Geological disaster dynamic early warning model construction method and system based on multi-parameter fusion

Abstract

The invention relates to the technical field of geological disaster monitoring and early warning, and discloses a geological disaster dynamic early warning model construction method and system based on multi-parameter fusion. The method comprises the following steps: acquiring geological environment and disaster induction parameters, and identifying a geological disaster scene mode; analyzing corresponding model structure elements and data interface specifications according to the scene mode; according to the method, multi-source monitoring data is preprocessed and aligned to form a structured geological feature sequence. Potential and dominant features are mined from the sequence, and relevance of the potential and dominant features is quantified to obtain a feature relevance matrix; using the matrix to iteratively correct the structural elements of the model, and generating an optimized model skeleton. And inputting the structured feature sequence into the skeleton, driving the model to carry out self-organizing evolution through a preset calculation process until output convergence, and completing construction of the dynamic early warning model. According to the invention, dynamic generation and self-optimization of the early warning model for a specific geological scene are realized.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and early warning technology, specifically to a method and system for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion. Background Technology

[0002] Current geological disaster early warning models generally adopt a pre-defined, general architecture, such as fixed machine learning algorithms or neural network structures. When processing multi-source heterogeneous geological data, these models typically rely on a unified data preprocessing workflow and feature engineering methods, and the model's structure and parameters are determined before training. Existing technical solutions attempt to improve early warning accuracy by introducing more monitoring parameters or optimizing algorithm hyperparameters, but their model kernels and data interfaces lack deep adaptation to the specific physical mechanisms and scenario characteristics of geological disasters.

[0003] This fixed architectural model has inherent flaws. The model struggles to fundamentally adapt its structure to different disaster types, such as landslides, collapses, and debris flows, or to the evolution of the same disaster under different geological environments. Furthermore, the model building process is often unidirectional: from feature extraction to model training, lacking a closed-loop mechanism that feeds back from the inherent correlations of data to the model's structural design. This results in insufficient generalization ability when facing complex and unsteady geological system evolution, leading to delayed early warning results or a high false alarm rate.

[0004] This invention aims to address the problem of insufficient model adaptability caused by the diversity of geological scenes. Simultaneously, it addresses the shortcomings of traditional one-time modeling processes, such as rigid model structures and the inability to self-optimize and evolve based on real-time data characteristics. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a method for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion, the method comprising:

[0007] Geological environment parameters and disaster-inducing parameters are acquired, and based on these parameters, the geological disaster scenario patterns on which the early warning model is constructed are identified.

[0008] Based on the geological disaster scenario model, the model structure elements and data interface specifications required to construct the geological disaster dynamic early warning model are analyzed;

[0009] Based on the model structural elements and data interface specifications, multi-source geological monitoring data are preprocessed and structurally aligned to form a structured geological feature sequence.

[0010] In the structured geological feature sequence, potential and explicit features related to the evolution of geological stability are extracted, and the correlation between the potential and explicit features is quantified to obtain a feature correlation matrix;

[0011] Using the feature correlation matrix, the model structure elements are iteratively corrected to generate an optimized model skeleton;

[0012] The structured geological feature sequence is input into the optimized model skeleton, and the geological disaster dynamic early warning model is driven to self-organize and evolve through a preset calculation process until the model output converges, thus completing the dynamic construction of the geological disaster dynamic early warning model.

[0013] Preferably, the step of parsing out the model structure elements and data interface specifications required for constructing the dynamic early warning model of geological disasters based on the geological disaster scenario pattern includes:

[0014] The geological disaster scene pattern is decomposed into multiple sub-scene pattern units;

[0015] Analyze the geological response characteristics and data requirement features corresponding to each of the sub-scene mode units;

[0016] Based on the geological response characteristics and data requirements, the functional components and their hierarchical relationships that should be included in the dynamic early warning model of geological disasters are mapped out, and the functional components and their hierarchical relationships are defined as the structural elements of the model.

[0017] Meanwhile, based on the data requirement characteristics, format conversion rules, time sequence alignment rules, and missing value handling rules are set for geological monitoring data from different sources when input into the geological disaster dynamic early warning model. The format conversion rules, time sequence alignment rules, and missing value handling rules are collectively referred to as the data interface specification.

[0018] By integrating the model structure elements and the data interface specifications, a preliminary blueprint for the construction of the dynamic early warning model for geological disasters is formed.

[0019] Preferably, the step of preprocessing and aligning the multi-source geological monitoring data according to the model structural elements and data interface specifications to form a structured geological feature sequence includes:

[0020] According to the format conversion rules in the data interface specification, the multi-source geological monitoring data are uniformly converted into a preset standard data format;

[0021] Based on the time alignment rules in the data interface specification, all converted geological monitoring data are synchronized in the time dimension to ensure that the timestamps of each data point are consistent.

[0022] By applying the missing value handling rules in the data interface specification, the time-aligned geological monitoring data is repaired to generate a complete continuous monitoring data stream.

[0023] Based on the hierarchical relationship of the functional components in the model structure elements, data segments corresponding to each level of functional components are extracted from the complete continuous monitoring data stream;

[0024] Feature extraction is performed on the data fragments to generate a primary feature set reflecting changes in geological conditions;

[0025] According to the hierarchical order defined by the model structural elements, the primary feature set is sorted and reorganized to construct the structured geological feature sequence.

[0026] Preferably, in the structured geological feature sequence, potential and explicit features related to the evolution of geological stability are extracted, and the correlation between the potential and explicit features is quantified to obtain a feature correlation matrix, including:

[0027] The structured geological feature sequence is segmented using a sliding window to obtain multiple consecutive local feature subsequences;

[0028] Principal component analysis and independent component analysis are performed on each of the local feature subsequences. The features extracted by principal component analysis are classified into a set of dominant features, and the features extracted by independent component analysis are classified into a set of latent features.

[0029] Calculate the nonlinear correlation coefficient between each dominant feature in the set of dominant features and each latent feature in the set of latent features;

[0030] For each pair of dominant and latent features, a comprehensive correlation strength value is calculated by combining their nonlinear correlation coefficients and the similarity of the changing trends of the dominant and latent features over their respective time series.

[0031] The comprehensive correlation strength values ​​between all dominant features and all latent features are arranged into a two-dimensional matrix according to the dominant feature index and the latent feature index. The two-dimensional matrix is ​​the feature correlation matrix.

[0032] Preferably, the step of iteratively refining the model structure elements using the feature correlation matrix to generate an optimized model skeleton includes:

[0033] From the feature association matrix, feature pairs whose comprehensive association strength value exceeds a preset threshold are identified, and the explicit and potential features constituting these feature pairs are marked as key features;

[0034] Based on the location distribution of the key features in the structured geological feature sequence, the necessity and connection weight of each functional component in the model structural elements are reassessed.

[0035] An optimization function is established with the goal of improving model structure stability and feature representation ability, using the feature correlation matrix as input constraints.

[0036] The optimization function is solved by gradient descent, and the number, type, and connection relationships of functional components in the model structure elements are iteratively adjusted.

[0037] When the output value of the optimization function converges or reaches the preset number of iterations, the iteration stops, and the functional components and their connection relationships determined at this time and adjusted are used as the optimized model skeleton.

[0038] Preferably, the step of inputting the structured geological feature sequence into the optimized model skeleton, and driving the dynamic early warning model of geological disasters to undergo self-organized evolution through a preset calculation process until the model output converges, thereby completing the dynamic construction of the dynamic early warning model of geological disasters, includes:

[0039] Initialization phase: The structured geological feature sequence is divided into blocks according to time sequence and used as an input stream, which is then injected into the optimized model skeleton. Each functional component in the optimized model skeleton is assigned a random initial internal state.

[0040] Feedforward propagation phase: At each time step, a data block in the input stream passes sequentially through the functional components of each level of the optimized model skeleton. Each functional component calculates its output based on its current internal state and input data and passes it to the next level until the model output of the time step is generated.

[0041] Feedback and adjustment phase: The model output is compared with the actual geological disaster observation records or historical early warning tags at the corresponding time step, and the output error is calculated. The output error is propagated back along the optimized model skeleton to update the internal state of each functional component and the connection strength between components.

[0042] Self-organizing iteration phase: The feedforward propagation phase and feedback adjustment phase are repeatedly executed, and the internal state and connection strength of the functional components of the optimized model skeleton are continuously and dynamically adjusted, so that the error of the model output gradually decreases.

[0043] Convergence Judgment and Completion Stage: When the error change rate of the model output is lower than a preset threshold or the total error is less than a set threshold for multiple consecutive time steps, it is determined that the model output has converged. At this time, the final internal state and connection relationship of all functional components in the optimized model skeleton are fixed, and the resulting stable computational structure is the completed geological disaster dynamic early warning model.

[0044] Preferably, in the feedback adjustment phase, the output error propagates backward along the optimized model skeleton to update the internal state of each functional component and the connection strength between components, specifically including:

[0045] The output error is distributed layer by layer from the output layer to the input layer according to the hierarchical relationship of the functional components in the optimized model skeleton, and the responsibility error that each functional component should bear is calculated.

[0046] For each functional component in the optimized model skeleton, obtain its input value, output value, and current internal state parameters at the time step;

[0047] The adjustment amount of the internal state parameters of the functional component is calculated based on the responsibility error, the input value, the output value, and the preset learning rate coefficient.

[0048] The adjustment amount is applied to the internal state parameters of the functional component to complete the update of its internal state;

[0049] Meanwhile, for the connection path connecting two functional components, the adjustment amount of the connection strength is calculated based on the output of the upstream functional component and the responsibility error of the downstream functional component in the connection path;

[0050] The adjustment amount of the connection strength is applied to the corresponding connection path to complete the update of the connection strength.

[0051] Preferably, the step of acquiring geological environmental parameters and disaster-inducing parameters, and identifying the geological disaster scenario patterns on which the early warning model is based based on the geological environmental parameters and disaster-inducing parameters, includes:

[0052] The geological environment parameters are collected in real time from the geological monitoring network or retrieved from historical databases. The geological environment parameters include topographic slope, rock and soil type, and geological structure information.

[0053] The disaster-inducing parameters are obtained from professional monitoring equipment or disaster case databases, including rainfall intensity sequences, groundwater level changes, and seismic event records.

[0054] The geological environment parameters are subjected to spatial grid interpolation to form a spatially continuously distributed geological environment field;

[0055] The disaster-inducing parameters are processed into a time series and matched and fused with the spatially continuously distributed geological environment field in time and space.

[0056] Based on the fused data, pattern recognition technology was used to delineate typical areas with different combinations of geological conditions and inducing factors.

[0057] Each typical region is labeled with the types and intensity levels of geological disasters that have occurred or are likely to occur in the past. The definitions of "typical region and its corresponding geological disaster type and intensity level" are combined as the geological disaster scenario pattern on which the early warning model is based.

[0058] Preferably, the method further includes the step of verifying and updating the dynamic early warning model of geological disasters after its construction is completed.

[0059] Model validation steps: The completed geological disaster dynamic early warning model is tested using an independent geological monitoring dataset that was not involved in the construction process. The independent geological monitoring dataset is input to obtain the model early warning result sequence.

[0060] The model's early warning result sequence is compared with the actual disaster occurrence records corresponding to the independent geological monitoring dataset to calculate the early warning accuracy, false alarm rate, missed alarm rate, and early warning lead time index of the geological disaster dynamic early warning model.

[0061] Model update trigger judgment steps: Set the lower limit threshold of the early warning accuracy, the upper limit threshold of the false alarm rate, and the lower limit threshold of the early warning advance time indicator;

[0062] The model validation step is performed periodically or when significant changes in geological conditions are detected. If any calculated index exceeds its corresponding threshold, the model update process is triggered.

[0063] Model update execution steps: When the model update process is triggered, the recently acquired new geological environmental parameters, disaster-induced parameters and their corresponding disaster observation records are used as incremental data;

[0064] The incremental data is merged with the original constructed data to form a new training dataset;

[0065] Using the currently constructed dynamic early warning model for geological disasters as the initial state, and using the new training dataset, the dynamic construction process of the dynamic early warning model for geological disasters is re-executed to generate an updated dynamic early warning model for geological disasters.

[0066] Preferably, the present invention also includes a geological disaster dynamic early warning model construction system based on multi-parameter fusion. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the geological disaster dynamic early warning model construction method based on multi-parameter fusion as described above.

[0067] Compared with the prior art, the beneficial effects of the present invention are:

[0068] Based on real-time identified geological hazard scene patterns, matching model structural elements and data interface specifications are dynamically generated. The inter-layer topological relationships, feature fusion nodes, and external data cleaning and alignment rules are all established in real-time according to the physical mechanisms of the scene. This allows the model to deeply integrate the evolutionary patterns of specific hazard types from the very beginning of its construction; its structure is no longer a fixed abstract framework, but a computational mapping of the scene's mechanisms. This method enhances the model's relevance to different combinations of geological conditions and triggering factors, improving the accuracy of the initial model's state representation.

[0069] By utilizing feature correlation matrices mined from structured data, the initially constructed model skeleton is iteratively refined, driving the model to complete its final self-organizing evolution. The quantified correlations between features serve as feedback signals, guiding the initialization of connection weights and the strengthening or weakening of important substructures within the model. In this dynamic closed loop, the model not only learns parameters, but its microstructure also adaptively fine-tunes based on the patterns continuously revealed in the data stream. The resulting early warning model possesses self-improvement capabilities in dynamic environments, and its decision boundary continuously optimizes with the accumulation of monitoring data, thereby enhancing the accuracy and timeliness of early warning signals in complex, time-varying processes. Attached Figure Description

[0070] Figure 1 This is a schematic diagram illustrating the working principle of the geological disaster dynamic early warning model construction method based on multi-parameter fusion described in this invention.

[0071] Figure 2 A flowchart for data preprocessing and structure alignment;

[0072] Figure 3 A flowchart for feature mining and correlation quantification;

[0073] Figure 4 A grouped bar chart showing the performance improvement of updating trigger conditions for geological disaster early warning models;

[0074] Figure 5 This is a heatmap showing the intensity of characteristic correlations in a geological disaster early warning model. Detailed Implementation

[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] Please see Figure 1 This invention provides a method for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion. The method includes: First, acquiring geological environmental parameters, including topographic slope, soil and rock type, and geological structure information, as well as disaster-inducing parameters, including rainfall intensity sequences, groundwater level changes, and seismic event records, from geological monitoring networks, historical databases, and professional monitoring equipment. Based on these parameters, the geological disaster scenario patterns upon which the early warning model relies are identified using pattern recognition technology. Subsequently, based on the identified geological disaster scenario patterns, the model structure elements and data interface specifications required for constructing the dynamic early warning model are analyzed. According to the analyzed model structure elements and data interface specifications, preprocessing and structure alignment operations, such as format conversion, temporal alignment, and missing value handling, are performed on multi-source geological monitoring data from different sensors to form a regular structured geological feature sequence. In the formed structured geological feature sequence, explicit and potential features related to geological stability evolution are mined using methods such as principal component analysis and independent component analysis, and the correlation between these two types of features is quantified to obtain a feature correlation matrix. Using the obtained feature correlation matrix, and aiming to optimize the model's structural stability and feature representation ability, the model's structural elements are iteratively corrected to generate an optimized model skeleton. Finally, the structured geological feature sequence is used as the input stream and injected into the optimized model skeleton. Through a computational process including initialization, feedforward propagation, feedback adjustment, and self-organizing iteration, the model is driven to undergo self-organizing evolution until the model output converges. At this point, the model structure is fixed, completing the dynamic construction of the geological disaster dynamic early warning model.

[0077] In one embodiment of the present invention, see [reference] Figure 2The identified geological hazard scenario patterns are decomposed into multiple sub-scenario pattern units. The geological response characteristics and data requirement features corresponding to each sub-scenario pattern unit are analyzed. Based on the analyzed geological response characteristics and data requirement features, the functional components and their hierarchical relationships that should be included in the dynamic early warning model of geological hazards are mapped, and these functional components and their hierarchical relationships are defined as model structural elements. Simultaneously, based on the data requirement characteristics, format conversion rules, time sequence alignment rules, and missing value handling rules are set for geological monitoring data from different sources when input into the dynamic early warning model of geological hazards; these rules are collectively referred to as data interface specifications. The model structural elements and data interface specifications are integrated to form a preliminary blueprint for the construction of the dynamic early warning model of geological hazards.

[0078] Following the format conversion rules in the data interface specification, multi-source geological monitoring data are uniformly converted into a preset standard data format. Based on the time alignment rules in the data interface specification, all converted geological monitoring data are synchronized in the time dimension to ensure consistency of timestamps for each data point. Applying the missing value handling rules in the data interface specification, data repair is performed on the time-aligned geological monitoring data to generate a complete continuous monitoring data stream. According to the hierarchical relationship of functional components in the model structure elements, data segments corresponding to each level of functional components are extracted from the complete continuous monitoring data stream. Feature extraction is performed on the extracted data segments to generate a primary feature set reflecting changes in geological state. The primary feature set is sorted and reorganized according to the hierarchical order defined by the model structure elements to construct a structured geological feature sequence. In practical implementation, geological disaster scenario models are parsed into model structural elements and data interface specifications. Based on these, data is processed to form a structured geological feature sequence. The geological disaster scenario model is decomposed into multiple sub-scenario model units. The geological response characteristics and data requirement characteristics corresponding to each sub-scenario model unit are analyzed. Based on the geological response characteristics and data requirement characteristics, the functional components and their hierarchical relationships that should be included in the dynamic early warning model of geological disasters are mapped. The functional components and their hierarchical relationships are defined as model structural elements. At the same time, based on the data requirement characteristics, the format conversion rules, time sequence alignment rules, and missing value handling rules for geological monitoring data from different sources when input into the dynamic early warning model of geological disasters are set. The format conversion rules, time sequence alignment rules, and missing value handling rules are collectively referred to as data interface specifications. The integration of model structural elements and data interface specifications forms the initial blueprint for the construction of the dynamic early warning model of geological disasters.

[0079] In some embodiments, multi-source geological monitoring data are uniformly converted into a preset standard data format according to the format conversion rules in the data interface specification. Based on the time alignment rules in the data interface specification, all converted geological monitoring data are synchronized in the time dimension to ensure that the timestamps of each data point are consistent. The missing value handling rules in the data interface specification are applied to repair the time-aligned geological monitoring data to generate a complete continuous monitoring data stream. It can be understood that the missing value handling rules involve interpolation calculations, such as using linear interpolation to fill missing data points. The linear interpolation formula is expressed as follows:

[0080] ;

[0081] in: Indicates time The interpolation results, and They represent time respectively and The known monitoring values, and Based on the hierarchical relationship of functional components in the model structure elements, data segments corresponding to each level of functional components are extracted from the complete continuous monitoring data stream. Feature extraction is performed on the data segments to generate a primary feature set reflecting changes in geological state. The primary feature set is sorted and reorganized according to the hierarchical order defined by the model structure elements to construct a structured geological feature sequence.

[0082] Optionally, the format conversion rules include the step of converting the raw binary data into a decimal floating-point array. Optionally, the feature extraction process calculates the mean, variance, or peak characteristics of the data segments based on statistical methods. In some embodiments, the functional components in the model structure elements correspond to different geological parameter processing modules, such as slope analysis modules and rainfall response modules. It can be understood that the time alignment rules in the data interface specification require all data sources to be resampled at one-minute intervals, and the missing value handling rules stipulate that data segments with more than three consecutive missing time points are filled using trend extrapolation.

[0083] In one embodiment of the present invention, see [reference] Figure 3The structured geological feature sequence is segmented using a sliding window to obtain multiple continuous local feature subsequences. Principal component analysis (PCA) and independent component analysis (ICA) are performed on each local feature subsequence. Features extracted through PCA are categorized into a dominant feature set, and features extracted through ICA are categorized into a latent feature set. The nonlinear correlation coefficient between each dominant feature in the dominant feature set and each latent feature in the latent feature set is calculated. For each pair of dominant and latent features, a comprehensive correlation strength value is calculated by combining their nonlinear correlation coefficients and the similarity of their temporal trends. The comprehensive correlation strength values ​​between all dominant and latent features are arranged into a two-dimensional matrix according to the dominant and latent feature indices; this two-dimensional matrix is ​​the feature correlation matrix. In practice, potential and explicit features related to geological stability evolution are extracted from the structured geological feature sequence and their correlation is quantified to obtain a feature correlation matrix. The structured geological feature sequence is segmented by a sliding window to obtain multiple continuous local feature subsequences. Principal component analysis and independent component analysis are performed on each local feature subsequence. Features extracted by principal component analysis are classified into a set of explicit features, and features extracted by independent component analysis are classified into a set of implicit features. The nonlinear correlation coefficient between each explicit feature in the explicit feature set and each implicit feature in the implicit feature set is calculated. For each pair of explicit and implicit features, a comprehensive correlation strength value is calculated by combining its nonlinear correlation coefficient and the similarity of the change trends of explicit and implicit features in their respective time series. The comprehensive correlation strength values ​​between all explicit features and all implicit features are arranged into a two-dimensional matrix according to the explicit feature index and the implicit feature index. The two-dimensional matrix is ​​the feature correlation matrix.

[0084] In some embodiments, sliding window segmentation uses a fixed-length time window to slide along a structured geological feature sequence, with each sliding step being smaller than the window length to generate overlapping local feature subsequences. It can be understood that the choice of window length is related to the periodic characteristics of geological hazard evolution. In some embodiments, principal component analysis is used to reduce dimensionality and extract dominant features representing the main variance direction of the data, while independent component analysis is used to separate statistically independent signal sources, i.e., latent features. The nonlinear correlation coefficient is calculated using the mutual information method to measure the statistical dependence between dominant and latent features. For the comparison of the changing trends of dominant and latent features, the time series of the two features are aligned using a dynamic time warping algorithm, and the minimum cumulative distance is calculated.

[0085] Optionally, the calculation of the comprehensive correlation strength value integrates two indicators: nonlinear correlation coefficient and trend similarity. Optionally, the formula for calculating the comprehensive correlation strength value is designed as follows:

[0086] ;

[0087] in: This represents the overall correlation strength value. This represents the nonlinear correlation coefficient calculated using mutual information. This represents the distance of the changing trend calculated through dynamic time warping. and The preset weighting coefficients are used to balance the contributions of the two parts. This represents an exponential function, which can be understood as a non-linear correlation coefficient. The value range is between 0 and positive numbers, and the trend of change is far from... For non-negative values, the exponential function maps distance to a similarity metric, with weighting coefficients. and The sum of is a fixed constant.

[0088] In practical implementation, the feature correlation matrix serves as the input constraint for subsequent iterative correction of model structural elements. The rows of the feature correlation matrix correspond to features in the explicit feature set, and the columns correspond to features in the potential feature set. The value of each element in the matrix represents the comprehensive correlation strength value between the corresponding explicit feature and the potential feature. It can be understood that feature pairs with higher comprehensive correlation strength values ​​indicate a strong coupling relationship between explicit monitoring indicators and hidden driving factors within the geological system, providing a quantitative basis for optimizing the connection relationship of functional components of the early warning model. The calculation process is entirely data-driven and does not require pre-setting a geomechanical model.

[0089] In one embodiment of the present invention, feature pairs with a comprehensive correlation strength value exceeding a preset threshold are identified from the feature correlation matrix, and the explicit and potential features constituting these feature pairs are marked as key features. Based on the positional distribution of the key features in the structured geological feature sequence, the necessity and connection weights of each functional component in the model structural elements are reassessed. An optimization function is established with the model structural stability and feature expressive ability as objectives, using the feature correlation matrix as input constraints. The optimization function is solved using the gradient descent method, iteratively adjusting the number, type, and connection relationships of functional components in the model structural elements. When the output value of the optimization function converges or reaches a preset number of iterations, the iteration stops, and the adjusted functional components and their connection relationships determined at this point are used as the optimized model skeleton. In practice, the feature correlation matrix is ​​used to iteratively correct the model structural elements to generate an optimized model skeleton. Feature pairs with a comprehensive correlation strength value exceeding a preset threshold are identified from the feature correlation matrix. The explicit and potential features constituting these feature pairs are marked as key features. The necessity and connection weight of each functional component in the model structural elements are re-evaluated based on the positional distribution of key features in the structured geological feature sequence. An optimization function is established with the model structural stability and feature expressive ability as the objectives, and the feature correlation matrix is ​​used as the input constraint. The optimization function is solved by the gradient descent method, and the number, type, and connection relationship of functional components in the model structural elements are iteratively adjusted. The iteration stops when the output value of the optimization function converges or reaches the preset number of iterations. The adjusted functional components and their connection relationships determined at this time are used as the optimized model skeleton.

[0090] In some embodiments, the preset threshold is set based on the statistical distribution of all comprehensive correlation strength values ​​in the feature correlation matrix. For example, the threshold is set as the sum of the average of all comprehensive correlation strength values ​​and twice the standard deviation. The key feature marking process generates a key feature list. Each entry in the key feature list records the explicit feature identifier, the potential feature identifier, and their comprehensive correlation strength value. The necessity of each functional component in the model structural elements is reassessed based on the location distribution of the key features in the structured geological feature sequence. The specific assessment method involves checking whether the data range processed by each functional component covers the location of the key features. If a functional component does not cover the location of any key features, its necessity is assessed as low.

[0091] Optionally, the design of the optimization function considers both the complexity of the model structure and the utilization of feature association information. Optionally, the mathematical form of the optimization function is expressed as follows:

[0092] ;

[0093] in: This represents the output value of the optimization function. This represents an adjacency matrix composed of the connection relationships between functional components. Representation matrix The rank is used to measure the structural complexity of a model. This represents the initial connection weight matrix between functional components. This represents the association constraint matrix derived from the feature association matrix and related to the mapping relationship of functional components. The symbol is... The Hadamard product of matrices is represented by element-wise multiplication. Denotes the Frobenius norm of a matrix. and To adjust the hyperparameters for the strength of the two penalty terms, Represents a matrix of all 1s, and an affinity constraint matrix. The larger the value of the element, the more the connection between the corresponding functional components should be weakened or removed. Gradient descent calculates the output value of the optimization function. For adjacency matrix and connection weight matrix The partial derivatives of the elements are used to guide iterative adjustments.

[0094] In practice, the iterative adjustment process adds new functional components to handle strongly correlated feature pairs and removes functional components deemed less necessary. The connections between functional components are added, deleted, or have their weights modified based on the gradient direction of the optimization function. When the output value of the optimization function... The iteration terminates when the change in consecutive iterations is less than the set convergence criterion or when the algorithm execution reaches the preset maximum number of iterations. At the time of termination, the set of functional components, the connection topology between functional components, and the numerical values ​​of the connection weights together constitute the optimized model skeleton.

[0095] In one embodiment of the present invention, the model self-organizing evolution construction stage begins with initialization. The structured geological feature sequence is divided into temporal blocks and injected as an input stream into the optimized model skeleton. Each functional component in the optimized model skeleton is assigned a random initial internal state. During the feedforward propagation stage, the input data block at each time step sequentially passes through the functional components of each level of the optimized model skeleton. Each functional component calculates its output based on its current internal state and input data and passes it to the next level until the model output for that time step is generated. During the feedback adjustment stage, the model output is compared with the actual geological disaster observation records or historical early warning tags for the corresponding time step to calculate the output error. The output error propagates backward along the optimized model skeleton to update the internal states of each functional component and the connection strength between components. Specifically, the output error is distributed layer by layer from the output layer to the input layer according to the hierarchical relationship of the functional components to calculate the responsibility error of each layer. For each functional component, the adjustment amount of the internal state parameters is calculated and updated based on its responsibility error, the input value of the current time step, the output value, and the preset learning rate coefficient. Simultaneously, for the connection paths between components, the adjustment amount of the connection strength is calculated and updated based on the output of the upstream component and the responsibility error of the downstream component. During the self-organizing iterative phase, feedforward propagation and feedback adjustment are repeatedly executed, continuously and dynamically adjusting the internal states and connection strengths of functional components, gradually reducing the model output error. In the convergence determination and completion phase, when the rate of change of the model output error over multiple consecutive time steps is lower than a preset threshold or the total error is less than a set threshold, the model output is determined to have converged. At this point, the final internal states and connection relationships of all functional components in the optimized model skeleton are fixed, and the resulting stable computational structure is the completed dynamic early warning model for geological disasters.

[0096] After the dynamic early warning model for geological disasters is constructed, it is validated and updated. The model validation step uses an independent geological monitoring dataset not involved in the construction process to test the completed model, obtaining a sequence of early warning results. This sequence is compared with actual disaster occurrence records to calculate the early warning accuracy, false alarm rate, missed alarm rate, and early warning lead time. The model update trigger step sets a lower threshold for early warning accuracy, an upper threshold for false alarm rate, and a lower threshold for early warning lead time. The model validation step is executed periodically or when geological conditions change significantly. If any calculated indicator exceeds its corresponding threshold, the model update process is triggered. The model update execution step, when the update process is triggered, uses recently acquired new parameters and disaster observation records as incremental data, merges them with the original construction data to form a new training dataset, and re-executes the entire process from identifying geological disaster scene patterns to completing the dynamic model construction using the new training dataset, using the currently constructed model as the initial state, to generate an updated dynamic early warning model for geological disasters.In practical implementation, the structured geological feature sequence is input into the optimized model skeleton and driven by a preset calculation process to self-organize and evolve until convergence. After model construction, verification and updates are performed. In the initialization phase, the structured geological feature sequence is divided into temporal blocks and injected into the optimized model skeleton as an input stream. Each functional component in the optimized model skeleton is assigned a random initial internal state. In the feedforward propagation phase, a data block in the input stream at each time step passes through the functional components at each level of the optimized model skeleton. Each functional component calculates its output based on its current internal state and input data and passes it to the next level until the model output for that time step is generated. In the feedback adjustment phase, the model output is compared with the actual geological disaster observation records or historical early warning tags at the corresponding time step to calculate the output error. The output error propagates backward along the optimized model skeleton to update the internal state of each functional component and the connection strength between components. Specifically, the update process involves distributing the output error layer by layer from the output layer to the input layer according to the hierarchical relationship of the functional components in the optimized model skeleton, calculating the responsibility error that each functional component should bear. Each functional component in the model skeleton acquires its input value, output value, and current internal state parameters at each time step. Based on the responsibility error, input value, output value, and preset learning rate coefficient, the adjustment amount of the internal state parameters of the functional component is calculated and applied to the internal state parameters of the functional component to complete the update of its internal state. At the same time, for the connection path connecting two functional components, the adjustment amount of the connection strength is calculated based on the output of the upstream functional component and the responsibility error of the downstream functional component and applied to the corresponding connection path to complete the update of the connection strength. In the self-organizing iteration stage, the feedforward propagation stage and the feedback adjustment stage are repeatedly executed, so that the internal state and connection strength of the functional components of the optimized model skeleton are continuously dynamically adjusted and the error of the model output gradually decreases. In the convergence judgment and completion stage, when the error change rate of the model output of multiple consecutive time steps is lower than the preset threshold or the total error is less than the set threshold, the model output is judged to have converged. At this time, the stable computational structure formed by the final internal state and connection relationship of all functional components in the optimized model skeleton is the completed geological disaster dynamic early warning model.

[0097] In some embodiments, the adjustment of the internal state parameters of functional components is calculated using a gradient-based method. It is understood that the adjustment of internal state parameters aims to reduce the contribution of the functional component at the current time step to the final output error. The adjustment of connection strength considers the efficiency of information transmission and the error requirements of downstream components. Internal state updates and connection strength updates can be performed synchronously or asynchronously. In a landslide early warning scenario, the optimized model skeleton contains a hierarchy of functional components for processing data from different sources such as rainfall seepage, surface displacement, and underground stress. Structured geological feature sequences are organized into continuous data blocks at hourly intervals and input into the model. The model output can be a landslide risk level such as "stable," "caution," "warning," or "alarm." The error of the model output is calculated by comparing the predicted risk level with the actual disaster state confirmed by field verification or monitoring equipment. The actual disaster state is calibrated by historical or real-time monitored landslide events.

[0098] Optionally, the formula for calculating the internal state parameter adjustment can be designed as follows:

[0099] ;

[0100] in: Represents the internal state parameter vector The adjustment amount at the current time step, This represents the preset learning rate coefficient. This indicates the error margin of responsibility currently borne by the functional component. This represents the output vector of the current functional component at the current time step, with the symbol... Indicates liability error For the output vector gradient, sign Represents the output vector For the internal state parameter vector The Jacobian matrix.

[0101] In practical implementation, the completed dynamic early warning model for geological disasters needs to undergo verification and update steps. The model verification step uses an independent geological monitoring dataset that was not involved in the construction process to test the completed dynamic early warning model for geological disasters. The independent geological monitoring dataset is input to obtain the model's early warning result sequence. The model's early warning result sequence is compared with the corresponding real disaster occurrence records in the independent geological monitoring dataset to calculate the early warning accuracy, false alarm rate, missed alarm rate, and early warning lead time index of the dynamic early warning model for geological disasters. The model update trigger judgment step sets the lower limit threshold for early warning accuracy, the upper limit threshold for false alarm rate, and the lower limit threshold for early warning lead time index. The model verification step is executed periodically or when a significant change in geological conditions is detected. If any of the calculated indicators exceeds its corresponding threshold, the model update process is triggered. When the model update process is triggered, the recently acquired new geological environmental parameters, disaster-inducing parameters, and their corresponding disaster observation records are used as incremental data. The incremental data is merged with the original construction data to form a new training dataset. The currently constructed dynamic early warning model for geological disasters is used as the initial state, and the dynamic construction process is re-executed using the new training dataset to generate the updated dynamic early warning model for geological disasters. See Table 1.

[0102] Table 1: Calculation Table of Model Validation Indicators

[0103] In some embodiments, the early warning accuracy rate is calculated based on the proportion of records in Table 1 whose "Does it correspond to an actual event" column is "Yes" and whose early warning risk level reaches a preset threshold to the total number of verification events. The false alarm rate is calculated based on the proportion of records whose "Does it correspond to an actual event" column is "No" to the total number of verification events. The false alarm rate is calculated based on the proportion of records whose actual events occurred but whose "Model early warning time" is empty to the total number of verification events. The early warning lead time index is usually calculated as the average or median of the early warning lead time of all successful early warning events.

[0104] See Figure 4 This is a bar chart showing the performance improvement of geological disaster early warning model update trigger conditions, grouped by performance. It primarily displays the changes in model performance and early warning lead time under different update trigger scenarios. Updates triggered when accuracy is below a threshold show the highest performance improvement and the greatest increase in early warning lead time. By comparing these two indicators, the chart intuitively demonstrates the differences in optimization effects under different update trigger conditions. This type of chart is used in the model update phase to help evaluate the effectiveness of update strategies and support the update scheduling logic of "prioritizing responses to high-impact trigger conditions." This chart reflects the "trigger priority" of model updates: updates triggered by anomalies in core indicators (such as accuracy) bring more significant performance improvements, providing a strategic basis for the dynamic maintenance of geological disaster early warning models.

[0105] In one embodiment of the present invention, geological environmental parameters are collected in real time from a geological monitoring network or retrieved from a historical database. These parameters include topographic slope, soil and rock type, and geological structural information. Disaster-inducing parameters, including rainfall intensity sequences, groundwater level changes, and seismic event records, are obtained from professional monitoring equipment or a disaster case database. Spatial grid interpolation is performed on the geological environmental parameters to form a spatially continuous geological environmental field. The disaster-inducing parameters are then processed into a time series and matched and fused with the spatially continuous geological environmental field in both time and space. Pattern recognition technology is used on the fused data to delineate typical regions with different combinations of geological conditions and inducing factors. Each typical region is labeled with the types and intensity levels of geological disasters that have occurred or are likely to occur in the past. The definition combination of "typical region and its corresponding geological disaster type and intensity level" is used as the geological disaster scenario pattern upon which the early warning model is constructed. In practice, geological environmental parameters and disaster-inducing parameters are acquired, and the geological disaster scenario patterns on which the early warning model is based are identified. Geological environmental parameters, including topographic slope, rock and soil type, and geological structure information, are collected in real time from the geological monitoring network or retrieved from historical databases. Disaster-inducing parameters, including rainfall intensity sequences, groundwater level changes, and seismic event records, are obtained from professional monitoring equipment or disaster case databases. Spatial grid interpolation is performed on the geological environmental parameters to form a spatially continuous geological environmental field. The disaster-inducing parameters are temporally sequenced and matched and fused with the spatially continuous geological environmental field in time and space. Pattern recognition technology is used to divide the fused data into typical areas with different combinations of geological conditions and inducing factors. Each typical area is labeled with the types and intensity levels of geological disasters that have occurred or are likely to occur in the past. The combination of the definitions of "typical areas and their corresponding geological disaster types and intensity levels" is used as the geological disaster scenario patterns on which the early warning model is based.

[0106] In some embodiments, spatial gridding interpolation uses the inverse distance weighting method to convert the geological environmental parameters measured at discrete points into regular grid data covering the entire study area. The inverse distance weighting interpolation formula is as follows:

[0107] ;

[0108] in: This represents the parameter estimates for the grid points to be interpolated. Indicates the first Parameter values ​​of a known measurement point Indicates the first Spatial distance from known measurement points to the grid points to be interpolated The distance decay power parameter is typically set to 2. This indicates the number of known measurement points involved in the interpolation calculation. The time-series processing unifies the disaster-induced parameters with different sampling frequencies to a common time reference through resampling. The matching and fusion process ensures that each spatial grid cell has the corresponding geological environment parameter value and the synchronous disaster-induced parameter value at each time step.

[0109] Optionally, the pattern recognition technology uses clustering algorithms, such as the K-means algorithm, to partition the fused spatiotemporal data. It can be understood that the clustering algorithm takes the geological environment parameter feature vector of each spatial grid unit and the disaster-induced parameter statistical feature vector within a time window as input. Each cluster output by the clustering algorithm corresponds to a typical region with similar geological conditions and response characteristics of inducing factors. The labeling of the geological disaster type and intensity level of each typical region is completed based on the historical disaster records and expert knowledge base of the region. In one implementation case, one typical region is labeled as "Type: shallow soil landslide; Intensity level: high frequency and low intensity", and another typical region is labeled as "Type: rock collapse; Intensity level: low frequency and high intensity".

[0110] In practical implementation, the geological environment field is a static or quasi-static spatial distribution field, while the disaster-inducing parameters are dynamic time series. Spatiotemporal matching and fusion create a multi-dimensional data cube with timestamp alignment. The dimensions of the data cube include spatial grid coordinates, time, and various parameter channels. Pattern recognition technology is used to divide typical regions, realizing the regional generalization of complex and heterogeneous geological environments. The geological disaster scenario patterns constituted by typical regions and their annotation information are the direct basis for subsequent analysis of model structure elements. It can be understood that different geological disaster scenario patterns will correspond to the analysis of early warning model construction blueprints with different functional component configurations and data interface specifications. In practical implementation, based on different geological disasters... The specific implementation of the scenario pattern analysis and early warning model construction blueprint is as follows: For each typical area divided and labeled by pattern recognition technology, such as the "shallow soil landslide" or "rock collapse" scenario pattern, the corresponding geological response characteristics are first analyzed. For example, the shallow soil landslide scenario pattern emphasizes the coupled response of rainfall seepage and surface displacement, while the rock collapse scenario pattern focuses on the relationship between vibration events and rock mass stress changes. Based on these characteristics, the functional component configuration that should be included in the dynamic early warning model for geological disasters is mapped. For example, a seepage analysis module and a displacement monitoring module are configured for the landslide scenario, and their hierarchical relationship is defined. For the collapse scenario, a stress calculation module and a vibration response module are configured, and the data flow direction is determined. At the same time, according to the data requirement characteristics of the scenario pattern, such as the landslide scenario requiring real-time processing of high-frequency rainfall data and the collapse scenario relying on the precise time series of vibration records, corresponding data interface specifications are set, including rules for converting raw monitoring data into a unified floating-point format, rules for time series alignment at minute intervals, and rules for repairing missing data using trend extrapolation. Ultimately, by integrating the configurations of these functional components and data interface specifications, a blueprint for building early warning models for specific scenarios is formed, ensuring that the model's structural elements and data flow processing are deeply adapted to the actual evolution mechanism of geological disasters.

[0111] See Figure 5 This is a heatmap showing the intensity of feature correlations in a geological disaster early warning model, primarily illustrating the degree of correlation between different geological features. The symmetrical heatmap visually presents the bidirectional correlations between features, with red areas clearly identifying core correlation pairs. This type of chart is used in the feature correlation matrix construction phase to help uncover potential correlation patterns among geological features, supporting feature fusion and structural optimization of the early warning model. This chart reflects the "synergistic influence logic" of geological features: strongly correlated feature combinations (such as topography-soil mass-geological structure) are the core driving factors for geological disasters, and the model needs to focus on integrating information from these features to improve early warning accuracy.

[0112] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0113] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion, characterized in that, include: Geological environment parameters and disaster-inducing parameters are acquired, and based on these parameters, the geological disaster scenario patterns on which the early warning model is constructed are identified. Based on the geological disaster scenario model, the model structure elements and data interface specifications required to construct the geological disaster dynamic early warning model are analyzed; Based on the model structural elements and data interface specifications, multi-source geological monitoring data are preprocessed and structurally aligned to form a structured geological feature sequence. In the structured geological feature sequence, potential and explicit features related to the evolution of geological stability are extracted, and the correlation between the potential and explicit features is quantified to obtain a feature correlation matrix; Using the feature correlation matrix, the model structure elements are iteratively corrected to generate an optimized model skeleton; The structured geological feature sequence is input into the optimized model skeleton, and the geological disaster dynamic early warning model is driven to self-organize and evolve through a preset calculation process until the model output converges, thus completing the dynamic construction of the geological disaster dynamic early warning model.

2. The method for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion as described in claim 1, characterized in that, The process involves analyzing the model structure elements and data interface specifications required for constructing the dynamic early warning model of geological disasters based on the geological disaster scenario pattern, including: The geological disaster scene pattern is decomposed into multiple sub-scene pattern units; Analyze the geological response characteristics and data requirement features corresponding to each of the sub-scene mode units; Based on the geological response characteristics and data requirements, the functional components and their hierarchical relationships that should be included in the dynamic early warning model of geological disasters are mapped out, and the functional components and their hierarchical relationships are defined as the structural elements of the model. Meanwhile, based on the data requirement characteristics, format conversion rules, time sequence alignment rules, and missing value handling rules are set for geological monitoring data from different sources when input into the geological disaster dynamic early warning model. The format conversion rules, time sequence alignment rules, and missing value handling rules are collectively referred to as the data interface specification. By integrating the model structure elements and the data interface specifications, a preliminary blueprint for the construction of the dynamic early warning model for geological disasters is formed.

3. The method for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion as described in claim 2, characterized in that, Based on the model structural elements and data interface specifications, the multi-source geological monitoring data is preprocessed and structurally aligned to form a structured geological feature sequence, including: According to the format conversion rules in the data interface specification, the multi-source geological monitoring data are uniformly converted into a preset standard data format; Based on the time alignment rules in the data interface specification, all converted geological monitoring data are synchronized in the time dimension to ensure that the timestamps of each data point are consistent. By applying the missing value handling rules in the data interface specification, the time-aligned geological monitoring data is repaired to generate a complete continuous monitoring data stream. Based on the hierarchical relationship of the functional components in the model structure elements, data segments corresponding to each level of functional components are extracted from the complete continuous monitoring data stream; Feature extraction is performed on the data fragments to generate a primary feature set reflecting changes in geological conditions; According to the hierarchical order defined by the model structural elements, the primary feature set is sorted and reorganized to construct the structured geological feature sequence.

4. The method for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion as described in claim 1, characterized in that, In the structured geological feature sequence, potential and explicit features related to the evolution of geological stability are extracted, and the correlation between these potential and explicit features is quantified to obtain a feature correlation matrix, including: The structured geological feature sequence is segmented using a sliding window to obtain multiple consecutive local feature subsequences; Principal component analysis and independent component analysis are performed on each of the local feature subsequences. The features extracted by principal component analysis are classified into a set of dominant features, and the features extracted by independent component analysis are classified into a set of latent features. Calculate the nonlinear correlation coefficient between each dominant feature in the set of dominant features and each latent feature in the set of latent features; For each pair of dominant and latent features, a comprehensive correlation strength value is calculated by combining their nonlinear correlation coefficients and the similarity of the changing trends of the dominant and latent features over their respective time series. The comprehensive correlation strength values ​​between all dominant features and all latent features are arranged into a two-dimensional matrix according to the dominant feature index and the latent feature index. The two-dimensional matrix is ​​the feature correlation matrix.

5. The method for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion as described in claim 4, characterized in that, The step of iteratively refining the model structure elements using the feature correlation matrix to generate an optimized model skeleton includes: From the feature association matrix, feature pairs whose comprehensive association strength value exceeds a preset threshold are identified, and the explicit and potential features constituting these feature pairs are marked as key features; Based on the location distribution of the key features in the structured geological feature sequence, the necessity and connection weight of each functional component in the model structural elements are reassessed. An optimization function is established with the goal of improving model structure stability and feature representation ability, using the feature correlation matrix as input constraints. The optimization function is solved by gradient descent, and the number, type, and connection relationships of functional components in the model structure elements are iteratively adjusted. When the output value of the optimization function converges or reaches the preset number of iterations, the iteration stops, and the functional components and their connection relationships determined at this time and adjusted are used as the optimized model skeleton.

6. The method for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion as described in claim 5, characterized in that, The process of inputting the structured geological feature sequence into the optimized model skeleton, and driving the dynamic early warning model of geological disasters to undergo self-organized evolution through a preset calculation process until the model output converges, thereby completing the dynamic construction of the dynamic early warning model of geological disasters, includes: Initialization phase: The structured geological feature sequence is divided into blocks according to time sequence and used as an input stream, which is then injected into the optimized model skeleton. Each functional component in the optimized model skeleton is assigned a random initial internal state. Feedforward propagation phase: At each time step, a data block in the input stream passes sequentially through the functional components of each level of the optimized model skeleton. Each functional component calculates its output based on its current internal state and input data and passes it to the next level until the model output of the time step is generated. Feedback and adjustment phase: The model output is compared with the actual geological disaster observation records or historical early warning tags at the corresponding time step, and the output error is calculated. The output error is propagated back along the optimized model skeleton to update the internal state of each functional component and the connection strength between components. Self-organizing iteration phase: The feedforward propagation phase and feedback adjustment phase are repeatedly executed, and the internal state and connection strength of the functional components of the optimized model skeleton are continuously and dynamically adjusted, so that the error of the model output gradually decreases. Convergence Judgment and Completion Stage: When the error change rate of the model output is lower than a preset threshold or the total error is less than a set threshold for multiple consecutive time steps, it is determined that the model output has converged. At this time, the final internal state and connection relationship of all functional components in the optimized model skeleton are fixed, and the resulting stable computational structure is the completed geological disaster dynamic early warning model.

7. The method for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion as described in claim 6, characterized in that, During the feedback adjustment phase, the output error propagates backward along the optimized model skeleton to update the internal state of each functional component and the connection strength between components, specifically including: The output error is distributed layer by layer from the output layer to the input layer according to the hierarchical relationship of the functional components in the optimized model skeleton, and the responsibility error that each functional component should bear is calculated. For each functional component in the optimized model skeleton, obtain its input value, output value, and current internal state parameters at the time step; The adjustment amount of the internal state parameters of the functional component is calculated based on the responsibility error, the input value, the output value, and the preset learning rate coefficient. The adjustment amount is applied to the internal state parameters of the functional component to complete the update of its internal state; Meanwhile, for the connection path connecting two functional components, the adjustment amount of the connection strength is calculated based on the output of the upstream functional component and the responsibility error of the downstream functional component in the connection path; The adjustment amount of the connection strength is applied to the corresponding connection path to complete the update of the connection strength.

8. The method for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion as described in claim 1, characterized in that, The process of acquiring geological environmental parameters and disaster-inducing parameters, and identifying the geological disaster scenario patterns upon which the early warning model is based, includes: The geological environment parameters are collected in real time from the geological monitoring network or retrieved from historical databases. The geological environment parameters include topographic slope, rock and soil type, and geological structure information. The disaster-inducing parameters are obtained from professional monitoring equipment or disaster case databases, including rainfall intensity sequences, groundwater level changes, and seismic event records. The geological environment parameters are subjected to spatial grid interpolation to form a spatially continuously distributed geological environment field; The disaster-inducing parameters are processed into a time series and matched and fused with the spatially continuously distributed geological environment field in time and space. Based on the fused data, pattern recognition technology was used to delineate typical areas with different combinations of geological conditions and inducing factors. Each typical region is labeled with the types and intensity levels of geological disasters that have occurred or are likely to occur in the past. The definitions of "typical region and its corresponding geological disaster type and intensity level" are combined as the geological disaster scenario pattern on which the early warning model is based.

9. The method for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion as described in claim 6, characterized in that, It also includes the steps of verifying and updating the dynamic early warning model for geological disasters after its construction is completed: Model validation steps: The completed geological disaster dynamic early warning model is tested using an independent geological monitoring dataset that was not involved in the construction process. The independent geological monitoring dataset is input to obtain the model's early warning result sequence. The model's early warning result sequence is compared with the actual disaster occurrence records corresponding to the independent geological monitoring dataset to calculate the early warning accuracy, false alarm rate, missed alarm rate, and early warning lead time index of the geological disaster dynamic early warning model. Model update trigger judgment steps: Set the lower limit threshold of the early warning accuracy, the upper limit threshold of the false alarm rate, and the lower limit threshold of the early warning advance time indicator; The model validation step is performed periodically or when significant changes in geological conditions are detected. If any calculated index exceeds its corresponding threshold, the model update process is triggered. Model update execution steps: When the model update process is triggered, the recently acquired new geological environmental parameters, disaster-induced parameters and their corresponding disaster observation records are used as incremental data; The incremental data is merged with the original constructed data to form a new training dataset; Using the currently constructed dynamic early warning model for geological disasters as the initial state, and using the new training dataset, the dynamic construction process of the dynamic early warning model for geological disasters is re-executed to generate an updated dynamic early warning model for geological disasters.

10. A system for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for constructing a dynamic early warning model for geological disasters based on multi-parameter fusion as described in any one of claims 1 to 9.

Images