Coal mine tunnel water inrush symptom identification method

Abstract

The invention discloses a coal mine tunnel water inrush symptom identification method. The method comprises the following steps: acquiring tunnel water inrush symptom data; performing short-time Fourier transform on the collected original data to convert the collected original data into a time-frequency graph; data layer generalization is carried out on the generated time-frequency graph, water inrush symptoms in different environments are simulated, and the diversity of data and the generalization ability of the model are improved; multi-scale and transformation combined feature extraction is carried out, the adaptability of the model to different data distributions is enhanced, and the generalization ability is improved; respectively inputting the original time-frequency graph and the time-frequency graph with the generalized data layer into a feature extraction network, and introducing a feature layer generalization module to improve the robustness and generalization ability of the model; and carrying out feature fusion on the original features and the generalized and enhanced features, constructing a classifier, and carrying out classification identification on the water inrush condition to realize water inrush symptom identification. According to the method, the water inrush identification generalization capability and accuracy of the identification model on different geological conditions, mining modes and environments can be effectively improved.

Description

Technical Field

[0001] The invention belongs to the technical field of coal mine tunnel water inrush symptom identification, and relates to a coal mine tunnel water inrush symptom identification method. Background Art

[0002] Water inrush is one of the most threatening hazards in coal mining operations, resulting in significant casualties and economic losses. Timely and accurate identification of water inrush events not only protects miners' lives but also ensures the continued stability of coal mine production. Managing water inrush in underground coal mines primarily focuses on both prevention before a water inrush occurs and post-inrush control. If trends can be identified before or during a weak water inrush phase, early warning can be provided to notify underground workers to evacuate promptly, minimizing casualties and equipment losses.

[0003] Traditional water inrush prediction methods, such as the water inrush coefficient method and the rock-water stress relationship method, analyze the water inrush mechanism. However, these methods fail to fully consider the factors affecting water inrush, resulting in unsatisfactory prediction results. In addition, due to the differences in geological conditions, mining methods, and environmental factors among different coal mines, the collected data has significant domain differences, requiring the recognition model to have strong generalization capabilities for different data distributions. However, due to the difficulty of underground data collection, it is difficult to collect enough samples to train the recognition model, resulting in poor results of existing coal mine roadway water inrush sign identification methods. Therefore, data sparsity and domain differences caused by different data distributions are the main challenges in the current identification of coal mine roadway water inrush signs. Summary of the Invention

[0004] In response to the shortcomings of the prior art, the present invention aims to provide a method for identifying signs of water inrush in coal mine tunnels, which is a method for identifying signs of water inrush in coal mine tunnels with strong generalization capabilities. Based on four typical water inrush symptom data, namely temperature, humidity, dust concentration and water noise, transfer learning and domain generalization theory are introduced, and data layer generalization and feature layer generalization methods are proposed to simulate various water inrush symptom characteristics in actual environments, thereby improving the robustness and generalization capabilities of the model. Even if there are domain differences between different coal mines, the recognition model can still maintain a high accuracy rate, thereby better ensuring the safety and stability of coal mine production. Based on this method, the present invention not only improves the accuracy and reliability of the water inrush symptom identification system, but also enhances its practicality, provides comprehensive decision support for project management, and helps to formulate more effective risk management strategies and emergency plans. It is of great significance to improving the safety management level of underground projects, ensuring the smooth progress of projects and the safety of personnel.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A method for identifying signs of water inrush in a coal mine tunnel comprises the following steps:

[0007] Step 1: Collecting water inrush symptom data in tunnels: Deploy multiple sensor nodes in coal mine tunnels to collect four types of water inrush symptom data: temperature, humidity, dust concentration, and water noise.

[0008] Step 2, data conversion: the collected raw data of water sounds, temperature, humidity and dust concentration are transformed into time-frequency graphs through short-time Fourier transform;

[0009] Step 3: Data layer generalization: Perform data layer generalization on the generated time-frequency graph. By changing the amplitude and phase of the time-frequency graph, more training data is generated to simulate water inrush symptoms in different environments, thereby improving data diversity and the generalization ability of the model.

[0010] Step 4: Feature extraction combining multi-scale and transformation: The feature extraction network uses a feature enhancement module. By combining compressed convolution and expanded convolution in parallel, it captures features at different scales, optimizes the computational efficiency of convolution operations, and implements fine-grained feature learning in the time-frequency domain through multi-scale feature extraction, enhancing the model's adaptability to different data distributions and improving generalization capabilities.

[0011] Step 5, feature layer generalization: The original time-frequency graph and the generalized time-frequency graph of the data layer are respectively input into the feature extraction network. Then, the feature layer generalization module is introduced to achieve feature perturbation by changing the mean and variance of the feature distribution parameters, thereby improving the robustness and generalization ability of the model.

[0012] Step 6, water inrush sign identification: The original features and the generalized and enhanced features are fused to construct a classifier. The input features are classified into no water state, roof water dripping state warning, roof water dripping state without warning, bottom plate water gushing state warning, or bottom plate water gushing state without warning, to achieve water inrush sign identification.

[0013] The present invention also includes the following technical features:

[0014] Specifically, in step 2, the mathematical expression of short-time Fourier transform is as follows:

[0015]

[0016] Among them, X(t,f) is the input signal, w(τ-t) is the window function, e -j2πfτ It is a complex sine wave, representing the frequency component, A(t,f) is the amplitude, and P(t,f) is the phase.

[0017] Specifically, in step 3, for the time-frequency diagram of time window i, its amplitude and phase are represented as A and i (t,f) and P i (t,f), the true label is represented by yi ; For the time-frequency diagram of time window j, its amplitude and phase are represented as A j (t,f) and P j (t,f), the true label is represented by y j ; Two parameters λ and γ are introduced to enhance the amplitude and phase respectively and generate virtual samples; the specific formula is as follows:

[0018]

[0019] in and represents the amplitude and phase of the newly generated time-frequency diagram, represents the label of the newly generated time-frequency graph, λ and γ are randomly generated weight coefficients, satisfying λ,γ∈[0,1]; in this way, a new time-frequency graph can be generated, thereby achieving the purpose of data generalization and enhancement.

[0020] Specifically, in step 4, the time-frequency map is first input into the convolution layer to obtain the initial features, and then input into the feature enhancement module to extract multi-scale features; in the implementation, a total of 8 feature enhancement modules are designed to generate multi-scale features in stages. First, 2 feature enhancement modules are used to extract high-scale features (56×56×128), which are then input into the 3rd-4th feature enhancement modules after the maximum pooling operation to extract medium-scale features (28×28×256), and then further input into the 5th-8th feature enhancement modules after the maximum pooling operation to extract low-scale features (14×14×512).

[0021] Specifically, the feature enhancement module includes a Squeeze layer and an Expand layer; the Squeeze layer is composed of a 1×1 convolutional layer for compressing the input, and the Expand layer contains two branches, which respectively use a 1×1 convolutional layer and a 3×3 convolutional layer to expand the compressed features. Finally, the features obtained from the two branches of the Expand layer are spliced as the output of the feature enhancement module.

[0022] Specifically, assume that the input of the feature enhancement module is represented as F in , the compressed feature obtained after the Squeeze layer is expressed as:

[0023]

[0024] in represents the weight of the 1×1 convolutional layer; the output of the feature enhancement module is expressed as:

[0025]

[0026] in and They represent the weights of the 1×1 and 3×3 convolutional layers of the two branches in the Expand layer, respectively. concat(·) means concatenating the features according to the channels.

[0027] Specifically, in step 5, it is assumed that the n feature vectors obtained from a set of time-frequency graph data are expressed as

[0028] [f1,f2,...,f i ,...,f n ], where f i Represents the i-th feature, the mean μ and variance σ of the set of feature vectors 2 Respectively expressed as:

[0029]

[0030] During the training process, a distribution is randomly selected from the Gaussian distribution, uniform distribution, and normal distribution. Two values are randomly determined based on the distribution and added to the mean and variance of the feature vector to obtain the new mean and variance. Assume that the new mean and variance are expressed as and variance Then for the feature f i , and its corresponding new feature is expressed as:

[0031]

[0032] Specifically, in step 6, the classifier is trained using the cross entropy loss function, which can measure the difference between the predicted result and the true label:

[0033]

[0034] Among them, y i is the true label of the data, is the predicted probability distribution and C is the number of categories.

[0035] Compared with the prior art, the present invention has the following technical effects:

[0036] The present invention proposes two strategies: data layer generalization and feature layer generalization, which can effectively improve the recognition model's generalization ability and accuracy for water inrush identification under different geological conditions, mining methods, and environments. In terms of data layer generalization, the present invention adjusts the amplitude and phase of the time-frequency graph to simulate water inrush symptom data under different conditions, which can improve the diversity of the data, thereby expanding the training data set and enabling the model to learn knowledge about water inrush symptoms in different coal mines. In terms of feature layer generalization, the present invention introduces Gaussian distribution, uniform distribution, and normal distribution disturbances to change the mean and variance of the features, simulate water inrush characteristics of different distributions, and assist the model in better capturing the dynamic change characteristics of the data, thereby improving the robustness and generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a model diagram of the method for identifying signs of water inrush of the present invention.

[0038] Figure 2 This is a schematic diagram of the feature extraction network of the present invention.

[0039] Figure 3 This is a schematic diagram of the process of the method for identifying signs of water inrush according to the present invention. DETAILED DESCRIPTION

[0040] Specific embodiments of the present invention are given below. It should be noted that the present invention is not limited to the following specific embodiments, and all equivalent modifications made on the basis of the technical solution of this application fall within the protection scope of the present invention.

[0041] Example 1:

[0042] This embodiment provides a method for identifying signs of water inrush in a coal mine tunnel, comprising the following steps:

[0043] Step 1: Collecting water inrush sign data in tunnels: Deploy multiple sensor nodes in coal mine tunnels to collect four types of water inrush sign data: temperature, humidity, dust concentration, and water noise, which serve as input for the recognition model.

[0044] Step 2, data conversion: The collected raw data of water sounds, temperature, humidity, and dust concentration are converted into time-frequency diagrams through short-time Fourier transform (STFT). STFT can process time and frequency information simultaneously, thereby helping the recognition model better capture the dynamic changes of the data.

[0045] The mathematical expression of STFT is as follows:

[0046]

[0047] Among them, X(t,f) is the input signal, w(τ-t) is the window function, e -j2πfτ is a complex sine wave, representing the frequency component, e is a natural constant, j is an imaginary unit, A(t,f) is the amplitude, and P(t,f) is the phase;

[0048] The advantage of a time-frequency graph is that it not only preserves the signal's time domain information but also includes frequency domain information, enabling subsequent recognition models to more comprehensively extract signal features. For example, for a water sound signal, a time-frequency graph can reveal how its frequency components change over time; for temperature and humidity signals, a time-frequency graph can reveal their periodic fluctuations; and for dust concentration signals, a time-frequency graph can reveal their instantaneous variations.

[0049] Step 3, data layer generalization: To improve the robustness and generalization ability of the model, the present invention performs data layer generalization on the generated time-frequency graph. By changing the amplitude and phase of the time-frequency graph, more training data is generated to simulate water inrush symptoms in different environments, thereby improving data diversity and the generalization ability of the model.

[0050] For the time-frequency diagram of time window i, its amplitude and phase are represented as A i (t,f) and P i (t,f), the true label is represented by y i ; For the time-frequency diagram of time window j, its amplitude and phase are represented as A j (t,f) and P j (t,f), the true label is represented by y j ; Two parameters λ and γ are introduced to enhance the amplitude and phase respectively and generate virtual samples; the specific formula is as follows:

[0051]

[0052] in and represents the amplitude and phase of the newly generated time-frequency diagram, represents the label of the newly generated time-frequency graph, λ and γ are randomly generated weight coefficients, satisfying λ,γ∈[0,1]; in this way, a new time-frequency graph can be generated, thereby achieving the purpose of data generalization and enhancement.

[0053] Through this method, the present invention can generate more virtual training samples, thereby better simulating various water inrush signs in real environments. By independently generalizing amplitude and phase, the model can learn richer and more diverse features, improving its stability and robustness in practical applications. Furthermore, this data generalization method can effectively alleviate the problem of data sparsity, allowing the model to maintain a high accuracy even in the absence of sufficient training data.

[0054] Step 4: Feature extraction combining multi-scale and transformation: The feature extraction network uses a feature enhancement module. By combining compressed convolution and expanded convolution in parallel, it efficiently captures features at different scales, optimizes the computational efficiency of convolution operations, and implements fine-grained feature learning in the time-frequency domain through multi-scale feature extraction, enhancing the model's adaptability to different data distributions and improving generalization capabilities.

[0055] The feature extractor synchronously extracts diversified features in the time-frequency domain from the original frequency map and the generalized frequency map data; specifically, the time-frequency map is first input into the convolution layer to obtain the initial features, and then it is input into the feature enhancement module to extract multi-scale features; in the implementation, a total of 8 feature enhancement modules are designed to generate multi-scale features in stages. First, 2 feature enhancement modules are used to extract high-scale features (56×56×128), which are then input into the 3rd-4th feature enhancement modules after the maximum pooling operation to extract medium-scale features (28×28×256), and then further input into the 5th-8th feature enhancement modules after the maximum pooling operation to extract low-scale features (14×14×512).

[0056] For each feature enhancement module, it contains a Squeeze layer and an Expand layer. The Squeeze layer consists of a 1×1 convolutional layer and is used to compress the input. The Expand layer contains two branches, which use a 1×1 convolutional layer and a 3×3 convolutional layer to expand the compressed features respectively. Finally, the features obtained from the two branches of the Expand layer are spliced as the output of the feature enhancement module.

[0057] Assume that the input of the feature enhancement module is represented as F in , the compressed feature obtained after the Squeeze layer is expressed as:

[0058]

[0059] in represents the weight of the 1×1 convolutional layer; the output of the feature enhancement module is expressed as:

[0060]

[0061] in and They represent the weights of the 1×1 and 3×3 convolutional layers of the two branches in the Expand layer, respectively. concat(·) means concatenating the features according to the channels.

[0062] Compared to the unidirectional feature degradation mechanism of traditional convolutional networks, this design imparts hierarchical network adaptability through a differentiable structural reorganization strategy: In the shallow feature extraction stage, it strengthens conservative compression of the channel dimension and detail enhancement dominated by spatial convolution to ensure complete capture of the zero-crossing features and pulse frontier morphology of the time-frequency graph. As the network depth increases, it gradually switches to a frequency-domain global feature extraction mode, extracting the envelope characteristics and energy distribution pattern of broadband signals by enhancing the intensity of channel interactions. This progressive feature expression conversion mechanism effectively overcomes the contradiction between detail loss and insufficient abstraction in conventional methods, achieving breakthrough improvements in both time-frequency coordinate alignment accuracy and cross-domain distribution adaptability.

[0063] Step 5, feature layer generalization: The original time-frequency graph and the generalized time-frequency graph of the data layer are respectively input into the feature extraction network. Then, the feature layer generalization module is introduced. By changing the distribution of features, the characteristics of water inrush signs in different environments can be simulated, further improving the robustness and generalization ability of the model.

[0064] To further improve the robustness and generalization ability of the model, this paper proposes a feature layer generalization module. The innovation of this module is to achieve feature perturbation by changing the distribution parameters (mean and variance) of the features, thereby helping the network optimize its learning ability under various distribution conditions.

[0065] Assume that the n eigenvectors obtained from a set of time-frequency graph data are represented as [f1,f2,...,f i ,...,f n ], where f i Represents the i-th feature, the mean μ and variance σ of the set of feature vectors 2 Respectively expressed as:

[0066]

[0067] The mean and variance can reflect the central tendency and dispersion of features. Therefore, the present invention changes the distribution of features by changing the mean and variance of the set of feature vectors, thereby helping the model learn more features with different distributions and improving its generalization ability. Specifically, during the training process, one distribution is randomly selected from the three distributions of Gaussian distribution, uniform distribution and normal distribution each time, and two values are further randomly determined based on the distribution. These values are added to the mean and variance of the feature vector to obtain the new mean and variance. Assume that the new mean and variance are expressed as and variance Then for the feature f i , and its corresponding new feature is expressed as:

[0068]

[0069] In this way, by perturbing the mean and variance of the features, we can effectively simulate the feature changes in different environments, helping the model learn more knowledge, which has important engineering application value and practical significance.

[0070] Step 6, water inrush sign identification: The extracted features of different data are integrated to construct a classifier, and the water inrush conditions are classified into five types: no water state, roof water dripping state warning, roof water dripping state without warning, bottom plate water gushing state warning, and bottom plate water gushing state without warning, so as to achieve the purpose of water inrush sign identification.

[0071] In this method, the original features and the generalized and enhanced features are first fused, and then the fused features are identified to determine whether a water inrush has occurred. Specifically, a five-category classifier is introduced to classify the input features into one of the following: no water, roof water warning, roof water no warning, floor water inrush warning, or floor water inrush no warning, thus achieving the purpose of identifying signs of water inrush.

[0072] The present invention uses the cross-entropy loss function to train the classifier, which measures the difference between the predicted result and the true label:

[0073]

[0074] Among them, y i is the true label of the data, is the predicted probability distribution, and C is the number of categories, which is equal to 5.

[0075] Experimental verification: To verify the technical effects of the embodiments of the present invention, a monitoring system was constructed and simulation experiments were carried out:

[0076] Sensors were deployed at the coal mining face to simultaneously collect four key data points: temperature, humidity, dust concentration, and water call signals. The data covered five typical conditions: no water, roof watering warning, roof watering without warning, floor water inrush warning, and floor water inrush without warning. 500 samples were collected for each condition (each containing 10 minutes of continuous time series data). The data was divided into training, validation, and test sets in a 3:1:1 ratio. Each dataset contained five conditions: no water, roof watering warning / no warning, and floor water inrush warning / no warning. The training set contained 300 samples for each condition, for a total of 1,500 samples; the validation set contained 100 samples for each condition, for a total of 500 samples; and the test set contained 100 samples for each condition, for a total of 500 samples. The results are shown in Table 1.

[0077] Table 1 Test data results

[0078]

[0079]

[0080] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various simple modifications can be made to the technical solution of the present invention, and these simple modifications all fall within the scope of protection of the present invention.

[0081] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any appropriate manner without contradiction. To avoid unnecessary repetition, the present invention will not further describe various possible combinations.

[0082] In addition, the various embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the present invention, they should also be regarded as the contents disclosed by the present invention.

Claims

1. A method for identifying signs of water inrush in coal mine tunnels, characterized in that: The following steps are involved: Step 1: Collecting water inrush symptom data in tunnels: Deploy multiple sensor nodes in coal mine tunnels to collect four types of water inrush symptom data: temperature, humidity, dust concentration, and water noise. Step 2, data conversion: the collected raw data of water sounds, temperature, humidity and dust concentration are transformed into time-frequency graphs through short-time Fourier transform; Step 3: Data layer generalization: Perform data layer generalization on the generated time-frequency graph. By changing the amplitude and phase of the time-frequency graph, more training data is generated to simulate water inrush symptoms in different environments, thereby improving data diversity and the generalization ability of the model. Step 4: Feature extraction combining multi-scale and transformation: The feature extraction network uses a feature enhancement module. By combining compressed convolution and expanded convolution in parallel, it captures features at different scales, optimizes the computational efficiency of convolution operations, and implements fine-grained feature learning in the time-frequency domain through multi-scale feature extraction, enhancing the model's adaptability to different data distributions and improving generalization capabilities. Step 5, feature layer generalization: The original time-frequency graph and the generalized time-frequency graph of the data layer are respectively input into the feature extraction network. Then, the feature layer generalization module is introduced to achieve feature perturbation by changing the mean and variance of the feature distribution parameters, thereby improving the robustness and generalization ability of the model. Step 6, water inrush sign identification: The original features and the generalized and enhanced features are fused to construct a classifier. The input features are classified into no water state, roof water dripping state warning, roof water dripping state without warning, bottom plate water gushing state warning, or bottom plate water gushing state without warning, to achieve water inrush sign identification.

2. The method for identifying water inrush signs in coal mine tunnels according to claim 1, characterized in that: In step 2, the mathematical expression of short-time Fourier transform is as follows: Among them, X(t,f) is the input signal, w(τ-t) is the window function, e -j2πfτ It is a complex sine wave, representing the frequency component, A(t,f) is the amplitude, and P(t,f) is the phase.

3. The method for identifying water inrush signs in coal mine tunnels according to claim 1, wherein: In step 3, for the time-frequency diagram of time window i, its amplitude and phase are represented as A and B respectively. i (t,f) and P i (t,f), the true label is represented by y i ; For the time-frequency diagram of time window j, its amplitude and phase are represented as A j (t,f) and P j (t,f), the true label is represented by y j ; Two parameters λ and γ are introduced to enhance the amplitude and phase respectively and generate virtual samples; the specific formula is as follows: in and represents the amplitude and phase of the newly generated time-frequency diagram, represents the label of the newly generated time-frequency graph, λ and γ are randomly generated weight coefficients, satisfying λ,γ∈[0,1]; in this way, a new time-frequency graph can be generated, thereby achieving the purpose of data generalization and enhancement.

4. The method for identifying water inrush signs in coal mine tunnels according to claim 1, wherein: In step 4, the time-frequency map is first input into the convolution layer to obtain the initial features, and then input into the feature enhancement module to extract multi-scale features. In the implementation, a total of 8 feature enhancement modules are designed to generate multi-scale features in stages. First, 2 feature enhancement modules are used to extract high-scale features (56×56×128). After the maximum pooling operation, the 3rd to 4th feature enhancement modules are input to extract medium-scale features (28×28×256). Then, after further maximum pooling operation, the 5th to 8th feature enhancement modules are input to extract low-scale features (14×14×512).

5. The method for identifying water inrush signs in coal mine tunnels according to claim 4, characterized in that: The feature enhancement module includes a Squeeze layer and an Expand layer; the Squeeze layer is composed of a 1×1 convolutional layer and is used to compress the input. The Expand layer contains two branches, which respectively use a 1×1 convolutional layer and a 3×3 convolutional layer to expand the compressed features. Finally, the features obtained from the two branches of the Expand layer are spliced as the output of the feature enhancement module.

6. The method for identifying water inrush signs in coal mine tunnels according to claim 5, characterized in that: Assume that the input of the feature enhancement module is represented as F in , the compressed feature obtained after the Squeeze layer is expressed as: in represents the weight of the 1×1 convolutional layer; the output of the feature enhancement module is expressed as: in and They represent the weights of the 1×1 and 3×3 convolutional layers of the two branches in the Expand layer, respectively. concat(·) means concatenating the features according to the channels.

7. The method for identifying water inrush signs in coal mine tunnels according to claim 1, wherein: In step 5, it is assumed that the n feature vectors obtained from a set of time-frequency graph data are represented as [f1, f2, ..., f i ,...,f n ], where f i Represents the i-th feature, the mean μ and variance σ of the set of feature vectors 2 Respectively expressed as: During the training process, a distribution is randomly selected from the Gaussian distribution, uniform distribution, and normal distribution. Two values are randomly determined based on the distribution and added to the mean and variance of the feature vector to obtain the new mean and variance. Assume that the new mean and variance are expressed as and variance Then for the feature f i , and its corresponding new feature is expressed as:

8. The method for identifying water inrush signs in coal mine tunnels according to claim 1, wherein: In step 6, the classifier is trained using the cross entropy loss function, which measures the difference between the predicted result and the true label: Among them, y i is the true label of the data, is the predicted probability distribution and C is the number of categories.

Images