Deep learning coal fire detection method based on ground penetrating radar

Abstract

The invention discloses a deep learning coal fire detection method based on a ground penetrating radar, and the method comprises the steps: collecting a radar signal, carrying out the underground detection of a coal mine region through employing the ground penetrating radar, and obtaining the radar signal data of a coal mine underground medium; preprocessing data, extracting feature information of a reflected signal, performing data enhancement on the data, expanding the sample size by adopting rotation, translation and scaling modes, and enhancing the generalization ability of the model; carrying out deep learning modeling, extracting local features of the radar signals through CNN, and extracting time sequence features of the radar signals through LSTM; coal fire detection analysis: identifying whether a coal fire trace exists or not; and outputting and verifying a result. According to the method, high-precision and intelligent identification of the coal fire hidden danger area can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of coal fire detection, and specifically refers to a deep learning coal fire detection method based on ground penetrating radar. Background Art

[0002] Coal fire is one of the most common and destructive disasters in coal mining. The occurrence of coal fire not only leads to a large waste of coal resources, but also causes a series of safety and environmental problems, such as the release of harmful gases, ground collapse and fire spread. Traditional coal fire detection methods rely on directly measuring underground temperature or collecting mine gas samples. However, temperature measurement has the problems of limited detection range and large environmental interference, while gas monitoring requires an accurate source of gas samples and is often limited by mine ventilation conditions. Although physical detection methods such as electromagnetic and gravity methods have certain applications in coal fire detection, they are limited by equipment complexity and analytical accuracy in actual operation. In addition, these methods usually rely on empirical analysis and cannot identify coal fire hazard areas with high precision and intelligence.

[0003] As a non-contact, high-resolution physical detection technology, ground penetrating radar can quickly obtain the structure and distribution information of underground media through the propagation and reflection characteristics of electromagnetic waves. Ground penetrating radar has great potential in the field of coal mine disaster detection due to its high efficiency, non-destructiveness and ability to adapt to complex geological conditions. However, due to the complex environment of coal mines, radar signals are easily affected by noise and interference, which makes data analysis difficult, and the application effect of traditional radar signal processing methods in coal fire detection is limited. In recent years, the rapid development of deep learning technology has provided a new way to process complex radar signals. By constructing a multi-layer neural network model, deep learning technology can learn feature representation from a large amount of radar signal data and achieve accurate classification and recognition. However, there is no mature research and application of combining ground penetrating radar technology with deep learning to establish an effective intelligent analysis method for coal fire detection. Summary of the invention

[0004] The purpose of the present invention is to provide a deep learning coal fire detection method based on ground penetrating radar, which can achieve high-precision and intelligent identification of coal fire hazard areas.

[0005] To achieve the above object, the present invention provides a deep learning coal fire detection method based on ground penetrating radar, comprising the following steps:

[0006] S1: radar signal acquisition, using ground penetrating radar to conduct underground detection in the coal mine area, and obtain radar signal data of the underground medium of the coal mine, wherein the radar signal data includes the reflection signal intensity and time delay information of the underground medium;

[0007] S2: Data preprocessing, preprocessing the radar signal data obtained in step S1, extracting characteristic information of the reflected signal, the characteristic information includes but is not limited to the spectrum characteristics, amplitude characteristics and time domain characteristics of the signal; performing data enhancement on the data, expanding the sample size by rotation, translation and scaling, and enhancing the generalization ability of the model;

[0008] S3: deep learning modeling, inputting the feature information extracted in step S2 into the deep learning model for training and recognition, wherein the deep learning model includes a convolutional neural network (CNN) and a long short-term memory network (LSTM): CNN extracts local features of the radar signal, and LSTM extracts time series features of the radar signal, and combines the outputs of the two to construct a joint feature representation, and uses a back propagation algorithm to optimize the model parameters to reduce prediction errors;

[0009] S4: coal fire detection analysis, analyzing the new radar signal data through the deep learning model trained in step S3 to identify whether there are traces of coal fire;

[0010] S5: Result output and verification. The deep learning model outputs the coal fire detection results and distribution map to verify the performance and reliability of the model.

[0011] As a further solution of the present invention: the coal mine area in step S1 includes a normal coal mine area and a coal mine area with coal fire phenomenon.

[0012] As a further solution of the present invention: the convolutional layer of CNN uses the ReLU activation function, the formula is:

[0013] f(x)=max(0,x)

[0014] Among them, is the linear combination result of the convolutional layer output, and f(x) is the value after ReLU activation.

[0015] As a further solution of the present invention: the input of LSTM is the time series feature of the radar signal, and the state update formula of the LSTM is:

[0016] f t =σ(W f ·[h t-1 ,xt]+b f )

[0017] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0018]

[0019] ot =σ(W o ·[h t-1 ,x t ]+b o )

[0020] h t =o t tanh(C t )

[0021] Among them, x t is the input at time step t, h t-1 is the hidden state of the previous time step, C t is the cell state at the current time step, f t For the forget gate, i t is the input gate, is the candidate unit state, o t is the output gate, σ and tanh are the sigmoid and tanh activation functions respectively, W and b are the weight matrix and bias term respectively.

[0022] As a further solution of the present invention: the pretreatment in step S2 comprises the following steps:

[0023] De-noising the collected original radar signal, and removing the background noise through wavelet transform or filter;

[0024] Normalize the denoised radar signal to keep its amplitude within a certain range, eliminating data deviation caused by acquisition equipment or environmental factors;

[0025] The radar signal is decomposed by time-frequency analysis method to obtain the time-frequency representation of different frequency components, and these components are used as input features.

[0026] As a further solution of the present invention: the basis for coal fire detection in step S4 includes:

[0027] The radar signal characteristics in the coal fire area are manifested as abnormal reflected signal intensity and obvious time domain waveform changes, especially sudden increases in amplitude within certain specific frequency ranges;

[0028] The occurrence of coal fires is usually accompanied by abnormal changes in underground temperature and dielectric constant, which can cause changes in the reflection characteristics of radar signals;

[0029] The coal fire detection model is trained on a large amount of signal data from known coal fire areas and areas without coal fires, and learns the differences between radar signal characteristics in coal fire areas and signals in normal areas.

[0030] As a further solution of the present invention: outputting the coal fire detection result in step S5 includes:

[0031] The probability value of coal fire detection indicates the possibility of detecting a coal fire, ranging from [0,1];

[0032] If the detection result is greater than the preset threshold, an alarm message indicating the presence of coal fire is output;

[0033] The confidence level of the output detection result indicates the reliability of the model for the result.

[0034] As a further solution of the present invention: the loss function used in the training process of the model in step S3 is a cross entropy loss function, and the formula is:

[0035]

[0036] Among them, y i is the actual label, is the predicted value output by the model, and N is the number of samples.

[0037] As a further solution of the present invention: the calculation formula of the reflected signal strength in step S1 is as follows:

[0038]

[0039] Among them, P r is the received reflected signal power, P t is the power of the transmitted signal, and R is the percentage of the reflected intensity.

[0040] As a further solution of the present invention: the coal fire detection result in step S5 is used to monitor the safety status of the coal mine in real time, update the coal fire detection status in real time, and issue a corresponding alarm signal according to the detection result.

[0041] Compared with the prior art, the present invention provides a deep learning coal fire detection method based on ground penetrating radar. This method achieves high-precision and intelligent identification of coal fire hazard areas by combining the high-resolution data acquisition capability of ground penetrating radar with the feature extraction and pattern recognition capabilities of deep learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of a deep learning coal fire detection method based on ground penetrating radar of the present invention.

[0043] Figure 2 It is a flow chart of acquiring radar signal collection and data preprocessing information of a predetermined area according to ground penetrating radar detection of the predetermined area in the present invention.

[0044] Figure 3 It is a flow chart of data modeling and training in the present invention.

[0045] Figure 4It is a schematic diagram of the spatial distribution of convolutional features in the present invention.

[0046] Figure 5 It is a schematic diagram of the three-dimensional distribution of the time domain characteristics of the radar signal in the present invention.

[0047] Figure 6 It is a schematic diagram of the change of model training loss in the present invention.

[0048] Figure 7 It is a schematic diagram of the spatial distribution of coal fire areas in the present invention. DETAILED DESCRIPTION

[0049] The present invention will be further described below in conjunction with the accompanying drawings.

[0050] like Figure 1 As shown, a deep learning coal fire detection method based on ground penetrating radar of the present invention comprises the following steps:

[0051] S1: radar signal acquisition, using ground penetrating radar to conduct underground detection in the coal mine area, and obtain radar signal data of the underground medium of the coal mine, wherein the radar signal data includes the reflection signal intensity and time delay information of the underground medium;

[0052] S2: Data preprocessing, preprocessing the radar signal data obtained in step S1, extracting characteristic information of the reflected signal, the characteristic information includes but is not limited to the spectrum characteristics, amplitude characteristics and time domain characteristics of the signal; performing data enhancement on the data, expanding the sample size by rotation, translation and scaling, and enhancing the generalization ability of the model;

[0053] S3: deep learning modeling, inputting the feature information extracted in step S2 into the deep learning model for training and recognition, wherein the deep learning model includes a convolutional neural network (CNN) and a long short-term memory network (LSTM): CNN extracts local features of the radar signal, and LSTM extracts time series features of the radar signal, and combines the outputs of the two to construct a joint feature representation, and uses a back propagation algorithm to optimize the model parameters to reduce prediction errors;

[0054] S4: coal fire detection analysis, analyzing the new radar signal data through the deep learning model trained in step S3 to identify whether there are traces of coal fire;

[0055] S5: Result output and verification. The deep learning model outputs the coal fire detection results and distribution map to verify the performance and reliability of the model.

[0056] Specific steps are as follows Figure 2 , Figure 3As shown in the figure, the original signal data of the ground penetrating radar is obtained, and the electromagnetic wave reflection signals of different depth ranges are collected by the radar equipment to obtain the data of reflection intensity and propagation characteristics. The obtained signal is denoised, and the interference noise caused by the external environment and the equipment itself is removed by filtering technology to obtain a clean radar signal. The processed signal is normalized, and the numerical range of the signal data is adjusted to a fixed standard range to ensure the comparability of various types of signals and obtain standardized data. The signal features are obtained and the image information is extracted. The radar signal is convoluted using a convolutional neural network to obtain multiple feature maps. The features output by the convolution layer are ReLU activated, and the ReLU activation function is applied to each value in the feature map to obtain the activated feature data. The activated features are obtained, and multiple activated feature maps are merged to obtain a complete feature set. The time series data is obtained for processing, and the radar signal data is arranged in time series to obtain the time series data. The time series data is processed by LSTM, and the signal at each moment is input into the LSTM network. Combined with the memory information of the previous moment, the state update at the current moment is obtained. The memory unit of the model is updated through the time series data to obtain the memory state after long-term dependent learning. The time series signal is output through the LSTM network to obtain the prediction results based on the time dynamic characteristics. The trained deep learning model is obtained, and the processed radar signal data is input into the trained convolutional neural network and LSTM model to obtain the coal fire detection results. According to the results output by the model, the prediction results are analyzed and the occurrence probability and location distribution map of the coal fire are generated to obtain the final output of the coal fire detection.

[0057] Get the feature map of the convolution layer output. The feature map extracts multi-level information from the input signal through the convolution operation. Process each value of each feature map. If the value is greater than zero, keep the value unchanged; if the value is less than or equal to zero, set it to zero. Through this process, all negative values ​​are set to zero and positive values ​​remain unchanged. The ReLU activation function makes the output of the convolution layer more sparse, which helps the model learn useful features more efficiently. This process enhances the nonlinear modeling ability of the network, improves the expressiveness of the feature map, and helps capture more complex patterns. The use of ReLU avoids the gradient vanishing problem, promotes the training of deep networks, and enables the network to better adapt to the learning needs of complex data.

[0058] Among them, the convolution layer uses the ReLU activation function, the formula is:

[0059] f(x)=max(0,x)

[0060] Among them, x is the linear combination result of the convolution layer output, and f(x) is the value after ReLU activation. Figure 4 As shown, the distribution of convolution features in space is demonstrated, providing support for the subsequent identification of coal fire areas.

[0061] The input data of the time step and the hidden state of the previous time step are obtained, and they are combined and input into different gated units of the LSTM network. The control signal is obtained through the forget gate to determine which information should be discarded from the memory. The input gate obtains the control signal to determine which new information should be added to the current memory unit. The candidate unit state generates a new candidate memory, and the current memory unit state is updated by combining the control of the forget gate and the input gate. The forget gate and the input gate work together to determine which memories remain unchanged and which memories are updated. The output gate generates the hidden state of the current time step based on the updated memory unit state and passes it to the next time step. Each step updates the current memory state and hidden state, ensuring that the network can effectively capture long-term dependencies when processing sequence data, while retaining important historical information, and determining the final prediction result through the control of the output gate.

[0062] Among them, the calculation formula for LSTM state update is:

[0063] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0064] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0065]

[0066] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0067] h t =o t tanh(C t )

[0068] Among them, x t is the input at time step t, h t-1 is the hidden state of the previous time step, C t is the cell state at the current time step, f t For the forget gate, i t is the input gate, is the candidate unit state, o tis the output gate, σ and tanh are the sigmoid and tanh activation functions respectively, W and b are the weight matrix and bias term respectively.

[0069] The preprocessing in step S2 includes the following steps:

[0070] De-noising the collected original radar signal, and removing the background noise through wavelet transform or filter;

[0071] Normalize the denoised radar signal to keep its amplitude within a certain range, eliminating data deviation caused by acquisition equipment or environmental factors;

[0072] like Figure 5 As shown, the radar signal is decomposed by time-frequency analysis method to obtain the time-frequency representation of different frequency components, and these components are used as input features.

[0073] The basis for coal fire detection in step S4 includes:

[0074] The radar signal characteristics in the coal fire area are manifested as abnormal reflected signal intensity and obvious time domain waveform changes, especially sudden increases in amplitude within certain specific frequency ranges;

[0075] The occurrence of coal fires is usually accompanied by abnormal changes in underground temperature and dielectric constant, which can cause changes in the reflection characteristics of radar signals;

[0076] The coal fire detection model is trained on a large amount of signal data from known coal fire areas and areas without coal fires, and is able to learn the differences between radar signal characteristics in coal fire areas and signals in normal areas.

[0077] Outputting the coal fire detection result in step S5 includes:

[0078] The probability value of coal fire detection indicates the possibility of detecting a coal fire, ranging from [0,1];

[0079] If the detection result is greater than the preset threshold, an alarm message indicating the presence of coal fire is output;

[0080] The confidence level of the output detection result indicates the reliability of the model for the result.

[0081] The loss function used in the training process of the model in step S3 is the cross-entropy loss function. Obtain the predicted probability and actual label output by the model, and calculate the loss of each sample. For each sample, determine the difference between its actual label and the predicted result. If the actual label is 1, calculate the logarithm of the predicted probability of the category and compare it with the actual label; if the actual label is 0, calculate the logarithm compensation value of the predicted probability of the category and compare it. Accumulate the losses of all samples to obtain the total loss of the entire data set. Through this process, the cross-entropy loss function effectively measures the gap between the model prediction results and the actual labels as the goal of model optimization. During the training process, the model updates parameters according to the total loss, and by minimizing the loss function, the prediction results gradually approach the actual labels, thereby improving the accuracy and robustness of the model. By continuously optimizing the loss function, the model gradually converges during the learning process to achieve better prediction performance. Among them, the calculation formula of the cross-entropy loss function is:

[0082]

[0083] Among them, y i is the actual label, is the predicted value output by the model, and N is the number of samples. Figure 6 As shown in the figure, we can intuitively understand the changes in the model optimization process. The X-axis represents the number of training rounds (unit: Epochs), and the Y-axis represents the loss value (Loss) of each training round. As the number of training rounds increases, the loss value gradually decreases, indicating that the model continues to improve its prediction ability. Finally, this figure can confirm that the model has achieved the expected training effect and provides accurate prediction results for coal fire detection.

[0084] The received radar signal power and the transmitted signal power are obtained to represent the strength of the echo signal and the transmitted signal. By comparing the two power values, their ratio is calculated, and the ratio is converted into a percentage to obtain the reflection intensity. The reflection intensity represents the reflection characteristics of the target object to the radar signal and reflects the physical properties of the target. A high reflection intensity indicates that the target object has a strong reflection ability to the radar signal, while a low reflection intensity indicates that the surface of the target may absorb or scatter the signal. Through this calculation process, the radar system can evaluate the characteristics of the target objects in the detection area, thereby providing accurate spatial positioning data for subsequent coal fire detection. The calculation of the reflection intensity can not only determine the size and shape of the target object, but also further analyze its physical properties and distribution state.

[0085] The calculation formula for the reflected signal strength is as follows:

[0086]

[0087] Among them, P r is the received reflected signal power, pt is the power of the transmitted signal, and R is the percentage of the reflected intensity.

[0088] The coal fire detection result in step S5 is used to monitor the safety status of the coal mine in real time, update the coal fire detection status in real time, and issue a corresponding alarm signal according to the detection result.

[0089] pass Figure 7 , which can clearly show the distribution of different coal fire areas and the reflection effect of coal fires of different intensities on radar signals.

Claims

1. A deep learning coal fire detection method based on ground penetrating radar, characterized in that: The following steps are involved: S1: radar signal acquisition, using ground penetrating radar to conduct underground detection in the coal mine area, and obtain radar signal data of the underground medium of the coal mine, wherein the radar signal data includes the reflection signal intensity and time delay information of the underground medium; S2: Data preprocessing: preprocessing the radar signal data obtained in step S1 to extract characteristic information of the reflected signal, including but not limited to the spectrum characteristics, amplitude characteristics and time domain characteristics of the signal; performing data enhancement on the data, expanding the sample size by rotation, translation and scaling, and enhancing the generalization ability of the model; S3: deep learning modeling, inputting the feature information extracted in step S2 into the deep learning model for training and recognition, wherein the deep learning model includes a convolutional neural network (CNN) and a long short-term memory network (LSTM): CNN extracts local features of the radar signal, and LSTM extracts time series features of the radar signal, and combines the outputs of the two to construct a joint feature representation, and uses a back propagation algorithm to optimize the model parameters to reduce prediction errors; S4: coal fire detection analysis, analyzing the new radar signal data through the deep learning model trained in step S3 to identify whether there are traces of coal fire; S5: Result output and verification. The deep learning model outputs the coal fire detection results and distribution map to verify the performance and reliability of the model.

2. A deep learning coal fire detection method based on ground penetrating radar according to claim 1, characterized in that: The coal mine area in step S1 includes a normal coal mine area and a coal mine area with coal fire.

3. A deep learning coal fire detection method based on ground penetrating radar according to claim 1 or 2, characterized in that: The convolutional layer of CNN uses the ReLU activation function, the formula is: f(x)=max(0,x) Among them, x is the linear combination result of the convolutional layer output, and f(x) is the value after ReLU activation.

4. A deep learning coal fire detection method based on ground penetrating radar according to claim 1 or 2, characterized in that: The input of the LSTM is the time series feature of the radar signal, and the state update formula of the LSTM is: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t ·tanh(C t ) Among them, x t is the input at time step t, h t-1 is the hidden state of the previous time step, C t is the cell state at the current time step, f t For the forget gate, i t is the input gate, is the candidate unit state, o t is the output gate, σ and tanh are the sigmoid and tanh activation functions respectively, W and b are the weight matrix and bias term respectively.

5. A deep learning coal fire detection method based on ground penetrating radar according to claim 1 or 2, characterized in that: The preprocessing in step S2 includes the following steps: De-noising the collected original radar signal, and removing the background noise through wavelet transform or filter; Normalize the denoised radar signal to keep its amplitude within a certain range, eliminating data deviation caused by acquisition equipment or environmental factors; The radar signal is decomposed by time-frequency analysis method to obtain the time-frequency representation of different frequency components, and these components are used as input features.

6. A deep learning coal fire detection method based on ground penetrating radar according to claim 1 or 2, characterized in that: The basis for coal fire detection in step S4 includes: The radar signal characteristics in the coal fire area are manifested as abnormal reflected signal intensity and obvious time domain waveform changes, especially sudden increases in amplitude within certain specific frequency ranges; The coal fire detection model is trained on a large amount of signal data from known coal fire areas and areas without coal fires, and learns the differences between radar signal characteristics in coal fire areas and signals in normal areas.

7. A deep learning coal fire detection method based on ground penetrating radar according to claim 1 or 2, characterized in that: Outputting the coal fire detection result in step S5 includes: The probability value of coal fire detection indicates the possibility of detecting a coal fire, ranging from [0,1]; If the detection result is greater than the preset threshold, an alarm message indicating the presence of coal fire is output; The confidence level of the output detection result indicates the reliability of the model for the result.

8. A deep learning coal fire detection method based on ground penetrating radar according to claim 1 or 2, characterized in that: The loss function used in the model training process in step S3 is the cross entropy loss function, and the formula is: Among them, y i is the actual label, is the predicted value output by the model, and N is the number of samples.

9. A deep learning coal fire detection method based on ground penetrating radar according to claim 1 or 2, characterized in that: The calculation formula of the reflected signal strength in step S1 is as follows: Among them, P r is the received reflected signal power, P t is the power of the transmitted signal, and R is the percentage of the reflected intensity.

10. A deep learning coal fire detection method based on ground penetrating radar according to claim 1 or 2, characterized in that: The coal fire detection result in step S5 is used to monitor the safety status of the coal mine in real time, update the coal fire detection status in real time, and issue a corresponding alarm signal according to the detection result.

Images