Mine disaster prediction and emergency response method based on artificial intelligence

Abstract

The invention relates to the technical field of data processing, in particular to a mine disaster prediction and emergency response method based on artificial intelligence, and the method comprises the steps: collecting original mine data; generating simulated mine data based on the original mine data; inputting the training data into a mine disaster prediction unit; acquiring new original mine data through a mine data acquisition unit, and performing feature extraction on the new original mine data by adopting the trained feature extraction model to obtain the new original mine data after feature extraction; and classifying the new original mine data after feature extraction by adopting the trained classification model to obtain a safety state corresponding to the new original mine data, and outputting a corresponding emergency response measure after the emergency response unit receives the safety state. An existing mine safety management system based on artificial intelligence is low in prediction accuracy, and the method is high in prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a mine disaster prediction and emergency response method based on artificial intelligence. Background Art

[0002] Mine disasters refer to a series of catastrophic events that endanger mine safety, production, and the lives and health of underground personnel caused by natural factors (such as geological conditions, groundwater, etc.) or human factors (such as improper mining methods, equipment failures, illegal operations, etc.) during the construction and mining of mines. They mainly include gas disasters, roof disasters, water inrush disasters, fire disasters, and dust disasters. Therefore, in order to ensure safe production in mines, it is necessary to predict possible disasters in advance and formulate corresponding emergency response measures.

[0003] With the continuous development of artificial intelligence, mine safety management systems based on artificial intelligence are often used in mines to predict disasters. The existing mine safety management systems based on artificial intelligence mainly extract features from mine data through deep learning models and then classify mine data through classification models to know their corresponding safety status. However, existing deep learning models may face problems such as gradient vanishing, gradient explosion, or falling into local optimal solutions, resulting in low feature extraction efficiency and affecting disaster prediction performance. Existing classification models are not able to adapt to dynamic changes in mine data, especially when facing new types of disasters or changes in data distribution, the prediction accuracy will drop significantly.

[0004] Therefore, the existing artificial intelligence-based mine safety management system has the problem of low prediction accuracy. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a mine disaster prediction and emergency response method based on artificial intelligence with high prediction accuracy.

[0006] In order to solve the above technical problems, the present invention provides an artificial intelligence-based mine disaster prediction and emergency response method, which includes:

[0007] S1. Collecting original mine data through a mine data collection unit, manually marking the safety status corresponding to the original mine data, and then storing it;

[0008] The safety status includes safety, potential gas disaster, potential roof disaster, potential water inrush disaster, potential fire disaster and potential dust disaster;

[0009] S2. Generate simulated mine data based on the original mine data by a mine data enhancement unit, and use the original mine data and the simulated mine data together as training data;

[0010] S3, inputting the training data into a mine disaster prediction unit to train a feature extraction model of the mine disaster prediction unit and a classification model of the mine disaster prediction unit;

[0011] The feature extraction model for training the mine disaster prediction unit adopts a neural network algorithm based on quantum potential field optimization;

[0012] The classification model for training the mine disaster prediction unit adopts an extreme learning machine algorithm based on a dynamic attractor;

[0013] S4. New original mine data is collected through the mine data collection unit, and features are extracted from the new original mine data using the trained feature extraction model to obtain new original mine data after feature extraction. The trained classification model is used to classify the new original mine data after feature extraction to obtain a safety status corresponding to the new original mine data. After receiving the safety status, the emergency response unit outputs corresponding emergency response measures.

[0014] As a further improvement of the present invention: the generation of simulated mine data in S2 adopts a bilinear interpolation algorithm based on a dynamic constraint set, and the generation of simulated mine data includes:

[0015] S201, initializing constraint set;

[0016] S202, dynamically adjusting the constraint set according to changes in environmental parameters during sample expansion, including:

[0017]

[0018] In the formula, is the updated constraint set, C c is the initial value of the constraint set, λ c is the constraint set adjustment factor, M ce is the number of risk factors, f c () is the adjustment function based on the risk assessment model, R j,c is the real-time data of the mine for the jth risk factor, Δt c is the time interval for sample collection;

[0019] S203, using a bilinear interpolation algorithm to fill in missing or incomplete mine data according to the current dynamic constraint set, expressed as:

[0020] V k,c =(1-a c )(1-b c )V i,c +a c (1-bc )V i+1,c +(1-a c )b c V j,c +a c b c V j+1,c ,

[0021] Where V k,c is the predicted value of the point to be interpolated, a c is the first interpolation coefficient adjusted according to the dynamic constraint, b c is the second interpolation coefficient adjusted according to the dynamic constraint, V i,c is the existing i-th data point,

[0022] V i+1,c is the existing i+1th data point, V j,c is the jth data point, V j+1,c is the existing j+1th data point;

[0023] S204, using an adaptive noise suppression mechanism to effectively identify and process the noise in the mine data, the calculation formula is:

[0024]

[0025] In the formula, h c () is the noise identification function, V i,c is the existing i-th data point, a c is the first interpolation coefficient adjusted according to the dynamic constraint, b c is the second interpolation coefficient adjusted according to the dynamic constraint.

[0026] Preferably, the initialization constraint set in S201 is expressed as:

[0027]

[0028] In the formula, C c is the initial value of the constraint set, N ce is the total number of samples in the mine training dataset before sample generation, d i,c is the attribute value of the i-th data point collected from the mine data set, μ c is the mean of the attributes of the ith data point collected from the mine data set, σ c is the standard deviation of the attribute of the ith data point collected from the mine dataset.

[0029] Preferably, the f in S202 c (R j,c ,Δt c ) is calculated as:

[0030]

[0031] In the formula, Δt c is the sample collection time interval, μ t,c is the mean of the sample collection time deviation, σ t,c is the standard deviation of sample collection time deviation, k c is the slope control parameter of the response curve, R j,c is the real-time data of the mine with the jth risk factor, θ c is the activation threshold for adjustment.

[0032] As a further improvement of the present invention: the feature extraction model for training the mine disaster prediction unit in S3 includes:

[0033] S311, initializing the neural network structure;

[0034] S312, in each iteration, dynamically adjusting the neural network parameters according to the feedback generated by the quantum potential field based on the mine data, the neural network parameters are the neural network weights, and the update amount of the neural network weights ΔW p The calculation formula is:

[0035]

[0036] Where W p represents the weight of the neural network, η p is the learning rate of the neural network, is the gradient of the neural network's loss function with respect to the weights, is the cross entropy loss function, D is the mine training data set, ξ p is the influence intensity of quantum potential field on parameter adjustment, Ψ p () is the quantum potential field function;

[0037] S313, dynamically adjusting the neural network structure according to error feedback during the training process;

[0038] S314, after a preset number of iterations, the feature extraction capability of the neural network is evaluated. The evaluation method of the feature extraction efficiency is expressed as:

[0039]

[0040] In the formula, Γ p Indicates the accuracy of feature extraction, V fe is the preset validation mine data set, v se The data in the preset verification mine data set, is the data point v se The direction angle of the true eigenvector of is the data point v se The direction angle of the feature vector predicted by the model, <,> represents the dot product of the vector, and ∥∥ is the L2 norm;

[0041] S315, repeating iterations S312 to S314 until a preset condition for stopping iteration is met, indicating that the model training is completed.

[0042] Preferably, the initializing the neural network structure in S311 includes:

[0043]

[0044] In the formula, represents the initial weight matrix of the neural network, σ p is the scale factor of the neural network parameters, randn() means generating random numbers, and randn(0,0.1) means generating random numbers in the interval (0,0.1).

[0045] Preferably, the Ψ in S312 p (W p ) is calculated as:

[0046]

[0047] Where K is the total number of quantum states, ζ k,p is the influence intensity of the kth quantum state, μ k,p is the center position of the kth quantum state, σ k,p is the dispersion of the quantum state.

[0048] As a further improvement of the present invention: the classification model for training the mine disaster prediction unit in S3 includes:

[0049] S321, initializing the weights and biases of the extreme learning machine;

[0050] S322, the training data after feature extraction is forward propagated through the initialized weight matrix of the extreme learning machine. During the forward propagation process, the output calculation formula of the hidden layer of the extreme learning machine is:

[0051] H u =Sig(W u X u +b u ),

[0052] In the formula, H u is the output of the hidden layer of the extreme learning machine, Sig() is the Sigmoid activation function, and W u is the weight matrix of the extreme learning machine, b uis the bias vector of the extreme learning machine;

[0053] S323, using Moore-Penrose pseudo-inverse to minimize the output error, the output layer weight of the extreme learning machine is solved, which is expressed as:

[0054]

[0055] In the formula, β u is the output layer weight of the extreme learning machine, H u is the hidden layer output of the extreme learning machine; H u + is the Moore-Penrose pseudo-inverse of the hidden layer output of the extreme learning machine; T u is the target output of the extreme learning machine;

[0056] S324. After each batch or a preset number of iterations, the weights from the hidden layer to the output layer are adjusted based on the dynamic attractor strategy. The weight adjustment method based on the dynamic attractor is expressed as:

[0057]

[0058] In the formula, is the weight of the extreme learning machine at the t+1th iteration, is the weight of the extreme learning machine at the tth iteration, η ur (t) is the learning rate of the extreme learning machine at the tth iteration, R u () is the regularization weight of the hidden layer output of the extreme learning machine, H u is the hidden layer output of the extreme learning machine, Attr() is the dynamic attractor function defined based on the current hidden layer output and error of the extreme learning machine, Error u is the error between the current model output and the target output;

[0059] S325, adopting a high-dimensional feature regularization mechanism based on energy mapping, by constructing an energy mapping function, regularizing the redundancy and instability that may appear in the high-dimensional feature space, while retaining important high-energy features, and calculating the regularization weights of the features output by the hidden layer of the extreme learning machine;

[0060] S326, repeating iterations S322 to S325 until a preset condition for stopping iteration is met, indicating that the model training is completed.

[0061] Preferably, the Moore-Penrose pseudo-inverse of the hidden layer output of the extreme learning machine in S323 is The calculation formula is:

[0062]

[0063] In the formula, is the transpose of the hidden layer output of the extreme learning machine, λ mr is the regularization parameter of the extreme learning machine, and I is the identity matrix.

[0064] Preferably, the Attr(H u ,Error u ) is calculated as:

[0065]

[0066] In the formula, δ u To adjust the parameters, tanh() is the hyperbolic tangent function, Error u is the error between the current model output and the target output.

[0067] The beneficial effects of the present invention are as follows: The artificial intelligence-based mine disaster prediction and emergency response method provided by the present invention has high prediction accuracy.

[0068] The method generates simulated mine data based on the original mine data through a mine data enhancement unit, and uses the original mine data and the simulated mine data as training data, thereby enhancing the diversity and authenticity of the training data, being able to simulate different disaster scenarios, and alleviating the problem of poor generalization ability of each model caused by insufficient training data. The method trains a feature extraction model through a neural network algorithm based on quantum potential field optimization, dynamically adjusts the neural network parameters through a quantum potential field function, so that the weight update not only depends on the gradient, but also combines the distribution characteristics of the quantum state, avoids the problem of gradient disappearance or gradient explosion, improves the feature extraction ability and training efficiency, and enhances the adaptability and stability of the model to complex mine data. The method trains the classification model of the mine disaster prediction unit through an extreme learning machine algorithm based on a dynamic attractor, uses a dynamic attractor to adjust the weight update strategy in real time according to the input data stream, and uses an energy mapping regularization mechanism to optimize the high-dimensional feature space, avoids the interference of redundant features, improves the disaster classification accuracy of the model, and enhances its adaptability and robustness to dynamic mine data. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a flowchart of the present invention. DETAILED DESCRIPTION

[0070] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0071] The features, operations or characteristics described in the present invention may be combined in any appropriate manner to form various embodiments. At the same time, the steps or actions described in the present invention may also be interchanged or adjusted in a manner that is obvious to those skilled in the art. Therefore, the various sequences in the present invention and the accompanying drawings, such as "S1", are only for the purpose of clearly describing a certain embodiment and are not intended to be a necessary sequence, unless otherwise specified that a certain sequence must be followed.

[0072] The artificial intelligence-based mine disaster prediction and emergency response system provided by the present invention includes a mine data acquisition unit for collecting and storing actual mine data, a mine data enhancement unit for generating simulated mine data, a mine disaster prediction unit for extracting features from actual mine data and classifying the actual mine data to obtain a safety status corresponding to the actual mine data, and an emergency response unit for receiving the safety status corresponding to the actual mine data output by the mine disaster prediction unit and outputting emergency response measures corresponding to the safety status.

[0073] like Figure 1 As shown, the mine disaster prediction and emergency response method based on artificial intelligence provided by the present invention includes:

[0074] S1. Collect original mine data from sources such as mine ventilation system, gas monitoring equipment, environmental monitoring system, equipment operation status monitoring, and historical disaster and accident records through the mine data collection unit, and manually mark the safety status corresponding to the original mine data and store it in JSON format;

[0075] The original mine data include gas concentration Ra, ventilation speed Da, temperature Ta, pressure Pa, humidity La, equipment operation status Sa, historical disaster frequency Ha, vibration index Va, CO concentration Ca and power consumption Ea; the record form is shown in Table 1 for the five pieces of original mine data. The mine data acquisition unit includes different types of sensors or acquisition equipment that can collect data with different attributes.

[0076] Table 1

[0077]

[0078] Safety status includes safety, potential gas disaster, potential roof disaster, potential water inrush disaster, potential fire disaster and potential dust disaster;

[0079] S2. Generate simulated mine data based on the original mine data through the mine data enhancement unit, and use the original mine data and the simulated mine data together as training data; Generate simulated mine data through this method, so that the total amount of training data increases. For example, if the original mine data is 800, and the data is expanded to obtain 200 simulated mine data, then the training data includes 1000 mine data.

[0080] The bilinear interpolation algorithm based on dynamic constraint set is used to generate simulated mine data; the bilinear interpolation algorithm based on dynamic constraint set means that when using the bilinear interpolation algorithm for sample interpolation generation, the changes in environmental parameters are monitored and the constraint set is dynamically adjusted according to these changes. The constraint set refers to a set of rules, conditions or ranges used to limit or guide certain operations during mine data processing or algorithm execution. The constraint set covers the basic distribution of mine data, the thresholds of key parameters, and specific patterns identified in historical disaster data. For example, if a sudden increase in gas concentration is detected, the interpolation constraint will be adjusted to reflect the increased possibility of such a dangerous state. This is to use a dynamic constraint set to dynamically adjust the constraints of the enhanced dangerous state in the interpolation constraint conditions according to the real-time changes in the mine environmental parameters.

[0081] The simulated mine data generated in S2 include:

[0082] S201, initialize the constraint set according to the statistical characteristics of the original mine training data set, expressed as:

[0083]

[0084] In the formula, C c is the initial value of the constraint set, N ce is the total number of samples in the mine training dataset before sample generation, d i,c is the attribute value of the i-th data point collected from the mine data set, μ c is the mean of the attributes of the ith data point collected from the mine data set, σ c is the standard deviation of the attribute of the ith data point collected from the mine dataset.

[0085] S202: Dynamically adjust the constraint set according to the changes in environmental parameters during the sample expansion process. The adjustment method is:

[0086]

[0087] In the formula, is the updated constraint set, C c is the initial value of the constraint set, λ c is the constraint set adjustment factor, M ce is the number of risk factors, f c () is the adjustment function based on the risk assessment model, R j,c is the real-time data of the mine for the jth risk factor, Δt c is the time interval for sample collection;

[0088] The adjustment function based on the risk assessment model is implemented based on the response curve of the multivariate risk factor, f c (Rj,c ,Δt c ) is calculated as:

[0089]

[0090] In the formula, Δt c is the sample collection time interval; μ t,c is the mean of the sample collection time deviation, σ t,c is the standard deviation of sample collection time deviation, which is used to adjust the impact weight of risk factors within the time window; k c is the slope control parameter of the response curve, k c Set to 2; R j,c is the real-time data of the mine with the jth risk factor, θ c is the activation threshold for adjustment, θ c Set to 0.1.

[0091] S203, using a bilinear interpolation algorithm to fill in missing or incomplete mine data according to the current dynamic constraint set to ensure the accuracy and consistency of the sample after expansion, expressed as:

[0092] V k,c =(1-a c )(1-b c )V i,c +a c (1-b c )V i+1,c +(1-a c )b c V j,c +a c b c V j+1,c ,

[0093] Where V k,c is the predicted value of the point to be interpolated, a c is the first interpolation coefficient adjusted according to the dynamic constraint, b c is the second interpolation coefficient adjusted according to the dynamic constraint, V i,c is the existing i-th data point,

[0094] V i+1,c is the existing i+1th data point, V j,c is the jth data point, V j+1,c is the existing j+1th data point;

[0095] The first interpolation coefficient a obtained by adjusting the dynamic constraint c and the second interpolation coefficient b c According to the dynamically adjusted constraint set and prediction error minimization strategy, a cWith b c The calculation formula is:

[0096]

[0097] Where N ce is the total number of samples in the mine training dataset before sample generation, w n,c According to the data point (x n,c ,y n,c ) to the distance function of the point to be predicted; n,c is the first coordinate of the nth interpolation reference point; y n,c is the second coordinate of the nth interpolation reference point;

[0098] According to the data point (x n,c ,y n,c ) to the distance function w from the point to be predicted n,c Used to enhance the local accuracy of interpolation, the calculation formula is:

[0099]

[0100] In the formula, is the updated constraint set, x k,c is the first coordinate of the kth interpolation reference point, y k,c is the second coordinate of the kth interpolation reference point, ∈ ces is a small positive number used to prevent division by zero, ∈ ces Set to 0.001.

[0101] S204, using an adaptive noise suppression mechanism to effectively identify and process the noise in the mine data, and reducing the impact of noise by adaptively adjusting the weight of the data point. Specifically, the noise identification function estimates the noise probability based on the historical statistical information of the data point and the current environmental parameters, and the calculation formula is:

[0102]

[0103] In the formula, h c () is the noise identification function, V i,c is the existing i-th data point, a c is the first interpolation coefficient adjusted according to the dynamic constraint, b c is a second interpolation coefficient obtained by adjusting according to the dynamic constraint;

[0104] Update the distance function to ensure that the influence of data points with large noise in the interpolation process is effectively reduced to suppress the influence of noise. The update method is expressed as:

[0105]

[0106] In the formula, is the updated distance function, w n,c According to the data point (x n,c ,y n,c ) to the distance function of the point to be predicted, V n,c is the nth data point, ∈ crw A small constant used to avoid division by zero errors; its value is 0.0001.

[0107] Adaptive noise suppression weights the adjusted data values The calculation formula is:

[0108]

[0109] Where N ce is the total number of samples in the mine training dataset before sample generation, is the updated distance function, V n,c is the nth data point already available.

[0110] S205, the expanded mine data set is post-processed to ensure that the quality of the mine data set meets the requirements of subsequent model training, which is expressed as:

[0111]

[0112] In the formula, is the final data value, g c () is the post-processing function, V k,c is the predicted value of the point to be interpolated, α cry is the parameter for the first adjusted post-processing intensity, β cry is the parameter of the second post-processing intensity adjustment, ⊙ is the element multiplication, The data value adjusted for adaptive noise suppression weights.

[0113] Considering the consistency of mine data and outlier processing, the post-processing function is designed to smooth the mine data and remove outliers. The calculation formula is:

[0114] g c (V k,c ,α cry ,β cry )=α cry tanh(β cry ·V k,c )+γ cry ·V k,c ,

[0115] Where tanh() is the hyperbolic tangent function, γ cry Parameters for adjusting the fidelity of mine data.

[0116] S3, inputting the training data into the mine disaster prediction unit, training the feature extraction model of the mine disaster prediction unit and the classification model of the mine disaster prediction unit;

[0117] The feature extraction model is built based on a fully connected neural network algorithm. The feature extraction model for training the mine disaster prediction unit adopts a neural network algorithm based on quantum potential field optimization. The neural network algorithm based on quantum potential field optimization uses the quantum state characteristics of the mine data in each iteration of the neural network. The parameters of the neural network are adaptively and dynamically adjusted through the quantum potential field function, allowing the model to self-optimize during the training process and adapt to different mine data distributions and characteristics. The feature extraction model for training the mine disaster prediction unit in S3 includes:

[0118] S311, initializing the neural network structure; setting the architecture of the fully connected neural network to include an input layer, four hidden layers and an output layer, and initializing the neural network structure in the following manner:

[0119]

[0120] In the formula, is the initial weight matrix of the neural network; σ p is the scaling factor of the neural network parameters, σ p Set to 0.5; randn() generates random numbers, and randn(0,0.1) generates random numbers in the interval (0,0.1).

[0121] S312. In each iteration, the neural network parameters are dynamically adjusted according to the feedback generated by the quantum potential field based on the mine data, so that the weight update is no longer a single path, but a superposition of probability distribution, which greatly increases the ability of the model to adapt to complex mine data structures and allows the model to self-optimize during the training process to adapt to different mine data distributions and characteristics. The update amount of the neural network weight is ΔW p The calculation formula is:

[0122]

[0123] Where W p is the weight of the neural network, η p is the learning rate of the neural network, is the gradient of the neural network's loss function with respect to the weights, is the cross entropy loss function, is the training data set; p is the influence intensity of the quantum potential field on parameter adjustment, ξ p Set to 0.2; Ψ p () is the quantum potential field function;

[0124] The quantum potential field function is used to simulate the potential changes of mine data characteristics under quantum physics environment, revealing the specific role of mine data in weight adjustment, Ψ p (W p ) is calculated as:

[0125]

[0126] Where K is the total number of quantum states, representing the potential diversity of the model in the weight space; ζ k,p is the influence intensity of the kth quantum state, representing the contribution of this quantum state in the weight adjustment; μ k,p is the center position of the kth quantum state, representing the ideal value of the weight vector in this state; σ k,p is the dispersion of the quantum state, which characterizes the variation range of the weight vector in this quantum state.

[0127] S313. According to the error feedback in the training process, the neural network structure is dynamically adjusted. For example, when the structural scoring function is less than the preset threshold, it indicates that the feature extraction of the domain is insufficient, and more nodes or layers need to be added to enhance the representation ability of the model. The calculation formula of the structural scoring function is:

[0128]

[0129] In the formula, S p () is the structural scoring function, which is used to evaluate the effectiveness of the current neural network structure; m per is the number of connections between neurons in the neural network; H per () is the Heaviside step function, which is used to activate or inhibit the connections in the neural network; λ i,p () is the importance score of the i-th connection, calculated by the normalized value of the connection weight;

[0130] S314. After a preset number of iterations, the feature extraction capability of the neural network is evaluated to ensure that the extracted features can effectively represent the key information in the input mine data and provide support for disaster prediction. The evaluation method of the feature extraction efficiency is expressed as:

[0131]

[0132] In the formula, Γ p is the accuracy of feature extraction, V fe is the preset validation mine data set, v se The data in the preset verification mine data set, is the data point v se The direction angle of the true eigenvector of is the data point v seThe direction angle of the feature vector predicted by the model, <,> is the dot product of the vector, and ∥∥ is the L2 norm;

[0133] S315, repeating iterations S312 to S314 until a preset stop iteration condition is met, indicating that the model training is completed. The preset stop iteration condition is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0134] The classification model is constructed based on the extreme learning machine algorithm. The classification model for training the mine disaster prediction unit adopts the extreme learning machine algorithm based on dynamic attractors. The dynamic attractor is a method of adjusting the weight update strategy in the learning process based on the data flow. The extreme learning machine algorithm based on dynamic attractors is used as a classification algorithm. When the weight of the extreme learning machine is updated, the dynamic attractor dynamically adjusts the weight of the hidden layer according to the changes in the input mine data, so that the model can more flexibly adapt to different types of mine data distribution, thereby improving the accuracy and robustness of the model's prediction of mine disaster types. The classification model for training the mine disaster prediction unit in S3 includes:

[0135] S321, initialize the weights and biases of the extreme learning machine; the initialization method is expressed as:

[0136] W u ←randn(0,0.1),

[0137] b u ←0,

[0138] Where W u is the weight matrix of the extreme learning machine, ← is the parameter update operation, randn() is to generate random numbers, randn(0,0.1) is to generate random numbers in the interval (0,0.1), b u is the bias vector of the extreme learning machine.

[0139] S322, the training data after feature extraction is forward propagated through the initialized weight matrix of the extreme learning machine, converted into the feature representation of the hidden layer, and the feature map is quickly generated. During the forward propagation process, the output calculation formula of the hidden layer of the extreme learning machine is:

[0140] H u =Sig(W u X u +b u ),

[0141] In the formula, H u is the output of the hidden layer of the extreme learning machine, Sig() is the Sigmoid activation function, and W u is the weight matrix of the extreme learning machine, b u is the bias vector of the extreme learning machine;

[0142] S323, using Moore-Penrose pseudo-inverse to minimize the output error, the output layer weight of the extreme learning machine is solved, which is expressed as:

[0143]

[0144] In the formula, β u is the output layer weight of the extreme learning machine, H u is the hidden layer output of the extreme learning machine; is the Moore-Penrose pseudo-inverse of the hidden layer output of the extreme learning machine; T u is the target output of the extreme learning machine;

[0145] Moore-Penrose pseudo-inverse of the hidden layer output of the extreme learning machine The calculation formula is:

[0146]

[0147] In the formula, is the transpose of the hidden layer output of the extreme learning machine; mr is the regularization parameter of the extreme learning machine, which is used to avoid the singularity problem in the inverse matrix calculation, λ mr Set to 0.2; I is the identity matrix.

[0148] S324. After each batch or a preset number of iterations, the weights from the hidden layer to the output layer are adjusted based on the dynamic attractor strategy to optimize the model's adaptability to new mine data and the accuracy of prediction. The weight adjustment method based on the dynamic attractor is expressed as:

[0149]

[0150] In the formula, is the weight of the extreme learning machine at the t+1th iteration, is the weight of the extreme learning machine at the tth iteration; η ur (t) is the learning rate of the extreme learning machine at the tth iteration, which is set by adaptive adjustment; R u () is the regularization weight of the hidden layer output of the extreme learning machine, H u is the hidden layer output of the extreme learning machine, Attr() is the dynamic attractor function defined based on the current hidden layer output and error of the extreme learning machine, which is used to adjust the weight of the extreme learning machine; Error u is the error between the current model output and the target output, Error u =T u -H u βu ;

[0151] Attr(H u ,Error u ) is calculated as:

[0152]

[0153] In the formula, δ u To adjust the parameters, tanh() is the hyperbolic tangent function, Error u is the error between the current model output and the target output;

[0154] The dynamic feedback mechanism of adaptive learning rate adjustment is used to dynamically adjust the learning rate by monitoring the rate of change of training error, thereby optimizing the learning process and ensuring that the model can adapt quickly and accurately when facing changes in mine data. The adjustment of the adaptive learning rate based on error feedback is defined based on the error of the current iteration and the error of the previous iteration. The calculation formula is:

[0155] η ur (t) = η ur (t-1)·exp(-γ yr ΔE(t)),

[0156] Where η ur (t) is the learning rate of the extreme learning machine at the tth iteration, η ur (t-1) is the learning rate of the extreme learning machine at the t-1th iteration; γ yr is the learning rate adjustment coefficient of the preset extreme learning machine, which controls the sensitivity of the learning rate adjustment. yr Set to 2; ΔE(t) is the absolute difference in error between two consecutive iterations, calculated as ΔE(t) = |E(t)-E(t-1)| is the absolute difference in error between two consecutive iterations of the extreme learning machine, and E(t-1) is the error of the extreme learning machine at the t-1th iteration.

[0157] S325. A high-dimensional feature regularization mechanism based on energy mapping is adopted. By constructing an energy mapping function, redundancy and instability that may appear in the high-dimensional feature space are regularized, while important high-energy features are retained. Regularization weights of features output by the hidden layer of the extreme learning machine are calculated. Specifically, an energy mapping function is constructed for each feature output by the hidden layer of the extreme learning machine, which is expressed as:

[0158]

[0159] In the formula, QE u () is the energy mapping function, QE u (H ui) is the energy proportion of the i-th feature output by the hidden layer of the extreme learning machine, H ui is the i-th feature output by the hidden layer of the extreme learning machine, H u is the hidden layer output of the extreme learning machine, ε uae is a small positive value used to prevent the denominator from being zero;

[0160] Based on the result of energy mapping, regularization weights are applied to the features. When the feature energy is high, the regularization weight of the hidden layer output of the extreme learning machine is close to 1; when the feature energy is low, the regularization weight of the hidden layer output of the extreme learning machine is close to 0, which weakens its influence on the model. The calculation formula is:

[0161] R u (H ui )=exp(-λ u ·(1-QE u (H ui ))),

[0162] In the formula, R u (H ui ) is the regularization weight of the i-th feature output by the hidden layer of the extreme learning machine, λ u It is the adjustment factor of the extreme learning machine, which controls the weight difference between high-energy features and low-energy features.

[0163] The regularization weight of the hidden layer output of the extreme learning machine represents the overall regularization effect of the calculation model, and the calculation formula is:

[0164]

[0165] In the formula, R u (H u ) is the regularization weight of the hidden layer output of the extreme learning machine, d hidden is the number of neurons in the hidden layer of the extreme learning machine.

[0166] S326, repeating iterations S322 to S325 until a preset stop iteration condition is met, indicating that the model training is completed. The preset stop iteration condition is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0167] S4. New original mine data is collected through the mine data collection unit, and features are extracted from the new original mine data using the trained feature extraction model to obtain new original mine data after feature extraction. The new original mine data after feature extraction is classified using the trained classification model to obtain the safety status corresponding to the new original mine data. After receiving the safety status, the emergency response unit outputs the corresponding emergency response measures.

[0168] When the safety status is safe, the emergency response measure is to maintain the original status. When the safety status is potential gas disaster, potential roof disaster, potential water inrush disaster, potential fire disaster and potential dust disaster, the emergency response measure is accident alarm and giving suggestions on personnel evacuation and preparatory rescue, accident site disposal, medical rescue and logistics support. The suggestions come from the mine's past disaster response experience or expert suggestions.

[0169] Alarm and communication include accident alarm and communication guarantee. Accident alarm is the first time for the emergency response unit to report to the mine dispatch room through the underground emergency alarm device (such as sound and light alarm, special alarm phone, etc.). The alarm content includes the type of accident, the location of the accident and the approximate scale of the accident. For example, when there is a potential gas disaster, the alarm content is: "There is a potential gas disaster at the XX position of the XX lane, and there is a risk of gas explosion." After receiving the alarm, the dispatch room must immediately notify relevant personnel, such as the mine manager, technical person in charge, and various emergency rescue teams, in accordance with the procedures specified in the emergency plan. At the same time, the situation must be reported to the superior competent department, local government safety regulatory agency, etc.

Claims

1. A mine disaster prediction and emergency response method based on artificial intelligence, characterized in that: include: S1. Collecting original mine data through a mine data collection unit, manually marking the safety status corresponding to the original mine data, and then storing it; The safety status includes safety, potential gas disaster, potential roof disaster, potential water inrush disaster, potential fire disaster and potential dust disaster; S2. Generate simulated mine data based on the original mine data by a mine data enhancement unit, and use the original mine data and the simulated mine data together as training data; S3, inputting the training data into a mine disaster prediction unit to train a feature extraction model of the mine disaster prediction unit and a classification model of the mine disaster prediction unit; The feature extraction model for training the mine disaster prediction unit adopts a neural network algorithm based on quantum potential field optimization; The classification model for training the mine disaster prediction unit adopts an extreme learning machine algorithm based on a dynamic attractor; S4. New original mine data is collected through the mine data collection unit, and features are extracted from the new original mine data using the trained feature extraction model to obtain new original mine data after feature extraction. The trained classification model is used to classify the new original mine data after feature extraction to obtain a safety status corresponding to the new original mine data. After receiving the safety status, the emergency response unit outputs corresponding emergency response measures.

2. The method for predicting mine disasters and emergency response based on artificial intelligence according to claim 1, characterized in that: The generation of simulated mine data in S2 adopts a bilinear interpolation algorithm based on a dynamic constraint set, and the generation of simulated mine data includes: S201, initializing constraint set; S202, dynamically adjusting the constraint set according to changes in environmental parameters during sample expansion, including: In the formula, is the updated constraint set, C c is the initial value of the constraint set, λ c is the constraint set adjustment factor, M ce is the number of risk factors, f c () is the adjustment function based on the risk assessment model, R j,c is the real-time data of the mine for the jth risk factor, Δt c is the time interval for sample collection; S203, using a bilinear interpolation algorithm to fill in missing or incomplete mine data according to the current dynamic constraint set, expressed as: V k,c =(1-a c )(1-b c )V i,c +a c (1-b c )V i+1,c +(1-a c )b c V j,c +a c b c V j+1,c , Where V k,c is the predicted value of the point to be interpolated, a c is the first interpolation coefficient adjusted according to the dynamic constraint, b c is the second interpolation coefficient adjusted according to the dynamic constraint, V i,c is the existing i-th data point, V i+1,c is the existing i+1th data point, V j,c is the jth data point, V j+1,c is the existing j+1th data point; S204, using an adaptive noise suppression mechanism to effectively identify and process the noise in the mine data, the calculation formula is: In the formula, h c () is the noise identification function, V i,c is the existing i-th data point, a c is the first interpolation coefficient adjusted according to the dynamic constraint, b c is the second interpolation coefficient adjusted according to the dynamic constraint.

3. The method for predicting mine disasters and emergency response based on artificial intelligence according to claim 2 is characterized in that: The initialization constraint set in S201 is expressed as: In the formula, C c is the initial value of the constraint set, N ce is the total number of samples in the mine training dataset before sample generation, d i,c is the attribute value of the i-th data point collected from the mine data set, μ c is the mean of the attributes of the ith data point collected from the mine data set, σ c is the standard deviation of the attribute of the ith data point collected from the mine dataset.

4. The method for predicting mine disasters and emergency response based on artificial intelligence according to claim 2, characterized in that: The S202 described in f c (R j,c ,Δt c ) is calculated as: In the formula, Δt c is the sample collection time interval, μ t,c is the mean of the sample collection time deviation, σ t,c is the standard deviation of sample collection time deviation, k c is the slope control parameter of the response curve, R j,c is the real-time data of the mine with the jth risk factor, θ c is the activation threshold for adjustment.

5. The method for predicting mine disasters and emergency response based on artificial intelligence according to claims 1 to 4, characterized in that: The feature extraction model for training the mine disaster prediction unit in S3 includes: S311, initializing the neural network structure; S312, in each iteration, dynamically adjusting the neural network parameters according to the feedback generated by the quantum potential field based on the mine data, the neural network parameters are the neural network weights, and the update amount of the neural network weights ΔW p The calculation formula is: Where W p represents the weight of the neural network, η p is the learning rate of the neural network, is the gradient of the neural network's loss function with respect to the weights, is the cross entropy loss function, is the mine training data set, ξ p is the influence intensity of quantum potential field on parameter adjustment, Ψ p () is the quantum potential field function; S313, dynamically adjusting the neural network structure according to error feedback during the training process; S314, after a preset number of iterations, the feature extraction capability of the neural network is evaluated. The evaluation method of the feature extraction efficiency is expressed as: In the formula, Γ p Indicates the accuracy of feature extraction, V fe is the preset validation mine data set, v se The data in the preset verification mine data set, is the data point v se The direction angle of the true eigenvector of is the data point v se The direction angle of the feature vector predicted by the model, <,> represents the dot product of the vector, and ∥∥ is the L2 norm; S315, repeating iterations S312 to S314 until a preset condition for stopping iteration is met, indicating that the model training is completed.

6. The method for predicting mine disasters and emergency response based on artificial intelligence according to claim 5, characterized in that: The initialization of the neural network structure in S311 includes: In the formula, represents the initial weight matrix of the neural network, σ p is the scale factor of the neural network parameters, randn() means generating random numbers, and randn(0,0.1) means generating random numbers in the interval (0,0.1).

7. The method for predicting mine disasters and emergency response based on artificial intelligence according to claim 5, characterized in that: The Ψ in S312 p (W p ) is calculated as: Where K is the total number of quantum states, ζ k,p is the influence intensity of the kth quantum state, μ k,p is the center position of the kth quantum state, σ k,p is the dispersion of the quantum state.

8. The method for predicting mine disasters and responding to emergencies based on artificial intelligence according to claims 1 to 4, characterized in that: The classification model for training the mine disaster prediction unit in S3 includes: S321, initializing the weights and biases of the extreme learning machine; S322, the training data after feature extraction is forward propagated through the initialized weight matrix of the extreme learning machine. During the forward propagation process, the output calculation formula of the hidden layer of the extreme learning machine is: H u =Sig(W u X u +b u ), In the formula, H u is the output of the hidden layer of the extreme learning machine, Sig() is the Sigmoid activation function, and W u is the weight matrix of the extreme learning machine, b u is the bias vector of the extreme learning machine; S323, using Moore-Penrose pseudo-inverse to minimize the output error, the output layer weight of the extreme learning machine is solved, which is expressed as: In the formula, β u is the output layer weight of the extreme learning machine, H u is the hidden layer output of the extreme learning machine; is the Moore-Penrose pseudo-inverse of the hidden layer output of the extreme learning machine; T u is the target output of the extreme learning machine; S324. After each batch or a preset number of iterations, the weights from the hidden layer to the output layer are adjusted based on the dynamic attractor strategy. The weight adjustment method based on the dynamic attractor is expressed as: In the formula, is the weight of the extreme learning machine at the t+1th iteration, is the weight of the extreme learning machine at the tth iteration, η ur (t) is the learning rate of the extreme learning machine at the tth iteration, R u () is the regularization weight of the hidden layer output of the extreme learning machine, H u is the hidden layer output of the extreme learning machine, Attr() is the dynamic attractor function defined based on the current hidden layer output and error of the extreme learning machine, Error u is the error between the current model output and the target output; S325, adopting a high-dimensional feature regularization mechanism based on energy mapping, by constructing an energy mapping function, regularizing the redundancy and instability that may appear in the high-dimensional feature space, while retaining important high-energy features, and calculating the regularization weights of the features output by the hidden layer of the extreme learning machine; S326, repeating iterations S322 to S325 until a preset condition for stopping iteration is met, indicating that the model training is completed.

9. The method for predicting mine disasters and emergency response based on artificial intelligence according to claim 8, characterized in that: The Moore-Penrose pseudo-inverse of the hidden layer output of the extreme learning machine in S323 The calculation formula is: In the formula, is the transpose of the hidden layer output of the extreme learning machine, λ mr is the regularization parameter of the extreme learning machine, and I is the identity matrix.

10. The method for predicting mine disasters and emergency response based on artificial intelligence according to claim 8, characterized in that: The Attr(H u ,Error u ) is calculated as: In the formula, δ u To adjust the parameters, tanh() is the hyperbolic tangent function, Error u is the error between the current model output and the target output.

Images