A regional geological disaster prediction and early warning method

Abstract

The present invention proposes a regional geological disaster prediction and early warning method, which relates to the field of machine learning, and proposes a TMPNet prediction model, which is applied to geological disaster prediction and early warning scenarios. The model includes a time series feature capture module, a multivariate feature capture module and a multi-layer perceptron stack. The time series feature capture module processes the time features in the input geological disaster related data, the multivariate feature capture module processes the implicit association in the geological disaster multivariate data, and the multi-layer perceptron stack is used to enhance the features of the geological disaster data and provide good feature representation for subsequent convolution operations and attention mechanisms, thereby improving the accuracy of model prediction.

Description

[0001] The present invention belongs to the technical field of machine learning, and in particular relates to a method for predicting and warning regional geological disasters. Background Art

[0002] Geological disasters refer to destructive activities of geological phenomena and geological bodies caused by changes in the geological environment due to natural or human factors, including landslides, mud-rock flows, earthquakes, and collapses. These disasters pose a serious threat to the safety of human life and property and social and economic development. The casualties and economic losses caused by geological disasters around the world are enormous every year. Therefore, it is particularly important to predict and warn of geological disasters in a timely and effective manner.

[0003] Traditional geological disaster prediction and early warning methods mainly rely on on-site surveys, historical data analysis and empirical judgment. Although these methods can provide disaster warnings to a certain extent, they have many limitations. First, the occurrence of geological disasters is often affected by a combination of factors, and traditional methods are difficult to fully consider all relevant factors. Secondly, traditional methods have limited data processing capabilities and cannot effectively respond to the analysis needs of massive data. Thirdly, the prediction accuracy and timeliness of these methods need to be improved, and they cannot meet the needs of rapid response.

[0004] Geological hazard data has highly nonlinear and complex relationships. At the same time, the occurrence of geological disasters is sudden and uncertain, which leads to instability in related data. Due to the complexity and suddenness of geological disasters, long-term prediction errors are easy to accumulate, and the accumulation of errors will significantly affect the prediction results, making prediction difficult. Therefore, the application of machine learning in geological disaster prediction has gradually become a research hotspot. By constructing complex mathematical models, machine learning can automatically learn and extract the laws and patterns of disaster occurrence from large amounts of data, thereby improving the accuracy and efficiency of predictions. The geological disaster prediction and early warning method based on machine learning can not only process multi-source heterogeneous data, but also achieve accurate predictions, providing strong support for early warning and emergency response of geological disasters. Summary of the invention

[0005] The present invention provides a regional geological disaster prediction and early warning method. Aiming at the geological disaster prediction task, a prediction model TMPNet is proposed, which is composed of a time series feature capture module (TFCM), a multivariate feature capture module (MFCM), and a multi-layer perceptron stack.

[0006] The technical method adopted by the present invention to achieve the above-mentioned purpose specifically comprises the following steps:

[0007] S1. Collect historical data related to geological hazards, including characteristics and target variables, and pre-process the collected data;

[0008] S2, adoption Methods The geological disaster data were standardized and divided into training set and test set.

[0009] S3, constructing a TFCM module for processing the temporal characteristics of the input geological hazard data, specifically including the following steps:

[0010] S31. Input geological disaster data , the time series representation of geological disaster data is obtained by enhancing the features through multi-layer perceptron stack ;

[0011] S32, using the convolution kernel The one-dimensional convolutional network combined with the multi-layer perceptron stack outputs to obtain the query ,key Sum , and and Perform a short-time Fourier transform and get , , calculate the relevance score ;

[0012] S33, perform time-delayed aggregation and introduce weight coefficients and weighting coefficient Enhance the time delay aggregation effect and obtain time series feature output Finally, the first prediction value of the TFCM module output is obtained through linear transformation , the second predicted value and the next TFCM submodule input ;

[0013] S4. Construct the MFCM module to process the implicit association in the multivariate data of geological hazards, which specifically includes the following steps:

[0014] S41, input second prediction value , the inter-variable feature representation of geological hazards is obtained by enhancing the features through multi-layer perceptron stack ;

[0015] S42, through the convolution kernel A one-dimensional convolutional network combines the output of a multilayer perceptron stack to compute the query ,key Sum , and use the multi-head self-attention mechanism to calculate the relationship between global variables of geological disaster data ;

[0016] S43. Using dilated convolution to calculate the relationship between local variables of geological hazard data , and add the outputs of the global variable relationship and the relationship between local variables to obtain the output , and after linear transformation, the third prediction value is obtained and the fourth predicted value and the next MFCM submodule input ;

[0017] S5. The TFCM module predicts the value Prediction value with MFCM module Add together to get the geological disaster prediction results .

[0018] Preferably, in step S1, the historical data related to geological disasters include the time of occurrence of geological disasters, the location of geological disasters, the type of geological disasters, the scale of geological disasters, meteorological data, geological data, topographic data, and hydrological data. The method deals with outliers in the data. The specific formula is as follows:

[0019] ;

[0020] In the formula, is an outlier, Features The mean of Features The standard deviation of .

[0021] Preferably, in step S2, using Methods The geological disaster data were standardized and divided into training set and test set in proportion. The specific formula is as follows:

[0022] ;

[0023] In the formula, is the mean value in the data, is the standard deviation, is the data value.

[0024] Preferably, the multivariate geological disaster data defined in steps S3 and S31 is ,in For the Time series values ​​of geological disasters, is the length of the historical geological disaster time series data, and the input is processed through a multi-layer perceptron stack , the enhanced features are used to obtain the time series representation of geological disaster data. The specific formula is as follows:

[0025] ;

[0026] In the formula, For the input geological disaster related data, is a multi-layer perceptron stack, is the number of perceptrons, is the index into the multilayer perceptron stack, A temporal representation of the output of a multilayer perceptron stack.

[0027] Preferably, the temporal characteristics of geological disaster data are enhanced by a multi-layer perceptron stack. Geological disasters often have temporal continuity and suddenness, and the multi-layer perceptron stack, as a deep neural network, can extract high-order features of data layer by layer through multi-layer neuron connections. Specifically, after the input geological disaster data is processed by the multi-layer perceptron stack, the temporal characteristics of capturing the time dependency in the geological disaster data are enhanced, and the dynamic changes of the geological disaster data are more effectively reflected. At the same time, it provides a richer feature representation for subsequent convolution operations and attention mechanisms, and enhances the TFCM module's ability to understand geological disaster time series data.

[0028] Preferably, the convolution kernel used in steps S3 and S32 is The convolutional network combined with the multilayer perceptron stack outputs to obtain the query ,key Sum , the specific formula is as follows:

[0029] ;

[0030] ;

[0031] ;

[0032] In the formula, , , are the learnable parameters of the convolutional layer, is the convolutional layer index, A temporal representation of the output of a multilayer perceptron stack;

[0033] right and Perform short-time Fourier transform, the specific formula is as follows:

[0034] ;

[0035] ;

[0036] In the formula, , For query, key at time point The value of is the window function, is the time window, is the frequency, is an imaginary unit;

[0037] Calculate the correlation score to obtain the time delay. The specific formula is as follows:

[0038] ;

[0039] In the formula, For time delay, is the length of geological disaster time series data, for , Delay in time The relevance score under , for , In time ,frequency The short-time Fourier transform frequency domain representation is shown below.

[0040] Preferably, the enhanced time series features are further processed by combining a one-dimensional convolutional network and a multi-layer perceptron stack. Specifically, the convolution operation of the one-dimensional convolutional network can efficiently extract local features in the geological disaster time series data, while the short-time Fourier transform can capture the periodicity and frequency characteristics in the geological disaster data. By obtaining queries, keys and values, the time series features are converted into a form suitable for attention mechanism processing. This multi-scale feature extraction method improves the model's ability to parse time series data and provides a basis for subsequent time delay aggregation and correlation score calculation.

[0041] Preferably, in steps S3 and S33, weight coefficients are introduced and weighting coefficient The model can more accurately integrate and utilize the characteristic information of different time delays, thereby improving the overall prediction performance, enhancing the time delay aggregation effect, and obtaining the TFCM module output. The specific calculation process is as follows:

[0042] ;

[0043] ;

[0044] ;

[0045] In the formula, For each time delay The weight coefficient of for , Delay in time The corresponding relevance scores are as follows, For a given time delay range, select The time delay with the highest correlation, is the first The time delay characteristic, is the weighting coefficient for each delay feature, For Perform time delay Scroll operation, It is the time series feature output;

[0046] Afterwards The first prediction value, the second prediction value and the next TFCM submodule input are obtained through linear transformation. The specific calculation steps are as follows:

[0047] ;

[0048] ;

[0049] ;

[0050] In the formula, is the block index in the TFCM module, , are the first predicted value and the second predicted value, , is the linear transformation function of the first predicted value and the second predicted value, is the time series feature output, Input for the next TFCM submodule.

[0051] Preferably, a time delay aggregation mechanism is introduced to enhance the comprehensive utilization effect of different time delay features through weight coefficients and weighting coefficients. Specifically, the geological disaster time series features after short-time Fourier transform can effectively integrate the characteristics of geological disaster data at multiple time points through time delay aggregation, and enhance the TFCM module's understanding of dynamic changes in time series. The introduction of weight coefficients and weighting coefficients enables the features of different time delays to be weighted according to their importance during the aggregation process, thereby improving the quality of aggregated features and enabling the TFCM module to more accurately capture the key features in the time series data. This mechanism can effectively reduce the error accumulation in long-term predictions and enhance the stability and reliability of the prediction model.

[0052] Preferably, in the steps S4 and S41, the second predicted value is input , the inter-variable feature representation of geological hazards is obtained by enhancing the features through multi-layer perceptron stack The specific calculation steps are as follows:

[0053] ;

[0054] In the formula, is the second predicted value of the input, is a multi-layer perceptron stack, is the number of perceptrons, is the index into the multilayer perceptron stack, Multivariate inter-variate representation of the output of a multilayer perceptron stack.

[0055] Preferably, the feature enhancement of the input second prediction value is performed through a multi-layer perceptron stack to extract the multivariate feature representation in the geological hazard data. Specifically, the implicit association in the multivariate data is crucial to the prediction of geological hazards, and the use of the multi-layer perceptron stack enables the MFCM module to capture and represent the multivariate features in complex geological hazards, improve the understanding and prediction capabilities of the implicit association in the geological hazard data, better understand and utilize the complex relationships in the multivariate data, and significantly improve the prediction accuracy of the prediction model and its adaptability to multivariate data.

[0056] Preferably, in steps S4 and S42, the convolution kernel is A one-dimensional convolutional network combines the output of a multilayer perceptron stack to compute the query ,key Sum , the specific calculation steps are as follows:

[0057] ;

[0058] ;

[0059] ;

[0060] In the formula, , , is the weight matrix, is the convolutional layer index, Inter-variable representation of the output of a multi-layer perceptron stack;

[0061] The relationship between global variables of geological disaster data is calculated through the multi-head self-attention mechanism. The specific formula is as follows:

[0062] ;

[0063] ;

[0064] In the formula, , , To query j, key and value, is the dimension of the key, For splicing operation, is the linear transformation matrix.

[0065] Preferably, a one-dimensional convolutional network is combined with the output of a multi-layer perceptron stack to calculate queries, keys, and values, and a multi-head self-attention mechanism is used to capture the relationship between global variables in geological hazard data. Specifically, the convolution operation of the one-dimensional convolutional network can extract some local features in multivariate geological hazard data, and the multi-head self-attention mechanism can simultaneously focus on different feature dimensions in geological hazard data, and can extract the complex relationship between global variables in geological hazard data. Through the combination of the above mechanisms, the prediction model's ability to parse multivariate data can be effectively improved, making the prediction results more accurate and reliable.

[0066] Preferably, in steps S4 and S43, dilated convolution is used to calculate the relationship between local variables of geological hazard data, and the specific calculation steps are as follows:

[0067] ;

[0068] In the formula, is the weight matrix of the dilated convolution kernel, A multivariate representation of the output of a multilayer perceptron stack;

[0069] Add the outputs of the relationship between global variables and the relationship between local variables. The specific calculation steps are as follows:

[0070] ;

[0071] In the formula, and is the learnable weight parameter, To extract the global variable relationship, is the extracted local variable relationship;

[0072] Afterwards After linear transformation, the third predicted value and the fourth predicted value as well as the next MFCM submodule input are obtained. The specific calculation steps are as follows:

[0073] ;

[0074] ;

[0075] ;

[0076] In the formula, is the block index in the MFCM module, , are the third and fourth predicted values, , is the linear transformation function of the third prediction value and the fourth prediction value, is the sum of the relationship between global variables and the relationship between local variables, It is the input of the next MFCM submodule.

[0077] Preferably, dilated convolution is used to calculate the relationship between local variables of geological hazard data, and the outputs of the relationship between global and local variables are added, and the predicted value is obtained through linear transformation. Specifically, dilated convolution can expand the receptive field with less computational effort and improve the ability to capture local variable relationships. By combining the relationship between global and local variables, the prediction model can more comprehensively utilize the information in multivariate data. This combination significantly enhances the accuracy and stability of the prediction, enabling the prediction model to better cope with complex geological hazard data.

[0078] Preferably, in step S5, the TFCM module prediction value and the MFCM module prediction value are added to obtain the geological disaster prediction result, and the specific formula is as follows:

[0079] ;

[0080] ;

[0081] ;

[0082] In the formula, , is the total number of blocks in the TFCM module and the MFCM module, , For TFCM module and MFCM module The predicted value of a block, , It is the final prediction result of TFCM module and MFCM module.

[0083] In summary, due to the adoption of the present technical solution, the beneficial effects of the present invention are as follows: The present invention proposes a TMPNet prediction model, which is applied to geological disaster prediction scenarios, including a time series feature capture module, a multivariate feature capture module and a multi-layer perceptron stack. The time series feature capture module processes the time features in the input geological disaster related data, the multivariate feature capture module processes the implicit associations in the geological disaster multivariate data, and the multi-layer perceptron stack is used to enhance the features of the geological disaster data and provide good feature representation for subsequent convolution operations and attention mechanisms, thereby improving the accuracy of model predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 This is a step-by-step diagram for a regional geological disaster prediction and early warning method.

[0085] Figure 2 This is the structure diagram of the TMPNet prediction model.

[0086] Figure 3 This is the TFCM module structure diagram.

[0087] Figure 4 This is the MFCM module structure diagram.

[0088] Figure 5 This is the fitting effect diagram of TMPNet model for geological disaster prediction. DETAILED DESCRIPTION

[0089] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0090] See also Figure 1-Figure 5 , the present invention provides a technical solution.

[0091] The present invention aims to propose a TMPNet prediction model, whose structure is as follows Figure 2 As shown, a regional geological disaster prediction and early warning method is provided, such as Figure 1 shown.

[0092] S1. Collect historical data related to geological disasters, including characteristics and target variables, and preprocess the collected data.

[0093] Furthermore, in step S1, the historical data related to geological disasters include the time of occurrence of geological disasters, the location of geological disasters, the type of geological disasters, the scale of geological disasters, meteorological data, geological data, topographic data, and hydrological data. The method deals with outliers in the data. The specific formula is as follows:

[0094] ;

[0095] In the formula, is an outlier, Features The mean of Features The standard deviation of .

[0096] S2, adoption Methods The geological hazard data were standardized and divided into training set and test set.

[0097] Furthermore, in step S2, using Methods The geological disaster data were standardized and divided into training set and test set in a ratio of 8:2. The specific formula is as follows:

[0098] ;

[0099] In the formula, is the mean value in the data, is the standard deviation, is the data value.

[0100] S3, constructing a TFCM module for processing the temporal characteristics of the input geological hazard data, specifically including the following steps:

[0101] S31. Input geological disaster data , the time series representation of geological disaster data is obtained by enhancing the features through multi-layer perceptron stack .

[0102] Furthermore, in step S3, the TFCM module structure is as follows: Figure 3 As shown in Figure 2, the multivariate geological disaster data is defined in S31 as ,in For the Time series values ​​of geological disasters, is the length of the historical geological disaster time series data, and the input is processed through a multi-layer perceptron stack , the enhanced features are used to obtain the time series representation of geological disaster data. The specific formula is as follows:

[0103] ;

[0104] In the formula, For the input geological disaster related data, is a multi-layer perceptron stack, is the number of perceptrons, is the index into the multilayer perceptron stack, A temporal representation of the output of a multilayer perceptron stack.

[0105] S32, using the convolution kernel The one-dimensional convolutional network combined with the multi-layer perceptron stack outputs to obtain the query ,key Sum , and and Perform a short-time Fourier transform and get , , calculate the relevance score .

[0106] Furthermore, the convolution kernel used in steps S3 and S32 is The convolutional network combined with the multilayer perceptron stack outputs to obtain the query ,key Sum , the specific formula is as follows:

[0107] ;

[0108] ;

[0109] ;

[0110] In the formula, , , are the learnable parameters of the convolutional layer, is the convolutional layer index, A temporal representation of the output of a multilayer perceptron stack;

[0111] right and Perform short-time Fourier transform, the specific formula is as follows:

[0112] ;

[0113] ;

[0114] In the formula, , For query, key at time point The value of is the window function, is the time window, is the frequency, is an imaginary unit;

[0115] Calculate the correlation score to obtain the time delay. The specific formula is as follows:

[0116] ;

[0117] In the formula, For time delay, is the length of geological disaster time series data, for , Delay in time The relevance score under , for , In time ,frequency The short-time Fourier transform frequency domain representation is shown below.

[0118] S33, perform time-delayed aggregation and introduce weight coefficients and weighting coefficient Enhance the time delay aggregation effect and obtain time series feature output Finally, the first prediction value of the TFCM module output is obtained through linear transformation , the second predicted value and the next TFCM submodule input .

[0119] Furthermore, in the steps S3 and S33, the weight coefficient and weighting coefficient The model can more accurately integrate and utilize the characteristic information of different time delays, thereby improving the overall prediction performance, enhancing the time delay aggregation effect, and obtaining the TFCM module output. The specific calculation process is as follows:

[0120] ;

[0121] ;

[0122] ;

[0123] In the formula, For each time delay The weight coefficient of for , Delay in time The corresponding relevance scores are as follows, For a given time delay range, select The time delay with the highest correlation, is the first The time delay characteristic, is the weighting coefficient for each delay feature, For Perform time delay Scroll operation, It is the time series feature output;

[0124] Afterwards The first prediction value, the second prediction value and the next TFCM submodule input are obtained through linear transformation. The specific calculation steps are as follows:

[0125] ;

[0126] ;

[0127] ;

[0128] In the formula, is the block index in the TFCM module, , are the first predicted value and the second predicted value, , is the linear transformation function of the first predicted value and the second predicted value, is the time series feature output, Input for the next TFCM submodule.

[0129] S4. Construct the MFCM module to process the implicit association in the multivariate data of geological hazards, which specifically includes the following steps:

[0130] S41, input second prediction value , the inter-variable feature representation of geological hazards is obtained by enhancing the features through multi-layer perceptron stack .

[0131] Furthermore, in step S4, the MFCM module structure is as follows: Figure 4 As shown, the second prediction value is input in S41 , the inter-variable feature representation of geological hazards is obtained by enhancing the features through multi-layer perceptron stack The specific calculation steps are as follows:

[0132] ;

[0133] In the formula, is the second predicted value of the input, is a multi-layer perceptron stack, is the number of perceptrons, is the index into the multilayer perceptron stack, Multivariate inter-variate representation of the output of a multilayer perceptron stack.

[0134] S42, through the convolution kernel A one-dimensional convolutional network combines the output of a multilayer perceptron stack to compute the query ,key Sum , and use the multi-head self-attention mechanism to calculate the relationship between global variables of geological disaster data .

[0135] Furthermore, in the steps S4 and S42, the convolution kernel is A one-dimensional convolutional network combines the output of a multilayer perceptron stack to compute the query ,key Sum , the specific calculation steps are as follows:

[0136] ;

[0137] ;

[0138] ;

[0139] In the formula, , , is the weight matrix, is the convolutional layer index, Inter-variable representation of the output of a multi-layer perceptron stack;

[0140] The relationship between global variables of geological disaster data is calculated through a multi-head self-attention mechanism. The specific formula is as follows:

[0141] ;

[0142] ;

[0143] In the formula, , , To query j, key and value, is the dimension of the key, For splicing operation, is the linear transformation matrix.

[0144] S43. Using dilated convolution to calculate the relationship between local variables of geological hazard data , and add the outputs of the global variable relationship and the relationship between local variables to obtain the output , and after linear transformation, the third prediction value is obtained and the fourth predicted value and the next MFCM submodule input .

[0145] Furthermore, in the steps S4 and S43, the relationship between local variables of the geological disaster data is calculated using dilated convolution, and the specific calculation steps are as follows:

[0146] ;

[0147] In the formula, is the weight matrix of the dilated convolution kernel, A multivariate representation of the output of a multilayer perceptron stack;

[0148] Add the outputs of the relationship between global variables and the relationship between local variables. The specific calculation steps are as follows:

[0149] ;

[0150] In the formula, and is the learnable weight parameter, To extract the global variable relationship, is the extracted local variable relationship;

[0151] Afterwards After linear transformation, the third predicted value and the fourth predicted value as well as the next MFCM submodule input are obtained. The specific calculation steps are as follows:

[0152] ;

[0153] ;

[0154] ;

[0155] In the formula, is the block index in the MFCM module, , are the third and fourth predicted values, , is the linear transformation function of the third prediction value and the fourth prediction value, is the sum of the relationship between global variables and the relationship between local variables, It is the input of the next MFCM submodule.

[0156] S5. The TFCM module predicts the value Prediction value with MFCM module Add together to get the geological disaster prediction results .

[0157] Furthermore, in step S5, the TFCM module prediction value and the MFCM module prediction value are added to obtain the geological disaster prediction result, and the specific formula is as follows:

[0158] ;

[0159] ;

[0160] ;

[0161] In the formula, , is the total number of blocks in the TFCM module and the MFCM module, , For TFCM module and MFCM module The predicted value of a block, , It is the final prediction result of TFCM module and MFCM module.

[0162] Furthermore, the TMPNet prediction model is written in Python, using the Pytorch framework under CUDA11.1, and tested on an NVIDIA RTX 4090 24GB graphics card server. 1825 days of geological disaster-related data are collected and preprocessed before input into the TMPNet prediction model. The input length is set to 96, and the TMPNet prediction model structure is set to \left [ {TFCM, MFCM, MFCM} \right ] , that is, a temporal feature capture module followed by two multivariate feature capture modules, setting the convolution kernel size to 3, the number of modules to 3, the hidden dimension to 256, and the initial learning rate to , batch size is 32, and Dropout is , the L2 loss is , set the learnable parameters and The size is 0.5, the weight coefficient and weighting coefficient It is a trainable parameter and its size is automatically learned through the training process.

[0163] Furthermore, the TMPNet prediction model fitting effect is as follows Figure 5 As shown, the ordinate is the probability of geological disasters (%), the abscissa is time (days), the black line represents the real geological disaster data, and the gray dots represent the predicted geological disaster data. It can be seen from the figure that the predicted data of the TMPNet model have a high degree of fit with the real data, indicating that the TMPNet model shows high accuracy in predicting the probability of geological disasters, and the fluctuation trend of the predicted value is basically consistent with the real value, especially in some areas with a high probability of geological disasters. The model can accurately reflect these changes. Although the predicted data has a certain degree of discreteness, the overall trend still closely follows the real data, indicating that the model has good stability and robustness. Even in the face of noise and outliers in the data, it can still maintain a relatively stable prediction performance. The TMPNet prediction model can well capture the trend and changes of geological disasters and has a high prediction accuracy, indicating that the TMPNet prediction model proposed in this invention performs well in identifying and learning complex geological disaster-related data.

Claims

1. A regional geological disaster prediction and early warning method, characterized in that: The following steps are involved: S1. Collect historical data related to geological hazards, including characteristics and target variables, and preprocess the collected data; S2. Use the Z-Score method to standardize the geological disaster data and divide the standardized data into training set and test set; S3, constructing a TFCM module for processing the temporal characteristics of the input geological hazard data, specifically including the following steps: S31. Input geological disaster data X, and obtain the time series representation of geological disaster data by multi-layer perceptron stack enhancement features. Use a convolutional network with a convolution kernel of 1×1 combined with a multilayer perceptron stack output to obtain the query Q TFC , key K TFC Sum value V TFC , the specific steps are as follows: Define multivariate geological hazard data as Where X N is the time series value of the Nth geological disaster, T x is the length of the historical geological disaster time series data. The input X is processed by a multi-layer perceptron stack to enhance the features and obtain the time series representation of the geological disaster data. The specific formula is as follows: Where X is the input geological disaster related data, is a multilayer perceptron stack, i is the number of perceptrons, p is the index of the multilayer perceptron stack, is the temporal representation of the multilayer perceptron stack output. A convolutional network with a convolution kernel of 1×1 is combined with the multilayer perceptron stack output to obtain the query Q TFC , key K TFC Sum value V TFC , the specific formula is as follows: In the formula, is the learnable parameter of the convolutional layer, k is the convolutional layer index, A temporal representation of the output of a multilayer perceptron stack; S32, Q TFC and K TFC Perform a short-time Fourier transform and get Calculating relevance scores Perform time-delayed aggregation and introduce weight coefficient w τ and weighting coefficient α i Enhance the time delay aggregation effect and obtain the timing feature output TFC out Finally, the first prediction value of the TFCM module output is obtained through linear transformation Second prediction value and the next TFCM submodule input The specific steps are as follows: Q TFC and K TFC Perform short-time Fourier transform, the specific formula is as follows: In the formula, Q TFC (t), K TFC (t) is the value of the query,key at time point t, w(t) is the window function, m is the time window, ω is the frequency, j is the imaginary unit, and the correlation score is calculated to obtain the time delay. The specific formula is as follows: Where τ is the time delay, T is the length of the geological disaster time series data, Q TFC , K TFC The correlation score at time delay τ, Q TFC , K TFC The short-time Fourier transform frequency domain representation at time t and frequency ω, finally, by introducing the weight coefficient w τ and weighting coefficient α i The model can more accurately integrate and utilize the characteristic information of different time delays, thereby improving the overall prediction performance, enhancing the time delay aggregation effect, and obtaining the TFCM module output. The specific calculation process is as follows: In the formula, w τ is the weight coefficient for each time delay τ, Q TFC , K TFC The corresponding relevance score at time delay τ, argTop k Select Top for a given time delay range k The time delay with the highest correlation, is the kth time delay feature after ReLU weighting, α i is the weighting coefficient of each delay feature, Roll(V TFC , τ i ) is for V TFC Perform time delay τ rolling operation, TFC out It is the timing feature output, and then TFC out The first prediction value, the second prediction value and the next TFCM submodule input are obtained through linear transformation. The specific calculation steps are as follows: Where Ptfc is the block index in the TFCM module, are the first predicted value and the second predicted value, is the linear transformation function of the first prediction value and the second prediction value, TFC out is the time series feature output, Input for the next TFCM submodule; S4. Construct the MFCM module to process the implicit association in the multivariate data of geological hazards, which specifically includes the following steps: S41, input second prediction value Enhanced features through multi-layer perceptron stacking to obtain inter-variable feature representation of geological hazards The query Q is calculated by combining the output of the multilayer perceptron stack with a 1×1 convolutional network. MFC , key K MFC Sum value V MFC ; S42. Use multi-head self-attention mechanism to calculate the relationship between global variables of geological disaster data MultiHead(Q MFC , K MFC , V MFC ), calculate the local variable relationship SConv of geological disaster data through dilated convolution, and add the output of global variable relationship and local variable relationship to obtain output S out , after linear transformation, the third prediction value is obtained and the fourth predicted value and the next MFCM submodule input S5. Use the TFCM module to predict Y TFC Predict Y with the MFCM module MFC Add them together to get the geological disaster prediction Y.

2. A regional geological disaster prediction and early warning method according to claim 1, characterized in that: In the step S4, the MFCM module is constructed. In the specific step S41, the second prediction value is input. Enhanced features through multi-layer perceptron stacking to obtain inter-variable feature representation of geological hazards The specific calculation steps are as follows: In the formula, is the second predicted value of the input, is a multilayer perceptron stack, i is the number of perceptrons, p is the index of the multilayer perceptron stack, The query Q is calculated by combining the output of the multi-layer perceptron stack with a one-dimensional convolutional network with a convolution kernel of 1×1. MFC , key K MFC Sum value V MFC , the specific calculation steps are as follows: In the formula, is the weight matrix, k is the convolutional layer index, Inter-variable representation of the output of a multi-layer perceptron stack.

3. A regional geological disaster prediction and early warning method according to claim 1, characterized in that: In the step S4, the MFCM module is constructed. In the specific step S42, the relationship between the global variables of the geological disaster data is calculated through the multi-head self-attention mechanism. The specific formula is as follows: MultiHead(Q MFC ,K MFC ,V MFC )=Concat(SAtt1,SAtt2,…,SAtt n )W o ; In the formula, Q MFC , K MFC 、V MFC To query j, key and value, d h is the dimension of the key, Concat is the concatenation operation, W o is a linear transformation matrix, and the relationship between local variables of geological hazard data is calculated using dilated convolution. The specific calculation steps are as follows: In the formula, is the weight matrix of the dilated convolution kernel, For the multivariate representation of the output of the multilayer perceptron stack, add the outputs of the global variable relationship and the local variable relationship. The specific calculation steps are as follows: S out =α·MultiHead(Q MFC ,K MFC ,V MFC )+β·SConv; Where α and β are learnable weight parameters, MultiHead(Q MFC , K MFC , V MFC ) is the extracted global variable relationship, SConv is the extracted local variable relationship, and then S out After linear transformation, the third predicted value and the fourth predicted value as well as the next MFCM submodule input are obtained. The specific calculation steps are as follows: Where Pmfc is the block index in the MFCM module, are the third and fourth predicted values, is the linear transformation function of the third prediction value and the fourth prediction value, S out is the sum of the relationship between global variables and the relationship between local variables, It is the input of the next MFCM submodule.

4. A regional geological disaster prediction and early warning method according to claim 1, characterized in that: In step S5, the predicted value of the TFCM module and the predicted value of the MFCM module are added to obtain the geological disaster prediction result, and the specific formula is as follows: Y=Y TFC +Y MFC ; Where ptfc and pmfc are the total number of blocks in the TFCM module and the MFCM module, is the predicted value of the TFCM module and the MFCM module for the i-th block, Y TFC , Y MFC It is the final prediction result of TFCM module and MFCM module.

Images