensive correlation strength. For each sample feature dimension i, the feature dimension j with the largest comprehensive correlation strength is selected. The index coordinates corresponding to the sample feature and the feature with the strongest correlation are added to the rotation plane candidate set. Through the rotation angle sequence and the rotation plane candidate set, a rotation matrix sequence is constructed. Each matrix in the rotation matrix sequence is multiplied to obtain an orthogonal matrix. The row dimension and column dimension of the orthogonal matrix are equal to the feature dimensions of the historical disaster sequence matrix. The orthogonal feature encoding unit is used to encode the historical disaster sequence matrix according to the orthogonal matrix to obtain the first disaster matrix; The orthogonal matrix is ​​transposed to obtain the transposed orthogonal matrix. The historical disaster sequence matrix is ​​then transformed and encoded using the transposed orthogonal matrix to obtain the transformed historical disaster sequence matrix, which serves as the first disaster matrix. The transformation encoding expression for the first disaster matrix is ​​as follows: ； in, This is the first disaster matrix. This is a matrix of historical disaster sequences. It is a transpose orthogonal matrix; The intelligent decision-making module includes a state construction unit, an action generation unit, and a probability evaluation unit; The probability evaluation unit is used to optimize the monitoring action using a greedy strategy to obtain an optimized monitoring sequence; Based on the remaining budget and iteration step size, iterate and optimize starting from the first monitoring round: Based on the state of the previous moment and the selected monitoring actions, the state construction unit is invoked, the current state component is recalculated, the monitoring action with the highest probability is selected, and the spatial unit is selected according to the priority score of the spatial unit in the regional spatial feature matrix based on the current monitoring action and the score, until the cumulative estimated cost of the spatial unit exceeds the preset remaining budget, then the iteration is terminated, and the selected actions in each iteration step are added to the optimized monitoring sequence in sequence to obtain the optimized monitoring sequence. The scheme decision module includes an orthogonal inverse transformation unit and an analysis unit; The orthogonal inverse transformation unit is used to restore the optimized monitoring sequence through orthogonal inverse transformation to obtain the optimized historical disaster sequence matrix; Encode each monitoring action in the optimized monitoring sequence by encoding the monitoring technology type, target spatial unit set, estimated cost vector, and time arrangement vector to generate an action coding vector. Then, concatenate the action codes of each round in chronological order to form an action coding sequence. The action coding sequence and the most recent L codes of the first disaster matrix are input into the encoder network. The latent variables are sampled using a reparameterized Gaussian distribution. An incremental feature matrix is ​​generated in the coding space through the decoder network. The incremental feature matrix is ​​inversely transformed by the orthogonal matrix output by the orthogonal matrix construction unit to calculate the scheme increment matrix of the original task space. The scheme increment matrix is ​​superimposed with the historical disaster sequence matrix to obtain the optimized historical disaster sequence matrix.

2. The cloud computing-based geological hazard risk assessment system as described in claim 1, characterized in that, The data processing module is used to acquire geological disaster monitoring data and perform data block encoding processing to obtain a historical disaster sequence matrix and a regional spatial feature matrix. The target monitoring area is divided into M two-dimensional grid units, and the monitoring process is discretized into T monitoring rounds. Geological disaster monitoring is carried out in each grid unit according to the round number. Topographic data, meteorological and hydrological data, soil and rock mechanics data, slope stability monitoring data and historical disaster record data of the target area in each round are collected and saved as a geological disaster monitoring sequence. The geological disaster monitoring data is subjected to coordinate unification and standardized coding to obtain the regional spatial feature matrix of the current round. The regional spatial feature matrix has a dimension of M*K, where K is the dimension of the geological disaster monitoring sequence. Each completed geological disaster monitoring task is recorded as a historical disaster sequence. The historical disaster sequence includes the task time, technology used, target area, cost, abnormal indicators, and result evaluation. The historical disaster sequence is arranged in chronological order to obtain a historical disaster sequence matrix. The topographic data includes slope, aspect, elevation, surface curvature, and lithology. The meteorological and hydrological data include rainfall intensity, rainfall duration, evaporation, runoff coefficient, and groundwater level; The soil and geotechnical data include soil type, pore water pressure, internal friction angle, cohesion, and permeability coefficient. The slope stability monitoring data includes displacement monitoring values, stress and strain data, anchor cable load, and crack width; The historical disaster record data includes disaster type, occurrence time, scope of impact, loss assessment, and recurrence index.

3. The cloud computing-based geological hazard risk assessment system as described in claim 2, characterized in that, The parameter integration unit is used to construct a monitoring environmental parameter vector using the regional spatial feature matrix and the historical disaster sequence matrix; Regional complexity parameters are statistically analyzed from the regional spatial feature matrix, including geological complexity and data noise complexity. The logic for handling data noise complexity is as follows: Meteorological and hydrological data and slope stability monitoring data are extracted from the regional spatial feature matrix and the regional spatial signal-to-noise ratio is obtained by moving average filtering. The reciprocal of the regional spatial signal-to-noise ratio is then taken and normalized to obtain the data noise complexity. Monitoring and evaluation parameters are statistically analyzed from the historical disaster sequence matrix, including monitoring cost parameters and monitoring success rate parameters. The processing logic for the monitoring cost parameters is as follows: Costs are extracted from the historical disaster sequence matrix to obtain cost sequences. The average value of the cost sequences is calculated and normalized to obtain monitoring cost parameters. The processing logic for the monitoring success rate parameter is as follows: The results evaluation sequence is extracted from the historical disaster sequence matrix. The results evaluation includes successful early warning and missed reporting. Successful early warning and missed reporting are mapped to 1 and 0, respectively. The success rate of the results evaluation sequence is calculated and normalized as a monitoring success parameter. The regional complexity parameter and the monitoring and evaluation parameter are concatenated into a monitoring environment parameter vector.

4. The cloud computing-based geological hazard risk assessment system as described in claim 3, characterized in that, The state construction unit is used to construct an extended monitoring state vector based on the first disaster matrix and the current monitoring round; Based on the regional spatial feature matrix of the current round, calculate the mean vector, standard deviation vector and corresponding outlier index vector of each feature dimension to obtain statistics. Then, concatenate the statistics with the remaining budget, remaining time window and environmental constraint indicators to obtain the current state components. The monitoring status component is obtained by calculating the anomaly index based on the historical disaster sequence matrix. The monitoring status component is used to represent the coverage and regional priority distribution of the currently monitored area. The monitoring status component includes the monitoring hit rate and the monitoring occupancy rate. The monitoring hit rate is the ratio of the number of abnormal areas with detected anomalies to the number of monitored areas. The monitoring occupancy rate is the ratio of the area of ​​the monitored area to the total area of ​​the target monitoring area. Extract the row vectors corresponding to the most recent L monitoring records from the first disaster matrix as historical state components; Save the current state component, the monitoring state component, and the historical state component as an extended monitoring state vector.

5. The cloud computing-based geological hazard risk assessment system as described in claim 4, characterized in that, The action generation unit is used to calculate the monitoring probability of the extended monitoring state vector through a lightweight decision network to obtain the monitoring action and target space unit set for the current round. The extended monitoring state vector is input into a lightweight decision network, which outputs the monitoring technology probability and spatial unit priority probability. The processing logic of the lightweight decision network is as follows: The extended monitoring state vector is input into a fully connected layer, where feature fusion is performed to obtain an intermediate feature vector. This intermediate feature vector is then input into the monitoring technology head, which includes a fully connected layer and a softmax activation function. The head outputs the probability of each monitoring technology. The feature vector of each grid cell is extracted from the regional spatial feature matrix of the current round. The feature vector of each grid cell has a dimension of K. The intermediate layer feature vector is projected onto the same dimension K as the feature vector of the grid cell through a fully connected layer to obtain a query vector. The dot product of the query vector and the feature vector of each grid cell in the regional spatial feature matrix is ​​calculated to obtain the priority score of each grid cell. The priority score is then subjected to a sigmoid transformation to obtain the spatial cell priority probability and the monitoring technology probability. The monitoring action with the highest probability is selected as the monitoring action for the current round. The spatial unit priority scores are sorted in descending order to obtain the target spatial unit set. The monitoring action for the current round and the target spatial unit set are saved as monitoring actions.

6. The cloud computing-based geological hazard risk assessment system as described in claim 5, characterized in that, The parsing unit is used to perform structured parsing on the optimized historical disaster sequence matrix to generate target monitoring tasks; The last row vector of the optimized historical disaster sequence matrix is ​​parsed to map the monitoring scheme record corresponding to the last row vector to the target monitoring task. The target monitoring task includes the monitoring round, the type of technology used, the set of target spatial units, the suggested sampling density, the budget allocation, the time arrangement, and the evaluation of the prediction results.

Images