Working face abnormal weighting prediction method based on data decomposition and deep learning

Abstract

The invention discloses a working face abnormal weighting prediction method based on data decomposition and deep learning. The method comprises the following steps: acquiring each data sequence of a coal mine stope working face; a CEEMDAN-VMD decomposition algorithm is adopted to decompose the IMF components into IMF components; dividing the data into a training set, a verification set and a test set; constructing a Seq2Seq model, and training and testing the Seq2Seq model by using the training data to obtain a prediction result; the prediction results of the IMF components are recombined to obtain a final prediction result; and performing clustering analysis on the predicted value of the hydraulic support pressure data of the coal mine stope face obtained by the final model to obtain a weighting type. According to the method, the non-stability of the data sequence is remarkably reduced, the accuracy of the prediction result is improved, and the abnormal weighting is subjected to clustering analysis, so that a production operation plan is adjusted for a coal mine to meet the safety production of a stope face.

Description

Technical Field

[0001] The present invention relates to the field of prediction of abnormal roof pressure of coal mine working faces, and in particular to a method for predicting abnormal roof pressure of working faces based on data decomposition and deep learning. Background Art

[0002] Coal mining often faces complex geological conditions. The mining face is affected by square fractures, strip crushing areas, upper goaf areas, and residual coal pillars. When the coal seam enters the mining impact area, it is affected by the special coal and rock occurrence conditions and repeated mining disturbances and sudden abnormal pressure. The hydraulic support of the mining face has the risk of pressure. The online monitoring system of the working resistance of the hydraulic support of the working face can analyze the pressure in real time, but it cannot predict the future pressure, and the on-site staff cannot make corresponding preparations in advance. Summary of the invention

[0003] Purpose of the invention: The purpose of the present invention is to provide a method for predicting abnormal working face pressure based on data decomposition and deep learning, which can reduce the non-stationarity of data sequences and improve prediction accuracy.

[0004] Technical solution: A method for predicting abnormal pressure on a working face based on data decomposition and deep learning, comprising the following steps:

[0005] (1) Obtain the working conditions, production conditions, and mining conditions of the hydraulic supports of the coal mine recovery working face, complete the missing values, and process the outliers of the above data to obtain the working resistance data series;

[0006] (2) Based on the working resistance data sequence of the hydraulic support of the mining face obtained in step (1), the CEEMDAN-VMD decomposition algorithm is used to decompose and reorganize the working resistance data sequence of the hydraulic support of the mining face into a set of IMF components as the hydraulic support pressure subsequence of the mining face, and the gray wolf algorithm is used to optimize the decomposition number and penalty factor of the VMD algorithm;

[0007] (3) based on the working condition, production condition, and mining condition of the hydraulic support of the mining face obtained in step (1) and a set of IMF components obtained in step (2), the above data and IMF components are divided into a training set, a validation set, and a test set;

[0008] (4) Construct a Seq2Seq model for each sub-IMF component in a group of IMF components, train and test the Seq2Seq model using training data, and obtain prediction results;

[0009] (5) The prediction results of each sub-IMF component are reorganized to obtain the final prediction result, realizing the prediction of the hydraulic support pressure data of the coal mine working face;

[0010] (6) Cluster analysis is performed on the predicted values ​​of the hydraulic support pressure data of the coal mine working face obtained by the Seq2Seq model to obtain the pressure type.

[0011] Furthermore, the working conditions of the hydraulic support of the coal mine recovery working face include the average working resistance of the hydraulic support in the middle of the recovery working face, time-weighted resistance, initial support force, final resistance, resistance increase rate in the initial pressurization stage, resistance increase rate in the relatively stable load-bearing stage, resistance increase rate in the pressurization stage before frame shifting, and the number of times the safety valve is opened;

[0012] The production conditions include hydraulic support advancement distance and cycle time;

[0013] The mining conditions include the square fracture position of the working face, the square fracture position of the upper section working face, the range of the strip crushing area, the distance of the remaining coal pillars from the goaf above the working face, the size range of the remaining coal pillars, the interlayer spacing, and the time when mining is stopped at the upper working face.

[0014] Furthermore, the missing value filling and outlier processing of the acquired data include the following steps:

[0015] (31) An interpolation method based on time series characteristics is used to fill in missing values ​​in the original data set to ensure the temporal continuity and consistency of the data;

[0016] (32) Use outlier detection algorithms to identify the completed data and mark potential outliers in the data set;

[0017] (33) Eliminate, replace or correct detected abnormal values;

[0018] (34) The data series after missing value filling and outlier processing are decomposed into IMF components using the CEEMDAM-VMD decomposition algorithm.

[0019] Furthermore, the CEEMDAN decomposition algorithm calculation includes the following steps:

[0020] (411) Add the adaptive standard normal distribution white noise y to the original sequence y(t) in the hydraulic support working resistance data sequence of the mining face i (t) = y(t) + ε o ω i (t), i=1, 2,...N, where, y i (t) is the sequence after adding adaptive standard normal distribution white noise to the original sequence, ε0 is the signal-to-noise ratio, N is the number of times white noise is added, ω i (t) is a white noise sequence;

[0021] (412) to y i(t) Perform EMD decomposition until the first modal component of CEEMDAN is obtained. in, Indicates y i (t) Perform the first modal component obtained by the i-th EMD decomposition and calculate the first residual sequence

[0022] (413) Using EMD to decompose the sequence r1(t)+ω i (t) Obtain the second modal component of CEEMDAN Among them, E1() represents the first modal component obtained by EMD decomposition of the sequence in the brackets;

[0023] (414) For the subsequent stage k=2,…,N, calculate the kth residual sequence Repeat the calculation process of step (413), the kth modal component of CEEMDAN

[0024] (415) Repeat step (414) to ensure that the residual sequence cannot be further decomposed, and finally obtain k modal components. The final residual signal of the decomposition is Among them, K represents the number of IMF components; therefore, the original sequence

[0025] The VMD algorithm calculation includes the following steps:

[0026] (421) Construct a variational problem; suppose the original sequence y(t) is decomposed into L components u l (t), l = 1, 2, ..., M, to ensure that the decomposed sequence has a central frequency ρ l The modal component of finite bandwidth, and the sum of the estimated bandwidths of each mode is minimized. The constraint condition is that the sum of all modes is equal to the original sequence. The corresponding constrained variational expression is: in, It represents minimization with u l , ρ l is the value of a function of a variable; represents the partial derivative with respect to t; δ(t) represents the Dirac distribution; c represents an imaginary number; u l (t) represents the lth component obtained by decomposition;

[0027] (422) Solve the constrained variational expression; introduce the penalty parameter α and the Lagrange multiplier λ, transform the constrained variational problem into an unconstrained variational problem, and obtain the augmented Lagrange expression:

[0028]

[0029] (423) Solve the saddle point of the augmented Lagrange expression in step (422); initialize the parameter u l ,ρ2,λ(I) and the number of iterations n, the initial value of n is set to 0, set the loop process, let n=n+1,u l ,ρ1,λ(t) is updated according to the following formula:

[0030]

[0031]

[0032]

[0033] in, Respectively represent u l n+1 The Fourier transform of (t), y(t), λ(t), represents the center frequency of the lth component in the n+1th iteration; γ represents the noise tolerance;

[0034] (424) When the components satisfy the formula When , the solution is completed; where ζ represents the accuracy convergence criterion.

[0035] Furthermore, in step (3), in order to increase the number of samples, the sliding window method is used to construct samples, and the IMF components of the adjacent working face, the hydraulic support working conditions, production conditions, and mining conditions of the mining working face are used as inputs, and the IMF components of the current working face are used as labels. The input vector X and output vector Y of the Seq2Seq model are as follows:

[0036] X=(X1,X2,…,X Q )

[0037] Y=(Y1,Y2,…,Y Q )

[0038] X q =(x q ,x q+1 ,…,x n′ )

[0039] Y q =(x n′+1 ), Q is the total number of samples; n′ is the amount of sample data, X q , Y q They are the input and output of the qth sample q=1,2,…,Q respectively.

[0040] Furthermore, the data segmentation in step (3) includes the following features:

[0041] Data segmentation ratio: determine the ratio of training set, validation set and test set according to the data structure and model complexity;

[0042] Continuity of time series. Considering the time series characteristics of data, data segmentation is performed in a time-continuous manner to ensure that the data points in the training set, validation set, and test set are arranged in chronological order;

[0043] Cross-validation method, uses time series cross-validation method to split the data into multiple training or test pairs;

[0044] Ensure data representativeness. When splitting data, ensure that each training set, validation set, and test set can represent the characteristics of the entire data set. When dealing with highly unbalanced data, use stratified sampling.

[0045] Furthermore, constructing the Seq2Seq model in step (4) includes the following steps:

[0046] (71) The sample (X q ,Y q ) is input into the Seq2Seq model. The encoder of the Seq2Seq model consists of a single-layer LSTM unit. The input sequence is read in step by step in chronological order. The hidden layer state h at each moment q The current input data x q and the hidden layer state h at the previous moment q-1 and cell state c q-1 Joint decision, q =f(x q ,h q-1 , c q-1 ), h q ,h q-1 are the hidden layer states of the LSTM neurons in the encoding layer at time q and time q-1, respectively. q is the input at time q, c q-1 is the LSTM neuron cell state of the encoding layer at time q-1;

[0047] (72) The LSTM unit is composed of an input gate, a forget gate, and an output gate. The calculation process inside the LSTM unit is as follows:

[0048] (721) The forget gate determines how many cell states s are discarded q-1 :f q =σ(W f (h q-1 , x q )+b f );

[0049] (722) The input gate determines how much current external input data x is retained q, and generate candidate cell states c q :v q =σ(W v (h q-1 , x q )+b v )

[0050]

[0051] (723) The output gate determines the output cell state c q Which features of the hidden layer state h are generated? q :o q =σ(W0(h q-1 , X q )+b0)

[0052] h q =o q tanhc q , where W f , W v , W c , W o are the weight vectors of the forget gate, input gate, output gate and gate control unit respectively; b f , b v , b c , b o are the bias vectors of the forget gate, input gate, output gate and gate control unit respectively, and f q , v q , , o q are the intermediate parameters of the model, σ(·) is the Sigmoid activation function, and tanh(·) is the hyperbolic tangent activation function;

[0053] (73) After the encoder is updated for g time steps, the input data is encoded into the hidden layer state h at the final time step g , the decoder consists of a single layer of LSTM units, which accepts the final state h of the encoder g As the initial input value; the decoder hidden layer state h′ at each moment q Through the current input h g and the hidden layer state h′ at the previous moment q-1 and cell state c′ q-1 Update, its expression is as follows: h′ q =f(h g , h′ q-1 , c′ q-1 ), h′ q , h′ q-1 Represents the hidden layer states of LSTM neurons in the decoding layer at time q and time q-1, respectively, c′ q-1The cell state of the LSTM neuron in the decoding layer at time q-1;

[0054] (74) Decoding layer updates h′ q-1 The specific operation and encoding layer update h q Consistent, hidden state h′ q The learning output of the fully connected layer is y q =F(h q ), y q is the output at time q, which is decoded step by step to form the final output sequence.

[0055] Furthermore, the calculation steps of the grey wolf algorithm include:

[0056] (81) Parameter definition: determine the key parameters in the VMD decomposition process—the number of decompositions K and the penalty factor α as the optimization objectives;

[0057] (82) Initialize the gray wolf population. Randomly initialize a group of gray wolves within a reasonable parameter range. Each individual represents a possible configuration of K and α.

[0058] (83) Fitness evaluation, which evaluates the performance of each individual by defining a fitness function, which is the reconstruction error between the IMF components after VMD decomposition and the original data sequence;

[0059] (84) Social hierarchy update: According to the fitness evaluation results, determine the Alpha, Beta, and Delta wolves, which represent the best parameter configuration currently found. Then, update the parameter configuration of other individuals in the population to explore the optimal solution.

[0060] (85) Iterative optimization, repeatedly executing the fitness evaluation and wolf group position update steps until the termination condition is met;

[0061] (86) The optimal K and α values ​​found by the Grey Wolf Algorithm are used to perform VMD decomposition on the original hydraulic support working resistance data series of the mining face to obtain the final IMF components.

[0062] Furthermore, the pressure types are classified according to the prediction results into abnormal pressure formed in squares, abnormal pressure formed in strip crushing areas, abnormal pressure formed through goaf areas, and abnormal pressure formed through residual coal pillars.

[0063] Compared with the prior art, the present invention has the following significant effects: the present invention uses a quadratic decomposition method to process air quality data sequences, significantly reducing the non-stationarity of the data sequences; constructs a Seq2Seq model, adds production status data and mining condition data as high-dimensional features to input data, and improves the accuracy of prediction results; in the coal mine recovery process, it can predict the working resistance of the hydraulic support in the future based on the hydraulic support working resistance monitoring data and the working face mining condition data, and perform cluster analysis on abnormal pressure, so as to adjust the production operation plan for the coal mine to meet the safety production needs of the recovery working face, and provide decision support at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic diagram of the process of the present invention;

[0065] Figure 2 Flow chart of VMD decomposition optimized by Grey Wolf Algorithm of the present invention;

[0066] Figure 3 This is a structural diagram of the Seq2Seq model of the present invention. DETAILED DESCRIPTION

[0067] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0068] The present invention provides a method for predicting abnormal pressure on a working face based on data decomposition and deep learning, taking the prediction and early warning of abnormal pressure on the roof of the 31317 mining working panel in Ordos Coal Mine as an example. Figure 1 As shown, the specific implementation steps are as follows:

[0069] Step 1, obtain the working condition, production status and mining conditions of the hydraulic support of the coal mine recovery working face, perform missing value filling and outlier processing operations on the acquired data, and obtain the working resistance data sequence, footage data, periodic pressure step data and periodic pressure continuous step data of the hydraulic support of the recovery working face respectively;

[0070] The working condition of the hydraulic support in the mining face mainly refers to the average working resistance of the hydraulic support in the middle of the mining face, time-weighted resistance, initial support force, final resistance, resistance increase rate in the initial pressurization stage, resistance increase rate in the relatively stable load-bearing stage, resistance increase rate in the pressurization stage before frame movement, and the number of times the safety valve is opened; the production status mainly refers to the advancement distance and cycle time of the hydraulic support; the mining conditions include the square breaking position of the working face, the square breaking position of the working face in the upper section, the range of the strip crushing area, the distance from the remaining coal pillar in the goaf above the working face, the size range of the remaining coal pillar, the interlayer spacing, and the time when the upper working face was stopped.

[0071] Step 2: Based on the hydraulic support working resistance data sequence of the mining face obtained in step 1, the CEEMDAN-VAM decomposition algorithm is used to decompose the hydraulic working resistance data sequence of the mining face into a set of IMF components, namely, the hydraulic support pressure subsequence of the mining face;

[0072] Step 21, using the CEEMDAN algorithm to decompose the working resistance data sequence of the hydraulic support of the mining face obtained in step 1 to obtain a number of sub-IMF components; the calculation of the CEEMDAN algorithm includes the following steps:

[0073] The steps of using the CEEMDAN-VAM decomposition algorithm are as follows:

[0074] Step 211, add the adaptive standard normal distribution white noise y to the original sequence y(t) in the hydraulic support working resistance data sequence of the mining face. i (y) = y(t) + ε o ω i (t), i=1, 2,...N, where, y i (t) is the sequence after adding adaptive standard normal distribution white noise to the original sequence, ε0 is the signal-to-noise ratio, N is the number of times white noise is added, ω i (t) is a white noise sequence;

[0075] Step 212, for y i (t) Perform EMD decomposition until the first modal component of CEEMDAN is obtained. in, Indicates y i (t) Perform the first modal component obtained by the i-th EMD decomposition and calculate the first residual sequence

[0076] Step 213, use EMD to decompose the sequence r1(t)+ω i (t) Obtain the second modal component of CEEMDAN Among them, E1() represents the first modal component obtained by EMD decomposition of the sequence in the brackets;

[0077] Step 214, for the subsequent stages k=2, ..., N, calculate the kth residual sequence Repeat the calculation process of step (413), the kth modal component of CEEMDAN

[0078] Step 215, repeat step 214 to ensure that the residual sequence cannot be further decomposed, and finally obtain k modal components, and the final residual signal of the decomposition is Among them, K represents the number of IMF components; therefore, the original sequence

[0079] Step 22, calculating the permutation entropy of each sub-IMF component, and sorting all components from large to small according to the permutation entropy; the permutation entropy calculation step includes:

[0080] Step 221, for a time series y(1), y(2), ..., y(t), set an embedding dimension m, a time delay t, and a number of reconstruction components T = t-(m-1)τ to ​​reconstruct the sequence in phase space: τ is the time delay, the time interval between the components when creating the phase space vector, which can be estimated by the autocorrelation function or the mutual information method;

[0081] Step 222, arrange the j=1, 2, ..., Tth reconstruction components in the reconstruction matrix in ascending order according to the numerical value, and the obtained sorted index values ​​constitute a set of symbol sequences S; if the reconstruction matrix of m columns is arranged in different arrangements (1, 2, ..., m), there are a total of m! symbol sequences in full arrangement;

[0082] Step 223, count the number of times each arrangement corresponding to the T reconstruction components appears in the m! full arrangement, and divide the number of times corresponding to each column by the total number of times in the column to obtain the probability P1, P2, ..., P of each corresponding position. j ,P T ;

[0083] Step 224, calculate the information entropy of these probabilities and obtain the permutation entropy H of the time series y(1), y(2), ..., y(t) PE (m):

[0084] Step 33, the VMD algorithm decomposes the component with the highest permutation entropy to obtain several sub-IMF components. The flowchart of the VMD algorithm optimized by the Grey Wolf Algorithm is shown in Figure 2 ; The VMD algorithm comprises the following steps:

[0085] Step 331, construct a variational problem; suppose the original sequence y(t) is decomposed into L components u l (t), l = 1, 2, ..., M, to ensure that the decomposed sequence has a central frequency ρ l The modal component of finite bandwidth, and the sum of the estimated bandwidths of each mode is minimized. The constraint condition is that the sum of all modes is equal to the original sequence. The corresponding constrained variational expression is:

[0086] in, It represents minimization with u l , ρ l is the value of a function of a variable; represents the partial derivative with respect to t; δ(t) represents the Dirac distribution; c represents an imaginary number; ul (t) represents the lth component obtained by decomposition;

[0087] Step 332, solving the constrained variational expression; introducing the penalty parameter α and the Lagrange multiplier λ, transforming the constrained variational problem into an unconstrained variational problem, and obtaining the augmented Lagrange expression as follows:

[0088]

[0089] Step 333, solve the saddle point of the augmented Lagrange expression; initialize the parameter u l ,ρ2,λ(I) and the number of iterations n, the initial value of n is set to 0, set the loop process, let n=n+1,u l ,ρ1,λ(t) is updated according to the following formula:

[0090]

[0091]

[0092]

[0093] in, Respectively represent u l n+1 The Fourier transform of (t), y(t), λ(t), represents the center frequency of the lth component in the n+1th iteration; γ represents the noise tolerance;

[0094] Step 334, when the components satisfy the formula When , the solution is completed; where ζ represents the accuracy convergence criterion.

[0095] Step 34, the components obtained by CEEMDAN decomposition except the component with the highest permutation entropy are divided into two groups according to the permutation entropy, and are reorganized to obtain an intermediate frequency IMF component and a low frequency IMF component;

[0096] Step 35, arranging the components decomposed by the VMD algorithm in step 3 and the intermediate frequency component and the low frequency component decomposed and recombined by the CEEMDAN algorithm in step 4 into a group of IMF components, ready for training and prediction;

[0097] Step 3: Based on the hydraulic support working condition, production condition, and mining condition of the mining face obtained in step 1, and a set of IMF components obtained in step 2, the IMF components and data are divided into a training set, a validation set, and a test set. In order to increase the number of samples, the sliding window method is used to construct samples. The IMF component data of the adjacent working face, the hydraulic support working condition data, production condition data, and mining condition data of the mining face are used as inputs, and the IMF component data of the current working face is used as a label. The input vector X and output vector Y of the model are as follows:

[0098] X=(X1,X2,…,X Q )

[0099] Y=(Y1,Y2,…,Y Q )

[0100] X q =(x q ,x q+1 ,…,x n′ )

[0101] Y q =(x n′+1 ), Q is the total number of samples; n′ is the amount of sample data, X q , Y q They are the input and output of the qth sample q=1,2,…,Q respectively.

[0102] Step 4: Build a Seq2Seq model for each sub-IMF component in a set of IMF components, and train the Seq2Seq model using the training data; the structure of the Seq2Seq model is referenced in Figure 3 ;

[0103] The Seq2Seq model includes the following steps:

[0104] Step 41, the sample (X q ,Y q ) is input into the Seq2Seq model. The encoder of the Seq2Seq model consists of a single-layer LSTM unit. The input sequence is read in step by step in chronological order. The hidden layer state h at each moment q The current input data x q and the hidden layer state h at the previous moment q-1 and cell state c q-1 Joint decision, q =f(x q ,h q-1 , c q-1 ), h q ,h q-1 are the hidden layer states of the LSTM neurons in the encoding layer at time q and time q-1, respectively.q is the input at time q, c q-1 is the LSTM neuron cell state of the encoding layer at time q-1;

[0105] Step 42, the LSTM unit is composed of an input gate, a forget gate, and an output gate. The calculation process inside the LSTM unit is as follows:

[0106] Step 421, the forget gate determines how many cell states s to discard q-1 :f q =σ(W f (h q-1 , x q )+b f );

[0107] Step 422, the input gate determines how much current external information (i.e., input data x) to retain q ), and generate candidate cell states c q :

[0108] v q =σ(W v (h q-1 , x q )+b v )

[0109]

[0110] Step 423, the output gate determines the output cell state c q Which features of the hidden layer state h are generated? q :o q =σ(W0(h q-1 , X q )+b0)

[0111] h q =o q tanhc q In the above formula, W f , W v , W c , W o are the weight vectors of the forget gate, input gate, output gate and gate control unit respectively; b f , b v , b c , b o are the bias vectors of the forget gate, input gate, output gate and gate control unit, fq, vq, , oq are the intermediate parameters of the model, σ(·) is the Sigmoid activation function, and tanh(·) is the hyperbolic tangent activation function;

[0112] Step 43, as Figure 3As shown, the encoder is updated for g time steps, and the input data is encoded into the hidden layer state h at the final time step. g , the decoder consists of a single layer of LSTM units, which accepts the final state h of the encoder g As the initial input value; the decoder hidden layer state h′ at each moment q Through the current input h g and the hidden layer state h′ at the previous moment q-1 and cell state c′ q-1 Update, its expression is as follows: h′ q =f(h g , h′ q-1 , c′ q-1 ), h′ q , h′ q-1 Represents the hidden layer states of LSTM neurons in the decoding layer at time q and time q-1, respectively, c′ q-1 The cell state of the LSTM neuron in the decoding layer at time q-1;

[0113] Step 44, the decoding layer updates h′ q-1 The specific operation and encoding layer update h q Consistent, hidden state h′ q The learning output of the fully connected layer is y q =F(h q ), y q is the output at time q, which is decoded step by step to form the final output sequence.

[0114] In addition, using R 2 , mean absolute error and mean square error are used as measurement indicators to quantify the prediction accuracy and error of each Seq2Seq model. The specific calculation formula is as follows:

[0115]

[0116]

[0117] In the formula, y i is the observed value, is the predicted value and y is the mean of the observed values.

[0118] Step 5: Test the trained Seq2Seq model on the test set;

[0119] Step 6, reorganize the prediction results of each sub-IMF component to obtain the final prediction result, and realize the prediction of the hydraulic support pressure data of the coal mine recovery working face;

[0120] Step 7, with the goal of better prediction results, the gray wolf optimization algorithm is used to optimize the number of decompositions and penalty factor parameters of VMD; Figure 2 As shown, the gray wolf algorithm includes the following steps:

[0121] Step 71, parameter definition, determining the key parameters in the VMD decomposition process - the number of decompositions (K) and the penalty factor (α) as optimization targets;

[0122] Step 72, initialize the gray wolf population, randomly initialize a group of gray wolves within a reasonable parameter range, and each individual represents a set of possible configurations of K and α;

[0123] Step 73, fitness evaluation, evaluates the performance of each individual by defining a fitness function, which can be based on the correlation between the IMF components after VMD decomposition and the original data sequence, the reconstruction error or other quality indicators;

[0124] Step 74, social hierarchy update, according to the fitness evaluation results, determine the Alpha, Beta, and Delta wolves, representing the best parameter configuration currently found, and then update the parameter configuration of other individuals in the population to explore the optimal solution;

[0125] Step 75, iterative optimization, repeatedly performing the fitness evaluation and wolf pack position update steps until a termination condition is met, such as reaching a maximum number of iterations or the fitness improvement is lower than a predetermined threshold;

[0126] Step 76, using the optimal K and α values ​​found by the Grey Wolf Algorithm, VMD decomposition is performed on the original mining face hydraulic support working resistance data sequence to obtain the final IMF component.

[0127] Step 8, cluster analysis is performed on the predicted values ​​of the hydraulic support pressure data of the coal mine recovery working face to obtain the pressure types, which include abnormal square pressure, abnormal pressure formed in the strip crushing area, abnormal pressure formed in the goaf area and abnormal pressure formed in the remaining coal pillar.

Claims

1. A method for predicting abnormal pressure on a working face based on data decomposition and deep learning, characterized in that: The following steps are involved: (1) Obtain the working conditions, production conditions, and mining conditions of the hydraulic supports of the coal mine recovery working face, complete the missing values, and process the outliers of the above data to obtain the working resistance data series; (2) Based on the working resistance data sequence of the hydraulic support of the mining face obtained in step (1), the CEEMDAN-VMD decomposition algorithm is used to decompose and reorganize the working resistance data sequence of the hydraulic support of the mining face into a set of IMF components as the hydraulic support pressure subsequence of the mining face, and the gray wolf algorithm is used to optimize the decomposition number and penalty factor of the VMD algorithm; (3) based on the working condition, production condition, and mining condition of the hydraulic support of the mining face obtained in step (1) and a set of IMF components obtained in step (2), the above data and IMF components are divided into a training set, a validation set, and a test set; (4) Construct a Seq2Seq model for each sub-IMF component in a group of IMF components, train and test the Seq2Seq model using training data, and obtain prediction results; (5) The prediction results of each sub-IMF component are reorganized to obtain the final prediction result, realizing the prediction of the hydraulic support pressure data of the coal mine working face; (6) Cluster analysis is performed on the predicted values ​​of the hydraulic support pressure data of the coal mine working face obtained by the Seq2Seq model to obtain the pressure type.

2. The method for predicting abnormal pressure on a working face based on data decomposition and deep learning according to claim 1 is characterized in that: The working conditions of the hydraulic support of the coal mine recovery working face include the average working resistance of the hydraulic support in the middle of the recovery working face, time-weighted resistance, initial support force, final resistance, resistance increase rate in the initial pressurization stage, resistance increase rate in the relatively stable load-bearing stage, resistance increase rate in the pressurization stage before frame transfer, and the number of times the safety valve is opened; The production conditions include hydraulic support advancement distance and cycle time; The mining conditions include the square fracture position of the working face, the square fracture position of the upper section working face, the range of the strip crushing area, the distance of the remaining coal pillars from the goaf above the working face, the size range of the remaining coal pillars, the interlayer spacing, and the time when mining is stopped at the upper working face.

3. The method for predicting abnormal pressure on a working face based on data decomposition and deep learning according to claim 1 is characterized in that: The missing value filling and outlier processing of the acquired data include the following steps: (31) An interpolation method based on time series characteristics is used to fill in missing values ​​in the original data set to ensure the temporal continuity and consistency of the data; (32) Use outlier detection algorithms to identify the completed data and mark potential outliers in the data set; (33) Eliminate, replace or correct detected abnormal values; (34) The data series after missing value filling and outlier processing are decomposed into IMF components using the CEEMDAM-VMD decomposition algorithm.

4. The method for predicting abnormal pressure on a working face based on data decomposition and deep learning according to claim 1 is characterized in that: The CEEMDAN decomposition algorithm calculation includes the following steps: (411) Add the adaptive standard normal distribution white noise y to the original sequence y(t) in the hydraulic support working resistance data sequence of the mining face i (t) = y(t) + ε o ω i (t), i=1, 2,...N, where, y i (t) is the sequence after adding adaptive standard normal distribution white noise to the original sequence, ε0 is the signal-to-noise ratio, N is the number of times white noise is added, ω i (t) is a white noise sequence; (412) to y i (t) Perform EMD decomposition until the first modal component of CEEMDAN is obtained. in, Indicates y i (t) Perform the first modal component obtained by the i-th EMD decomposition and calculate the first residual sequence (413) Using EMD to decompose the sequence r1(t)+ω i (t) Obtain the second modal component of CEEMDAN Among them, E1() represents the first modal component obtained by EMD decomposition of the sequence in the brackets; (414) For the subsequent stage k = 2, ..., N, calculate the kth residual sequence Repeat the calculation process of step (413), the kth modal component of CEEMDAN (415) Repeat step (414) to ensure that the residual sequence cannot be further decomposed, and finally obtain k modal components. The final residual signal of the decomposition is Among them, K represents the number of IMF components; therefore, the original sequence The VMD algorithm calculation includes the following steps: (421) Construct a variational problem; suppose the original sequence y(t) is decomposed into L components u l (t), I = 1, 2, ..., M, to ensure that the decomposed sequence has a central frequency ρ l The modal component of finite bandwidth, and the sum of the estimated bandwidths of each mode is minimized. The constraint condition is that the sum of all modes is equal to the original sequence. The corresponding constrained variational expression is: in, It represents minimization with u l , ρ l is the value of a function of a variable; represents the partial derivative with respect to t; δ(t) represents the Dirac distribution; c represents an imaginary number; u l (t) represents the lth component obtained by decomposition; (422) Solve the constrained variational expression; introduce the penalty parameter α and the Lagrange multiplier λ, transform the constrained variational problem into an unconstrained variational problem, and obtain the augmented Lagrange expression: (423) Solve the saddle point of the augmented Lagrange expression in step (422); initialize the parameter u l ,ρ2,λ(l) and the number of iterations n, the initial value of n is set to 0, set the loop process, let n=n+1,u l ,ρ1,λ(t) is updated according to the following formula: in, Respectively represent u l n+1 The Fourier transform of (t), y(t), λ(t), represents the center frequency of the lth component in the n+1th iteration; γ represents the noise tolerance; (424) When the components satisfy the formula When , the solution is completed; where ζ represents the accuracy convergence criterion.

5. The method for predicting abnormal pressure on a working face based on data decomposition and deep learning according to claim 1 is characterized in that: In step (3), in order to increase the number of samples, the sliding window method is used to construct samples. The IMF components of the adjacent working face, the hydraulic support working conditions, production conditions, and mining conditions of the mining working face are used as inputs, and the IMF components of the current working face are used as labels. The input vector X and output vector Y of the Seq2Seq model are as follows: X=(X1,X2,…,X Q ) Y=(Y1,Y2,…,Y Q ) X q =(x q ,x q+1 ,…,x n′ ) Y q =(x n′+1 ), Q is the total number of samples; n′ is the amount of sample data, X q , Y q They are the input and output of the qth sample q=1,2,…,Q respectively.

6. The method for predicting abnormal pressure on a working face based on data decomposition and deep learning according to claim 1 is characterized in that: The data segmentation in step (3) includes the following features: Data segmentation ratio: determine the ratio of training set, validation set and test set according to the data structure and model complexity; Continuity of time series. Considering the time series characteristics of data, data segmentation is performed in a time-continuous manner to ensure that the data points in the training set, validation set, and test set are arranged in chronological order; Cross-validation method, uses time series cross-validation method to split the data into multiple training or test pairs; Ensure data representativeness. When splitting data, ensure that each training set, validation set, and test set can represent the characteristics of the entire data set. When dealing with highly unbalanced data, use stratified sampling.

7. The method for predicting abnormal pressure on a working face based on data decomposition and deep learning according to claim 6 is characterized in that: The step (4) of constructing the Seq2Seq model includes the following steps: (71) The sample (X q ,Y q ) is input into the Seq2Seq model. The encoder of the Seq2Seq model consists of a single-layer LSTM unit. The input sequence is read in step by step in chronological order. The hidden layer state h at each moment q The current input data x q and the hidden layer state h at the previous moment q-1 and cell state c q-1 Joint decision, q =f(x q ,h q-1 , c q-1 ), h q ,h q-1 are the hidden layer states of the LSTM neurons in the encoding layer at time q and time q-1, respectively. q is the input at time q, c q-1 is the LSTM neuron cell state of the encoding layer at time q-1; (72) The LSTM unit is composed of an input gate, a forget gate, and an output gate. The calculation process inside the LSTM unit is as follows: (721) The forget gate determines how many cell states s are discarded q-1 :f q =σ(W f (h q-1 , x q )+b f ); (722) The input gate determines how much current external input data x is retained q , and generate candidate cell states c q : v q =σ(W v (h q-1 ,x q )+b v ) (723) The output gate determines the output cell state c q Which features of the hidden layer state h q : about q =σ(W0(h q-1 ,X q )+b0) h q =o q tanhc q , where W f , W v , W c , W o are the weight vectors of the forget gate, input gate, output gate and gate control unit respectively; b f , b v , b c , b o are the bias vectors of the forget gate, input gate, output gate and gate control unit respectively, and f q , v q ,c q ,o q are the intermediate parameters of the model, σ(·) is the Sigmoid activation function, and tanh(·) is the hyperbolic tangent activation function; (73) After the encoder is updated for g time steps, the input data is encoded into the hidden layer state h at the final time step g , the decoder consists of a single layer of LSTM units, which accepts the final state h of the encoder g As the initial input value; the decoder hidden layer state h′ at each moment q Through the current input h g and the hidden layer state h′ at the previous moment q-1 and cell state c′ q-1 Update, its expression is as follows: h′ q =f(h g , h′ q-1 , c′ q-1 ), h′ q , h′ q-1 Represents the hidden layer states of LSTM neurons in the decoding layer at time q and time q-1, respectively, c′ q-1 The cell state of the LSTM neuron in the decoding layer at time q-1; (74) Decoding layer updates h′ q-1 The specific operation and encoding layer update h q Consistent, hidden state h′ q The learning output of the fully connected layer is y q =F(h q ), y q is the output at time q, which is decoded step by step to form the final output sequence.

8. The method for predicting abnormal pressure on a working face based on data decomposition and deep learning according to claim 1 is characterized in that: The calculation steps of the gray wolf algorithm include: (81) Parameter definition: determine the key parameters in the VMD decomposition process—the number of decompositions K and the penalty factor α as the optimization objectives; (82) Initialize the gray wolf population. Randomly initialize a group of gray wolves within a reasonable parameter range. Each individual represents a possible configuration of K and α. (83) Fitness evaluation, which evaluates the performance of each individual by defining a fitness function, which is the reconstruction error between the IMF components after VMD decomposition and the original data sequence; (84) Social hierarchy update: According to the fitness evaluation results, determine the Alpha, Beta, and Delta wolves, which represent the best parameter configuration currently found. Then, update the parameter configuration of other individuals in the population to explore the optimal solution. (85) Iterative optimization, repeatedly executing the fitness evaluation and wolf group position update steps until the termination condition is met; (86) The optimal K and α values ​​found by the Grey Wolf Algorithm are used to perform VMD decomposition on the original hydraulic support working resistance data series of the mining face to obtain the final IMF components.

9. The method for predicting abnormal pressure on a working face based on data decomposition and deep learning according to claim 1 is characterized in that: The pressure types are classified according to the prediction results into abnormal pressure formed in squares, abnormal pressure formed in strip crushing areas, abnormal pressure formed through goaf areas, and abnormal pressure formed through residual coal pillars.

Images