Method and system for evaluating ecological restoration effect of abandoned stone coal mine

Abstract

The application discloses a kind of abandoned stone coal mine ecological restoration effect evaluation method and system, method includes: the dataset of the multi-source ecology of repair area is established, and the multidimensional time series matrix is formed after data set is spatially registered, time marker and abnormality is eliminated;The time series matrix is analyzed in time sequence consistency, and the time sequence dependence between ecological factors is calculated using the time lag correction algorithm based on mutual information, and the coupling feature set is generated using variable step delay matching to correct time difference;The coupling feature set is input into the graph convolution network model, the node is ecological factor, and the edge weight is time correlation, and the associated features of space and time are extracted by graph convolution.The application realizes the comprehensive evaluation of precision, stability and interpretability of abandoned mine restoration effect by constructing multi-source ecological coupling model and dynamic intelligent analysis mechanism.

Description

Technical Field

[0001] This application relates to the field of ecological restoration assessment technology, and in particular to a method and system for evaluating the ecological restoration effect of abandoned coal mines. Background Technology

[0002] In the ecological restoration of abandoned coal mines, the scientific evaluation of restoration effects is a key basis for subsequent governance decisions and resource inputs. However, existing methods often focus on single-dimensional analysis of vegetation restoration or soil physicochemical indicators, neglecting the dynamic evolution characteristics of the mine ecosystem in terms of time scale and spatial structure, making it difficult to fully reflect the stability and sustainability of the restored ecosystem.

[0003] On the one hand, traditional evaluation methods lack a mechanism for identifying the temporal consistency among multi-source ecological data. Data such as vegetation indices, hydrological characteristics, and soil activity factors obtained from different monitoring periods have significant time lags and interferences. Without temporal feature coupling and dynamic equilibrium analysis, the evaluation results are easily distorted, leading to misjudgments of the true ecological state of the restoration area.

[0004] On the other hand, existing methods often treat environmental factors equally when integrating multidimensional indicators, neglecting the role of negative ecological contribution factors. That is, although some restored areas show high vegetation coverage in the short term, negative factors such as soil heavy metal migration and groundwater chemical anomalies are not fully identified and weighted, resulting in a systematic overestimation of the overall ecological quality. This problem is particularly prominent in the high-sulfur and high-vanadium environment unique to abandoned coal mines.

[0005] In summary, the current evaluation system for the restoration of abandoned mines has the problem of insufficient identification of temporal dynamics and negative ecological coupling effects, and lacks a comprehensive evaluation method that can simultaneously consider the consistency of ecological element evolution and the weight balance of negative effect factors. Summary of the Invention

[0006] To address the aforementioned problems, embodiments of the present invention provide a method for evaluating the ecological restoration effect of abandoned coal mines, the method comprising: A dataset of multi-source ecosystems in the restoration area was established. After spatial registration, temporal labeling, and anomaly removal, the dataset was transformed into a multi-dimensional time-series matrix. Temporal consistency analysis is performed on the time series matrix. A time delay correction algorithm based on mutual information is used to calculate the temporal dependence between ecological factors. Variable step size delay matching is used to correct the time difference and generate a coupled feature set. The coupled feature set is input into a graph convolutional network model, where nodes are ecological factors and edge weights are temporal correlations. Spatial and temporal correlation features are extracted through graph convolution. A negative ecological contribution module is constructed on the graph convolution output. A bidirectional attention mechanism is used to identify negative ecological factors in a weighted manner, and the weighted result is output. The training is optimized by inverse gradient constraint. The associated features and weighted results are input into the fusion module, and nonlinear fusion is performed using a multi-channel autoencoder structure to output a comprehensive ecological state vector. The comprehensive ecological state vector is input into the joint model of long short-term memory and adaptive fuzzy inference. The gating parameters are adjusted and the prediction is corrected through dynamic membership function to obtain the comprehensive evaluation result of the restoration effect.

[0007] Furthermore, the time delay correction algorithm includes a mutual information entropy calculation module and a time synchronization correction module. The mutual information entropy calculation module is used to measure the correlation strength between ecological variables. The time synchronization correction module determines the time delay corresponding to the maximum mutual information based on a variable step-size window search algorithm, and performs synchronization alignment and interpolation compensation on multi-source time series within a sliding window to improve the temporal consistency between different data sources.

[0008] Furthermore, the graph convolutional network model includes spectral domain convolutional layers, temporal convolutional layers, and pooling layers; the spectral domain convolutional layers perform feature decomposition on the Laplacian matrix of the ecological factor graph to extract global correlations; the temporal convolutional layers adopt a gated convolutional structure to capture multi-scale time-dependent features; the pooling layers generate compressed graph feature embeddings through weighted averaging and downsampling operations for subsequent module input.

[0009] Furthermore, in the negative ecological contribution module, the bidirectional attention layer includes a temporal attention sublayer and a spatial attention sublayer. The temporal attention sublayer calculates the influence intensity of each ecological factor in different time windows, and the spatial attention sublayer calculates the degree of spatial interference between factors. The negative weight matrix output by the negative ecological contribution module is regularized by the inverse gradient constraint unit and used to weight the corresponding factor components in the main feature set.

[0010] Furthermore, the multi-channel autoencoder structure of the fusion module includes: a feature branch layer, an encoding layer, a decoding layer, and a channel fusion layer; the feature branch layer divides the input data into three channels according to spatial features, temporal features, and environmental background features; the encoding layer independently extracts high-order nonlinear representations in each channel; the decoding layer reconstructs the data using deconvolution operations with shared weights; and the channel fusion layer fuses the outputs of each channel using an attention weighting mechanism to generate a comprehensive ecological state vector.

[0011] Furthermore, the joint model of long short-term memory and adaptive fuzzy inference includes a bidirectional coupling structure between a long short-term memory temporal prediction unit and a fuzzy self-adjusting unit. The long short-term memory temporal prediction unit is responsible for extracting the temporal change patterns of ecological factors, while the fuzzy self-adjusting unit describes the uncertainty of input features by calculating dynamic membership functions in real time and generates weight adjustment signals based on membership changes, which are then fed back to the gating control layer of the long short-term memory. During training, the fuzzy rule base automatically adjusts the membership function boundaries and weight distribution based on the prediction residuals, achieving synergistic optimization of fuzzy logic and temporal memory to improve the model's adaptive prediction capability for complex ecological processes.

[0012] Furthermore, the comprehensive ecological state vector is normalized and denoised before output. The normalization adopts the hierarchical standard deviation constraint method to map ecological features of different dimensions to a unified scale. The denoising adopts the multi-scale wavelet threshold method to suppress high-frequency noise components, so as to reduce the interference of monitoring errors on the final prediction results.

[0013] Furthermore, in the restoration effect output stage, based on the joint feature space of graph convolutional network and LSTM model, an adaptive threshold segmentation and multiple aggregation mechanism are introduced. The adaptive threshold segmentation module uses statistical mean square error to dynamically determine the restoration level boundary; the multiple aggregation module uses a hierarchical weighted decision structure to fuse and calculate the time series indicators, spatial structure indicators and negative ecological weight results to obtain a multidimensional and consistent restoration effect evaluation result.

[0014] A method for evaluating the ecological restoration effect of abandoned coal mines, the method further includes: Introducing time-series dynamic smoothing and adaptive residual correction, including: After the joint model of long short-term memory and adaptive fuzzy inference outputs the comprehensive evaluation result of the repair effect, the comprehensive evaluation result of the repair effect is dynamically smoothed. By calculating the weighted moving average of the prediction results within the continuous time window, the impact of short-term disturbances on the overall prediction trend is reduced. At the same time, an adaptive residual correction mechanism is established to fine-tune the current model output using the prediction error of the previous period, so as to improve the model's performance in terms of temporal continuity and prediction stability.

[0015] A multimodal fusion intelligent question answering and knowledge retrieval system, comprising: The data construction module establishes a dataset of multi-source ecosystems in the restoration area. After spatial registration, temporal labeling, and anomaly removal, the dataset forms a multi-dimensional time series matrix. The time series analysis module performs time series consistency analysis on the time series matrix, uses a time delay correction algorithm based on mutual information to calculate the time series dependence between ecological factors, and uses variable step size delay matching to correct time differences and generate a coupled feature set. The feature extraction module inputs the coupled feature set into the graph convolutional network model, where nodes are ecological factors and edge weights are temporal correlations, and extracts spatial and temporal correlation features through graph convolution. The contribution identification module constructs a negative ecological contribution module on the graph convolution output, uses a bidirectional attention mechanism to weightedly identify negative ecological factors, outputs the weighted result, and optimizes the training through inverse gradient constraints. A multi-channel fusion module inputs the associated features and weighted results into the fusion module, performs nonlinear fusion using a multi-channel autoencoder structure, and outputs a comprehensive ecological state vector. The fuzzy evaluation module inputs the comprehensive ecological state vector into the joint model of long short-term memory and adaptive fuzzy inference, adjusts the gating parameters and corrects the prediction through a dynamic membership function, and obtains a comprehensive evaluation result of the restoration effect.

[0016] The technical effects and advantages of the method for evaluating the ecological restoration effect of abandoned coal mines provided by this invention are as follows: This invention achieves a comprehensive evaluation of the restoration effects of abandoned mines with precision, stability, and interpretability by constructing a multi-source ecological coupling model and a dynamic intelligent analysis mechanism. The invention addresses the temporal misalignment and coupling imbalance between ecological monitoring data through multi-source dataset construction and temporal consistency analysis, improving the temporal coordination and data integrity of the evaluation. It employs graph convolutional networks to extract spatial-temporal correlation features between ecological factors, achieving a structured expression of complex ecological relationships in the restoration area. A negative ecological contribution module identifies and weights potential adverse ecological factors, ensuring that the evaluation results reflect both the positive restoration process and reveal potential ecological risks. A multi-channel self-encoding fusion mechanism enhances the nonlinear comprehensive capability of spatial, temporal, and environmental characteristics, making the restoration state vector globally representative. A joint model of long short-term memory and adaptive fuzzy inference is introduced to improve the system's dynamic prediction and uncertainty adaptation capabilities for complex ecological processes. After temporal smoothing and residual correction, the model output exhibits higher continuity and stability, providing a reliable basis for the phased assessment and dynamic decision-making of ecological restoration. Attached Figure Description

[0017] Figure 1 This is a flowchart of an evaluation method for the ecological restoration effect of an abandoned coal mine, as shown in Example 1. Figure 2 This is a flowchart of a method for evaluating the ecological restoration effect of abandoned coal mines in Example 2; Figure 3 This is a schematic diagram of the connection of an ecological restoration effect evaluation system for abandoned coal mines in Example 3. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 Please see Figure 1 As shown, an embodiment of the present invention provides a method for evaluating the ecological restoration effect of abandoned coal mines, comprising the following steps: S1. Establish a dataset of multi-source ecosystems in the restoration area. After spatial registration, temporal labeling and anomaly removal, the dataset forms a multi-dimensional time series matrix. S2. Perform time series consistency analysis on the time series matrix, use a time delay correction algorithm based on mutual information to calculate the time series dependence between ecological factors, and use variable step size delay matching to correct the time difference and generate a coupled feature set. S3. Input the coupled feature set into the graph convolutional network model, where the nodes are ecological factors and the edge weights are temporal correlations, and extract the spatial and temporal correlation features through graph convolution; S4. Construct a negative ecological contribution module on the graph convolution output, use a bidirectional attention mechanism to identify negative ecological factors in a weighted manner, output the weighted results, and optimize the training through inverse gradient constraints. S5. Input the associated features and weighted results into the fusion module, and perform nonlinear fusion using a multi-channel autoencoder structure to output a comprehensive ecological state vector. S6. Input the comprehensive ecological state vector into the joint model of long short-term memory and adaptive fuzzy inference, adjust the gating parameters and correct the prediction through the dynamic membership function, and obtain the comprehensive evaluation result of the restoration effect.

[0020] The dataset in S1 includes vegetation spectral indices, hydrological factors, and soil physicochemical and microbial activity indicators.

[0021] The time-delay correction algorithm in S2 includes a mutual information entropy calculation module and a time synchronization correction module. The mutual information entropy calculation module is used to measure the correlation strength between ecological variables. The time synchronization correction module determines the time delay corresponding to the maximum mutual information based on a variable step-size window search algorithm, and performs synchronization alignment and interpolation compensation on multi-source time series within a sliding window to improve the temporal consistency between different data sources. The data after time-delay correction provides more accurate information on the temporal evolution of ecological factors, enabling subsequent graph convolutional networks to capture more accurate spatiotemporal relationships when modeling these data. In addition, the corrected data also improves the accuracy of the negative ecological contribution identification module and avoids the negative factor identification bias caused by time-delay errors.

[0022] When performing time series correction, the mutual information entropy calculation module is used to measure the correlation strength between ecological factors. Mutual information entropy is a way to measure the dependency between variables. Its basic idea is that the richer the information between two sets of time series data, the stronger the correlation between them. In the context of ecological restoration, this calculation helps to discover the temporal relationship between different ecological factors, such as the coupling between vegetation growth and soil change.

[0023] Mutual information entropy calculation calculates the degree of dependence between two time series by analyzing the amount of overlapping information within the same time window. This amount of information is used to provide a preliminary basis for time delay correction. Specifically, a period with higher mutual information entropy indicates that the two factors have a strong dependence within that time window and their time steps need to be adjusted synchronously.

[0024] After obtaining the mutual information entropy value, the time synchronization correction module continues to perform the critical task of correcting time delay deviations. To this end, a variable step-size window search algorithm is employed. This algorithm searches for the optimal time delay correction value on the time axis by sliding the window. The time periods within the window are weighted according to their correlation strength to calculate an optimal time delay. The core of this process lies in determining the optimal time delay by maximizing the mutual information value, i.e., determining the time alignment point by finding the deviation at the position of maximum mutual information. Once the time delay value is determined, the time synchronization correction module aligns the data to ensure consistency of time series between different data sources. To enhance the correction effect, an interpolation compensation step is applied to the data within the time window, especially when data is lost or missing. The interpolation algorithm calculates the missing value using preceding and following data points, thereby reducing the impact of time series errors.

[0025] For example: In a practical application, consider two ecological factors—soil moisture and vegetation growth index—derived from different sensor systems with varying sampling frequencies. Soil moisture data is sampled 24 hours a day, while the vegetation growth index is sampled hourly. In this scenario, the collected vegetation growth index data may exhibit a significant time lag, meaning that vegetation changes at certain points in time may not reflect changes in soil moisture. Mutual information entropy calculations can identify a strong correlation between these two variables over certain time periods. However, due to the different sampling periods, the data is not perfectly aligned. Applying a time synchronization correction module automatically adjusts the timestamps of the vegetation growth index data based on the maximum mutual information criterion and corrects for time discrepancies using a variable step-size window, ensuring the time synchronization of soil moisture and vegetation growth, thereby improving the accuracy of subsequent analyses.

[0026] The graph convolutional network model in S3 includes spectral domain convolutional layers, temporal convolutional layers, and pooling layers. The spectral domain convolutional layers perform feature decomposition on the Laplacian matrix of the ecological factor graph to extract global correlations. Each node in the ecological factor graph represents an ecological factor (such as soil moisture, vegetation index, etc.), while the edges represent the temporal correlations between factors. The temporal convolutional layers use a gated convolutional structure to capture multi-scale time-dependent features. The pooling layers generate compressed graph feature embeddings through weighted averaging and downsampling operations, which are used as inputs for subsequent modules.

[0027] In the structure of graph convolutional networks, temporal convolutional layers are used to capture multi-scale temporal dependencies. The evolution of ecological factors depends not only on spatial relationships but also on temporal change patterns; therefore, temporal convolutional layers introduce gated convolutional structures, which can effectively capture and model long-term temporal dependencies in data, especially showing significant advantages when processing time series data. The key advantage of gated convolutional structures lies in their ability to flexibly control data updates at each time step. Specifically, the gating mechanism can dynamically adjust the intensity of convolutional operations according to the characteristics of the data, avoiding overfitting caused by unnecessary temporal dependencies. For example, ecological factors may change rapidly within some time windows, while changes are more gradual in other time windows. The gating mechanism can automatically adjust its processing method according to these different temporal dependencies.

[0028] The final layer of a graph convolutional network is the pooling layer, whose main function is to reduce the dimensionality and aggregate the graph features after convolution, generating more compact graph embedding features. The pooling layer uses a combination of weighted averaging and downsampling operations. By weighting the neighboring nodes of each graph node, the most important features in the graph are extracted, and the data dimensionality is further compressed by downsampling to reduce the amount of computation. The key to the pooling process is that weighted averaging can reasonably weight features according to the influence of different nodes, thereby ensuring that the model can effectively capture important local information; downsampling helps the model ignore unimportant features and focus on the ecological factors that have the greatest impact on the restoration effect.

[0029] For example: Suppose we have an ecological restoration area comprising three ecological factors: soil moisture, vegetation index, and temperature. We aim to capture the temporal evolution of these factors using a graph convolutional network. First, we construct an ecological factor graph, treating each factor as a node. Edges between nodes represent their temporal correlations. For example, soil moisture and vegetation index may have a strong temporal dependency, while the effect of temperature on vegetation growth may have a lag effect. Through spectral domain convolutional layers, the model can extract global correlation patterns between these factors, understanding their long-term relationships. Next, temporal convolutional layers, through gated convolution operations, capture the changing patterns over different time periods. For instance, in spring and summer, vegetation growth may be more sensitive to the influence of soil moisture, while in winter, the influence of temperature may be more significant. Temporal convolutional layers can automatically adjust their convolution operations based on these changes, thus more accurately modeling the temporal variations of ecological factors. Finally, pooling layers, through weighted averaging and downsampling, extract the most critical features and generate compressed graph feature embeddings. These embedded features will serve as input to subsequent modules for further evaluation of restoration effectiveness.

[0030] Through the structure of the graph convolutional network model, the model can effectively extract the spatiotemporal correlation features between ecological factors and provide accurate data features for subsequent evaluation of restoration effects. Especially when processing spatiotemporal data, the graph convolutional network can greatly improve the accuracy and reliability of the ecological restoration process through adaptive learning from a global perspective and time scale.

[0031] In the negative ecological contribution module of S4, the operating region of the bidirectional attention mechanism, namely the bidirectional attention layer, includes a temporal attention sublayer and a spatial attention sublayer. The temporal attention sublayer calculates the influence intensity of each ecological factor in different time windows, and the spatial attention sublayer calculates the degree of spatial interference between factors. The negative weight matrix output by the negative ecological contribution module is regularized by the inverse gradient constraint unit and used to weight the corresponding factor components in the principal feature set. By performing dual modeling of the spatiotemporal dependencies between factors, the bidirectional attention layer can better capture the complex interactions between ecological factors.

[0032] The primary task of the time attention sublayer is to calculate the influence intensity of each ecological factor within different time windows. Based on the input time-series data, this sublayer quantifies the relative importance of ecological factors within different time windows by calculating the correlation between different time steps. During ecological restoration, the role of certain factors may change significantly over different time periods. For example, some factors may have a significant impact on restoration effectiveness in the short term, but become less important in the long term. Through the time attention mechanism, the negative ecological contribution module can adaptively adjust its weights, strengthening the focus on ecological factors during critical time periods and suppressing the influence of less important time periods. This allows for more accurate identification of negative ecological factors affecting restoration effectiveness. The role of time attention is to dynamically capture the evolution of time-series characteristics and adjust the data weighting within each time step according to the current restoration status. The role of the spatial attention sublayer is to calculate the degree of spatial interference between factors. The spatial relationships between ecological factors are crucial to the restoration effect, especially in the process of ecological restoration, where ecological factors in different spatial regions may have different effects. For example, the diffusion range and concentration distribution of heavy metal pollution in soil may be uneven in different geographical regions, resulting in differences in the restoration effect in different regions. The spatial attention sublayer learns the dependencies between different spatial locations, quantifies these spatial interference factors, and thus effectively identifies and weights spatial factors that have a negative impact on the restoration effect. The role of the spatial attention mechanism is to dynamically adjust the weighted allocation of spatial information, so that the negative ecological contribution module can focus on the spatial regions that affect the restoration effect, while ignoring irrelevant regions.

[0033] During the training of the negative ecological contribution identification module, a set of negative weight matrices is generated. To prevent overfitting and gradient vanishing, a back gradient constraint unit is used to regularize the negative weight matrix. This unit controls the gradient update magnitude during backpropagation, enabling the negative ecological contribution identification module to adjust weights more stably during training and ensuring that the identification results of negative factors are not disturbed during training. Specifically, the back gradient constraint unit adjusts the gradient magnitude to avoid excessive weight updates, thereby maintaining the stability of the negative weight matrix. Through this regularization mechanism, the model can more accurately identify and weight negative influencing factors, avoiding misidentification due to overfitting.

[0034] For example: Assuming that vegetation index, soil moisture, and soil heavy metal pollution are the three main ecological factors in a certain remediation area, we first obtain the temporal and spatial correlation characteristics of these three factors through graph convolutional network processing. In the negative ecological contribution identification module, firstly, the temporal attention sublayer calculates the changing trend of vegetation index in different time windows and finds that the vegetation index changes significantly in spring and summer. Therefore, the weight of this factor will be enhanced during these time periods. Next, the spatial attention sublayer identifies the diffusion differences of soil heavy metal pollution in different regions, focusing on severely polluted areas, thus adding more negative influences to the weight allocation of these areas. Finally, through the inverse gradient constraint unit, negative ecological factors (such as soil heavy metal pollution) are correctly weighted, and the negative weight matrix is ​​output to correct the final remediation effect evaluation results.

[0035] The multi-channel autoencoder structure of the fusion module in S5 includes a feature branch layer, an encoding layer, a decoding layer, and a channel fusion layer. The feature branch layer divides the input data into three channels according to spatial features, temporal features, and environmental background features. The encoding layer independently extracts high-order nonlinear representations in each channel. The decoding layer uses deconvolution operations with shared weights for reconstruction. The channel fusion layer fuses the outputs of each channel using an attention weighting mechanism to generate a comprehensive ecological state vector.

[0036] The feature branching layer is the first layer of a multi-channel autoencoder network. Its function is to branch the input raw data according to different feature types. Specifically, the input data includes spatial features, temporal features, and environmental background features, which reflect different levels of ecological impact factors in the ecological restoration process. The feature branching layer divides the input data into three independent channels based on different data attributes, with each channel specifically processing one type of feature, including: Spatial characteristics: These typically reflect information such as geographical distribution and environmental structure, and these characteristics are crucial for describing ecological differences in different regions.

[0037] Temporal characteristics: These reflect the changing trends of ecological factors at different time scales and are an important component of time series data.

[0038] Environmental background features: including background information such as climate, pollution sources, and remediation measures, to help the model understand the external conditions during the remediation process.

[0039] In this way, the feature branching layer ensures that each feature type can be processed independently, thereby avoiding information interference and loss.

[0040] The task of the encoding layer is to extract nonlinear features from the data in each channel. In each channel, the network processes the data through convolution operations to extract a high-order representation of each feature. This process is mainly carried out by convolutional neural networks (CNNs), which extract deep information of different features in the channels of spatial, temporal and environmental features respectively. The encoding layer of each channel obtains a high-order nonlinear representation of the data through a series of convolution and activation function operations, thereby capturing more complex spatiotemporal correlation information. The role of this layer is to map the input data from the original space to a more abstract and compact feature space, preserving important spatiotemporal dependencies and patterns.

[0041] The function of the decoding layer is to reconstruct the output of the encoding layer, restore the original structure of the data, and provide the necessary information for the subsequent fusion layer. The decoding layer uses deconvolution operation, which processes each channel by sharing weights. Deconvolution operation can effectively map low-dimensional high-order features back to a higher-dimensional space, thereby restoring the structural information of the input data. Through the processing of the decoding layer, the information of features in the spatiotemporal dimension can be further revealed, and fusionable feature information can be provided for the subsequent channel fusion layer. The advantage of deconvolution is that it can not only perform upsampling operations, but also maintain the original temporal and spatial structure when restoring data.

[0042] The channel fusion layer is a key part of the fusion module. Its main function is to integrate the outputs of different channels to generate the final comprehensive ecological state vector. In this layer, through an attention weighting mechanism, the model can dynamically adjust the weights of different channels according to the importance of the input features of each channel. Specifically, the fusion layer first weights the output features of each channel and automatically adjusts the weight allocation according to the importance of each feature. This weighting process is adaptive. Based on the patterns learned by the network, the weights are automatically adjusted during training so that the most influential features can be given more weight in the restoration effect evaluation. Through weighted fusion, the model can comprehensively consider the influence of spatial, temporal, and environmental factors, providing a comprehensive reference for the prediction of restoration effect. Finally, the comprehensive ecological state vector output by the channel fusion layer will be used as the input of subsequent modules for further restoration effect evaluation. This vector integrates all spatiotemporal and environmental background features and is an efficient representation of the overall ecological health status of the restored area.

[0043] For example: Suppose we have three main ecological factors within a restoration area: vegetation index, soil moisture, and temperature. These factors possess spatial characteristics (geographical distribution), temporal characteristics (seasonal variation), and environmental background characteristics (climate change), respectively. After the input data passes through the feature branching layer, the vegetation index, soil moisture, and temperature enter different channels for independent processing. The encoding layer of each channel extracts the deep information of its respective features through convolution operations. Subsequently, the decoding layer restores the high-order features of each channel through deconvolution operations, ensuring that the information of each channel is fully preserved. In the channel fusion layer, the model automatically adjusts the weights of each channel through an attention weighting mechanism. For example, in certain time periods, soil moisture may have a greater impact on ecological restoration, so this channel will be assigned a higher weight. Finally, the weighted results of all channels are fused into a comprehensive ecological state vector, which is provided to the subsequent restoration effect evaluation module for prediction.

[0044] The joint model of Long Short-Term Memory (LSTM) and Adaptive Fuzzy Inference in S6 includes a bidirectional coupling structure between an LSTM temporal prediction unit and a fuzzy self-adjusting unit. The LSTM temporal prediction unit is responsible for extracting the temporal variation patterns of ecological factors, while the fuzzy self-adjusting unit describes the uncertainty of input features by calculating dynamic membership functions in real time and generates weight adjustment signals based on membership changes, which are then fed back to the LSTM gating control layer. During training, the fuzzy rule base automatically adjusts the membership function boundaries and weight distribution based on the prediction residuals, achieving synergistic optimization of fuzzy logic and temporal memory to improve the model's adaptive prediction capability for complex ecological processes. The LSTM temporal prediction unit and the fuzzy self-adjusting unit work together, adjusting the model's temporal prediction capability through mutual feedback. The LSTM unit is responsible for extracting temporal information from historical ecological data, while the fuzzy self-adjusting unit is responsible for adjusting the gating control of the LSTM unit through dynamic membership functions, optimizing the model's adaptive capability during ecological restoration.

[0045] The main task of the LSTM time series prediction unit is to extract the changing patterns of ecological factors over time. Ecological factors (such as soil moisture, vegetation growth, and climate conditions) are usually represented as time series data, which contain the changing patterns at different time steps. Through its internal long short-term memory units, LSTM can capture long-term dependencies in the time series, thereby accurately predicting the future ecological state. Specifically, at each time step, LSTM combines the current input data with the memory state of the previous time step, and uses gating mechanisms (including forget gate, input gate, and output gate) to determine how much historical information to retain, ultimately predicting the state at the next time step. This mechanism enables LSTM to efficiently learn the evolutionary patterns of ecological factors over time, especially in complex ecosystems with long-term dependencies.

[0046] The role of the fuzzy self-adjusting unit is to handle the uncertainty in ecological data by calculating a dynamic membership function to describe the fuzziness of the input features. Due to the complexity of ecosystems, there are uncertain relationships between many factors, and traditional models may struggle to accurately capture these relationships. However, fuzzy inference systems can effectively handle this fuzziness. The fuzzy self-adjusting unit reflects the uncertainty of features by calculating the membership function of the input data in real time. The membership function represents the degree of membership between data and a certain feature. For example, in the process of soil moisture change, the moisture change in some areas may show a clear trend, while other areas may experience large fluctuations. The fuzzy self-adjusting unit calculates the membership degree of each input feature through the membership function, thereby reflecting its influence on the model's prediction results.

[0047] After the fuzzy self-adjusting unit calculates the membership degree of each input feature, it generates a weight adjustment signal and feeds it back to the gating control layer of the LSTM. The purpose of this mechanism is to dynamically adjust the gating parameters of the LSTM according to the uncertainty of the input data, so that the model can more accurately capture important time dependencies. Specifically, if the uncertainty of ecological factors is high in certain periods, the gating layer of the LSTM unit will reduce the memory and influence of these data accordingly, thereby avoiding incorrect predictions. This dynamic adjustment mechanism ensures that the model can flexibly and adaptively adjust its prediction strategy when facing noise and uncertainty in ecological data, thereby improving the model's generalization ability and prediction accuracy.

[0048] During training, the fuzzy rule base automatically adjusts the boundaries and weight distribution of the membership function based on the prediction residuals. The prediction residuals refer to the differences between the model's predictions and the actual values. The fuzzy rule base uses these residuals to optimize the membership function in the fuzzy inference process, enabling the model to better adapt to changes in data. For example, in certain periods, changes in soil moisture may be significantly affected by other environmental factors, leading to prediction bias. By updating the boundaries of the membership function in real time, the fuzzy rule base enables the system to more accurately understand and respond to changes in different ecological factors, thereby improving prediction accuracy.

[0049] For example: Suppose we have the following main ecological factors in a certain ecological restoration area: soil moisture, vegetation growth index, and climate change data. The changes in these factors are influenced by a variety of environmental factors and exhibit obvious temporal characteristics.

[0050] In this context, the LSTM time-series prediction unit first extracts the temporal information of these factors and learns their change patterns at different points in time. For example, soil moisture may change rapidly in spring and summer, while it changes more gradually in autumn and winter. At the same time, the fuzzy self-adjusting unit calculates the membership degree of each factor based on the changes in the input features to reflect its influence under different environmental conditions. For example, in a dry season, there may be significant uncertainty between the changes in vegetation growth index and soil moisture. In this case, the fuzzy inference system will adjust the attention given to these factors by dynamically adjusting the membership function.

[0051] In S5 and S6, the integrated ecological state vector is normalized and denoised before output. The normalization adopts the hierarchical standard deviation constraint method to map ecological features of different dimensions to a unified scale. The denoising adopts the multi-scale wavelet threshold method to suppress high-frequency noise components, so as to reduce the interference of monitoring errors on the final prediction results.

[0052] The methods for using the stratified standard deviation constraint method include: The comprehensive ecological state vector is divided into three layers according to its source semantics: Layer A: spatial feature component (reflecting the neighborhood structure and spatial gradient of land parcels), Layer B: temporal feature component (reflecting the evolutionary trajectory under multiple time windows), and Layer C: environmental background component (reflecting climate, human intervention, and background constraints). The components within each layer are similar in statistical characteristics, and the standard deviation can be calculated within each layer to apply scale constraints to the components within that layer. The steps include: The scaling factor is estimated separately for each layer, with the standard deviation as the main factor. Alternatively, a scaling factor that is more robust to outliers can be used as an alternative to obtain a unified scaling benchmark within the layer. All components of the layer are linearly scaled according to this scale reference, so that the components within the layer fall into similar numerical ranges, avoiding the overwhelming of other information by a single component due to its excessively large unit or value range. To avoid the overall weakening or amplification of information between layers, interlayer constraints are introduced, including setting upper and lower bounds on the overall energy of the three layers (such as the sum of variances within each layer) to ensure that spatial, temporal, and background information remain comparable at a unified scale. For components with a large number of outliers, a lightweight and robust truncation is first performed without introducing a precise threshold, only limiting the influence range of extreme values, before participating in intra-layer scaling to ensure that normalization is not dominated by a small number of outliers.

[0053] After normalization, the normalized vector is subjected to multi-scale wavelet decomposition and thresholding on a component-by-component basis. A mother wavelet suitable for the smoothness of the ecological time series is selected, and detail coefficients and approximation coefficients are obtained at multiple decomposition levels. The noise level is estimated based on the highest frequency detail coefficients, and thresholding is applied to the detail coefficients at each scale (using a soft thresholding method to reduce ringing effects). The denoised components are then reconstructed.

[0054] For example: First, the scale is estimated and scaled separately for each of the three layers: the temporal component fluctuates significantly during wet and dry seasons, and the peaks and valleys are not excessively amplified after scaling within the layer; the spatial component is easily affected by individual extreme grids, so it is first robustly truncated and then scaled to avoid hotspot dominance; the background component is generally flat, and inter-layer constraints are used to ensure that it is not compressed to near zero; then wavelet multi-scale decomposition is performed: sensor noise and sampling jitter are reduced at the highest frequency scale, the threshold is carefully set at the mid-frequency scale to preserve seasonal rhythms, and the trend is basically preserved at the lowest frequency scale; when the reconstructed vectors are fed into the joint model for training, the gradient update is more stable and the feature contribution is more balanced.

[0055] In the restoration effect output stage, based on the joint feature space of graph convolutional network and LSTM model, an adaptive threshold segmentation and multiple aggregation mechanism are introduced. The adaptive threshold segmentation module uses statistical mean square error to dynamically determine the restoration level boundary. The multiple aggregation module uses a hierarchical weighted decision structure to fuse and calculate the time series indicators, spatial structure indicators and negative ecological weight results to obtain a multidimensional and consistent restoration effect evaluation result.

[0056] Example 2 like Figure 3 As shown, this embodiment further improves upon the design of Embodiment 1. The difference lies in the fact that, in the actual operation of Embodiment 1, it was found that when the ecological factor data of the restoration area exhibits uneven sampling intervals and large fluctuations in noise intensity, the joint model of long short-term memory and adaptive fuzzy inference is prone to local overfitting in the time-series prediction stage. This leads to a slight shift in the temporal continuity of the comprehensive evaluation results of the restoration effect, failing to fully reflect the true dynamic changes in the restoration process. Based on this, a method for evaluating the ecological restoration effect of abandoned coal mines further includes: S7. Introducing time-series dynamic smoothing and adaptive residual correction, including: After the joint model of long short-term memory and adaptive fuzzy inference (S6 step) outputs the comprehensive evaluation result of the repair effect, the comprehensive evaluation result of the repair effect is dynamically smoothed. By calculating the weighted moving average of the prediction results within the continuous time window, the impact of short-term disturbances on the overall prediction trend is reduced. At the same time, an adaptive residual correction mechanism is established to fine-tune the current model output using the prediction error of the previous period, so as to improve the model's performance in terms of temporal continuity and prediction stability.

[0057] The so-called time-series dynamic smoothing refers to weighted shifting and aggregation of the comprehensive ecological state sequence within a continuous time window to reduce the impact of short-term disturbances on trend characterization, while avoiding smoothing out true inflection points. To this end, this embodiment adopts a weighting system that adapts to time and uncertainty, including: The window length and sliding step size are consistent with the temporal granularity of the multidimensional time series matrix formed in stage S1. The window is not forcibly applied with a fixed length, but is adaptively adjusted based on the stability index of the most recent time period. The stability index consists of two parts: first, the confidence cues output by the S6 model (such as the opening degree of the gating unit and the dispersion of the fuzzy membership degree distribution); second, the fluctuation amplitude of ecological observations (such as vegetation index, surface temperature, and soil moisture) within the same window at the fusion scale. The window shrinks in the high fluctuation range and widens appropriately in the stable range. Two types of weighting factors are assigned to the state value at each moment within the window: time decay factor and uncertainty suppression factor. The time decay factor reflects the proximate cause priority and decreases monotonically with the time interval from the current moment. The uncertainty suppression factor comes from the membership degree entropy of the fuzzy inference in stage S6 and the variance information of the LSTM output. The larger the entropy or the higher the variance, the smaller the weight. The two types of factors are multiplied after normalization to obtain the final weight.

[0058] To prevent smoothing from weakening the real ecological transition, this embodiment introduces gradient sensitivity in the weight calculation, including: if there is a significant state gradient in adjacent time intervals within the window and it is consistent with the exogenous event indicated by the coupled feature set (the spatial and temporal correlation features obtained by S2 / S3), such as precipitation transition or disturbance operation log, then the cross-window weights that cross the gradient are reduced, and aggregation is only performed on one side of the gradient to retain the inflection point.

[0059] Example 3 like Figure 3 As shown, based on the same inventive concept as the method for evaluating the ecological restoration effect of abandoned coal mines in the foregoing embodiments, this application provides a system for evaluating the ecological restoration effect of abandoned coal mines. The system and method embodiments in this application are based on the same inventive concept. The system includes: Data Construction Module: The data construction module establishes a dataset of multi-source ecosystems in the restoration area. After spatial registration, temporal labeling, and anomaly removal, the dataset forms a multi-dimensional time series matrix. The time series analysis module performs time series consistency analysis on the time series matrix, uses a time delay correction algorithm based on mutual information to calculate the time series dependencies between ecological factors, and uses variable step size delay matching to correct time differences and generate a coupled feature set. Feature extraction module: The feature extraction module inputs the coupled feature set into the graph convolutional network model, where nodes are ecological factors and edge weights are temporal correlations, and extracts spatial and temporal correlation features through graph convolution; Contribution identification module: The contribution identification module constructs a negative ecological contribution module on the graph convolution output, uses a bidirectional attention mechanism to weighted identify negative ecological factors, outputs the weighted results, and optimizes the training through inverse gradient constraints; Multi-channel fusion module: The multi-channel fusion module inputs the associated features and weighted results into the fusion module, uses a multi-channel autoencoder structure for nonlinear fusion, and outputs a comprehensive ecological state vector; Fuzzy evaluation module: The fuzzy evaluation module inputs the comprehensive ecological state vector into the joint model of long short-term memory and adaptive fuzzy inference, adjusts the gating parameters and corrects the prediction through dynamic membership function, and obtains the comprehensive evaluation result of the restoration effect.

[0060] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0061] The above description is merely a preferred embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present application, based on the technical solution and concept of the present application, should be covered within the scope of protection of the present application.

Claims

1. A method for evaluating the ecological restoration effect of abandoned coal mines, characterized in that, The methods include: A dataset of multi-source ecosystems in the restoration area was established. After spatial registration, temporal labeling, and anomaly removal, the dataset was transformed into a multi-dimensional time-series matrix. Temporal consistency analysis is performed on the time series matrix. A time delay correction algorithm based on mutual information is used to calculate the temporal dependence between ecological factors. Variable step size delay matching is used to correct the time difference and generate a coupled feature set. The coupled feature set is input into a graph convolutional network model, where nodes are ecological factors and edge weights are temporal correlations. Spatial and temporal correlation features are extracted through graph convolution. A negative ecological contribution module is constructed on the graph convolution output. A bidirectional attention mechanism is used to identify negative ecological factors in a weighted manner, and the weighted result is output. The training is optimized by inverse gradient constraint. The associated features and weighted results are input into the fusion module, and nonlinear fusion is performed using a multi-channel autoencoder structure to output a comprehensive ecological state vector. The comprehensive ecological state vector is input into the joint model of long short-term memory and adaptive fuzzy inference. The gating parameters are adjusted and the prediction is corrected through dynamic membership function to obtain the comprehensive evaluation result of the restoration effect.

2. The method for evaluating the ecological restoration effect of abandoned coal mines according to claim 1, characterized in that, The time delay correction algorithm includes a mutual information entropy calculation module and a time synchronization correction module. The mutual information entropy calculation module is used to measure the correlation strength between ecological variables. The time synchronization correction module determines the time delay corresponding to the maximum mutual information based on a variable step-size window search algorithm, and performs synchronization alignment and interpolation compensation on multi-source time series within a sliding window to improve the temporal consistency between different data sources.

3. The method for evaluating the ecological restoration effect of abandoned coal mines according to claim 1, characterized in that, The graph convolutional network model includes spectral domain convolutional layers, temporal convolutional layers, and pooling layers; the spectral domain convolutional layers perform feature decomposition on the Laplacian matrix of the ecological factor graph to extract global correlations; the temporal convolutional layers employ a gated convolutional structure to capture multi-scale temporal dependent features; The pooling layer generates compressed graph feature embeddings through weighted averaging and downsampling operations, which are then used as input for subsequent modules.

4. The method for evaluating the ecological restoration effect of abandoned coal mines according to claim 1, characterized in that, In the negative ecological contribution module, the bidirectional attention layer includes a temporal attention sublayer and a spatial attention sublayer. The temporal attention sublayer calculates the influence intensity of each ecological factor in different time windows, and the spatial attention sublayer calculates the degree of spatial interference between factors. The negative weight matrix output by the negative ecological contribution module is regularized by the inverse gradient constraint unit and used to weight the corresponding factor components in the main feature set.

5. The method for evaluating the ecological restoration effect of abandoned coal mines according to claim 1, characterized in that, The multi-channel autoencoder structure of the fusion module includes: a feature branch layer, an encoding layer, a decoding layer, and a channel fusion layer; the feature branch layer divides the input data into three channels according to spatial features, temporal features, and environmental background features; the encoding layer independently extracts high-order nonlinear representations in each channel; the decoding layer reconstructs the data using deconvolution operations with shared weights; and the channel fusion layer fuses the outputs of each channel using an attention weighting mechanism to generate a comprehensive ecological state vector.

6. The method for evaluating the ecological restoration effect of abandoned coal mines according to claim 1, characterized in that, The joint model of long short-term memory and adaptive fuzzy inference includes a bidirectional coupling structure between a long short-term memory temporal prediction unit and a fuzzy self-adjusting unit. The long short-term memory temporal prediction unit is responsible for extracting the temporal change patterns of ecological factors, while the fuzzy self-adjusting unit describes the uncertainty of input features by calculating dynamic membership functions in real time and generates weight adjustment signals based on membership changes, which are then fed back to the gating control layer of the long short-term memory. During training, the fuzzy rule base automatically adjusts the membership function boundaries and weight distribution based on the prediction residuals, achieving synergistic optimization of fuzzy logic and temporal memory to improve the model's adaptive prediction capability for complex ecological processes.

7. The method for evaluating the ecological restoration effect of abandoned coal mines according to claim 1, characterized in that, The comprehensive ecological state vector is normalized and denoised before output. The normalization adopts the hierarchical standard deviation constraint method to map ecological characteristics of different dimensions to a unified scale. The noise reduction process employs a multi-scale wavelet thresholding method to suppress high-frequency noise components, thereby reducing the interference of monitoring errors on the final prediction results.

8. The method for evaluating the ecological restoration effect of abandoned coal mines according to claim 1, characterized in that, In the repair effect output stage, based on the joint feature space of graph convolutional network and LSTM model, an adaptive threshold segmentation and multiple aggregation mechanism are introduced. The adaptive threshold segmentation module uses statistical mean square error to dynamically determine the repair level boundary. The multi-aggregation module integrates time-series indicators, spatial structure indicators, and negative ecological weight results through a hierarchical weighted decision structure to obtain a multi-dimensional and consistent assessment result of the restoration effect.

9. The method for evaluating the ecological restoration effect of abandoned coal mines according to claim 1, characterized in that, The method also includes: Introducing time-series dynamic smoothing and adaptive residual correction, including: After the joint model of long short-term memory and adaptive fuzzy inference outputs the comprehensive evaluation result of the repair effect, the comprehensive evaluation result of the repair effect is dynamically smoothed. By calculating the weighted moving average of the prediction results within the continuous time window, the impact of short-term disturbances on the overall prediction trend is reduced. At the same time, an adaptive residual correction mechanism is established to fine-tune the current model output using the prediction error of the previous period, so as to improve the model's performance in terms of temporal continuity and prediction stability.

10. A multimodal fusion intelligent question-answering and knowledge retrieval system, characterized in that, The system includes: The data construction module establishes a dataset of multi-source ecosystems in the restoration area. After spatial registration, temporal labeling, and anomaly removal, the dataset forms a multi-dimensional time series matrix. The time series analysis module performs time series consistency analysis on the time series matrix, uses a time delay correction algorithm based on mutual information to calculate the time series dependence between ecological factors, and uses variable step size delay matching to correct time differences and generate a coupled feature set. The feature extraction module inputs the coupled feature set into the graph convolutional network model, where nodes are ecological factors and edge weights are temporal correlations, and extracts spatial and temporal correlation features through graph convolution. The contribution identification module constructs a negative ecological contribution module on the graph convolution output, uses a bidirectional attention mechanism to weightedly identify negative ecological factors, outputs the weighted result, and optimizes the training through inverse gradient constraints. A multi-channel fusion module inputs the associated features and weighted results into the fusion module, performs nonlinear fusion using a multi-channel autoencoder structure, and outputs a comprehensive ecological state vector. The fuzzy evaluation module inputs the comprehensive ecological state vector into the joint model of long short-term memory and adaptive fuzzy inference, adjusts the gating parameters and corrects the prediction through a dynamic membership function, and obtains a comprehensive evaluation result of the restoration effect.

Images