MSADBO-CNN-GRU-Attention-based gas concentration time sequence prediction method

Abstract

The invention relates to the technical field of mine gas disaster prediction, in particular to a gas concentration time sequence prediction method based on MSADBO-CNN-GRU-Attention. Firstly, dynamic Pearson correlation analysis is used, weighting is conducted on all indexes based on a time window of one hour, key factors leading gas concentration in different time periods are revealed, and therefore interference of irrelevant factors is weakened, and meanwhile the signal strength of the key factors is enhanced. Thirdly, performing local feature extraction on the data by applying a convolutional neural network (CNN), capturing multi-scale information, learning a time sequence data long-term dependency relationship and transmitting state information between time steps in combination with a gating cycle unit (GRU); an Attention mechanism is added to weight the hidden state output by the GRU so as to highlight the importance of key time nodes and improve the accuracy of multi-step prediction. And then, an improved sine algorithm, adaptive Gaussian-Cauchy mixed variation disturbance and a Bernoulli chaotic mapping improved dung beetle search algorithm are fused to obtain MSADBO, and then model hyper-parameters are globally and adaptively optimized. And finally, training the model, establishing a prediction model based on MSADBO-CNN-GRU-Attention, and verifying the performance of the prediction model by comparing the model. According to the gas concentration multi-index multi-step time sequence prediction method provided by the invention, parameter adaptability, feature mining and depth time sequence modeling are combined, the prediction performance of the gas concentration is improved, and a solution is provided for high-precision multi-step time sequence prediction of the gas concentration under complex working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of mine gas disaster prediction, and in particular to a gas concentration time series prediction method based on MSADBO-CNN-GRU-Attention. Background Art

[0002] Gas is one of the major risk sources that restrict underground production safety. Abnormal gas accumulation will not only form a suffocating and dangerous environment in the working area, but also easily trigger a chain disaster effect due to the strong coupling correlation between the time series characteristics of gas concentration and the dynamic evolution of gas disasters, causing gas combustion, explosion, coal and gas outburst and other accidents. The gas concentration sequence contains the multi-physical field coupling information of coal body stress field-fracture field-seepage field, and its multi-dimensional time series fluctuation characteristics have key characterization value for the identification of precursors of coal-rock gas composite dynamic disasters. Therefore, accurate prediction of gas concentration is the core link of mine gas disaster warning and prevention and control.

[0003] The deep learning fusion model has the advantages of automatic feature learning, strong robustness and expressiveness. Among them, CNN extracts the local fluctuation pattern and spatial correlation of the gas concentration sequence through multi-level convolution kernels. The GRU network models the long-term and short-term temporal dependencies and transmits hidden state information. CNN is combined with BiLSTM, and the Attention mechanism is added to dynamically allocate weight coefficients at different time steps to fully mine the deep information of gas concentration. Combined with data preprocessing, multi-module collaboration and adaptive hyperparameter optimization, the prediction accuracy and robustness of the model are improved. Summary of the invention

[0004] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0005] The present invention dynamically weights the time window based on Pearson correlation analysis, enhances the quality of initial features and improves the robustness of the model; performs global search optimization of hyperparameters based on the improved dung beetle search algorithm (MSADBO), so that key parameters such as learning rate converge to the global optimal solution domain; combines the module advantages of CNN-GRU-Attention, extracts the local fluctuation pattern and spatial correlation of the gas concentration sequence through CNN through multi-level convolution kernels, GRU network models long-term and short-term temporal dependencies and transmits hidden state information, and the Attention mechanism dynamically allocates weight coefficients of different time steps to comprehensively mine deep information on gas concentration. Through data preprocessing, multi-module collaboration and adaptive hyperparameter optimization, the prediction accuracy and robustness of the model are improved, providing a new method support for building an intelligent early warning system for coal mine gas disasters.

[0006] To achieve the above object, according to one aspect of the present invention, the present invention provides the following technical solution:

[0007] A gas concentration time series prediction method based on MSADBO-CNN-GRU-Attention is characterized by being implemented in the following steps:

[0008] Step S1: The sensor collects multi-source time series data underground and obtains multi-index data including gas concentration. The sensors are arranged at multiple monitoring points in the working face of the mine, among which the gas concentration data to be predicted is used as the target indicator, and the data collected by the remaining N monitoring points is used as the influencing indicator data. Then, Pearson correlation analysis is used to perform dynamic weighting based on a 1h time window;

[0009] Step S2: Apply CNN to extract local features of the data and capture multi-scale information. Combine GRU to learn the long-term dependencies of time series data and transfer state information between time steps. Add Attention mechanism to weight the hidden state of GRU output to highlight the importance of key time nodes and improve the accuracy of multi-step prediction.

[0010] Step S3: The improved sine algorithm, the adaptive Gauss-Cauchy mixed variation perturbation and the Bernoulli chaotic mapping are integrated to improve the dung beetle search algorithm to obtain MSADBO, and then the model hyperparameters are globally adaptively optimized;

[0011] Step S4: Train the model, establish a prediction model based on MSADBO-CNN-GRU-Attention, and use the test set to perform multi-index and multi-step prediction of gas concentration for performance verification.

[0012] The specific process of step S1 is:

[0013] Step S101, using Pearson coefficient to perform correlation analysis, the calculation formula is as follows:

[0014]

[0015] Where R is the Pearson coefficient; X i is the characteristic variable; i is the target variable; X represents the mean of the characteristic variable; σ Y represents the mean of the target variable; n represents the number of data samples;

[0016] Step S102, setting the sliding window to 1h, constructing a dynamic Pearson correlation analysis matrix;

[0017] Step S103, taking the absolute value of the correlation coefficient in each time window of the dynamic correlation matrix;

[0018] Step S104, normalize the correlation coefficients in each time window so that their sum is 1, and obtain the weight of each influencing factor in the time window, the formula of which is:

[0019]

[0020] Step S105, in each time window, the weights of the influencing factors are used to multiply their values ​​to obtain weighted values, forming a dynamic enhanced feature matrix, and the weighted feature matrix is ​​used as the input of the prediction model.

[0021] The specific process of step S2 is:

[0022] Step S201: Input multi-indicator time series data in the shape of (T, N), where T is the time step and N is the number of indicators. Perform convolution operation to extract local spatial features, the formula is:

[0023] X conv =f(W conv *X input +b conv );

[0024] Among them, W conv and b conv is the convolution kernel parameter, f(·) is the activation function, and * is the convolution operation. Standardize the convolution output to speed up training:

[0025]

[0026] Apply the ReLU activation function, use the maximum pooling to reduce the dimension and retain important features, use the flattening layer to flatten the multi-dimensional features into a one-dimensional vector, and use the fully connected layer to map to the feature representation space:

[0027] X fc1 =W fc1 X flat1 +b fc1 ;

[0028] Step S202, flatten the input sequence into one-dimensional features, and use GRU to extract time-dependent features. The update formula of GRU is:

[0029]

[0030] Use the feature concatenation layer to concatenate the feature vectors output by CNN and GRU;

[0031] Step S203: Use the attention mechanism for weighting. The attention weight calculation formula is:

[0032]

[0033] Among them, Q, K, V are query, key, and value matrices, d k is the dimension of the key. Finally, the attention output is mapped to the target dimension to get the final prediction output:

[0034] X output =W fc X att +b fc .

[0035] The specific process of step S3 is:

[0036] Step S301, set the population size N, the maximum number of iterations T max , search space boundary [X min ,X max ] Use Bernoulli chaos map to generate the position X of the initial dung beetle population i (0) and speed V i (0), generate the initial solution:

[0037]

[0038] Among them, the Bernoulli sequence is generated by iteration:

[0039]

[0040] Calculate the fitness value f(X i ), and the individual historical optimal P i best and the global optimal G best Keep records;

[0041] Step S302, update the position by combining the sine function and the adaptive inertia weight w(t):

[0042]

[0043] Among them, A, B, C are the amplitude, frequency and phase of the sine function, φ is the acceleration factor. ω(t) is the inertia weight of linear decay:

[0044]

[0045] A larger inertia weight helps global search in the early stages of optimization, then the inertia weight gradually decreases and the algorithm focuses on local development;

[0046] Step S303, applying Gaussian-Cauchy hybrid mutation to the updated position:

[0047]

[0048] GN is Gaussian noise, used for local fine search; CN is Cauchy noise, used for global jump. λ controls the weight decay rate to ensure that the early stage of iteration is dominated by Cauchy perturbation and the later stage is dominated by Gaussian perturbation;

[0049] Step S304, applying Bernoulli chaotic perturbation to the mutated position:

[0050]

[0051] BM is the random perturbation generated by Bernoulli chaotic mapping, δ is the intensity of chaotic perturbation, which decreases with iteration;

[0052] Step S305, calculate the new position X i final The fitness value f(X i final ), update the individual optimal P i best and the global optimal G i best Repeat steps 2 to 4 until the maximum number of iterations T is reached. max , output the global optimal solution G best .

[0053] The specific process of step S4 is:

[0054] Step S401: Use the training set to train the model and establish a gas concentration time series prediction model based on MSADBO-CNN-GRU-Attention;

[0055] Step S402: Use the test set to test the MSADBO-CNN-GRU-Attention gas concentration time series prediction model to verify the performance of the model.

[0056] Due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0057] It is proposed to establish a multi-index and multi-step time series prediction model for gas concentration by combining data preprocessing, improved hyperparameter optimization and deep learning fusion network. The time series features are extracted and processed by combining the CNN-GRU-Attention deep learning fusion network. Through training, a gas concentration time series prediction model based on MSADBO-CNN-GRU-Attention was constructed; MSADBO was used to optimize the model hyperparameters to ensure the global search ability and convergence of the model; a comparative model was established to perform performance evaluation and verify its superiority. Compared with previous prediction models, the present invention improves the ability of data utilization and feature mining, and has significant advantages in model prediction accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solution of the implementation mode of the present invention, the present invention will be described in detail with reference to the accompanying drawings and examples. Obviously, the drawings described below are only for the present invention and protect some embodiments. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0059] Figure 1 It is a basic flow chart according to the embodiment of this specification;

[0060] Figure 2 It is a general structure diagram of CNN-GRU-Attention according to the embodiments of this specification;

[0061] Figure 3 is a general structure diagram of MSADBO according to the embodiment of this specification;

[0062] Figure 4 It is the prediction result of the model test set shown in the embodiment of this specification. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] The embodiment of the present invention discloses a gas concentration time series prediction method based on MSADBO-CNN-GRU-Attention, such as Figure 1 As shown, the following steps are included:

[0065] Step 1: The sensor collects multi-source time series data underground and obtains multi-index data including gas concentration. The sensors are arranged at multiple monitoring points in the working face of the mine, among which the gas concentration data to be predicted is used as the target indicator, and the data collected by the remaining N monitoring points are used as the influencing indicator data.

[0066] Then, Pearson correlation analysis was used to perform dynamic weighting based on a 1h time window.

[0067] Step 1.1. In this embodiment, multi-source underground monitoring data of a coal mine from March 2 to May 27, 2014 are selected, of which MM264 is the target indicator and 27 influencing factors are used. A total of 499,920 time step data are obtained for time series prediction.

[0068] Step 1.2: In this embodiment, data with gas concentration greater than 5% and less than 0 are identified as abnormal values ​​and eliminated, and the moving average method is used to perform data interpolation to obtain complete data.

[0069] Step 1.3, in this embodiment, the training set and the test set are divided according to the ratio of 8:2, that is, among the 499920 time step data, 399840 time step data (1666h) is used as the training set, and 100080 time step data (417h) is used as the test set.

[0070] Step 1.4: In this embodiment, the sliding window is set to 1 hour to construct a dynamic Pearson correlation analysis.

[0071] For the dynamic correlation matrix formed using the sliding window method, the absolute value of the correlation coefficient in each time window is taken to reflect the degree of linear correlation between the influencing factors and the prediction indicators (every 240 data are 1 h time windows, and there are 2083 1 h time windows in total).

[0072] Then, the correlation coefficients in each time window are normalized so that their sum is 1, and the weight of each influencing factor in the time window is obtained.

[0073] Then, in each time window, the weights of the influencing factors are used to multiply their values ​​to obtain weighted values, so that high-weight factors are amplified and low-weight factors are suppressed, forming a dynamic enhancement feature matrix.

[0074] Step 2: In this embodiment, a CNN-GRU-Attention prediction model framework is constructed, such as Figure 2shown.

[0075] In this example, a 13-layer deep learning network is established, and the information of each layer is shown in Table 1.

[0076] Among them, the input data is the sequence data of 27 influencing factors in 1h (240 sampling points);

[0077] CNN feature extraction includes two-dimensional convolution layer, batch normalization layer, ReLU activation layer and two-dimensional maximum pooling layer; full connection and dimensionality reduction include the first flattening layer and the fully connected layer; data reconstruction and time series modeling include the second flattening layer and the GRU layer; feature integration and attention mechanism include the concatenation layer and the self-attention layer; output mapping and regression prediction include the fully connected layer and the regression output layer. The fully connected layer maps the 45-dimensional features extracted by the attention layer to 240 dimensions, corresponding to the prediction results of 1h, and the regression output layer outputs the prediction results of continuous values.

[0078] Table 1 Network initialization parameters

[0079]

[0080] Step 3. In this example, the improved sine algorithm, adaptive Gaussian-Cauchy mixed mutation perturbation and Bernoulli chaotic mapping are integrated to improve the dung beetle search algorithm to obtain MSADBO. The sine function and adaptive inertia weight update the position, apply Gaussian-Cauchy mixed mutation to the updated position, and apply Bernoulli chaotic perturbation to the mutated position, and finally output the global optimal solution, such as Figure 3 shown.

[0081] Use MSADBO to globally adaptively optimize model hyperparameters.

[0082] Step 3.1: In this example, there are 4 hyperparameters in the prediction model that need to be optimized, namely: learning rate, number of convolution kernels, number of neurons in GRU, and key value of attention mechanism.

[0083] Step 3.2: In this example, the population size is set to 5 and the maximum number of iterations is 15. After training, the optimal values ​​of each parameter are obtained.

[0084] Step 3.3: In this example, after training, the optimal value of each parameter is obtained.

[0085] Table 2 Initial range and optimal value of hyperparameters

[0086]

[0087] Rounded to integers, the number of convolution kernels is 25, the number of GRU neurons is 21, and the key value of the attention mechanism is 37.

[0088] Step 3.4: In this example, the optimal value of the hyperparameter is substituted into the model for training to obtain the MSADBO-CNN-GRU-Attention gas concentration multi-indicator multi-step time series prediction model.

[0089] Step 4: In this example, the model is predicted using the test set, and the prediction results are compared with the actual results to verify the prediction accuracy of the model. Figure 4 shown.

[0090] In this example, 100080 time step data is used as the test set sample, and its MSE, RMSE, MAE, MAPE, R 2 The values ​​are 0.001877, 0.043328, 0.014720, 4.419097% and 0.975077 respectively. The prediction effect of the model based on the test set data is good.

[0091] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A gas concentration time series prediction method based on MSADBO-CNN-GRU-Attention, characterized in that: Follow the steps below to implement it: Step S1: The sensor collects multi-source time series data underground and obtains multi-index data including gas concentration. The sensors are arranged at multiple monitoring points in the working face of the mine, among which the gas concentration data to be predicted is used as the target indicator, and the data collected by the remaining N monitoring points is used as the influencing indicator data. Then, Pearson correlation analysis is used to perform dynamic weighting based on a 1h time window; Step S2: Apply CNN to extract local features of the data and capture multi-scale information. Combine GRU to learn the long-term dependencies of time series data and transfer state information between time steps. Add Attention mechanism to weight the hidden state of GRU output to highlight the importance of key time nodes and improve the accuracy of multi-step prediction. Step S3: The improved sine algorithm, the adaptive Gauss-Cauchy mixed variation perturbation and the Bernoulli chaotic mapping are integrated to improve the dung beetle search algorithm to obtain MSADBO, and then the model hyperparameters are globally adaptively optimized; Step S4: Train the model, establish a prediction model based on MSADBO-CNN-GRU-Attention, and use the test set to perform multi-index and multi-step prediction of gas concentration for performance verification.

2. According to claim 1, a gas concentration time series prediction method based on MSADBO-CNN-GRU-Attention is characterized in that: The specific process of step S1 is as follows: Step S101, using Pearson coefficient to perform correlation analysis, the calculation formula is as follows: Where R is the Pearson coefficient; X i is the characteristic variable; i is the target variable; X represents the mean of the characteristic variable; σ Y represents the mean of the target variable; n represents the number of data samples; Step S102, setting the sliding window to 1h, constructing a dynamic Pearson correlation analysis matrix; Step S103, taking the absolute value of the correlation coefficient in each time window of the dynamic correlation matrix; Step S104, normalize the correlation coefficients in each time window so that their sum is 1, and obtain the weight of each influencing factor in the time window, the formula of which is: Step S105, in each time window, the weights of the influencing factors are used to multiply their values ​​to obtain weighted values, forming a dynamic enhanced feature matrix, and the weighted feature matrix is ​​used as the input of the prediction model.

3. According to claim 1, a gas concentration time series prediction method based on MSADBO-CNN-GRU-Attention is characterized in that: The specific process of step S2 is: Step S201, input multi-indicator time series data, the shape is (T, N), where T is the time step and N is the number of indicators. Perform convolution operation to extract local spatial features. Standardize the convolution output to accelerate training, apply ReLU activation function, use maximum pooling to reduce dimension and retain important features, use flattening layer to flatten multi-dimensional features into one-dimensional vectors, and use fully connected layer to map to feature representation space; Step S202, flattening the input sequence into one-dimensional features, then using GRU to extract time-dependent features, and using a feature concatenation layer to concatenate the feature vectors output by CNN and GRU; Step S203: Use the Attention mechanism to weight and map the attention output to the target dimension to obtain the final prediction output.

4. According to claim 1, a gas concentration time series prediction method based on MSADBO-CNN-GRU-Attention is characterized in that: The specific process of step S3 is: Step S301, setting the population size, maximum number of iterations, and search space boundary. Using the iteratively generated Bernoulli chaotic map to generate the position and velocity of the initial dung beetle population, generate an initial solution, then calculate the fitness value of each dung beetle individual, and record the individual historical optimum and global optimum; Step S302, updating the position by combining the sine function and the adaptive inertia weight; Step S303, applying Gauss-Cauchy hybrid mutation to the updated position; Step S304, applying Bernoulli chaotic perturbation to the mutated position; Step S305, calculate the fitness value of the new position, update the individual optimum and the global optimum, repeat steps S302 to S304 until the maximum number of iterations is reached, and output the global optimal solution.

5. The method for predicting gas concentration time series based on MSADBO-CNN-GRU-Attention according to claim 1, characterized in that: The specific process of step S4 is as follows: Step S401: Use the training set to train the model and establish a gas concentration time series prediction model based on MSADBO-CNN-GRU-Attention; Step S402: Use the test set to test the MSADBO-CNN-GRU-Attention gas concentration time series prediction model to verify the performance of the model.

6. The method for predicting gas concentration time series based on MSADBO-CNN-GRU-Attention according to claim 4 is characterized in that: In S302, the formula for updating the position by combining the sine function and the adaptive inertia weight w(t) is: Among them, A, B, C are the amplitude, frequency and phase of the sine function, and φ is the acceleration factor.

7. The method for predicting gas concentration time series based on MSADBO-CNN-GRU-Attention according to claim 4 is characterized in that: In S303, the Gauss-Cauchy mixed mutation formula is: Among them, GN is Gaussian noise, which is used for local fine search; CN is Cauchy noise, which is used for global jump. λ controls the weight decay rate to ensure that the Cauchy perturbation is dominant in the early stage of iteration and Gaussian perturbation is dominant in the later stage.

Images