Multi-index self-adaptive fusion grading early warning method for coal rock gas dynamic disasters

Abstract

The invention provides a multi-index self-adaptive fusion grading early warning method for coal rock gas dynamic disasters. The method comprises the following steps: (1) multi-source data collaborative preprocessing: collecting multi-source data, identifying and marking abnormal values by using a semi-supervised collaborative training model based on xLSTM, denoising acoustic emission and electromagnetic radiation signals through a PSO-wavelet threshold-OPTICS model, and interpolating missing values in real time through a BayOTIDE model; (2) multi-disaster feature fusion and risk prediction: inputting the preprocessed data into an MTAFM model, processing different data by a core expert module and an auxiliary expert module respectively, and dynamically fusing features through a gating unit to generate a risk probability; and (3) dynamic grading early warning and composite disaster judgment: determining an early warning threshold value based on Bayesian dynamic optimization, grading by combining with a quantile method, and judging coal rock and gas dynamic disasters as composite disasters during early warning. The method is high in adaptive data preprocessing capability, accurate in multi-index fusion early warning and reliable in dynamic grading early warning, and the dependence on professionals is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of coal mine safety monitoring and early warning, and particularly to a multi-index adaptive fusion hierarchical early warning method for coal-rock gas dynamic disasters. Background Technique

[0002] As one of the important fossil fuels globally, the safety of coal mining is directly related to the stability of energy supply and the lives of miners. With the gradual depletion of shallow coal resources, deep mining has become an inevitable trend in the development of the coal industry. However, the geological conditions in deep mining environments are more complex, and the risks of coal-rock dynamic disasters (such as rock bursts, roof collapses, etc.) and gas dynamic disasters (such as coal and gas outbursts, gas gushes, etc.) have increased significantly, and may even trigger coal-rock gas compound disasters, posing a serious threat to the safe mining of coal.

[0003] To effectively address these disaster risks, it is particularly important to accurately and efficiently process and analyze various monitoring data in coal mine excavation and working faces. Traditional disaster early warning methods mostly rely on statistical discrimination methods that study the risks of coal-rock gas dynamic disasters based on time-frequency characteristics, such as trend methods, threshold methods, etc. These methods calculate the statistical laws of each monitoring index independently. Although they can identify single disaster risks to a certain extent and achieve early warning, their limitations become more obvious in the complex deep excavation and working face environments.

[0004] On the one hand, due to the mutual coupling of coal-rock gas dynamic disasters, single-index early warning is difficult to accurately quantify the evolution and catastrophic characteristics of disasters, has low generalization ability, and is prone to false negatives or false positives. On the other hand, existing multi-index fusion early warning methods are mostly based on independent disaster early warning models, which are difficult to comprehensively consider the correlations between coal-rock gas dynamic disasters, and are insufficient in adaptively quantifying the complex relationships between different indicators and different disaster risks, and are difficult to dynamically optimize parameters to adapt to changes in the coal mine environment, thus limiting the early warning performance.

[0005] In addition, the quality of monitoring data directly affects the accuracy of disaster early warning. However, in practical applications, monitoring data often has problems such as outliers, missing values, and noise, which seriously affect the data quality and thus reduce the accuracy of disaster risk assessment. When dealing with these problems, traditional methods mostly rely on the experience of professionals to adjust parameters, with limited generalization ability. Although deep learning methods have improved the processing accuracy and generalization ability to a certain extent, due to the complex distribution of abnormal data and high annotation costs, supervised learning is restricted, while unsupervised learning is difficult to adapt to complex working conditions, affecting the accuracy of abnormal data processing. Summary of the Invention

[0006] The purpose of the present invention is to provide a multi-index adaptive fusion hierarchical early warning method for coal-rock gas dynamic disasters, including the following steps:

[0007] (1) Multi-source data collaborative preprocessing, specifically including:

[0008] Collect gas concentration, support resistance, electromagnetic radiation, acoustic emission, bolt and cable stress, borehole stress, microseismic energy, wind speed and temperature data of coal mine excavation faces to form an initial data set;

[0009] Identify outliers through a semi-supervised collaborative training model based on xLSTM and label them as missing values;

[0010] Denoise acoustic emission and electromagnetic radiation signals through a PSO-wavelet threshold-OPTICS adaptive denoising model;

[0011] Interpolate the labeled missing values using the BayOTIDE missing value online real-time interpolation model;

[0012] (2) Disaster feature fusion and risk prediction, specifically:

[0013] Input the preprocessed data into the multi-index fusion early warning model MTAFM. The MTAFM model processes gas concentration and support resistance data through the core expert module, processes acoustic emission, electromagnetic radiation, bolt and cable stress, microseismic, wind speed and temperature data through the auxiliary expert module, and uses a gated unit to dynamically fuse the output features of the two types of modules to generate the risk probability of coal and rock gas dynamic disasters;

[0014] (3) Dynamic hierarchical early warning and compound disaster determination, including:

[0015] Determine the lowest early warning threshold of disaster risk based on Bayesian dynamic optimization;

[0016] Combine the quantile method to divide the risk levels: no early warning below the threshold, yellow early warning above the threshold and below the 85th percentile, orange early warning for the 85-90th percentile, and red early warning above the 90th percentile; if coal and rock and gas disasters are early warned at the same time, it is determined as a compound disaster risk.

[0017] Furthermore, the outlier identification model in step (1) is based on the collaborative training framework of xLSTM, specifically including:

[0018] Construct sLSTM and mLSTM as dual-view learning modules, where sLSTM optimizes the information flow processing through an exponential gated unit and a scalar update mechanism, and mLSTM enhances the memory and parallel processing capabilities through matrix operations;

[0019] Perform sLSTM and mLSTM predictions on unlabeled data respectively, and cross-select samples with confidence higher than the preset threshold and add them to the training set of the other model;

[0020] Optimize the model parameters through iterative training and error backpropagation until the prediction error converges or the maximum number of iterations is reached.

[0021] Furthermore, the denoising process of the PSO-wavelet threshold-OPTICS denoising model includes the following steps:

[0022] Wavelet decomposition: Decompose the original signal into multi-level approximation coefficients a j [k] and detail coefficients d j [k], and the formula is:

[0023] s[n] = y[n] + x[t]

[0024] a j [k] = ∑s[n]f[2k - n]

[0025] d j [k] = ∑s[n]g[2k - n],

[0026] where s[n] is the original signal, composed of the effective signal y[n] and the noise x[t], f[n] is the low-pass filter, and g[n] is the high-pass filter;

[0027] PSO optimization: Initialize the particle swarm. The position of each particle represents the wavelet threshold λ and the decomposition level level, and the objective function is the number of noise points identified after OPTICS clustering;

[0028] The particle velocity and position update formulas are:

[0029]

[0030]

[0031] where k is the number of iterations, c1 and c2 are the individual particle learning factor and the particle swarm learning factor, and r1 and r2 are the particle random search factors, with the value range [0, 1];

[0032] where, is the velocity of the i-th particle in the m-th dimensional search space, is the position of the i-th particle in the m-th dimensional search space; ω (k) is the inertia weight; is the global optimal position of the i-th particle in the m-th dimensional search space; is the historical optimal position of the group in the m-th dimension;

[0033] where the inertia weight ω (k) The calculation formula is as follows:

[0034]

[0035] where k max is the maximum number of iterations, and ω max , ω min are the maximum and minimum inertia weights respectively. ω is proportional to the global search ability. The larger the former value, the greater the convergence speed and amplitude of the search space. On the contrary, the stronger the local fine search ability, and a higher-precision solution can be calculated;

[0036] Soft threshold denoising: The detail coefficients are shrunk, and the formula is:

[0037]

[0038] where λ is the manually set threshold, is the denoising result using the soft threshold for high frequencies, and sgn(·) is the sign function;

[0039] Noise feedback: The discrete points that are not clustered in the denoised signal are determined as noise through OPTICS clustering and fed back to PSO to adjust the threshold λ until the number of noise points is lower than the preset threshold.

[0040] Furthermore, applying the BayOTIDE missing value online real-time imputation model to impute the labeled missing values includes the following specific steps:

[0041] Function decomposition and joint modeling:

[0042] Data decomposition: For the M-dimensional multivariate time series data with missing values, it is decomposed into a linear combination of low-rank and independent latent factors;

[0043] Latent factor modeling: Using the Gaussian process (GP) combined with the Matérn kernel function to model the non-periodic long-term change characteristics in the data to obtain the trend factors; using the Gaussian process combined with the periodic kernel function to model the periodic fluctuation characteristics in the data to obtain the seasonality factors;

[0044] Constructing a joint probability model: Gaussian priors (U_trend, U_season) and gamma priors (τ) are set for the trend factors, seasonality factors, weight coefficients, and noise terms respectively, and a joint probability model is constructed based on these prior distributions;

[0045] Online parameter inference:

[0046] Bayesian online learning: During the continuous input of data, the posterior distributions of the trend factors, seasonality factors, and noise are dynamically updated to adapt to the changes in the data;

[0047] Optimization of the RTS-Smoother algorithm: The RTS-Smoother algorithm is used to perform backward distribution smoothing on the Gaussian process parameters, thereby optimizing the factor estimation results and improving the accuracy of the estimation.

[0048] Collaborative optimization of Conditional Expectation Propagation (CEP): Through the Message Passing & Merging mechanism, real-time collaborative optimization of the model parameters is achieved to ensure the consistency and rationality of the model parameters at different time points.

[0049] Real-time imputation of missing values:

[0050] Based on the updated posterior distribution and the GP prior adjoint form Z(t), probabilistic imputation is performed on the missing values at any time stamp.

[0051] Dynamic imputation of non-uniformly sampled data is supported through continuous-time modeling, and a complete time-series data stream is output.

[0052] Online update of the model:

[0053] The new monitoring data is fed back to the joint probability model, and the Gaussian process prior and posterior distributions are updated iteratively in a loop to achieve the dynamic adaptability of the imputation model.

[0054] Furthermore, the construction of the MTAFM model includes:

[0055] Module definition: The core expert module processes the gas concentration and support resistance data and outputs The auxiliary expert module processes the acoustic emission, electromagnetic radiation, anchor cable stress, microseismic, wind speed, and temperature data and outputs

[0056] Gating fusion: In each layer l, the weight vector of the core module is calculated through a gating unit, and the formula is:

[0057]

[0058] where is a learnable parameter matrix; softmax(.) is the normalized exponential function; is the input of the auxiliary module in the (l - 1)-th layer, representing the output matrix of task t in the (l - 1)-th layer;

[0059] Residual connection: The output of the core expert module is added to the gating result, and the formula is:

[0060]

[0061] where It represents the output of the gating unit, indicating the gating fusion output of task t for the expert network at layer l; It is the residual connection coefficient;

[0062] Multi-layer stacking: Repeat the above steps, and use multi-level cascaded fusion modules to mine the correlation features between metrics and disasters.

[0063] Furthermore, the execution process of the RTS-Smoother algorithm includes the following steps:

[0064] Use BayOTIDE to regard each functional factor as a discrete equivalent linear Gaussian state space model, and its system equation and observation equation are respectively:

[0065] φ t+1 = M t φ t + q t , q t ~ N(0, Q t )

[0066] y t = L t φ t + r t , r t ~ N(0, R t )

[0067] Among them, φ t is the hidden state of the factor at the t-th moment, y t is the observed value, M t , L t are system matrices, Q t , R t are the system variance and the observation noise covariance respectively;

[0068] The following variables can be obtained through online Kalman filtering:

[0069] Filtered state estimate and its filtering covariance and P t|t ; Predicted state value Predicted covariance P t+1|t = A t P t|t A t + Q t ;

[0070] Based on the complete time series data, the following steps are executed backward from t = T - 1 to t = 0 through the RTS smoother:

[0071] Calculate the smoothing gain matrix: G t = P t|t A t (Pt+1|t ) -1 ;

[0072] Update the posterior state estimate:

[0073] Update the posterior covariance: P t|T = P t|t + G t (P t+1|T - P t+1|t )G t .

[0074] Furthermore, the step (3) of determining the lowest early warning threshold of disaster risk based on Bayesian dynamic optimization includes the following steps:

[0075] Determine the objective function: Use the AUC value of the early warning result as the classification effect evaluation index, and the calculation formula is as follows:

[0076]

[0077] where ∑ T rank is the sum of the ranks of the positive sample scores after the overall ranking, M is the positive sample, and N is the number of negative samples;

[0078] Construct a Gaussian process surrogate model:

[0079] Use the Matern kernel function to construct the covariance function

[0080] Define the mean function μ(x) and the Gaussian process prior F(x), and the calculation formulas are respectively:

[0081] μ(x) = E(F(x))

[0082] F(x) ∼ GP(m(x), K(x,x′))

[0083] where the covariance function uses the Matern kernel function, and the expression is:

[0084]

[0085] where v is the smoothing parameter, x and x′ are data points at different positions, l is the length scale parameter, Γ(v) is the gamma function, and K v is the second-kind modified Bessel function;

[0086] Dynamic parameter optimization:

[0087] Optimize the kernel function parameters through the Expected Improvement (EI) strategy, and the predicted function value is calculated using the following formula:

[0088] p(f(x)|f) = N(f(x)|μ, σ2 )

[0089] μ = k T K -1 f

[0090] σ 2 = k(0) - k T K -1 k

[0091] where k is the covariance vector between the prediction point and the known sample points, K is the covariance matrix, f is the objective function, μ is the mean, and σ is the variance.

[0092] Compared with traditional methods, the present invention proposes a method for adaptive data preprocessing and multi-index fusion early warning of coal mine coal-rock gas dynamic disasters. Compared with the prior art, its advantages are as follows:

[0093] 1. Adaptive data preprocessing ability: This application introduces a semi-supervised collaborative training model based on xLSTM to identify outliers and label them as missing values. The acoustic emission and electromagnetic radiation signals are denoised by the PSO-wavelet threshold-OPTICS adaptive denoising model, and the labeled missing values are imputed by the BayOTIDE missing value online real-time imputation model. These methods enhance the adaptive parameter adjustment ability of statistical methods, reduce the labeling requirements of deep learning methods, and improve the online incremental learning and abnormal data processing ability of the model, effectively optimizing the data quality.

[0094] 2. Multi-index fusion early warning model: This application constructs a gas and roof disaster adaptive fusion early warning model (MTAFM) based on an expert network and a gated unit. The model processes gas concentration and support resistance data through the core expert module, processes acoustic emission, electromagnetic radiation, bolt and cable stress, microseismic, wind speed, and temperature data through the auxiliary expert module, and dynamically fuses the output features of the two types of modules using the gated unit to generate the probability of coal-rock gas dynamic disaster risk. This model can adaptively fuse the data of each index, quantify the correlations between indexes, between indexes and disasters, and between coal-rock gas dynamic disaster risks, improving the early warning performance and generalization ability.

[0095] 3. Dynamic hierarchical early warning and compound disaster determination: This application determines the lowest early warning threshold for disaster risk based on Bayesian dynamic optimization and divides the risk levels by combining the quantile method, achieving dynamic hierarchical early warning. At the same time, if both coal-rock and gas disasters are early warned, it is determined as a compound disaster risk, further improving the accuracy and reliability of the early warning.

[0096] 4. Reducing the dependence on professionals: The models and methods proposed in this application can automatically adjust key parameters to adapt to different environments, reducing the dependence on professionals and providing a new theoretical support and technical path for accurate early warning of multi-disaster risks.

[0097] 5. Improve resource utilization efficiency and generalization ability: As the number of indicators and data volume increase, the fusion advantage of this application becomes more significant, and the generalization and early warning effects are further enhanced. When applied to the new coal mine environment, the on-site historical data can be used to fine-tune MTAFM, enabling it to learn and output the risk probability distribution, and adaptively selecting the optimal threshold through hierarchical early warning to ensure the best early warning performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention.

[0099] Figure 1 It is a flowchart of a multi-index adaptive fusion hierarchical early warning method for coal-rock gas dynamic disasters in this application.

[0100] Figure 2 It is the preprocessing model architecture and execution process of this application.

[0101] Figure 3 It is the architecture diagram of the MTAFM model in this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0102] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. Of course, the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0103] Combined with the attached Figure 1 - attached Figure 3 , this application provides a multi-index adaptive fusion hierarchical early warning method for coal-rock gas dynamic disasters, including the following steps:

[0104] (1) Collaborative preprocessing of multi-source data, specifically including:

[0105] Collect the gas concentration, support resistance, electromagnetic radiation, acoustic emission, bolt and cable stress, borehole stress, microseismic energy, wind speed and temperature data of the coal mine excavation face to form an initial data set;

[0106] Identify outliers through a semi-supervised collaborative training model based on xLSTM and label them as missing values;

[0107] Denoise the acoustic emission and electromagnetic radiation signals through a PSO-wavelet threshold-OPTICS adaptive denoising model;

[0108] Interpolate the labeled missing values using the BayOTIDE missing value online real-time interpolation model;

[0109] (2) Disaster feature fusion and risk prediction, specifically:

[0110] Input the preprocessed data into the multi-index fusion early warning model MTAFM. The MTAFM model processes the gas concentration and support resistance data through the core expert module, processes the acoustic emission, electromagnetic radiation, bolt and cable stress, microseismic, wind speed, and temperature data through the auxiliary expert module, and uses a gated unit to dynamically fuse the output features of the two types of modules to generate the risk probability of coal-rock gas dynamic disasters.

[0111] (3) Dynamic hierarchical early warning and compound disaster determination, including:

[0112] Determine the lowest early warning threshold for disaster risk based on Bayesian dynamic optimization;

[0113] Combine the quantile method to divide the risk levels: no early warning if below the threshold, yellow early warning if above the threshold and below the 85th percentile, orange early warning if between the 85th and 90th percentiles, and red early warning if above the 90th percentile; if both coal-rock and gas disasters are early warned, it is determined as a compound disaster risk.

[0114] Furthermore, the outlier identification model in step (1) is based on the collaborative training framework of xLSTM, specifically including: constructing sLSTM and mLSTM as dual-view learning modules, where sLSTM optimizes the information flow processing through an exponential gated unit and a scalar update mechanism, and mLSTM improves the memory and parallel processing capabilities through matrix operations; respectively perform sLSTM and mLSTM predictions on the unlabeled data, and cross-select the samples with confidence higher than the preset threshold and add them to the training set of the other model; optimize the model parameters through iterative training and error backpropagation until the prediction error converges or reaches the maximum number of iterations.

[0115] This application proposes a semi-supervised early warning model based on collaborative training of xLSTM. Here, xLSTM is an extended version that enables LSTM to handle large-scale data. The main idea is to optimize LSTM by introducing exponential gating units and a new memory mixing technique to obtain new memory cells. Stacking these memory cells can achieve better information filtering, memory, storage effects, and higher data processing efficiency. The optimized memory cell variants include sLSTM and mLSTM. The former adds a scalar update mechanism and can process the information flow more accurately, stably, and with high performance through exponential gating units and normalization methods. The latter uses matrix operations to improve the performance of LSTM in memory and parallel processing, enabling it to be applied to larger-scale datasets and effectively enhancing the model's performance in processing rare label predictions. The basic method of semi-supervised learning is to use a small amount of unlabeled data and a large amount of unlabeled data to train the model, aiming to reduce the dependence on labeled data and improve the learning efficiency, prediction performance, and generalization ability of the model under a small amount of labeled data. Since different features can be extracted from unlabeled data under different network views, resulting in divergent prediction results, the cross-complementation between the divergent results of multi-view learning can improve the model's prediction performance. Collaborative training is one of the representative methods of multi-view learning. Its basic idea is to train separately from different views by establishing independent prediction models, then predict the unlabeled data, select credible samples from the prediction results of each model and cross-add them to the training samples of other views for the next training until the model converges or the iteration terminates. The present invention first separately establishes xLSTM based on sLSTM and mLSTM memory cells to learn the features of abnormal sequence data from two different views. Then, a semi-supervised learning framework based on collaborative training is constructed to train the model. Finally, the best model is selected for on-site application through error evaluation and performance evaluation metrics.

[0116] This application proposes an adaptive denoising model based on PSO-wavelet threshold denoising-OPTICS. Among them, wavelet threshold denoising is a non-linear denoising method that uses the time-frequency local decomposition ability of orthogonal wavelet decomposition to decompose the original signal into wavelet coefficients of multiple levels and different scales. Among them, the wavelet coefficients containing noise usually have high-frequency components and low amplitudes, which are different from the low-frequency or stationary components and high amplitudes exhibited by the wavelet coefficients of the effective signal. Based on this feature, the wavelet coefficients of the effective signal and noise can be distinguished by artificially setting a threshold, and the noise signal can be removed. The methods of using the set threshold to denoise the signal include soft threshold and hard threshold. Among them, the hard threshold method sets the wavelet coefficients smaller than the threshold to 0 and retains the coefficients higher than the threshold as they are. This method is prone to loss of signal details, resulting in signal distortion or mutation. The soft threshold method can provide a smoother and more stable denoising result by shrinking the wavelet coefficients with absolute values higher than the threshold. Finally, the denoised wavelet coefficients can be restored by inverse wavelet transform to obtain the denoised signal. The denoising process of the PSO-wavelet threshold-OPTICS denoising model includes the following steps:

[0117] Wavelet decomposition: Decompose the original signal into multi-level approximation coefficients a j [k] and detail coefficients dj[k], and the formula is:

[0118] s[n] = y[n] + x[t]

[0119] a j [k] = ∑s[n]f[2k - n]

[0120] d j [k] = ∑s[n]g[2k - n],

[0121] Soft threshold denoising: Shrink the detail coefficients, and the formula is:

[0122]

[0123] Among them, s[n] is the original signal, which consists of the effective signal y[n] and the noise x[t]. a j [k] is the j-th layer approximation coefficient (including low-frequency components), and f[n] is the low-pass filter. d j [k] is the j-th layer detail coefficient (including high-frequency components), and g[n] is the high-pass filter. λ is the artificially set threshold, is the denoising result using the soft threshold for high frequencies, and sgn(·) is the sign function. Since the threshold needs to be artificially set before applying the threshold denoising method, this is prone to problems such as loss of effective signal or insufficient noise separation. Therefore, the selection of the threshold has an important impact on the noise removal effect, and using an optimization algorithm to calculate the optimal threshold can effectively address these problems.

[0124] PSO Optimization: Initialize the particle swarm. The position of each particle represents the wavelet threshold λ and the decomposition level level, and the objective function is the number of noise points identified after OPTICS clustering. The PSO (Particle Swarm Optimization) is an optimal solution search algorithm designed based on the phenomenon of bird foraging behavior. The basic idea is to generate N particles with two characteristic attributes, namely moving speed and direction. The particles search for the best solution in the solution space through the method of collective information sharing. Assume that the position vector of particle i in the solution space is represented as X im =(x i1 ,x i2 ,…,x iM ), and the movement speed vector is represented as V im =(v i1 ,v i2 ,…,v iM ). During the particle search process, it is necessary to use the objective function to calculate the corresponding positions of the optimal fitness values of individual particles and all particles respectively to dynamically adjust their own movement speeds. Define the position vectors of individual particles and all particles calculated according to the optimal fitness during the search process as P im,pbest =(p i1,pbest ,p i2,pbest ,…,p iM,pbest ) and P m,gbest =(p 1,gbest ,p 2,gbest ,…,p M,gbest ), then the particle speed and position can be updated using the following formula:

[0125]

[0126] where k is the number of iterations, c1 and c2 are the individual particle learning factor and the particle swarm learning factor, and r1 and r2 are the particle random search factors, with the value range of [0,1].

[0127] The calculation formula of the inertia weight ω(k) is as follows:

[0128]

[0129] where k max is the maximum number of iterations, ω max , ω min are the maximum and minimum inertia weights respectively. ω is proportional to the global search ability. The larger the former value, the greater the convergence speed and amplitude of the search space; conversely, the stronger the local fine search ability, and a higher-precision solution can be calculated;

[0130] Soft Threshold Denoising: Perform a shrinking process on the detail coefficients, and the formula is:

[0131]

[0132] where λ is a threshold set manually, is the denoising result using the soft threshold for high frequencies, and sgn(·) is the sign function;

[0133] Noise feedback: The discrete points that are not clustered in the denoised signal are determined as noise through OPTICS clustering, and fed back to PSO to adjust the threshold λ until the number of noise points is lower than the preset threshold.

[0134] To better select the optimal threshold for wavelet threshold denoising, the present invention selects the OPTICS clustering method to identify the noise points in the signal, and takes minimizing the number of noise points as the objective function of the particle swarm algorithm. The core idea of OPTICS for identifying noise is to distinguish noise points by using the characteristics that noise is not easy to cluster and is scattered. The main method is to regard the data points in the high-density clusters obtained after clustering as valid signal points, and the data points that do not belong to any cluster can be regarded as noise or outliers.

[0135] Calculation process of the wavelet threshold denoising model based on PSO: First, wavelet threshold denoising is performed using the optimized threshold obtained by the particle swarm algorithm; then OPTICS is applied to the signal after wavelet threshold denoising to obtain the distribution of noise points; finally, the noise recognition result is used to further guide the optimization direction and speed of the particle swarm algorithm, and a better denoising threshold is calculated. The optimization is terminated when the threshold selection tends to be stable or the maximum number of PSO optimization iterations is reached.

[0136] Furthermore, applying the BayOTIDE missing value online real-time imputation model to impute the labeled missing values includes the following specific steps:

[0137] Function decomposition and joint modeling:

[0138] Data decomposition: For the M-dimensional multivariate time series data with missing values, it is decomposed into a linear combination of low-rank and independent latent factors;

[0139] Latent factor modeling: Using the Gaussian process (GP) combined with the Matérn kernel function to model the non-periodic long-term change characteristics in the data to obtain the trend factors; using the Gaussian process combined with the periodic kernel function to model the periodic fluctuation characteristics in the data to obtain the seasonality factors;

[0140] Constructing a joint probability model: Gaussian priors (U_trend, U_season) and gamma priors (τ) are set for the trend factor, seasonality factor, weight coefficient, and noise term respectively, and a joint probability model is constructed based on these prior distributions;

[0141] Online parameter inference:

[0142] Bayesian Online Learning: During the continuous input of data, dynamically update the posterior distributions of the trend factor, seasonal factor, and noise to adapt to the changes in the data;

[0143] Optimization of the RTS-Smoother algorithm: Use the RTS-Smoother algorithm to perform backward distribution smoothing on the Gaussian process parameters, thereby optimizing the factor estimation results and improving the accuracy of the estimation;

[0144] Collaborative optimization of Conditional Expectation Propagation (CEP): Through the Message Passing & Merging mechanism, achieve real-time collaborative optimization of the model parameters, ensuring the consistency and rationality of the model parameters at different time points;

[0145] Real-time imputation of missing values: Based on the updated posterior distribution and the GP prior adjoint form Z(t), perform Probabilistic Imputation on the missing values at any timestamp; Support the dynamic imputation of non-uniformly sampled data through continuous-time modeling, and output a complete time-series data stream;

[0146] Online update of the model: Feed the newly monitored data back to the joint probability model, and iteratively update the Gaussian process prior and posterior distributions to achieve the dynamic adaption of the imputation model.

[0147] This application uses BayOTIDE to achieve online real-time imputation of missing values for each indicator. BayOTIDE is a Bayesian online multi-variable time series missing value imputation method based on functional decomposition. The basic modeling idea is to first decompose the M-dimensional data function with observed missing values into a linear combination of a group of low-rank and independent latent factors, and then add the priors of their weights and noise to obtain the joint probability distribution of the model. Among them, the latent factors include trend factors and seasonality factors obtained by separately selecting the matérn kernel and periodic kernel for the Gaussian process (GP). Secondly, use Bayesian Online Learning and RTS-Smoother to infer and estimate the posterior distribution of the model parameters online. Finally, use the current posterior distribution and GP prior to perform probabilistic imputation on the missing values at any timestamp, and achieve an imputation model that can model continuous time and be updated online.

[0148] Furthermore, the present application proposes a coal-rock-gas dynamic disaster risk fusion early warning model MTAFM (Multi-hazard Task-Adaptive Fusion Model).

[0149] The present invention proposes a coal-rock-gas dynamic disaster risk fusion early warning model MTAFM (Multi-hazard Task-Adaptive Fusion Model). This model constructs a core expert network module and an auxiliary expert module, quantitatively learns the index data of different disasters, and uses a gating mechanism to fuse the outputs of the core expert module and the auxiliary expert module to form a hierarchical fusion structure. Through the stacking of multi-layer fusion modules, this model can deeply mine the correlation features between different index data and the risks of gas and roof disasters, and achieve knowledge sharing and fusion.

[0150] In view of the different response degrees of each index to the risks of gas and roof disasters, this study selects gas concentration and support resistance as the core early warning indicators, and uses acoustic emission, electromagnetic radiation, bolt and cable stress, microseismic, wind speed, and temperature as auxiliary indicators. The core indicators are respectively input into the core expert modules of the gas and roof early warning tasks, while all auxiliary indicators perform feature learning through the auxiliary expert module. For the fusion early warning of gas and roof disaster risks, the MTAFM model can adaptively integrate the outputs of the core expert module and the auxiliary expert module, fully mine the potential correlations between indicators, between indicators and risks, and between the two types of disaster risks, improve the accuracy and stability of multi-disaster collaborative early warning, and thus effectively improve the adaptability and generalization ability of the early warning model.

[0151] Its principle is as follows:

[0152] Let the input monitoring index data be x, t1 and t2 be the gas and roof disaster risk early warning tasks, and the fusion module stacking layer be L = {l1, l2,..., l n}, and the t-th core fusion module in the l-th layer is denoted as The output is The feature vectors output by the core expert module and the auxiliary expert module are concatenated to form a feature matrix:

[0153]

[0154] Specifically, the construction of the MTAFM model includes:

[0155] Module definition: The core expert module processes gas concentration and support resistance data and outputs The auxiliary expert module processes acoustic emission, electromagnetic radiation, bolt and cable stress, microseismic, wind speed, and temperature data and outputs

[0156] Gated fusion: In each layer l, the weight vector of the core module is calculated by a gating unit, and the formula is:

[0157]

[0158] where is a learnable parameter matrix; then the output of the risk warning task t in layer l through the gating unit is:

[0159]

[0160] Residual connection: Considering the important impact of core indicators on disaster risk warning, the present invention linearly fuses the outputs of multiple core experts in the risk warning task expert module and adds them to the output of the gating unit in the form of residuals. The formula is:

[0161]

[0162] Multi-layer stacking: Repeat the above steps to mine the correlation features between indicators and disasters through a multi-level cascaded fusion module. Through multi-layer stacking training of the fusion module, the weights of the indicators closely related to disaster risk warning gradually increase, and the weights of relatively independent indicators gradually decrease.

[0163] Based on the warning results output by MTAFM, the present invention further proposes an adaptive hierarchical identification method for coal and rock gas dynamic disaster risk based on dynamic Bayesian optimization and quantiles. Bayesian optimization is an efficient algorithm for solving function optimization and is widely used in hyperparameter optimization of machine learning 58, 59. It has high sample efficiency, adaptive search, and real-time optimization capabilities, making it an ideal method for threshold selection. The core of Bayesian optimization includes a surrogate model and an optimization strategy, where the surrogate model is used to model the objective function f(x). Essentially, the surrogate model is a learning model that can be trained based on all observed data and estimate the function value F(x) corresponding to any x. Bayesian optimization usually uses a Gaussian process (GP) to model the surrogate function, which is jointly defined by a mean function and a covariance function (i.e., a kernel function). The former provides the expected performance, and the latter calculates the uncertainty of the estimated value, controlling the smoothness and behavior of the estimated function.

[0164] Specifically, the step of determining the lowest warning threshold for disaster risk based on Bayesian dynamic optimization in step (3) includes the following steps:

[0165] Determine the objective function: Use the AUC value of the warning result as the classification effect evaluation index, and the calculation formula is as follows:

[0166]

[0167] where ∑ Trank is the sum of ranks after the overall ranking of the positive sample scores, M is the positive sample, and N is the number of negative samples;

[0168] Construct a Gaussian process surrogate model:

[0169] Use the Matern kernel function to construct the covariance function

[0170] Define the mean function μ(x) and the Gaussian process prior F(x), and the calculation formulas are respectively:

[0171] μ(x) = E(F(x))

[0172] F(x) ~ GP(m(x), K(x, x'))

[0173] Since AUC has characteristics such as non - smooth and non - convex, the present invention uses the Matern kernel as the kernel function, where the covariance function uses the Matern kernel function, and the expression is:

[0174]

[0175] where ν is the smoothness parameter, l is the length scale parameter, Γ(ν) is the gamma function, and K ν is the second - kind modified Bessel function;

[0176] Dynamic parameter optimization:

[0177] Optimize the kernel function parameters through the Expected Improvement (EI) strategy, and the predicted function value is calculated using the following formula:

[0178] p(f(x)|f) = N(f(x)|μ, σ 2 )

[0179] μ = k T K -1 f

[0180] σ 2 = k(0) - k T K -1 k

[0181] where k is the covariance vector between the prediction point and the known sample points, K is the covariance matrix, f is the objective function, μ is the mean, and σ is the variance. Since there is uncertainty in GP when predicting the next sampling point, the Expected Improvement (EI) strategy can be used for optimization. This method guides the sampling point selection by quantifying the expected improvement relative to the current optimal value, thus effectively balancing "Exploration" and "Exploitation".

[0182] The lowest warning threshold can be dynamically obtained by using Bayesian optimization. Based on this, the adaptive classification warning thresholds for gas and roof disasters are further divided by quantiles. From the probability distribution of the disaster risk warning by the model, it can be seen that for the risk probability with precursor characteristics, it gradually increases with the risk manifestation intensity and reaches the maximum probability when the risk occurs. Therefore, the 90% quantile of all warning probability distributions can be used as the red warning, the 85% quantile as the orange warning, and the rest higher than the lowest warning threshold as the yellow warning, and no warning is given when it is lower than the lowest threshold. Suppose the thresholds of the gas and coal and rock dynamic samples obtained by Bayesian optimization are threshold-gas and threshold-rock respectively. The warning probability ranges for each level of the gas and coal and rock dynamic samples are shown in the following table:

[0183]

[0184] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A multi-index adaptive fusion classification and early warning method for coal-rock gas dynamic disasters, characterized in that, It includes the following steps: (1) Multi-source data collaborative preprocessing, specifically including: Collecting gas concentration, support resistance, electromagnetic radiation, acoustic emission, bolt and cable stress, borehole stress, microseismic energy, wind speed and temperature data of coal mine excavation working face to form an initial data set; Identifying outliers through a semi-supervised collaborative training model based on xLSTM and labeling them as missing values; Denosing acoustic emission and electromagnetic radiation signals through a PSO-wavelet threshold-OPTICS adaptive denoising model; Imputing the labeled missing values using the BayOTIDE missing value online real-time imputation model; (2) Disaster feature fusion and risk prediction, specifically: Inputting the preprocessed data into the multi-index fusion early warning model MTAFM. The MTAFM model processes gas concentration and support resistance data through the core expert module, processes acoustic emission, electromagnetic radiation, bolt and cable stress, microseismic, wind speed and temperature data through the auxiliary expert module, and dynamically fuses the output features of the two types of modules using a gated unit to generate the risk probability of coal and rock gas dynamic disasters; (3) Dynamic hierarchical early warning and compound disaster determination, including: Determining the lowest early warning threshold for disaster risk based on Bayesian dynamic optimization; Combining the quantile method to divide the risk levels: no early warning below the threshold, yellow early warning above the threshold and below the 85th percentile, orange early warning at the 85th - 90th percentile, and red early warning above the 90th percentile; if both coal and rock and gas disasters are early warned simultaneously, it is determined as a compound disaster risk.

2. The multi-index adaptive fusion classification and early warning method for coal-rock gas dynamic disasters according to claim 1, wherein The outlier identification model in step (1) is based on the collaborative training framework of xLSTM, specifically including: Constructing sLSTM and mLSTM as dual-view learning modules, where sLSTM optimizes the information flow processing through an exponential gated unit and a scalar update mechanism, and mLSTM enhances the memory and parallel processing capabilities through matrix operations; Performing sLSTM and mLSTM predictions on unlabeled data respectively, and cross-selecting samples with confidence higher than the preset threshold and adding them to the training set of the other model; Optimizing the model parameters through iterative training and error backpropagation until the prediction error converges or reaches the maximum number of iterations.

3. A multi-index adaptive fusion grading early warning method for coal-rock gas dynamic disasters according to claim 1, characterized in that The denoising process of the PSO-wavelet threshold-OPTICS denoising model includes the following steps: Wavelet decomposition: Decompose the original signal into multi-level approximation coefficients a j [k] and detail coefficients d j [k], and the formula is: s[n] = y[n] + x[t] where s[n] is the original signal, composed of the effective signal y[n] and the noise x[t], f[n] is the low-pass filter, and g[n] is the high-pass filter; PSO optimization: Initializing the particle swarm, where the position of each particle represents the wavelet threshold λ and the decomposition level level, and the objective function is the number of noise points identified after OPTICS clustering; The particle velocity and position update formulas are: where k is the number of iterations, c1 and c2 are the individual particle learning factor and the particle swarm learning factor, and r1 and r2 are the particle random search factors, with the value range of [0, 1]; wherein, is the velocity of the i-th particle in the m-th dimensional search space, is the position of the i-th particle in the m-th dimensional search space; ω (k) is the inertia weight; is the global optimal position of the i-th particle in the m-th dimensional search space; is the historical optimal position of the population in the m-th dimension; Among them, the inertia weight ω (k) The calculation formula is as follows: where k max is the maximum number of iterations, ω max , ω min are the maximum and minimum inertia weights respectively. ω is proportional to the global search ability. The larger the former value, the greater the convergence speed and amplitude of the search space. On the contrary, the stronger the local fine search ability, the higher-precision solution can be calculated; Soft threshold denoising: Shrinking the detail coefficients, the formula is: where λ is a threshold set manually, is the denoising result using the soft threshold for high-frequency components, and sgn(·) is the sign function; Noise feedback: Judging the unclustered discrete points in the denoised signal as noise through OPTICS clustering, and feeding back to PSO to adjust the threshold λ until the number of noise points is lower than the preset threshold.

4. A multi-index adaptive fusion classification and early warning method for coal-rock gas dynamic disasters according to claim 1, characterized in that Imputing the labeled missing values using the BayOTIDE missing value online real-time imputation model includes the following specific steps: Function decomposition and joint modeling: Data decomposition: For the M-dimensional multivariate time series data with missing values, it is decomposed into a linear combination of low-rank and independent latent factors; Latent factor modeling: Using Gaussian process (GP) combined with the Matérn kernel function to model the non-periodic long-term change characteristics in the data to obtain trend factors; using Gaussian process combined with the periodic kernel function (PeriodicKernel) to model the periodic fluctuation characteristics in the data to obtain seasonality factors; Constructing a joint probability model: Setting Gaussian priors (U_trend, U_season) and gamma priors (τ) for the trend factor, seasonality factor, weight coefficient, and noise term respectively, and constructing a joint probability model based on these prior distributions; Online parameter inference: Bayesian Online Learning: During the continuous input of data, dynamically update the posterior distributions of the trend factor, seasonality factor, and noise to adapt to the changes in the data; Optimization of the RTS-Smoother algorithm: Using the RTS-Smoother algorithm to perform backward distribution smoothing on the Gaussian process parameters, thereby optimizing the factor estimation results and improving the accuracy of the estimation; Conditional Expectation Propagation (CEP) collaborative optimization: Through the Message Passing & Merging mechanism, achieve real-time collaborative optimization of the model parameters to ensure the consistency and rationality of the model parameters at different time points; Real-time imputation of missing values: Based on the updated posterior distribution and the GP prior adjoint form Z(t), perform Probabilistic Imputation on the missing values at any timestamp; Support dynamic imputation of non-uniformly sampled data through continuous time modeling, and output a complete time series data stream; Model online update: Feed the newly monitored data back to the joint probability model, and iteratively update the Gaussian process prior and posterior distributions to achieve the dynamic adaptability of the imputation model.

5. A multi-index adaptive fusion classification and early warning method for coal-rock gas dynamic disasters according to claim 1, characterized in that The construction of the MTAFM model includes: Module definition: The core expert module processes gas concentration and support resistance data and outputs The auxiliary expert module processes acoustic emission, electromagnetic radiation, anchor cable stress, microseismic, wind speed and temperature data and outputs Gated fusion: In each layer l, calculate the weight vector of the core module through the gated unit, and the formula is: where is a learnable parameter matrix; softmax(.) is the softmax function; is the input of the auxiliary module in the (l - 1)-th layer, representing the output matrix of task t in the (l - 1)-th layer; Residual connection: Add the output of the core expert module and the gated result, and the formula is: Among them is the output of the gating unit, representing the gating fusion output of task t for the expert network at layer l; is the residual connection coefficient; Multi-layer stacking: Repeat the above steps, and mine the correlation characteristics between indicators and disasters through multi-level cascaded fusion modules.

6. The multi-index adaptive fusion grading early warning method for coal-rock gas dynamic disasters according to claim 5, characterized in that The execution process of the RTS-Smoother algorithm includes the following steps: Using BayOTIDE to regard each functional factor as a discrete equivalent linear Gaussian state space model, and its system equation and observation equation are respectively: φ t+1 = M t φ t + q t , q t ~N(0, Q t ) y t = L t φ t + r t , r t ~ N(0, R t ) where, φ t is the hidden state of the factor at the t-th moment, y t is the observed value, M t , L t are the system matrices, Q t , R t are the system variance and the observation noise covariance, respectively; The following variables can be obtained through online Kalman filtering: Filtered state estimate value and its filtering covariance and P t|t ; Predicted state value Predicted covariance P t+1|t = A t P t|t A t + Q t ; Based on the complete time series data, perform the following steps backward from t = T - 1 to t = 0 through the RTS smoother: Calculate the smoothing gain matrix: G t = P t|t A t (P t+1|t ) -1 ; Update the posterior state estimate: Update the posterior covariance: P t|T = P t|t + G t (P t+1|T - P t+1|t ) G t .

7. A multi-index adaptive fusion classification and early warning method for coal-rock gas dynamic disasters according to claim 1, characterized in that, In step (3), determining the lowest early warning threshold for disaster risk based on Bayesian dynamic optimization includes the following steps: Determine the objective function: Use the AUC value of the early warning result as the evaluation index for classification effect, and the calculation formula is as follows: where ∑ T rank is the sum of the ranks of the positive sample scores after overall ranking, M is the number of positive samples, and N is the number of negative samples; Construct a Gaussian process surrogate model: Use the Matern kernel function to construct the covariance function Define the mean function μ(x) and the Gaussian process prior F(x), and the calculation formulas are respectively: μ(x) = E(F(x)) F(x) ∼ GP(m(x), K(x, x')) Among them, the covariance function uses the Matern kernel function, and the expression is: where \(v\) is the smoothing parameter, \(x\) and \(x'\) are data points at different positions, \(l\) is the length scale parameter, \(\Gamma(v)\) is the gamma function, and \(K\) v is the modified Bessel function of the second kind; Dynamic parameter optimization: Optimize the kernel function parameters through the expected improvement strategy (EI), and the predicted function value is calculated using the following formula: p(f(x)|f) = N(f(x)|μ,σ 2 ) μ = k T K -1 f σ 2 = k(0) - k T K -1 k Where k is the covariance vector between the prediction point and the known sample points, K is the covariance matrix, f is the objective function, μ is the mean, and σ is the variance.

Images