A method for predicting landslide disasters in coal mining areas and a multi-layered adversarial generation model

Abstract

The present invention provides a method for predicting landslide disasters in coal mining areas and a multi-layer adversarial generation model, belonging to the technical field of landslide disaster prediction. The method solves the problem that existing landslide disaster monitoring and evaluation methods in coal mining areas are time-consuming and labor-intensive. The method comprises the following steps: mapping terrain data and remote sensing image data acquired by an unmanned aerial vehicle (UAV) into a constructed frequency domain space; constructing a hierarchical generator using an LSTM model, embedding a time series prediction model and a Bessel-based fitting model in each layer of the generator before inputting data, and inputting the input data into an N-layer time series prediction model in layers; embedding a convolutional neural network in each layer of the discriminator to judge the accuracy of the generated sequence data in the frequency domain space; and assigning weights to and reordering landslide images at each layer based on the output data of each layer of the discriminator and a random forest regression model. The method is applicable to landslide disaster prediction in coal mining areas.

Description

Technical Field

[0001] The invention provides a method for predicting landslide disasters in coal mining areas and a multi-layered confrontation generation model, belonging to the technical field of landslide disaster prediction. Background Art

[0002] Coal resources are primarily distributed in mountainous areas with complex geological conditions and poor accessibility. Therefore, large-scale coal mining has resulted in numerous and extensive goaf areas, which, over time, can lead to geological disasters of varying sizes, scales, and types. In recent years, Shanxi Province has experienced frequent heavy and torrential rains, resulting in landslides of varying sizes and causing significant property losses to numerous coal mines. Traditional disaster monitoring and assessment methods are time-consuming, labor-intensive, and inaccurate, hindering the effectiveness of disaster management. Remote sensing, with its effective temporal and spatial resolution, provides powerful technical support for monitoring disaster changes in remote areas with limited accessibility. However, compared to the rapidly increasing rate of high-resolution satellite launches, the application of these images has generally stagnated due to time-consuming, labor-intensive, and costly manual interpretation. Therefore, automated and intelligent recognition of remote sensing images has become an urgent challenge.

[0003] With the rapid development of artificial intelligence technology, especially the widespread application of deep learning algorithms in the field of artificial intelligence, experts and scholars in various fields have seen the potential for industrial intelligence in various fields. Therefore, this paper is dedicated to the development of an intelligent landslide disaster prediction model that can effectively shorten emergency response time after a landslide disaster occurs and formulate reasonable plans in a timely manner. At the same time, there is currently little research on the use of deep learning algorithms to predict landslide disasters in mining areas. Therefore, a reasonable method is urgently needed to predict the time of landslide disaster occurrence and analyze the influencing factors. Summary of the Invention

[0004] In order to solve the problem that the existing landslide disaster monitoring and evaluation methods in coal mining areas are time-consuming and labor-intensive, the present invention proposes a method for predicting landslide disasters in coal mining areas and a multi-layered confrontation generation model.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for predicting landslide disasters in coal mining areas, comprising the following steps:

[0006] S1: Frequency domain hierarchical mapping of multimodal data: Map terrain data and remote sensing image data acquired by UAVs into the constructed frequency domain space, and couple the above data into the set data structure in layers according to the time scale;

[0007] S2: Constructing a hierarchical generative model: Using the LSTM model to construct a hierarchical generator, before inputting data, a time series prediction model and a Bessel-based fitting model are embedded in each layer of the generator, and the input data is hierarchically input into the N-layer time series prediction model;

[0008] S3: Build a hierarchical discriminant model: embed a convolutional neural network in each layer of the discriminator to determine the accuracy of the generated sequence data in the frequency domain;

[0009] S4: Construct a coupled-driven prediction model: Based on the output data of each layer of discriminators and the random forest regression model, the images of landslides at each level are weighted and re-ranked, and each layer is comprehensively ranked according to the strategy score.

[0010] The remote sensing image in step S1 is obtained by a drone, divided into data of a size of 512*512*3, and the remote two-dimensional sensing image is expanded in the spectral dimension.

[0011] The terrain data in step S1 includes DEM data of topography, geological structure, groundwater, precipitation, and historical records of landslides. The DEM data is converted into a grid type, and the frequency domain space is constructed based on the grid value.

[0012] Specifically, step S1 maps the different dimensional data of two-dimensional remote sensing data and three-dimensional terrain data to a one-dimensional frequency domain space using a layered restoration model, and uses the time when landslide disasters occurred in previous years as label values.

[0013] The Bessel-based fitting model in step S2 is used to increase the time scale of the data in the frequency domain space to days, and the hierarchical generation model inputs the frequency domain data with the increased time scale into the time series prediction model in layers to generate hierarchical sequence data.

[0014] The coupled-driven prediction model processes the data generated by each layer through a discriminator learning process, and arranges the influence of each layer on the occurrence of landslides in descending order based on random forest regression.

[0015] A multi-layered adversarial generative model for predicting landslide hazards in coal mining areas, including:

[0016] Hierarchical generator: The generator embeds an N-layer time series prediction model and generates time series prediction data hierarchically based on real-world multimodal data;

[0017] Hierarchical discriminator: The discriminator is embedded with a classification discrimination module, which judges the truth or falsehood of each layer of time series prediction model based on the input hierarchical time series prediction data and outputs the prediction data of each layer;

[0018] Coupling-driven prediction module: couples the layered prediction data together and comprehensively ranks each layer based on the strategy score.

[0019] The multimodal data includes remote sensing images, topography, geological structure, groundwater, and precipitation data; N in the N-layer time series prediction model represents the category of the multimodal data.

[0020] The multimodal data are uniformly mapped to a one-dimensional frequency domain space through a hierarchical restoration model before being input into the hierarchical generator.

[0021] The layer generator also embeds a Bessel-based fitting model to increase the time scale of the data to days.

[0022] The beneficial effects of the present invention compared to the prior art are:

[0023] 1. The hierarchical generator constructed by the patent of this invention can directly generate data for different types of multimodal data, breaking the data barriers that make it difficult to comprehensively analyze different data in prediction problems. This is mainly because the generator has become a parallel mode for multiple time series models.

[0024] 2. The discriminator constructed by the patent of this invention transforms the time series prediction problem into a sequence similarity comparison problem, maps the complex multivariate prediction problem to the spectrum space, and simplifies the complexity of the prediction model, mainly due to the processing of the multimodal hierarchical spectrum mapping of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The present invention will be further described below with reference to the accompanying drawings:

[0026] Figure 1 Schematic diagram of the structure of the hierarchical adversarial generation model of the present invention. DETAILED DESCRIPTION

[0027] like Figure 1 As shown, the present invention provides a method for predicting landslide disasters in coal mining areas, comprising the following steps:

[0028] S1: Frequency domain hierarchical mapping of multimodal data: Map terrain data and remote sensing image data acquired by UAVs into the constructed frequency domain space, and couple the above data into the set data structure in layers according to the time scale;

[0029] S2: Constructing a hierarchical generative model: Using the LSTM model to construct a hierarchical generator, before inputting data, a time series prediction model and a Bessel-based fitting model are embedded in each layer of the generator, and the input data is hierarchically input into the N-layer time series prediction model;

[0030] S3: Build a hierarchical discriminant model: embed a convolutional neural network in each layer of the discriminator to determine the accuracy of the generated sequence data in the frequency domain;

[0031] S4: Construct a coupled-driven prediction model: Based on the output data of each layer of discriminators and the random forest regression model, the images of landslides at each level are weighted and re-ranked, and each layer is comprehensively ranked according to the strategy score.

[0032] The remote sensing image in step S1 is obtained by a drone, divided into data of a size of 512*512*3, and the two-dimensional remote sensing image is expanded in the spectral dimension.

[0033] The terrain data in step S1 includes DEM data of topography, geological structure, groundwater, precipitation, and historical records of landslides. The DEM data is converted into a grid type, and the frequency domain space is constructed based on the grid value.

[0034] Specifically, step S1 maps the different dimensional data of two-dimensional remote sensing data and three-dimensional terrain data to a one-dimensional frequency domain space using a layered restoration model, and uses the time when landslide disasters occurred in previous years as label values.

[0035] The Bessel-based fitting model in step S2 is used to increase the time scale of the data in the frequency domain space to days, and the hierarchical generation model inputs the frequency domain data with the increased time scale into the time series prediction model in layers to generate hierarchical sequence data.

[0036] The coupled-driven prediction model processes the data generated by each layer through a discriminator learning process, and arranges the influence of each layer on the occurrence of landslides in descending order based on random forest regression.

[0037] A multi-layered adversarial generative model for predicting landslide hazards in coal mining areas, including:

[0038] Hierarchical generator: The generator embeds an N-layer time series prediction model and generates time series prediction data hierarchically based on real-world multimodal data;

[0039] Hierarchical discriminator: The discriminator is embedded with a classification discrimination module, which judges the truth or falsehood of each layer of time series prediction model based on the input hierarchical time series prediction data and outputs the prediction data of each layer;

[0040] Coupling-driven prediction module: couples the layered prediction data together and comprehensively ranks each layer based on the strategy score.

[0041] The multimodal data includes remote sensing images, topography, geological structure, groundwater, and precipitation data; N in the N-layer time series prediction model represents the category of the multimodal data.

[0042] The multimodal data are uniformly mapped to a one-dimensional frequency domain space through a hierarchical restoration model before being input into the hierarchical generator.

[0043] The layer generator also embeds a Bessel-based fitting model to increase the time scale of the data to days.

[0044] The method for predicting landslide hazards in coal mining areas, provided by the present invention, primarily involves layered mapping of multimodal raw data to the frequency domain and performing matching coupling. Before inputting into the layered generator, the multimodal real input undergoes layered frequency domain restoration and is then fed into the time series prediction model of each layer of the generator for data generation. Prior to performing discrimination, multimodal data from different layers is coupled to facilitate weight assignment for the layered discriminator.

[0045] The multimodal data input by the present invention includes data of different dimensions such as two-dimensional remote sensing data and three-dimensional terrain data. The data are uniformly mapped to a one-dimensional frequency domain space using a layered restoration model, and the time when landslide disasters occurred in previous years is used as the label value.

[0046] The hierarchical generator of the present invention embeds an N-layer time series prediction model (N is the category of multimodal data) and generates time series prediction data in a hierarchical manner based on real-world multimodal data.

[0047] The classification and discrimination module is embedded in the discriminant model. According to the input generated data, the discriminator of each layer judges the truth or falsehood of the time series prediction model of each layer to obtain the optimal prediction data.

[0048] Before obtaining the output, the layered prediction data are coupled together through a coupling driving model, and each layer is comprehensively ranked according to the strategy score.

[0049] like Figure 1 As shown, the main steps of constructing the hierarchical adversarial generative model of the present invention are as follows:

[0050] Step 1: Frequency domain layered mapping of multimodal data: Remote sensing images are acquired by drones and divided into data of size 512*512*3, and the two-dimensional remote sensing images are expanded in the spectral dimension; the terrain data is mainly represented by DEM data, including precipitation data, geological structure data, and historical record data of landslides, and the above DEM data is converted into raster type; at the same time, based on the raster value, a frequency domain space is constructed, and the above data are layered and coupled into the HDF5 data file according to the time scale to obtain input data in a unified data format, where each layer of data in the HDF5 data file is arranged according to the time series.

[0051] Step 2: Improve the time scale of time series data: Perform layered Bessel fitting on the H5 data constructed in step 1 and increase the time scale unit to days.

[0052] Step 3: Constructing a hierarchical generative model: Five time series prediction models are embedded. This paper primarily uses the LSTM model to construct the generator. Before inputting data, a hierarchical reduction module and a Bessel-based fitting model are embedded, reducing the time scale to days and inputting the data into the time series prediction model in layers.

[0053] Step 4: Construct a hierarchical discriminant model: Embed a convolutional neural network in each layer of the discriminator to determine the accuracy of the generated sequence in the frequency domain. The hierarchical discriminant model of the present invention is based on a convolutional neural network and performs a discriminant score on the sequence data generated in step 3, transforming the prediction problem into a classification problem. If the result is 1, the data is output; if the result is 0, the data is discarded.

[0054] Step 5: Build a coupled-driven prediction model: Based on the output data from each discriminator layer, a random forest regression model is used to weight and reorder the landslide images at each layer. After each layer of data is processed by the discriminator and reaches optimal accuracy, the random forest regression model ranks each layer in descending order of its impact on landslide occurrence.

[0055] The main hardware and software requirements of the present invention are as follows:

[0056] Hardware environment: The processor is Intel i7-8700, the main frequency is 3.2GHz, the memory is 32G, and the graphics card is NVIDIA Quadro P4000.

[0057] The main software environment of the present invention is: Windows 10, Anaconda 3 development environment, and Notepad++ editor.

[0058] This paper proposes a multi-layered adversarial generative model for predicting landslide disasters in coal mining areas, which can empower coal mine disaster warning and management. The key technologies are as follows:

[0059] 1. Fusion of multimodal data;

[0060] 2. Extract time series features of multimodal data in the frequency domain;

[0061] 3. Initialize network input based on multimodal time series features;

[0062] 4. Generate time series data in the generator module of the network;

[0063] 5. The discriminator module of the network determines whether the generated time series data is true or false;

[0064] 6. For multimodal input, output the probability of landslide disasters occurring in the future and the corresponding weight ratio.

[0065] In view of the above-mentioned key technologies, the present invention mainly solves the following key problems:

[0066] 1. The multimodal data of remote sensing images, precipitation, wind speed, groundwater content, topography, landslide disaster records, etc. are uniformly mapped into the frequency domain space, solving the dimensionality, dimension, and representation differences of multimodal data;

[0067] 2. We introduce the technical concept of adversarial generation into landslide disaster prediction in mining areas. We use a layered initialization method to solve the problem that traditional methods cannot input real-world multimodal data into the generator. At the same time, we construct a time series prediction model as a generation module based on the weights initialized from multimodal data.

[0068] 3. In the discriminator, since the generator has mapped the multimodal input data to the frequency domain, the layered convolutional neural network is used to distinguish the authenticity of the one-dimensional generated data, transforming the prediction problem into a classification problem, solving the problem of multimodal input being difficult to predict;

[0069] 4. According to the output of different modal data, the weight ratio of the factors affecting disaster prediction is given, which simplifies the analysis steps of driving factors in the prediction problem.

[0070] The present invention can provide an intelligent model for landslide disaster prediction in mining areas, and disaster emergency management departments can formulate focused disaster prevention plans based on multi-modal weight outputs.

[0071] Regarding the specific structure of the present invention, it should be noted that the connection relationship between the various component modules adopted in the present invention is definite and feasible. Except for those specifically described in the embodiments, the specific connection relationship can bring about corresponding technical effects and solve the technical problems raised by the present invention without relying on the execution of corresponding software programs. The components, modules, models of specific components appearing in the present invention, the connection methods between each other, and the conventional usage methods and expected technical effects brought about by the above-mentioned technical features, except for those specifically described, all belong to the disclosed contents in patents, journal articles, technical manuals, technical dictionaries, and textbooks that can be obtained by technical personnel in this field before the application date, or belong to the existing technologies such as conventional technology and common knowledge in this field. There is no need to elaborate, so that the technical solution provided in this case is clear, complete, and feasible, and the corresponding physical products can be reproduced or obtained based on this technical means.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting landslide disasters in coal mining areas, characterized by: The steps include: S1: Frequency domain hierarchical mapping of multimodal data: 2D remote sensing image data and 3D terrain data are uniformly mapped to 1D frequency domain space using a hierarchical restoration model. The above data are then hierarchically coupled into HDF5 data files according to the time scale, and the time when landslide disasters occurred in previous years is used as the label value. S2: Constructing a hierarchical generative model: A hierarchical generator is constructed using an LSTM model. Before inputting data, a time series prediction model and a Bessel-based fitting model are embedded in each layer of the generator. The Bessel-based fitting model is used to increase the time scale of the data in the one-dimensional frequency domain space to days. The hierarchical generative model inputs the frequency domain data with increased time scale into an N-layer time series prediction model in layers to generate hierarchical sequence data. S3: Build a hierarchical discriminant model: embed a convolutional neural network in each layer of the discriminator to determine the accuracy of the generated sequence data in the frequency domain; S4: Construct a coupled-driven prediction model: Based on the output data of each layer of discriminators and the random forest regression model, the images of landslides at each level are weighted and re-sorted, and the degree of influence of each layer on the occurrence of landslides is arranged in descending order according to the strategy score.

2. The method for predicting landslide disasters in coal mining areas according to claim 1, characterized in that: The remote sensing image in step S1 is obtained by a drone, divided into data of a size of 512*512*3, and the two-dimensional remote sensing image is expanded in the spectral dimension.

3. The method for predicting landslide disasters in coal mining areas according to claim 1, characterized in that: The terrain data in step S1 includes DEM data of topography, geological structure, groundwater, precipitation, and historical records of landslides. The DEM data is converted into a grid type, and the frequency domain space is constructed based on the grid value.

Images