A mine water inflow prediction method based on artificial intelligence

Abstract

The present application belongs to the technical field of data processing, and particularly relates to a mine water inflow prediction method based on artificial intelligence. The steps are as follows: collecting multi-modal data of mine geological structure, hydrology, meteorology and mining working conditions, aligning with the water inflow data according to time, and obtaining labeled and unlabeled data sets; preprocessing through multi-modal space-time alignment, improved isolated forest to remove outliers and complete missing data; constructing a Transformer self-supervised pre-training model containing cross-modal attention, learning general features on the total unlabeled data set; screening unlabeled data to form a pseudo-labeled data set by cosine similarity; training the model by combining labeled and pseudo-labeled data through a double-branch deep neural network, and outputting the mine water inflow. This method solves the problems of data space-time deviation, abnormal missing, etc., and improves the prediction accuracy and generalization.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a method for predicting mine water inflow based on artificial intelligence. Background Technology

[0002] Mine water inflow is a critical aspect of mine safety. While data-driven methods have become a research hotspot, current data processing suffers from significant technical bottlenecks, failing to meet high-precision requirements. Multimodal data used for prediction, including geological structure, hydrology, meteorology, and mining conditions, exhibits timestamp differences due to varying monitoring equipment collection frequencies. Existing methods lack a unified time benchmark for data resampling and matching verification, leading to biases in the foundational data for subsequent modeling. Complex on-site environments also introduce outliers or missing data. Traditional isolated forest methods for handling outliers fail to consider the varying importance of features across modalities, easily resulting in the deletion of valid data. Mean imputation and linear interpolation methods also fail to consider historical correlations, leading to significant discrepancies between the imputed and actual results. Labeled data requires long-term on-site collection, resulting in high costs and limited quantity. Existing purely supervised learning models rely on only a small amount of labeled data for training, making it difficult to learn universal features across multimodal data. Furthermore, a large amount of unlabeled data remains idle, wasting resources. These issues collectively contribute to insufficient dataset quality and weak model generalization ability, severely impacting the accuracy of mine water inflow prediction. Summary of the Invention

[0003] To address the data processing problems existing in the aforementioned background technology, this invention proposes an artificial intelligence-based method for predicting mine water inflow.

[0004] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0005] Multimodal data of the mining area were collected, including geological structure data, hydrological data, meteorological data and mining condition data. Existing water inflow data were collected and aligned based on the data collection time to obtain labeled datasets with water inflow data labels and unlabeled datasets without water inflow data labels.

[0006] The labeled and unlabeled datasets are preprocessed, and an improved isolated forest is used to remove outliers and fill in missing data to obtain standardized labeled and unlabeled datasets.

[0007] A self-supervised pre-trained model based on Transformer is constructed. The labeled dataset is combined with the unlabeled dataset after removing the water flow label. The model is then pre-trained on the unlabeled dataset to learn the general feature representation of multimodal data. A cross-modal attention mechanism is introduced to calculate the correlation weight of each modality feature and output the feature vector sequence.

[0008] The feature vectors output by the pre-trained model are divided into the original labeled dataset and the unlabeled dataset according to the index label. A threshold is set to filter the unlabeled data whose feature vectors are more or less similar to the labeled data, thus forming a pseudo-labeled dataset.

[0009] A deep neural network prediction model is constructed, taking the labeled dataset and the pseudo-labeled dataset as inputs. The deep neural network contains two parallel branches. The first branch takes the pseudo-labeled dataset as input and outputs the predicted pseudo-inflow data labels. The pseudo-inflow data labels are combined with the pseudo-labeled dataset and used together with the labeled dataset as inputs to the second branch to train the prediction model. The trained prediction model is then used to predict the mine water inflow data.

[0010] As a preferred option, the geological structure data in the multimodal data includes lithology and fault density, the hydrological data includes groundwater level and permeability coefficient, the meteorological data includes precipitation and evaporation, and the mining condition data includes mining depth and working face advance speed.

[0011] Preferably, before preprocessing the labeled and unlabeled datasets and using an improved isolation forest to remove outliers, a multimodal spatiotemporal alignment operation should be performed on the entire dataset.

[0012] First, the timestamp sequence of the original data of each modality is extracted. Then, the data of each modality is resampled using a sliding time window to obtain the resampled data sequence and the corresponding preliminary timestamp.

[0013] Building a unified timestamp Calculate the resampling timestamp of each module and Matching degree The calculation method is as follows ,in Modal resampling timestamp and The number of matches, for Total length, This is the i-th timestamp after modal resampling. for The i-th timestamp;

[0014] Matching degree Modal data below a set threshold are supplemented using linear interpolation. The values ​​at the specified time points are used to obtain the spatiotemporally aligned overall dataset. .

[0015] As a preferred approach, an improved isolated forest is used to remove outliers from the dataset and fill in missing data, resulting in a standardized implementation of both labeled and unlabeled datasets:

[0016] First, calculate the representation of each modality in the overall dataset. Feature importance weights Based on information gain calculation of random forest, ,in This represents the information gain of the modal data.

[0017] Initialize the number of trees, starting from each isolated tree. Randomly select a subset of samples ;

[0018] During sample splitting, first follow Select the mode to be split, and then generate an initial splitting threshold within the feature value range of that mode. And corrected by modal variation coefficient ,in, For correction factor, The coefficient of variation for this mode is obtained by dividing the mode standard deviation by the mode mean; repeat the sample splitting step until the sample subset cannot be split or the tree depth is reached. ,in for Given the number of samples, construct K trees.

[0019] right Each sample Calculate its corrected path length in each isolated tree. ,in This is the original path length. For the k-th tree, For the sample Weights of the associated mode;

[0020] Calculate samples Average corrected path length Isolated scores are calculated using the average corrected path length. ,in, This is the path length correction value, calculated as follows: ;

[0021] Set an outlier score threshold to exclude abnormal and isolated cases. Samples with an anomaly score greater than or equal to the anomaly score threshold are considered an anomaly-free dataset. ;

[0022] For missing data in a dataset without anomalies, the missing parts are selected from a five-year historical dataset based on time, and the average value is used to fill in the missing data.

[0023] Finally, the complete dataset is standardized to obtain a standardized labeled dataset and an unlabeled dataset.

[0024] As a preferred approach, a self-supervised pre-trained model based on Transformer is constructed. This involves combining the labeled dataset (excluding the water flow label) with the unlabeled dataset to obtain a total unlabeled dataset. The model is then pre-trained on this total unlabeled dataset to learn general feature representations for multimodal data. The specific implementation of this method is as follows:

[0025] First, the preprocessed labeled dataset with the water flow label removed and the unlabeled dataset are combined to form a total unlabeled dataset.

[0026] The total unlabeled dataset is input into the Transformer self-supervised pre-trained model, which outputs a sequence of feature vectors. The model includes an input adaptation layer, a LayerNorm layer, and a cross-modal attention layer.

[0027] The input adaptation layer is used to transform the raw multimodal features into a high-dimensional vector format that the Transformer can accept;

[0028] The LayerNorm layer is used to normalize the feature vector output by the input adaptation layer on a modal basis, eliminating training fluctuations caused by differences in the numerical distribution of features across different modalities.

[0029] The cross-modal attention layer is used to quantify the feature dependencies between any two modalities, dynamically allocate modal association weights, and achieve collaborative fusion of multimodal features.

[0030] Preferably, the cross-modal attention layer is used to quantify the feature dependencies between any two modalities, dynamically allocate modal association weights, and achieve the collaborative fusion of multimodal features. The specific implementation is as follows:

[0031] For the feature vector of each modality feature output by the LayerNorm layer, a query vector, a key vector, and a value vector are generated respectively through three independent linear mappings.

[0032] For each attention head Calculate the cross-modal attention scores for any two modalities A and B, and calculate the base attention scores. ,in For the feature dimensions of a single attention head, For mode A in the th... The query vector under each attention head For mode A in the th... Each attention head key vector;

[0033] Modality importance weights are introduced for initialization, and the attention score is corrected. The calculation method is as follows: ,in These are the feature importance weights for modes A and B, respectively.

[0034] Softmax normalization is applied to the corrected attention scores to obtain the association weights of mode A relative to mode B. ;

[0035] Multimodal feature fusion and final feature output are performed. For each attention head, the value vectors of all modalities are weighted and summed through association weights to obtain the fused features of that head. The fused features of all attention heads are concatenated, and dimensionality is restored and features are integrated through linear mapping to obtain the output features of the cross-modal attention module. Where n is the number of attention heads, To output the weight matrix, For bias terms;

[0036] Finally, the output features of the cross-modal attention module are input into the fully connected layer to output a sequence of general feature vectors.

[0037] Preferably, the feature vectors output by the pre-trained model are divided into the original labeled dataset and the unlabeled dataset according to the index label. A threshold is set to filter the unlabeled data whose similarity to the feature vectors of the labeled data is greater than or equal to the threshold, forming a pseudo-labeled dataset. The similarity is calculated using the cosine similarity method. Specifically, the cosine similarity between the feature vectors of each unlabeled dataset and the feature vectors of each labeled dataset is calculated. Data in the unlabeled dataset that is greater than or equal to the preset similarity threshold is filtered to form a pseudo-labeled dataset.

[0038] Compared with existing technologies, the advantages and positive effects of this invention are as follows: It proposes a method for predicting mine water inflow by combining multimodal data processing, self-supervised learning, and a bi-branch prediction structure. First, through multimodal spatiotemporal alignment and an improved isolated forest algorithm, it achieves consistency and high-quality construction of multi-source data such as geological, hydrological, meteorological, and mining conditions, avoiding prediction biases caused by data anomalies and missing data in traditional methods. Second, this invention utilizes a Transformer self-supervised pre-trained model, combined with a cross-modal attention mechanism, to fully mine the potential information of unlabeled data under limited labeled sample conditions, improving feature representation capabilities. Furthermore, it generates a pseudo-labeled dataset through cosine similarity filtering and introduces a pseudo-label purification and joint supervised training mechanism into the bi-branch deep neural network, achieving collaborative modeling of real labeled and high-quality pseudo-labeled samples, effectively alleviating the problem of insufficient labeling. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of the structural process of a method for predicting mine water inflow based on artificial intelligence. Detailed Implementation

[0041] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0042] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0043] In this example, mine water inflow is a core technical aspect of safe mine production and efficient resource extraction, directly impacting drainage system scheduling, mining plan formulation, and safety risk prevention. While data-driven prediction methods are gradually replacing traditional empirical formulas in the existing technological system, they still face multi-dimensional technical bottlenecks. Therefore, an artificial intelligence-based mine water inflow prediction method is proposed, with the specific implementation process as follows: Figure 1 As shown.

[0044] First, multimodal data of the mining area were collected, including geological structure data, hydrological data, meteorological data, and mining condition data. Existing water inflow data were also collected and aligned based on the data acquisition time to obtain labeled datasets with water inflow data labels and unlabeled datasets without water inflow data labels. Among the multimodal data, geological structure data included lithology and fault density, hydrological data included groundwater level and permeability coefficient, meteorological data included precipitation and evaporation, and mining condition data included mining depth and working face advance speed.

[0045] To overcome the timestamp differences between different modalities and ensure the consistency of input data, a multimodal spatiotemporal alignment operation is proposed. First, the timestamp sequences of the original data for each modality are extracted. Then, a sliding time window is used to resample the data for each modality, obtaining the resampled data sequence and its corresponding preliminary timestamp. Finally, a unified timestamp is constructed. Calculate the resampling timestamp of each module and Matching degree The calculation method is as follows ,in Modal resampling timestamp and The number of matches, for Total length, This is the i-th timestamp after modal resampling. for The i-th timestamp; for matching degree Modal data below a set threshold are supplemented using linear interpolation. The values ​​at the specified time points are used to obtain the spatiotemporally aligned overall dataset. This approach ensures the alignment of different modal data on a unified time axis, enabling the model to receive consistent input and avoiding prediction bias caused by time series differences.

[0046] To improve data quality and reduce model bias caused by outliers and missing data, an improved isolation forest approach was adopted to remove outliers and impute missing data, resulting in a standardized labeled dataset and an unlabeled dataset. First, the feature importance of each modality in the overall dataset was calculated, and its weight in anomaly detection was determined based on information gain. When constructing the isolation tree, the modality variation coefficient was introduced to correct the split threshold, and the isolation score was calculated using the corrected path length to remove outliers. Missing values ​​were imputed by incorporating the temporal correlation of five years of historical data, and finally, standardization was performed. Specifically, the feature importance of each modality in the overall dataset was first calculated... Feature importance weights Based on information gain calculation of random forest, ,in The information gain of this modality data; initialize the number of trees, starting from [the value] for each isolated tree. Randomly select a subset of samples During sample splitting, first follow... Select the mode to be split, and then generate an initial splitting threshold within the feature value range of that mode. And corrected by modal variation coefficient ,in, For correction factor, The coefficient of variation for this mode is obtained by dividing the mode standard deviation by the mode mean; repeat the sample splitting step until the sample subset cannot be split or the tree depth is reached. ,in for Given the number of samples, construct K trees; for Each sample Calculate its corrected path length in each isolated tree. ,in This is the original path length. For the k-th tree, For the sample Weights of the respective modes; calculate samples Average corrected path length Isolated scores are calculated using the average corrected path length. ,in, This is the path length correction value, calculated as follows: Set an abnormal isolated score threshold to exclude [illegible] cases. Samples with an anomaly score greater than or equal to the anomaly score threshold are considered an anomaly-free dataset. For missing data in the dataset without anomalies, the missing parts are selected from the five-year historical dataset according to time, and the missing values ​​are filled in using the average value. Finally, the completed dataset is standardized to obtain a standardized labeled dataset and an unlabeled dataset.

[0047] To fully utilize the large amount of unlabeled data and address the problem of insufficient labeled data, this invention constructs a self-supervised pre-trained model based on Transformer. The labeled dataset is combined with the unlabeled dataset after removing the inflow volume label, resulting in a total unlabeled dataset. The model is then pre-trained on this total unlabeled dataset to learn general feature representations for multimodal data. A cross-modal attention mechanism is introduced to calculate the association weights of features across different modalities, outputting a sequence of feature vectors. First, the preprocessed labeled dataset (without the inflow label) and the unlabeled dataset are extracted and combined to form a total unlabeled dataset. This total unlabeled dataset is then input into a Transformer self-supervised pre-trained model, which outputs a sequence of feature vectors. The model includes an input adaptation layer, a LayerNorm layer, and a cross-modal attention layer. The input adaptation layer transforms the original multimodal features into a high-dimensional vector format acceptable to Transformer. The LayerNorm layer normalizes the output feature vectors from the input adaptation layer modally, eliminating training fluctuations caused by differences in the numerical distribution of features across different modalities. The cross-modal attention layer quantifies the feature dependencies between any two modalities, dynamically assigns modal association weights, and achieves collaborative fusion of multimodal features. Specifically, to achieve a unified representation of the original data from different modalities, this invention sets an input adaptation layer at the front end of the self-supervised pre-trained model. Specifically, the input adaptation layer first receives multimodal data from geology, hydrology, meteorology, and mining conditions, each type of data having different numerical ranges, dimensional structures, and physical meanings during the original acquisition. To enable the Transformer to process this data uniformly, the input adaptation layer employs a multi-channel linear projection approach, mapping the original features of each modality to a high-dimensional vector space of the same dimension. For example, for geological modal inputs, a linear transformation using the weight matrix and bias term converts the original lithology density, fault parameters, and other features into d-dimensional embedding vectors. Similarly, for meteorological modalities such as precipitation and evaporation, vectors of the same dimension are obtained through corresponding mapping layers. These vectors are then concatenated into a multimodal sequence input, with added positional encoding information to maintain temporal order. In this way, the input adaptation layer completes the transformation from "modal features" to "unified high-dimensional representation," enabling the subsequent Transformer encoder to perform cross-modal modeling within the same vector space. This layer ensures that the original heterogeneous data can be integrated into the same neural network architecture, laying the foundation for cross-modal feature learning.

[0048] After mapping multimodal features to a unified high-dimensional vector space in the input adaptation layer, directly entering deep network training often faces the problem of large differences in numerical distributions between modalities. For example, the groundwater level in hydrological data and the precipitation in meteorological data have different numerical scales by orders of magnitude. This can lead to gradient fluctuations during model training, and even cause some modal features to interfere too strongly or too weakly with the overall training process. To solve this problem, this invention introduces a modal-wise LayerNorm layer after the adaptation layer. Specifically, it is implemented as follows: first, the mean and variance of the feature vector for each modality are calculated independently; then, all feature components of that modality are standardized so that their numerical distribution approaches zero mean and unit variance; furthermore, a learnable scaling factor and bias factor are used to linearly transform the standardization result to ensure that the feature distribution can flexibly adapt to downstream tasks. Unlike traditional BatchNorm, LayerNorm remains stable at each time step and within each sample, making it particularly effective for time series prediction and mini-batch training. Through this normalization operation, the model can better eliminate the fluctuations caused by modal differences, making the training process more stable and the feature representation more balanced, thus creating a good precondition for the subsequent calculation of cross-modal attention mechanisms.

[0049] After normalization is completed in the LayerNorm layer, the model enters the cross-modal attention layer, which is a crucial step in achieving deep fusion of multimodal data. The process involves taking the feature vector of each modal feature output from the LayerNorm layer and generating a query vector, key vector, and value vector through three independent linear mappings; for each attention head... Calculate the cross-modal attention scores for any two modalities A and B, and calculate the base attention scores. ,in For the feature dimensions of a single attention head, For mode A in the th... The query vector under each attention head For mode A in the th... Each attention head has a key vector; modal importance weights are introduced for initialization, and the attention score is corrected, calculated as follows: ,in These are the feature importance weights for modes A and B, respectively; the modified attention scores are then softmax normalized to obtain the association weight of mode A relative to mode B. Multimodal feature fusion and final feature output are performed. For each attention head, the value vectors of all modalities are weighted and summed through association weights to obtain the fused features of that head. The fused features of all attention heads are concatenated, and dimensionality is restored and features are integrated through linear mapping to obtain the output features of the cross-modal attention module. Where n is the number of attention heads, To output the weight matrix, This is the bias term; finally, the output features of the cross-modal attention module are input into the fully connected layer to output a general feature vector sequence.

[0050] The feature vectors output by the pre-trained model are divided into labeled and unlabeled datasets according to their index labels. A threshold is set to filter unlabeled data whose similarity to the labeled data feature vectors is greater than or equal to the threshold, forming a pseudo-labeled dataset. The similarity is calculated using cosine similarity. The method involves first calculating the cosine similarity between the feature vectors of each unlabeled dataset and the feature vectors of each labeled dataset, and then filtering out data from the unlabeled dataset whose similarity is greater than or equal to the preset similarity threshold to form the pseudo-labeled dataset. Specifically, the feature vectors output by the pre-trained model are first divided into labeled and unlabeled datasets according to their original index labels. Then, for each feature vector of unlabeled data, the cosine similarity between it and all the feature vectors of labeled data is calculated. Cosine similarity measures the degree of similarity by the angle between vectors, with a value range between -1 and 1. The closer the value is to 1, the closer the direction of the feature vectors is, and the stronger the correlation. To ensure the reliability of the screening results, this invention sets a similarity threshold. When the similarity between an unlabeled data point and any labeled data point is greater than or equal to this threshold, the unlabeled data point is determined to be similar in features to the labeled sample, and can be assigned a pseudo-label and included in the pseudo-labeled dataset. In this way, the model can expand the training sample size under the condition of limited real labeled samples, enhance its learning ability on feature distribution, and thus effectively improve the accuracy and robustness of mine water inflow prediction.

[0051] Finally, a deep neural network prediction model is constructed, taking both the labeled dataset and the pseudo-labeled dataset as input. The deep neural network contains two parallel branches. The first branch takes the pseudo-labeled dataset as input and outputs predicted pseudo-inflow data labels. These pseudo-inflow data labels are combined with the pseudo-labeled dataset and used as input along with the labeled dataset, fed into the second branch to train the prediction model. The trained prediction model is then used to predict the mine inflow data. Specifically, the model includes a first branch and a second branch. The first branch primarily targets the pseudo-labeled dataset, with its input consisting of pseudo-labeled samples selected through cosine similarity. This branch's task is to learn the mapping relationship between the pseudo-labeled data and the inflow labels, and output the corresponding "pseudo-inflow prediction value." During training, this branch uses a multilayer perceptron to perform regression modeling on the pseudo-labeled samples and uses mean squared error as the loss function for parameter updates. These predicted labels are then combined with the pseudo-labeled features. This mechanism effectively reduces the negative impact of low-quality pseudo-labels, ensuring that the pseudo-labeled data plays a positive role in the overall training. The second branch is the core prediction model, whose input includes both real labeled data and data combined by the first branch. This branch is trained using a more complex deep network structure, leveraging strong supervision signals from real-labeled samples and supplemented by high-quality pseudo-labeled samples to establish a deep model of the relationship between multimodal features and water inflow. During training, the loss function of the second branch consists of two parts: a rigorously supervised loss for real-labeled samples and a weighted loss for pseudo-labeled samples, with the weighting factor related to the confidence level of the pseudo-labels. This allows the model to converge stably with real-labeled data in the early stages of training, while gradually incorporating information from pseudo-labeled data in later stages, thus achieving comprehensive learning of large-scale data distributions. The resulting prediction model not only maintains high prediction accuracy under limited labeling conditions but also exhibits better generalization ability and robustness in complex mining environments.

[0052] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A mine water inflow prediction method based on artificial intelligence, characterized in that, The method comprises the following steps: Collecting multi-modal data of the mine area, including geological structure data, hydrological data, meteorological data and mining working condition data, and collecting existing water inrush dataAligning the data collection time to obtain a labeled data set with water inrush data labels and an unlabeled data set without water inrush data labels; Pretreating the labeled data set and the unlabeled data set, removing abnormal values in the data set by improved isolation forest, and completing missing data to obtain standardized labeled data set and unlabeled data set; Building a self-supervised pre-training model based on Transformer, combining the labeled data set without water inrush labels and the unlabeled data set to obtain the total unlabeled data set, pre-training the model on the total unlabeled data set, learning the general feature representation of multi-modal data, introducing cross-modal attention mechanism, calculating the correlation weight of each modal feature, and outputting the feature vector sequence; The feature vectors output by the pre-training model are divided into the original labeled data set and the unlabeled data set according to the serial number label, and a threshold is set to filter the unlabeled data with a similarity greater than or equal to the threshold with the labeled data feature vector to form a pseudo-labeled data set; Building a deep neural network prediction model, taking the labeled data set and the pseudo-labeled data set as input; the deep neural network contains two parallel branches, the first branch inputs the pseudo-labeled data set, outputs the predicted pseudo-water inrush data label, combines the pseudo-water inrush data label with the pseudo-labeled data set, and inputs them into the second branch together as input to train the prediction model, and uses the trained prediction model to predict the mine water inrush data; Before pretreating the labeled data set and the unlabeled data set, the multi-modal spatio-temporal alignment operation needs to be performed on the whole data set: First, extract the timestamp sequence of each modal original data, resample each modal data by using a sliding time window, and obtain the resampled data sequence and the corresponding preliminary timestamp; Constructing unified timestamps , calculating the matching degree of each module resampling timestamp with , the calculation method is , wherein is the matching number of modal resampling timestamps and is the total length of , is the i th timestamp after modal resampling , is the i th timestamp of ​​​​ degree of matching Modal data below a set threshold is supplemented with linear interpolation The values at the exact time points are obtained, resulting in a spatio-temporally aligned overall dataset ; The specific implementation of removing abnormal values in the data set and completing missing data to obtain standardized labeled data set and unlabeled data set is as follows: First, the feature importance weight of each modality in the overall dataset is calculated based on the random forest information gain calculation, where is the information gain of the modality data; The number of trees is initialized, and for each isolated tree, a sample subset is randomly drawn from the set of samples ; In the sample splitting, first according to Select the mode to be split, and then generate the initial splitting threshold in the characteristic value range of the mode And correct by the coefficient of variation of the mode Wherein, The correction coefficient is The coefficient of variation of the mode is obtained by dividing the mode standard deviation by the mode mean; repeat the sample splitting step until the sample subset cannot be split or the tree depth is reached Wherein The The number of samples is completed K trees right Each sample Calculate its corrected path length in each isolated tree. ,in This is the original path length. For the k-th tree, For the sample Weights of the associated mode; Computing the average corrected path length of the sample Computing the isolated fraction by the average corrected path length Computing the isolated fraction by the average corrected path length wherein is the path length correction value, which is computed as ; Set an abnormal isolated score threshold, eliminate Samples greater than or equal to the abnormal score threshold, get the non- abnormal data set ; For the missing data in the abnormal data set, the missing value data is supplemented by using the average value according to the time selected from the five-year historical data set; Finally, the complete data set is standardized to obtain the standardized labeled data set and the unlabeled data set. 2.The mine water inflow prediction method based on artificial intelligence according to claim 1, characterized in that, The geological structure data in the multi-modal data includes lithology and fault density, the hydrological data includes groundwater level and permeability coefficient, the meteorological data includes precipitation and evaporation, and the mining working condition data includes mining depth and working face advancing speed. 3.The mine water inflow prediction method based on artificial intelligence according to claim 1, characterized in that, The specific implementation of building a self-supervised pre-training model based on Transformer, combining the labeled data set without water inrush labels and the unlabeled data set to obtain the total unlabeled data set, pre-training the model on the total unlabeled data set, and learning the general feature representation of multi-modal data is as follows: First, the pre-processed labeled data set and unlabeled data set without the gushing water amount label are combined to form a total unlabeled data set; The total unlabeled data set is input into a Transformer self-supervised pre-training model to output a sequence of feature vectors, the model comprising an input adaptation layer, a LayerNorm layer, and a cross-modal attention layer; The input adaptation layer is used to convert the original multi-modal features into a high-dimensional vector format that can be accepted by the Transformer; The LayerNorm layer is used to normalize the feature vectors output by the input adaptation layer, eliminating training fluctuations caused by differences in the numerical distribution of different modal features; The cross-modal attention layer is used to quantify the feature dependency between any two modalities, dynamically allocate modal correlation weights, and achieve collaborative fusion of multi-modal features.

4. The artificial intelligence-based mine water inflow prediction method of claim 3, wherein The specific implementation of the cross-modal attention layer for quantifying the feature dependency between any two modalities, dynamically allocating modal correlation weights, and achieving collaborative fusion of multi-modal features is as follows: For each modal feature vector output by the LayerNorm layer, three independent linear mappings are used to generate query vectors, key vectors, and value vectors, respectively; For each attention head , the cross-modal attention score of any two modalities A and B is calculated, and the base attention score is calculated , where is the feature dimension of a single attention head, is the query vector of modality A under the th attention head, is the key vector of modality A under the th attention head; The modal importance weight initialization is introduced to correct the attention score, and the calculation manner is wherein respectively are the feature importance weights of modal A and B. Softmax normalization is performed on the modified attention scores to obtain the correlation weight of modality A relative to modality B ; The multi-modal feature fusion and the final feature output are performed, and the value vectors of all modes are weighted and summed through the association weight to obtain the fusion feature of each attention head The fusion features of all attention heads are spliced, and the dimension is reduced and the features are integrated through linear mapping to obtain the output feature of the cross-modal attention module Wherein n is the number of attention heads, is an output weight matrix, is a bias term Finally, the output features of the cross-modal attention module are input into a fully connected layer to output a sequence of general feature vectors. 5.The mine water inflow prediction method based on artificial intelligence according to claim 1, wherein, The feature vectors output by the pre-training model are divided into the original labeled data set and the unlabeled data set according to the sequence number label, and a threshold is set to filter the unlabeled data with a similarity greater than or equal to the threshold value with the labeled data feature vectors to form a pseudo-labeled data set, the similarity is calculated using the cosine similarity calculation method, the specific implementation method is as follows: first, calculate the cosine similarity between each unlabeled data set feature vector and each labeled data set feature vector, filter the data in the unlabeled data set with a similarity greater than or equal to the preset similarity threshold to form the pseudo-labeled data set.

Images