A coal body damage and destruction precursor information early warning method based on a BP neural network

Abstract

The application discloses a coal body damage and destruction precursor information early warning method based on a BP neural network, and comprises the following steps: performing principal component analysis on an acoustic emission event data set, and dividing time period points of an energy storage period and a destruction period in a coal body damage process according to the principal component analysis; according to the divided time period, the coal body damage comprehensive evaluation index CDCI of the acoustic emission event data set is calculated in sections; a four-layer BP neural network is constructed; the connection weight and the neuron bias term of the BP neural network are iteratively optimized by adopting a sparrow search algorithm; the real-time collected acoustic emission multivariate parameters are input into the optimized BP neural network, the coal body damage comprehensive evaluation index and the damage grade at the current time are predicted, and the coal body damage and destruction precursor information early warning is performed. In view of the low prediction accuracy caused by the high-dimensional sparsity of the acoustic emission signal in the coal body damage and destruction precursor in the prior art, the application effectively identifies the key features of different damage modes, and provides accurate precursor early warning for the coal body damage.

Description

Technical Field

[0001] This application relates to the field of coal body analysis, and in particular to a method for early warning of coal body damage and destruction precursors based on BP neural network. Background Technology

[0002] Acoustic emission (AE) technology, as a non-destructive testing method, can monitor the elastic wave signals released during the generation and propagation of microcracks within coal seams in real time, and has become an important means of monitoring coal damage evolution. During the loading and failure process of coal, multi-dimensional acoustic emission parameters are generated, including ring count, RA value (rise time / amplitude), AF value (average frequency), dominant frequency, and amplitude. These parameters reflect the damage state inside the coal seam from different perspectives: ring count reflects the frequency of crack activity, RA and AF values ​​can be used to distinguish between tensile and shear failure modes, dominant frequency reflects energy concentration characteristics, and amplitude represents the failure intensity. Theoretically, comprehensive analysis of these multi-dimensional parameters can fully characterize the coal damage evolution process.

[0003] However, in practical applications, the acoustic emission signals from coal body failure exhibit significant high-dimensional sparsity. Specifically: First, the acoustic emission signals are highly non-uniformly distributed over time. The coal body remains relatively stable for most of the time, with weak acoustic emission activity. High-energy pulse signals are only generated at the moment when stress concentration leads to rapid microcrack propagation, and the effective data usually accounts for less than 5%. Second, there are complex nonlinear coupling relationships among the multidimensional parameters. The sensitivity and dominance of different parameters change dynamically at different damage stages, making it difficult for a single parameter to accurately characterize the damage state. Third, the heterogeneity of the internal structure of the coal body leads to diverse failure modes, with tensile, shear, and mixed failure modes coexisting. The acoustic emission characteristics of different failure modes differ significantly, increasing the difficulty of pattern recognition.

[0004] Traditional coal seam damage early warning methods mainly employ threshold discrimination or simple statistical analysis methods. These methods typically consider only a single parameter or a simple combination of parameters, ignoring the intrinsic correlation and time-varying characteristics between parameters. In recent years, backpropagation (BP) neural networks have been introduced into the field of coal seam damage prediction due to their powerful nonlinear mapping capabilities. However, standard BP neural networks face serious challenges when processing high-dimensional sparse data: on the one hand, sparse data leads to extremely uneven distribution of training samples in the high-dimensional feature space, making the network prone to overfitting to local samples and exhibiting poor generalization ability; on the other hand, high-dimensional input increases network complexity, making the gradient descent algorithm highly susceptible to local extremum traps formed by sparse data, resulting in slow convergence speed and low accuracy; furthermore, the fixed network structure is difficult to adapt to the characteristic changes at different stages of coal seam damage evolution, and its sensitivity to identifying precursory damage is insufficient.

[0005] While existing research has attempted to improve the performance of BP neural networks by increasing hidden layer nodes and adjusting the learning rate, it has not fundamentally solved the problems caused by high-dimensional sparse data. Although dimensionality reduction techniques such as principal component analysis have been used for data preprocessing, most studies treat them only as independent preprocessing steps, failing to deeply integrate them with neural network models or consider the stage-specific characteristics of coal damage evolution. Furthermore, traditional optimization algorithms are inefficient in searching high-dimensional parameter spaces, making it difficult to find the global optimum. These problems result in low accuracy and high false alarm rates in existing coal damage early warning systems, particularly their insufficient ability to identify precursory damage stages, thus failing to meet the actual needs of safe coal mine production.

[0006] Therefore, there is an urgent need to develop a new method that can effectively process high-dimensional sparse acoustic emission data, accurately identify the stages of coal damage evolution, and achieve precise early warning of damage precursors. Summary of the Invention

[0007] To address the low prediction accuracy caused by the high-dimensional sparseness of acoustic emission signals in the precursors of coal body damage in existing technologies, this application provides a method for early warning of precursors of coal body damage based on BP neural network. By using principal component analysis to reduce the dimensionality and extract features of the high-dimensional sparse acoustic emission multivariate parameters, the method effectively identifies the key features of different damage modes and provides accurate early warning of coal body damage.

[0008] One aspect of this application provides a method for early warning of coal body damage precursors based on a BP neural network, comprising: S1, collecting acoustic emission signals during the coal body failure process, extracting features from the acoustic emission signals to obtain multivariate parameters, including: ring count, RA value, AF value, dominant frequency, and amplitude; preprocessing the multivariate parameters to form an m×5 dimensional acoustic emission event dataset, where m is the number of acoustic emission events; S2, performing principal component analysis on the acoustic emission event dataset, and dividing the coal body failure process into energy storage period and failure period based on the principal component analysis, wherein the energy storage period is considered as the precursor stage of coal body damage; S3, based on... According to the divided time periods, the comprehensive coal damage evaluation index (CDCI) of the acoustic emission event dataset is calculated segment by segment; S4, a four-layer BP neural network is constructed to establish the mapping relationship between multivariate parameters and damage level; S5, the sparrow search algorithm is used to iteratively optimize the connection weights and neuron bias terms of the BP neural network to minimize the mean square error between the CDCI prediction value output by the BP neural network and the CDCI value calculated in step S3; S6, the real-time collected acoustic emission multivariate parameters are input into the optimized BP neural network to predict the comprehensive coal damage evaluation index and damage level at the current moment, so as to provide early warning of coal damage precursor information.

[0009] Furthermore, in S2, principal component analysis is performed on the acoustic emission event dataset. Based on the principal component analysis, the time nodes of the energy storage period and the failure period during the coal body failure process are determined, including: calculating the temporal change rate of the scores of each principal component, and marking the moment when the temporal change rate first exceeds the threshold as the starting point of the energy storage period. The moment when the principal component score reaches its maximum value is marked as the starting point of the destruction period. ; Time period Defined as the precursor stage of coal body damage and destruction;

[0010] Furthermore, the temporal rate of change of the scores of each principal component is calculated, and the moment when the temporal rate of change first exceeds the threshold is marked as the starting point of the energy storage period. This includes: standardizing the m×5 dimensional acoustic emission event dataset to obtain the standardized matrix X: ;in, This represents the element in the i-th row and j-th column of the normalized matrix X; Let j be the j-th parameter value of the i-th acoustic emission event, where j=1, 2, 3, 4, 5 correspond to the ring count, RA value, AF value, master control frequency, and amplitude, respectively. and Let X be the mean and standard deviation of the j-th parameter across all m acoustic emission events, respectively; calculate the covariance matrix of the normalized matrix X. Where C is a 5×5 symmetric matrix. Let X be the transpose of X; perform eigenvalue decomposition on the covariance matrix C to obtain 5 eigenvalues. and the corresponding feature vectors ,in, This represents the k-th eigenvalue. Indicates and The corresponding unit eigenvector; calculate the first principal component. Time series score: ,in, The first eigenvector The j-th component, Indicates the time when the i-th acoustic emission event occurs. express The score of the first principal component at time step 1; calculate the temporal rate of change of the score of the first principal component: ;in, express The rate of change over time, This indicates the time when the (i-1)th acoustic emission event occurs;

[0011] Determine the threshold: Calculate the average of the time series change rates over the first 20% of the time period as the benchmark value. Set threshold ; Start traversing from the starting point of the time series, when First time exceeding the threshold And when the rate of change at three consecutive time points exceeds the threshold, then that time point... Marked as the starting point of the energy storage period Score the first principal component. The moment when the maximum value is reached is marked as the start point of the destruction period. ,Right now .

[0012] In particular, the damage evolution of coal under external forces has obvious stage characteristics: in the initial elastic stage, acoustic emission activity is weak and stable; when the stress reaches a certain level, microcracks begin to initiate and slowly expand, and acoustic emission activity gradually increases; when failure is imminent, cracks expand rapidly and penetrate each other, and acoustic emission activity grows exponentially.

[0013] Traditional methods typically use the absolute values ​​of acoustic emission parameters as criteria, setting fixed thresholds to classify damage stages. This method has fundamental flaws: first, the differences in the physical and mechanical properties of different coal bodies lead to significant variations in the absolute values ​​of acoustic emission, making the fixed thresholds lack universality; second, the absolute value criterion only reflects the state at a specific moment and cannot capture the dynamic process of damage evolution; and third, single-parameter criteria ignore the coupling relationships between multiple parameters, resulting in insufficient information utilization.

[0014] This scheme calculates the time-series rate of change of principal component scores. The rate of change directly reflects the speed of damage accumulation within the coal body and is an essential characteristic of damage evolution dynamics. When the coal body transitions from the elastic stage to the damage accumulation stage, the initiation and propagation of microcracks lead to a sharp increase in acoustic emission activity, manifested as a sudden increase in the rate of change of principal component scores. This abrupt change is an external manifestation of the transformation of the internal structure of the coal body from stability to instability and has clear physical significance.

[0015] More importantly, the first principal component It is the projection of five acoustic emission parameters onto the direction of maximum variance, automatically extracting the information components most relevant to damage evolution from each parameter. Dimensionality reduction effectively solves the problem of processing high-dimensional sparse data. The temporal change rate of the principal component scores comprehensively reflects the coordinated change characteristics of multiple parameters, and is more stable and reliable than the change rate of a single parameter.

[0016] Furthermore, the comprehensive evaluation index of coal body damage was calculated from the acoustic emission event dataset. : ;in, Let be the value of the i-th parameter at time t. for The average value of the i-th parameter within the time period. for The maximum value of the i-th parameter within the time period. The weight of the i-th parameter is determined based on the contribution rate of the principal components;

[0017] In particular, CDCI in this application uses a dynamic baseline value. and dynamic maximum value Instead of traditional fixed benchmarks or global normalization methods.

[0018] Traditional damage assessment methods typically employ two benchmark strategies: one is to use the parameter values ​​at the start of the experiment as the zero benchmark. This method ignores the natural fluctuations in the initial state of the coal body and is prone to misinterpreting normal fluctuations as damage. The other is to use the minimum and maximum values ​​of the entire data set for normalization. This method obscures the characteristic differences between different evolution stages and cannot accurately reflect the gradual nature of damage. A common drawback of both methods is that they use static benchmarks, which cannot adapt to the dynamic characteristics of coal body damage evolution.

[0019] This application will use the benchmark value Defined as The average value over the period of time The time period corresponds to the initial elastic stage of the coal body. During this stage, the coal body structure is intact, and acoustic emission activity mainly originates from the random activity of internal stress adjustment and minor defects, exhibiting stable statistical characteristics. Using the average value of this time period as a benchmark eliminates the influence of instantaneous fluctuations while accurately characterizing the initial state of a specific coal body.

[0020] Pick The maximum value within the time period, with time t2 corresponding to the critical point of macroscopic damage to the coal body, at which point all acoustic emission parameters reach their peak values. Through The normalization process maps parameters with different dimensions and numerical ranges to the interval [0, 1], where 0 represents the undamaged state and 1 represents the critical damage state.

[0021] This application establishes a quantitative relationship between acoustic emission parameters and the degree of damage. When CDCI approaches 0, it corresponds to a no-damage state; when Approaching When the CDCI approaches 1, it corresponds to a critical failure state.

[0022] Furthermore, The weight of the i-th parameter, determined based on the principal component contribution rate, is calculated using the following formula: Based on the standardized matrix X from step S2, the m×m correlation coefficient matrix R is calculated: ;in, This represents the correlation coefficient between the i-th parameter and the j-th parameter, where i, j = 1, 2, ..., 5, and Perform eigenvalue decomposition on the correlation coefficient matrix R: Where E is a 5×5 identity matrix, and 5 eigenvalues ​​are obtained by solving. and the corresponding feature vectors ;

[0023] Calculate the contribution rate of each principal component and the cumulative contribution rate: Contribution rate: Cumulative contribution rate: When the cumulative contribution rate reaches 80% or more, select the top p principal components; construct the factor loading matrix ρ: ;in, This represents the factor loading of the j-th parameter on the i-th principal component. For the i-th eigenvector The j-th component; calculate the coefficient matrix of the principal component analysis function: ;in, Let represent the coefficient of the j-th parameter in the first principal component. Let j be the factor loading of the j-th parameter on the first principal component; normalize the coefficients to obtain the weights of each parameter: Where i=1, 2, 3, 4, 5 correspond to the weights of ring count, RA value, AF value, main control frequency, and amplitude, respectively.

[0024] Further, in step S4, a four-layer BP neural network is constructed, including: constructing an input layer with five input nodes that respectively receive the normalized ring count, RA value, AF value, master control frequency, and amplitude; and simultaneously processing the time nodes determined in step S2. , Converted to time series labels [0, 1, 2], respectively corresponding to , , Three time periods; construct a feature fusion layer, set n hidden nodes, and each hidden node is activated by a function. Process the input layer data, where, For connection weights, As a bias term; construct a feature output layer, and adjust the connection weights according to the comprehensive coal damage evaluation index CDCI and the current coal stress value σ(t). and bias terms And based on the adjusted connection weights and bias terms Output CDCI prediction values; construct a damage decision layer and classify damage levels according to CDCI prediction values, including no damage, weak damage and strong damage.

[0025] Furthermore, the connection weights are adjusted based on the comprehensive coal damage evaluation index CDCI and the current coal stress value σ(t). and bias terms This includes: determining the baseline parameters corresponding to the current time t based on the CDCI calculation formula in step S3; when hour, Take the average value of the i-th parameter within the time interval [0, t]; when hour, Pick The average value of the i-th parameter within the time period; when hour, Pick The average value of the i-th parameter within the time period. Pick Find the maximum value of the i-th parameter within the time period; map the output vector of the fusion layer to the CDCI predicted value using the CDCI calculation formula; obtain the coal stress value σ(t) at the current moment and calculate the normalized stress level: ;when And when the CDCI predicted value is <0.33, or when And when the CDCI predicted value is <0.66, or When the error is large, the CDCI prediction value is fed back to the fusion layer, and the connection weights and bias terms are adjusted through the backpropagation algorithm; when the error is small, the CDCI prediction value is passed to the decision layer.

[0026] In particular, this application constructs an adaptive learning mechanism based on physical constraints in the design of the BP neural network. By embedding the physical laws and temporal constraints of coal damage evolution into the feature output layer, it achieves a deep integration of data-driven and physical mechanisms, fundamentally improving the reliability and physical rationality of the neural network in coal damage prediction.

[0027] Traditional backpropagation (BP) neural networks have a fatal flaw when dealing with coal damage prediction: the network relies entirely on data-driven approaches, treating the input-output mapping as a purely mathematical fitting problem, ignoring the physical laws that coal damage evolution must follow. This black-box learning method can lead to predictions that violate physical principles, such as predicting strong damage during the energy storage period or predicting low damage levels under high stress.

[0028] This application introduces a dual physical constraint mechanism in the feature output layer. The first constraint is a temporal rationality constraint, which constrains the network output by setting reasonable intervals for CDCI at different evolution stages: when During the energy storage period, CDCI should be in the range of [0.33, 0.66]; when During the (damage period), CDCI should be greater than 0.66. This constraint is based on the objective laws of coal damage evolution—damage is necessarily an increasing and irreversible process, and the degree of damage at different stages has clear physical boundaries.

[0029] The second constraint is the stress-damage consistency constraint, which is achieved by introducing a normalized stress level. This establishes a physical correlation between external loads and internal damage. According to the principles of damage mechanics, under the same material conditions, stress level and damage degree should maintain a positive correlation. When This indicates that the predicted damage level is seriously inconsistent with the actual stress level, violating basic mechanical laws.

[0030] This dual constraint mechanism enables real-time network correction through an error assessment mechanism. When a physically unreasonable prediction is detected, the system does not simply accept the network output but identifies it as a state with a large error, triggering a backpropagation mechanism to adjust the network weights accordingly. This design transforms physical constraints into learning signals for the network, guiding the network to optimize parameters while adhering to physical laws.

[0031] This application, firstly, improves the physical plausibility of predictions, ensuring that the network output consistently conforms to the fundamental laws of coal damage evolution, avoiding potentially absurd results from purely data-driven methods. Secondly, it accelerates network convergence; physical constraints provide clear directional guidance for network learning, reducing ineffective searches in physically unreasonable regions and significantly improving training efficiency. Finally, it enhances the model's generalization ability; through the constraints of physical laws, the network learns not only the statistical relationships between data but also the underlying physical mechanisms, enabling the model to provide reasonable predictions even when facing new working conditions.

[0032] Furthermore, damage levels are classified based on CDCI predicted values, including: when When it is classified as no damage; when When classified as weak damage; when The time is classified as severe damage.

[0033] Furthermore, in step S5, the sparrow search algorithm is used to optimize the weights and thresholds of the BP neural network, including: adjusting the connection weights in the four-layer BP neural network constructed in step S4. Bias terms of the feature fusion layer The position vectors of sparrow individuals are encoded, and N sparrow individuals are initialized to form a population. The fitness function is defined as the mean squared error between the CDCI predicted value output by the BP neural network and the CDCI true value calculated in step S3: F = (1 / M) × Σ(CDCI predicted(t) - CDCI true(t))²; where M is the number of training samples, and CDCI true(t) is calculated according to the formula in step S3. During the iteration process, the population is divided into discoverers and followers based on the fitness value: the discoverer performs a global search based on its current position and updates its position to find the optimal solution; the follower performs a local search based on the discoverer's position and searches near the optimal solution. The population position is updated through multiple iterations, and the iteration terminates when the preset convergence condition is met. The position vector of the best individual is encoded as the optimized connection weight. Bias terms of the feature fusion layer ;

[0034] In particular, this application introduces the exploration-development dynamic equilibrium mechanism of the sparrow search algorithm to optimize the parameters of the BP neural network, which fundamentally solves the problem that the traditional gradient descent method is prone to getting trapped in local optima when processing high-dimensional sparse acoustic emission data, and realizes efficient global search of the neural network parameter space.

[0035] Specifically, the core challenge faced by BP neural networks in coal seam damage prediction is the complexity of parameter optimization. For the four-layer network structure of this scheme, the parameters that need to be optimized include 5×n weights from the input layer to the fusion layer and n bias terms, forming a high-dimensional parameter search space. In this space, due to the high-dimensional sparsity of acoustic emission data, the objective function (mean squared error) exhibits a large number of local extrema. Traditional gradient descent methods update parameters based on local gradient information, which are easily trapped in suboptimal solutions in the steep valleys formed by sparse data, making it impossible to find the globally optimal parameter configuration.

[0036] Furthermore, this application encodes neural network parameters as sparrow individual position vectors. This encoding method maps the discrete network structure to a continuous search space, enabling the optimization mechanism of swarm intelligence algorithms to directly optimize network parameters. Each sparrow individual represents a complete network configuration, and the evolutionary process of the population is the optimization process of network parameters.

[0037] Another aspect of this application provides a coal seam damage precursor information early warning system based on BP neural network, comprising: a data acquisition module, which acquires acoustic emission signals during the coal seam damage process, extracts features from the acoustic emission signals to obtain multivariate parameters, the multivariate parameters including ring count, RA value, AF value, master control frequency and amplitude, and preprocesses the multivariate parameters to form an m×5 dimensional acoustic emission event dataset, where m is the number of acoustic emission events; and a stage division module, which performs principal component analysis on the acoustic emission event dataset, calculates the temporal change rate of the first principal component score, and marks the moment when the temporal change rate first exceeds a threshold as the energy storage period start point. The moment when the principal component score reaches its maximum value is marked as the starting point of the destruction period. and time period Defined as the precursor stage of coal body damage and failure; the index calculation module calculates the comprehensive coal body damage evaluation index (CDCI) for the acoustic emission event dataset in segments according to the divided time periods; the BP neural network construction module constructs a four-layer BP neural network, including: an input layer that receives normalized multivariate parameters and time series labels, a feature fusion layer with n hidden nodes, and a layer that calculates the comprehensive coal body damage evaluation index (CDCI) based on the CDCI and coal body stress values. The system comprises a feature output layer for physical constraint determination and a damage decision layer for damage level classification; a parameter optimization module that uses a sparrow search algorithm to iteratively optimize the connection weights and neuron biases of the BP neural network, encoding the network parameters as sparrow individual position vectors, and using the mean square error between the CDCI predicted value output by the BP neural network and the CDCI true value calculated by the damage evaluation index calculation module as the fitness function, and finding the optimal parameter configuration through a discoverer-follower collaborative search mechanism; and a real-time early warning module that inputs the real-time collected acoustic emission multivariate parameters into the optimized BP neural network to predict the coal body damage comprehensive evaluation index and damage level at the current moment.

[0038] Compared to existing technologies, the advantages of this application are:

[0039] This application collects multivariate acoustic emission parameters (ring count, RA value, AF value, master frequency, and amplitude) during the coal body failure process, accurately divides the time nodes of the energy storage period and the failure period using principal component analysis, and determines the weight of each parameter based on the principal component contribution rate to construct the comprehensive evaluation index CDCI for coal body damage; then, it establishes a four-layer BP neural network model and uses the sparrow search algorithm to optimize the network parameters to achieve a precise mapping between multivariate parameters and damage level.

[0040] This application accurately identifies the precursor stages of coal body damage and uses the energy storage period as a key early warning period; it comprehensively utilizes multivariate acoustic emission parameters, eliminates parameter redundancy through principal component analysis, and extracts key features; it establishes a dynamic damage evaluation system, adaptively adjusting the evaluation benchmark according to different damage stages; and it predicts the comprehensive evaluation index and damage level of coal body damage in real time, providing accurate early warnings for coal body damage. Attached Figure Description

[0041] This application will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0042] Figure 1 This is an exemplary flowchart illustrating a method for early warning of coal body damage precursors based on a BP neural network, according to some embodiments of this application.

[0043] Figure 2 This is a coal seam damage and destruction precursor information early warning model based on a BP neural network, as shown in some embodiments of this application;

[0044] Figure 3 This is a schematic diagram of the data input and model analysis and prediction process according to some embodiments of this application;

[0045] Figure 4 This involves comparing the correlation between model-predicted values ​​and actual values ​​as shown in some embodiments of this application;

[0046] Figure 5 This is a comparison of the model-predicted CDCI and the actual CDCI combined risk index shown in some embodiments of this application;

[0047] Figure 6 This involves comparing the predicted CDCI values ​​of the GA-BP model with the actual CDCI values, as shown in some embodiments of this application.

[0048] Figure 7 This is a comparison of the predicted CDCI values ​​and the actual CDCI values ​​of the PSO-BP model shown in some embodiments of this application;

[0049] Figure 8 This is a comparison of the predicted CDCI values ​​and the actual CDCI values ​​of the SSA-BP model shown in some embodiments of this application;

[0050] Figure 9 These are fitness curves of the GA-BP model shown in some embodiments of this application;

[0051] Figure 10These are fitness curves of the PSO-BP model shown in some embodiments of this application;

[0052] Figure 11 The fitness curves of the SSA-BP model are shown in some embodiments of this application. Detailed Implementation

[0053] The methods and systems provided in the embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0054] like Figure 1 As shown, acoustic emission signals during the coal body failure process are collected, and feature extraction is performed on the acoustic emission signals to obtain multivariate parameters, including ring count, RA value, AF value, master control frequency, and amplitude. The multivariate parameters are preprocessed to form an m×5 dimensional acoustic emission event dataset, where m is the number of acoustic emission events. Principal component analysis is performed on the acoustic emission event dataset to divide the coal body failure process into energy storage period and failure period, with the energy storage period serving as the precursor stage of coal body damage. Based on the divided time periods, the comprehensive coal body damage evaluation index (CDCI) of the acoustic emission event dataset is calculated segment by segment. A four-layer BP neural network is constructed to establish the mapping relationship between multivariate parameters and damage level. A sparrow search algorithm is used to iteratively optimize the connection weights and neuron bias terms of the BP neural network to minimize the mean square error between the CDCI prediction value output by the BP neural network and the CDCI value calculated in step S3. The real-time collected acoustic emission multivariate parameters are input into the optimized BP neural network to predict the comprehensive coal body damage evaluation index and damage level at the current moment, providing early warning information for precursors of coal body damage.

[0055] First, for different individual multivariate information parameters, their individual damage risk indices are determined. Then, based on the closely correlated quantity analysis results constructed from PCA principal component scores in the previous section, the final calculation formula for the coal body damage comprehensive evaluation index (CDCI) after multivariate information fusion is determined as follows: ; In the formula, The risk index is a single parameter indicator. Let t be the value of a single information parameter emitted by acoustic emission at time t; This is the average value of the single parameter; This is the maximum value of the parameter; The single hazard index weight value for the i-th parameter information at time t is determined based on the PCA contribution rate in the previous section. Specifically, in this embodiment, the coal body damage degree classification based on the comprehensive evaluation index is detailed in Table 1.

[0056] Table 1 Classification of Coal Body Damage Degree Based on Comprehensive Evaluation Index

[0057]

[0058] This application uses a BP neural network model as the main carrier and the coal damage discrimination index (CDCI) based on acoustic emission multivariate information as the core output index of the model.

[0059] like Figure 2 As shown, the main structure of this model consists of four parts: input layer: extraction of individual feature parameters of acoustic emission multivariate information parameters; fusion layer: fusion of features of different parameter information; output layer: verification and output based on coal damage degree discrimination index (CDCI); and decision layer: stability judgment result based on coal damage precursor information.

[0060] The feature input layer uses five characteristic parameters from the acoustic emission multivariate information parameters—ringing count, acoustic emission RA, AF value, acoustic emission master control frequency, and amplitude—as the main precursor information parameters for judging coal body instability and failure. The verification parameters are input after processing the results of the coal body damage and failure stage division into energy storage period III and failure period IV based on acoustic emission multivariate information fusion, as well as the corresponding stress-time curve characteristic values ​​of the coal body, to verify and check the judgment results in the subsequent output layer.

[0061] Feature fusion layer: For multiple different input parameter feature values, the comprehensive risk index of each individual parameter is first calculated, and then the comprehensive risk index is calculated based on the weight values. Within the fusion layer, learning is performed based on a large amount of associated datasets and result information to dynamically adjust and determine the weight values ​​between different information.

[0062] Feature Output Layer: The output coal damage discrimination index (CDCI) results are verified with the coal damage and failure stage division results of energy storage period III and failure period IV based on acoustic emission multivariate information fusion and the corresponding coal stress as verification parameters. If the judgment information error is large, it is returned to the fusion layer; if the judgment error is small, the result is output to the decision layer.

[0063] Damage Decision Layer: Based on the coal damage discrimination index (CDCI) output by the output layer, and according to the risk classification results of the comprehensive evaluation index of coal damage, the risk level of coal damage and instability is classified, thereby achieving the purpose of early warning of coal damage and instability at the current moment.

[0064] like Figure 3The diagram illustrates the overall process of the coal body damage precursor early warning model based on the constructed BP neural network model architecture. First, acoustic emission feature information is extracted from the acoustic emission signal dataset obtained from the uniaxial compression test acoustic emission monitoring of the coal body, i.e., acoustic emission multivariate information parameter extraction. The occurrence time nodes before and after the coal body failure energy storage period III, obtained through PCA processing and analysis, are used as the basis for classifying different damage degrees (no damage, weak damage, strong damage) of the coal body sample. Based on the time span, the actual comprehensive damage index of the coal body reflected by acoustic emission events in different failure stages is determined. The extracted acoustic emission multivariate information parameters are input into the neural network model, and the model is used for comprehensive hazard index evaluation optimization and subsequent prediction and early warning work.

[0065] Multivariate acoustic emission data of coal samples at different stages of loading failure—initial compaction, elasticity, and yielding—were analyzed and processed. The sample data were divided into a training set (70%) and a test set (30%) in a 7:3 ratio. A BP neural network-based early warning model for coal damage precursors was established to determine the type of precursor information for coal damage. The predicted coal damage results are shown in Table 2. The correlation between the model predictions and actual values ​​is shown in... Figure 4 As shown, the comparison results under different test samples are as follows: Figure 5 As shown.

[0066] Based on the above Figure 4 , Figure 5 The results in Table 2 show that the linear relationship between the model's predicted values ​​and the actual coal body comprehensive damage index is relatively good, indicating that the current BP neural network model can make certain predictions about coal body damage. However, the judgment results are still not accurate enough. Therefore, we seek an intelligent optimization algorithm to optimize the hyperparameters of the established BP neural network model and obtain a more accurate and scientific early warning model for coal body damage and destruction through the optimization model.

[0067] Table 2. Acoustic emission multi-factor information input and coal body damage prediction

[0068]

[0069] This application uses the Sparrow Search algorithm to optimize the weights and thresholds of the BP neural network.

[0070] Specifically, the connection weights and neuron bias terms in the four-layer BP neural network constructed in step S4 are encoded into sparrow individual position vectors, including the connection weights from the input layer to the feature fusion layer. Bias terms of the feature fusion layer Initialize N sparrow individuals to form a population;

[0071] The fitness function is defined as the mean squared error between the CDCI predicted value output by the BP neural network and the CDCI true value calculated in step S3:

[0072] F = (1 / M) × Σ(CDCI prediction (t) - CDCI actual (t))²

[0073] Where M is the number of training samples, and CDCI true(t) is calculated according to the formula in step S3;

[0074] In each iteration, sparrow individuals are sorted according to their fitness values, and the top 20% of individuals with the lowest fitness values ​​are selected as discoverers, whose positions are updated as follows:

[0075] ,when hour

[0076] ,when hour

[0077] in, Let be the position of the i-th individual in the t-th iteration, α be the contraction factor, and T be the maximum number of iterations. Let S be a random number in the range [0, 1], ST be the safety threshold, Q be a normally distributed random number, and L be a vector of all 1s;

[0078] The remaining individuals, acting as followers, update based on the discoverer's location: ;in, Let A be the current optimal individual position, and let A be a random vector in the range [-1, 1].

[0079] When the rate of change of fitness value is less than the set threshold for 5 consecutive iterations or the maximum number of iterations is reached, the iteration is terminated, and the position vector of the best individual is decoded into the connection weights and neuron bias terms of the BP neural network.

[0080] In detail, the population is initialized, and the positions and numbers of sparrows in the population can be represented by matrix X: ;

[0081] The fitness value of a sparrow is then expressed as: In the formula: f is the fitness value.

[0082] After initializing the population, the optimal foraging direction coordinates are determined for the top 20% of the sparrows in terms of fitness. Their position updates are shown in the following formula: In the formula: t is the current iteration number; itermax is the maximum iteration number; Let represent the position information of the i-th sparrow in dimension j; a is a random number between (0, 1]; Q is a random number following a normal distribution; L is a 1×n matrix where all elements are 1; The warning value is ST ∈ (0.5, 1], which is the safe value.

[0083] The follower follows the discoverer and performs local optimization within the discoverer's neighborhood solution space. Its position update formula is: In the formula: It is the optimal position for the current discoverer; This represents the worst-case position globally at the t-th iteration. Given a 1×d matrix, where each element is randomly assigned the value 1 or -1, and n represents the population size.

[0084] The vigilant achieves effective global search through position updates, and its position update formula is: In the formula: is the global optimal position in the t-th iteration; β is the step size control parameter, which is a random number that follows a normal distribution with a mean of 0 and a variance of 1; Current fitness; ε represents the fitness of the global optimum and the global worst, respectively; K is a random number between [-1, 1]; ε is a small constant to avoid the case where the denominator is 0.

[0085] In this embodiment, the sample dataset is divided into a training set:test set ratio of 7:3. The model performance is better evaluated by analyzing the training set accuracy, test set accuracy, and fitness curves. Coal body damage precursor information early warning models GA-BP, PSO-BP, and GWO-BP are constructed respectively. According to Table 3, the training set accuracy of the three models (GA-BP, PSO-BP, and SSA-BP) is 88.6%, 91.9%, and 93.1%, respectively, while the test set accuracy is 73.2%, 76.6%, and 82.3%, respectively. The SSA-BP model shows relatively optimal accuracy on both the training and test sets, demonstrating the unique advantage of the Sparrow Search algorithm's powerful global search and generalization capabilities in hyperparameter optimization of the established BP neural network model. Combined with... Figure 6 , Figure 7 and Figure 8 It can be seen that SSA achieved convergence within 5 generations, while GA and PSO achieved convergence within 5 and 20 generations respectively. Figure 9 , Figure 10 and Figure 11 It can be seen that the predicted results of the comprehensive coal damage index trained by the SSA-BP model have a better correlation with the actual results, and the highest degree of agreement. In comprehensive comparison, the coal damage precursor information early warning model based on the SSA-BP neural network is the current optimal model result.

[0086] Table 3 Test results of the three models

[0087]

[0088] In summary, based on the comprehensive evaluation index of coal body damage using acoustic emission multivariate information, a coal body damage precursor information early warning model based on BP neural network was established. The model hyperparameters were optimized using GA, PSO, and SSA, and the performance of the three optimized models was compared and evaluated. The optimal coal body damage precursor information early warning model based on SSA-BP neural network was obtained.

[0089] A comprehensive coal damage assessment index (CDCI) based on acoustic emission multivariate information parameters was proposed, and a coal damage precursor warning model based on a backpropagation neural network was established. This model mainly consists of four parts: extraction of multivariate information features from single acoustic emission signals, fusion of different multivariate information features, verification and output of the comprehensive damage degree discrimination index, and judgment of coal damage status. It can realize real-time inversion analysis of the coal damage process, thereby providing early warning information for coal instability. A feedback mechanism is used to judge the reliability of the acoustic emission multivariate information fusion warning, providing an effective and scientific reference model and approach for realizing early warning discrimination of coal damage precursors in carbon-storing areas.

[0090] The hyperparameters of the model were optimized using three intelligent optimization algorithms: GA, PSO, and SSA. The performance of the three optimized models was compared and evaluated. The results showed that the SSA algorithm has strong global search and generalization capabilities. The SSA-BP model achieved prediction accuracies of 93.1% and 82.3% on the training and test sets, respectively, which is relatively optimal. Moreover, the convergence speed is relatively fast, indicating that the prediction results of the coal body comprehensive damage index under the SSA-BP model training have a better correlation with the actual results and the highest degree of agreement, making it the optimal model result.

[0091] The foregoing illustrative description of the present application and its embodiments is not restrictive and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. The accompanying drawings are only one embodiment of the present application, and the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present application, such designs should fall within the scope of protection of this application. Furthermore, the word "comprising" does not exclude other elements or steps, and the word "a" preceding an element does not exclude the inclusion of "a plurality" of that element. Terms such as "first," "second," etc., are used to indicate names and do not indicate any specific order.

Claims

1. A method for early warning of precursor information of coal damage and destruction based on a BP neural network, characterized in that, The method comprises the following steps: S1, collecting acoustic emission signals in the process of coal body damage, extracting features of the acoustic emission signals, and obtaining multiple parameters including ring count, RA value, AF value, main control frequency and amplitude; preprocessing the multiple parameters to form an acoustic emission event data set with m*5 dimensions, wherein m is the number of acoustic emission events; S2, performing principal component analysis on the acoustic emission event data set, and determining time nodes of the energy storage period and the damage period in the process of coal body damage according to the principal component analysis, wherein the energy storage period is a precursor stage of coal body damage; S3, segmenting the acoustic emission event data set according to the divided time period to calculate a coal body damage comprehensive evaluation index CDCI; S4, constructing a four-layer BP neural network to establish a mapping relationship between the multiple parameters and the damage grade; S5, using the sparrow search algorithm to iteratively optimize the connection weights and neuron bias of the BP neural network, so that the mean square error between the CDCI prediction value output by the BP neural network and the CDCI value calculated in step S3 is minimized; S6, inputting the real-time collected acoustic emission multiple parameters into the optimized BP neural network to predict the coal body damage comprehensive evaluation index and the damage grade at the current time, so as to perform early warning on the precursor information of coal body damage; S2, performing principal component analysis on the acoustic emission event data set, and determining time nodes of the energy storage period and the damage period in the process of coal body damage according to the principal component analysis, comprising: The time sequence change rate of each principal component score is calculated, and the moment when the time sequence change rate exceeds a threshold for the first time is marked as the starting point of the energy storage period ; The moment when the principal component score reaches the maximum value is marked as the starting point of the damage period ; Time period defined as the precursor stage of coal body damage and destruction; The time sequence change rate of each principal component score is calculated, and the moment when the time sequence change rate exceeds a threshold for the first time is marked as the starting point of the energy storage period , comprising: The m x 5-dimensional acoustic emission event dataset is standardized to obtain a standardized matrix X: ; wherein, represents an element in the i-th row and the j-th column of the standardized matrix X; is a j-th parameter value of the i-th acoustic emission event; and are a mean value and a standard deviation of the j-th parameter in all m acoustic emission events, respectively. computing the covariance matrix of the standardized matrix X where C is a 5x5 symmetric matrix, denotes the transpose matrix of X; Eigenvalue decomposition of the covariance matrix C yields 5 eigenvalues. and the corresponding feature vectors ,in, This represents the k-th eigenvalue. Indicates and The corresponding unit eigenvector; computing a timing score of the first principal component wherein is the jth component of the first feature vector denotes the time instant at which the ith acoustic emission event occurred, denotes the first principal component score at the time instant​​ a time series rate of change of the first principal component score: ; wherein, represents a time series rate of change of the first principal component score at a time t, represents a time of occurrence of the i-1th acoustic emission event; marking the time instant when the rate of temporal variation first exceeds the threshold as the start of the accumulation phase ; The first principal component score is calculated as follows: The time at which the maximum is reached is marked as the start of the breakdown phase ; Coal body damage comprehensive evaluation index of computed acoustic emission event data set : ; wherein, is the i th parameter value at time t, is is the average value of the i th parameter in the period, is is the maximum value of the i th parameter in the period, is the weight of the i th parameter determined based on the principal component contribution rate; The weight of the i-th parameter determined based on the principal component contribution rate is determined by the following equation: Based on the standardized matrix X in step S2, a correlation coefficient matrix R between parameters is constructed, the correlation coefficient matrix R is a 5x5 symmetric matrix, wherein elements represent the correlation coefficient between the ith parameter and the jth parameter; Perform eigenvalue decomposition on the correlation coefficient matrix R to obtain the eigenvalues. and the corresponding feature vectors ; calculating the contribution rate of each principal component based on the eigenvalue, and determining the number of principal components according to the cumulative contribution rate, wherein the threshold of the cumulative contribution rate is set to 80%; constructing a factor loading matrix using the eigenvalue and the eigenvector, wherein the factor loading matrix reflects the loading degree of each parameter on the principal component; The coefficients of each parameter are calculated based on the factor load of the first principal component, and the weights of each parameter are obtained through normalization processing where i = 1, 2, 3, 4, 5 respectively correspond to the weights of the ring count, the RA value, the AF value, the master control frequency and the amplitude value.

2. The coal body damage precursor information early warning method based on the BP neural network according to claim 1, wherein: S4, constructing a four-layer BP neural network, comprising: The input layer is constructed, and 5 input nodes are set to receive the normalized ring count, RA value, AF value, main control frequency and amplitude value respectively, and the time node determined in step S2 is also set 、 converted into a time sequence label [0, 1, 2], which corresponds to 、 、 three time periods respectively; Construct a feature fusion layer, set n hidden nodes, and each hidden node passes through an activation function Processing the input layer data, wherein, is a connection weight, is a bias term; The characteristic output layer is constructed, and the connection weight is adjusted according to the coal mass damage comprehensive evaluation index CDCI and the coal mass stress value σ(t) at the current moment and the bias term , and the CDCI prediction value is output according to the adjusted connection weight and the bias term ; constructing a damage decision layer to divide the damage grade according to the CDCI prediction value, wherein the damage grade includes no damage, weak damage and strong damage.

3. The coal body damage precursor information early warning method based on the BP neural network according to claim 2, wherein: According to the coal body damage comprehensive evaluation index And the coal body stress value at the current moment Adjust the connection weight And the bias term , comprising: According to the calculation formula in step S3 , the reference parameter corresponding to the current time t is determined: when , , the average value of the ith parameter in the period [0, t] is taken; when , , the average value of the ith parameter in the period [t, t+T] is taken , the average value of the ith parameter in the period [t, t+T] is taken; when , , the average value of the ith parameter in the period [t, t+T] is taken , the average value of the ith parameter in the period [t, t+T] is taken, , the maximum value of the ith parameter in the period [t, t+T] is taken , the maximum value of the ith parameter in the period [t, t+T] is taken The output vector of the fusion layer is passed through The computational formula is mapped to the CDCI prediction value obtaining a coal stress value at a current time , calculating a normalized stress level: ; When and CDCI predicted value < 0.33, or when and CDCI predicted value < 0.66, or |, it is determined that the error is large when the determination error is large, the CDCI prediction value is fed back to the fusion layer, and the connection weight and the bias item are adjusted through the back propagation algorithm when the determination error is small, the CDCI prediction value is transmitted to the decision layer.

4. The coal body damage precursor information early warning method based on the BP neural network according to claim 3, wherein: dividing the damage grade according to the CDCI prediction value, comprising: When time is divided into non-invasive; When weak damage; When strong impairment.

5. The coal body damage precursor information early warning method based on the BP neural network according to claim 2, wherein: S5, using the sparrow search algorithm to optimize the weight and threshold of the BP neural network, comprising: The connection weight in the four-layer BP neural network constructed in step S4 The bias term of the feature fusion layer Encode as sparrow individual position vector, initialize N sparrow individuals to form a population; defining the fitness function as the mean square error between the CDCI prediction value output by the BP neural network and the CDCI real value calculated in step S3: F=(1 / M)×Σ(CDCI prediction(t)-CDCI real(t))² wherein M is the number of training samples, and CDCI real(t) is calculated according to the formula in step S3. In the iteration process, the population is divided into discoverers and followers according to the fitness value: The discoverers perform global search according to the current position, and update the position to find the optimal solution; The followers perform local search according to the position of the discoverers, and search within the preset range of the optimal solution; The population position is updated through multiple iterations, and the iteration is terminated when a preset convergence condition is met, and a position vector of an optimal individual is encoded as an optimized connection weight and a bias term of the feature fusion layer .

6. A BP neural network-based coal mass damage and failure precursor information early warning system for implementing the method of any one of claims 1 to 5, characterized in that, It comprises: A data acquisition module acquires acoustic emission signals during the coal failure process, extracts features from the acoustic emission signals to obtain multiple parameters, the multiple parameters include ring count, RA value, AF value, main control frequency and amplitude, and the multiple parameters are preprocessed to form an acoustic emission event data set with m 5 dimensions, wherein m is the number of acoustic emission events; The phase division module performs principal component analysis on the acoustic emission event data set, calculates a time sequence change rate of a first principal component score, marks a time when the time sequence change rate first exceeds a threshold value as a starting point of the accumulation period , marks a time when the principal component score reaches a maximum value as a starting point of the damage period , and defines a time period as a precursor stage of coal body damage and destruction. An index calculation module calculates the coal damage comprehensive evaluation index CDCI of the acoustic emission event data set in segments according to the divided time period; The BP neural network construction module constructs a four-layer BP neural network, which comprises an input layer receiving normalized multi-element parameters and time sequence labels, a feature fusion layer with n hidden nodes, a CDCI and coal body stress value determination feature output layer, and a damage decision layer for damage grade division; A parameter optimization module uses the sparrow search algorithm to iteratively optimize the connection weight and neuron bias of the BP neural network, encodes the network parameters as a sparrow individual position vector, takes the mean square error between the CDCI prediction value output by the BP neural network and the CDCI true value calculated by the damage evaluation index calculation module as the fitness function, and finds the optimal parameter configuration through the discoverer-follower collaborative search mechanism; A real-time warning module inputs the real-time collected acoustic emission multiple parameters into the optimized BP neural network to predict the coal damage comprehensive evaluation index and damage grade at the current time.

Images