A coal mine underground medium identification method, system, terminal and storage medium based on heterogeneous model fusion

Abstract

The application relates to the technical field of medium identification underground coal mines, and discloses a coal mine underground medium identification method, system, terminal and storage medium based on heterogeneous model fusion, the method comprising the following steps: acquiring medium data to perform signal modulation, obtaining a digital signal; performing standardization and extension on the digital signal to obtain standard features and extended features, splicing the standard features and the extended features to obtain an input feature vector; inputting the input feature vector into a first base model, a second base model and a third base model respectively to perform reasoning, obtaining a first probability vector, a second probability vector and a third probability vector; and fusing the first probability vector, the second probability vector and the third probability vector through a meta-learner to obtain a classification result. The application performs reasoning on sensor data through three kinds of heterogeneous models, and fuses the reasoning results of the three kinds of models through a meta-learner to obtain a medium classification result, thereby improving the accuracy of medium identification.

Description

Technical Field

[0001] This invention relates to the field of underground coal mine media identification technology, and in particular to a method, system, terminal, and computer-readable storage medium for underground coal mine media identification based on heterogeneous model fusion. Background Technology

[0002] In underground coal mining and tunneling operations, real-time and accurate perception of the working face environment is crucial for achieving intelligent and unmanned mining. For example, the adaptive adjustment of the coal mining machine drum height requires precise identification of the coal-rock interface to avoid cutting the roof and floor rocks, which would cause equipment wear and increase the gangue content of the coal. Traditional manual inspection and experience-based judgment methods have inherent defects such as low efficiency, poor real-time performance, and high risk, and can no longer meet the development requirements of smart mines.

[0003] Existing coal mine underground media identification methods typically construct a physical model based on the relationship between the equivalent dielectric constant of the mixed medium and the volume fraction of each component. The equivalent dielectric constant of the multi-medium mixture under test is measured using sensors. Based on the physical model and the measured equivalent dielectric constant, the proportion of each component in the multi-medium mixture under test is deduced. However, this coal mine underground media identification method can usually only perform simple component inversion and cannot accurately distinguish the detailed state of the medium, such as distinguishing between "high-concentration coal slurry water" and "low-concentration coal slurry water", which leads to a decrease in the accuracy of media identification.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a method, system, terminal, and computer-readable storage medium for identifying underground coal mine media based on heterogeneous model fusion. This invention aims to solve the problem that existing methods for identifying underground coal mine media can only perform simple component inversion and cannot accurately distinguish the detailed states of the media, resulting in low accuracy in media identification.

[0006] To achieve the above objectives, the present invention provides a method for identifying underground media in coal mines based on heterogeneous model fusion, comprising the following steps: Acquire the medium data of the medium to be tested in the coal mine, and perform signal modulation processing on the medium data to obtain a digital signal; The digital signal is standardized to obtain standard features, the standard features are expanded to obtain extended features, and the standard features and the extended features are concatenated to obtain the input feature vector. Determine the first base model, the second base model, and the third base model, and input the input feature vector into the first base model, the second base model, and the third base model respectively for inference to obtain the first probability vector, the second probability vector, and the third probability vector. A meta-learner is determined, and the first probability vector, the second probability vector, and the third probability vector are fused using the meta-learner to obtain the medium classification result in the coal mine.

[0007] Optionally, in the coal mine underground medium identification method based on heterogeneous model fusion, the medium data includes dielectric property data and environmental data; The process of acquiring media data of the medium to be detected in the coal mine and performing signal modulation processing on the media data to obtain a digital signal specifically includes: The equivalent capacitance, resonant frequency, impedance, and phase shift of the medium under test are collected by a capacitance sensor, and the equivalent capacitance, resonant frequency, impedance, and phase shift are used as the dielectric characteristic data. Ambient temperature and humidity are collected by environmental sensors, and the ambient temperature and humidity are used as the environmental data. The dielectric property data and the environmental data are filtered, amplified, and converted to obtain a digital signal.

[0008] Optionally, the coal mine underground medium identification method based on heterogeneous model fusion, wherein the standardization processing of the digital signal to obtain standard features, the expansion processing of the standard features to obtain expanded features, and the concatenation processing of the standard features and the expanded features to obtain an input feature vector, specifically includes: Calculate the arithmetic mean and standard deviation of the digital signal, and perform range calculation processing based on the arithmetic mean and standard deviation to obtain the normal signal range; The digital signal is subjected to outlier replacement processing based on the normal signal range to obtain the original signal; The original signal is subjected to mean normalization to obtain standard features; Obtain the length of the sliding window, and expand the standard feature according to the length of the sliding window to obtain the expanded feature; The standard features and the extended features are concatenated to obtain the input feature vector.

[0009] Optionally, the coal mine underground media identification method based on heterogeneous model fusion, wherein obtaining the sliding window length and expanding the standard features according to the sliding window length to obtain expanded features specifically includes: Obtain the length of the sliding window, and perform segmentation processing on the standard feature according to the length of the sliding window to obtain multiple feature segments; Statistical feature calculations are performed on the core physical quantities of multiple feature segments to obtain multiple extended statistical features; The extended features are obtained based on multiple extended statistical features.

[0010] Optionally, the coal mine underground media identification method based on heterogeneous model fusion, wherein determining the first basic model, the second basic model, and the third basic model, and inputting the input feature vector into the first basic model, the second basic model, and the third basic model respectively for inference to obtain the first probability vector, the second probability vector, and the third probability vector, specifically includes: A first base model is determined, and a first probability vector is obtained by performing multi-level decision tree prediction processing based on the input feature vector using the first base model. A second base model is determined, and a second probability vector is obtained by performing one-dimensional convolutional prediction processing based on the input feature vector using the second base model. A third base model is determined, and a third probability vector is obtained by performing kernel mapping classification processing on the input feature vector using the third base model.

[0011] Optionally, the coal mine underground media identification method based on heterogeneous model fusion, wherein the determination of the meta-learner, by fusing the first probability vector, the second probability vector, and the third probability vector through the meta-learner to obtain the coal mine underground media classification result, specifically includes: The first probability vector, the second probability vector, and the third probability vector are concatenated to obtain the meta-feature vector. A meta-learner is determined, and the meta-feature vector is linearly transformed by the meta-learner to obtain a score vector; The score vector is then subjected to probability normalization to obtain the target probability vector; Based on the target probability vector, a category decision is made to obtain the medium classification result.

[0012] Optionally, the coal mine underground media identification method based on heterogeneous model fusion, wherein the determination of the meta-learner, by performing a linear transformation on the meta-feature vector through the meta-learner to obtain a score vector, further includes: Obtain a validation set, determine the initial meta-learner and loss function, and iteratively train the initial meta-learner based on the validation set, the first probability vector, the second probability vector, the third probability vector, and the loss function. When the loss function converges, the target linear transformation parameters are obtained. The initial meta-learner is parameterized according to the target linear transformation parameters to obtain the meta-learner.

[0013] Furthermore, to achieve the above objectives, the present invention also provides a coal mine underground media identification system based on heterogeneous model fusion, wherein the coal mine underground media identification system based on heterogeneous model fusion includes: The signal modulation module is used to acquire the medium data of the medium to be detected in the coal mine, and to perform signal modulation processing on the medium data to obtain a digital signal; The feature generation module is used to standardize the digital signal to obtain standard features, expand the standard features to obtain expanded features, and concatenate the standard features and the expanded features to obtain an input feature vector. The heterogeneous model inference module is used to determine the first base model, the second base model and the third base model, and to input the input feature vector into the first base model, the second base model and the third base model respectively for inference to obtain the first probability vector, the second probability vector and the third probability vector. The medium classification module is used to determine the meta-learner, and to fuse the first probability vector, the second probability vector, and the third probability vector through the meta-learner to obtain the medium classification result in the coal mine.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a coal mine underground media identification program based on heterogeneous model fusion stored in the memory and executable on the processor. When the coal mine underground media identification program based on heterogeneous model fusion is executed by the processor, it implements the steps of the coal mine underground media identification method based on heterogeneous model fusion as described above.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a coal mine underground media identification program based on heterogeneous model fusion, and the coal mine underground media identification program based on heterogeneous model fusion, when executed by a processor, implements the steps of the coal mine underground media identification method based on heterogeneous model fusion as described above.

[0016] In this invention, medium data of the medium to be detected in a coal mine is acquired. The medium data is then subjected to signal modulation processing to obtain a digital signal. The digital signal is standardized to obtain standard features. These standard features are then expanded to obtain extended features. The standard features and the extended features are concatenated to obtain an input feature vector. A first, second, and third basic model are determined. The input feature vector is input into each of these models for inference, resulting in a first, second, and third probability vectors. A meta-learner is then determined, and these three probability vectors are fused to obtain a medium classification result for the coal mine. This invention uses three heterogeneous models to infer from sensor data and fuses the inference results of the three models using a meta-learner to obtain a medium classification result, thus improving the accuracy of medium identification. Attached Figure Description

[0017] Figure 1 This is a flowchart of a preferred embodiment of the coal mine underground media identification method based on heterogeneous model fusion of the present invention; Figure 2 This is a system overall flowchart of the coal mine underground media identification method based on heterogeneous model fusion according to the present invention; Figure 3 This is a comparison chart of data before and after standardization for the coal mine underground media identification method based on heterogeneous model fusion according to the present invention; Figure 4 This is a schematic diagram of sliding window feature extraction in the coal mine underground medium identification method based on heterogeneous model fusion of the present invention. Figure 5 This is a schematic diagram of the feature set composition of the coal mine underground media identification method based on heterogeneous model fusion according to the present invention; Figure 6 This is a flowchart of the XGBoost gradient boosting process for the coal mine underground media identification method based on heterogeneous model fusion, as described in this invention. Figure 7 This is a schematic diagram of one-dimensional convolutional feature extraction in the coal mine underground media identification method based on heterogeneous model fusion of the present invention; Figure 8 This is a schematic diagram illustrating the Gaussian kernel support vector machine classification principle of the coal mine underground media identification method based on heterogeneous model fusion according to the present invention. Figure 9 This is a structural diagram of a preferred embodiment of the coal mine underground media identification system based on heterogeneous model fusion of the present invention; Figure 10 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] In underground coal mining and tunneling operations, real-time and accurate perception of the working face environment is crucial for achieving intelligent and unmanned mining. For example, the adaptive adjustment of the coal mining machine drum height requires precise identification of the coal-rock interface to avoid cutting the roof and floor rocks, which would cause equipment wear and increase the gangue content of the coal. Traditional manual inspection and experience-based judgment methods have inherent defects such as low efficiency, poor real-time performance, and high risk, and can no longer meet the development requirements of smart mines.

[0020] Existing coal mine underground media identification methods typically construct a physical model based on the relationship between the equivalent dielectric constant of the mixed medium and the volume fraction of each component. The equivalent dielectric constant of the multi-medium mixture under test is measured using sensors. Based on the physical model and the measured equivalent dielectric constant, the proportion of each component in the multi-medium mixture under test is deduced. However, this coal mine underground media identification method can usually only perform simple component inversion and cannot accurately distinguish the detailed state of the medium, such as distinguishing between "high-concentration coal slurry water" and "low-concentration coal slurry water", which leads to a decrease in the accuracy of media identification.

[0021] To address the aforementioned issues, this invention proposes a method for identifying underground coal mine media based on heterogeneous model fusion. This method uses three heterogeneous models to infer from sensor data, and then fuses the inference results of the three models using a meta-learner to obtain the media classification result, thereby improving the accuracy of media identification.

[0022] The preferred embodiment of the coal mine underground media identification method based on heterogeneous model fusion described in this invention, such as... Figure 1 As shown, the coal mine underground media identification method based on heterogeneous model fusion includes the following steps: Step S10: Obtain the medium data of the medium to be tested in the coal mine, and perform signal modulation processing on the medium data to obtain a digital signal.

[0023] To address the shortcomings of existing technologies, such as poor environmental adaptability of single physical models, coarse multi-media classification, and lack of dynamic feature extraction capabilities, this invention first collects multi-dimensional raw features using capacitive sensors and auxiliary sensors, and constructs an extended feature set containing statistical features. Then, it runs three heterogeneous base models in parallel: XGBoost gradient boosting tree, one-dimensional convolutional neural network (1D-CNN), and kernel support vector machine (SVM). These models extract differentiated features and output predicted probabilities from three dimensions: static nonlinear mapping, temporal dynamic fluctuations, and small-sample generalization. Finally, through a stacking ensemble strategy, a secondary logistic regression meta-learner is trained to dynamically fuse the advantages of the three base models, outputting the final media classification result.

[0024] Specifically, the dielectric data includes dielectric property data and environmental data; the equivalent capacitance, resonant frequency, impedance, and phase shift of the dielectric to be tested are collected by a capacitance sensor, and the equivalent capacitance, resonant frequency, impedance, and phase shift are used as the dielectric property data; the ambient temperature and ambient humidity are collected by an environmental sensor, and the ambient temperature and ambient humidity are used as the environmental data.

[0025] like Figure 2 As shown, the medium data of the medium to be identified is first collected from the capacitive sensor and the environmental sensor to form a 6-dimensional raw signal.

[0026] The sensor of the present invention is deployed in a sensing terminal at a key location on the body of a coal mining machine, a hydraulic support, or a roadway. It includes a capacitance sensing unit and an environmental sensing unit. The capacitance sensing unit is used to collect the equivalent capacitance (C), resonant frequency (f), impedance (Z), and phase shift (θ) that reflect the dielectric properties of the medium. The environmental sensing unit is used to simultaneously collect the underground ambient temperature (T) and ambient humidity (H).

[0027] Furthermore, the dielectric property data and the environmental data are filtered, amplified, and converted to obtain a digital signal.

[0028] The acquired 6-dimensional raw analog signal is subjected to multi-stage filtering (to suppress electromagnetic pulse interference from downhole frequency converters, etc.), amplification, and analog-to-digital conversion to output a digital signal.

[0029] Step S20: Standardize the digital signal to obtain standard features, expand the standard features to obtain expanded features, and concatenate the standard features and the expanded features to obtain the input feature vector.

[0030] Specifically, the arithmetic mean and standard deviation of the digital signal are calculated, and a range calculation is performed based on the arithmetic mean and standard deviation to obtain a normal signal range; outlier replacement processing is performed on the digital signal based on the normal signal range to obtain the original signal; mean normalization processing is performed on the original signal to obtain standard features.

[0031] like Figure 2 As shown, outliers in the original 6-dimensional signal are removed and the data is standardized. The standardized signal is then expanded using the sliding window statistical method, resulting in a total output of a 15-dimensional feature vector.

[0032] The raw 6-dimensional signals directly output by capacitive and environmental sensors have three inherent limitations that prevent them from being directly used as input for subsequent recognition models: 1. The frequent start-stop of high-power electromechanical equipment (such as frequency converters, coal mining machine motors, etc.) in underground mines will generate strong electromagnetic pulses. After these pulses are captured by sensors, they will form short-term, large-amplitude abnormal jump points (commonly known as outliers) in the continuous signal. If they are directly sent into the model without processing, it will lead to occasional errors in the recognition results. 2. Six-dimensional signals have completely different physical dimensions and numerical ranges. For example, the impedance can be on the order of 10⁵ to 10⁵ ohms or more, while the phase shift is only within a few degrees. If such features with huge numerical differences are directly fed into the machine learning model, the features with large values ​​will unreasonably dominate the model's decision, causing the features with small values ​​but important physical meaning to be ignored, thus reducing the classification accuracy. 3. The single-point sampling value at the current moment only reflects the static characteristics of the measured medium at this moment, and cannot characterize the dynamic evolution process of the medium state that is taking place. For example, key patterns such as rapid capacitance fluctuations caused by water flow and slow signal drift caused by gradual coal dust deposition cannot be captured, resulting in lag and misjudgment in the identification of dynamic operating conditions.

[0033] To overcome the above-mentioned shortcomings, this invention specifically designs feature engineering and feature set construction steps after signal acquisition and before model inference.

[0034] First, outliers caused by external interference such as electromagnetic pulses are automatically identified and removed to ensure the continuity and stability of the input data. Specifically, the three sigma (3σ) criterion in statistics is used to achieve this.

[0035] The Three Sigma criterion is an outlier identification method based on the assumption of a normal distribution. For a continuously acquired signal sequence, its arithmetic mean (denoted as μ) and standard deviation (denoted as σ) are first calculated over an observation interval. The standard deviation reflects the amplitude of normal fluctuations in the signal. According to the statistical laws of the normal distribution, approximately 99.7% of the sampling points should fall within the interval [μ...]. Within [3σ, μ+3σ]. This invention defines this range as the "normal fluctuation range". Once the value of the current sampling point exceeds this range, it is determined that the point is an abnormal value caused by external interference such as electromagnetic pulses, rather than a real signal change caused by changes in the medium state.

[0036] A separate sliding statistical window is maintained for each signal channel (such as capacitance, frequency, impedance, etc.), and its mean and standard deviation are updated in real time. A 3σ interval test is performed on the current sampled value. Points determined to be abnormal are removed and replaced with the last valid sampled value or linear interpolation result before that point, so that the signal waveform remains smooth and continuous.

[0037] Next, Z-score standardization is used to eliminate the influence caused by inconsistencies in the dimensions and numerical ranges of various features, making all features numerically comparable and ensuring fair learning of subsequent models. Z-score standardization refers to performing the following linear transformation on all sampled values ​​of a feature: New value = (Original value - Mean of all historical samples of this feature) / Standard deviation of all historical samples of this feature.

[0038] After this transformation, the numerical distribution of any feature will be adjusted to a standard normal distribution with a mean of 0 and a standard deviation of 1. The original dimensions are eliminated, and all features become dimensionless pure numerical values ​​with the same central location and degree of dispersion.

[0039] like Figure 3 As shown, Figure 3 (a) in the figure is a schematic diagram of the original dimensions of impedance and phase shift, showing the original data of impedance and phase shift before normalization; Figure 3 (b) in the figure is a schematic diagram of the standardized dimensions of impedance and phase shift. After Z-score standardization of impedance (Z) and phase shift (θ) respectively, the dimensions of the two features are eliminated and adjusted to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0040] This invention performs Z-score standardization on the 6 original features after outlier removal to obtain the corresponding standardized features (denoted as the standardized equivalent capacitance). Standardized resonant frequency Standardized impedance Standardized phase shift Standardized downhole ambient temperature and standardized downhole environmental humidity ), which serve as the basic features for subsequent processing.

[0041] Further, the sliding window length is obtained, and the standard feature is segmented according to the sliding window length to obtain multiple feature segments; the core physical quantities of the multiple feature segments are statistically calculated to obtain multiple extended statistical features; and the extended features are obtained based on the multiple extended statistical features.

[0042] This invention introduces a sliding time window mechanism to endow the recognition model with the ability to "perceive dynamic changes," enabling it to make comprehensive judgments not only based on the current value but also on the changing trends and fluctuation characteristics within a certain time range. The sliding time window can be figuratively understood as a fixed-length time interval that continuously moves forward with the current time, always covering the signal sampling points of the most recent consecutive moments. By performing statistical analysis on a segment of signal within the window, statistical features reflecting the signal's changing patterns during that time period can be extracted.

[0043] In this invention, three core physical quantities that best reflect the dielectric properties of the measured medium are selected: equivalent capacitance C, impedance Z, and phase offset θ.

[0044] Within the sliding time window, the following three types of statistical characteristics are calculated for the standardized sequences of these three physical quantities: 1. Mean: The arithmetic mean of all sampled values ​​within a window, reflecting the average signal strength during that period, and used to characterize the overall state level of the medium; 2. Variance: The average of the squares of the deviations of each sampled value from its mean within a window. It reflects the degree of fluctuation of the signal during that period and is used to distinguish between "static" media (such as static water accumulation or static frost) and "dynamic" media (such as flowing water or drifting coal dust). 3. Mean of First-order Difference: The first-order difference refers to the difference between the values ​​of two adjacent sampling points. Its mean represents the average rate of change and direction (rising or falling) of the signal. It is used to capture the evolution trend of the medium state, such as rising water level, thickening of frost, and continuous increase in coal dust concentration.

[0045] like Figure 4 As shown, Figure 4 (a) in the diagram is a schematic diagram of sliding window statistical feature extraction of equivalent capacitance data, using standardized equivalent capacitance data. For example, first set a window value W, such as W=20, and then use a sliding window to capture a local data segment; Figure 4 (b) in the figure is a schematic diagram of the distribution of equivalent capacitance data within the sliding window. The mean value (Mean) is extracted based on the distribution of data within the sliding window, which is approximately -0.76. Figure 4(c) in the figure is the first-order difference sequence plot of the equivalent capacitance data within the window. The first-order difference mean Diff is extracted from the first-order difference sequence plot, which is approximately -0.001. Finally, the statistical features are output, and the mean Mean_C of the equivalent capacitance is -0.7577, the variance Var_C is 0.0001, and the first-order difference mean Diff_C is -0.0012.

[0046] right , and Three statistical measures were calculated for the three standardized physical quantities, resulting in a total of 9 extended statistical features, denoted as the capacitance mean (Mean_C), capacitance variance (Var_C), capacitance first-order difference mean (Diff_C), impedance mean (Mean_Z), impedance variance (Var_Z), impedance first-order difference mean (Diff_Z), phase shift mean (Mean_θ), phase shift variance (Var_θ), and phase shift first-order difference mean (Diff_θ).

[0047] Furthermore, the standard features and the extended features are concatenated to obtain the input feature vector.

[0048] like Figure 5 As shown, the 6-dimensional standardized features and the 9-dimensional extended statistical features obtained through the sliding window are concatenated together to obtain a 15-dimensional high-dimensional input feature vector.

[0049] The high-dimensional input feature vector contains both the current static properties and recent dynamic evolution information of the measured medium, providing a sufficient data foundation for accurately identifying five types of media: air, clean water, coal slurry, coal dust, and ice / frost.

[0050] Step S30: Determine the first basic model, the second basic model, and the third basic model. Input the input feature vector into the first basic model, the second basic model, and the third basic model respectively for inference to obtain the first probability vector, the second probability vector, and the third probability vector.

[0051] Based on the constructed 15-dimensional high-dimensional input features, this invention employs three independent data processing pipelines to run three machine learning base models with completely different mathematical principles in parallel.

[0052] Specifically, a first base model is determined, and a first probability vector is obtained by performing multi-level decision tree prediction processing based on the input feature vector using the first base model.

[0053] like Figure 2As shown, the input features are processed by XGBoost gradient boosting and static nonlinear mapping, outputting the first probability vector P1; the dynamic temporal features of the input features are extracted by 1D-CNN convolutional neural network, outputting the second probability vector P2; and the input features are robustly generalized to small samples by kernel SVM support vector machine, outputting the third probability vector P3.

[0054] First, we build a dataset to train three machine learning base models.

[0055] Specifically, a labeled dataset covering five categories—"air," "clean water," "coal slurry," "coal dust," and "frost"—was created by combining precise laboratory calibration (pure media, mixtures with known concentration ratios) with on-site data collection in the mine (labeled team leader inspection records). Each sample contains 6-dimensional raw sensor data at the same time. Three base models were then trained using the constructed dataset.

[0056] The first base model is the XGBoost gradient boosting tree, which is specifically designed for high-precision mapping of stable, static operating conditions, such as stable water accumulation zones exposed at the mining face or slow-forming static frost on the outer walls of equipment. In these scenarios, the state of the medium hardly changes over time, eliminating the need for time-series analysis capabilities. The classification task is essentially a complex multidimensional nonlinear regression problem, and XGBoost happens to have a very strong modeling capability for this type of structured data.

[0057] like Figure 6 As shown, XGBoost receives a 15-dimensional feature vector X as input, without involving temporal information. Its core employs a gradient boosting framework, progressively constructing multiple decision trees through a sequential iterative process. First, an initial tree is used to make preliminary predictions on the training data, and then the prediction residuals are calculated. Subsequently, each newly added tree is specifically designed to fit the overall residuals of the current model. Finally, the prediction scores of all K decision trees are summed and transformed into probability values ​​for five media types using a Softmax function.

[0058] This additive integration approach enables the model to automatically uncover deep, non-linear interactions between features, such as how temperature, humidity, and capacitance values ​​couple to affect the dielectric constant.

[0059] The assumed model includes The first decision tree, the tree to the first Input features of each sample The predicted score is The final predicted score The sum of the outputs for all trees: .

[0060] The score vector is transformed into a prediction probability vector P1 for five media types using the Softmax function. The five media types are air, clean water, coal slurry, coal dust, and ice / frost.

[0061] In this invention, the number of trees is preferably 100 to 300, the learning rate is preferably 0.01 to 0.1, the maximum depth of a single tree is preferably 3 to 6, and the regularization coefficient is preferably 0.1 to 1.0.

[0062] Further, a second base model is determined, and a second probability vector is obtained by performing one-dimensional convolutional prediction processing on the input feature vector using the second base model.

[0063] The second base model of this invention is a 1D-CNN one-dimensional convolutional neural network, specifically designed to capture the temporal features of dynamically changing operating conditions. Many critical operating conditions in underground mines evolve gradually, rather than undergoing instantaneous changes: water flow generates high-frequency oscillations, dust dispersion corresponds to slow signal drift, and frost freezing is accompanied by phased step changes. These dynamic patterns are crucial for determining the type of medium, but they cannot be perceived from numerical values ​​at a single moment; they must be identified by observing the "shape" and "trend" of the signal over time. The initial design intent of 1D-CNN is precisely to automatically extract such local patterns from time series data.

[0064] The 1D-CNN takes as input a continuous signal segment X (size T×15) within a sliding time window, representing the complete sequence of the current time and a preceding historical interval. It outputs a predicted probability vector P2 for five media types at the end of the window. The model internally uses multiple one-dimensional convolutional kernels that slide along the time axis. Each kernel acts as a local pattern detector; some are sensitive to fast rising edges, some to slow falling trends, and others to periodic fluctuations. By stacking multiple layers of convolutions and pooling, lower layers detect simple rising and falling changes, while higher layers combine to represent more complex dynamic processes. Finally, after a fully connected layer and a softmax transform, the model outputs the media type probability at the end of the current window.

[0065] like Figure 7 As shown, Figure 7 (a) in the figure is a schematic diagram of the original signal time sequence segment, where the horizontal axis represents the sampling point (corresponding to the time sequence) and the vertical axis represents the capacitance value. The segmentation marks intuitively present the dynamic evolution process of the medium state from coal dust, the transition section (rapid rise) to the stable section of clear water, providing the original signal background for the subsequent time sequence feature extraction of the 1D-CNN model. Figure 7 (b) in the diagram is a schematic diagram of rising edge detection of the convolution kernel, where the horizontal axis is the convolution kernel weight index and the vertical axis is the weight value. Through the weight distribution design of "negative weight + positive weight", the local pattern detection of the rapid rising trend of the signal is realized. Figure 7(c) in the figure is a schematic diagram of the convolution output features, where the horizontal axis represents the sampling points and the vertical axis represents the convolution response values. The green positive response area in the figure corresponds to the rising edge stage (transition segment) of the signal, and the red negative response area corresponds to the stable or falling stage, which intuitively shows how the convolution kernel accurately identifies the dynamic change features in the signal.

[0066] In this invention, the preferred time window length is 20-50, the preferred kernel length is 3-7, the preferred number of kernels in each layer is 16-64, and the preferred pooling window size is 2-3.

[0067] Furthermore, a third basic model is determined, and a third probability vector is obtained by performing kernel mapping classification processing on the input feature vector using the third basic model.

[0068] The third base model in this invention is a Gaussian kernel SVM support vector machine, which is specifically designed for robust identification of rare, small-sample extreme working conditions in mines. During long-term operation, hazardous samples such as thick frost and high-concentration coal slurry water occur very infrequently, accounting for a negligible proportion in the training set. Models that aim to minimize the overall error rate (such as XGBoost and 1D-CNN) naturally tend to "ignore" these niche samples, misclassifying them as large classes and causing serious underreporting. Kernel SVM, due to its unique decision-making mechanism (the classification boundary is determined only by a few support vectors, and the majority of samples far from the boundary do not affect the boundary position), is naturally resistant to small samples and class imbalance, and can establish reliable classification boundaries with very few labeled samples.

[0069] The input to the kernel SVM is consistent with XGBoost, consisting of a 15-dimensional feature vector X at a single time step. The SVM itself is a binary classifier; to handle the 5-class classification task of this invention, a "one-to-one" strategy is adopted, constructing a total of 10 binary SVMs, with one classifier trained between every two classes. During the inference phase, all binary classifiers vote individually, and finally, a unified 5-dimensional probability vector P3 is output through plat scaling. When dealing with nonlinearly separable problems, the SVM implicitly maps the original low-dimensional features to a high-dimensional space using a Gaussian radial basis kernel function, constructing a linear dividing surface in the high-dimensional space, which is equivalent to generating a complex nonlinear decision boundary in the original space.

[0070] like Figure 8 As shown in the diagram, the classification principle of the Gaussian kernel support vector machine (SVM) model intuitively demonstrates its robustness to small sample conditions. Figure 8In the example shown, the number of majority class (n) in the medium sample is 174, the number of rare class (26) is 26, the number of support vectors is 39, and the decision boundary (f) is 0. This demonstrates the nonlinear decision boundary constructed by the Gaussian kernel function and the core characteristic of "the boundary position is determined only by a few support vectors and is not sensitive to the majority of samples". This reflects the model's recognition advantage in small sample and class imbalance conditions.

[0071] Step S40: Determine the meta-learner, and fuse the first probability vector, the second probability vector, and the third probability vector through the meta-learner to obtain the medium classification result in the coal mine.

[0072] Specifically, the first probability vector, the second probability vector, and the third probability vector are concatenated to obtain the meta-feature vector.

[0073] like Figure 2 As shown, P1, P2, and P3 are stacked and fused using a logistic regression meta-learner. These three sets of probability vectors are concatenated into 15-dimensional meta-features, which automatically learn the credibility weights of each base model under different working conditions. Finally, an optimal medium classification result after fusion is output, which is determined as air, clean water, coal slurry, coal dust, or frost, and triggers the corresponding underground equipment linkage control actions to perform drainage, dust removal, or defrosting.

[0074] By concatenating the three 5-dimensional probability vectors P1, P2, and P3 sequentially, a 15-dimensional vector is formed, which this invention calls the meta-feature vector. This vector completely preserves all the prediction information of the three base models for the current sample and is the sole basis for the meta-learner to make fusion decisions.

[0075] Further, a validation set is obtained, an initial meta-learner and a loss function are determined, and the initial meta-learner is iteratively trained based on the validation set, the first probability vector, the second probability vector, the third probability vector, and the loss function. When the loss function converges, the target linear transformation parameters are obtained. The initial meta-learner is then parameterized based on the target linear transformation parameters to obtain the meta-learner.

[0076] This invention selects logistic regression as the meta-learner, mainly based on the following considerations: 1. Simple and highly interpretable. Logistic regression is a classic linear classification model whose internal parameters are the weight coefficients corresponding to each input dimension. These weight coefficients can be directly interpreted as "how credible the meta-learner is of a base model's prediction for a certain category." Positive weights with large absolute values ​​indicate that the base model's judgment of that category is highly trusted, while weights close to zero indicate that its opinion is basically ignored. This interpretability is crucial for troubleshooting and model debugging in industrial applications.

[0077] 2. Suitable for fusion scenarios with probabilistic inputs. Logistic regression receives continuous numerical inputs and outputs normalized class probabilities through the Softmax function. It is naturally suitable for processing probability vectors output by three base models. The calculation process is simple, the values ​​are stable, and the output will not be distorted by small fluctuations in the input probabilities.

[0078] 3. High training efficiency and less prone to overfitting. The training data volume of meta-learners is relatively limited (limited by the total number of labeled samples). Logistic regression has few parameters, requiring only 15×5=75 weight parameters to be learned. Compared with complex models such as neural networks, it has a lower requirement for training data volume and better generalization ability.

[0079] First, supervised learning is used to train the meta-learner on an independent validation set, resulting in its weight matrix. W And bias b.

[0080] The training process is as follows: Three base models are trained using the training set; then, each sample from the validation set is fed into the three trained base models to obtain three sets of probability vectors, which are then concatenated into a meta-feature vector according to a formula; finally, using the true class labels of the validation set as the supervision signal and the cross-entropy loss function as the optimization objective, a logistic regression meta-learner is trained to solve for the optimal... W and b.

[0081] The cross-entropy loss function is: ; in, Represents cross-entropy loss, This represents the total number of samples in the validation set. This represents the sample index of the validation set. Indicates the category index of the sample. Indicates the first The one-hot encoding of the true class of the nth sample, where the true class of the sample is the nth sample. The value is 1 if the condition is met, otherwise it is 0. The meta-learner will represent the samples Predicted as the first The probability of a class.

[0082] The cross-entropy loss function is minimized using gradient descent, guiding the meta-learner to automatically learn the optimal weight allocation for each base model under different operating conditions. For example, when the samples are in a stable static condition, the output probability distribution of XGBoost is usually quite sharp (the probability of one class is close to 1, and the rest are close to 0), while the output of 1D-CNN is relatively smooth. In this case, logistic regression will naturally learn that XGBoost should be given a larger positive coefficient for the corresponding dimension. When the signal exhibits typical dynamic fluctuations, the output of 1D-CNN is more reliable, and logistic regression will increase the weight of the corresponding dimension of 1D-CNN. For rare operating conditions, the probability vector output by the kernel SVM becomes a key clue, and the meta-learner will also adjust the weights accordingly to rely on the judgment of the SVM.

[0083] The parameters of the initial meta-learner are set according to the optimal weight matrix and bias vector to obtain the meta-learner used for fusion.

[0084] Furthermore, the meta-feature vector is linearly transformed by the meta-learner to obtain a score vector; the score vector is then probability-normalized to obtain a target probability vector; and a category decision is made based on the target probability vector to obtain a media classification result.

[0085] The meta-learner processes the meta-feature vector in two steps: linear transformation and probability normalization.

[0086] Specifically, the linear transformation multiplies a 15-dimensional vector by a weight matrix. W Adding the bias vector b, we obtain a 5-dimensional score vector. S Each dimension corresponds to a type of medium.

[0087] The score vector is obtained through the Softmax function. S Transform it into the final probability vector, so that the sum of the probabilities of each category is 1.

[0088] Finally, a category decision is made, and the category corresponding to the dimension with the largest value in the fused probability vector is taken as the recognition result. ypred , ypred The values ​​1 to 5 correspond to air, clean water, coal slurry, coal dust, and frost, respectively.

[0089] The final output of this invention is a classification label. ypred The classification label is used to clearly indicate which of the five categories the current tested medium belongs to, and the confidence score is the probability value corresponding to that category in the output probability of the meta-learner, reflecting the reliability of the classification result.

[0090] The two pieces of information output by this invention will serve as the decision-making basis for subsequent equipment linkage control. When the information is determined to be coal slurry and the confidence level exceeds a preset threshold, the underground drainage pump will be triggered to start; when the information is determined to be coal dust, the spray dust suppression system will be activated; and when the information is determined to be frost, the equipment preheating and defrosting program will be initiated. Simultaneously, the classification results and confidence levels will be uploaded to the coal mine safety monitoring platform in real time for monitoring and traceability by the dispatch center.

[0091] The key innovations of this invention include: 1. For the first time, it proposes and integrates three completely heterogeneous models in terms of mechanism: XGBoost (static fitting), 1D-CNN (dynamic features), and kernel SVM (small sample robustness). Through stacking fusion, it achieves complementary and enhanced perception capabilities, surpassing the performance boundaries of any single model or traditional homogeneous ensemble (such as random forest).

[0092] 2. It abandons the traditional single-point, static threshold discrimination and introduces statistical features based on time windows and 1D-CNN model. It incorporates the "dynamic temporal change pattern" itself as a core feature into the recognition system, which greatly improves the ability to identify dynamic processes and state transitions and solves the problems of false alarms and lag.

[0093] 3. A stacking two-layer learning mechanism is adopted to train a "meta-learner" that can adaptively adjust the decision weights of the three base models. This end-to-end learning weight allocation strategy is far more refined and better suited to complex and ever-changing working conditions than manually set fixed voting rules.

[0094] 4. Based on accurate classification, a scheme with an additional lightweight regression model is proposed. This model can directly regress quantitative indicators such as concentration and thickness according to the identified media type, upgrading the primary perception of "what it is" to the secondary perception of "what it is and how much it is", thus providing a data foundation for precise control.

[0095] This invention uses three heterogeneous models to infer sensor data, and then fuses the inference results of the three models through a meta-learner to obtain media classification results, thereby improving the accuracy of media identification.

[0096] Possible design modifications to this invention include: 1. Model Replacement and Upgrade: All three base models can be replaced with similar but more advanced versions. For example, 1D-CNN can be replaced with a Long Short-Term Memory network or a Transformer temporal model, which are more powerful for handling temporal sequences; XGBoost can be replaced with LightGBM or CatBoost; SVM can be replaced with a meta-learning algorithm suitable for small sample sizes. Logistic Regression can be replaced with a simple multilayer perceptron, gradient boosting tree, or weighted averaging strategy to seek different balances between computational power and accuracy.

[0097] 2. Task Expansion and Refinement: Based on specific mine needs, media categories can be added or removed, such as adding "emulsion" or "lubricating oil." Only the corresponding samples need to be supplemented, and the model fine-tuned or retrained. The classification task can be upgraded to instance segmentation or semantic segmentation. For example, on an electrode array, instead of outputting an overall category, a "distribution map" can be output, showing whether each location is coal or rock, achieving more refined interface rendering.

[0098] 3. Input Modality Fusion: In addition to capacitance and temperature / humidity sensors, other intrinsically safe sensors, such as micro-vibration sensors, ultrasonic sensors, and low-light cameras, can be seamlessly integrated. At this point, a 1D-CNN or a newly added 2D-CNN can process vibration signals and image signals separately, and the generated deep features are then fed together into a meta-learner for fusion, further improving reliability under extreme conditions.

[0099] 4. Simplified Training and Deployment Strategies: To deploy on low-cost microcontrollers, a pre-trained heterogeneous fusion "large model" can be used as the "teacher model." Through knowledge distillation, a lightweight "student model" with a simple structure is trained, achieving significant model compression and acceleration with minimal loss of accuracy. To protect data privacy and utilize heterogeneous data distributed across various mines, a federated learning framework can be adopted. Edge nodes in each mine only update the model locally, uploading only encrypted gradient information to the central server for aggregation, achieving collaborative evolution where "the data remains still while the model moves."

[0100] Furthermore, such as Figure 9 As shown, based on the above-mentioned method for identifying underground coal mine media based on heterogeneous model fusion, this invention also provides a system for identifying underground coal mine media based on heterogeneous model fusion, wherein the system for identifying underground coal mine media based on heterogeneous model fusion includes: The signal modulation module 51 is used to acquire the medium data of the medium to be detected in the coal mine, and to perform signal modulation processing on the medium data to obtain a digital signal; The feature generation module 52 is used to standardize the digital signal to obtain standard features, expand the standard features to obtain expanded features, and concatenate the standard features and the expanded features to obtain an input feature vector. The heterogeneous model inference module 53 is used to determine the first base model, the second base model and the third base model, and to input the input feature vector into the first base model, the second base model and the third base model respectively for inference to obtain the first probability vector, the second probability vector and the third probability vector. The medium classification module 54 is used to determine the meta-learner, and to fuse the first probability vector, the second probability vector and the third probability vector through the meta-learner to obtain the medium classification result of the coal mine.

[0101] Furthermore, such as Figure 10 As shown, based on the above-mentioned coal mine underground media identification method and system based on heterogeneous model fusion, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 10 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0102] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal. Further, the memory 20 may include both internal and external storage units of the terminal. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a coal mine underground media identification program 40 based on heterogeneous model fusion, which can be executed by the processor 10 to implement the coal mine underground media identification method based on heterogeneous model fusion in this application.

[0103] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the coal mine underground media identification method based on heterogeneous model fusion.

[0104] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface.

[0105] In one embodiment, when the processor 10 executes the coal mine underground media identification program 40 based on heterogeneous model fusion in the memory 20, the following steps are performed: Acquire the medium data of the medium to be tested in the coal mine, and perform signal modulation processing on the medium data to obtain a digital signal; The digital signal is standardized to obtain standard features, the standard features are expanded to obtain extended features, and the standard features and the extended features are concatenated to obtain the input feature vector. Determine the first base model, the second base model, and the third base model, and input the input feature vector into the first base model, the second base model, and the third base model respectively for inference to obtain the first probability vector, the second probability vector, and the third probability vector. A meta-learner is determined, and the first probability vector, the second probability vector, and the third probability vector are fused using the meta-learner to obtain the medium classification result in the coal mine.

[0106] The dielectric data includes dielectric property data and environmental data; The process of acquiring media data of the medium to be detected in the coal mine and performing signal modulation processing on the media data to obtain a digital signal specifically includes: The equivalent capacitance, resonant frequency, impedance, and phase shift of the medium under test are collected by a capacitance sensor, and the equivalent capacitance, resonant frequency, impedance, and phase shift are used as the dielectric characteristic data. Ambient temperature and humidity are collected by environmental sensors, and the ambient temperature and humidity are used as the environmental data. The dielectric property data and the environmental data are filtered, amplified, and converted to obtain a digital signal.

[0107] Specifically, the process of standardizing the digital signal to obtain standard features, expanding the standard features to obtain expanded features, and concatenating the standard features and the expanded features to obtain the input feature vector includes: Calculate the arithmetic mean and standard deviation of the digital signal, and perform range calculation processing based on the arithmetic mean and standard deviation to obtain the normal signal range; The digital signal is subjected to outlier replacement processing based on the normal signal range to obtain the original signal; The original signal is subjected to mean normalization to obtain standard features; Obtain the length of the sliding window, and expand the standard feature according to the length of the sliding window to obtain the expanded feature; The standard features and the extended features are concatenated to obtain the input feature vector.

[0108] Specifically, obtaining the sliding window length and expanding the standard features based on the sliding window length to obtain expanded features includes: Obtain the length of the sliding window, and perform segmentation processing on the standard feature according to the length of the sliding window to obtain multiple feature segments; Statistical feature calculations are performed on the core physical quantities of multiple feature segments to obtain multiple extended statistical features; The extended features are obtained based on multiple extended statistical features.

[0109] Specifically, determining the first base model, the second base model, and the third base model involves inputting the input feature vector into the first base model, the second base model, and the third base model respectively for inference to obtain the first probability vector, the second probability vector, and the third probability vector, which includes: A first base model is determined, and a first probability vector is obtained by performing multi-level decision tree prediction processing based on the input feature vector using the first base model. A second base model is determined, and a second probability vector is obtained by performing one-dimensional convolutional prediction processing based on the input feature vector using the second base model. A third base model is determined, and a third probability vector is obtained by performing kernel mapping classification processing on the input feature vector using the third base model.

[0110] The determined meta-learner fuses the first probability vector, the second probability vector, and the third probability vector to obtain the medium classification result in the coal mine, specifically including: The first probability vector, the second probability vector, and the third probability vector are concatenated to obtain the meta-feature vector. A meta-learner is determined, and the meta-feature vector is linearly transformed by the meta-learner to obtain a score vector; The score vector is then subjected to probability normalization to obtain the target probability vector; Based on the target probability vector, a category decision is made to obtain the medium classification result.

[0111] The determination of the meta-learner involves performing a linear transformation on the meta-feature vector to obtain a score vector, and the process further includes: Obtain a validation set, determine the initial meta-learner and loss function, and iteratively train the initial meta-learner based on the validation set, the first probability vector, the second probability vector, the third probability vector, and the loss function. When the loss function converges, the target linear transformation parameters are obtained. The initial meta-learner is parameterized according to the target linear transformation parameters to obtain the meta-learner.

[0112] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a coal mine underground media identification program based on heterogeneous model fusion, and the coal mine underground media identification program based on heterogeneous model fusion implements the steps of the coal mine underground media identification method based on heterogeneous model fusion as described above when executed by a processor.

[0113] In summary, this invention provides a method, system, and terminal for identifying underground coal mine media based on heterogeneous model fusion. The method includes: acquiring media data of the medium to be detected in the underground coal mine; performing signal modulation processing on the media data to obtain a digital signal; standardizing the digital signal to obtain standard features; expanding the standard features to obtain expanded features; and concatenating the standard features and the expanded features to obtain an input feature vector; determining a first basic model, a second basic model, and a third basic model; inputting the input feature vector into the first basic model, the second basic model, and the third basic model respectively for inference to obtain a first probability vector, a second probability vector, and a third probability vector; determining a meta-learner; and fusing the first probability vector, the second probability vector, and the third probability vector through the meta-learner to obtain a media classification result for the underground coal mine. This invention improves the accuracy of media identification by using three heterogeneous models to infer sensor data and fusing the inference results of the three models through a meta-learner.

[0114] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one" does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0115] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0116] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for identifying underground media in coal mines based on heterogeneous model fusion, characterized in that, The coal mine underground media identification method based on heterogeneous model fusion includes: Acquire the medium data of the medium to be tested in the coal mine, and perform signal modulation processing on the medium data to obtain a digital signal; The digital signal is standardized to obtain standard features, the standard features are expanded to obtain extended features, and the standard features and the extended features are concatenated to obtain the input feature vector. Determine the first base model, the second base model, and the third base model, and input the input feature vector into the first base model, the second base model, and the third base model respectively for inference to obtain the first probability vector, the second probability vector, and the third probability vector. A meta-learner is determined, and the first probability vector, the second probability vector, and the third probability vector are fused through the meta-learner to obtain the medium classification result in the coal mine. The process of standardizing the digital signal to obtain standard features, expanding the standard features to obtain expanded features, and concatenating the standard features and the expanded features to obtain an input feature vector specifically includes: Calculate the arithmetic mean and standard deviation of the digital signal, and perform range calculation processing based on the arithmetic mean and standard deviation to obtain the normal signal range; The digital signal is subjected to outlier replacement processing based on the normal signal range to obtain the original signal; The original signal is subjected to mean normalization to obtain standard features; Obtain the length of the sliding window, and expand the standard feature according to the length of the sliding window to obtain the expanded feature; The standard features and the extended features are concatenated to obtain the input feature vector; The process of obtaining the sliding window length and expanding the standard features based on the sliding window length to obtain expanded features specifically includes: Obtain the length of the sliding window, and perform segmentation processing on the standard feature according to the length of the sliding window to obtain multiple feature segments; Statistical feature calculations are performed on the core physical quantities of multiple feature segments to obtain multiple extended statistical features; The extended features are obtained based on multiple extended statistical features; The process of determining the first, second, and third basic models, and inputting the input feature vectors into the first, second, and third basic models respectively for inference to obtain the first, second, and third probability vectors, specifically includes: A first base model is determined, and a first probability vector is obtained by performing multi-level decision tree prediction processing based on the input feature vector using the first base model. A second base model is determined, and a second probability vector is obtained by performing one-dimensional convolutional prediction processing based on the input feature vector using the second base model. A third base model is determined, and a third probability vector is obtained by performing kernel mapping classification processing on the input feature vector using the third base model.

2. The method for identifying underground media in coal mines based on heterogeneous model fusion according to claim 1, characterized in that, The medium data includes dielectric property data and environmental data; The process of acquiring media data of the medium to be detected in the coal mine and performing signal modulation processing on the media data to obtain a digital signal specifically includes: The equivalent capacitance, resonant frequency, impedance, and phase shift of the medium under test are collected by a capacitance sensor, and the equivalent capacitance, resonant frequency, impedance, and phase shift are used as the dielectric characteristic data. Ambient temperature and humidity are collected by environmental sensors, and the ambient temperature and humidity are used as the environmental data. The dielectric property data and the environmental data are filtered, amplified, and converted to obtain a digital signal.

3. The method for identifying underground media in coal mines based on heterogeneous model fusion according to claim 1, characterized in that, The determined meta-learner fuses the first probability vector, the second probability vector, and the third probability vector to obtain the medium classification result in the coal mine, specifically including: The first probability vector, the second probability vector, and the third probability vector are concatenated to obtain the meta-feature vector. A meta-learner is determined, and the meta-feature vector is linearly transformed by the meta-learner to obtain a score vector; The score vector is then subjected to probability normalization to obtain the target probability vector; The target probability vector is used to perform category decision processing to obtain the medium classification result.

4. The method for identifying underground media in coal mines based on heterogeneous model fusion according to claim 3, characterized in that, The determined meta-learner performs a linear transformation on the meta-feature vector to obtain a score vector, and the process further includes: Obtain a validation set, determine an initial meta-learner and a loss function, iteratively train the initial meta-learner based on the validation set, the first probability vector, the second probability vector, the third probability vector, and the loss function, and obtain the target linear transformation parameters when the loss function converges. The initial meta-learner is parameterized according to the target linear transformation parameters to obtain the meta-learner.

5. A coal mine underground media identification system based on heterogeneous model fusion, wherein the coal mine underground media identification system based on heterogeneous model fusion is used to implement the coal mine underground media identification method based on heterogeneous model fusion as described in any one of claims 1-4, characterized in that, The underground coal mine media identification system based on heterogeneous model fusion includes: The signal modulation module is used to acquire the medium data of the medium to be detected in the coal mine, and to perform signal modulation processing on the medium data to obtain a digital signal; The feature generation module is used to standardize the digital signal to obtain standard features, expand the standard features to obtain expanded features, and concatenate the standard features and the expanded features to obtain an input feature vector. The heterogeneous model inference module is used to determine the first base model, the second base model and the third base model, and to input the input feature vector into the first base model, the second base model and the third base model respectively for inference to obtain the first probability vector, the second probability vector and the third probability vector. The medium classification module is used to determine the meta-learner, and to fuse the first probability vector, the second probability vector, and the third probability vector through the meta-learner to obtain the medium classification result in the coal mine.

6. A terminal, characterized in that, The terminal includes a memory, a processor, and a coal mine underground media identification program based on heterogeneous model fusion stored in the memory and executable on the processor. When the processor executes the coal mine underground media identification program based on heterogeneous model fusion, it implements the steps of the coal mine underground media identification method based on heterogeneous model fusion as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a coal mine underground media identification program based on heterogeneous model fusion. When the coal mine underground media identification program based on heterogeneous model fusion is executed by a processor, it implements the steps of the coal mine underground media identification method based on heterogeneous model fusion as described in any one of claims 1-4.

Images