CNN (Convolutional Neural Network) fused Transform working face gas emission anomaly identification and dynamic early warning algorithm
A gas outburst anomaly identification model was constructed by using CNN and Transformer deep learning algorithms, which solved the problem of identifying and warning of gas outburst anomalies in coal mines. This model enables rapid and accurate identification and warning of gas outbursts, thereby reducing the risk of gas disasters.
Patent Information
- Application Number
- CN202411500639.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-25
- Publication Date
- 2025-01-10
AI Technical Summary
Existing technologies are insufficient to effectively identify and provide early warning of abnormal gas outbursts at coal mining faces, especially in high-gas and outburst-prone mines, where gas vents and abnormal outbursts pose a serious threat to miners' safety.
A deep learning algorithm that combines CNN and Transformer, along with a convolutional neural network and a Transformer encoder, is used to construct a gas outburst anomaly identification model. Through data preprocessing, model training, and real-time data analysis, the model can quickly and accurately identify and warn of abnormal gas outbursts.
It enables real-time, rapid, and accurate identification and early warning of gas outbursts at coal mine working faces, improving the accuracy of gas anomaly identification and the timeliness of early warning, and reducing the risk of gas disasters.
Smart Images
Figure CN119293693A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of gas disaster prevention and control, and relates to a CNN-fused Transformer working face gas outburst anomaly recognition and dynamic early warning algorithm. Background Art
[0002] Gas is a common dangerous gas in coal mines. Especially in high-gas and outburst mines, gas blowholes and abnormal gas outflow are very likely to occur, posing a great threat to the safety of miners. Abnormal gas outflow is an important factor leading to gas disasters and seriously threatens the safe production of mines.
[0003] Abnormal identification and dynamic warning of gas outburst in coal mining working faces is one of the key technologies for predicting and warning of gas outburst hazards. Abnormal gas outburst in coal mines mainly includes phenomena such as gas fluctuations, abnormal peak trend, abnormal trend during non-operating periods, abnormal frequency, abnormal peak-to-average ratio, and abnormal wave width ratio. Coal mine gas outburst data is time series data with strong temporal characteristics. The current research mainly uses neural network learning methods to predict gas outburst data. Its multi-layer network can be used to automatically extract features from the original data, and has better fitting ability under the same parameter order. Summary of the invention
[0004] The purpose of the present invention is to provide a CNN fusion Transformer working face gas emission abnormality recognition and dynamic early warning algorithm, which can recognize and warn the abnormal gas emission data of coal mine mining working face in real time, quickly and accurately.
[0005] The technical solution adopted by the present invention is that CNN integrates the Transformer working face gas outburst anomaly recognition and dynamic early warning algorithm, including the following steps: Step 1: Obtaining gas outburst data of working face; Step 2: Preprocessing of gas outburst data; Step 3: Construct a CNN-Transformer gas abnormal emission recognition model; Step 4: Train the CNN-Transformer gas abnormal emission recognition model; Step 5: The CNN-Transformer abnormal gas emission recognition model recognizes abnormal gas emission online.
[0006] The present invention is also characterized in that: In step 1, the coal mine working face gas emission data obtained include: gas concentration, air volume, excavation or mining mileage and tunnel cross section.
[0007] Step 2 is as follows: Step 2.1, prepare data: the gas outburst data of the working face obtained in step 1 is used as the data set, and the data set is divided into a training set and a test set; Step 2.2, calculate the distance: use Euclidean distance or other distance measurement methods to calculate the distance between each point in the training set and each point in the test set; Step 2.3, find the nearest k points: For each point in the test set, find the nearest k points in the training set; Step 2.4, calculate the mean: For each missing value, use the mean of the k neighbors found for interpolation; Step 2.5: Apply the interpolation results: Use the interpolated dataset for model training and prediction.
[0008] In step 3, the CNN-Transformer gas abnormal emission recognition model includes a convolutional layer, a Transformer encoder, and an MLP layer. The functions of each layer are as follows: 1) Convolutional layer: Use a 1-dimensional CNN convolution kernel to perform convolution operations on the time series, project each local window into an embedding vector, which is regarded as a token. Each token carries the short-term time series features extracted by the convolution kernel on the corresponding local time window; 2) Transformer encoder: The embedding vector Token output by the convolutional layer is added with the position embedding of the corresponding position and then input into the Transformer encoder. The Transformer encoder is composed of multiple Transformer encoding blocks. Each Transformer encoding block includes two sublayers, which are sequentially set as a multi-head self-attention layer and a fully connected layer. The data needs to be layer-normalized before entering each sublayer, and the sublayers are connected by residual connections. After passing through the Transformer encoder, the long-term dependencies between these Tokens are extracted and output by the Transformer encoder as a hidden embedding vector. The hidden embedding vector contains the short-term sequence features extracted from the convolutional layer and the long-term sequence features extracted from the Transformer encoder. 3) MLP layer: The hidden embedding vector output by the Transformer encoder eventually enters a multi-layer perceptron with a Softmax activation function, and finally outputs real-time data on gas outflow in the next 30 minutes.
[0009] In step 3, the configuration and hyperparameters of the CNN-Transformer gas abnormal outburst recognition model are shown in Table 1: Table 1 CNN-Transformer gas abnormal emission recognition model configuration
[0010] .
[0011] In step 4, the cross entropy loss function is used as the model loss function to train the CNN-Transformer gas abnormal outburst recognition model in step 3, and the gas outburst data preprocessed in step 2 is used to train the model. AdamW is used as the model optimizer to solve the parameters in the model. During training, the loss value of the cross entropy loss function decreases with the increase in the number of training times; AdamW uses L2 regularization and weight decay regularization to perform stochastic gradient descent on the CNN-Transformer gas abnormal outburst recognition model until the model loss value is less than the preset value.
[0012] Step 5 is as follows: input the real-time gas concentration data into the CNN-Transformer gas abnormal emission recognition model trained in step (4), use the model to predict the gas emission data of the working face in the next 30 minutes, establish three warning levels: general warning, yellow warning, and red warning, calculate the difference between the current gas concentration and the model prediction value to establish the warning threshold, as shown in Table 2; at the same time, different gas concentration sensor measurement points on the working face are separately set with general warning and over-limit thresholds. If the increase in the time series data in the next 30 minutes is greater than 0.2, a gas hazard warning is issued; Table 2 Warning threshold relationship
[0013] .
[0014] The beneficial effects of the present invention are: The algorithm of the present invention applies CNN and Transformer deep learning technology to predict gas outbursts in mining working faces, and establishes an abnormal recognition model to assess the risk of gas outbursts. According to the time series characteristics of gas outburst data, K nearest neighbor interpolation is used to complete the gas outburst data to solve the problem of data loss caused by power outages and transmission. The CNN and Transformer model architecture is used for optimization to identify and predict abnormalities in gas outburst data. The CNN-Transformer gas outburst abnormality recognition model is used to predict gas concentration data within the next 30 minutes. The difference between the future predicted value and the current true value is calculated to establish the gas outburst abnormality threshold relationship, and the model is used to warn of abnormal gas outbursts. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 This is a flow chart of the CNN-fused Transformer working face gas outburst anomaly recognition and dynamic early warning algorithm of the present invention; Figure 2This is a structural diagram of the CNN-Transformer gas abnormal emission recognition model of the present invention; Figure 3 The loss drop graph of the cross entropy loss function for model training of the present invention. DETAILED DESCRIPTION
[0016] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0017] The present invention provides a CNN-fused Transformer working face gas outburst anomaly recognition and dynamic early warning algorithm, such as Figure 1 As shown, the following steps are included: Step 1: Obtaining gas outburst data of working face; The coal mine working face gas outburst data obtained include: gas concentration, air volume, excavation or mining mileage and tunnel cross-section, these structured and unstructured data. The structured data is accessed using MQTT, and the real-time time series data stream is obtained through the coal mine safety monitoring system and the coal mine gas extraction system. The structured data includes: gas concentration, air volume and extraction volume; the data format uses the Json string format; the unstructured data includes: tunnel cross-section and excavation or mining mileage, and the unstructured data is entered manually.
[0018] Gas outburst refers to: gas exhaust volume = gas concentration × air volume × cross-section (fixed value).
[0019] Step 2: Preprocessing of gas outburst data; The coal mine gas online monitoring system realizes real-time transmission and collection of data by installing sensors underground. However, the underground working environment is complex, and the monitoring system is susceptible to many factors such as the power system, data transmission network, and sensor equipment, which often cause failures, resulting in the loss of some gas concentration data. Therefore, it is necessary to reasonably deal with the missing data. The present invention uses K nearest neighbor interpolation to fill in the missing gas data, ensure data quality, and improve the accuracy of gas anomaly analysis and identification.
[0020] K nearest neighbor interpolation identifies adjacent points through distance measurement and uses the complete values of adjacent observations to estimate missing values. It identifies K spatially similar or close samples in the data set and then uses these samples to estimate the values of missing data points. The missing value of each sample is interpolated using the average value of the K fields found in the data set. The K nearest neighbor interpolation algorithm is used to preprocess the gas outburst data. The specific steps are as follows: Step 2.1, prepare data: Take the gas outburst data of the working face obtained in step 1 as a data set (all types of data are one data set, and all data are put into a matrix to calculate the distance; such as the gas concentration, air volume, extraction volume, and footage of a certain tunnel in 3 months), and divide the data set into a training set and a test set; Step 2.2, calculate the distance: use Euclidean distance or other distance measurement methods to calculate the distance between each point in the training set and each point in the test set; Step 2.3, find the nearest k points: For each point in the test set, find the nearest k points in the training set; Step 2.4, calculate the mean: For each missing value, use the mean of the k neighbors found for interpolation; Step 2.5: Apply the interpolation results: Use the interpolated dataset for model training and prediction.
[0021] Step 3: Construct a CNN-Transformer gas abnormal emission recognition model; In step 3, the CNN-Transformer gas abnormal outburst recognition model: combines the local feature extraction capability of the convolutional neural network (CNN) and the long time series modeling advantage of the Transformer to improve the model's ability to analyze time series data streams and solve complex problems that are difficult to handle with traditional single network structures. It uses a self-attention mechanism to model dependencies in sequences, avoiding sequential calculations in traditional recurrent neural networks.
[0022] In step 3, the CNN-Transformer gas abnormal outburst recognition model includes a convolutional layer, a Transformer encoder (Encoder), and an MLP layer (multi-layer perceptron layer), such as Figure 2 As shown, the functions of each layer of the structure are as follows: 1) Convolution layer: Use a 1-dimensional CNN convolution kernel to perform convolution operations on the time series, projecting each local window into an embedding vector. The embedding vector is regarded as a token. Each token carries the short-term time series features extracted by the convolution kernel on the corresponding local time window. 2) Transformer Encoder: The embedding vector Token output by the convolutional layer is added with the position embedding of the corresponding position and then input into the Transformer Encoder. The Transformer Encoder is composed of multiple Transformer Encoding Blocks. Each Transformer Encoding Block includes two sub-layers, which are sequentially set as a multi-head self-attention layer and a fully connected layer. Before the data enters each sub-layer, layer normalization is required, and the sub-layers are connected with residual connections. After the Token passes through the Transformer Encoder, the long-term dependency between these Tokens is extracted and the Token is output as a hidden embedding vector by the Transformer Encoder. The hidden embedding vector contains the short-term sequence features extracted from the convolutional layer and the long-term sequence features extracted from the Transformer Encoder. 3) MLP layer: The hidden embedding vector output by the Transformer encoder eventually enters a multi-layer perceptron with a Softmax activation function, and finally outputs real-time data on gas outflow in the next 30 minutes.
[0023] The configuration and hyperparameters of the CNN-Transformer gas abnormal outburst recognition model are shown in Table 1: Table 1 CNN-Transformer gas abnormal emission recognition model configuration
[0024] Step 4: Train the CNN-Transformer gas abnormal emission recognition model; The present invention uses the cross-entropy loss function as the model loss function to train the CNN-Transformer gas abnormal emission recognition model in step 3, uses the pre-processed gas emission data in step 2 to train the model, and uses AdamW as the model optimizer to solve the parameters in the model. During the training, the loss value of the cross-entropy loss function decreases with the increase of the number of training times, such as Figure 3 As shown in the figure, AdamW uses L2 regularization and weight decay regularization to perform stochastic gradient descent on the CNN-Transformer gas abnormal outburst recognition model until the model loss value is less than the preset value.
[0025] Step 5: The CNN-Transformer gas abnormal emission recognition model recognizes the gas abnormal emission online; The real-time gas concentration data of the coal mine safety monitoring system is accessed, and the real-time gas concentration data is input into the CNN-Transformer gas abnormal emission recognition model trained in step (4). The CNN-Transformer gas abnormal emission recognition model is packaged into a computer program module. The program module has a time series data recognition input interface and an output interface, which can be stored and used on any computer device. For the time series data of gas emission of the working face that needs to be identified, the computer calls the program time series recognition input interface to input data. After the program operation, the real-time gas emission data of the working face in the next 30 minutes can be output on the output interface.
[0026] The model is used to predict the gas outburst data of the working face in the next 30 minutes, and three warning levels are established: general warning, yellow warning, and red warning. The difference between the current gas concentration and the model prediction value is calculated to establish the warning threshold, as shown in Table 2. At the same time, different gas concentration sensor measuring points on the working face are separately set with general warning and over-limit thresholds. If the increase in the time series data in the next 30 minutes is greater than 0.2, a gas hazard warning will be issued.
[0027] Table 2 Warning threshold relationship
[0028] Example 1 The CNN-fused Transformer working face gas outburst anomaly recognition and dynamic warning algorithm includes the following steps: Step 1: Obtaining gas outburst data of working face; Step 2: Preprocessing of gas outburst data; Step 3: Construct a CNN-Transformer gas abnormal emission recognition model; Step 4: Train the CNN-Transformer gas abnormal emission recognition model; Step 5: The CNN-Transformer abnormal gas emission recognition model recognizes abnormal gas emission online.
[0029] Example 2 The CNN-fused Transformer working face gas outburst anomaly recognition and dynamic warning algorithm includes the following steps: Step 1: Obtaining gas outburst data of working face; In step 1, the coal mine working face gas emission data obtained include: gas concentration, air volume, excavation or mining mileage and tunnel cross section.
[0030] Step 2: Preprocessing of gas outburst data; Step 3: Construct a CNN-Transformer gas abnormal emission recognition model; Step 4: Train the CNN-Transformer gas abnormal emission recognition model; Step 5: The CNN-Transformer abnormal gas emission recognition model recognizes abnormal gas emission online.
[0031] Example 3 The CNN-fused Transformer working face gas outburst anomaly recognition and dynamic warning algorithm includes the following steps: Step 1: Obtaining gas outburst data of working face; In step 1, the coal mine working face gas emission data obtained include: gas concentration, air volume, excavation or mining mileage and tunnel cross section.
[0032] Step 2: Preprocessing of gas outburst data; Step 2 is as follows: Step 2.1, prepare data: the gas outburst data of the working face obtained in step 1 is used as the data set, and the data set is divided into a training set and a test set; Step 2.2, calculate the distance: use Euclidean distance or other distance measurement methods to calculate the distance between each point in the training set and each point in the test set; Step 2.3, find the nearest k points: For each point in the test set, find the nearest k points in the training set; Step 2.4, calculate the mean: For each missing value, use the mean of the k neighbors found for interpolation; Step 2.5: Apply the interpolation results: Use the interpolated dataset for model training and prediction.
[0033] Step 3: Construct a CNN-Transformer gas abnormal emission recognition model; Step 4: Train the CNN-Transformer gas abnormal emission recognition model; Step 5: The CNN-Transformer abnormal gas emission recognition model recognizes abnormal gas emission online.
Claims
1. CNN fusion Transformer working face gas outburst anomaly recognition and dynamic warning algorithm, characterized by: The following steps are involved: Step 1: Obtaining gas outburst data of working face; Step 2: Preprocessing of gas outburst data; Step 3: Construct a CNN-Transformer gas abnormal emission recognition model; Step 4: Train the CNN-Transformer gas abnormal emission recognition model; Step 5: The CNN-Transformer abnormal gas emission recognition model recognizes abnormal gas emission online.
2. The CNN fusion Transformer working face gas outburst anomaly recognition and dynamic early warning algorithm according to claim 1 is characterized in that: In step 1, the coal mine working face gas emission data obtained include: gas concentration, air volume, excavation or mining mileage and tunnel cross section.
3. The CNN fusion Transformer working face gas outburst anomaly recognition and dynamic early warning algorithm according to claim 2 is characterized in that: Step 2 is as follows: Step 2.1, prepare data: the gas outburst data of the working face obtained in step 1 is used as the data set, and the data set is divided into a training set and a test set; Step 2.2, calculate the distance: use Euclidean distance or other distance measurement methods to calculate the distance between each point in the training set and each point in the test set; Step 2.3, find the nearest k points: For each point in the test set, find the nearest k points in the training set; Step 2.4, calculate the mean: For each missing value, use the mean of the k neighbors found for interpolation; Step 2.5: Apply the interpolation results: Use the interpolated dataset for model training and prediction.
4. The CNN fusion Transformer working face gas outburst anomaly recognition and dynamic early warning algorithm according to claim 3 is characterized in that: In step 3, the CNN-Transformer gas abnormal outburst recognition model includes a convolutional layer, a Transformer encoder, and an MLP layer. The functions of each layer are as follows: 1) Convolution layer: Use a 1-dimensional CNN convolution kernel to perform convolution operations on the time series, project each local window into an embedding vector, which is regarded as a token. Each token carries the short-term time series features extracted by the convolution kernel on the corresponding local time window; 2) Transformer encoder: The embedding vector Token output by the convolutional layer is added with the position embedding of the corresponding position and then input into the Transformer encoder. The Transformer encoder is composed of multiple Transformer encoding blocks. Each Transformer encoding block includes two sublayers, which are sequentially set as a multi-head self-attention layer and a fully connected layer. The data needs to be layer-normalized before entering each sublayer, and the sublayers are connected by residual connections. After passing through the Transformer encoder, the long-term dependencies between these Tokens are extracted and output by the Transformer encoder as a hidden embedding vector. The hidden embedding vector contains the short-term sequence features extracted from the convolutional layer and the long-term sequence features extracted from the Transformer encoder. 3) MLP layer: The hidden embedding vector output by the Transformer encoder eventually enters a multi-layer perceptron with a Softmax activation function, and finally outputs real-time data on gas outflow in the next 30 minutes.
5. The CNN fusion Transformer working face gas outburst anomaly recognition and dynamic early warning algorithm according to claim 4 is characterized in that: In step 3, the configuration and hyperparameters of the CNN-Transformer gas abnormal outburst recognition model are shown in Table 1: Table 1 CNN-Transformer gas abnormal emission recognition model configuration 。 6. The CNN fusion Transformer working face gas outburst anomaly recognition and dynamic early warning algorithm according to claim 4 is characterized in that: In step 4, the cross entropy loss function is used as the model loss function to train the CNN-Transformer gas abnormal emission recognition model in step 3. The gas emission data preprocessed in step 2 is used to train the model. AdamW is used as the model optimizer to solve the parameters in the model. During the training, the loss value of the cross entropy loss function decreases with the increase of the number of training times. AdamW uses L2 regularization and weight decay regularization to perform stochastic gradient descent on the CNN-Transformer gas abnormal outburst recognition model until the model loss value is less than the preset value.
7. The CNN fusion Transformer working face gas outburst anomaly recognition and dynamic early warning algorithm according to claim 4 is characterized in that: Step 5 is as follows: input the real-time gas concentration data into the CNN-Transformer gas abnormal emission recognition model trained in step (4), use the model to predict the gas emission data of the working face in the next 30 minutes, establish three warning levels: general warning, yellow warning, and red warning, calculate the difference between the current gas concentration and the model prediction value to establish the warning threshold, as shown in Table 2; at the same time, different gas concentration sensor measurement points on the working face are separately set with general warning and over-limit thresholds. If the increase in the time series data in the next 30 minutes is greater than 0.2, a gas hazard warning is issued; Table 2 Warning threshold relationship 。