Water inflow prediction method based on adaptive two-type fuzzy information granulation calculation

Abstract

The application relates to a water inrush amount prediction method based on adaptive two-type fuzzy information granulation calculation, which comprises the following steps: preprocessing original water inrush amount data collected; adopting a perception important point algorithm to dynamically segment the processed water inrush amount data; constructing an interval two-type fuzzy set for the data in each adaptive window; extracting five-dimensional key features of each adaptive window based on the interval two-type fuzzy set to form a feature tensor; constructing a water inrush amount prediction network to obtain a prediction lower limit of the water inrush amount at a future moment and a prediction upper limit of the water inrush amount at the future moment; determining a water inrush amount prediction interval according to the prediction lower limit of the water inrush amount at the future moment and the prediction upper limit of the water inrush amount at the future moment, and triggering multidimensional alarm when the prediction upper limit of the water inrush amount at the future moment exceeds a safety threshold. The application not only solves the technical pain points of traditional methods, but also realizes the dual targets of high-precision prediction and safety warning, and can be directly applied to coal mine water inrush amount monitoring and water prevention and control decision-making.

Description

Technical Field

[0001] This application relates to the field of coal mine safety production, specifically to a method for predicting water inflow based on adaptive type-two fuzzy information granularization calculation. Background Technology

[0002] Accurate prediction of water inflow from the roof of coal mines is a crucial step in ensuring safe production, directly impacting the safety of mining personnel, the integrity of equipment, and production efficiency. In the western region's massive sandstone-rich mining areas prone to water hazards, the water inflow data exhibits significant noise interference, strong nonlinear characteristics, and multi-scale fluctuations due to complex geological conditions and dynamic hydrological changes, posing a significant challenge to prediction work.

[0003] Existing methods for predicting water inflow mainly suffer from the following shortcomings: First, traditional prediction models often use fixed time windows to divide data, making it difficult to align key turning points in the water inflow sequence (such as peak water inrush), which can easily lead to feature smoothing and loss of important fluctuation information. Second, existing methods do not adequately model "second-order uncertainty" (i.e., the uncertainty of the fluctuation itself), and cannot effectively quantify the ambiguity of data fluctuations, resulting in poor adaptability of prediction results to complex working conditions. Third, single feature extraction networks (such as TCN or LSTM alone) cannot fully explore the high-dimensional spatiotemporal features and long- and short-term evolution patterns of water inflow data, resulting in limited prediction accuracy and reliability. Fourth, most models only output single-point prediction results, lacking effective estimation of the prediction interval, and cannot provide a comprehensive risk reference for mine water hazard early warning. Summary of the Invention

[0004] To overcome at least one deficiency in the prior art, this application provides a method for predicting water inflow based on adaptive type-two fuzzy information granularization calculation.

[0005] Firstly, a method for predicting water inflow based on adaptive type-II fuzzy information granularization calculation is provided, including: The collected raw water flow data is preprocessed to obtain processed water flow data; The processed water inflow data is dynamically segmented using a perceptual importance point algorithm to form multiple adaptive windows; for the data within each adaptive window, an interval type II fuzzy set is constructed; based on the interval type II fuzzy set, five-dimensional key features of each adaptive window are extracted to form a feature tensor; A water inflow prediction network is constructed, which consists of a multi-channel temporal convolutional network, a long short-term memory network, and a support vector regression model connected in sequence. The feature tensor is input into the water inflow prediction network, passes through the multi-channel temporal convolutional network to obtain a comprehensive feature vector, and is input into the long short-term memory network to obtain the hidden state. The hidden state is input into the support vector regression model to obtain the lower limit and upper limit of the predicted water inflow at future times. The water inflow prediction range is determined based on the lower limit and upper limit of the predicted water inflow at future times, and a multi-dimensional alarm is triggered when the upper limit of the predicted water inflow at future times exceeds the safety threshold.

[0006] In one embodiment, the processed water inflow data is dynamically segmented using a significance point sensing algorithm to form multiple adaptive windows, including: The water inflow corresponding to the start and end times of the processed water inflow data is selected as the initial key control points to form a set of key control points. For any two adjacent critical control points in the current set of critical control points and Calculate the distance from each data point to the line segment between two key control points. The Euclidean distance is used to select the data point with the largest Euclidean distance as the new key control point and add it to the key control point set to form a new key control point set; the process is repeated until the Euclidean distance corresponding to the new key control point is less than or equal to the preset threshold, and the final key control point set is obtained. The final set of critical control points consists of data points between any two adjacent critical control points, forming an adaptive window.

[0007] In one embodiment, for the data within each adaptive window, an interval type II fuzzy set is constructed, including: Calculate the window mean and window standard deviation based on the data within the adaptive window; The upper and lower membership functions are constructed based on the window mean and window standard deviation, and are expressed by the following formula:

[0008] in, For the membership function, To adapt to the data within the window, The mean of the window. For the expanded window standard deviation, , The standard deviation of the window; ) in, For the lower membership function, The standard deviation of the narrowed window. .

[0009] In one embodiment, the five key features include: the minimum value of data within the adaptive window. The maximum value U of the data within the adaptive window, and the 90% confidence lower bound. 90% confidence upper bound Uncertain width ; Among them, the lower bound of 90% confidence To adapt the membership function values ​​within the window The smallest , To adapt to the data within the window; 90% confidence upper bound To adapt the membership function values ​​within the window The largest ; Uncertain width ,in, To adapt the membership function value within the window The largest , To adapt the membership function value within the window The smallest .

[0010] In one embodiment, the multi-channel temporal convolutional network includes: 5 TCN networks and a cross-channel feature fusion module; Each TCN network processes a key feature of one dimension of the feature tensor; The cross-channel feature fusion module is used to concatenate the feature maps output by each TCN network along the feature dimension to obtain a comprehensive feature vector.

[0011] In one embodiment, the TCN network includes a convolutional layer, a batch normalization layer, a ReLU activation layer, and a dilated convolutional layer connected in sequence; the dilation coefficient sequence of the dilated convolutional layer is set as follows: .

[0012] In one embodiment, the support vector regression model includes three support vector regression units, namely a lower bound regressor, a trend regressor, and an upper bound regressor; The lower bound regressor generates a lower bound for predicting the future inflow rate based on the input latent state, while the upper bound regressor generates an upper bound for predicting the future inflow rate based on the input latent state. The trend regressor outputs the trend value of the future inflow rate.

[0013] In one embodiment, the inflow prediction network is a trained network. During training, the lower bound regressor uses the minimum value of the data within each adaptive window. To train the training labels, the trend regressor is trained using the mean of the data within each adaptive window as the training label, while the upper bound regressor is trained using the maximum value U of the data within each adaptive window as the training label.

[0014] In one embodiment, preprocessing includes data continuity restoration, outlier identification and removal, and data validity verification.

[0015] Secondly, a water inflow prediction device based on adaptive type-II fuzzy information granularization calculation is provided, comprising: The data preprocessing module is used to preprocess the collected raw water flow data to obtain processed water flow data. The feature tensor construction module is used to dynamically segment the processed water flow data using the perceptual importance point algorithm to form multiple adaptive windows; for the data within each adaptive window, an interval type II fuzzy set is constructed; based on the interval type II fuzzy set, five-dimensional key features of each adaptive window are extracted to form a feature tensor; The prediction module is used to construct a water inflow prediction network, which includes a multi-channel temporal convolutional network, a long short-term memory network, and a support vector regression model connected in sequence. The feature tensor is input into the water inflow prediction network, passes through the multi-channel temporal convolutional network to obtain a comprehensive feature vector, and is input into the long short-term memory network to obtain the hidden state. The hidden state is input into the support vector regression model to obtain the lower limit and upper limit of the predicted water inflow at future times. The prediction range determination and alarm module is used to determine the water inflow prediction range based on the lower limit and upper limit of the predicted water inflow at future times, and to trigger a multi-dimensional alarm when the upper limit of the predicted water inflow at future times exceeds the safety threshold.

[0016] Compared with the prior art, this application has the following beneficial effects: This application solves the technical pain points of traditional methods and achieves the dual goals of high-precision prediction and safety early warning by using key technical steps such as data preprocessing to ensure input quality, adaptive type II fuzzy set to quantify core uncertainty, multi-channel TCN to extract spatiotemporal features, LSTM to remember long-term trends, and granular SVM output range and early warning. It can be directly applied to coal mine water inflow monitoring and water control decision-making. Attached Figure Description

[0017] This application can be better understood by referring to the description given below in conjunction with the accompanying drawings, which, together with the detailed description below, are incorporated in and form part of this specification. In the drawings: Figure 1 A flowchart of a water inflow prediction method based on adaptive type-II fuzzy information granularization calculation is shown. Figure 2 A schematic diagram of the inflow prediction network is shown; Figure 3 A schematic diagram of a multi-channel temporal convolutional network is shown; Figure 4 The study area is shown in the comprehensive hydrogeological profile and columnar section. Figure 5The raw data of mine water inflow monitoring is shown; Figure 6 The diagram illustrates the contribution of each module to the prediction performance in the ablation experiment; Figure 7 The performance changes of each module after removal are shown in the ablation experiment; Figure 8 The prediction details for each variant during the water inrush event in the ablation experiment are shown; Figure 9 The data representation of different granulation methods is shown in comparison; Figure 10 The impact of different granulation methods on prediction performance is shown; Figure 11 The comparison of prediction performance metrics for different granulation methods is shown; Figure 12 The comparison of predictions using different granulation methods during water inrush events is shown; Figure 13 The comparison of prediction results of different prediction models on the test set is shown; Figure 14 The performance metrics of different prediction models are compared. Figure 15 The radar charts showing the overall performance of each model are displayed; Figure 16 This illustrates the interval prediction and safety warning of this application; Figure 17 Sensitivity analysis of the granulation window size parameter is shown; Figure 18 The prediction performance of the strong baseline model and the method in this application are compared. Detailed Implementation

[0018] Exemplary embodiments of the present application will be described below with reference to the accompanying drawings. For clarity and brevity, not all features of the actual embodiments are described in the specification. However, it should be understood that many embodiment-specific decisions can be made in the development of any such actual embodiment to achieve the developer’s specific objectives, and these decisions may vary as the embodiments differ.

[0019] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the device structure closely related to the solution of this application is shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0020] It should be understood that this application is not limited to the described embodiments by virtue of the following description with reference to the accompanying drawings. In this application, embodiments may be combined with each other, features may be substituted or borrowed between different embodiments, and one or more features may be omitted in one embodiment, where feasible.

[0021] This application provides a method for predicting water inflow based on adaptive type-II fuzzy information granularization calculation. Figure 1 A flowchart of a water inflow prediction method based on adaptive type-II fuzzy information granularization calculation is shown. (See attached diagram) Figure 1 The method mainly includes the following steps: Step S1: Preprocess the collected raw water flow data to obtain processed water flow data.

[0022] Raw water flow data collected by downhole IoT sensors (such as ultrasonic flow meters) is susceptible to equipment vibration, humidity changes, and momentary malfunctions, resulting in issues such as null values ​​and extreme outliers. A simplified yet rigorous processing procedure is required to ensure input reliability. The specific steps are as follows: Data continuity restoration: First, the original data (sampling frequency typically 1 time / hour to 1 time / day) is time-axis verified to remove discrete null values ​​caused by sensor offline. For continuous null values ​​(length ≤ 3 sampling points, to avoid disrupting data sequence), linear interpolation is used to complete them, specifically including: Assume that the adjacent valid data before and after control are respectively (corresponding to time t-1) and (corresponding to time t+1), then null value The calculation formula is: This ensures the continuity of data sequence.

[0023] Outlier identification and removal: Combining two robust statistical methods to locate outliers, taking into account both normal and non-normal data scenarios.

[0024] The 3σ criterion is used to process stationary data that are approximately normally distributed: the mean μ and standard deviation σ of the data are calculated, and if a data point x satisfies the 3σ criterion... (99.73% of the data in the normal distribution fall within this interval), which is identified as an extreme outlier; Using box plots to process non-normal or fluctuating data containing extreme values: Calculate the lower quartiles. Upper quartiles and interquartile range If data point x satisfies or If the identified outlier is less than 5% (to ensure data integrity), it is directly removed. If the length of a continuous outlier segment is less than or equal to 2, it is replaced with the moving average of the previous and next 5 normal data to avoid local time series breaks.

[0025] Data validity verification: After processing, the data is verified by "data integrity ≥ 95%" (ensuring sufficient sample size) and "preliminary verification of time series stationarity" (using the ADF test, with a p-value < 0.1; excessive stationarization is unnecessary to avoid losing fluctuation characteristics). The final output is a clean inflow input sequence. (N is the total length of the data), providing noise-free basic data for the subsequent granulation process.

[0026] Step S2: The processed water inflow data is dynamically segmented using the Perception Important Points algorithm to form multiple adaptive windows; for the data within each adaptive window, an interval type II fuzzy set is constructed; based on the interval type II fuzzy set, the five-dimensional key features of each adaptive window are extracted to form a feature tensor.

[0027] Step S3: Construct a water inflow prediction network. The water inflow prediction network includes a multi-channel temporal convolutional network, a long short-term memory network, and a support vector regression model connected in sequence. The feature tensor is input into the water inflow prediction network, passes through the multi-channel temporal convolutional network to obtain a comprehensive feature vector, and is input into the long short-term memory network to obtain the hidden state. The hidden state is input into the support vector regression model to obtain the lower limit and upper limit of the predicted water inflow at future times.

[0028] Figure 2 A schematic diagram of the inflow prediction network is shown.

[0029] Step S4: Determine the water inflow prediction range based on the predicted lower limit and the predicted upper limit of the water inflow at future times, and trigger a multi-dimensional alarm when the predicted upper limit of the water inflow at future times exceeds the safety threshold.

[0030] This embodiment addresses the technical pain points of traditional methods by employing key technical steps such as data preprocessing to ensure input quality, adaptive type-two fuzzy set quantification of core uncertainties, multi-channel TCN extraction of spatiotemporal features, LSTM memory of long-term trends, and granular SVM output range and early warning. It achieves the dual goals of high-precision prediction and safety early warning, and can be directly applied to coal mine water inflow monitoring and water control decision-making.

[0031] In one embodiment, to address the multi-scale characteristics of water inflow—characterized by "drastic fluctuations in some periods and gradual changes in others"—a Perceptually Important Points (PIPs) algorithm is employed for dynamic segmentation to ensure accurate capture of key fluctuations. In step S2, the processed water inflow data is dynamically segmented using the PIPs algorithm, forming multiple adaptive windows, including: The water inflow at the start and end times of the processed water inflow data are selected as the initial key control points to form a set of key control points, thereby anchoring the global change trend of the sequence and avoiding the loss of overall information.

[0032] For any two adjacent critical control points in the current critical control point set and , Calculate each data point between two key control points. , to line segment European distance, This distance reflects the degree to which a data point deviates from the "trend of the line connecting two points." The larger the distance, the more critical the point is to be in local fluctuations. The data point with the largest Euclidean distance is selected as a new critical control point and added to the critical control point set, forming a new set of critical control points. This iterative process is repeated until the Euclidean distance corresponding to the newly added critical control point is less than or equal to a preset threshold. This yields the final set of critical control points; Typically, 1 / 5 of the standard deviation of the input sequence is taken. This avoids excessive subdivision leading to computational redundancy while ensuring that key fluctuations with amplitudes ≥ 1 / 5 of the standard deviation are captured. When the Euclidean distance is less than or equal to the threshold, it indicates that the remaining data points have relatively small deviations and no longer have segmentation value, at which point the iteration terminates.

[0033] The final set of critical control points uses data points between any two adjacent critical control points to form an adaptive window. (k=1,2,…,m, where m is the total number of windows), k represents the window number.

[0034] This results in a dynamic adaptation effect where there are more windows and finer divisions in areas of intense fluctuation, and fewer windows and coarser divisions in areas of gentle fluctuation, perfectly matching the multi-scale fluctuation characteristics of the inflow.

[0035] In one embodiment, in each adaptive window Within the data, there is still a certain degree of random fluctuation and uncertainty (such as small sensor errors and small changes in the hydrological environment). Traditional Type-1 fuzzy sets can only quantify "single uncertainty," while Interval Type-2 Fuzzy Sets (IT2FS) can quantify second-order uncertainty through the region enclosed by "upper and lower membership functions." In step S2, for the data within each adaptive window, an Interval Type-2 fuzzy set is constructed, including: Calculate the window mean and window standard deviation based on the data within the adaptive window; The Upper Membership Function (UMF) and Lower Membership Function (LMF) are constructed based on the window mean and window standard deviation, and are expressed by the following formula:

[0036] in, For the membership function, To adapt to the data within the window, The mean of the window. For the expanded window standard deviation, , The standard deviation of the window; ) in, For the subordinate membership function, The standard deviation of the narrowed window. .

[0037] Here, the upper membership function is used to describe the "maximum possible range" of data fluctuations within the window, and the lower membership function is used to describe the "core confidence range" of data fluctuations within the window.

[0038] Quantitative Uncertainty Footprint (FOU): The region bounded by the upper and lower membership functions on the data domain is called the uncertainty footprint (FOU). , For the uncertainty footprint of the k-th window, This indicates uncertainty. The larger the area of ​​FOU, the larger the window. The greater the uncertainty of the internal data (e.g., the larger the FOU area during periods of drastic fluctuation in water flow and the smaller the FOU area during periods of calm), the more effective it is to visualize and quantify the uncertainty of the data.

[0039] In one embodiment, in step S2, the five key features include: the minimum value of the data within the adaptive window. Maximum value U of data within the adaptive window, 90% confidence lower bound 90% confidence upper bound Uncertain width ; Among them, the minimum value of data within the adaptive window The value U reflects the lower limit of the inflow during that period; the maximum value U of the data within the adaptive window reflects the upper limit of the inflow during that period. Among them, the lower bound of 90% confidence To adapt the membership function values ​​within the window The smallest , To adapt to the data within the window, reflecting the lower limit of the core data within the window; 90% confidence upper bound To adapt the membership function values ​​within the window The largest ; Reflects the upper limit of the core data within the window; Uncertain width ,in, To adapt the membership function value within the window The largest , To adapt the membership function value within the window The smallest Directly quantify the degree of uncertainty in window data.

[0040] Arrange the five-dimensional feature vectors of all windows in chronological order to form a dimension of Feature tensor ( (where is the total number of windows and 5 is the feature dimension). This tensor preserves the temporal sequence of water inflow and integrates the statistical boundaries and uncertainty information of each time period, providing structured input for subsequent models to extract deep features.

[0041] In one embodiment, the five dimensions of the feature tensor correspond to different types of information such as "boundary, core range, and uncertainty." Directly mixing these dimensions and inputting them into the model can easily lead to information interference. This step uses a multi-channel temporal convolutional network to achieve targeted extraction of features in each dimension and capture of their spatiotemporal dependencies. Figure 3 A schematic diagram of a multi-channel temporal convolutional network is shown. The multi-channel temporal convolutional network includes: 5 TCN networks and a cross-channel feature fusion module; Each TCN network processes key features of one dimension of the feature tensor; here, five independent TCN (Temporal Convolutional Network) networks are constructed, each TCN processing only one dimension of the feature tensor (e.g., TCN 1 processes all windows). TCN2 handles all windows. (and so on), ensuring that the information in each dimension is not interfered with by other dimensions, thus achieving "precise extraction by dimension".

[0042] The TCN network consists of sequentially connected convolutional layers, batch normalization layers, ReLU activation layers, and dilated convolutional layers; the dilation coefficient sequence of the dilated convolutional layers is set as follows: (i.e., 1, 2, 4, 8). By using an exponentially increasing dilation coefficient, the receptive field is expanded exponentially with a small number of convolutional kernels. For example, with a kernel size of 3, the receptive field of 4 layers of dilated convolutions can reach... This is sufficient to cover more than 80% of the number of windows, effectively capturing the long-term spatiotemporal dependencies between different windows (such as the correlation between the uncertainty of a certain window and the fluctuations of the previous 10 windows).

[0043] The cross-channel feature fusion module is used to concatenate the feature maps output by each TCN network along the feature dimension to obtain a comprehensive feature vector.

[0044] After each TCN channel is processed, the output dimension is: Feature map ( The feature dimension of each channel is typically set to 64 to balance feature expressiveness and computational cost. The feature maps from the five channels are concatenated along this feature dimension to obtain a result with dimension [missing information]. The comprehensive feature vector H integrates deep spatiotemporal features from five dimensions, providing rich information for subsequent time series modeling.

[0045] In one embodiment, although the comprehensive feature vector H contains spatiotemporal dependencies, it is still necessary to further explore the "long-term temporal evolution pattern" of the water inflow (such as seasonal variations, long-term increasing / decreasing trends). A Long Short-Term Memory (LSTM) network is employed, using a gating mechanism to address the "vanishing gradient" problem of traditional recurrent neural networks (RNNs), thus achieving long-term trend memory.

[0046] The core function of LSTM gating mechanism: LSTM regulates the information flow through the coordinated operation of forget gate, input gate, and output gate, without the need for complex mathematical derivation.

[0047] Model parameters and training settings: Network structure: Two hidden layers are set (to balance model capacity and risk of overfitting), with 128 hidden units per layer (sufficient to express complex temporal features), and a dropout layer is introduced (dropout rate = 0.2 to prevent overfitting). Optimized configuration: The Adam optimizer (adaptive learning rate, fast convergence speed) is adopted, the learning rate is set to 0.001, the mean squared error (MSE) is used as the loss function, and overtraining is avoided by using the "early stopping strategy" (training stops if the validation set loss does not decrease for 10 consecutive epochs).

[0048] Output key hidden state: After the LSTM processes the comprehensive feature vector H of all windows, it outputs the hidden state at the last time step. (The dimension is 128, consistent with the number of hidden units). This hidden state encapsulates complete information on "spatiotemporal characteristics - uncertainty evolution - long-term trend" from the beginning of the sequence to the current moment, and is the core decision basis for subsequent interval prediction.

[0049] In one embodiment, traditional prediction methods often output single-point values ​​and cannot provide fluctuation boundaries. By using three independent Support Vector Regression (SVR) models to construct a "lower bound-trend-upper bound" interval prediction system, the inflow boundary output and safety early warning can be achieved.

[0050] The support vector regression model includes three support vector regression units: a lower bound regressor, a trend regressor, and an upper bound regressor. All three SVR models use radial basis functions (RBF) as the kernel function (adapting to the nonlinear characteristics of inflow, with fewer parameters and strong generalization ability). The kernel parameters are optimized through 5-fold cross-validation (typically ranging from 0.1 to 1.0), and the penalty coefficient... (Sensitivity and fitting accuracy of the balanced model to outliers) The lower bound regressor generates a lower bound for predicting the future inflow rate based on the input latent state, while the upper bound regressor generates an upper bound for predicting the future inflow rate based on the input latent state. The trend regressor outputs the trend value of the future inflow rate.

[0051] The inflow prediction network is a trained network. During training, the hidden states output by the LSTM are used. Using these as inputs, three SVR models were trained with different labels respectively: The lower bound regressor uses the minimum value of the data within each adaptive window. Train the system with training labels to generate a lower bound for predicting future inflow rates. ; The trend regressor is trained using the mean of the data within each adaptive window as the training label, and outputs the trend value of the inflow at future time. Used to verify the reasonableness of the interval, if or If necessary, adjust the model parameters to ensure the interval logic is correct. The upper bound regressor is trained using the maximum value U of the data within each adaptive window as the training label to generate the upper limit of the predicted inflow rate at future times. .

[0052] In one embodiment, the 90% confidence prediction interval for the inflow is [ , (Matching the 90% confidence range of LMF, balancing interval coverage and accuracy).

[0053] Safety warning: Preset a safe threshold for water inflow based on the mine's geological conditions (such as aquifer thickness and drainage capacity). (usually 200~300) (The specific amount will be determined based on the actual conditions of the mine). If the upper limit is predicted... The system immediately triggers multi-dimensional early warnings (such as on-site audible and visual alarms, sending SMS notifications to management personnel, and activating backup pumps in conjunction with the drainage system), buying critical response time for coal mine water control.

[0054] Employing the same inventive concept as the inflow prediction method based on adaptive type-two fuzzy information granularization calculation, this embodiment also provides a corresponding inflow prediction device based on adaptive type-two fuzzy information granularization calculation, comprising: The data preprocessing module is used to preprocess the collected raw water flow data to obtain processed water flow data. The feature tensor construction module is used to dynamically segment the processed water flow data using the perceptual importance point algorithm to form multiple adaptive windows; for the data within each adaptive window, an interval type II fuzzy set is constructed; based on the interval type II fuzzy set, five-dimensional key features of each adaptive window are extracted to form a feature tensor; The prediction module is used to construct a water inflow prediction network, which includes a multi-channel temporal convolutional network, a long short-term memory network, and a support vector regression model connected in sequence. The feature tensor is input into the water inflow prediction network, passes through the multi-channel temporal convolutional network to obtain a comprehensive feature vector, and is input into the long short-term memory network to obtain the hidden state. The hidden state is input into the support vector regression model to obtain the lower limit and upper limit of the predicted water inflow at future times. The prediction range determination and alarm module is used to determine the water inflow prediction range based on the lower limit and upper limit of the predicted water inflow at future times, and to trigger a multi-dimensional alarm when the upper limit of the predicted water inflow at future times exceeds the safety threshold.

[0055] The inflow prediction device based on adaptive type II fuzzy information granular calculation in this embodiment has the same inventive concept as the inflow prediction method based on adaptive type II fuzzy information granular calculation described above. Therefore, the specific implementation of this device can be found in the embodiment section of the inflow prediction method based on adaptive type II fuzzy information granular calculation described above, and its technical effects correspond to the technical effects of the above method, so it will not be repeated here.

[0056] To further verify the effectiveness of the method in this application, the following experimental analysis was conducted.

[0057] The experimental data came from the automated monitoring system of a mining area from January 2022 to June 2024, collecting a total of 900 daily average water inflow data points. The dataset included three significant water inrush events triggered by water-conducting faults exposed during mining activities, providing valuable extreme condition samples for model validation. Figure 4 The integrated hydrogeological profile and columnar section of the study area are shown. Figure 5 The raw data of mine water inflow monitoring is shown.

[0058] To systematically verify the effectiveness of the proposed method, four progressive verification experiments were designed: ablation experiments, comparison experiments of different granulation methods, comparison experiments of mainstream prediction models, and strong baseline comparison experiments, supplemented by interval prediction analysis and parameter sensitivity analysis. The dataset was divided into a training set (first 630 days) and a test set (last 270 days) in a 7:3 ratio. The evaluation metrics used were the mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²).

[0059] Experimental objectives and verification methods The experiment focuses on "stable prediction of water inflow time series under sudden water inrush disturbance" and answers three key questions: First, can adaptive type II fuzzy granulation better characterize the trend, fluctuation and uncertainty of water inflow sequence? Second, can the TCN-LSTM-SVM coupled model maintain high accuracy under both normal operating conditions and peak water inrush stage? Third, can the interval prediction results provide interpretable upper and lower bounds for safety early warning?

[0060] Therefore, the experiment not only compares the average error across the entire test set, but also separately examines peak tracking capability near the water inrush event, the contribution of the granulation module, the differences between different granulation strategies, model structure ablation, and the sensitivity of the granulation window parameters. This design avoids drawing one-sided conclusions based on a single metric.

[0061] Table 1. Correspondence between Experimental Design and Validation Objectives

[0062] Table 2 Dataset, Preprocessing, and Leakage Prevention Settings

[0063] Table 3 Evaluation Indicators and Experimental Output

[0064] Experimental Procedure and Result Interpretation Guidelines The specific experimental procedure is as follows: First, the original daily average water inflow sequence is time-aligned, missing data is checked, and training set statistics are normalized; then, PIPs are used to adaptively divide the granular window to perceive important points, and five-dimensional key features are constructed, including minimum value, maximum value U, 90% confidence lower bound, 90% confidence upper bound, and uncertainty width; then, the granular features are input into the TCN-LSTM structure to extract local fluctuations and long-term memory, and prediction boundary and safety warning information are output through SVM; finally, error index, interval index, and local index of water inrush event are uniformly calculated on the test set.

[0065] The interpretation of results should follow a three-tiered logic: "global accuracy + sudden flood stage + uncertainty." If a model has low error during stable periods but significantly lags or underestimates the peak value near the sudden flood, its engineering application value remains limited. If the prediction interval coverage is high but the interval is too wide, it indicates that the early warning results are conservative but lack operability. Therefore, the advantages of the method in this application should be reflected simultaneously in low error, peak tracking, and reasonable interval width.

[0066] Experimental Hypothesis and Judgment Criteria To make the experimental verification more targeted, this application decomposes the effectiveness of the method into several testable hypotheses. Each hypothesis corresponds to a specific control experiment or evaluation index, avoiding reliance on a single result figure for generalization.

[0067] Table 4 Research Hypotheses, Corresponding Experiments, and Pass Criteria

[0068] Data statistics, event segmentation and evaluation units The challenge in predicting water inflow lies not only in fitting the small error during the steady-state phase, but also in capturing the risks near high inflow and peak values. Therefore, in addition to the full test set indicators, this application uses 250 m³ / h as a safety threshold and treats processes exceeding the threshold as separate event-level evaluation units.

[0069] Table 5. Data Overview and Training / Test Split Statistics

[0070] Table 6 Classification and Peak Information of High Water Inrush Events

[0071] Fair comparison, parameter tuning agreement and leakage prevention measures To ensure fairness in comparisons between different models, this experiment requires all methods to use the same training / test split, the same evaluation metric, and the same prediction step size. Parameter tuning and data preprocessing can only be performed on the training set or a reserved validation set; the model structure or thresholds cannot be repeatedly modified based on the performance on the test set.

[0072] Table 7 Training, Coordination, and Leakage Prevention Protocol

[0073] Evaluation Indicator Formulas and Reporting Standards Table 8. Formulas and Explanations of Key Evaluation Indicators

[0074] Experimental reliability control and conclusion boundaries This dataset comes from engineering monitoring records of a single mining area, and the boundaries of the conclusions must be clearly defined in the report: First, the research conclusions are mainly applicable to the current geological conditions, monitoring scope, and mining stage of this mining area; second, fields such as drainage capacity and pump operation can only be included in the model if they can be obtained in advance at the prediction time or constructed as lagging features; third, deep model training is affected by random initialization, and the results reporting adopts unified partitioning, unified indicators, and fixed random seed control. The above control measures are used to ensure the reproducibility of experimental comparisons and the consistency of engineering interpretation.

[0075] Table 9 Experimental Reliability Risks and Control Measures

[0076] Experiment 1: Ablation Experiment Ablation experiments are a key means of verifying the contribution of each core module of the model. This experiment constructs four experimental variants by sequentially removing key modules in this application to quantify the independent contribution of each module to the final prediction performance. The four variants are: (1) IT2FGC-TCN-LSTM-SVM, i.e., the complete method; (2) TCN-LSTM-SVM, removing the type II fuzzy granulation module and directly inputting the original time series into the prediction network; (3) IT2FGC-LSTM-SVM, removing the temporal convolutional network (TCN) module and directly processing the granulated features by LSTM; (4) IT2FGC-TCN-LSTM, removing the SVM interval output module and directly outputting the single-point prediction value by LSTM.

[0077] Figure 6 The diagram illustrates the contribution of each module to the prediction performance in the ablation experiment. Figure 7The performance changes after removing each module in the ablation experiment are shown. The complete method has a MAE of 3.27 m³ / h and an R² of 0.9883, which is the best among all variants. After removing the granulation module (variant 2), the MAE increases to 6.45 m³ / h, and the R² decreases to 0.9507, showing the largest performance decline. This indicates that the Type II fuzzy granulation module is the core driving force for the performance improvement of this application, and its structured representation and uncertainty quantification of the original data provide an important information foundation for subsequent predictions. After removing the TCN module (variant 3), the MAE increases to 4.88 m³ / h, and the R² decreases to 0.9726, indicating that TCN plays an important role in multi-dimensional spatiotemporal feature extraction. After removing the SVM interval output (variant 4), the MAE increases slightly to 3.54 m³ / h, and the R² decreases to 0.9855. Although the decrease in single-point prediction accuracy is relatively small, this variant loses the core functions of interval prediction and safety warning, which limits its engineering application value.

[0078] Figure 8 The diagram illustrates the prediction details for various variants during the water inrush event in the ablation experiment. It is clearly visible that removing the granulation module significantly reduces the model's ability to capture the water inrush peak, and the deviation between the predicted and measured curves increases significantly. In contrast, the complete method, thanks to its type-II fuzzy granulation's keen perception of abrupt change information, maintains high prediction accuracy in the peak region.

[0079] Experiment 2: Comparison of different granulation methods This experiment aims to compare the prediction performance of "adaptive type II fuzzy granulation" with other granulation strategies. Under the premise of keeping the back-end prediction model (TCN-LSTM-SVM) completely consistent, this experiment only replaced the front-end granulation method and constructed four sets of comparative experiments: (1) IT2FGC (adaptive type II fuzzy granulation), which is the strategy adopted in this application; (2) T1FIG (type I fuzzy granulation), which uses the traditional type I Gaussian fuzzy set for granulation; (3) fixed window granulation, which uses an equal-length fixed window for granulation; (4) no granulation, which directly inputs the original time series into the model.

[0080] To visually demonstrate the differences in how different granulation methods process raw data, Figure 9 The data representation comparison of different granulation methods is shown. It can be seen that Type I fuzzy granulation (T1FIG) uses a fixed equal-length window, which can only generate a single fuzzy interval consisting of an upper and lower bound. In contrast, the Type II fuzzy granulation (IT2FGC) of this application uses a PIPs adaptive window, the window size of which automatically adjusts with data fluctuations, and constructs a Type II fuzzy set containing the FOU (Uncertainty Footprint) within each window, which can additionally quantify the "uncertainty of uncertainty itself".

[0081] Figure 10The impact of different granulation methods on prediction performance is shown. Figure 11 The comparison of prediction performance metrics for different granulation methods is shown.

[0082] The IT2FGC granulation strategy adopted in this application achieved the best performance, with an MAE of 3.27 m³ / h and an R² of 0.9883. Type-I fuzzy granulation (T1FIG) achieved an MAE of 4.98 m³ / h and an R² of 0.9722, a 52% increase compared to this application, indicating that the advantage of Type-II fuzzy sets in uncertainty representation directly translates into improved prediction accuracy. Fixed-window granulation achieved an MAE of 5.22 m³ / h and an R² of 0.9703, showing a further decline in performance, indicating that PIPs adaptive window partitioning has better adaptability than fixed-window partitioning. The non-granulation scheme achieved the highest MAE (6.45 m³ / h) and the lowest R² (0.9507), further confirming the importance of granulation for improving prediction performance.

[0083] In a local comparison during a sudden water inrush event, the IT2FGC granularization scheme showed the most accurate tracking of the peak value, while other granularization schemes all exhibited varying degrees of peak value underestimation or response lag. Figure 12 The comparison of predictions using different granulation methods during water inrush events is shown.

[0084] Experiment 3: Comparison with mainstream prediction models After verifying the effectiveness of each module and granulation strategy, this application compares the complete method with four commonly used deep learning prediction models. The comparison models include: (1) LSTM, the classic long short-term memory network; (2) CNN-LSTM, an LSTM model that combines convolutional neural networks for feature extraction; (3) LSTM-Transformer, an LSTM model that incorporates the Transformer attention mechanism; and (4) SMA-LSTM-GRU-Attention, a model that combines multiple optimization strategies and hybrid network structures.

[0085] Figure 13 The comparison of prediction results of different prediction models on the test set is shown. Figure 14 The performance metrics of different prediction models are compared.

[0086] This application outperforms other comparative models across all evaluation metrics. Specifically, the MAE of this application is 3.27 m³ / h, RMSE is 4.04 m³ / h, MAPE is 1.33%, and R² is 0.9883. Compared to the worst-performing LSTM model, the MAPE of this application is reduced by approximately 78%; compared to the second-best-performing SMA-LSTM-GRU-Attention model, the MAPE is also reduced by approximately 38%. This demonstrates the significant advantage of this application in prediction accuracy.

[0087] Figure 15 The radar chart showing the overall performance of each model is displayed. Figure 15 The performance of each model across five dimensions—MAE, RMSE, MAPE, R², and water inrush response—is comprehensively presented in radar chart format. This application (green area) achieves optimal performance across all dimensions, covering the largest area, and its overall performance surpasses that of other models.

[0088] Interval forecasting and safety early warning capabilities Besides its high single-point prediction accuracy, the greatest advantage of this application lies in its ability to provide prediction intervals, offering a basis for risk management decisions. As shown in the figure below, the 90% prediction interval generated by this application successfully includes the vast majority of measured values ​​with a coverage rate of 90.4%. When the safety threshold is set at 250 m³ / h, the system will automatically trigger an early warning (red area in the figure) once the upper limit of the prediction interval exceeds this threshold. This function enables mine managers to anticipate potential water inrush risks and take corresponding preventative measures, demonstrating significant engineering application value. Figure 16 The diagram illustrates the interval prediction and safety warning of this application.

[0089] Parameter sensitivity analysis To verify the impact of the key parameter in this application—the granulation window size—on prediction performance, this experiment tested the prediction effects with window sizes of 5, 10, 15, 20, 30, 50, and 80, while keeping other parameters constant. As shown in the figure below, when the window is too small (e.g., 5), the granulated information particles are too fragmented, making it difficult to effectively extract trend features, resulting in a high MAE. When the window is too large (e.g., 50, 80), too much detailed information is smoothed out, and the prediction performance decreases significantly. The optimal performance occurs in the window size range of approximately 15-20, which coincides with the average window size adaptively determined by the PIPs algorithm in this application, verifying the rationality and effectiveness of the PIPs adaptive window partitioning strategy. Figure 17 Sensitivity analysis of the granulation window size parameter is shown.

[0090] Experiment 4: Strong Baseline Contrast Experiment To avoid overestimating the method's advantages by comparing it only with complex deep learning models, this application further sets up four strong baselines: ridge regression autoregression, previous day persistence prediction, 7-day moving average, and 30-day moving average. All baselines use only historical information available before the test date, and the training / test split is consistent with the main experiment. The results show that the proposed method outperforms the strong baselines in all four metrics: MAE, RMSE, MAPE, and R², indicating that the performance improvement is not solely due to sequence autocorrelation. Figure 18 The prediction performance of the strong baseline model and the method in this application are compared.

[0091] Table 10 Comparison of prediction performance between the strong baseline model and the method in this application

[0092] In summary, based on 900 days of daily monitoring data from a mining area from January 2022 to June 2024, this application constructs an adaptive type-II fuzzy particle computation and TCN-LSTM-SVM coupled model verification framework for mine water inflow prediction. Test set results show that the proposed method achieves prediction performance of MAE=3.27 m³ / h, RMSE=4.04 m³ / h, MAPE=1.33%, and R²=0.9883, with low overall error and stable fitting effect.

[0093] Ablation experiments show that the adaptive type-II fuzzy granular computation module contributes the most to the structured characterization and uncertainty assessment of the sequence; the TCN module enhances the ability to capture local fluctuations and peak edges during the water inrush stage; and the SVM output further provides explanations for interval boundaries and early warnings. Comparison results with different granulation strategies and mainstream prediction models collectively demonstrate that the proposed method has good predictive advantages in both the normal fluctuation stage and the high water inrush risk stage.

[0094] Interval prediction results show that 90% of the prediction intervals can include the measured values ​​with a coverage rate of 90.4%. When the safety threshold is set at 250 m³ / h, the upper bound of the prediction can provide an advance warning for water inrush risk. Parameter sensitivity analysis indicates that a PIPs granulation window of around 20 performs optimally overall; a window that is too small will lead to information fragmentation, while a window that is too large will weaken the details of abrupt changes. The conclusions of this application are applicable to the current monitoring caliber and data period of this mining area. During engineering deployment, it is necessary to continue to verify the time sequence and strictly exclude information leakage caused by response variables on the same day.

[0095] The beneficial effects of this application are mainly reflected in the accurate solution to the core pain point of mine water inflow prediction and the realization of its practical application value, which can be summarized in three aspects: 1. It can efficiently adapt to complex data characteristics and deeply quantify uncertainty. Addressing the characteristics of non-stationary, high-noise, and multi-scale fluctuations in water inflow, the PIPs algorithm is used to achieve adaptive window partitioning—the window is denser during periods of drastic data fluctuation to capture key changes, while the window is sparser during periods of calm to avoid redundant computation, completely solving the information loss problem caused by traditional fixed windows, making the data partitioning more closely match the actual evolution of water inflow; simultaneously, it utilizes Interval Type II Fuzzy Sets (IT2FS) to construct a multi-layered description system of "upper membership function - lower membership function - uncertainty footprint," accurately quantifying the "uncertainty of uncertainty" (second-order uncertainty), significantly improving the model's prediction stability in high-noise scenarios compared to traditional Type I fuzzy sets.

[0096] 2. Multi-model collaboration significantly improves prediction accuracy and enables interval prediction and proactive early warning. Breaking through the limitations of single models, multi-channel TCN can extract spatiotemporal dependencies across dimensions, LSTM can effectively memorize the long-term evolution trend of water inflow (such as seasonal patterns), and granular SVM outputs robust prediction boundaries. The coupling of these three models results in significantly higher prediction accuracy than traditional single models. Furthermore, through the interval design of "lower bound - trend - upper bound," it not only provides a reference for optimizing drainage system energy consumption (with the lower bound guiding pump scheduling), but also triggers multi-dimensional early warnings in real time when the upper limit of the prediction exceeds the safety threshold, gaining crucial response time for water control measures. This upgrades the model from a simple data prediction tool to a core support for coal mine safety decision-making.

[0097] 3. It combines strong practicality and comprehensive benefits, demonstrating significant potential for widespread application. The entire technical process is simplified and efficient, with data preprocessing and model training supporting end-to-end operation. It can directly connect to real-time mine monitoring systems, significantly reducing the operational threshold for on-site personnel. The core technology is not limited to coal mine water inflow prediction but can also be applied to similar time-series prediction scenarios such as groundwater levels and tunnel water inflow, demonstrating strong generalization capabilities. Economically, it can effectively reduce energy consumption through dynamic scheduling of the drainage system; in terms of safety, it can reduce losses such as equipment damage and operational interruptions caused by water hazards, truly achieving a balance between economic and safety benefits, and is of significant practical importance for ensuring safe coal mine production.

[0098] The above descriptions are merely various embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting water inflow based on adaptive type-II fuzzy information granularization calculation, characterized in that, include: The collected raw water flow data is preprocessed to obtain processed water flow data; The processed water inflow data is dynamically segmented using a perceptual importance point algorithm to form multiple adaptive windows; for the data within each adaptive window, an interval type II fuzzy set is constructed; based on the interval type II fuzzy set, five-dimensional key features of each adaptive window are extracted to form a feature tensor; A water inflow prediction network is constructed, comprising a multi-channel temporal convolutional network, a long short-term memory network, and a support vector regression model connected in sequence. The feature tensor is input into the water inflow prediction network, passes through the multi-channel temporal convolutional network to obtain a comprehensive feature vector, and is input into the long short-term memory network to obtain a hidden state. The hidden state is input into the support vector regression model to obtain the lower limit and upper limit of the predicted water inflow at future times. The predicted water inflow range is determined based on the predicted lower limit and the predicted upper limit of the water inflow at future times, and a multi-dimensional alarm is triggered when the predicted upper limit of the water inflow at future times exceeds the safety threshold.

2. The method as described in claim 1, characterized in that, in, The processed water inflow data is dynamically segmented using a significance point perception algorithm to form multiple adaptive windows, including: The water inflow corresponding to the start and end times of the processed water inflow data is selected as the initial key control points to form a set of key control points. For any two adjacent critical control points in the current critical control point set and Calculate the distance from each data point to the line segment between two key control points. The Euclidean distance is used to select the data point with the largest Euclidean distance as the new key control point and add it to the key control point set to form a new key control point set; the process is repeated until the Euclidean distance corresponding to the new key control point is less than or equal to the preset threshold, and the final key control point set is obtained. The data points between any two adjacent key control points in the final key control point set form an adaptive window.

3. The method as described in claim 1, characterized in that, in, For the data within each adaptive window, construct interval type II fuzzy sets, including: Calculate the window mean and window standard deviation based on the data within the adaptive window; Based on the window mean and window standard deviation, upper and lower membership functions are constructed, expressed by the following formulas: in, For the membership function, To adapt to the data within the window, The mean of the window. For the expanded window standard deviation, , The standard deviation of the window; ) in, For the subordinate membership function, The standard deviation of the narrowed window. .

4. The method as described in claim 1, characterized in that, The five key features include: the minimum value of data within the adaptive window. Maximum value U of data within the adaptive window, 90% confidence lower bound 90% confidence upper bound Uncertain width ; Among them, the lower bound of 90% confidence To adapt the membership function values ​​within the window The smallest , To adapt to the data within the window; 90% confidence upper bound To adapt the membership function values ​​within the window The largest ; Uncertain width ,in, To adapt the membership function value within the window The largest , To adapt the membership function value within the window The smallest .

5. The method as described in claim 1, characterized in that, The multi-channel temporal convolutional network includes: 5 TCN networks and a cross-channel feature fusion module; Each TCN network processes key features of one dimension of the feature tensor; The cross-channel feature fusion module is used to concatenate the feature maps output by each TCN network along the feature dimension to obtain a comprehensive feature vector.

6. The method as described in claim 5, characterized in that, The TCN network comprises a convolutional layer, a batch normalization layer, a ReLU activation layer, and a dilated convolutional layer connected in sequence; the dilation coefficient sequence of the dilated convolutional layer is set as follows: .

7. The method as described in claim 1, characterized in that, The support vector regression model includes three support vector regression units: a lower bound regressor, a trend regressor, and an upper bound regressor. The lower bound regressor generates a lower bound for predicting the future inflow rate based on the input latent state, and the upper bound regressor generates an upper bound for predicting the future inflow rate based on the input latent state; the trend regressor is used to output the trend value of the future inflow rate.

8. The method as described in claim 7, wherein the inflow prediction network is a trained network, and during training, the lower bound regressor uses the minimum value of the data within each adaptive window. To train the training labels, the trend regressor is trained using the mean of the data within each adaptive window as the training label, and the upper bound regressor is trained using the maximum value U of the data within each adaptive window as the training label.

9. The method as described in claim 1, characterized in that, The preprocessing includes data continuity restoration, outlier identification and removal, and data validity verification.

10. A water inflow prediction device based on adaptive type-II fuzzy information granular calculation, characterized in that, include: The data preprocessing module is used to preprocess the collected raw water flow data to obtain processed water flow data. The feature tensor construction module is used to dynamically segment the processed water inflow data using a perceptual importance point algorithm to form multiple adaptive windows; for the data within each adaptive window, an interval type II fuzzy set is constructed; based on the interval type II fuzzy set, five-dimensional key features of each adaptive window are extracted to form a feature tensor; The prediction module is used to construct a water inflow prediction network, which includes a multi-channel temporal convolutional network, a long short-term memory network, and a support vector regression model connected in sequence. The feature tensor is input into the water inflow prediction network, passes through the multi-channel temporal convolutional network to obtain a comprehensive feature vector, and is input into the long short-term memory network to obtain a hidden state. The hidden state is input into the support vector regression model to obtain the lower limit and upper limit of the predicted water inflow at future times. The prediction interval determination and alarm module is used to determine the water inflow prediction interval based on the predicted lower limit and the predicted upper limit of the water inflow at the future time, and to trigger a multi-dimensional alarm when the predicted upper limit of the water inflow at the future time exceeds the safety threshold.

Images