A risk early warning method and system for mine ecological restoration areas based on environmental perception

Abstract

This application provides a risk early warning method and system for mine ecological restoration areas based on environmental perception. It collects multi-source environmental datasets through a constructed four-level environmental perception system (space-air-ground-depth). Spatial spectral features and temporal dependence features are extracted from the multi-source environmental datasets, and cross-modal feature fusion is performed to obtain a fused environmental feature tensor. The environmental feature tensor is decomposed and dimensionality reduced, and multi-task temporal prediction is performed based on the dimensionality-reduced feature sequence to obtain trend prediction values ​​for four risk indices in the mine ecological restoration area: geological hazards, vegetation degradation, soil and water pollution, and degradation of restoration effectiveness. Based on the trend prediction values ​​of each risk index, the risk early warning level is determined and corresponding response strategies are triggered. Based on the above scheme, multi-dimensional environmental perception and cross-modal feature fusion can achieve multi-task risk trend prediction to complete accurate early warning level determination and differentiated response.

Description

Technical Field

[0001] This application relates to the field of ecological and environmental risk early warning technology, and more specifically, to a risk early warning method and system for mine ecological restoration areas based on environmental perception. Background Technology

[0002] Mine ecological restoration areas are the core areas for ecological reconstruction and geological safety control in mining areas. Affected by multiple factors such as underground rock disturbance, changes in water and soil environment, and natural vegetation succession, they are prone to frequent geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness. There is an urgent need for intelligent risk early warning technology adapted to the characteristics of mine restoration areas to support the long-term management and maintenance of restoration projects and regional ecological safety control.

[0003] Existing risk early warning technologies for mine ecological restoration areas mostly rely on single-modal feature analysis based on a single sensing data source, determining the risk status at the current moment by setting fixed thresholds. However, these methods lack a comprehensive sensing system that integrates satellite remote sensing, UAV inspections, ground sensor networks, and deep microseismic monitoring. This prevents the cross-modal, bidirectional, mutually reinforcing fusion of spatial spectral features and time-dependent features, and makes it difficult to simultaneously predict multiple risks such as geological disasters, vegetation degradation, soil and water pollution, and the degradation of restoration effectiveness. These technologies suffer from incomplete sensing dimension coverage, fragmented feature information representation, a lack of ability to predict future risk evolution trends, and a lack of multi-indicator fusion mechanisms for early warning level classification. Therefore, how to achieve multi-task risk trend prediction through multi-dimensional environmental perception and cross-modal feature fusion to complete accurate early warning level determination and differentiated responses has become a challenge for the industry. Summary of the Invention

[0004] This application provides a risk early warning method and system for mine ecological restoration areas based on environmental perception. It can realize multi-task risk trend prediction through multi-dimensional environmental perception and cross-modal feature fusion to complete accurate early warning level determination and differentiated response.

[0005] Firstly, this application provides a risk early warning method for mine ecological restoration areas based on environmental perception, comprising the following steps:

[0006] By constructing a four-level environmental perception system (air-space-ground-depth), multi-dimensional environmental perception data collection and spatiotemporal benchmark registration are carried out in the mine ecological restoration area to obtain a multi-source environmental dataset.

[0007] Spatial spectral features of remote sensing images and UAV images in the multi-source environmental dataset are extracted. Temporal dependence features of sensing time series data and microseismic signals are extracted from the multi-source environmental dataset. Then, cross-modal feature fusion is performed on the spatial spectral features and the temporal dependence features to obtain the fused environmental feature tensor.

[0008] The environmental feature tensor is decomposed and dimensionality reduced, and multi-task time series prediction is performed based on the dimensionality-reduced feature sequence to obtain the trend prediction values ​​of the mine ecological restoration area on four types of risk indices: geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness.

[0009] The trend forecast values ​​of each risk index are compared with the preset four-level early warning threshold system to determine the risk warning level and trigger the corresponding response strategy.

[0010] In some embodiments, a multi-dimensional environmental perception data collection and spatiotemporal benchmark registration are performed on the mine ecological restoration area through a constructed four-level environmental perception system (space-air-ground-depth), resulting in a multi-source environmental dataset, specifically including:

[0011] Remote sensing images of the target mine ecological restoration area were obtained through a satellite remote sensing data interface.

[0012] The drone aerial photography transmission module is used to acquire drone images covering the target mine ecological restoration area.

[0013] By deploying a sensor array within the target mine ecological restoration area, soil physicochemical parameters, groundwater level and water quality indicators, and three-dimensional surface displacement were collected to obtain sensing time-series data.

[0014] Microseismic signals of rock strata within the target mine ecological restoration area are collected through a network of deployed deep microseismic sensors.

[0015] The remote sensing images, the UAV images, the sensor time series data, and the microseismic signals are converted to the same geographic coordinate system and unified with a time reference to obtain a multi-source environmental dataset with spatiotemporal reference registration.

[0016] In some embodiments, extracting the spatial spectral features of remote sensing images and UAV images from the multi-source environmental dataset specifically includes:

[0017] Multi-scale spatial feature extraction is performed on the remote sensing images and UAV images in the multi-source environmental dataset to obtain the initial feature map of the surface of the target mine ecological restoration area;

[0018] Eliminate the radiation distortion of spectral information caused by terrain undulation in the initial feature map to obtain a terrain-corrected spectral feature map;

[0019] Geometric deformation modeling is performed on the spectral feature map to obtain spatial spectral features.

[0020] In some embodiments, extracting the time-series dependency features between sensing time-series data and microseismic signals from the multi-source environmental dataset specifically includes:

[0021] The sensor time series data in the multi-source environmental dataset are spliced ​​together according to the sampling timestamp to obtain the ground sensor time series observation stream;

[0022] The microseismic signals in the multi-source environmental dataset are subjected to event-triggered interception and time-out picking to obtain a microseismic event waveform sequence;

[0023] Multi-scale temporal feature extraction is performed by temporal convolution on the sensor time-series observation stream and the microseismic event waveform sequence to obtain sensor time-series features and microseismic time-series features;

[0024] The sensing time-series features and the microseismic time-series features are aggregated to obtain time-dependent features.

[0025] In some embodiments, the cross-modal feature fusion of the spatial spectral features and the temporal dependency features to obtain the fused environmental feature tensor specifically includes:

[0026] The spatial spectral features are flattened along the channel dimension and linearly projected to obtain the spatial spectral query vector and the spatial spectral key vector.

[0027] The time-dependent features are flattened and linearly projected along the time dimension to obtain the time-series query vector and the time-series key-value vector.

[0028] The spatial spectrum query vector and the temporal key value vector are subjected to cross-modal dot product attention calculation to obtain the attention weight matrix of spatial spectrum to temporal.

[0029] The temporal query vector and the spatial spectral key vector are subjected to cross-modal dot product attention calculation to obtain the attention weight matrix of temporal to spatial spectrum;

[0030] The spatial spectral features for time enhancement are determined based on the attention weight matrix of spatial spectra on time and the time-dependent features.

[0031] The temporal dependence features of spatial spectral enhancement are determined based on the attention weight matrix of the spatial spectrum on the time sequence and the spatial spectral features.

[0032] The temporally enhanced spatial spectral features and the temporally dependent features of the spatial spectral enhancement are concatenated and then subjected to cross-modal information compression and nonlinear mapping to obtain the fused environmental feature tensor.

[0033] In some embodiments, the environmental feature tensor is decomposed and dimensionality reduced, and multi-task time-series prediction is performed based on the dimensionality-reduced feature sequence to obtain trend prediction values ​​for four risk indices of the mine ecological restoration area: geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness. Specifically, these include:

[0034] The temporal components of each dimension of the environmental feature tensor are sequentially subjected to variational mode decomposition to obtain multiple intrinsic mode components.

[0035] Nonlinear dimensionality reduction is performed on all intrinsic mode components to form a dimensionality-reduced feature sequence;

[0036] The reduced feature sequence is input into an attention-enhanced long short-term memory network for multi-task temporal prediction, and trend prediction values ​​of four risk indices of mine ecological restoration areas, namely geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness, are obtained.

[0037] In some embodiments, comparing the trend forecast values ​​of each risk index with a preset four-level early warning threshold system to determine the risk warning level and trigger the corresponding response strategy specifically includes:

[0038] For the trend prediction value of each risk index, the trend prediction value is compared with the warning threshold of the corresponding risk dimension in the preset four-level warning threshold system index by index to obtain the warning level of the risk index, and then the warning level of each risk index is obtained.

[0039] The current risk warning level of the target mine ecological restoration area is determined based on the warning levels of all risk indices;

[0040] Based on the risk warning level, the corresponding response measures are retrieved from the preset differentiated warning response strategy library and triggered for execution.

[0041] Secondly, this application provides a risk early warning system for mine ecological restoration areas based on environmental perception, including:

[0042] The data acquisition module is used to collect multi-dimensional environmental perception data and perform spatiotemporal benchmark registration in the mine ecological restoration area through the constructed four-level environmental perception system of air-space-ground-depth, so as to obtain multi-source environmental datasets.

[0043] The processing module is used to extract the spatial spectral features of remote sensing images and UAV images in the multi-source environmental dataset, extract the temporal dependence features of sensing time series data and microseismic signals in the multi-source environmental dataset, and then perform cross-modal feature fusion on the spatial spectral features and the temporal dependence features to obtain the fused environmental feature tensor.

[0044] The processing module is also used to decompose and reduce the dimensionality of the environmental feature tensor, and perform multi-task time series prediction based on the dimensionality-reduced feature sequence to obtain the trend prediction values ​​of the mine ecological restoration area on four types of risk indices: geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness.

[0045] The execution module is used to compare the trend prediction values ​​of each risk index with the preset four-level early warning threshold system to determine the risk warning level and trigger the corresponding response strategy.

[0046] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described method for risk early warning of mine ecological restoration areas based on environmental perception.

[0047] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described risk warning method for mine ecological restoration areas based on environmental perception.

[0048] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0049] This application provides a risk early warning method and system for mine ecological restoration areas based on environmental perception. The system utilizes a constructed four-level environmental perception system (space-air-ground-depth) to collect multi-dimensional environmental perception data and perform spatiotemporal benchmark registration in the mine ecological restoration area, obtaining a multi-source environmental dataset. Spatial spectral features of remote sensing images and UAV images are extracted from the multi-source environmental dataset. Temporal dependence features of sensor time-series data and microseismic signals are extracted from the multi-source environmental dataset. Cross-modal feature fusion is then performed on the spatial spectral features and the temporal dependence features to obtain a fused environmental feature tensor. The environmental feature tensor is decomposed and dimensionality reduced, and multi-task temporal prediction is performed based on the dimensionality-reduced feature sequence to obtain trend prediction values ​​for four risk indices: geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness in the mine ecological restoration area. The trend prediction values ​​of each risk index are compared with a preset four-level early warning threshold system to determine the risk warning level and trigger corresponding response strategies.

[0050] Therefore, in this application, the trend prediction values ​​of each risk index are compared with the preset four-level early warning threshold system to determine the risk warning level and trigger the corresponding response strategy. First, by determining the multi-source environmental dataset, a multi-dimensional environmental perception data set for the mine ecological restoration area can be obtained, which provides a complete spatiotemporal aligned data foundation for cross-modal feature extraction and fusion. This multi-source environmental dataset combines the regional-scale surface deformation and vegetation cover macro-spatial distribution provided by satellite remote sensing images, the fine-scale surface micro-topography and abnormal vegetation patches provided by UAV aerial photography, and the real-time continuous soil data provided by ground IoT sensors. This study integrates the spatiotemporal changes of chemical parameters and hydrological indicators, as well as the precursory information of rock microfracture activity provided by deep microseismic signals. This breaks down the spatiotemporal barriers and modal barriers of multi-dimensional sensing information across air, space, ground, and depth at the data source. This allows subsequent feature extraction steps to simultaneously mine surface spatial spectral features and subsurface temporal dependence features based on a unified spatiotemporal benchmark of multi-dimensional data. It provides a complete, aligned, and jointly analyzable multi-source heterogeneous data foundation for cross-modal bidirectional mutual enhancement fusion, thereby fundamentally overcoming the inherent limitations of incomplete information representation from single data sources and the incompatibility between spatial coverage and temporal continuity. Subsequently, by determining the environmental feature tensor, a cross-modal fusion feature tensor that uniformly represents the multi-source heterogeneous environmental sensing data of the mine ecological restoration area can be obtained. This environmental feature tensor integrates the spatial distribution information of surface ecological damage reflected by remote sensing images and UAV aerial photography with the temporal evolution patterns of water and soil environment and rock strata activity revealed by ground IoT sensors and deep microseismic signals. At the feature level, cross-modal semantic alignment and adaptive weight allocation are performed, enabling spatial modes to perceive the dynamic evolution patterns of their temporal correlations, and temporal modes to locate their corresponding spatial ecological damage areas. This realizes the integration of air-space-ground-deep multi-source sensing data in the special... The bidirectional mutual enhancement fusion of the feature layer effectively solves the technical defects of fragmented feature representation of multi-source environmental perception data in the existing technology; thus, it provides a unified feature representation with rich semantics and full interaction between modes for subsequent variational mode decomposition, so that the decomposed intrinsic mode components can carry both surface spatial structure information and underground temporal activity information. This enables the attention-enhanced long short-term memory network to make full use of ground-deep correlation features to model risk evolution trends in the process of multi-task temporal prediction, which significantly improves the accuracy and forward-looking nature of trend prediction for four types of risk indices: geological disasters, vegetation degradation, water and soil pollution, and degradation of remediation effectiveness.Finally, by determining the trend prediction value, a continuous numerical sequence reflecting the dynamic evolution direction of various risk indices in the target mine ecological restoration area within a preset future time period can be obtained. Compared with the static early warning mode of existing technologies that rely solely on a single threshold at the current moment to determine the risk status, the introduction of trend prediction values ​​upgrades the early warning mechanism from post-event alarm to pre-event prediction. This allows the early warning system to detect the inflection point of accelerated deterioration of risks before they reach the critical threshold, providing a quantitative decision-making basis for the early deployment of tiered early warning response strategies and the provision of preparation time windows for emergency response measures. This effectively compensates for the technical deficiency of existing technologies in predicting the evolution trend of risks in future time periods. In summary, based on the above scheme, multi-task risk trend prediction can be achieved through multi-dimensional environmental perception and cross-modal feature fusion to complete accurate early warning level determination and differentiated response. Attached Figure Description

[0051] Figure 1 This is an exemplary flowchart of a risk early warning method for mine ecological restoration areas based on environmental perception, according to some embodiments of this application;

[0052] Figure 2 This is an exemplary flowchart illustrating the determination of spatial spectral features according to some embodiments of this application;

[0053] Figure 3 This is a flowchart illustrating the operation of determining trend prediction values ​​according to some embodiments of this application;

[0054] Figure 4 This is a structural schematic diagram of a risk early warning system unit for mine ecological restoration areas based on environmental perception, as shown in some embodiments of this application;

[0055] Figure 5 This is a schematic diagram of the structure of a computer device for implementing a risk early warning method for mine ecological restoration areas based on environmental perception, according to some embodiments of this application. Detailed Implementation

[0056] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] refer to Figure 1 This figure is an exemplary flowchart of a risk early warning method for mine ecological restoration areas based on environmental perception, according to some embodiments of this application. The figure mainly includes the following steps:

[0058] In step 101, a multi-dimensional environmental perception data collection and spatiotemporal benchmark registration are carried out on the mine ecological restoration area through the constructed four-level environmental perception system of air-space-ground-depth to obtain a multi-source environmental dataset.

[0059] In some embodiments, the following steps can be taken to collect multi-dimensional environmental perception data and register spatiotemporal references in the mine ecological restoration area through the constructed four-level environmental perception system (space-air-ground-depth) to obtain a multi-source environmental dataset:

[0060] Remote sensing images of the target mine ecological restoration area were obtained through a satellite remote sensing data interface.

[0061] The drone aerial photography transmission module is used to acquire drone images covering the target mine ecological restoration area.

[0062] By deploying a sensor array within the target mine ecological restoration area, soil physicochemical parameters, groundwater level and water quality indicators, and three-dimensional surface displacement were collected to obtain sensing time-series data.

[0063] Microseismic signals of rock strata within the target mine ecological restoration area are collected through a network of deployed deep microseismic sensors.

[0064] The remote sensing images, the UAV images, the sensor time series data, and the microseismic signals are converted to the same geographic coordinate system and unified with a time reference to obtain a multi-source environmental dataset with spatiotemporal reference registration.

[0065] It should be noted that in this application, remote sensing imagery refers to periodic macroscopic observation data covering the target mine ecological restoration area. This data can be used to extract the spatiotemporal distribution characteristics of large-scale surface deformation rates and vegetation cover, providing the scheme with macroscopic perception capabilities of regional-scale geological hazard risks and ecological degradation trends. Unmanned aerial vehicle (UAV) imagery refers to high-resolution orthophotos of the target mine ecological restoration area, which can be used to identify local micro-topographic changes and patches of abnormal vegetation growth, providing the scheme with fine-scale surface condition verification and risk target area locking capabilities. Sensor time-series data is used to characterize the target... The multivariate time series of the continuous dynamic evolution of water and soil environmental parameters in the target mine ecological restoration area can provide the scheme with real-time and continuous monitoring capabilities for surface environmental changes; the microseismic signal is a waveform sequence characterizing the micro-fracture activity of the rock strata above the target mine ecological restoration area, which can be used to locate the spatial distribution and energy release rate of rock mass fracture events, providing the scheme with deep perception of the precursors of underground rock strata instability and early identification capabilities for geological hazards; the multi-source environmental dataset is a multi-dimensional environmental perception data set of the mine ecological restoration area that provides a complete spatiotemporal aligned data foundation for cross-modal feature extraction and fusion.

[0066] In specific implementation, firstly, Sentinel-2 multispectral images covering the target mine ecological restoration area can be acquired via a satellite remote sensing data interface as remote sensing images; secondly, a multispectral imager mounted on an UAV aerial photography transmission module can be used to conduct gridded aerial photography of the target mine ecological restoration area at a preset flight altitude and heading overlap, and the motion recovery structure algorithm is used to perform 3D reconstruction and orthorectification of the aerial images to obtain UAV images covering the target mine ecological restoration area; then, a multi-parameter soil sensor array consisting of soil pH sensors, soil heavy metal ion selective electrodes, and soil moisture sensors, a hydrological monitoring sensor array consisting of water level gauges and multi-parameter water quality analyzers, and a global navigation satellite system can be deployed at key monitoring points within the target mine ecological restoration area. A surface displacement monitoring station, composed of a satellite system displacement monitoring station and a crack displacement meter, continuously collects and processes outlier removal data on soil physicochemical parameters, groundwater level and water quality indicators, and three-dimensional surface displacement at a sampling frequency of once per hour to obtain sensor time-series data. Next, a three-component microseismic sensor array, deployed in a grid above known goaf areas within the target mine ecological restoration area, continuously collects rock micro-fracture signals at a sampling rate of once per hour. The raw microseismic signals are then processed by bandpass filtering for noise reduction and time-based picking to obtain microseismic signals. Finally, remote sensing images, UAV images, sensor time-series data, and microseismic signals are uniformly converted to the 2000 National Geodetic Coordinate System and time-aligned and spatially registered based on Coordinated Universal Time timestamps to obtain a multi-source environmental dataset with spatiotemporal reference registration.

[0067] In step 102, the spatial spectral features of remote sensing images and UAV images in the multi-source environmental dataset are extracted, and the temporal dependence features of sensing time series data and microseismic signals are extracted in the multi-source environmental dataset. Then, the spatial spectral features and the temporal dependence features are fused across modes to obtain the fused environmental feature tensor.

[0068] In some embodiments, reference Figure 2 The figure is an exemplary flowchart illustrating the determination of spatial spectral features according to some embodiments of this application. The extraction of spatial spectral features from remote sensing images and UAV images in the multi-source environmental dataset can be achieved by the following steps:

[0069] In step 1021, multi-scale spatial feature extraction is performed on the remote sensing images and UAV images in the multi-source environmental dataset to obtain an initial feature map of the surface of the target mine ecological restoration area;

[0070] In step 1022, the radiation distortion caused by terrain undulation to the spectral information in the initial feature map is eliminated to obtain the terrain-corrected spectral feature map;

[0071] In step 1023, the spectral feature map is geometrically deformed to obtain spatial spectral features.

[0072] It should be noted that in this application, the initial feature map is a multi-scale initial feature expression that includes semantic information of land cover type and preliminary characterization of ecological damage. It can integrate multi-level land cover information from fine-scale mining pit cracks to coarse-scale vegetation and water bodies, overcoming the shortcomings of single-scale feature extraction in taking into account both the global geomorphological background and local ecological damage details. The spectral feature map is a spectral feature expression that reflects the true reflectance characteristics of land cover. It can eliminate the interference of radiation distortion such as shadow shading and overexposure of sunny slopes caused by complex mine terrain undulations, so that spectral information can objectively characterize ecological restoration status indicators such as vegetation coverage and soil exposure, avoiding false alarms or missed alarms caused by terrain noise. The spatial spectral feature refers to the spatial spectral feature representation that integrates geometric deformation adaptability and spectral resolution. It can adaptively capture the irregular linear distribution of ground fissures, the concave contour of mining pits, and the conical accumulation shape of slag heaps, which are unique complex geometric structures of mining areas, significantly improving the accuracy of subsequent risk index prediction in identifying the spatial morphology of geological disaster hazards.

[0073] In practical implementation, firstly, object-oriented multi-scale segmentation technology can be used to segment remote sensing images and UAV images. By setting multi-level segmentation scale parameters, image object layers of different granularities are generated. Fine-scale segmentation layers are used to identify small-area ecologically damaged features such as mining pits, slag heaps, and ground fissures, while coarse-scale segmentation layers are used to extract large-area land cover types such as vegetation restoration areas, water bodies, and exposed surfaces. Then, the spectral mean, spectral standard deviation, shape index, and texture entropy value of each image object layer are statistically analyzed as multi-dimensional feature descriptors. All multi-dimensional feature descriptors are then input into a pre-trained deep residual network for layer-by-layer convolution and pooling. The process involves a step-by-step extraction of high-level abstract features, from edge contours to semantic categories. Simultaneously, a feature pyramid network is used to enhance the feature maps output from each layer of the deep residual network from top to bottom, and to perform lateral connections. This fuses shallow detail features with deep semantic features, generating an initial feature map with multi-scale representation capabilities. Then, digital elevation model (DEM) data with coverage consistent with remote sensing and UAV imagery is acquired. Based on the solar zenith angle, solar azimuth angle, and slope and aspect information provided by the DEM, a cosine correction model is used to calculate the modulation coefficient of the terrain on incident solar radiation pixel by pixel. Finally, the pixel values ​​of each band in the initial feature map are adjusted... The system incorporates an illumination compensation factor and utilizes calculated modulation coefficients to enhance the radiation of shadowed pixels and suppress the radiation of overly bright pixels on sunny slopes. For example, when the angle between the slope orientation of a pixel and the direction of solar incidence exceeds a preset threshold, the pixel is classified as a shadowed region and compensated based on the average spectral value of its neighboring non-shadowed pixels. Then, the corrected pixel values ​​are normalized twice using the atmospheric radiative transfer equation to eliminate residual atmospheric path radiation interference, resulting in a topographically corrected spectral feature map. Finally, a deformable convolution operation can be applied to the spectral feature map, i.e., a modulated convolution operation is applied to each sampling point on the regular sampling grid of the standard convolution kernel. A set of learnable two-dimensional spatial offsets is input, enabling the convolutional kernel to adaptively adjust the shape and position of the receptive field according to the content of the input feature map. The two-dimensional spatial offsets and network weight parameters are jointly learned through the backpropagation algorithm, so that the deformable convolutional module automatically focuses on complex geometric structures such as the irregular linear distribution of ground fissures in the mining area, the concave contour of the mining pit, and the conical accumulation shape of the slag heap during the training process. Then, the geometrically corrected features output by the deformable convolution are concatenated with the original spectral features channel by channel, and cross-channel information interaction and dimensionality compression are performed through point convolution to obtain spatial spectral features that integrate geometric deformation adaptability and spectral resolution.

[0074] In some embodiments, extracting the time-series dependency features between sensing time-series data and microseismic signals from the multi-source environmental dataset can be achieved using the following steps:

[0075] The sensor time series data in the multi-source environmental dataset are spliced ​​together according to the sampling timestamp to obtain the ground sensor time series observation stream;

[0076] The microseismic signals in the multi-source environmental dataset are subjected to event-triggered interception and time-out picking to obtain a microseismic event waveform sequence;

[0077] Multi-scale temporal feature extraction is performed by temporal convolution on the sensor time-series observation stream and the microseismic event waveform sequence to obtain sensor time-series features and microseismic time-series features;

[0078] The sensing time-series features and the microseismic time-series features are aggregated to obtain time-dependent features.

[0079] It should be noted that in this application, the sensor time-series observation stream is a continuous time-series data formed by aligning and stitching multiple environmental variables along a unified time axis. It can integrate surface sensing data with different physical meanings into a time-consistent and dimensionally unified observation sequence, solving the problem of fragmentation and asynchrony of multi-source ground sensor data in the time dimension, which makes joint analysis difficult. The microseismic event waveform sequence is a waveform dataset composed of discrete microseismic event waveforms extracted from microseismic signals and arranged in chronological order. It can accurately extract effective microseismic signal segments related to rock strata fracturing and mining subsidence activation from massive continuous background noise, transforming continuous monitoring data into discrete event sequences that can be feature-learned, significantly reducing the data redundancy and computational overhead of subsequent time-series modeling. The sensor time-series features are multi-scale surface environmental feature representations that simultaneously include short-term environmental fluctuation features and long-term trend change features. They can capture instantaneous soil physicochemical parameters and hydrological parameters in parallel. Abnormal fluctuations and seasonal gradual trends provide multi-temporal environmental change information for accurate prediction of vegetation degradation risk index and soil and water pollution risk index; microseismic time series features are multi-scale underground seismic feature representations that simultaneously include short-term pulse features, medium-term envelope features, and long-term series energy evolution features. They can simultaneously characterize the instantaneous rupture characteristics of a single microseismic event and the energy accumulation trend of microseismic activity on the time axis, providing sensitive deep sensing information for the geological disaster risk index to predict the instability of mining subsidence areas and the expansion of rock fissures; time-dependent features are cross-modal time series feature representations that integrate the complementary correlation characteristics of surface environmental modes and underground deep sensing modes. They can adaptively suppress redundant information between the two modes through gating mechanisms and enhance the causal correlation characteristics between surface environmental changes and underground microseismic activity, enabling subsequent cross-modal feature fusion steps to carry out higher-level semantic fusion based on fully exploring the ground-deep correlation relationship.

[0080] In practical implementation, firstly, using Coordinated Universal Time (UTC) as the benchmark, the pH, heavy metal ion concentration, and moisture content data collected by multi-parameter soil IoT sensors from multi-source environmental datasets are time-aligned to the hour. Similarly, the groundwater level, turbidity, and conductivity data collected by hydrological monitoring sensors are time-aligned to the hour. Finally, the three-dimensional displacement component data collected by global navigation satellite system displacement monitoring stations are time-aligned to the minute. Next, a weighted moving average-based missing value imputation technique is used to make the aligned multi-source time-series data continuous. For segments with continuous missing values ​​exceeding a preset duration threshold, cubic spline interpolation is used for smooth completion. For example, when the soil moisture sensor detects a missing value at a certain time... When a data breakpoint occurs due to communication interruption within a segment, interpolation estimation is performed using the mean change trends of ten effective sampling points before and after the breakpoint. The completed pH, heavy metal ion concentration, water content, groundwater level, turbidity, conductivity, and three-dimensional displacement component sequences are then stitched together row by row according to timestamps to form a multivariate matrix. Each row corresponds to the values ​​of all ground sensor variables at a unified sampling time, forming a continuous sensor time-series observation stream. Secondly, the microseismic signal is divided into waveform segments to be detected according to fixed time windows. The signal-to-noise ratio (SNR) of each waveform segment is calculated, and waveform segments with an SNR exceeding a preset trigger threshold are identified as suspected microseismic events. The SNR calculation uses the energy ratio criterion between long and short time windows, with the short time window length determined by... Based on the main frequency range of typical microseismic events in the target mine ecological restoration area, for example, the short time window length is taken as the number of sampling points corresponding to half of the main period of the microseismic signal, and the long time window length is taken as five times the short time window length. Then, for the identified suspected microseismic events, the first arrival and arrival times of the waveform are automatically picked up using the Akaike information criterion. The three-channel waveform data, with the first arrival and arrival times as the center, are truncated forward and backward for a preset number of seconds to form a complete microseismic event waveform sample. Then, all microseismic event waveform samples are arranged into a microseismic event waveform sequence according to the chronological order of event triggering. Each microseismic event waveform sample is accompanied by its time position label and multi-channel waveform amplitude information. Finally, the sensor time series observation stream is input into a multi-layer temporal convolutional network. In this study, causal convolutional layers with different dilation factors are used to progressively expand the temporal receptive field. Shallow branches with small dilation factors are used to capture short-term environmental fluctuation features, while deep branches with large dilation factors are used to extract long-term trend change features. The outputs of each branch are then normalized, concatenated channel by channel, and transformed through a fully connected mapping layer to form a sensing temporal feature with multiple time scales. Simultaneously, the waveform sequence of microseismic events is input into a multi-layer temporal convolutional network with the same structure. Causal convolutional layers with different dilation factors are used to extract short-term pulse features, mid-term envelope features, and long-term sequence energy evolution features of the microseismic waveforms. After normalization and channel-by-channel concatenation, the microseismic waveform is transformed through a fully connected mapping layer to form a microseismic temporal feature with multiple time scales.Finally, the sensing time-series features and the microseismic time-series features are concatenated along the feature dimension to form a composite time-series feature vector. This composite time-series feature vector is then input into a gated linear unit for feature interaction and information filtering. The gated linear unit uses a gating mechanism to selectively retain or suppress information in each dimension of the composite time-series feature vector, eliminating redundant information between the sensing modes and the microseismic modes and enhancing their complementary correlation features. The feature vector output by the gated linear unit is then used as the time-dependent feature vector.

[0081] In some embodiments, the cross-modal feature fusion of the spatial spectral features and the temporal dependency features to obtain the fused environmental feature tensor can be achieved by the following steps:

[0082] The spatial spectral features are flattened along the channel dimension and linearly projected to obtain the spatial spectral query vector and the spatial spectral key vector.

[0083] The time-dependent features are flattened and linearly projected along the time dimension to obtain the time-series query vector and the time-series key-value vector.

[0084] The spatial spectrum query vector and the temporal key value vector are subjected to cross-modal dot product attention calculation to obtain the attention weight matrix of spatial spectrum to temporal.

[0085] The temporal query vector and the spatial spectral key vector are subjected to cross-modal dot product attention calculation to obtain the attention weight matrix of temporal to spatial spectrum;

[0086] The spatial spectral features for time enhancement are determined based on the attention weight matrix of spatial spectra on time and the time-dependent features.

[0087] The temporal dependence features of spatial spectral enhancement are determined based on the attention weight matrix of the spatial spectrum on the time sequence and the spatial spectral features.

[0088] The temporally enhanced spatial spectral features and the temporally dependent features of the spatial spectral enhancement are concatenated and then subjected to cross-modal information compression and nonlinear mapping to obtain the fused environmental feature tensor.

[0089] It should be noted that, in this application, the spatial spectral query vector is a query representation vector used to actively retrieve key time patterns related to spatial spectral information in time-dependent features; the spatial spectral key value vector is an index representation vector of spatial spectral information matched by time-dependent feature retrieval; the time-series query vector is a query representation vector used to actively retrieve key spatial locations related to time-series changes in spatial spectral features; the time-series key value vector is an index representation vector of time-series information matched by spatial spectral feature retrieval; the spatial spectral attention weight matrix for time series is a cross-modal correlation strength matrix characterizing the degree of dependence of different spatial locations on each time step in the time series; and the time series attention weight matrix for spatial spectral is a cross-modal correlation strength matrix characterizing the degree of attention of different time steps to each location in the spatial spectrum.

[0090] In specific implementation, firstly, a channel-dimensional flattening operation is performed on the spatial spectral features, transforming the three-dimensional feature map into a two-dimensional feature matrix. This two-dimensional feature matrix is ​​then linearly mapped through a first fully connected linear projection layer to obtain the spatial spectral query vector and the spatial spectral key-value vector. Simultaneously, a time-dimensional flattening operation is performed on the temporal-dependent features, expanding the temporal feature sequence into a two-dimensional feature matrix. This two-dimensional feature matrix is ​​then linearly mapped through a second fully connected linear projection layer to obtain the temporal query vector and the temporal key-value vector. Secondly, cross-modal dot-product attention is calculated on the spatial spectral query vector and the temporal key-value vector. This involves matrix multiplication of the transposes of the spatial spectral query vector and the temporal key-value vector, followed by scaling and normalization exponential function activation to obtain the attention weight matrix of the spatial spectral features on the temporal-dependent features. Simultaneously, cross-modal dot-product attention is also performed on the temporal query vector and the spatial spectral key-value vector. The calculation involves matrix multiplication of the transpose of the temporal query vector and the spatial spectral key vector, followed by scaling and normalization exponential function activation to obtain the attention weight matrix of temporal dependent features on spatial spectral features. Then, based on the attention weight matrix of spatial spectral features on temporal dependent features, the feature vectors of each time step of the temporal dependent features are weighted and aggregated to obtain temporally enhanced spatial spectral features. Finally, based on the attention weight matrix of temporal dependent features on spatial spectral features, the feature vectors of each spatial location of the spatial spectral features are weighted and aggregated to obtain spatially enhanced temporally dependent features. Finally, the temporally enhanced spatial spectral features and the spatially enhanced temporally dependent features are concatenated along the feature dimension. The concatenated feature vector is then input into a multilayer perceptron consisting of two fully connected layers for cross-modal information compression and nonlinear mapping. The output of the multilayer perceptron is used as the fused environmental feature tensor.

[0091] It should be noted that in this application, the temporally enhanced spatial spectral features are multimodal spatial representation vectors with cross-modal temporal semantic enhancement injected into the original spatial spectral information; the temporally dependent features of the spatial spectral enhancement are enhanced temporal representation vectors with cross-modal spatial location information integrated into the original temporal change patterns; the environmental feature tensor is a cross-modal fusion feature tensor that uniformly represents the multi-source heterogeneous environmental perception data of the mine ecological restoration area. The temporally enhanced spatial spectral features enable the spatial spectral modes to perceive the dynamic evolution of their temporal correlations, and the temporally dependent features of the spatial spectral enhancement enable the temporal modes to locate their corresponding ecological damage areas in space. The environmental feature tensor formed by the two realizes the bidirectional mutual enhancement fusion of air-space-ground-depth multi-source perception data at the feature layer, which can eliminate the information fragmentation of single-modal independent representations and provide a semantically rich and cross-modal aligned unified feature representation for subsequent variational mode decomposition and multi-task temporal prediction.

[0092] In step 103, the environmental feature tensor is decomposed and dimensionality reduced, and multi-task time series prediction is performed based on the dimensionality-reduced feature sequence to obtain the trend prediction values ​​of the mine ecological restoration area on four types of risk indices: geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness.

[0093] In some embodiments, reference Figure 3 The figure is a flowchart illustrating the operation of determining trend prediction values ​​according to some embodiments of this application. In this application, the environmental feature tensor is decomposed and dimensionality reduced, and multi-task time series prediction is performed based on the dimensionality-reduced feature sequence to obtain the trend prediction values ​​of the mine ecological restoration area on four risk indices: geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness. This can be achieved by the following steps:

[0094] The temporal components of each dimension of the environmental feature tensor are sequentially subjected to variational mode decomposition to obtain multiple intrinsic mode components.

[0095] Nonlinear dimensionality reduction is performed on all intrinsic mode components to form a dimensionality-reduced feature sequence;

[0096] The reduced feature sequence is input into an attention-enhanced long short-term memory network for multi-task temporal prediction, and trend prediction values ​​of four risk indices of mine ecological restoration areas, namely geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness, are obtained.

[0097] It should be noted that in this application, the attention-enhanced long short-term memory network is a composite deep neural network structure that cascades a temporal attention mechanism as an independent functional layer with long short-term memory units. It consists of three core parts: the first part is the input layer, which receives the feature sequence after dimensionality reduction via variational mode decomposition and kernel principal component analysis, and feeds the feature vectors of each time step into the network in chronological order; the second part is the temporal attention layer, located between the input layer and the long short-term memory units. This layer performs adaptive weighting operations on the feature vectors of each time step in the input feature sequence using a trainable attention weight matrix. For each time step's feature vector, this layer calculates an attention weight coefficient between zero and one, and multiplies the original feature vector element-wise with the corresponding attention weight coefficient to obtain a weighted feature vector. This enhances the network's attention to information at inflection points in risk evolution and periods of abnormal fluctuations, while suppressing noise during stable periods. The third part is a stacked long short-term memory (LSM) unit layer, consisting of two or three LSM units stacked in series. Each LSM unit contains three gating structures: a forget gate, an input gate, and an output gate. The forget gate controls which information in the cell state at the previous time step needs to be discarded, the input gate controls which new information in the input features at the current time step needs to be written into the cell state, and the output gate controls which information in the cell state at the current time step needs to be output to the next layer or the next time step. Through the synergistic effect of the three gating structures, the gradient vanishing and gradient exploding problems are effectively alleviated in the long-term modeling process, and the long-range dependencies with large spans in the environmental parameter sequence are captured. After the LSM unit stacked layer, four independent fully connected output heads are set. Each output head consists of one or two fully connected layers, corresponding to the predicted outputs of the geological disaster risk index, vegetation degradation risk index, water and soil pollution risk index, and ecological restoration effectiveness degradation risk index, respectively.

[0098] In practice, firstly, variational mode decomposition (VMD) is used to adaptively decompose the temporal components of each dimension in the environmental feature tensor. Under a preset constraint on the number of decomposition levels, each temporal component is non-recursively analyzed into multiple intrinsic mode components with different center frequencies and finite bandwidths. Then, kernel principal component analysis (KPCA) is used to perform nonlinear dimensionality reduction on all intrinsic mode components across all dimensions. For example, Gaussian radial basis function (RBF) is used to map each intrinsic mode component to a high-dimensional feature space to extract principal components. Based on a preset cumulative information retention rate threshold, the key information components that contribute the most to the risk index prediction are automatically selected. The selected key information components are then concatenated in temporal order to form the dimensionality reduction matrix. The feature sequence is divided into four dimensions. The risk index prediction contribution is a measure of the importance of each intrinsic modal component in improving the accuracy of trend predictions for the four risk indices. This contribution can be quantified by the eigenvalues ​​of each kernel principal component in kernel principal component analysis. Larger eigenvalues ​​indicate a greater variance explained by the kernel principal component in the original data, thus contributing more to subsequent risk index predictions. Finally, the dimensionality-reduced feature sequence is input into an attention-enhanced long short-term memory network for multi-task temporal prediction. A temporal attention layer adaptively assigns weights to the feature vectors at different time steps to enhance the focus on information at inflection points in risk evolution. The long short-term memory unit then performs weighted analysis. The subsequent feature sequences are modeled for long-range temporal dependencies to capture the nonlinear coupling dynamics between environmental parameters. Then, four independent fully connected output heads, set in parallel at the end of an attention-enhanced long short-term memory network, synchronously output the trend prediction values ​​of the geological disaster risk index, vegetation degradation risk index, water and soil pollution risk index, and ecological restoration effectiveness degradation risk index of the target mine ecological restoration area within a preset future time period. The trend prediction values ​​of each risk index are calculated as follows: first, the hidden state vector is output through the last time step of the attention-enhanced long short-term memory network; then, this hidden state vector is used as input through the fully connected output heads corresponding to each risk index, and the calculation formula is... The trend prediction value corresponding to the risk index is obtained by calculating the weight vector of the fully connected output head · the hidden state vector + the bias term of the fully connected output head. The above four risk indices all have corresponding weight vectors and bias terms in the attention-enhanced long short-term memory network. The geological disaster risk index is a quantitative indicator reflecting the comprehensive probability of landslides, collapses and ground fissures. The vegetation degradation risk index is a quantitative indicator reflecting the abnormal change rate of vegetation coverage and normalized vegetation index. The water and soil pollution risk index is a quantitative indicator reflecting the multiple of heavy metal exceedance in soil and the magnitude of change in pollutant concentration in water. The ecological restoration effectiveness degradation risk index is a comprehensive quantitative indicator reflecting the degree of overall ecological function decline in the restoration area.

[0099] It should be noted that the training process of the attention-enhanced long short-term memory network in this application is as follows: First, multi-source environmental perception data of the target mine ecological restoration area during historical periods are collected. A training sample set is constructed according to the above-mentioned variational mode decomposition and kernel principal component analysis dimensionality reduction process. Each training sample includes a dimensionality-reduced feature sequence of a continuous time step and the corresponding true label values ​​of four risk indices. The true label value of the geological disaster risk index is determined based on historical geological disaster event records and expert scores. The true label value of the vegetation degradation risk index is calculated based on the normalized vegetation index and historical monitoring data of vegetation coverage and their rate of change. The true label value of the water and soil pollution risk index is calculated based on historical monitoring data of soil heavy metal concentration and water pollutant concentration and their exceedance multiples. The true label value of the comprehensive risk index of ecological restoration effectiveness degradation is determined by the above three... The risk indices are weighted and synthesized according to preset weights. Then, the training sample set is divided into a training set and a validation set according to the proportion. The mean squared error loss function is used to measure the total error between the predicted values ​​of the four output heads and the corresponding true label values. The specific form of the mean squared error loss function is the weighted sum of the prediction errors of the four output heads. The error weight of each output head is preset according to the importance of the corresponding risk index in the actual early warning work. The adaptive moment estimation optimizer is selected as the parameter optimization algorithm. The initial learning rate is set to 0.1%, and the batch size is set to 32 or 64. In each training iteration, the gradient of the loss function with respect to the weight parameters of each layer is calculated and the weight parameters are updated along the gradient descent direction. At the same time, an early stopping strategy is used to monitor the loss function value on the validation set. When the loss on the validation set has not decreased for ten consecutive training iterations, the training is terminated early to prevent the model from overfitting to the training samples.

[0100] It should be noted that in this application, the intrinsic modal components are oscillating signal components, each with a single dominant vibration frequency and finite bandwidth. They can effectively separate the coupled and superimposed multi-scale fluctuation characteristics in mining environmental parameters into independent modes with clear physical meanings. This allows the seasonal growth cycle fluctuations of vegetation, the short-term impulse response caused by rainfall events, and the long-term trend changes caused by the slow deformation of the goaf to each correspond to different intrinsic modal components, providing a set of input features with a clear structure and strong separability for subsequent kernel principal component analysis. The dimensionality reduction feature sequence is a low-dimensional feature vector sequence assembled in chronological order. The trend prediction value is a continuous numerical sequence reflecting the dynamic evolution direction of each risk index in the target mine ecological restoration area within a future preset period. It can upgrade the static alarm mode based on a single threshold judgment at the current moment in traditional early warning methods to an advanced prediction mode based on the risk evolution trend in the future period. This enables the early warning system to detect the inflection point of accelerated deterioration of risks before the risk reaches the critical threshold, providing a quantitative decision-making basis for the early deployment of graded early warning response strategies and the reserved preparation time window for emergency response measures.

[0101] In step 104, the trend prediction values ​​of each risk index are compared with the preset four-level early warning threshold system to determine the risk warning level and trigger the corresponding response strategy.

[0102] In some embodiments, the trend forecast values ​​of each risk index are compared with a preset four-level early warning threshold system to determine the risk warning level and trigger the corresponding response strategy. This can be achieved by the following steps:

[0103] For the trend prediction value of each risk index, the trend prediction value is compared with the warning threshold of the corresponding risk dimension in the preset four-level warning threshold system index by index to obtain the warning level of the risk index, and then the warning level of each risk index is obtained.

[0104] The current risk warning level of the target mine ecological restoration area is determined based on the warning levels of all risk indices;

[0105] Based on the risk warning level, the corresponding response measures are retrieved from the preset differentiated warning response strategy library and triggered for execution.

[0106] It should be noted that in this application, the warning level is an indicator of the degree to which the mine ecological restoration area deviates from the safety benchmark under a single risk dimension; the risk warning level is an indicator that uniformly represents the overall comprehensive risk situation currently faced by the mine ecological restoration area.

[0107] In practical implementation, firstly, a pre-constructed four-level early warning threshold system can be used to compare the trend prediction values ​​of the geological disaster risk index, vegetation degradation risk index, water and soil pollution risk index, and the comprehensive risk index of ecological restoration effectiveness degradation, item by item, to obtain the early warning level independently mapped to each risk index. The four-level early warning threshold system is set according to each risk dimension. For example, the geological disaster risk index uses surface displacement rate as the grading indicator, with its blue, yellow, orange, and red early warning thresholds set to 5 mm, 10 mm, 15 mm, and 20 mm per day, respectively. The vegetation degradation risk index uses the rate of change of the normalized vegetation index from the benchmark value as the grading indicator, with corresponding thresholds set to decreases of 20%, 30%, 40%, and 50% respectively. The water and soil pollution risk index uses the soil heavy metal cadmium content as the grading indicator, with corresponding thresholds of 0.5% per kilogram. The comprehensive risk index for ecological restoration effectiveness degradation was divided into four threshold levels based on the normalized comprehensive score: 0.6 mg, 0.9 mg, 1.2 mg, and 1.5 mg. Then, according to the multi-indicator fusion judgment rule of "choosing the higher one," the warning levels mapped independently by the four risk indices were compared, and the highest level was determined as the current risk warning level of the mine ecological restoration area. Finally, based on the risk warning level, strategies were matched in a preset differentiated warning response strategy library, and corresponding warning response measures were retrieved. In this differentiated warning response strategy library, the blue warning level corresponds to the archiving of monitoring data records in the restoration area and continuous tracking of risk trends; the yellow warning level corresponds to the increase in the sampling frequency of related risk dimension sensors and the judgment of risk evolution trends; the orange warning level corresponds to the issuance of drone emergency inspection instructions and the generation of on-site emergency response plans; and the red warning level corresponds to the sending of alarm signals by the emergency command center, the push of personnel evacuation notices in the affected area, and the activation of multi-department joint emergency response.

[0108] Furthermore, in another aspect of this application, in some embodiments, this application provides a risk early warning system for mine ecological restoration areas based on environmental perception, with reference to... Figure 4 The figure is a schematic diagram of the structure of a mine ecological restoration area risk early warning system based on environmental perception, according to some embodiments of this application. It includes: a data acquisition module 201, a processing module 202, and an execution module 203, which are described below:

[0109] The acquisition module 201 in this application is mainly used to collect multi-dimensional environmental perception data and register spatiotemporal benchmarks for the mine ecological restoration area through the constructed four-level environmental perception system of air-space-ground-depth, so as to obtain a multi-source environmental dataset.

[0110] Processing module 202, in this application, is used to extract the spatial spectral features of remote sensing images and UAV images in the multi-source environmental dataset, extract the temporal dependence features of sensing time series data and microseismic signals in the multi-source environmental dataset, and then perform cross-modal feature fusion on the spatial spectral features and the temporal dependence features to obtain the fused environmental feature tensor.

[0111] It should be noted that the processing module 202 is also used to decompose and reduce the dimensionality of the environmental feature tensor, and perform multi-task time series prediction based on the dimensionality-reduced feature sequence to obtain the trend prediction values ​​of the mine ecological restoration area on four types of risk indices: geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness.

[0112] The execution module 203 in this application is mainly used to compare the trend prediction values ​​of each risk index with the preset four-level early warning threshold system to determine the risk warning level and trigger the corresponding response strategy.

[0113] The foregoing detailed examples of the risk early warning method and system for mine ecological restoration areas based on environmental perception provided in the embodiments of this application. It is understood that the corresponding apparatus, in order to achieve the above functions, includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specified application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specified application, but such implementation should not be considered beyond the scope of this application.

[0114] In some embodiments, this application also provides a computer device, the computer device including a memory and a processor, the memory for storing a computer program, and the processor for calling and running the computer program from the memory, so that the computer device executes the above-described method for risk early warning of mine ecological restoration areas based on environmental perception.

[0115] In some embodiments, reference Figure 5 The dashed lines in the figure indicate that the unit or module is optional. This figure is a structural schematic diagram of a computer device for implementing a mine ecological restoration area risk early warning method based on environmental perception, according to an embodiment of this application. The mine ecological restoration area risk early warning method based on environmental perception described in the above embodiments can... Figure 5The computer device shown is used to implement this, and the computer device includes at least one processor 301, a memory 302 and at least one communication unit 305. The computer device may be a terminal device, a server or a chip.

[0116] Processor 301 can be a general-purpose processor or a special-purpose processor. For example, processor 301 can be a central processing unit (CPU), which can be used to control computer devices, execute software programs, and process data from software programs. The computer device may also include a communication unit 305 for inputting (receiving) and outputting (transmitting) signals.

[0117] For example, the computer device may be a chip, and the communication unit 305 may be the input and / or output circuit of the chip, or the communication unit 305 may be the communication interface of the chip, which may be a component of a terminal device, network device or other device.

[0118] For example, the computer device may be a terminal device or a server, and the communication unit 305 may be a transceiver of the terminal device or the server, or the communication unit 305 may be a transceiver circuit of the terminal device or the server.

[0119] The computer device may include one or more memories 302 storing a program 304. The program 304 can be executed by a processor 301 to generate instructions 303, causing the processor 301 to execute the method described in the above method embodiments according to the instructions 303. Optionally, the memory 302 may also store data (such as a target audit model). Optionally, the processor 301 may also read data stored in the memory 302, which may be stored at the same storage address as the program 304, or it may be stored at a different storage address than the program 304.

[0120] The processor 301 and memory 302 can be configured separately or integrated together, for example, integrated on the system on chip (SOC) of the terminal device.

[0121] It should be understood that each step of the above method embodiment can be completed by hardware logic circuits or software instructions in the processor 301. The processor 301 can be a CPU, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, such as discrete gates, transistor logic devices, or discrete hardware components.

[0122] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0123] For example, in some embodiments, this application also provides a computer-readable storage medium storing instructions or code that, when executed on a computer, cause the computer to implement the above-described environmentally-aware mine ecological restoration area risk early warning method.

[0124] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0125] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A mine ecological restoration area risk early warning method based on environmental perception, characterized in that, Includes the following steps: By constructing a four-level environmental perception system (air-space-ground-depth), multi-dimensional environmental perception data collection and spatiotemporal benchmark registration are carried out in the mine ecological restoration area to obtain a multi-source environmental dataset. Spatial spectral features of remote sensing images and UAV images in the multi-source environmental dataset are extracted. Temporal dependence features of sensing time series data and microseismic signals are extracted from the multi-source environmental dataset. Then, cross-modal feature fusion is performed on the spatial spectral features and the temporal dependence features to obtain the fused environmental feature tensor. The environmental feature tensor is decomposed and dimensionality reduced, and multi-task time series prediction is performed based on the dimensionality-reduced feature sequence to obtain the trend prediction values ​​of the mine ecological restoration area on four types of risk indices: geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness. The trend forecast values ​​of each risk index are compared with the preset four-level early warning threshold system to determine the risk warning level and trigger the corresponding response strategy.

2. The method of claim 1, wherein, By constructing a four-level environmental perception system (space-air-ground-depth), multi-dimensional environmental perception data collection and spatiotemporal benchmark registration were carried out in the mine ecological restoration area, resulting in a multi-source environmental dataset, specifically including: Remote sensing images of the target mine ecological restoration area were acquired through a satellite remote sensing data interface. The drone aerial photography transmission module is used to acquire drone images covering the target mine ecological restoration area. By deploying a sensor array within the target mine ecological restoration area, soil physicochemical parameters, groundwater level and water quality indicators, and three-dimensional surface displacement were collected to obtain sensing time-series data. Microseismic signals of rock strata within the target mine ecological restoration area are collected through a network of deployed deep microseismic sensors. The remote sensing images, the UAV images, the sensor time series data, and the microseismic signals are converted to the same geographic coordinate system and unified with a time reference to obtain a multi-source environmental dataset with spatiotemporal reference registration.

3. The method of claim 1, wherein, Extracting the spatial spectral features of remote sensing images and UAV images from the multi-source environmental dataset specifically includes: Multi-scale spatial feature extraction is performed on the remote sensing images and UAV images in the multi-source environmental dataset to obtain the initial feature map of the surface of the target mine ecological restoration area; Eliminate the radiation distortion of spectral information caused by terrain undulation in the initial feature map to obtain a terrain-corrected spectral feature map; Geometric deformation modeling is performed on the spectral feature map to obtain spatial spectral features.

4. The method as described in claim 1, characterized in that, Extracting the time-series dependency features between sensing time-series data and microseismic signals from the multi-source environmental dataset specifically includes: The sensor time series data in the multi-source environmental dataset are spliced ​​together according to the sampling timestamp to obtain the ground sensor time series observation stream; The microseismic signals in the multi-source environmental dataset are subjected to event-triggered interception and time-out picking to obtain a microseismic event waveform sequence; Multi-scale temporal feature extraction is performed by temporal convolution on the sensor time-series observation stream and the microseismic event waveform sequence to obtain sensor time-series features and microseismic time-series features; The sensing time-series features and the microseismic time-series features are aggregated to obtain time-dependent features.

5. The method as described in claim 1, characterized in that, The cross-modal feature fusion of the spatial spectral features and the temporal dependency features to obtain the fused environmental feature tensor specifically includes: The spatial spectral features are flattened along the channel dimension and linearly projected to obtain the spatial spectral query vector and the spatial spectral key vector. The time-dependent features are flattened and linearly projected along the time dimension to obtain the time-series query vector and the time-series key-value vector. The spatial spectrum query vector and the temporal key value vector are subjected to cross-modal dot product attention calculation to obtain the attention weight matrix of spatial spectrum to temporal. The temporal query vector and the spatial spectral key vector are subjected to cross-modal dot product attention calculation to obtain the attention weight matrix of temporal to spatial spectrum; The spatial spectral features for time enhancement are determined based on the attention weight matrix of spatial spectra on time and the time-dependent features. The temporal dependence features of spatial spectral enhancement are determined based on the attention weight matrix of the spatial spectrum on the time sequence and the spatial spectral features. The temporally enhanced spatial spectral features and the temporally dependent features of the spatial spectral enhancement are concatenated and then subjected to cross-modal information compression and nonlinear mapping to obtain the fused environmental feature tensor.

6. The method as described in claim 1, characterized in that, The environmental feature tensor is decomposed and dimensionality reduced, and multi-task time series prediction is performed based on the dimensionality-reduced feature sequence to obtain the trend prediction values ​​of the mine ecological restoration area on four types of risk indices: geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness. Specifically, these include: The temporal components of each dimension of the environmental feature tensor are sequentially subjected to variational mode decomposition to obtain multiple intrinsic mode components. Nonlinear dimensionality reduction is performed on all intrinsic mode components to form a dimensionality-reduced feature sequence; The reduced feature sequence is input into an attention-enhanced long short-term memory network for multi-task temporal prediction, and trend prediction values ​​of four risk indices of mine ecological restoration areas, namely geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness, are obtained.

7. The method as described in claim 1, characterized in that, The trend forecast values ​​of each risk index are compared with the preset four-level early warning threshold system to determine the risk warning level and trigger the corresponding response strategy. Specifically, this includes: For the trend prediction value of each risk index, the trend prediction value is compared with the warning threshold of the corresponding risk dimension in the preset four-level warning threshold system index by index to obtain the warning level of the risk index, and then the warning level of each risk index is obtained. The current risk warning level of the target mine ecological restoration area is determined based on the warning levels of all risk indices; Based on the risk warning level, the corresponding response measures are retrieved from the preset differentiated warning response strategy library and triggered for execution.

8. A risk early warning system for mine ecological restoration areas based on environmental perception, employing the risk early warning method for mine ecological restoration areas based on environmental perception as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to collect multi-dimensional environmental perception data and perform spatiotemporal benchmark registration in the mine ecological restoration area through the constructed four-level environmental perception system of air-space-ground-depth, so as to obtain multi-source environmental datasets. The processing module is used to extract the spatial spectral features of remote sensing images and UAV images in the multi-source environmental dataset, extract the temporal dependence features of sensing time series data and microseismic signals in the multi-source environmental dataset, and then perform cross-modal feature fusion on the spatial spectral features and the temporal dependence features to obtain the fused environmental feature tensor. The processing module is also used to decompose and reduce the dimensionality of the environmental feature tensor, and perform multi-task time series prediction based on the dimensionality-reduced feature sequence to obtain the trend prediction values ​​of the mine ecological restoration area on four types of risk indices: geological disasters, vegetation degradation, water and soil pollution, and degradation of restoration effectiveness. The execution module is used to compare the trend prediction values ​​of each risk index with the preset four-level early warning threshold system to determine the risk warning level and trigger the corresponding response strategy.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the environmental perception-based risk early warning method for mine ecological restoration areas as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the risk early warning method for mine ecological restoration areas based on environmental perception as described in any one of claims 1 to 7.

Images