Non-ferrous metal mining area karst underground water analysis method based on deep machine learning

Abstract

The invention relates to the technical field of artificial intelligence, and provides a non-ferrous metal mining area karst groundwater analysis method based on deep machine learning, and the method comprises the steps: obtaining a multi-source time sequence of a source mining area in a target time period, and combining the multi-source time sequence into an input-output sample pair; constructing an LSTM analysis model of the source mining area, and performing hyper-parameter optimization on the LSTM analysis model by using a Bayesian optimization algorithm to obtain an optimized LSTM model; using the input-output sample pair to train the optimized LSTM model to obtain an initial groundwater analysis model; migrating the initial groundwater analysis model from the source mining area to a target mining area by adopting a migration learning method to obtain a final groundwater analysis model matched with the target mining area; and outputting an underground water analysis result corresponding to the real-time rainfall of the target mining area through the final underground water analysis model. The method can overcome the defects that an existing karst groundwater dynamic prediction model must be trained on a large scale again after a mining area is replaced, and the generalization ability is weak.

Description

Technical Field

[0001] This invention relates to a method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning, belonging to the field of artificial intelligence technology. Background Technology

[0002] Most non-ferrous metal deposits are located in carbonate rocks rich in groundwater. Water hazards are a significant factor affecting their safe mining. Large-scale dewatering during mining alters the groundwater recharge, runoff, and drainage systems to varying degrees. Blind mining can easily lead to mine flooding accidents, seriously threatening mining safety. Furthermore, a drop in groundwater levels not only depletes underground resources in adjacent aquifers and interrupts surface water flow, severely impacting people's water supply for production and daily life, but may also induce karst ground subsidence, causing serious environmental and geological problems. Therefore, taking a typical non-ferrous metal mine as an example, fully utilizing hydrogeological survey and monitoring data of the mining area to analyze the spatial characteristics and evolution of aquifers at different mining stages, and predicting the dynamic changes in groundwater levels and inflow in the mining area, is of great significance for water hazard prevention and geological environmental protection in karst metal mining areas of our province.

[0003] Traditional methods for predicting dynamic karst hydrology primarily rely on numerical simulation of groundwater. This approach demands high-quality hydrogeological data, particularly quantitative analysis of karst hydrogeological parameters. However, the hydrogeological conditions in mining areas are often influenced by both regional geological background and human activities. Local hydrogeological parameters struggle to accurately reflect the spatial variability of the karst water system within the mining area. Consequently, the accuracy of traditional numerical simulations is subject to uncertainty. In recent years, the rapid development of artificial intelligence (AI) technology has led to the emergence of information-based AI prediction models. These models, based on big data-driven digital twin technology, offer high prediction accuracy and have been widely applied in environmental and hydrological fields. Research on AI models for dynamic prediction of karst water in non-ferrous metal mines responds to the national AI development strategy. By intelligently processing karst water monitoring data, multiple sets of spatiotemporal measured data, including karst water level, temperature, flow rate, and rainfall, are combined with AI. This approach has low dependence on fundamental karst hydrogeological parameters, significantly reducing time and economic costs. Furthermore, it enriches the innovative application system of AI and is of great significance for promoting the deep integration of AI with the hydrogeological and environmental fields, fostering data-driven, human-machine collaborative intelligent applications.

[0004] The karst aquifer in the mining area is composed of multiple media, including conduits, fissures, and pores. Its hydraulic conductivity exhibits abrupt changes in space and time, and its flow rate and head show strong randomness. Differences in lithology, structure, and karst development make the groundwater distribution far more uneven than that of pore or fissure water. It also undergoes rapid conversion with surface water and significant fluctuations over time. Domestic and international studies generally treat such aquifers as random media and employ methods such as Monte Carlo, stochastic inversion, Kalman filtering, and statistical moment equations to characterize parameter uncertainties. However, there are disagreements regarding the flow properties; some studies believe... To address the failure of Darcy's law due to fractured channel flow, it is argued that pipe flow, diffuse flow, and Darcy flow can coexist within the same system, leading to the proposal of triple or quadruple media models. Some studies have also verified the existence of non-Darcy flow through non-Darcy conversion of permeability coefficients, dual-media coupling networks, or experimental methods. In contrast, resource evaluation in uniform fractured karst areas in northern China can be conducted using a unified water table and Darcy's law. In some provinces' metal mines, tunnel excavation creates artificial water collection channels and potential sinks, resulting in a sharp increase in the near-field hydraulic gradient, which easily leads to non-Darcy flow. In areas with small far-field disturbances and relatively uniform fractures, Darcy flow characteristics can still be maintained.

[0005] Karst aquifers are inherently heterogeneous, anisotropic, and exhibit both laminar and turbulent flow. Their circulation paths are neither simple lines nor single-surface structures. Traditional finite difference or finite element methods often fall short in such environments. With the advancement of computers, numerical methods and supporting software such as characteristic finite element method, stochastic finite element method, multi-medium coupling, discrete fracture network, confluence calculation, dual-triple-quadruple medium models, continuous-discontinuous coupling, and GIS-3D visualization have emerged. Simulation has progressed from two-dimensional to three-dimensional, from single-medium to multi-medium, from black box to distributed, and from conceptual to real-time visualization. The current recognized challenge is how to improve simulation accuracy to a usable level while keeping data requirements manageable.

[0006] Due to the strong adaptability of stochastic methods, machine learning has rapidly become a hot topic in model research in the fields of groundwater resource evaluation and management and mine water hazard prevention in the past two or three decades. Commonly used stochastic model methods in the field of hydrogeology mainly include: regression model, grey model, DM(n,h) model, neural network model, and time series model.

[0007] In recent years, artificial intelligence models have been widely used in the study of groundwater flow and the spatiotemporal distribution of pollutants due to their significant advantages in data analysis and prediction, including high accuracy, low cost, and high efficiency. For example, algorithms such as K-means clustering, random forest, and support vector machine (SVM) have successfully predicted regional groundwater levels; convolutional neural network-long short-term memory (CNN-LSTM) and deep learning models based on spatiotemporal attention mechanisms have been used for real-time monitoring of groundwater flow changes; artificial neural networks and support vector machines have been applied to groundwater modeling; and hybrid methods of adaptive neural fuzzy inference system (ANFIS) and decision tree (DT) algorithms have been used to accurately predict groundwater quality indices. These models provide strong technical support for the rational utilization and protection of water resources.

[0008] However, due to their unique geological structure and hydrological characteristics, karst water systems exhibit significant heterogeneity in groundwater flow and solute migration processes (WRR karst water system literature), making dynamic prediction more challenging. RNNs can handle long sequences with strong temporal correlations, thus they are most widely used in karst water dynamic prediction. However, RNNs encounter gradient vanishing or exploding problems, limiting their ability to learn long-term dependencies. The gating mechanism of LSTM effectively alleviates this problem and has been successfully applied to simulate the dynamic prediction of karst groundwater pipe flow in the Ribeaucourt system in the Meuse region of France. By combining a long short-term memory encoder-decoder neural network with a Savitzky-Golay filter, Bi, Jing, et al. developed a sequence-enhanced long short-term memory model, which demonstrated high accuracy and reliability on time-series data of dissolved oxygen levels and permanganate content in karst water. Although these machine learning algorithms have achieved relatively satisfactory results, their detection efficiency remains poor. In particular, the generalization ability of these models is weak, which limits their application in karst hydrological dynamic prediction tasks.

[0009] Therefore, existing dynamic prediction models for karst groundwater must be retrained on a large scale after a change of mining area, and their generalization ability is weak. Summary of the Invention

[0010] This invention provides a method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning. Its main purpose is to overcome the shortcomings of existing dynamic prediction models for karst groundwater, which require large-scale retraining after changing mining areas and have weak generalization ability.

[0011] To achieve the above objectives, the present invention provides a method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning, comprising: Obtain multi-source time series of the source mining area within the target time period, and combine the multi-source time series into input-output sample pairs, wherein the multi-source time series includes rainfall, groundwater level, water temperature and conductivity; An LSTM analysis model of the source mining area is constructed, and the hyperparameters of the LSTM analysis model are optimized using the Bayesian optimization algorithm to obtain the optimized LSTM model. The hyperparameters of the LSTM analysis model include the number of LSTM layers, learning rate, number of training rounds, batch size, and number of hidden units. The optimized LSTM model is trained using the input-output sample pair to obtain the initial groundwater analysis model; The initial groundwater analysis model is transferred from the source mining area to the target mining area using the transfer learning method, resulting in a final groundwater analysis model that is adapted to the target mining area. The final groundwater analysis model outputs the groundwater analysis results corresponding to the real-time rainfall in the target mining area.

[0012] Optionally, the hyperparameters of the LSTM analysis model are optimized using a Bayesian optimization algorithm to obtain an optimized LSTM model, including: Obtain the validation set from the input-output sample pairs; Based on the validation set, the hyperparameters of the LSTM analysis model are iteratively optimized using the Bayesian optimization algorithm; The combination of hyperparameters corresponding to the minimum average relative error obtained from each observation after reaching the preset number of iterations is taken as the optimal hyperparameters. The LSTM analysis model is reconstructed using the optimal hyperparameters to obtain the optimized LSTM model.

[0013] Optionally, based on the validation set, the hyperparameters of the LSTM analysis model are iteratively optimized using a Bayesian optimization algorithm, including: Multiple sets of hyperparameters are initialized within the search space of the hyperparameters of the LSTM analysis model, and the LSTM analysis model is trained group by group on the validation set using the multiple sets of hyperparameters. Record the average relative error of the LSTM analysis model output with respect to the validation set; Based on the average relative error of the validation set, the hyperparameters of the LSTM analysis model are iteratively optimized using a Bayesian optimization algorithm.

[0014] Optional Bayesian optimization algorithms include: A Gaussian process surrogate model is constructed using hyperparameters as input variables and average relative error as the objective function value. The posterior distribution of the Gaussian process surrogate model is updated using the observation samples formed by the hyperparameters and the corresponding average relative errors, so as to select the next set of hyperparameters through the acquisition function; The next set of hyperparameters is substituted into the LSTM analysis model to obtain a new average relative error, until the preset number of iterations is reached; After reaching the preset number of iterations, the process of iteratively optimizing the hyperparameters of the LSTM analysis model is completed.

[0015] Optionally, an LSTM analysis model of the source mining area is constructed, including: The number of grid outputs is set according to the number of monitoring stations in the source mining area, so that each grid outputs the water level analysis value of the corresponding monitoring station; Establish an LSTM analysis model on a general computing platform.

[0016] Optionally, the LSTM analysis model includes an LSTM layer, an activation function, and a fully connected layer. The LSTM layer consists of multiple neural units, each of which includes a forget gate, an input gate, and an output gate. The number of output nodes in the fully connected layer is configured to be equal to the number of grid outputs.

[0017] Optionally, a transfer learning method is used to transfer the initial groundwater analysis model from the source mining area to the target mining area to obtain a final groundwater analysis model adapted to the target mining area, including: After transferring the initial groundwater analysis model to the target mining area using the transfer learning method, the initial groundwater analysis model is fine-tuned to obtain the fine-tuned model. The fine-tuned model is used for forward calculation to obtain the predicted water level map; Based on the predicted water level map, the final groundwater analysis model is determined.

[0018] Optionally, based on the predicted water level map, a final groundwater analysis model is determined, including: Calculate the depth error, correlation, similarity, and probability distribution overlap between the predicted water level map and the measured water level map; The final groundwater analysis model is determined based on the depth error, the correlation, the similarity, and the overlap of the probability distributions.

[0019] Optionally, the depth error, correlation, similarity, and probability distribution overlap between the predicted water level map and the measured water level map are calculated, including: The depth error between the predicted water level map and the measured water level map is calculated using the average relative error method. The correlation and similarity between the predicted water level map and the measured water level map were analyzed using the two-dimensional correlation coefficient method and the structural similarity method. The probability distribution overlap between the predicted water level map and the measured water level map was analyzed using the Batachalia distance method and the histogram intersection distance method.

[0020] To address the aforementioned problems, this invention also provides a karst groundwater analysis system for non-ferrous metal mining areas based on deep machine learning, the system comprising: The sample combination module is used to acquire multi-source time series of the source mining area within the target time period and combine the multi-source time series into input-output sample pairs. The multi-source time series includes rainfall, groundwater level, water temperature and conductivity. The parameter optimization module is used to construct the LSTM analysis model of the source mining area and use the Bayesian optimization algorithm to optimize the hyperparameters of the LSTM analysis model to obtain the optimized LSTM model. The hyperparameters of the LSTM analysis model include the number of LSTM layers, learning rate, number of training rounds, batch size, and number of hidden units. The model training module is used to train the optimized LSTM model using the input-output sample pair to obtain the initial groundwater analysis model. The model transfer module is used to transfer the initial groundwater analysis model from the source mining area to the target mining area using the transfer learning method, so as to obtain a final groundwater analysis model that is adapted to the target mining area. The groundwater analysis module is used to output the groundwater analysis results corresponding to the real-time rainfall in the target mining area through the final groundwater analysis model.

[0021] Compared to the problems described in the background art, this embodiment of the invention acquires multi-source time series data from the source mining area within a target time period and combines these multi-source time series data into input-output sample pairs to collect data that may affect groundwater levels. After analyzing this data, factors influencing groundwater levels can be selected for subsequent prediction. This embodiment of the invention constructs an LSTM analysis model of the source mining area to build a model that can be used to predict groundwater levels. Furthermore, this embodiment of the invention utilizes a Bayesian optimization algorithm to optimize the hyperparameters of the LSTM analysis model, automatically finding the optimal number of layers, learning rate, batch size, and other hyperparameters in one step, eliminating the need for repeated manual calculations and reducing... To minimize parameter tuning time and reduce validation set error while ensuring model prediction accuracy, this invention employs input-output samples to train and optimize the LSTM model under optimal hyperparameters provided by Bayesian optimization. The resulting initial groundwater analysis model achieves minimum mean relative error on the validation set of the source mining area, demonstrating stable and high-precision water level prediction capabilities. This provides directly reusable high-quality initial weights for subsequent migration to the target mining area, significantly shortening the overall modeling cycle. Furthermore, this invention utilizes transfer learning to migrate the initial groundwater analysis model from the source mining area to the target mining area, overcoming the shortcomings of existing karst groundwater dynamic prediction models that require large-scale retraining after changing mining areas and exhibit weak generalization ability. Therefore, this invention overcomes the shortcomings of existing karst groundwater dynamic prediction models that require large-scale retraining after changing mining areas and exhibit weak generalization ability. Attached Figure Description

[0022] Figure 1 A flowchart illustrating a method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning, provided in an embodiment of the present invention. Figure 2 A schematic diagram illustrating the implementation steps of the deep machine learning-based karst groundwater analysis method for non-ferrous metal mining areas, provided as an embodiment of the present invention; Figure 3 A schematic diagram of an LSTM network for implementing the deep machine learning-based karst groundwater analysis method in non-ferrous metal mining areas, provided as an embodiment of the present invention; Figure 4 A schematic diagram of a single neural unit for implementing the deep machine learning-based karst groundwater analysis method in non-ferrous metal mining areas, provided in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the Bayesian optimization of the deep machine learning-based karst groundwater analysis method for non-ferrous metal mining areas, provided as an embodiment of the present invention. Figure 6A schematic diagram illustrating the transfer learning technique for implementing the deep machine learning-based karst groundwater analysis method in non-ferrous metal mining areas, as provided in an embodiment of the present invention. Figure 7 A schematic diagram of the modules for implementing the deep machine learning-based karst groundwater analysis method in non-ferrous metal mining areas, provided as an embodiment of the present invention; Figure 8 A schematic diagram of a computer device for a method of analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning, provided in an embodiment of the present invention. The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0024] This application provides a method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning. The execution entity of this method includes, but is not limited to, at least one electronic device that can be configured to execute the method provided in this application, such as a server or a terminal. In other words, the method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0025] Reference Figure 1 The diagram shown is a flowchart illustrating a method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning, according to an embodiment of the present invention. In this embodiment, the method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning includes: S1. Obtain multi-source time series of the source mining area within the target time period, and combine the multi-source time series into input-output sample pairs, wherein the multi-source time series includes rainfall, groundwater level, water temperature and conductivity.

[0026] This invention provides an embodiment of the invention that acquires multi-source time series data of a source mining area within a target time period and combines these multi-source time series data into input-output sample pairs to collect data that may affect groundwater levels. After analyzing these data, factors that can be used to predict groundwater levels are then selected.

[0027] In one embodiment of the present invention, combining the multi-source time series into input-output sample pairs includes: analyzing the spatiotemporal correlation between rainfall, water temperature, and conductivity in the multi-source time series and groundwater level; screening key influencing factors affecting the groundwater level from the rainfall, water temperature, and conductivity based on the spatiotemporal correlation; and combining the key influencing factors with the groundwater level into input-output sample pairs.

[0028] The multi-source time series refers to parallel measured data sequences of rainfall, water temperature, electrical conductivity, groundwater level, etc., acquired from the same monitoring network within the same time period. The spatiotemporal correlation refers to the strength of the correlation and lag relationship between variables at spatial locations and time nodes. The input-output sample pair refers to the training sample composed of the time series of key influencing factors as model input and the groundwater level at the corresponding time as the expected output.

[0029] For example, the process of analyzing the spatiotemporal correlation between rainfall, water temperature, and conductivity in the multi-source time series and groundwater level is as follows: After aligning rainfall, water temperature, conductivity, and groundwater level with the same spatiotemporal resolution, the spatiotemporal correlation can be calculated using existing technologies, including but not limited to Pearson correlation coefficient, Spearman correlation coefficient, etc. Furthermore, the process of screening key influencing factors affecting groundwater level from rainfall, water temperature, and conductivity based on the spatiotemporal correlation is as follows: Set a threshold for the absolute value of the correlation coefficient, retain factors greater than the threshold, remove water temperature and conductivity with low correlation, and use rainfall as the final input variable.

[0030] See Figure 2 The diagram shown illustrates the implementation steps of the deep machine learning-based karst groundwater analysis method for non-ferrous metal mining areas, according to an embodiment of the present invention. Figure 2 In this model, the rainfall time series is used as the only input. The multi-layer LSTM network outputs a predicted water level map under real-time parameter tuning by the Bayesian optimizer. The relative error between the predicted map and the measured water level (GT) map is calculated pixel by pixel. The error is fed back to the optimizer to update the hyperparameters. The process is repeated until the verification error is minimized, and the optimal LSTM model is obtained.

[0031] S2. Construct an LSTM analysis model for the source mining area, and use the Bayesian optimization algorithm to optimize the hyperparameters of the LSTM analysis model to obtain an optimized LSTM model. The hyperparameters of the LSTM analysis model include the number of LSTM layers, learning rate, number of training rounds, batch size, and number of hidden units.

[0032] This invention provides an embodiment of an LSTM analysis model for the source mining area to construct a model that can be used to predict groundwater levels.

[0033] In one embodiment of the present invention, constructing the LSTM analysis model of the source mining area includes: setting the corresponding number of grid outputs based on the number of monitoring stations in the source mining area, such that each grid outputs the water level analysis value corresponding to the monitoring station; establishing an LSTM analysis model on a general computing platform, wherein the LSTM analysis model includes an LSTM layer, an activation function, and a fully connected layer, the LSTM layer is composed of multiple neural units, each neural unit includes a forget gate, an input gate, and an output gate, and the number of output nodes of the fully connected layer is configured to be equal to the number of grid outputs.

[0034] The general-purpose computing platform refers to a scalable hardware-software environment capable of running numerical computation, algorithm development, and model training. The general-purpose computing platform includes, but is not limited to, general-purpose scientific computing environments such as MATLAB 2021a (MathWorks Inc., Natick, MA, US) and Python.

[0035] See Figure 3 The diagram shown is a schematic representation of an LSTM network used in an embodiment of the present invention to implement the deep machine learning-based method for analyzing karst groundwater in non-ferrous metal mining areas. Figure 3 In this context, the LSTM analysis model includes LSTM layers, activation functions, and fully connected layers.

[0036] See Figure 4 The diagram shown is a schematic representation of a single neural unit implementing the deep machine learning-based karst groundwater analysis method for non-ferrous metal mining areas, according to an embodiment of the present invention. Figure 4 In this LSTM layer, multiple neural units are composed of each unit, including a forget gate, an input gate, and an output gate. The number of output nodes in the fully connected layer is configured to be equal to the number of grid outputs. The forget gate determines how many unit states are retained at time (t-1) until time (t). The input gate determines the update of the unit state. The output of the LSTM neural unit state is determined by a nonlinear activation function and the output gate. Generally, an input (x) passes through a neural unit to obtain an output (h). Specifically, the computation process of a single LSTM neural unit is as follows:

[0037]

[0038]

[0039]

[0040]

[0041]

[0042] In the formula, It is the output of the forget gate. and It consists of the weight matrix and bias of the forget gate, and and It is the output (time(t-1)) of the previous neuron and the current input (time(t)). It is the output of the input gate, and and These represent the unit states of the current input and the current time, respectively. It is the output of the output gate, and It is the output of a neural unit over time (t).

[0043] Furthermore, in this embodiment of the invention, the hyperparameters of the LSTM analysis model are optimized using the Bayesian optimization algorithm. The optimal number of layers, learning rate, batch size, and other hyperparameters are automatically found in one go using Bayesian optimization, eliminating the need for repeated manual calculations, reducing parameter tuning time, and minimizing validation set error to ensure the prediction accuracy of the model.

[0044] In one embodiment of the present invention, the step of optimizing the hyperparameters of the LSTM analysis model using a Bayesian optimization algorithm to obtain an optimized LSTM model includes: obtaining a validation set from the input-output sample pairs; iteratively optimizing the hyperparameters of the LSTM analysis model using a Bayesian optimization algorithm based on the validation set; taking the combination of hyperparameters corresponding to the minimum average relative error obtained from each observation after reaching a preset number of iterations as the optimal hyperparameters; and reconstructing the LSTM analysis model using the optimal hyperparameters to obtain the optimized LSTM model.

[0045] In another embodiment of the present invention, the step of iteratively optimizing the hyperparameters of the LSTM analysis model using a Bayesian optimization algorithm based on the validation set includes: initializing multiple sets of hyperparameters in the search space of the hyperparameters of the LSTM analysis model, and training the LSTM analysis model group by group on the validation set using the multiple sets of hyperparameters; recording the average relative error of the LSTM analysis model output with respect to the validation set; and iteratively optimizing the hyperparameters of the LSTM analysis model using a Bayesian optimization algorithm based on the average relative error of the validation set.

[0046] In another embodiment of the present invention, the Bayesian optimization algorithm includes: constructing a Gaussian process surrogate model with hyperparameters as input variables and average relative error as the objective function value; updating the posterior distribution of the Gaussian process surrogate model using the observation samples formed by the hyperparameters and the corresponding average relative error, so as to select the next set of hyperparameters through a collection function; substituting the next set of hyperparameters into the LSTM analysis model to obtain a new average relative error, until a preset number of iterations is reached; and completing the iterative optimization process of the hyperparameters of the LSTM analysis model after reaching the preset number of iterations.

[0047] It should be noted that the acquisition function mentioned above refers to the following: Figure 5 The get function that appears in the text.

[0048] See Figure 5 The diagram shown illustrates a Bayesian optimization implementation of the deep machine learning-based karst groundwater analysis method for non-ferrous metal mining areas, according to an embodiment of the present invention. Figure 5 One problem with the aforementioned LSTM network is that its structure layers, learning rate, number of training epochs, mini-batch size, and number of neurons are all unknown. Manually selecting and fine-tuning these hyperparameters from scratch can be extremely difficult and time-consuming. Bayesian optimization (BO) is an algorithm that can automatically search for the optimal combination of hyperparameters. BO is a continuously updated probabilistic model that assumes that the probability of event B occurring in the prior condition is proportional to the probability of event B occurring in the posterior condition. That is, for consecutively occurring events, the probability of the latter event is proportional to the probability of all previous events. This is a potential hyperparameter optimization scheme, meaning that the most likely parameter combination is inferred through multiple prior trials (i.e., training network models with different structures) by adjusting the objective function. Multiple evaluations are used to update the posterior probability of the optimization function to obtain the optimal parameter combination, which can provide a reference for subsequent model attempts. Based on prior conditions (i.e., historical evaluation records, the average relative error in the network models tried in this project), the algorithm makes full use of previous evaluation information when selecting the next set of parameter combinations, reducing the parameter search time. Specifically, multiple search ranges for hyperparameters are designed, and the BO algorithm automatically selects values ​​from the search range, continuously trying network models with different structures, and then recording the error. In this project, the hyperparameters to be optimized include the number of LSTM layers, learning rate, epoch rate, mini-batch size, and number of hidden units. The search ranges for these five parameters are set to [1-5], [10-4-1], [0-600], [0-100], and [0-100], respectively. Finally, BO infers the possible optimal network combination based on historical error information. The selection process is shown in the following equation:

[0049]

[0050] In the formula, and These are the posterior probability and prior probability of event A, respectively. It is the probability of observation points obtained from previous events. It is the objective function (i.e., the average relative error). It is the optimal parameter combination, and It refers to the range of values ​​for the parameters. Specifically, (1) within the range of hyperparameter values, randomly select a combination of hyperparameters. (e.g., MaxEpochs and learning rate) a set, (2) will The input is fed into the network for training to obtain the corresponding objective function. (3) Using all input values ​​(x, f(x)), the probability distribution of f(x) corresponding to x is calculated and predicted through a Gaussian process. (4) The optimal x is determined by the acquisition function in the probability distribution. (5) The x obtained in step (4) is used as the hyperparameter combination of the network for training and calculation of the objective function f(x). (6) Before reaching the maximum number of iterations, the (x, f(x)) obtained in step (5) is used as the input of the Gaussian process to continuously update the probability model and obtain new (x, f(x)). Once the maximum number of iterations is reached, the x corresponding to the minimum value of f(x) is taken as the optimal hyperparameter combination. .

[0051] S3. Use the input-output sample pair to train the optimized LSTM model to obtain the initial groundwater analysis model.

[0052] The embodiments of the present invention, by using input-output samples to train and optimize the LSTM model under the optimal hyperparameters given by Bayesian optimization, have obtained an initial groundwater analysis model that has achieved the minimum average relative error on the validation set of the source mining area. It has stable and high-precision water level prediction capabilities, and provides high-quality initial weights that can be directly reused for subsequent migration to the target mining area, significantly shortening the overall modeling cycle.

[0053] In one embodiment of the present invention, training the optimized LSTM model using the input-output sample pairs to obtain an initial groundwater analysis model includes: feeding the training set from the input-output sample pairs into the optimized LSTM model in batches; using an adaptive moment estimation optimizer to backpropagate the optimized LSTM model based on the loss function value output by the optimized LSTM model with respect to the training set, so as to update the network weights of the optimized LSTM model; and determining the initial groundwater analysis model by using the updated network weights corresponding to the final training round after reaching a preset number of training rounds.

[0054] The Adaptive Moment Estimator (Adam) is a first-order gradient optimization algorithm that dynamically adjusts the learning rate of each parameter by calculating the first moment (mean) and second moment (uncentered variance) of the gradient. It features adaptive step size, bias correction, and momentum acceleration, thereby speeding up convergence, alleviating gradient sparsity problems, and reducing sensitivity to the initial learning rate value. It is suitable for training deep neural networks. The calculation method of the Adaptive Moment Estimator is as follows:

[0055]

[0056]

[0057] In the formula, and These are the gradient decay factor (0.9) and the squared gradient decay factor (0.999), respectively. E(θ) is the loss function, m and v are momentum terms, ε = 10⁻⁸, n is the number of samples, and θ is the network weight.

[0058] Optionally, the loss function value of the optimized LSTM model output with respect to the training set is calculated as follows:

[0059] In the formula, and These are the predicted results and the actual results, respectively.

[0060] In one embodiment of the present invention, after training the optimized LSTM model using the input-output samples to obtain the initial groundwater analysis model, the method further includes: obtaining three sets of predicted water level maps output by the optimized LSTM model, the artificial neural network, and the convolutional neural network on the same validation set of the input-output samples; and evaluating the consistency between the three sets of predicted water level maps and the measured water level maps using the average relative error method, the two-dimensional correlation coefficient method, the structural similarity method, the Patacharian distance method, and the histogram intersection distance method to determine the effectiveness of the optimized LSTM model.

[0061] Optionally, the process of evaluating the consistency between the predicted and measured water level maps of the three groups using the average relative error method, two-dimensional correlation coefficient method, structural similarity method, Patacharian distance method, and histogram intersection distance method to determine the effectiveness of the optimized LSTM model is as follows: if the optimized LSTM model outperforms artificial neural networks and convolutional neural networks in all five indicators (average relative error method, two-dimensional correlation coefficient method, structural similarity method, Patacharian distance method, and histogram intersection distance method), then the optimized LSTM model is effective.

[0062] S4. The initial groundwater analysis model is transferred from the source mining area to the target mining area using the transfer learning method to obtain the final groundwater analysis model that is adapted to the target mining area.

[0063] This invention employs a transfer learning method to transfer the initial groundwater analysis model from the source mining area to the target mining area, thereby overcoming the shortcomings of existing karst groundwater dynamic prediction models, which require large-scale retraining after changing mining areas and have weak generalization ability.

[0064] In one embodiment of the present invention, the step of using transfer learning to transfer the initial groundwater analysis model from the source mining area to the target mining area to obtain a final groundwater analysis model adapted to the target mining area includes: after transferring the initial groundwater analysis model to the target mining area using transfer learning, fine-tuning the initial groundwater analysis model to obtain a fine-tuned model; performing forward computation on the fine-tuned model to obtain a predicted water level map; calculating the depth error between the predicted water level map and the measured water level map using the average relative error method; analyzing the correlation and similarity between the predicted water level map and the measured water level map using the two-dimensional correlation coefficient method and the structural similarity method; analyzing the probability distribution overlap between the predicted water level map and the measured water level map using the Batachalia distance method and the histogram intersection distance method; and determining the final groundwater analysis model based on the depth error, the correlation, the similarity, and the probability distribution overlap.

[0065] It should be noted that the calculation methods for the depth error, correlation, similarity, and probability distribution overlap mentioned above are as follows:

[0066]

[0067]

[0068]

[0069]

[0070] In the formula, and denoted as the average pixel values ​​of images I and J, respectively. µI, µJ, σI, σJ, and σIJ represent the local mean, standard deviation, and cross-covariance of pixels in images I and J, respectively. C1 and C2 are 6.5 and 58.5, respectively. p(x) and q(x) represent the probability distributions of pixels in images I and J, respectively. X is the structural domain of p(x) and q(x). PR represents the predicted water level map, GT represents the measured water level map, Mre represents the depth error, 2D-CC represents the correlation calculated by the two-dimensional correlation coefficient method, SS represents the similarity calculated by the structural similarity method, BD represents the overlap of the first probability distribution calculated by the Batachalia distance method, and HID represents the overlap of the second probability distribution calculated by the histogram intersection distance method.

[0071] Optionally, the process of performing forward calculation on the fine-tuned model is implemented using test data from the target mining area that was not used in the training. Further, the process of determining the final groundwater analysis model based on the depth error, the correlation, the similarity, and the probability distribution overlap is as follows: if all five indicators meet the preset consistency threshold, the fine-tuned model is considered to have met the accuracy standard in the target scenario and can be directly determined as the final groundwater analysis model; otherwise, fine-tuning continues until the standard is met.

[0072] See Figure 6 The diagram shown is a schematic representation of the transfer learning technique used in an embodiment of the deep machine learning-based karst groundwater analysis method for non-ferrous metal mining areas, according to an embodiment of the present invention. Figure 6 One of the main challenges of data-driven models is their compatibility, as the model developed in this study seemed only applicable to the locations where the water levels were investigated. This problem can be addressed by using transfer learning (TL) techniques to achieve dynamic prediction of karst hydrology at new sites. TL, or learning from experience, can significantly expand the application scope of intelligent algorithms. TL is a method of transferring knowledge from one domain (source domain) to another (target domain). By training a source model (pre-trained network) using source data (water level monitoring point A), the pre-trained network can acquire strong feature extraction capabilities in similar tasks. Subsequently, through fine-tuning on new data (water level monitoring point B), the pre-trained network can quickly adapt to new sites in different scenarios. The pre-trained network serves as the target model. This method can save a significant amount of training time for the target domain (new water level monitoring point) and achieve better training results, especially when the training samples in the target domain are limited. In this project, TL is used to transfer the LSTM network obtained from the current site to a second case study site with data from the new site to expand the compatibility and generalization ability of the proposed method.

[0073] S5. Output the groundwater analysis results corresponding to the real-time rainfall in the target mining area through the final groundwater analysis model.

[0074] like Figure 7 The diagram shown is a functional module diagram of the non-ferrous metal mining area karst groundwater analysis system based on deep machine learning, according to the present invention.

[0075] The deep machine learning-based karst groundwater analysis system 700 for non-ferrous metal mining areas described in this invention can be installed in an electronic device. Depending on the functions implemented, the deep machine learning-based karst groundwater analysis system for non-ferrous metal mining areas includes a sample combination module 701, a parameter optimization module 702, a model training module 703, a model transfer module 704, and a groundwater analysis module 705. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.

[0076] In this embodiment of the invention, the functions of each module / unit are as follows: The sample combination module 701 is used to acquire multi-source time series of the source mining area within the target time period and combine the multi-source time series into input-output sample pairs, wherein the multi-source time series includes rainfall, groundwater level, water temperature and conductivity. The parameter optimization module 702 is used to construct an LSTM analysis model of the source mining area and use a Bayesian optimization algorithm to optimize the hyperparameters of the LSTM analysis model to obtain an optimized LSTM model. The hyperparameters of the LSTM analysis model include the number of LSTM layers, learning rate, number of training rounds, batch size, and number of hidden units. The model training module 703 is used to train the optimized LSTM model using the input-output sample pair to obtain an initial groundwater analysis model. The model transfer module 704 is used to transfer the initial groundwater analysis model from the source mining area to the target mining area using a transfer learning method, so as to obtain a final groundwater analysis model that is adapted to the target mining area. The groundwater analysis module 705 is used to output the groundwater analysis results corresponding to the real-time rainfall in the target mining area through the final groundwater analysis model.

[0077] In detail, the modules in the non-ferrous metal mining area karst groundwater analysis system 700 based on deep machine learning described in this embodiment of the invention employ the same methods as described above. Figure 1 The method used is the same as the one described in the article on the analysis of karst groundwater in non-ferrous metal mining areas based on deep machine learning, and can produce the same technical effects, so it will not be repeated here.

[0078] In one embodiment, a computer device is provided, which may be a server or a client, and its internal structure diagram may be as follows: Figure 8 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external clients via a network connection. When the computer program is executed by the processor, it implements server-side or client-side functions or steps of a method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning.

[0079] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain multi-source time series of the source mining area within the target time period, and combine the multi-source time series into input-output sample pairs, wherein the multi-source time series includes rainfall, groundwater level, water temperature and conductivity; An LSTM analysis model of the source mining area is constructed, and the hyperparameters of the LSTM analysis model are optimized using the Bayesian optimization algorithm to obtain the optimized LSTM model. The hyperparameters of the LSTM analysis model include the number of LSTM layers, learning rate, number of training rounds, batch size, and number of hidden units. The optimized LSTM model is trained using the input-output sample pair to obtain the initial groundwater analysis model; The initial groundwater analysis model is transferred from the source mining area to the target mining area using the transfer learning method, resulting in a final groundwater analysis model that is adapted to the target mining area. The final groundwater analysis model outputs the groundwater analysis results corresponding to the real-time rainfall in the target mining area.

[0080] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain multi-source time series of the source mining area within the target time period, and combine the multi-source time series into input-output sample pairs, wherein the multi-source time series includes rainfall, groundwater level, water temperature and conductivity; An LSTM analysis model of the source mining area is constructed, and the hyperparameters of the LSTM analysis model are optimized using the Bayesian optimization algorithm to obtain the optimized LSTM model. The hyperparameters of the LSTM analysis model include the number of LSTM layers, learning rate, number of training rounds, batch size, and number of hidden units. The optimized LSTM model is trained using the input-output sample pair to obtain the initial groundwater analysis model; The initial groundwater analysis model is transferred from the source mining area to the target mining area using the transfer learning method, resulting in a final groundwater analysis model that is adapted to the target mining area. The final groundwater analysis model outputs the groundwater analysis results corresponding to the real-time rainfall in the target mining area.

[0081] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0082] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0083] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0084] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0085] Finally, it should be noted that in the above embodiments, each embodiment can be combined with each other or independent. Deleting any one of them will not affect the technical implementation of other embodiments. The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for analyzing karst groundwater in a non-ferrous metal mine area based on deep machine learning, characterized in that, The method includes: Obtain multi-source time series of the source mining area within the target time period, and combine the multi-source time series into input-output sample pairs, wherein the multi-source time series includes rainfall, groundwater level, water temperature and conductivity; An LSTM analysis model of the source mining area is constructed, and the hyperparameters of the LSTM analysis model are optimized using the Bayesian optimization algorithm to obtain the optimized LSTM model. The hyperparameters of the LSTM analysis model include the number of LSTM layers, learning rate, number of training rounds, batch size, and number of hidden units. The optimized LSTM model is trained using the input-output sample pair to obtain the initial groundwater analysis model; The initial groundwater analysis model is transferred from the source mining area to the target mining area using the transfer learning method, resulting in a final groundwater analysis model that is adapted to the target mining area. The final groundwater analysis model outputs the groundwater analysis results corresponding to the real-time rainfall in the target mining area.

2. The deep machine learning-based non-ferrous mine karst groundwater analysis method according to claim 1, wherein, The hyperparameters of the LSTM analysis model are optimized using the Bayesian optimization algorithm to obtain the optimized LSTM model, including: Obtain the validation set from the input-output sample pairs; Based on the validation set, the hyperparameters of the LSTM analysis model are iteratively optimized using the Bayesian optimization algorithm; The combination of hyperparameters corresponding to the minimum average relative error obtained from each observation after reaching the preset number of iterations is taken as the optimal hyperparameters. The LSTM analysis model is reconstructed using the optimal hyperparameters to obtain the optimized LSTM model.

3. The method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning as described in claim 2, characterized in that, Based on the validation set, the hyperparameters of the LSTM analysis model are iteratively optimized using a Bayesian optimization algorithm, including: Multiple sets of hyperparameters are initialized within the search space of the hyperparameters of the LSTM analysis model, and the LSTM analysis model is trained group by group on the validation set using the multiple sets of hyperparameters. Record the average relative error of the LSTM analysis model output with respect to the validation set; Based on the average relative error of the validation set, the hyperparameters of the LSTM analysis model are iteratively optimized using a Bayesian optimization algorithm.

4. The method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning as described in claim 3, characterized in that, Bayesian optimization algorithms include: A Gaussian process surrogate model is constructed using hyperparameters as input variables and average relative error as the objective function value. The posterior distribution of the Gaussian process surrogate model is updated using the observation samples formed by the hyperparameters and the corresponding average relative errors, so as to select the next set of hyperparameters through the acquisition function; The next set of hyperparameters is substituted into the LSTM analysis model to obtain a new average relative error, until the preset number of iterations is reached; After reaching the preset number of iterations, the process of iteratively optimizing the hyperparameters of the LSTM analysis model is completed.

5. The method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning as described in claim 1, characterized in that, Constructing an LSTM analysis model for the source mining area includes: The number of grid outputs is set according to the number of monitoring stations in the source mining area, so that each grid outputs the water level analysis value of the corresponding monitoring station; Establish an LSTM analysis model on a general computing platform.

6. The method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning as described in claim 5, characterized in that, The LSTM analysis model includes an LSTM layer, an activation function, and a fully connected layer. The LSTM layer consists of multiple neural units, each of which includes a forget gate, an input gate, and an output gate. The number of output nodes in the fully connected layer is configured to be equal to the number of grid outputs.

7. The method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning as described in claim 1, characterized in that, The initial groundwater analysis model is transferred from the source mining area to the target mining area using a transfer learning method to obtain a final groundwater analysis model adapted to the target mining area, including: After transferring the initial groundwater analysis model to the target mining area using the transfer learning method, the initial groundwater analysis model is fine-tuned to obtain the fine-tuned model. The fine-tuned model is used for forward calculation to obtain the predicted water level map; Based on the predicted water level map, the final groundwater analysis model is determined.

8. The method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning as described in claim 7, characterized in that, Based on the predicted water level map, the final groundwater analysis model is determined, including: Calculate the depth error, correlation, similarity, and probability distribution overlap between the predicted water level map and the measured water level map; The final groundwater analysis model is determined based on the depth error, the correlation, the similarity, and the overlap of the probability distributions.

9. The method for analyzing karst groundwater in non-ferrous metal mining areas based on deep machine learning as described in claim 8, characterized in that, Calculating the depth error, correlation, similarity, and probability distribution overlap between the predicted and measured water level maps includes: The depth error between the predicted water level map and the measured water level map is calculated using the average relative error method. The correlation and similarity between the predicted water level map and the measured water level map were analyzed using the two-dimensional correlation coefficient method and the structural similarity method. The probability distribution overlap between the predicted water level map and the measured water level map was analyzed using the Batachalia distance method and the histogram intersection distance method.

10. A karst groundwater analysis system for non-ferrous metal mining areas based on deep machine learning, characterized in that, The system includes: The sample combination module is used to acquire multi-source time series of the source mining area within the target time period and combine the multi-source time series into input-output sample pairs. The multi-source time series includes rainfall, groundwater level, water temperature and conductivity. The parameter optimization module is used to construct the LSTM analysis model of the source mining area and use the Bayesian optimization algorithm to optimize the hyperparameters of the LSTM analysis model to obtain the optimized LSTM model. The hyperparameters of the LSTM analysis model include the number of LSTM layers, learning rate, number of training rounds, batch size, and number of hidden units. The model training module is used to train the optimized LSTM model using the input-output sample pair to obtain the initial groundwater analysis model. The model transfer module is used to transfer the initial groundwater analysis model from the source mining area to the target mining area using the transfer learning method, so as to obtain a final groundwater analysis model that is adapted to the target mining area. The groundwater analysis module is used to output the groundwater analysis results corresponding to the real-time rainfall in the target mining area through the final groundwater analysis model.

Images