Mine ecological risk prediction method and system based on artificial intelligence

Abstract

The invention provides a mine ecological risk prediction method and system based on artificial intelligence, and the method comprises the steps: collecting multi-source ecological monitoring data through a distributed sensor network disposed in a mine region, carrying out the time-space alignment processing to generate a time-space correlation feature set, calling a pre-trained ecological risk prediction model to predict the features, and carrying out the prediction of the features. Generating a risk conduction mode set including a risk propagation path, an initial node, a propagation direction and strength, constructing an ecological risk evolution network based on the risk conduction mode set, and displaying a topological structure and a time sequence dependency relationship of risk nodes, and according to the ecological risk evolution network, generating a risk early warning instruction set containing a risk grade division result and an ecological restoration strategy, and sending the risk early warning instruction set to an ecological management platform, thereby comprehensively and accurately predicting the ecological risk of the mine, providing an effective risk early warning and ecological restoration strategy, and ensuring the ecological safety of the mine.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and in particular, to a method and system for predicting the ecological risk of mines based on artificial intelligence. Background Art

[0002] During the process of mine exploitation, due to the disturbance of mining activities to the environment, the mine area often faces ecological risks, such as soil erosion, water pollution, vegetation damage, etc. These ecological risks not only affect the ecological environment quality of the mine area, but may also cause chain reactions to the surrounding areas, threatening ecological security and human health. At present, the prediction of mine ecological risks mainly relies on manual monitoring and empirical judgment. This method is not only inefficient, but also difficult to accurately capture the dynamic changes of ecological risks in time and space. In addition, the existing prediction methods often lack in-depth analysis of the risk conduction mode, and are unable to effectively predict the propagation path and intensity of risks, resulting in the lack of pertinence and timeliness in the implementation of ecological restoration measures. Therefore, developing a method that can comprehensively and accurately predict the ecological risks of mines and provide effective risk warnings and ecological restoration strategies has become an urgent technical problem to be solved in the current field of mine environmental protection. Summary of the Invention

[0003] In view of the problems mentioned above, in combination with the first aspect of the present invention, embodiments of the present invention provide a method for predicting the ecological risk of mines based on artificial intelligence. The method includes: Collect multi-source ecological monitoring data through a distributed sensor network deployed in the mine area. The multi-source ecological monitoring data includes a time-continuous sequence of environmental indicators and spatially associated geographical coordinate information; Perform spatio-temporal alignment processing on the multi-source ecological monitoring data to generate a spatio-temporal correlation feature set. The spatio-temporal correlation feature set includes trend change features in the time dimension and regional correlation features in the space dimension; Call a pre-trained ecological risk prediction model to perform ecological risk prediction on the spatio-temporal correlation feature set, and generate a risk conduction mode set. The risk conduction mode set includes the starting node, propagation direction, and propagation intensity of the risk propagation path; Construct an ecological risk evolution network for the mine area based on the risk conduction mode set. The ecological risk evolution network includes the topological structure of risk nodes and the temporal dependence relationship of risk diffusion; Generate a risk warning instruction set according to the ecological risk evolution network. The risk warning instruction set includes the risk level classification result and the corresponding ecological restoration strategy, and send the risk warning instruction set to the ecological management platform.

[0004] In another aspect, an embodiment of the present invention further provides an artificial intelligence-based mine ecological risk prediction system, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0005] Based on the above aspects, the embodiment of the present invention collects multi-source ecological monitoring data through a distributed sensor network deployed in the mine area, performs spatio-temporal alignment processing, generates a spatio-temporal correlation feature set including time-dimensional trend change features and space-dimensional regional correlation features, calls a pre-trained ecological risk prediction model to deeply analyze these features, generates a risk conduction mode set, reveals the starting node, propagation direction and propagation intensity of the risk propagation path. Based on these risk conduction modes, this method constructs an ecological risk evolution network in the mine area, clearly shows the topological structure of risk nodes and the time-dependent relationship of risk diffusion. According to this ecological risk evolution network, a risk warning instruction set including risk level division results and corresponding ecological restoration strategies can be generated and sent to the ecological management platform in a timely manner, thereby significantly improving the accuracy and timeliness of mine ecological risk prediction and effectively ensuring the ecological safety of the mine area. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 is a schematic execution flowchart of an artificial intelligence-based mine ecological risk prediction method provided by an embodiment of the present invention.

[0007] Figure 2 is a schematic diagram of exemplary hardware and software components of an artificial intelligence-based mine ecological risk prediction system provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0008] The present invention will be specifically described below with reference to the accompanying drawings of the specification. Figure 1 is a schematic flowchart of an artificial intelligence-based mine ecological risk prediction method provided by an embodiment of the present invention. The artificial intelligence-based mine ecological risk prediction method will be introduced in detail below.

[0009] Step S110: Collect multi-source ecological monitoring data through a distributed sensor network deployed in the mine area. The multi-source ecological monitoring data includes an environmental index sequence with continuous time and geographical coordinate information with spatial correlation.

[0010] To carry out ecological risk prediction work in the mining area, it is necessary to obtain relevant basic data first. In order to collect data comprehensively and accurately, a distributed sensor network will be deployed in the mining area. This distributed sensor network consists of multiple sensor nodes, which are distributed at various key locations in the mine, such as the mining area, around the tailings pond, near the water source, etc.

[0011] Each sensor node has multiple monitoring functions and can collect various environmental indicators in real time. These environmental indicators cover multiple aspects such as the atmosphere, water, soil, etc. For example, the pollutant concentration in the atmosphere, the pH value and dissolved oxygen content of water, the pH value and heavy metal content of soil, etc. The sensors will record these environmental indicators at regular time intervals, thus forming a time-continuous sequence of environmental indicators.

[0012] At the same time, each sensor node is equipped with a positioning device, which can accurately record its own geographical coordinate information. When the sensor collects environmental indicator data, it will associate this data with the corresponding geographical coordinate information. In this way, each set of environmental indicator data has its clear spatial location identifier, forming spatially associated geographical coordinate information.

[0013] For example, sensor A is located in the eastern mining area of the mine. It has collected a series of atmospheric pollutant concentration data at different time points, and at the same time recorded the longitude and latitude coordinates of this location. As time goes by, sensor A continuously collects data, obtaining a multi-source ecological monitoring data that includes a time-continuous sequence of atmospheric pollutant concentrations and the corresponding geographical coordinate information.

[0014] Step S120: Perform spatio-temporal alignment processing on the multi-source ecological monitoring data to generate a spatio-temporal association feature set, and the spatio-temporal association feature set includes trend change features in the time dimension and regional association features in the spatial dimension.

[0015] Since there are differences in the time points and positions of data collected by different sensors, in order to better analyze the data and mine the potential information therein, it is necessary to perform spatio-temporal alignment processing on the multi-source ecological monitoring data.

[0016] Step S121: Perform sliding window segmentation processing on the time series data in the multi-source ecological monitoring data to generate a set of data segments aligned in the time dimension, and each data segment corresponds to the change of environmental indicators within a preset time length.

[0017] For the time series data in the multi-source ecological monitoring data, the sliding window segmentation method is used for processing. The sliding window is a time period with a fixed length, which slides on the time series data. The preset time length is determined according to actual needs and analysis purposes.

[0018] In the specific operation, the sliding window starts from the starting point of the time series data and slides backward in sequence according to a certain step length. The data covered by each sliding window constitutes a data segment. Each data segment contains the change information of the environmental indicators within a preset time length.

[0019] For example, assume that the preset time length is T and the step size of the sliding window is S. Starting from the starting point of the time series data, the time range covered by the first sliding window is from t1 to t1+T, and the environmental indicator data within this time range constitutes the first data segment. Then, the sliding window moves backward by S time units, and the covered time range becomes t1+S to t1+S+T, which forms the second data segment. And so on, until the sliding window traverses the entire time series data, thereby generating a set of data segments aligned in the time dimension.

[0020] Step S122: Mapping the data segment set to the three-dimensional spatial coordinate system of the mining area to generate a monitoring data set associated with spatial dimensions, wherein each data unit in the monitoring data set includes a geographic coordinate identifier and a corresponding environmental indicator value.

[0021] After obtaining the data segment set aligned in the time dimension, it is necessary to associate these data segments with the spatial information of the mining area. To this end, the data segment set is mapped to the three-dimensional spatial coordinate system of the mining area.

[0022] The three-dimensional spatial coordinate system of the mining area is established according to the actual geographical conditions of the mine, which can accurately represent the spatial coordinates of each location in the mine. For each data fragment, according to the geographical coordinate information of its corresponding sensor node, it is mapped to the corresponding position in the three-dimensional spatial coordinate system.

[0023] In this way, each data segment corresponds to a specific location in the three-dimensional space coordinate system, forming a monitoring data set associated with the spatial dimension. In this monitoring data set, each data unit contains not only the geographic coordinate identifier, but also the environmental indicator value of the location within a preset time length.

[0024] For example, the sensor node corresponding to data segment A is located at point (x1, y1, z1) in the three-dimensional space coordinate system, and the data segment contains the environmental index values ​​such as the concentration of air pollutants and the pH value of water at the location within a preset time length. After mapping data segment A to the three-dimensional space coordinate system, a data unit containing the geographic coordinate identifier (x1, y1, z1) and the corresponding environmental index value is obtained.

[0025] Step S123: performing outlier detection processing on the monitoring data set, identifying and removing abnormal data points that exceed a preset fluctuation range, and generating a cleaned spatiotemporal correlation data set.

[0026] In the set of monitoring data associated with spatial dimensions, there may be some abnormal data points, which may be caused by sensor failures, external interferences, etc. If these abnormal data points are not processed, it will affect the subsequent analysis and prediction results. Therefore, it is necessary to perform outlier detection and processing on the set of monitoring data.

[0027] The preset fluctuation range is determined based on historical data and actual experience, and it represents the fluctuation range of environmental indicators under normal circumstances. For each data unit in the set of monitoring data, check whether the value of its environmental indicator exceeds the preset fluctuation range.

[0028] If the value of the environmental indicator of a certain data unit exceeds the preset fluctuation range, then mark this data unit as an abnormal data point. Then, remove these abnormal data points from the set of monitoring data, so as to generate a cleaned spatio-temporal association data set.

[0029] For example, for the environmental indicator of atmospheric pollutant concentration, the preset fluctuation range is [Cmin, Cmax]. If the value C of the atmospheric pollutant concentration of a certain data unit exceeds this preset fluctuation range, that is, C < Cmin or C > Cmax, then this data unit is marked as an abnormal data point and removed.

[0030] Step S124: Extract the trend change characteristics in the time dimension and the regional association characteristics in the spatial dimension from the cleaned spatio-temporal association data set, and merge them to generate a spatio-temporal association characteristic set.

[0031] After obtaining the cleaned spatio-temporal association data set, it is necessary to extract the trend change characteristics in the time dimension and the regional association characteristics in the spatial dimension from it.

[0032] Step S1241: Perform sliding window linear regression processing on the time-continuous environmental indicator sequence with each geographical coordinate identifier in the cleaned spatio-temporal association data set, calculate the fitting slope and residual variance of the environmental indicator within each sliding window, and generate the trend change characteristics in the time dimension.

[0033] For the time-continuous environmental indicator sequence corresponding to each geographical coordinate identifier in the cleaned spatio-temporal association data set, use the sliding window for processing again. Within each sliding window, perform linear regression analysis on the environmental indicator data.

[0034] The purpose of linear regression is to find a straight line to fit the environmental indicator data within the window, so that the error between the straight line and the data points is minimized. Through linear regression, the slope and residual variance of the fitting straight line can be calculated.

[0035] The fitting slope represents the change trend of the environmental indicator within the sliding window. A positive slope indicates an upward trend of the environmental indicator, while a negative slope indicates a downward trend. The residual variance represents the degree of dispersion between the data points and the fitting line, reflecting the stability of the change of the environmental indicator.

[0036] For example, for the time series of the concentration of atmospheric pollutants identified by a certain geographical coordinate, linear regression analysis is performed within a sliding window, and the fitting slope is k and the residual variance is σ². The fitting slope and the residual variance within each sliding window are combined to generate the trend change characteristics in the time dimension corresponding to the geographical coordinate.

[0037] Step S1242: Perform Fourier spectrum analysis on the time-continuous environmental indicator sequence, extract the amplitude peak value and phase offset within a preset frequency band as the periodic fluctuation sub-features in the time dimension, and perform time alignment and splicing on the periodic fluctuation sub-features and the trend change characteristics to generate the trend change characteristics in the time dimension.

[0038] In addition to the linear trend, there may also be periodic fluctuations in the environmental indicator data. To capture this periodic fluctuation information, Fourier spectrum analysis is performed on the time-continuous environmental indicator sequence.

[0039] Fourier spectrum analysis can convert the signal in the time domain into a signal in the frequency domain, so as to analyze the amplitude and phase of different frequency components in the signal. The preset frequency band is determined according to actual needs and experience, and it contains the frequency range where periodic fluctuations may exist.

[0040] Within the preset frequency band, the amplitude peak value and phase offset are extracted. The amplitude peak value represents the intensity of this frequency component, and the phase offset represents the starting position of this frequency component.

[0041] The extracted amplitude peak value and phase offset are used as the periodic fluctuation sub-features in the time dimension. Then, these periodic fluctuation sub-features are subjected to time alignment and splicing with the previously obtained trend change characteristics. Specifically, each data point in the periodic fluctuation sub-features is spliced with the data at the corresponding time point in the trend change characteristics, so as to generate the complete trend change characteristics in the time dimension.

[0042] For example, through Fourier spectrum analysis, the amplitude peak value within the preset frequency band is A and the phase offset is φ, and they are spliced with the previous fitting slope and residual variance in chronological order to form the new trend change characteristics in the time dimension.

[0043] Step S1243: Construct a regional association graph according to the spatial distribution of the geographical coordinate information, calculate the environmental indicator difference degree and spatial autocorrelation coefficient between each node and its adjacent nodes in the regional association graph, and generate the regional association characteristics in the spatial dimension.

[0044] To analyze the regional association characteristics in the spatial dimension, a regional association graph is constructed based on the spatial distribution of geographic coordinate information. The regional association graph is a graph structure, where each node represents a location identified by a geographic coordinate, and the edges between nodes represent the spatial adjacency relationship between nodes.

[0045] For each node in the regional association graph, calculate the difference degree of environmental indicators and the spatial autocorrelation coefficient between it and its adjacent nodes. The difference degree of environmental indicators represents the degree of difference in the environmental indicator values between two adjacent nodes, which can be obtained by calculating the absolute value of the difference between the environmental indicator values of the two nodes.

[0046] The spatial autocorrelation coefficient reflects the correlation between environmental indicators at adjacent locations in space. Common calculation methods include Moran's I coefficient, etc. By calculating the spatial autocorrelation coefficient, it can be determined whether the environmental indicators at adjacent locations are similar (positive correlation) or different (negative correlation).

[0047] Combine the difference degree of environmental indicators and the spatial autocorrelation coefficient of each node to generate the regional association characteristics in the spatial dimension.

[0048] For example, if the atmospheric pollutant concentrations of node i and its adjacent node j are Ci and Cj respectively, then the difference degree of environmental indicators between them is |Ci - Cj|. By calculating the difference degree of environmental indicators and the spatial autocorrelation coefficient between node i and all its adjacent nodes, the regional association characteristics of node i in the spatial dimension are obtained.

[0049] Step S1244: After standardizing the trend change characteristics and the regional association characteristics, align them according to the geographic coordinate identifier to generate a spatio-temporal association feature set.

[0050] To make the trend change characteristics and the regional association characteristics comparable, they need to be standardized. The purpose of standardization is to convert the data into a distribution with a mean of 0 and a standard deviation of 1.

[0051] For each feature vector in the trend change characteristics and the regional association characteristics, calculate its mean and standard deviation. Then, subtract the mean from each data point in each feature vector and divide by the standard deviation to obtain the standardized feature vector.

[0052] After completing the standardization process, align the trend change characteristics and the regional association characteristics according to the geographic coordinate identifier. Specifically, splice the trend change characteristics and the regional association characteristics with the same geographic coordinate identifier to form a new feature vector.

[0053] Combine the new feature vectors corresponding to all geographical coordinate identifiers to generate a spatio-temporal correlation feature set, which contains the trend change features in the time dimension and the regional correlation features in the spatial dimension.

[0054] Step S130: Invoke the pre-trained ecological risk prediction model to perform ecological risk prediction on the spatio-temporal correlation feature set, and generate a risk conduction mode set, which contains the starting node, propagation direction, and propagation intensity of the risk propagation path.

[0055] After obtaining the spatio-temporal correlation feature set, it is necessary to use the pre-trained ecological risk prediction model to perform ecological risk prediction on it. The pre-trained ecological risk prediction model is trained with a large number of sample data, and can effectively analyze the information in the spatio-temporal correlation feature set and predict the ecological risks in the mining area.

[0056] Step S131: Input the spatio-temporal correlation feature set into the time feature extraction module of the ecological risk prediction model, and deeply mine the trend change features in the time dimension through a multi-layer convolutional neural network to generate dynamic risk evolution features.

[0057] First, input the spatio-temporal correlation feature set into the time feature extraction module of the ecological risk prediction model. The time feature extraction module is mainly composed of a multi-layer convolutional neural network, and its purpose is to deeply mine the trend change features in the time dimension.

[0058] For example, step S1311: Segment the trend change features in the time dimension of the spatio-temporal correlation feature set according to the geographical coordinate identifier into time series segments, and each time series segment contains the fitting slope and periodic fluctuation sub-features within a preset time window.

[0059] For the trend change features in the time dimension of the spatio-temporal correlation feature set, segment them according to the geographical coordinate identifier. Each geographical coordinate identifier corresponds to a time series, and the time series is divided according to a preset time window to obtain multiple time series segments.

[0060] Each time series segment contains the fitting slope and periodic fluctuation sub-features within a preset time window. These features reflect the change trend and periodic fluctuation of the environmental indicators at the geographical coordinate identifier in the time dimension.

[0061] For example, for the time series corresponding to the geographical coordinate identifier (x, y, z), divide it into multiple segments according to a preset time window, and each segment contains the fitting slope and periodic fluctuation sub-features within that time period.

[0062] Step S1312: Input the time series segment into the first convolutional layer of the multi-layer convolutional neural network for local temporal pattern extraction to generate a set of primary temporal features. The first convolutional layer includes a one-dimensional convolutional kernel and a non-linear activation function.

[0063] Input the segmented time series segment into the first convolutional layer of the multi-layer convolutional neural network. The first convolutional layer includes a one-dimensional convolutional kernel and a non-linear activation function.

[0064] The one-dimensional convolutional kernel slides on the time series segment and performs a convolutional operation on the data within each window. The convolutional operation can extract local temporal patterns in the time series segment, such as short-term upward or downward trends.

[0065] The non-linear activation function then performs a non-linear transformation on the convolutional result to increase the expressive power of the model. Commonly used non-linear activation functions include the ReLU function, etc.

[0066] Through the processing of the first convolutional layer, each time series segment is converted into a primary temporal feature vector. Combine the primary temporal feature vectors corresponding to all time series segments to generate a set of primary temporal features.

[0067] For example, for a time series segment, after the convolution and non-linear activation processing of the first convolutional layer, a primary temporal feature vector is obtained. Combine the primary temporal feature vectors of all time series segments together to form a set of primary temporal features.

[0068] Step S1313: Input the set of primary temporal features into the second convolutional layer of the multi-layer convolutional neural network for cross-window temporal dependence modeling to generate a set of high-order temporal features. The second convolutional layer expands the temporal receptive field through dilated convolution.

[0069] Input the set of primary temporal features into the second convolutional layer of the multi-layer convolutional neural network. The main task of the second convolutional layer is to perform cross-window temporal dependence modeling, that is, to capture the dependence relationships between different time windows.

[0070] To expand the temporal receptive field, the second convolutional layer uses the method of dilated convolution. When performing the convolutional operation, there is a certain interval between the elements of the convolutional kernel, which can expand the receptive range of the convolutional kernel without increasing the size of the convolutional kernel.

[0071] Through dilated convolution, the second convolutional layer can capture longer-term temporal dependence relationships, thereby generating a set of high-order temporal features. Each feature vector in the set of high-order temporal features contains richer time-dimensional information.

[0072] For example, through dilated convolution, the second convolutional layer can capture the trend continuation or turning relationship between adjacent time windows, generating more representative high-order time series feature vectors. After all the primary time series feature vectors are processed by the second convolutional layer, a high-order time series feature set is obtained.

[0073] Step S1314: Perform residual connection processing on the high-order time series feature set, add the high-order time series feature set and the primary time series feature set element by element to generate an enhanced time series feature set.

[0074] To avoid the problem of gradient vanishing or gradient explosion in a multi-layer convolutional neural network, residual connection processing is performed on the high-order time series feature set. The basic idea of residual connection is to add the input features to the output features after being processed by the convolutional layer.

[0075] Specifically, the high-order time series feature set and the primary time series feature set are added element by element. The advantage of doing this is that some important information in the primary time series feature set can be retained, and at the same time, new information in the high-order time series feature set can also be fused.

[0076] Through residual connection processing, an enhanced time series feature set is generated. The feature vectors in the enhanced time series feature set contain more comprehensive time dimension information, which helps to improve the performance of the model.

[0077] For example, for a feature vector h in the high-order time series feature set and the corresponding feature vector p in the primary time series feature set, they are added element by element to obtain an enhanced time series feature vector e = h + p. After performing residual connection processing on all high-order time series feature vectors and primary time series feature vectors, an enhanced time series feature set is obtained.

[0078] Step S1315: Perform temporal max pooling processing on the enhanced time series feature set, extract the maximum response feature within each time window to generate dynamic risk evolution features.

[0079] To further extract important information from the enhanced time series feature set, temporal max pooling processing is performed on it. The operation of temporal max pooling is to select the maximum feature value within each time window as the representative feature of that window.

[0080] Through temporal max pooling processing, the dimension of the features can be reduced while retaining the most representative features within each time window. Combining the maximum response features of all time windows generates dynamic risk evolution features, which reflect the dynamic changes of environmental indicators in the time dimension. For example, for the feature vectors within a time window in the enhanced time series feature set, select the maximum feature value as the representative feature of that window. Combining all the maximum feature values of the time windows forms the dynamic risk evolution features.

[0081] Step S132: Input the spatio-temporal correlation feature set into the spatial feature extraction module of the ecological risk prediction model, and perform node relationship modeling on the regional association features in the spatial dimension through a graph neural network to generate spatial risk diffusion features.

[0082] In addition to feature extraction in the time dimension, it is also necessary to process the regional association features in the spatial dimension. Input the spatio-temporal correlation feature set into the spatial feature extraction module of the ecological risk prediction model. The spatial feature extraction module is mainly composed of a graph neural network and is used to perform node relationship modeling on the regional association features in the spatial dimension.

[0083] For example, step S1321: According to the geographical coordinate identifiers and regional association features in the spatio-temporal correlation feature set, construct a node feature set of the regional association graph, where each node corresponds to a unique geographical coordinate and the node features include the environmental index difference degree and the spatial autocorrelation coefficient.

[0084] Construct a node feature set of the regional association graph according to the geographical coordinate identifiers and regional association features in the spatio-temporal correlation feature set. In the regional association graph, each node corresponds to a unique geographical coordinate, and the features of the node include the environmental index difference degree and the spatial autocorrelation coefficient at that location.

[0085] For example, for the geographical coordinate identifier (x1, y1, z1), its corresponding node features include the environmental index difference degree and the spatial autocorrelation coefficient between this location and adjacent nodes. Combine the node features corresponding to all geographical coordinate identifiers to form the node feature set of the regional association graph.

[0086] Step S1322: Input the regional association graph into the first graph convolutional layer of the graph neural network for neighborhood information aggregation to generate a primary node feature set. The first graph convolutional layer fuses the features of adjacent nodes through a weighted summation method, and the weights are determined by the product of the normalized environmental index difference degree and the spatial autocorrelation coefficient.

[0087] Input the constructed region association graph into the first graph convolutional layer of the graph neural network. The main task of the first graph convolutional layer is to perform neighborhood information aggregation, that is, to fuse the feature information of each node and its adjacent nodes. When performing neighborhood information aggregation, the first graph convolutional layer uses the method of weighted summation. The weights here are determined by the product of the normalized environmental index difference degree and the spatial autocorrelation coefficient. First, perform normalization processing on the environmental index difference degree, aiming to convert environmental index difference degree data of different magnitudes to a unified scale to ensure that in subsequent calculations, no data will have too large or too small an impact on the results due to magnitude differences. Let the environmental index difference degree of node i be D_i, the spatial autocorrelation coefficient be R_i, and the normalized environmental index difference degree be denoted as D_i_normalized. Normalization can be achieved using common linear normalization methods, such as subtracting the minimum value of the environmental index difference degree and then dividing by the difference between the maximum value and the minimum value. Assuming the minimum value of the environmental index difference degree is D_min and the maximum value is D_max, then D_i_normalized = (D_i - D_min) / (D_max - D_min). Then calculate the weight W_i = D_i_normalized × R_i. For node i, its primary node feature is obtained by performing a weighted sum of the features of all its adjacent nodes. Let the set of adjacent nodes of node i be N(i), and the feature vector of adjacent node j be F_j. Then the primary node feature F_i_primary output by node i in the first graph convolutional layer can be expressed as F_i_primary = Σ(W_j × F_j), where j belongs to N(i). Combine the primary node features of all nodes to generate a primary node feature set.

[0088] Step S1323: Input the primary node feature set into the second graph convolutional layer of the graph neural network for cross-level feature propagation to generate a high-order node feature set. The second graph convolutional layer dynamically adjusts the information transfer intensity between nodes through an attention mechanism.

[0089] Input the set of primary node features into the second graph convolutional layer of the graph neural network. The main function of this layer is to perform cross-level feature propagation and further explore more complex relationships between nodes. To achieve this goal, the second graph convolutional layer introduces an attention mechanism. The attention mechanism can dynamically adjust the intensity of information transmission according to the feature correlation between nodes. For each node, the attention mechanism calculates the attention weights between this node and other nodes. Specifically, first perform a linear transformation on the features of the node to map the primary node feature vector to a new feature space. Let the primary node feature vector of node i be \(F_{i\_primary}\), and it is transformed into \(F_{i\_transformed}=A\times F_{i\_primary}\) through the linear transformation matrix A. Then calculate the attention score \(S_{ij}\) between node i and node j, which reflects the degree of attention of node i to the information of node j. The attention score can be calculated through methods such as dot product and concatenation. For example, using the dot product method, \(S_{ij}=F_{i\_transformed}\times F_{j\_transformed}^T\). Then normalize the attention score, usually using the softmax function, to obtain the attention weight \(\alpha_{ij}=\frac{\exp(S_{ij})}{\sum_{k}\exp(S_{ik})}\), where k belongs to all relevant node sets. With the attention weight, the output feature \(F_{i\_secondary}\) of node i in the second graph convolutional layer can be expressed as \(F_{i\_secondary}=\sum_{j}(\alpha_{ij}\times F_{j\_transformed})\), where j belongs to all relevant node sets. Combine the output features of all nodes in the second graph convolutional layer to generate a set of high-order node features.

[0090] Step S1324: Perform graph pooling on the set of high-order node features, screen key region nodes based on the node feature amplitude and the spatial autocorrelation coefficient, retain the nodes whose node feature amplitude is higher than the preset amplitude threshold and the spatial autocorrelation coefficient is greater than the critical value, and generate a set of regional node embedding features.

[0091] Perform graph pooling on the high-order node feature set, aiming to screen out the nodes in the key regions, reduce data redundancy, and at the same time retain the node information that is of great significance for risk prediction. When performing graph pooling, nodes are mainly screened based on the node feature amplitude and the spatial autocorrelation coefficient. The node feature amplitude reflects the importance of the node feature, and the spatial autocorrelation coefficient reflects the correlation between the node and its surrounding nodes. The preset amplitude threshold and critical value are determined according to the actual situation and experience. For each node in the high-order node feature set, calculate its feature amplitude. For example, the modulus length of the feature vector can be used to represent the feature amplitude. Let the high-order node feature vector of node i be F_i_secondary, and its feature amplitude be ||F_i_secondary||. At the same time, obtain the spatial autocorrelation coefficient R_i of this node. If ||F_i_secondary|| is higher than the preset amplitude threshold and R_i is greater than the critical value, then this node is retained as a key region node. Combine the feature vectors of all the retained key region nodes to generate the regional node embedding feature set.

[0092] Step S1325: Multiply each element of the regional node embedding feature set by the spatial autocorrelation coefficient in the spatio-temporal correlation feature set to generate the spatial risk diffusion feature.

[0093] Perform an element-wise multiplication operation on the generated regional node embedding feature set and the spatial autocorrelation coefficient in the spatio-temporal correlation feature set. Since the spatial autocorrelation coefficient reflects the spatial correlation between nodes, through this element-wise multiplication method, the spatial correlation information can be incorporated into the regional node embedding features, thereby generating features that can reflect the spatial risk diffusion situation. Let the feature vector of node i in the regional node embedding feature set be E_i, and the spatial autocorrelation coefficient of node i in the spatio-temporal correlation feature set be R_i. Then the spatial risk diffusion feature S_i of node i = E_i × R_i. Combine the spatial risk diffusion features of all nodes to obtain the spatial risk diffusion feature set.

[0094] Step S133: Perform two-way attention interaction processing on the dynamic risk evolution feature and the spatial risk diffusion feature through the feature interaction module of the ecological risk prediction model to generate the attention interaction feature.

[0095] In order to fully integrate the feature information in the time dimension and the spatial dimension, use the feature interaction module of the ecological risk prediction model to perform two-way attention interaction processing on the dynamic risk evolution feature and the spatial risk diffusion feature.

[0096] Step S1331: Perform a linear transformation on the dynamic risk evolution feature to generate a query vector set, and perform a linear transformation on the spatial risk diffusion feature to generate a key vector set.

[0097] First, perform a linear transformation on the dynamic risk evolution features, map them to a new feature space, and generate a set of query vectors. The linear transformation can be achieved through a linear transformation matrix. Let the feature vector of node i in the dynamic risk evolution features be D_i, and the linear transformation matrix be Q. Then the query vector q_i of node i is q_i = Q × D_i. Similarly, perform a linear transformation on the spatial risk diffusion features to generate a set of key vectors. Let the feature vector of node i in the spatial risk diffusion features be S_i, and the linear transformation matrix be K. Then the key vector k_i of node i is k_i = K × S_i. Combine the query vectors and key vectors of all nodes respectively to obtain a set of query vectors and a set of key vectors.

[0098] Step S1332: Divide the dynamic risk evolution features and the spatial risk diffusion features into regions according to geographical coordinates, calculate the local similarity matrix of the set of query vectors and the set of key vectors within each region, and generate a set of spatio-temporal correlation attention weights for the divided regions.

[0099] Divide the dynamic risk evolution features and the spatial risk diffusion features according to geographical coordinates, so that feature interaction analysis can be performed separately for different regions. Within each region, calculate the local similarity matrix of the set of query vectors and the set of key vectors. Each element in the similarity matrix represents the similarity between a query vector and a key vector within the region. Methods such as dot product can be used to calculate the similarity. Let the set of query vectors in region r be Q_r, the set of key vectors be K_r, and the element S_r_ij of the similarity matrix S_r represents the similarity between the query vector q_r_i and the key vector k_r_j, that is, S_r_ij = q_r_i × k_r_j^T. Then normalize the similarity matrix, for example, use the softmax function, to obtain a set of spatio-temporal correlation attention weights for the divided regions. Let the normalized attention weight matrix be A_r, and its element A_r_ij = exp(S_r_ij) / Σexp(S_r_ik), where k belongs to all the key vector sets within the region.

[0100] Step S1333: Perform parallel weighted aggregation processing on the spatial risk diffusion features according to the set of spatio-temporal correlation attention weights for the divided regions to generate spatial attention features.

[0101] Perform parallel weighted aggregation processing on the spatial risk diffusion characteristics according to the sub-region spatio-temporal correlation attention weight set. For each region, perform weighted summation on the spatial risk diffusion feature vectors within the region according to the attention weights. Let the set of spatial risk diffusion feature vectors in region r be S_r, and the attention weight matrix be A_r. Then the spatial attention feature vector SA_r of this region can be expressed as SA_r = Σ(A_r_ij × S_r_j), where j belongs to all the spatial risk diffusion feature vector sets within the region. Combine the spatial attention feature vectors of all regions to generate a spatial attention feature set.

[0102] Step S1334: Perform residual connection processing on the spatial attention feature and the dynamic risk evolution feature to generate an attention interaction feature.

[0103] In order to retain the original information in the dynamic risk evolution feature and at the same time fuse the spatial attention feature, perform residual connection processing on the two to obtain an attention interaction feature set.

[0104] Step S134: Input the attention interaction feature into the prediction module of the ecological risk prediction model to generate a set of risk conduction patterns.

[0105] Input the generated attention interaction feature set into the prediction module of the ecological risk prediction model. The prediction module is a trained neural network module, which can predict the risk conduction pattern according to the input attention interaction feature. The prediction module analyzes and learns various information in the attention interaction feature and outputs a set of risk conduction patterns. This set of risk conduction patterns contains information such as the starting node, propagation direction, and propagation intensity of the risk propagation path. During the training process, the prediction module learns the mapping relationship between the attention interaction feature and the risk conduction pattern through a large number of sample data, so it can accurately predict the risk conduction pattern according to the input attention interaction feature.

[0106] Step S140: Construct an ecological risk evolution network for the mining area based on the set of risk conduction patterns. The ecological risk evolution network includes the topological structure of risk nodes and the temporal dependence relationship of risk diffusion.

[0107] After obtaining the set of risk conduction patterns, use this information to construct an ecological risk evolution network for the mining area to more intuitively display the spread and evolution of risks in the mining area.

[0108] Step S141: Extract the starting node and propagation direction of the risk propagation path from the set of risk conduction patterns to construct an initial topological structure of risk nodes.

[0109] Extract the starting nodes and propagation direction information of each risk propagation path from the set of risk conduction patterns. The starting node represents the location where the risk begins to spread, and the propagation direction indicates the direction of risk spread. Based on this information, construct the initial topological structure of the risk nodes. In this topological structure, each node represents a possible risk location, and the directed edges between the nodes represent the direction of risk spread. For example, if the set of risk conduction patterns indicates that the risk spreads from node A to node B, then there will be a directed edge from node A to node B in the initial topological structure. Integrate the starting nodes and propagation direction information of all risk propagation paths to construct the initial topological structure of the risk nodes.

[0110] Step S142: Calculate the connection weights between risk nodes according to the propagation intensity in the set of risk conduction patterns, and generate a set of weighted risk propagation edges.

[0111] The set of risk conduction patterns also contains the intensity information of risk spread. According to this propagation intensity information, calculate the connection weights between risk nodes. The connection weight reflects the possibility and intensity of risk spread between two nodes. The value of the propagation intensity can be used as the basis for the connection weight. Let the propagation intensity of the risk from node i to node j be I_ij, then the connection weight W_ij between node i and node j can be directly set to I_ij. Combine the connection weights between all risk nodes to generate a set of weighted risk propagation edges. Each edge in this set of weighted risk propagation edges has a corresponding connection weight, which is used to describe the intensity of risk spread between nodes.

[0112] Step S143: Perform iterative optimization processing on the initial topological structure based on the set of weighted risk propagation edges to generate an ecological risk evolution network containing multi-level risk diffusion paths.

[0113] Step S1431: Perform clustering analysis on the risk nodes in the initial topological structure to generate a set of risk node communities. Each risk node community contains risk nodes that are spatially adjacent and have similar risk propagation intensities.

[0114] Cluster analysis is performed on the risk nodes in the initial topological structure. The purpose of cluster analysis is to divide risk nodes that are spatially adjacent and have similar risk propagation intensities into the same community. A density-based clustering algorithm, such as the DBSCAN algorithm, can be used. This algorithm determines core points and border points by defining a neighborhood radius and a minimum number of points, thereby dividing the nodes into different communities. When performing clustering, consider the spatial location and risk propagation intensity of the nodes. The spatial location can be represented by geographical coordinates, and the risk propagation intensity can use the connection weights calculated previously. For each node in the initial topological structure, calculate the spatial distance and the difference in risk propagation intensity between it and other nodes. If the spatial distance between two nodes is less than the neighborhood radius and the difference in risk propagation intensity is within a certain range, then they may be divided into the same community. Through cluster analysis, all risk nodes are divided into multiple sets of risk node communities, and the nodes in each community have similar spatial locations and risk propagation intensities.

[0115] Step S1432: Calculate the cross-community propagation probability between different risk node communities according to the weighted risk propagation edge set, and generate an inter-community propagation path.

[0116] Calculate the cross-community propagation probability between different risk node communities according to the weighted risk propagation edge set. For two different risk node communities C_i and C_j, first find all the connection edges between the nodes in community C_i and the nodes in community C_j. Then, sum up the connection weights of these connection edges to obtain the total connection weight W_ij_total from community C_i to community C_j. At the same time, calculate the total connection weight W_i_total of all nodes in community C_i (that is, the sum of the connection weights between the nodes in community C_i and all other nodes). Then the cross-community propagation probability P_ij from community C_i to community C_j is P_ij = W_ij_total / W_i_total. According to the calculated cross-community propagation probability, determine the inter-community propagation path. If the cross-community propagation probability is greater than a preset threshold, then it is considered that there is a propagation path from community C_i to community C_j. Combine all the inter-community propagation paths that meet the conditions to generate an inter-community propagation path set.

[0117] Step S1433: Add the inter-community propagation path to the initial topological structure to generate an extended risk topological structure.

[0118] Add the generated inter-community transmission paths to the initial topological structure to obtain an extended risk topological structure. When adding inter-community transmission paths, new directed edges need to be added to the topological structure to represent these paths. At the same time, assign corresponding connection weights to these new directed edges, and the connection weights can be set as the corresponding cross-community transmission probabilities. By adding inter-community transmission paths, the extended risk topological structure can more comprehensively reflect the spread of risks between different communities.

[0119] Step S1434: Perform path redundancy elimination processing on the extended risk topological structure to generate an optimized ecological risk evolution network.

[0120] There may be some path redundancy situations in the extended risk topological structure. For example, some paths can be indirectly represented by other paths. To simplify the network structure and improve the analysis efficiency, perform path redundancy elimination processing on the extended risk topological structure. Graph theory path search algorithms, such as depth-first search algorithm or breadth-first search algorithm, can be used to find all possible paths. Then, compare and analyze these paths to determine whether there are redundant paths. If there are redundant paths, delete them from the topological structure. After path redundancy elimination processing, an optimized ecological risk evolution network is obtained. This network contains multi-level risk diffusion paths and can more accurately describe the spread and evolution of risks in the mining area.

[0121] Step S144: Model the temporal dependence relationship in the ecological risk evolution network to generate a time series prediction result of risk diffusion.

[0122] Step S1441: Extract the historical risk intensity sequences of each risk node from the ecological risk evolution network to generate a set of temporal characteristics of risk nodes. The historical risk intensity sequences contain the risk propagation intensity values recorded at preset time intervals.

[0123] Extract the historical risk intensity sequences of each risk node from the optimized ecological risk evolution network. The historical risk intensity sequences are the risk propagation intensity values recorded at preset time intervals. The preset time interval is determined according to the actual situation and data collection frequency. For each risk node, collect its risk propagation intensity values at different time points to form a time series. For example, if the risk propagation intensities recorded at time points t_1, t_2, t_3, etc. for node i are I_i1, I_i2, I_i3, etc., then the historical risk intensity sequence of node i is [I_i1, I_i2, I_i3,...]. Combine the historical risk intensity sequences of all risk nodes to generate a set of temporal characteristics of risk nodes.

[0124] Step S1442: Construct a temporal graph network based on the topological structure of the risk nodes and the historical risk intensity sequence. The node attributes in the temporal graph network include the temporal risk intensity characteristics of the corresponding risk nodes, and the edge weights are determined by the connection weights in the weighted risk propagation edge set.

[0125] Construct a temporal graph network based on the topological structure of risk nodes and the historical risk intensity sequence. In the temporal graph network, each node represents a risk node, and the attributes of the node include the temporal risk intensity characteristics of the risk node, that is, the historical risk intensity sequence. The weights of the edges are determined by the connection weights in the weighted risk propagation edge set. In this way, the temporal graph network contains both the spatial relationship between risk nodes (represented by the connection and weights of the edges) and the time series information of the risk nodes (represented by the node attributes). By constructing the temporal graph network, the spatial and time information in the ecological risk evolution network can be integrated.

[0126] Step S1443: Input the temporal graph network into a pre-trained temporal graph neural network model, and perform causal convolution processing on the node attributes in the time dimension through a spatio-temporal convolutional layer to extract the risk intensity change patterns between adjacent time steps.

[0127] Input the constructed temporal graph network into a pre-trained temporal graph neural network model. The spatio-temporal convolutional layer of this model will perform causal convolution processing on the node attributes in the time dimension. Causal convolution is a special convolution operation that only considers the information of the current time step and the previous time steps, which conforms to the causal relationship of time series data. When performing causal convolution, the convolutional kernel slides from front to back in the time dimension, and performs convolution calculation on the node attributes of each time step. Through causal convolution processing, the risk intensity change patterns between adjacent time steps can be extracted. For example, by analyzing the convolution results, it can be found whether the risk intensity shows an upward trend, a downward trend or a fluctuating trend, etc.

[0128] Step S1444: Calculate the dynamic association weights between risk nodes at different time steps through the spatio-temporal attention layer of the temporal graph neural network model to generate a spatio-temporal attention weight matrix.

[0129] The spatio-temporal attention layer of the temporal graph neural network model calculates the dynamic association weights between risk nodes at different time steps. At different time steps, the degree of association between risk nodes may change. The spatio-temporal attention layer dynamically calculates the attention weights between nodes based on node attributes and edge weight information. Specifically, for each time step t, first perform a linear transformation on the node attributes to obtain query vectors, key vectors, and value vectors. Then calculate the similarity between the query vector and the key vector, and perform normalization processing through the softmax function to obtain the attention weights. Combine the attention weights between different nodes at all time steps to generate a spatio-temporal attention weight matrix. This spatio-temporal attention weight matrix can reflect the dynamic association relationship between risk nodes at different time steps, and helps to capture the propagation law of risks in time and space.

[0130] Step S1445: Aggregate the node attributes in the temporal graph network across time steps according to the spatio-temporal attention weight matrix to generate a set of temporal hidden states of each risk node.

[0131] Based on the generated spatio-temporal attention weight matrix, perform cross-time-step feature aggregation on the node attributes in the temporal graph network. For each risk node, according to the association weights between this node and other nodes at different time steps in the spatio-temporal attention weight matrix, perform weighted summation on the attributes of other nodes at different time steps. Let the attribute vector of node i at time step t be A_it, and the association weight between node i and node j at time steps t and t' in the spatio-temporal attention weight matrix be W_ij_tt'. Then the temporal hidden state H_it of node i at time step t can be expressed as H_it = Σ(W_ij_tt' × A_jt'), where j is all nodes and t' is all time steps. Combine the temporal hidden states of all risk nodes at all time steps to generate a set of temporal hidden states of each risk node. This set of temporal hidden states contains the comprehensive feature information of risk nodes in the time dimension.

[0132] Step S1446: Input the set of temporal hidden states into the multi-step prediction layer of the temporal graph neural network model, and decode step by step to generate the risk intensity prediction values of each risk node within a preset future time period.

[0133] Input the set of temporal hidden states of each risk node into the multi-step prediction layer of the temporal graph neural network model. The main function of the multi-step prediction layer is to decode and generate the predicted risk intensity values of each risk node within a preset future time period step by step based on the input temporal hidden states. By learning the patterns and regularities in the temporal hidden states and leveraging the non-linear mapping ability of the neural network, this layer predicts the future risk intensity. Specifically, during operation, starting from the current time step, the risk intensity of the next time step is predicted based on the current temporal hidden state, and then the prediction result is used as the new input to continue predicting the risk intensity of the next time step, and so on, until all the predicted risk intensity values within the preset future time period are generated. For example, if the preset time period is the next T time steps, then the multi-step prediction layer will sequentially generate the predicted risk intensity values of each risk node from the 1st time step to the Tth time step after the current time step.

[0134] Step S1447: Generate the risk diffusion time series prediction result of the ecological risk evolution network according to the predicted risk intensity value and the spatial positions of the risk nodes in the topological structure.

[0135] Combine the predicted risk intensity value and the spatial position information of the risk nodes in the topological structure to generate the risk diffusion time series prediction result of the ecological risk evolution network. For the predicted risk intensity value at each time step, analyze the spatial diffusion of the risk according to the spatial positions of the risk nodes. For example, if the predicted risk intensity value of a certain risk node is high and there are connection edges between this node and other nodes, then it can be predicted that the risk will spread along these connection edges to adjacent nodes. At the same time, consider the connection weights between different nodes. The greater the connection weight, the higher the possibility and intensity of risk diffusion. Through the comprehensive analysis of the predicted risk intensity value and the node spatial position at each time step, obtain the spatial diffusion situation of the risk at different time steps, thereby generating the risk diffusion time series prediction result of the ecological risk evolution network. This risk diffusion time series prediction result can intuitively display the dynamic diffusion process of the risk in the mining area over time.

[0136] Step S150: Generate a risk warning instruction set according to the ecological risk evolution network. The risk warning instruction set includes the risk level classification result and the corresponding ecological restoration strategy, and send the risk warning instruction set to the ecological management platform.

[0137] Step S151: Perform a priority evaluation process on the risk nodes in the ecological risk evolution network to generate a risk level distribution map.

[0138] Perform a priority assessment on the risk nodes in the ecological risk evolution network to determine the risk levels of each node. The assessment process comprehensively considers multiple factors, including the risk intensity of the risk node, the risk propagation speed, the influence range, and the potential harm degree to the surrounding ecological environment and human activities, etc. For nodes with higher risk intensity, faster propagation speed, larger influence range, and greater potential harm to the surrounding area, a higher priority is assigned; conversely, a lower priority is assigned. Corresponding weights can be set for each assessment factor, and then the comprehensive priority score of each risk node is calculated by weighted summation. For example, let the weight of risk intensity be w1, the weight of risk propagation speed be w2, the weight of influence range be w3, and the weight of potential harm degree be w4. The risk intensity score of node i is S1i, the risk propagation speed score is S2i, the influence range score is S3i, and the potential harm degree score is S4i (all are normalized converted score values). Then the comprehensive priority score Pi of node i = w1×S1i + w2×S2i + w3×S3i + w4×S4i. According to the comprehensive priority score, the risk nodes are divided into different risk levels, such as high-risk, medium-risk, and low-risk levels (here is just an example to illustrate the level concept, and the high, medium, and low expressions are not actually used). Integrate the risk level information of all risk nodes to generate a risk level distribution map, which visually shows the risk level distribution of different locations in the mining area.

[0139] Step S152: Identify the spatial coordinate ranges of the high-risk areas and potential risk areas from the risk level distribution map.

[0140] Identify the spatial coordinate ranges of the high-risk areas and potential risk areas from the generated risk level distribution map. High-risk areas refer to areas with relatively high risk levels, where the risk intensity is high, the propagation speed is fast, and they may cause serious harm to the ecological environment and human activities. Potential risk areas refer to areas that, although the current risk level is relatively low, have certain risk hidden dangers and may develop into high-risk areas in the future. By analyzing the positions and risk levels of each risk node in the risk level distribution map, determine the boundaries of the high-risk areas and potential risk areas. Clustering analysis and other methods can be used to divide adjacent risk nodes with similar risk levels into the same area. For high-risk areas and potential risk areas, record the spatial coordinates of their boundary nodes to determine the spatial coordinate ranges of these areas.

[0141] Step S153: Match the spatial coordinate range of the high-risk area with the preset emergency response strategy library to generate a high-risk area warning instruction.

[0142] The preset emergency response strategy library stores emergency response strategies for high-risk areas with different types and different spatial scopes. Match the spatial coordinate range of the identified high-risk area with the strategies in the emergency response strategy library. The matching process can be carried out according to the characteristics such as the location, scale, and risk type of the high-risk area. For example, if the high-risk area is a soil pollution area caused by mine exploitation, and the spatial coordinate range and scale of the soil pollution area match the situation targeted by a certain strategy in the emergency response strategy library, then select this strategy as the emergency response strategy for this high-risk area. Generate a warning instruction for the high-risk area according to the matching result. The warning instruction contains emergency measures for the high-risk area, such as restricting personnel entry, starting pollution treatment equipment, and strengthening the monitoring frequency, etc.

[0143] Step S154: Extract the environmental index characteristics and spatial distribution characteristics of the potential risk area to generate a potential risk feature vector.

[0144] Extract the environmental index characteristics and spatial distribution characteristics of the potential risk area. The environmental index characteristics include information such as the concentration of air pollutants, water quality indicators, and soil acidity in this area, and this information can be obtained from the multi-source ecological monitoring data collected before. The spatial distribution characteristics include information such as the shape, area, and positional relationship with the surrounding environment of the potential risk area. Organize and combine these environmental index characteristics and spatial distribution characteristics to generate a potential risk feature vector. For example, assume that the concentration of air pollutants in the potential risk area is C1, the water quality indicator is C2, the soil acidity is C3, the area of the area is A, the distance from the surrounding water source is D, etc., then the potential risk feature vector V = [C1, C2, C3, A, D,...]. This potential risk feature vector can comprehensively describe the characteristics of the potential risk area.

[0145] Step S155: Retrieve a set of sample cases that match the potential risk feature vector from the sample repair case database.

[0146] A large number of historical ecological restoration cases are stored in the sample restoration case database. Each case contains the environmental index characteristics, spatial distribution characteristics of the case, and the corresponding restoration plan. Match the generated potential risk feature vector with the cases in the sample restoration case database. The matching process can adopt the method of similarity calculation, such as calculating the cosine similarity between the potential risk feature vector and the feature vector of each case. Let the potential risk feature vector be V, and the feature vector of case i be Vi, and the cosine similarity Si between them is Si=(V·Vi) / (||V||×||Vi||). According to the similarity size, select the case with a higher similarity as the matching sample case. Combine all the matching sample cases to generate a sample case set. The cases in this sample case set are similar to the characteristics of the potential risk area, and their restoration plans may have reference value for the restoration of the potential risk area.

[0147] Step S156: Perform an effect evaluation process on the restoration plans in the sample case set to generate a result of sorting the priorities of the plans.

[0148] Perform an effect evaluation on the restoration plans in the sample case set. The evaluation process considers multiple factors, such as the cost of the restoration plan, the restoration time, the durability of the restoration effect, etc. Set corresponding weights for each evaluation factor, and then calculate the comprehensive evaluation score of each restoration plan by weighted summation. For example, let the weight of the restoration cost be w5, the weight of the restoration time be w6, the weight of the durability of the restoration effect be w7, the restoration cost score of plan j be S5j, the restoration time score be S6j, and the restoration effect durability score be S7j, then the comprehensive evaluation score Pj' of plan j is Pj'=w5×S5j+w6×S6j+w7×S7j. Sort the priorities of the restoration plans in the sample case set according to the comprehensive evaluation score. The higher the score, the higher the priority of the plan.

[0149] Step S157: Select a target restoration plan according to the result of sorting the priorities of the plans and generate a risk prevention and control recommendation instruction.

[0150] According to the result of sorting the priorities of the plans, select the restoration plan with the highest priority as the target restoration plan. The target restoration plan is the most suitable plan for the restoration of the potential risk area, and it has good comprehensive performance in terms of cost, time, and effect. Based on the target restoration plan, generate a risk prevention and control recommendation instruction. The risk prevention and control recommendation instruction contains specific prevention and control measures for the potential risk area, such as what restoration technology to adopt, when to start the restoration, the resources to be invested, etc. Combine the high-risk area warning instruction and the risk prevention and control recommendation instruction to form a risk warning instruction set. Finally, send the risk warning instruction set to the ecological management platform so that the ecological management department can take corresponding measures in time to prevent and control and restore the ecological risks in the mine area.

[0151] Step S210: The training method of the pre-trained ecological risk prediction model includes obtaining a sample mine ecological data set, which contains multiple groups of spatio-temporally correlated sample multi-source ecological monitoring data and corresponding risk annotation results.

[0152] To train the ecological risk prediction model, it is necessary to obtain a sample mine ecological data set first. This data set is obtained by deploying a sensor network in multiple mine areas to collect a large amount of multi-source ecological monitoring data. These data contain a time-continuous sequence of environmental indicators and spatially correlated geographical coordinate information, and after spatio-temporal alignment processing, they form spatio-temporally correlated sample multi-source ecological monitoring data. At the same time, in order to provide the model with labels for supervised learning, risk annotation is performed on these sample data. The risk annotation results include information such as the starting node, propagation direction, and propagation intensity of the risk propagation path. These annotation results can be obtained through methods such as expert experience and historical data statistical analysis. Combining multiple groups of spatio-temporally correlated sample multi-source ecological monitoring data and corresponding risk annotation results, a sample mine ecological data set is obtained.

[0153] Step S220: Perform data augmentation processing on the sample multi-source ecological monitoring data, and after merging the augmented multi-source ecological monitoring data and the sample multi-source ecological monitoring data, extract a standardized spatio-temporal correlation feature set to generate a training sample set, where the data augmentation processing includes time series interpolation, noise injection, and restricted spatial coordinate perturbation, and the maximum offset distance of the spatial coordinate perturbation does not exceed a preset ratio of the adjacent sensor spacing.

[0154] Perform data augmentation processing on the sample multi-source ecological monitoring data to increase the diversity and richness of the data and improve the generalization ability of the model. The data augmentation processing includes the following methods: Time series interpolation: For time series data, interpolation methods are used to insert new data points between the original time points. For example, linear interpolation, spline interpolation, etc. can be used. Linear interpolation calculates the data value of the inserted point based on the linear relationship between the data values of two adjacent time points. Let the data values at time points t1 and t2 be y1 and y2 respectively, and to insert a time point t between t1 and t2, the data value y of the inserted point is y = y1+(y2 - y1)×(t - t1) / (t2 - t1). Through time series interpolation, the density of time series data can be increased, enabling the model to learn more detailed time change patterns.

[0155] Noise injection: Inject a certain amount of noise into the multi-source ecological monitoring data of samples. The noise can be random noise such as Gaussian noise. The purpose of noise injection is to simulate the measurement errors and interferences that may exist in the actual data and improve the robustness of the model. Let the original data value be x, and the injected noise be ε (ε follows a Gaussian distribution), then the data value after injecting noise is x' = x + ε.

[0156] Restricted space coordinate perturbation: Perturb the geographical coordinate information in the multi-source ecological monitoring data of samples. The maximum offset distance of the perturbation does not exceed a preset ratio of the adjacent sensor spacing. For example, let the adjacent sensor spacing be d and the preset ratio be r, then the maximum offset distance of the coordinate perturbation is r×d. When performing coordinate perturbation, randomly make a small offset to the geographical coordinates to generate new spatial coordinates. This can increase the diversity of data in the spatial dimension.

[0157] Merge the enhanced multi-source ecological monitoring data obtained through data augmentation processing and the original multi-source ecological monitoring data of samples. Then, perform spatio-temporal correlation feature extraction and normalization processing on the merged data. The method of spatio-temporal correlation feature extraction is the same as the method described in the previous steps S121 - S124, including trend change feature extraction in the time dimension and regional correlation feature extraction in the spatial dimension. Normalization processing is to normalize the extracted features so that the features have the same scale, which is convenient for model learning. Combine the normalized spatio-temporal correlation feature set with the corresponding risk annotation results to generate a training sample set.

[0158] Step S230: Construct an initial ecological risk prediction model, and the initial ecological risk prediction model includes a time feature extraction module, a spatial feature extraction module, a feature interaction module, and a prediction module.

[0159] In this embodiment, the time feature extraction module: mainly consists of a multi-layer convolutional neural network and is used to deeply mine the trend change features in the time dimension. The specific structure and processing process of this module are the same as those described in step S131, including operations such as local time series pattern extraction, cross-window time series dependence modeling, residual connection processing, and time series max pooling processing on time series segments, and finally generating dynamic risk evolution features.

[0160] The spatial feature extraction module: consists of a graph neural network and is used to model the node relationships of the regional correlation features in the spatial dimension. The specific structure and processing process of this module are the same as those described in step S132, including operations such as constructing a node feature set of the regional correlation graph, neighborhood information aggregation, cross-level feature propagation, and graph pooling processing, and finally generating spatial risk diffusion features.

[0161] Feature Interaction Module: It is used to perform bidirectional attention interaction processing on the dynamic risk evolution features and spatial risk diffusion features. The specific structure and processing process of this module are the same as those described in step S133, including operations such as generating a query vector set and a key vector set, calculating a regional spatio-temporal correlation attention weight set, performing parallel weighted aggregation processing and residual connection processing, and finally generating attention interaction features.

[0162] Prediction Module: It is a trained neural network module that predicts the risk conduction pattern based on the input attention interaction features. This module learns the mapping relationship between the attention interaction features and the risk conduction pattern during the training process and outputs a set of risk conduction patterns.

[0163] Step S240: Input the spatio-temporal correlation feature set in the training sample set into the time feature extraction module and the spatial feature extraction module of the initial ecological risk prediction model, and respectively output the training dynamic risk evolution features and the training spatial risk diffusion features.

[0164] Input the spatio-temporal correlation feature set in the training sample set into the time feature extraction module and the spatial feature extraction module of the initial ecological risk prediction model respectively. The time feature extraction module processes the trend change features in the time dimension of the spatio-temporal correlation feature set according to its internal processing flow and outputs the training dynamic risk evolution features. The spatial feature extraction module processes the regional correlation features in the spatial dimension of the spatio-temporal correlation feature set and outputs the training spatial risk diffusion features. The processing processes of these two modules are the same as those described in the previous actual prediction process, except that the input data is the spatio-temporal correlation feature set in the training sample set.

[0165] Step S250: Input the training dynamic risk evolution features and the training spatial risk diffusion features into the feature interaction module, generate training interaction features through bidirectional attention interaction processing, and input the training interaction features into the prediction module to output a set of predicted risk conduction patterns.

[0166] Input the training dynamic risk evolution features and the training spatial risk diffusion features into the feature interaction module. The feature interaction module performs bidirectional attention interaction processing on these two features. The specific process is the same as that described in step S133, including operations such as generating a query vector set and a key vector set, calculating a regional spatio-temporal correlation attention weight set, performing parallel weighted aggregation processing and residual connection processing, and finally generating training interaction features. Then, input the training interaction features into the prediction module. The prediction module predicts the risk conduction pattern based on the training interaction features and outputs a set of predicted risk conduction patterns. This set of predicted risk conduction patterns contains information such as the starting node, propagation direction, and propagation intensity of the predicted risk propagation path.

[0167] Step S260: Calculate the loss function value based on the difference between the predicted risk conduction mode set and the true risk conduction mode set labeled in the risk annotation result.

[0168] Calculate the difference between the predicted risk conduction mode set and the true risk conduction mode set labeled in the risk annotation result, and use this to calculate the loss function value. The loss function is used to measure the error between the model prediction result and the true result. Commonly used loss functions include the mean squared error loss function, cross-entropy loss function, etc. Taking the mean squared error loss function as an example, assume that the predicted value of the i-th sample in the predicted risk conduction mode set is Pi, and the true value of the i-th sample in the true risk conduction mode set is Ti. Then the mean squared error loss function L = (1 / n) × Σ(Pi - Ti)^2, where n is the number of samples. By calculating the loss function value, the performance of the model under the current parameters can be evaluated. The smaller the loss function value, the closer the model's prediction result is to the true result.

[0169] Step S270: Synchronously update the convolution kernel parameters of the time feature extraction module, the graph node weight parameters of the spatial feature extraction module, and the attention weight parameters of the feature interaction module according to the loss function value through backpropagation gradients, and iteratively adjust until the loss function value converges to a preset threshold to generate the pre-trained ecological risk prediction model.

[0170] According to the calculated loss function value, update the parameters of the model. Using the backpropagation algorithm, calculate the gradients of the loss function with respect to the convolution kernel parameters of the time feature extraction module, the graph node weight parameters of the spatial feature extraction module, and the attention weight parameters of the feature interaction module. The gradient represents the rate of change of the loss function in the parameter space. Through the gradient, it can be known how to adjust the parameters to reduce the loss function value.

[0171] For the convolution kernel parameters of the time feature extraction module, update them according to the calculated gradients. The update method is to adjust the parameters along the opposite direction of the gradient because the opposite direction of the gradient is the direction in which the loss function value drops fastest. Assume the convolution kernel parameter is W, the learning rate is α, and the gradient is ∇L(W). Then the updated convolution kernel parameter W' = W - α × ∇L(W). The learning rate α controls the step size of parameter update. An overly large learning rate may cause excessive parameter update and inability to converge to the optimal value; an overly small learning rate will result in too slow convergence speed.

[0172] Similarly, for the graph node weight parameters of the spatial feature extraction module, also update them according to the gradients. Assume the graph node weight parameter is V, and the gradient is ∇L(V). Then the updated graph node weight parameter V' = V - α × ∇L(V).

[0173] For the attention weight parameters of the feature interaction module, update them in the same way. Let the attention weight parameter be U and the gradient be ∇L(U), then the updated attention weight parameter U' = U - α × ∇L(U).

[0174] When updating the parameters, synchronously update the convolution kernel parameters of the time feature extraction module, the graph node weight parameters of the spatial feature extraction module, and the attention weight parameters of the feature interaction module to ensure that the parameters of each module can be optimized collaboratively.

[0175] After completing one parameter update, apply the updated parameters to the model, and then input the spatio-temporal correlation feature set in the training sample set into the model again. After being processed by the time feature extraction module, the spatial feature extraction module, the feature interaction module, and the prediction module, output a new set of predicted risk conduction patterns. Then, calculate the loss function value between the new set of predicted risk conduction patterns and the set of true risk conduction patterns again.

[0176] Continuously repeat the above process of parameter update and loss function calculation for iterative adjustment. Each iteration will change the parameters of the model in the direction of reducing the loss function value. As the number of iterations increases, the loss function value will gradually decrease. When the loss function value converges to a preset threshold, it indicates that the performance of the model has reached a relatively ideal state. At this time, stop the iteration and generate a pre-trained ecological risk prediction model.

[0177] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of a mine ecological risk prediction system 100 based on artificial intelligence that can implement the idea of the present application. For example, the processor 120 can be used on the mine ecological risk prediction system 100 based on artificial intelligence and is used to execute the functions in the present application.

[0178] The mine ecological risk prediction system 100 based on artificial intelligence can be a general server or a special-purpose server, both of which can be used to implement the mine ecological risk prediction method based on artificial intelligence of the present application. Although only one server is shown in the present application, for convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0179] For example, the artificial intelligence-based mine ecological risk prediction system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, ROM, or RAM, or any combination thereof. Exemplarily, the artificial intelligence-based mine ecological risk prediction system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of the present application can be implemented according to these program instructions. The artificial intelligence-based mine ecological risk prediction system 100 further includes an I / O interface 150 between the computer and other input / output devices.

[0180] For ease of explanation, only one processor is described in the artificial intelligence-based mine ecological risk prediction system 100. However, it should be noted that the artificial intelligence-based mine ecological risk prediction system 100 in the present application may also include multiple processors. Therefore, the steps performed by one processor described in the present application may also be jointly performed or separately performed by multiple processors. For example, if the processor of the artificial intelligence-based mine ecological risk prediction system 100 performs step A and step B, it should be understood that step A and step B may also be jointly performed by two different processors or separately performed in one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0181] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned artificial intelligence-based mine ecological risk prediction method is implemented.

[0182] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the previous description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. A mine ecological risk prediction method based on artificial intelligence, characterized in that, The method includes: Collecting multi-source ecological monitoring data through a distributed sensor network deployed in a mining area, where the multi-source ecological monitoring data includes a time-continuous sequence of environmental indicators and spatially associated geographic coordinate information; Performing spatio-temporal alignment processing on the multi-source ecological monitoring data to generate a spatio-temporal association feature set, where the spatio-temporal association feature set includes trend change features in the time dimension and regional association features in the space dimension; Invoking a pre-trained ecological risk prediction model to perform ecological risk prediction on the spatio-temporal association feature set, generating a risk conduction mode set, where the risk conduction mode set includes the starting node, propagation direction, and propagation intensity of the risk propagation path; Constructing an ecological risk evolution network for the mining area based on the risk conduction mode set, where the ecological risk evolution network includes the topological structure of risk nodes and the temporal dependence relationship of risk diffusion; Generating a risk warning instruction set according to the ecological risk evolution network, where the risk warning instruction set includes the risk level classification result and the corresponding ecological restoration strategy, and sending the risk warning instruction set to the ecological management platform.

2. The method for predicting the ecological risk of mines based on artificial intelligence according to claim 1, wherein The performing spatio-temporal alignment processing on the multi-source ecological monitoring data to generate a spatio-temporal association feature set includes: Performing sliding window segmentation processing on the time series data in the multi-source ecological monitoring data to generate a set of data segments aligned in the time dimension, and each data segment corresponds to the change of environmental indicators within a preset time length; Mapping the set of data segments to a three-dimensional space coordinate system of the mining area to generate a set of monitoring data associated in the space dimension, and each data unit in the set of monitoring data includes a geographic coordinate identifier and the corresponding environmental indicator value; Performing outlier detection processing on the set of monitoring data to identify and remove abnormal data points beyond the preset fluctuation range, generating a cleaned spatio-temporal association data set; Extracting trend change features in the time dimension and regional association features in the space dimension from the cleaned spatio-temporal association data set, and merging them to generate a spatio-temporal association feature set.

3. The method for predicting the ecological risk of mines based on artificial intelligence according to claim 2, characterized in that The extracting trend change features in the time dimension and regional association features in the space dimension from the cleaned spatio-temporal association data set, and merging them to generate a spatio-temporal association feature set includes: Performing sliding window linear regression processing on the time-continuous environmental indicator sequence of each geographic coordinate identifier in the cleaned spatio-temporal association data set, calculating the fitting slope and residual variance of the environmental indicators within each sliding window, and generating trend change features in the time dimension; Performing Fourier spectrum analysis on the time-continuous environmental indicator sequence, extracting the amplitude peak value and phase offset within a preset frequency band as periodic fluctuation sub-features in the time dimension, and performing time alignment splicing on the periodic fluctuation sub-features and the trend change features to generate trend change features in the time dimension; Constructing a regional association graph according to the spatial distribution of the geographic coordinate information, calculating the environmental indicator difference degree and spatial autocorrelation coefficient between each node and its adjacent nodes in the regional association graph, and generating regional association features in the space dimension; After standardizing the trend change features and the regional association features, align them according to the geographical coordinate identifiers to generate a spatio-temporal association feature set.

4. The method for predicting the ecological risk of mines based on artificial intelligence according to claim 1, wherein, Call the pre-trained ecological risk prediction model to perform ecological risk prediction on the spatio-temporal association feature set, and generate a set of risk conduction patterns, including: Input the spatio-temporal association feature set into the time feature extraction module of the ecological risk prediction model, and deeply mine the trend change features in the time dimension through a multi-layer convolutional neural network to generate dynamic risk evolution features; Input the spatio-temporal association feature set into the spatial feature extraction module of the ecological risk prediction model, and model the node relationships of the regional association features in the spatial dimension through a graph neural network to generate spatial risk diffusion features; Perform bidirectional attention interaction processing on the dynamic risk evolution features and the spatial risk diffusion features through the feature interaction module of the ecological risk prediction model to generate attention interaction features; Input the attention interaction features into the prediction module of the ecological risk prediction model to generate a set of risk conduction patterns.

5. The method for predicting the ecological risk of mines based on artificial intelligence according to claim 4, wherein, The bidirectional attention interaction processing on the dynamic risk evolution features and the spatial risk diffusion features through the feature interaction module of the ecological risk prediction model to generate attention interaction features includes: Perform a linear transformation on the dynamic risk evolution features to generate a set of query vectors, and perform a linear transformation on the spatial risk diffusion features to generate a set of key vectors; Divide the dynamic risk evolution features and the spatial risk diffusion features into regions according to geographical coordinates, calculate the local similarity matrix of the query vector set and the key vector set within each region, and generate a set of spatio-temporal association attention weights for sub-regions; Perform parallel weighted aggregation processing on the spatial risk diffusion features according to the set of spatio-temporal association attention weights for sub-regions to generate spatial attention features; Perform residual connection processing on the spatial attention features and the dynamic risk evolution features to generate attention interaction features.

6. The method for predicting the ecological risk of mines based on artificial intelligence according to claim 1, wherein Construct an ecological risk evolution network for the mine area based on the set of risk conduction patterns, including: Extract the starting nodes and propagation directions of the risk propagation paths from the set of risk conduction patterns to construct an initial topological structure of the risk nodes; Calculate the connection weights between the risk nodes according to the propagation intensity in the set of risk conduction patterns to generate a set of weighted risk propagation edges; Perform iterative optimization processing on the initial topological structure based on the set of weighted risk propagation edges to generate an ecological risk evolution network containing multi-level risk diffusion paths; Model the time series dependence relationship in the ecological risk evolution network to generate a time series prediction result of risk diffusion.

7. The method for predicting the ecological risk of mines based on artificial intelligence according to claim 6, wherein, The iterative optimization processing on the initial topological structure based on the set of weighted risk propagation edges to generate an ecological risk evolution network containing multi-level risk diffusion paths includes: Perform clustering analysis on the risk nodes in the initial topological structure to generate a set of risk node communities, and each risk node community contains risk nodes that are spatially adjacent and have similar risk propagation intensities; Calculate the cross-community transmission probability between different risk node communities based on the weighted risk propagation edge set, and generate the inter-community transmission path; Add the inter-community transmission path to the initial topological structure to generate an extended risk topological structure; Perform path redundancy elimination processing on the extended risk topological structure to generate an optimized ecological risk evolution network; Model the temporal dependence relationship in the ecological risk evolution network to generate a time series prediction result of risk diffusion, including: Extract the historical risk intensity sequence of each risk node from the ecological risk evolution network to generate a risk node temporal feature set, and the historical risk intensity sequence includes risk propagation intensity values recorded at preset time intervals; Construct a temporal graph network based on the topological structure of the risk node and the historical risk intensity sequence. The node attributes in the temporal graph network include the temporal risk intensity characteristics of the corresponding risk node, and the edge weights are determined by the connection weights in the weighted risk propagation edge set; Input the temporal graph network into a pre-trained temporal graph neural network model, and perform causal convolution processing on the node attributes in the time dimension through a spatio-temporal convolutional layer to extract the risk intensity change pattern between adjacent time steps; Calculate the dynamic association weights between risk nodes at different time steps through the spatio-temporal attention layer of the temporal graph neural network model to generate a spatio-temporal attention weight matrix; Aggregate the node attributes in the temporal graph network across time steps according to the spatio-temporal attention weight matrix to generate a set of temporal hidden states for each risk node; Input the set of temporal hidden states into the multi-step prediction layer of the temporal graph neural network model, and decode step by step to generate the risk intensity prediction values of each risk node within a preset future time period; Generate a time series prediction result of risk diffusion for the ecological risk evolution network according to the risk intensity prediction value and the spatial position of the risk nodes in the topological structure.

8. The method for predicting the ecological risk of mines based on artificial intelligence according to claim 1, wherein, The method for generating a risk warning instruction set according to the ecological risk evolution network includes: Perform a priority evaluation process on the risk nodes in the ecological risk evolution network to generate a risk level distribution map; Identify the spatial coordinate ranges of high-risk areas and potential risk areas from the risk level distribution map; Match the spatial coordinate range of the high-risk area with a preset emergency response strategy library to generate a high-risk area warning instruction; Extract the environmental index characteristics and spatial distribution characteristics of the potential risk area to generate a potential risk feature vector; Retrieve a set of sample cases matching the potential risk feature vector from the sample repair case database; Perform an effect evaluation process on the repair schemes in the set of sample cases to generate a scheme priority ranking result; Select a target repair scheme according to the scheme priority ranking result and generate a risk prevention and control recommendation instruction.

9. The method for predicting the ecological risk of mines based on artificial intelligence according to claim 1, wherein, The training method of the pre-trained ecological risk prediction model includes: Obtain a sample mine ecological data set, which includes multiple groups of spatio-temporally correlated sample multi-source ecological monitoring data and corresponding risk annotation results; Perform data augmentation processing on the multi-source ecological monitoring data of the sample, and merge the augmented multi-source ecological monitoring data and the multi-source ecological monitoring data of the sample, and then extract a standardized set of spatio-temporal correlation features to generate a training sample set. Among them, the data augmentation processing includes time series interpolation, noise injection, and restricted spatial coordinate perturbation, where the maximum offset distance of the spatial coordinate perturbation does not exceed a preset ratio of the adjacent sensor spacing; Construct an initial ecological risk prediction model, which includes a time feature extraction module, a spatial feature extraction module, a feature interaction module, and a prediction module; Input the set of spatio-temporal correlation features in the training sample set into the time feature extraction module and the spatial feature extraction module of the initial ecological risk prediction model, and output the training dynamic risk evolution features and the training spatial risk diffusion features respectively; Input the training dynamic risk evolution features and the training spatial risk diffusion features into the feature interaction module, generate training interaction features through bidirectional attention interaction processing, and input the training interaction features into the prediction module to output a set of predicted risk conduction patterns; Calculate the loss function value based on the difference between the set of predicted risk conduction patterns and the set of true risk conduction patterns marked in the risk annotation result; Perform synchronous backpropagation gradient update on the convolution kernel parameters of the time feature extraction module, the graph node weight parameters of the spatial feature extraction module, and the attention weight parameters of the feature interaction module according to the loss function value, and iteratively adjust until the loss function value converges to a preset threshold to generate the pre-trained ecological risk prediction model.

10. An artificial intelligence-based mine ecological risk prediction system, characterized in that, It includes a processor and a memory. The memory is connected to the processor. The memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the artificial intelligence-based mine ecological risk prediction method according to any one of claims 1-9 above.

Images