Coal spontaneous combustion three-zone division method based on multilayer sensor

Abstract

The invention belongs to the technical field of coal mine safety, and aims to solve the problems that an oxygen volume fraction index division method is generally adopted at present, but the method is not accurate enough in division and difficult to deal with complex coal mine environments and dynamic changes. The invention provides a coal spontaneous combustion three-zone division method based on a multilayer perceptron, which comprises the following steps of: taking a processed coal mine area characteristic data set as input data, constructing a multilayer perceptron model, training the model by using the coal mine area characteristic data set, and then inputting real-time coal mine area characteristic data into the trained multilayer perceptron model to obtain a coal spontaneous combustion three-zone division result. According to the method, the coal spontaneous combustion three-zone region division result is obtained, the coal spontaneous combustion three-zone region is efficiently identified in a complex data environment, so that spontaneous combustion risk assessment can be performed more quickly, a visual division result is provided, field decision making is facilitated, and errors caused by human participation are reduced.

Description

Technical Field

[0001] The invention relates to the technical field of coal mine safety, and in particular to a method for dividing coal spontaneous combustion into three zones based on a multi-layer perceptron. Background Art

[0002] At present, coal resources still occupy a major position in my country's resource structure, and coal spontaneous combustion is a common safety hazard in coal mining. The spontaneous combustion of coal is caused by the fact that during the mining process of the working face, oxygen enters the goaf through the mining cracks, causing the residual coal to accumulate heat, oxidize and spontaneously combust. In order to prevent the spontaneous combustion of the residual coal in the goaf, it is necessary to divide the heat dissipation zone, oxidation zone and asphyxiation zone, namely the "three zones" of spontaneous combustion.

[0003] The traditional method of dividing the spontaneous combustion zone mainly relies on the experience of underground workers, judging the potential area of ​​coal spontaneous combustion by identifying odors through human sense of smell, or roughly dividing the goaf by building a simple physical model, and the divided area is too large. At present, the oxygen volume fraction index division method is widely used, but this method is not accurate enough and it is difficult to cope with the complex coal mine environment and dynamic changes. Therefore, it is urgent to invent a method that can efficiently complete the division of the "three zones" of spontaneous combustion in a complex environment, so as to effectively avoid the spontaneous combustion of coal seams. Summary of the invention

[0004] In order to solve the above technical problems or at least partially solve the above technical problems, the present disclosure provides a method for dividing coal spontaneous combustion into three zones based on a multi-layer perceptron.

[0005] The present invention provides a method for dividing coal spontaneous combustion into three zones based on a multi-layer perceptron, the method comprising: obtaining a coal mine regional characteristic data set, and preprocessing the coal mine regional characteristic data set to obtain standard characteristic data, wherein the coal mine regional characteristic data at least includes temperature, oxygen concentration, carbon monoxide concentration and air leakage;

[0006] Construct a multi-layer perceptron model, use the standard feature data as the model input, and the probability division results of the three zones of coal spontaneous combustion of the multi-layer perceptron model; train and verify the multi-layer perceptron model based on the above coal mine regional feature data;

[0007] The real-time coal mine area characteristic data is input into the trained multi-layer perceptron model to obtain the three-zone division results of coal spontaneous combustion.

[0008] Preferably, the multilayer perceptron model includes at least an input layer, two hidden layers and an output layer, the input nodes of the input layer are set to 4, corresponding to temperature, oxygen concentration, carbon monoxide concentration and air leakage; the first hidden layer and the second hidden layer are used to extract high-level representations in the coal mine area characteristic data; the number of neurons in the output layer is set to 3, corresponding to the heat dissipation zone, oxidation zone and asphyxiation zone in the three zones.

[0009] The optimized multilayer perceptron model outputs the probability distribution of the heat dissipation zone, the oxidation zone and the suffocation zone, and selects the zone with the highest corresponding probability as the final output zone of the model.

[0010] Preferably, if P(heat dissipation zone)>P(oxidation zone) and P(heat dissipation zone)>P(suffocation zone), then the zone is determined to belong to the heat dissipation zone;

[0011] If P(oxidation zone)>P(heat dissipation zone) and P(oxidation zone)>P(asphyxia zone), then the zone is judged to be an oxidation zone;

[0012] If P(suffocation zone)>P(heat dissipation zone) and P(suffocation zone)>P(oxidation zone), then the zone is determined to belong to the suffocation zone.

[0013] Preferably, the coal mine area feature data set is preprocessed, and the specific steps are as follows:

[0014] The mean imputation method is used to process the regional characteristic data of coal mines and replace the missing values;

[0015] Use box plots to identify and process outliers in coal mine regional characteristic data;

[0016] The Gaussian kernel density function is used to estimate the regional characteristic data of the coal mine after the above processing, and then the estimated results are logarithmically transformed;

[0017] The logarithmic transformation results are normalized to the maximum and minimum values ​​to obtain standard feature data.

[0018] Preferably, the multilayer perceptron model is trained using a training set, specifically comprising:

[0019] A loss function is used to calculate the difference between the predicted probability distribution and the true probability distribution to evaluate the prediction accuracy of the multilayer perceptron model; an optimizer is used to process the coal mine area characteristic data to automatically adjust the learning speed of the multilayer perceptron model; the training set is divided into several small training sets, and the multilayer perceptron model parameters are updated batch by batch using a small batch gradient descent method.

[0020] Preferably, the monitored coal mine area characteristic data set is divided into a training set and a validation set, specifically including: using the K-fold cross-validation method to evenly divide the coal mine area characteristic data set into K subsets, using each subset as a validation set in turn, and the rest as training sets.

[0021] After the trained multi-layer perceptron model is obtained, the method further includes: evaluating the multi-layer perceptron model according to preset indicators, and adjusting and optimizing the multi-layer perceptron model according to the evaluation results.

[0022] Preferably, the preset indicators include but are not limited to accuracy, recall, precision and F1 score; accuracy is the proportion of correctly predicted, recall is the proportion of correctly identified positive samples, precision is the proportion of predicted positive samples that are actually positive samples, and F1 score is the harmonic average of precision and recall, where:

[0023]

[0024] Where TP and TN are the number of correctly predicted positive classes and the number of correctly predicted negative classes, respectively; FP and FN are the number of incorrectly predicted positive classes and the number of incorrectly predicted negative classes, respectively.

[0025] Optimally, the positive class represents the target that the multi-layer perception model focuses on. In the classification task, each time a specific area is predicted, the area is the positive class, and other areas are the negative class. The positive and negative classes change dynamically according to the current prediction target.

[0026] Compared with the prior art, the technical solution provided by the embodiments of the present disclosure has the following advantages:

[0027] The present invention constructs a multi-layer perceptron model by taking the processed coal mine area feature data set as input data, and uses data to train the model. The trained multi-layer perceptron model is used to efficiently identify the three zones of coal spontaneous combustion in a complex data environment, thereby enabling faster spontaneous combustion risk assessment and providing intuitive division results to facilitate on-site decision-making. In addition, data training can also improve the accuracy and robustness of the model and reduce errors caused by human participation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0029] Figure 1 4 is a flow chart of a method for dividing coal spontaneous combustion into three zones based on a multi-layer perceptron according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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 those skilled in the art without creative work are within the scope of protection of the present invention.

[0031] The division of the three zones of coal spontaneous combustion is based on the existence of the coal spontaneous combustion area. According to the ventilation oxygen supply and coal combustion composite theory, the goaf is divided into different areas. The three zones include the heat dissipation zone, the oxidation zone and the asphyxiation zone. Among them, the heat dissipation zone is close to the working face, and the residual coal will not spontaneously combust; the oxidation zone is close to the heat dissipation zone, which is suitable for coal oxidation reaction and eventually leads to spontaneous combustion; the asphyxiation zone is located in the deepest part of the goaf, and the oxygen concentration is very small, which does not meet the conditions required for the spontaneous combustion of the residual coal, so the coal spontaneous combustion phenomenon will not occur. Therefore, it is necessary to accurately determine the division of the three zones of coal spontaneous combustion, so as to take appropriate measures to prevent and control coal spontaneous combustion and ensure the safety of coal mines.

[0032] At present, the division of the three zones of coal spontaneous combustion is mainly based on the experience of underground workers, using human sense of smell to identify odors to judge the potential areas of coal spontaneous combustion, or by building a simple physical model to roughly divide the goaf, and the divided areas are too large. At present, the oxygen volume fraction index division method is generally used, but this method is not accurate enough and it is difficult to cope with the complex coal mine environment and dynamic changes. The present invention provides a method for dividing the three zones of coal spontaneous combustion based on a multi-layer perceptron, which can accurately identify the three zones of coal spontaneous combustion in a complex data environment, and provide intuitive division results, so that the risk of spontaneous combustion can be more accurately assessed.

[0033] In this embodiment, a method for dividing coal spontaneous combustion into three zones based on a multi-layer perceptron is provided, which can be used in the above-mentioned mobile terminals, such as integrated computers, tablet computers, etc. Figure 1 is a flow chart of a method for dividing coal spontaneous combustion into three zones based on a multi-layer perceptron according to an embodiment of the present invention. Figure 1 As shown, the process includes the following steps:

[0034] Step S100: Acquire a coal mine area characteristic data set, and preprocess the coal mine area characteristic data set to obtain standard characteristic data, wherein the coal mine area characteristic data at least includes temperature, oxygen concentration, carbon monoxide concentration and air leakage.

[0035] In this embodiment, monitoring points are set in the coal mine goaf area, and each monitoring point is installed with a temperature sensor, an oxygen sensor, a carbon monoxide sensor and an anemometer to obtain four key characteristic data of the coal mine goaf area: temperature, oxygen concentration, carbon monoxide concentration and air leakage.

[0036] Optionally, the coal mine area characteristic data set is preprocessed, and the specific steps are: use the mean filling method to replace the missing values ​​in the coal mine area characteristic data; use the box plot to identify and process the outliers in the coal mine area characteristic data; use the Gaussian kernel density function to estimate the coal mine area characteristic data after the above processing, and then perform a logarithmic transformation on the estimated result; perform maximum and minimum normalization processing on the logarithmic transformation result to obtain standard characteristic data.

[0037] In this embodiment, the mean filling method adopts multi-feature mean filling to obtain the characteristics of each data in the coal mine area feature data set, calculate the correlation between each feature and other features, determine the weighting coefficient according to the correlation, and then use the weighted mean to fill the missing values, so that the filled data is more in line with the actual working conditions of the coal mine.

[0038] The embodiment of the present invention uses the relationship between multiple features of the coal mine regional feature data set to calculate the weighted mean, which can fill the missing values ​​more accurately. For example, in practical applications, the missing values ​​can be filled based on the correlation between oxygen concentration and temperature, combined with the weighted mean of the features of the respective data, so that the processed data is more consistent with the actual working conditions of the coal mine.

[0039] In this embodiment, the mean filling method is used to process the missing values ​​in the coal mine area characteristic data set, and the missing values ​​in the data set are replaced by the mean of the corresponding data; then the box plot is used to identify and process the outliers in the data set, and the quartiles of the data are calculated, including the lower quartile (Q1) and the upper quartile (Q3), and then the interquartile range (IQR) is obtained to determine the outlier limits, and the lower limit is calculated as Q1-1.5×IQR and the upper limit is Q3+1.5×IQR. Values ​​in the data that exceed these limits are identified as outliers, and these outliers are deleted or adjusted according to actual needs.

[0040] In this embodiment, the Gaussian kernel density function formula is: Where n is the sample size; h is the bandwidth; x is the evaluation point; x i is the data point; K is the Gaussian kernel function.

[0041] In this embodiment, the logarithmic transformation formula is: y = log (x + c), where c is the adjustment coefficient. In practical applications, c is a preset coefficient and can be set to 1 × 10 -10 , as long as it can be ensured that the problem of undefined logarithms does not occur when the data is transformed, the present invention is not limited to this.

[0042] In this embodiment, the maximum and minimum normalization is used to process the data, and the data is scaled to the interval [0, 1] to unify the feature scale. The maximum and minimum normalization formula is: Among them, x is the normalized data; x' is the normalized data; xmin is the minimum value in the data set; x max is the maximum value in the data set.

[0043] Step S200: Construct a multilayer perceptron model, which includes at least an input layer, two hidden layers and an output layer. The input layer is used to receive the standard feature data, the hidden layer is used to process the standard feature data, and the output layer is used to determine the probability division results of the three zones of coal spontaneous combustion based on the data processing results of the hidden layer.

[0044] Optionally, a multilayer perceptron model is constructed, specifically including: the input nodes of the input layer are set to 4, corresponding to temperature, oxygen concentration, carbon monoxide concentration and air leakage; the number of neurons in the first hidden layer is set to 32, and the number of neurons in the second hidden layer is set to 16, which are used to extract high-level representations in the coal mine area characteristic data; the number of neurons in the output layer is set to 3, corresponding to the heat dissipation zone, oxidation zone and asphyxiation zone in the three zones.

[0045] In this embodiment, the multilayer perceptron model consists of an input layer, two hidden layers and an output layer. The input layer receives standard feature data. The number of input nodes is 4, corresponding to temperature, oxygen concentration, carbon monoxide concentration and air leakage respectively. The number of neurons in the first hidden layer is set to 32, and the number of neurons in the second hidden layer is set to 16. The high-level representation of the features is further extracted, and the ReLU activation function is used to capture the nonlinear relationship of the features. The number of neurons in the output layer is 3, corresponding to the heat dissipation zone, oxidation zone and asphyxiation zone respectively. The Softmax activation function is used to output the probability of each category, and the output is converted into a probability distribution to achieve the probability division of the three zones.

[0046] In this embodiment, the high-level representation formula is x = [x1, x2, x3, x4], where x1 is temperature, x2 is oxygen concentration, x3 is carbon monoxide concentration, and x4 is air leakage. There are 32 neurons in the first hidden layer, each of which performs weighted summation on the input through a weight matrix W1 and a bias vector b1, and then applies the ReLU activation function to capture the nonlinear relationship between each feature. The weighted summation formula is: h1 = RELU (W1 x + b1), where h1 is the output of the first hidden layer, that is, the high-level representation of the first hidden layer; W1 is a weight matrix of size 32×4; b1 is a bias vector of size 32×1.

[0047] In this embodiment, the formula of the ReLU activation function is: f(x)=max(0,x), where f(x) is the output after activation and x represents the input value.

[0048] In this embodiment, the second hidden layer has 16 neurons, and further feature extraction is performed on the output of the first hidden layer. The formula is: h2=RELU(W2·h1+b2), where h2 is the output of the second hidden layer, that is, the high-level representation of the second hidden layer; W2 is a weight matrix of size 16×4; b2 is a 16×1 bias vector.

[0049] In this embodiment, the output layer has 3 neurons, corresponding to the heat dissipation zone, oxidation zone and suffocation zone, and the calculation formula is:

[0050] z=W3·h2+b3, where z is the original score of the output layer, i.e., logits, W3 is a weight matrix of size 3×16, and b3 is a bias vector of size 3×1. After the logits of the output layer pass through the Softmax activation function, the probabilities corresponding to the heat dissipation zone, oxidation zone, and suffocation zone are output.

[0051] In this embodiment, each neuron will output a logits, which represents the activation intensity of the class. In order to convert these scores into probabilities, the Softmax activation function is used. Output the probability distribution and select the band with high corresponding probability as the final output band. The logits output by the model will be converted into three probability values ​​P (heat dissipation zone), P (oxidation zone), and P (suffocation zone), and the sum of these probability values ​​is 1. The model will output a probability value belonging to each spontaneous combustion zone, and the output of the model will clearly indicate which spontaneous combustion zone each data point or coal mine area belongs to. The specific judgment principles are:

[0052] If P(heat dissipation zone)>P(oxidation zone) and P(heat dissipation zone)>P(suffocation zone), then the position is determined to belong to the heat dissipation zone.

[0053] If P(oxidation zone)>P(heat dissipation zone) and P(oxidation zone)>P(suffocation zone), then the position is determined to belong to the oxidation zone.

[0054] If P(suffocation zone)>P(heat dissipation zone) and P(suffocation zone)>P(oxidation zone), then the position is determined to belong to the suffocation zone.

[0055] In this embodiment, the Softmax activation function formula is: Among them, P(y i ) is the probability of the i-th category, that is, the probability of a spontaneous combustion zone; z i are the logits corresponding to the i-th category; It indicates that there are three categories, corresponding to the heat dissipation zone, oxidation zone, and asphyxiation zone.

[0056] In this embodiment, the nonlinear relationship of features captured by the ReLU activation function includes nonlinear feature mapping, for example: 1. Nonlinear transformation of linear combination: neurons in each layer process input through weighted sum and activation function. 2. Feature interaction: Through the multi-layer neural network structure, the multi-layer perceptron model can automatically capture the complex interaction relationship between features.

[0057] In this embodiment, the nonlinear relationship between temperature, oxygen concentration, carbon monoxide concentration and air leakage can be learned through nonlinear mapping of hidden layer neurons. Specifically, as follows:

[0058] Nonlinear relationship between temperature and oxygen concentration: The effect of temperature increase and the acceleration of oxygen consumption. In the spontaneous combustion area of ​​coal mines, there is a significant nonlinear relationship between temperature and oxygen concentration. Generally, as the temperature increases, the oxygen consumption rate accelerates, leading to the formation of oxidation zones and asphyxiation zones. Through the ReLU activation function, the multi-layer perceptron model can learn a nonlinear pattern in which oxygen consumption is slower at lower temperatures, but the oxygen concentration drops sharply after the temperature rises to a certain threshold.

[0059] Nonlinear relationship between temperature and carbon monoxide concentration: Carbon monoxide is an early indicator gas of the spontaneous combustion process. The carbon monoxide concentration will gradually increase in the early stage of coal spontaneous combustion (i.e., the oxidation zone). As the temperature further increases, the amount of carbon monoxide generated increases rapidly. The multi-layer perceptron model can capture the complex relationship between carbon monoxide concentration and temperature through nonlinear activation functions. (For example, when the coal temperature is low, the carbon monoxide concentration is not significant, but when the temperature approaches the critical temperature of spontaneous combustion, the carbon monoxide concentration will increase sharply, forming a nonlinear "peak effect.")

[0060] Multi-feature interactive nonlinear relationship: The combined effect of temperature, oxygen concentration and carbon monoxide concentration: The increase in temperature will accelerate the oxidation reaction, resulting in a decrease in oxygen concentration and an increase in carbon monoxide concentration; while the decrease in oxygen concentration will inhibit the further intensification of the oxidation reaction. The multi-layer perceptron model can capture this complex interaction through the neuron connection weights and ReLU activation function in the hidden layer, helping the multi-layer perceptron model to judge the progress of the spontaneous combustion process.

[0061] Coupling effect between temperature and air leakage: The increase in coal seam temperature will accelerate the spontaneous combustion reaction, while the change in air leakage will affect the heat dissipation and oxygen supply of the coal seam, resulting in different spontaneous combustion rates in different areas. The multi-layer perceptron model can identify the impact of air leakage on the formation of spontaneous combustion zones of different coals through this nonlinear relationship.

[0062] Nonlinear distribution characteristics of the three zones of coal spontaneous combustion: Nonlinear boundaries of the heat dissipation zone, oxidation zone and asphyxiation zone: As the coal temperature increases, the spontaneous combustion area transitions from the heat dissipation zone to the oxidation zone and then to the asphyxiation zone. The multilayer perceptron model can capture the complex transformation process between these three zones through nonlinear transformations of multiple hidden layers, and output the probability of each zone through the final Softmax layer.

[0063] Step S300: Divide the coal mine area feature data set into a training set and a validation set, use the training set to train the multilayer perceptron model, and then use the validation set to determine the optimal parameters to obtain a trained multilayer perceptron model.

[0064] In this embodiment, the relevant parameters of the multilayer perceptron model during model training can be set to: learning rate is 0.001, batch size is 32, and number of training rounds is 50. At the same time, L2 regularization is used to prevent overfitting, and the Xavier initialization method is used to ensure that the initial weight is reasonable.

[0065] In this embodiment, the coal mine area characteristic data set is divided into a training set and a validation set according to 80% and 20%, and random noise is added to the data training set, and a Gaussian noise filter is used to smooth the data.

[0066] Optionally, the training set is divided into several small training sets, and the multilayer perceptron model parameters are updated batch by batch using a small batch gradient descent method.

[0067] Optionally, the coal mine area feature dataset is evenly divided into K subsets using a K-fold cross-validation method, and each subset is used as a validation set in turn, and the rest are used as training sets.

[0068] Optionally, the validation set is used to determine the best parameters, specifically including: using the accuracy, recall, precision, and F1 score of the validation set to determine the best parameters.

[0069] Optionally, the multilayer perceptron model is trained using a training set, specifically including: using a loss function to calculate the difference between the predicted probability distribution and the true probability distribution to evaluate the prediction accuracy of the multilayer perceptron model; using an optimizer to process coal mine area characteristic data to automatically adjust the learning speed of the multilayer perceptron model.

[0070] In this embodiment, the optimizer can select the Adam optimizer, which converges quickly due to its adaptive learning rate characteristics and is suitable for processing such complex data. The optimizer update rule is: t =β1m t-1 +(1-β1)g t ,

[0071] Among them, m t is the first-order momentum of the gradient; Vt is the second-order momentum of the gradient; β1 and β2 are the exponential decay rates of the first-order and second-order momentum, which are 0.9 and 0.999 respectively; g t is the current gradient; and are the momentum and second-order moment after bias correction, respectively; α is the learning rate, which is 0.001; θ t is the updated parameter, θ represents all trainable parameters in the multi-layer perceptual model, that is, the set of weight matrices and bias vectors of all layers in the multi-layer perceptual model.

[0072] Step S400: inputting the real-time coal mine area characteristic data into the trained multi-layer perceptron model to obtain the three-zone division result of coal spontaneous combustion.

[0073] The embodiment of the present invention constructs a multilayer perceptron model by taking the processed coal mine area feature data set as input data, and uses data to train the model. The trained multilayer perceptron model is used to efficiently identify the three zones of coal spontaneous combustion in a complex data environment, thereby enabling faster spontaneous combustion risk assessment and providing intuitive division results to facilitate on-site decision-making. In addition, data training also improves the accuracy and robustness of the model, reducing errors caused by human participation.

[0074] Step S500: Evaluate the multi-layer perception model according to preset indicators, and adjust and optimize the multi-layer perception model according to the evaluation results.

[0075] In this embodiment, the performance of the multi-layer perception model is evaluated by preset indicators such as accuracy, recall rate, precision and F1 score, and the model is adjusted and optimized.

[0076] In this embodiment, the accuracy rate is the proportion of correctly predicted samples, the recall rate is the proportion of correctly identified positive samples, the precision rate is the proportion of predicted positive samples that are actually positive samples, and the F1 score is the harmonic mean of the precision rate and the recall rate, where:

[0077]

[0078] Where TP and TN are the number of correctly predicted positive classes and the number of correctly predicted negative classes, respectively; FP and FN are the number of incorrectly predicted positive classes and the number of incorrectly predicted negative classes, respectively.

[0079] In this embodiment, the positive class represents the target that the multi-layer perception model focuses on, that is, the specific category that you want to identify. In the classification task, each time a specific area is predicted, the area is the positive class, and other areas are the negative class. The positive class and the negative class change dynamically according to the current prediction target.

[0080] For example, if we are currently focusing on the oxidation zone, the oxidation zone is a positive class, and the other two zones are negative classes; if we are currently focusing on the heat dissipation zone, the heat dissipation zone is a positive class, and the other two zones are negative classes.

[0081] TP: The multi-layer perception model predicts that the sample is positive (oxidation zone), and the sample is actually positive (in the oxidation zone).

[0082] TN: The multi-layer perception model predicts that the sample is a negative class (i.e., a cooling zone or a suffocation zone), and the sample is actually a negative class (in the cooling zone or the suffocation zone).

[0083] FP: The multi-layer perception model predicts that the sample is a positive class (oxidation zone), but the sample is actually a negative class (that is, it is actually a cooling zone or a suffocation zone).

[0084] FN: The multi-layer perception model predicts that the sample is a negative class (i.e., a heat dissipation zone or a suffocation zone), but the sample is actually a positive class (i.e., it is actually an oxidation zone).

[0085] The embodiment of the present invention evaluates the trained multilayer perceptron model through preset indicators, and further optimizes and adjusts the constructed model, so that the multilayer perceptron model can more accurately obtain the division results of the three zones of coal spontaneous combustion under complex environments.

[0086] In order to verify the effectiveness and feasibility of the method of the present invention, the present invention conducts experiments by taking data collected in a coal mine area as an example.

[0087] The collected data include: temperature (T): 50℃; oxygen concentration (O2): 16%; carbon monoxide concentration (CO): 0.05%; air leakage (Q): 12m 3 / min. Through the multi-layer perceptron model designed by the present invention, the first hidden layer extracts the following nonlinear feature combinations:

[0088] Feature 1: When the temperature rises, the oxygen concentration drops rapidly (non-linear relationship).

[0089] Feature 2: The carbon monoxide concentration increases nonlinearly with increasing temperature and air leakage.

[0090] Feature 3: The critical transition point between the oxidation zone and the asphyxiation zone at a certain temperature.

[0091] The above features are further combined in the second hidden layer, and the final output probability distribution is: heat dissipation zone: 0.1; oxidation zone: 0.7; suffocation zone: 0.2.

[0092] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A method for dividing coal spontaneous combustion into three zones based on a multi-layer perceptron, characterized in that: The method comprises: Acquire a coal mine area characteristic data set, and preprocess the coal mine area characteristic data set to obtain standard characteristic data, wherein the coal mine area characteristic data at least includes temperature, oxygen concentration, carbon monoxide concentration, and air leakage; A multi-layer perceptron model was constructed, and the standard feature data was used as the input of the model. The probability division results of the three zones of coal spontaneous combustion of the multi-layer perceptron model were obtained. The multi-layer perceptron model was trained and verified based on the training set composed of the above coal mine regional feature data. The real-time coal mine area characteristic data is input into the trained multi-layer perceptron model to obtain the three-zone division results of coal spontaneous combustion.

2. The method for dividing coal spontaneous combustion into three zones based on multi-layer perceptron according to claim 1 is characterized in that: The multilayer perceptron model includes at least an input layer, two hidden layers and an output layer. The input nodes of the input layer are set to 4, corresponding to temperature, oxygen concentration, carbon monoxide concentration and air leakage; the first hidden layer and the second hidden layer are used to extract high-level representations in the coal mine area feature data; the number of neurons in the output layer is set to 3, corresponding to the heat dissipation zone, oxidation zone and asphyxiation zone in the three zones.

3. The method for dividing coal spontaneous combustion into three zones based on multi-layer perceptron according to claim 2 is characterized in that: The multi-layer perceptron model outputs the probability distribution of the heat dissipation zone, oxidation zone, and suffocation zone, and selects the zone with the highest corresponding probability as the final output zone of the model.

4. The method for dividing coal spontaneous combustion into three zones based on multi-layer perceptron according to claim 3 is characterized in that: If P(heat dissipation zone)>P(oxidation zone) and P(heat dissipation zone)>P(suffocation zone), then the zone is judged to belong to the heat dissipation zone; If P(oxidation zone)>P(heat dissipation zone) and P(oxidation zone)>P(asphyxia zone), then the zone is judged to be an oxidation zone; If P(suffocation zone)>P(heat dissipation zone) and P(suffocation zone)>P(oxidation zone), then the zone is determined to belong to the suffocation zone.

5. The method for dividing coal spontaneous combustion into three zones based on multi-layer perceptron according to claim 1 is characterized in that: The coal mine area feature data set is preprocessed, and the specific steps are as follows: The mean imputation method is used to process the regional characteristic data of coal mines and replace the missing values; Use box plots to identify and process outliers in coal mine regional characteristic data; The Gaussian kernel density function is used to estimate the regional characteristic data of the coal mine after the above processing, and then the estimated results are logarithmically transformed; The logarithmic transformation results are normalized to the maximum and minimum values ​​to obtain standard feature data.

6. The method for dividing coal spontaneous combustion into three zones based on multi-layer perceptron according to claim 1 is characterized in that: The multilayer perceptron model is trained using the training set, specifically including: A loss function is used to calculate the difference between the predicted probability distribution and the true probability distribution to evaluate the prediction accuracy of the multilayer perceptron model; an optimizer is used to process the coal mine area characteristic data to automatically adjust the learning speed of the multilayer perceptron model; the training set is divided into several small training sets, and the multilayer perceptron model parameters are updated batch by batch using a small batch gradient descent method.

7. The method for dividing coal spontaneous combustion into three zones based on multi-layer perceptron according to claim 1 is characterized in that: The monitored coal mine area characteristic data set is divided into a training set and a validation set, including: The K-fold cross-validation method is used to divide the coal mine area feature dataset into K subsets on average. Each subset is used as a validation set in turn, and the rest are used as training sets. The accuracy, recall, precision and F1 score of the validation set are used to determine the optimal parameters of the multilayer perceptron model.

8. The method for dividing coal spontaneous combustion into three zones based on multi-layer perceptron according to claim 1 is characterized in that: After obtaining the trained multi-layer perceptron model, the method further includes: evaluating the multi-layer perceptron model according to preset indicators, and adjusting and optimizing the multi-layer perceptron model according to the evaluation results.

9. The method for dividing coal spontaneous combustion into three zones based on multi-layer perceptron according to claim 8 is characterized in that: The preset indicators include but are not limited to accuracy, recall, precision and F1 score; accuracy is the proportion of correctly predicted, recall is the proportion of correctly identified positive samples, precision is the proportion of predicted positive samples that are actually positive samples, and F1 score is the harmonic mean of precision and recall, where: Where TP and TN are the number of correctly predicted positive classes and the number of correctly predicted negative classes, respectively; FP and FN are the number of incorrectly predicted positive classes and the number of incorrectly predicted negative classes, respectively.

10. The method for dividing coal spontaneous combustion into three zones based on multi-layer perceptron according to claim 9, characterized in that: The positive class represents the target that the multi-layer perception model focuses on. In the classification task, each time a specific area is predicted, the area is the positive class, and other areas are the negative class. The positive and negative classes change dynamically according to the current prediction target.

Images