Method for identifying and extracting rock noise signals of different operation procedures based on CNN (Convolutional Neural Network)

Abstract

The invention discloses a CNN-based method for identifying and extracting rock noise signals of different operation procedures, and belongs to the technical field of mine gas control. The method comprises the following steps: S1, collecting rock noise signals of different operation processes, and preprocessing the rock noise signals to obtain time-frequency images of the rock noise signals; s2, establishing a CNN-based rock noise signal classification model; s3, training and testing the classification model; s4, establishing a signal extraction model, extracting a multi-scale feature map of the mixed signal through the signal extraction model, reconstructing a time domain signal of a target process, and separating signal segments; and S5, inputting the signal segments obtained in the step S4 into a CNN-based rock noise signal classification model, obtaining a classification model output result, and completing rock noise signal identification and extraction. According to the invention, the efficiency and accuracy of coal and gas outburst monitoring and early warning can be improved, and the intelligent process of coal and gas outburst rock noise monitoring is further promoted.

Description

Technical Field

[0001] This invention relates to the field of mine gas control technology, and in particular to a method for identifying and extracting ground sound signals for different work processes based on CNN. Background Technology

[0002] Ground sound is an elastic stress wave released when coal and rock masses are damaged and destroyed under load. It is an external manifestation of coal and rock damage and is directly related to the development of micro-fractures and defects within the coal and rock mass. By monitoring coal and rock fracture information, it reflects precursory information of coal and gas outbursts and is widely used in the prevention and control of coal and gas outbursts. However, due to its sensitivity and wide monitoring frequency domain, ground sound monitoring faces the challenge of extracting coal and rock fracture signals, leading to a need to improve the accuracy of coal and gas outburst monitoring and early warning.

[0003] Currently, the extraction of ground acoustic signals from coal and rock fractures largely relies on manual methods to determine their frequency range and decide whether to retain or filter them. However, given the numerous and complex operational procedures during longwall mining and tunneling, continuing to use traditional manual decision-making methods would consume a significant amount of time and effort. Therefore, there is an urgent need for an intelligent method for extracting target ground acoustic signals. Summary of the Invention

[0004] Based on the above-mentioned technical problems, this invention proposes a method for identifying and extracting ground sound signals for different work processes based on CNN.

[0005] The technical solution adopted in this invention is: A method for identifying and extracting ground sound signals from different work processes based on CNN includes the following steps: S1. Collect ground sound signals from different work processes, perform preprocessing, and obtain time-frequency images of the ground sound signals; S2. Establish a CNN-based classification model for ground sound signals; S3. Based on the time-frequency images of ground sound signals from different work processes obtained in S1, train the CNN-based ground sound signal classification model established in S2. S4. Based on the convolutional and pooling layers in the classification model trained in S3, extract multi-scale feature maps of mixed signals under known multiple work processes, and reconstruct the temporal signal of the target process using deconvolutional layers. S5. Input the time-domain signal obtained in S4 into the CNN-based ground sound signal classification model, obtain the output result of the classification model, and complete the ground sound signal recognition and extraction.

[0006] Furthermore, in S1, sound signals from coal mining, tunneling, blasting, belt conveying, coal removal, support, and no-operation are collected respectively. The sound signals from different operation procedures are converted into time and frequency using fast Fourier transform to obtain time and frequency images of the sound signals from different operation procedures.

[0007] Furthermore, in S2, the established CNN-based ground sound signal classification model adopts a multi-branch convolutional network. The main branch extracts the local time-frequency features of the ground sound signal time-frequency image through a 3×3 convolution kernel, while the parallel branch expands the receptive field through dilated convolution to capture long-term signal patterns. The output layer is designed as a multi-label classifier.

[0008] Furthermore, in step S3, the time-frequency images of the ground sound signals corresponding to different work procedures are marked, divided into training and test sets according to the proportion, and input into the classification model built in step S2 for training.

[0009] Furthermore, in S4, the convolutional and pooling layers of the classification model are used as encoders to extract multi-scale feature maps of the mixed signal; the temporal signal of the target process is reconstructed by upsampling the feature maps through deconvolutional layers and combining the shallow detail features and deep semantic features of the encoder with skip connections.

[0010] Furthermore, in step S5, the time-domain signal of the reconstructed target process is input into the classification model. If the output result of the classification model does not match the operation process contained in the mixed signal, the model is returned to be re-optimized.

[0011] The beneficial technical effects of this invention are: This invention establishes a feature library of ground sound signals under different work processes, and combines it with a convolutional neural network algorithm to identify ground sound signals under different work processes and extract ground sound signals for specific work processes, thereby improving the efficiency and accuracy of coal and gas outburst monitoring and early warning, and further promoting the intelligent process of coal and gas outburst ground sound monitoring.

[0012] Compared to traditional manual methods that require significant time and effort, this invention avoids reliance on manual screening of raw signals, greatly improving the efficiency of early warning operations. Noise filtering technology eliminates strong noise interference from construction equipment, preventing equipment downtime or excessive maintenance due to signal misinterpretation. It also has the ability to reduce economic losses from accidents; high-purity ground acoustic data can provide early warnings of risks such as roof collapse and gas outbursts. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the method for identifying and extracting ground sound signals based on different work processes according to the present invention. Detailed Implementation

[0014] To address the interference problem caused by multiple coexisting work processes in monitoring coal and gas outbursts, traditional effective signal extraction methods often employ manual noise filtering by dividing work processes, which consumes a significant amount of time and effort. In contrast, this invention utilizes neural network algorithms to achieve intelligent identification and extraction of coal and rock fracture signals during coal mining and tunneling, thereby improving the efficiency of coal and gas outburst monitoring and early warning.

[0015] This invention utilizes a classification model to identify multiple work processes contained in a mixed signal. The process of reconstructing the time-domain signal is the extraction process of the corresponding work process signal, during which the desired signal can be extracted. The process of identifying and extracting the target signal in this method is the noise filtering process.

[0016] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0017] like Figure 1 As shown, a method for identifying and extracting ground sound signals from different work processes based on CNN includes the following steps:

[0018] S1. Collect ground sound signals from different work processes, perform preprocessing, and obtain time-frequency images of the ground sound signals.

[0019] Specifically, ground sound signals were collected for different work processes, such as coal mining, tunneling, blasting, belt conveyor, coal removal and support, and no-work. The ground sound signals for different work processes were converted into time and frequency using fast Fourier transform to obtain time and frequency images of the ground sound signals for different work processes.

[0020] S2. Establish a CNN-based classification model for ground sound signals.

[0021] This classification model employs a multi-branch convolutional network. The main branch extracts local time-frequency features from the time-frequency image of the ground sound signal using a 3×3 convolutional kernel, while the parallel branch expands the receptive field through dilated convolution (dilation coefficient 2-4) to capture long-term signal patterns. The output layer is designed as a multi-label classifier (Sigmoid activation function) that simultaneously predicts all work operation categories contained in the mixed signal (e.g., "tunneling + belt conveyor"), addressing the multi-class labeling problem when different work operations overlap.

[0022] S3. Based on the time-frequency images of ground sound signals from different work processes obtained in S1, train the CNN-based ground sound signal classification model established in S2.

[0023] The time-frequency images of the ground sound signals corresponding to different work processes are labeled and divided into training and test sets in an 8:2 ratio. These sets are then input into the classification model built in S2 for training.

[0024] Steps S2 and S3 mainly involve training the classification model using existing ground sound signals from different work processes, enabling it to better identify the corresponding work processes from different ground sound signals.

[0025] S4. Based on the convolutional and pooling layers of the classification model trained in S3, extract multi-scale feature maps of mixed signals under known multiple work processes, and reconstruct the time-domain signal of the target process using deconvolutional layers.

[0026] The convolutional and pooling layers of the classification model are used as encoders to extract multi-scale feature maps of the mixed signal. The feature maps are upsampled by the deconvolutional layer and combined with the shallow detail features and deep semantic features of the encoder by skip connections to reconstruct the temporal signal of the target process.

[0027] S5. Input the signal segment obtained in S4 into the CNN-based ground sound signal classification model, obtain the output result of the classification model, and complete the ground sound signal recognition and extraction.

[0028] When performing S5, the time-domain signal or signal segment of the target process is input into the classification model. If the output of the classification model does not match the work process contained in the mixed signal, the classification model is returned to be optimized again to improve the accuracy of the classification model and the signal extraction model.

[0029] Steps S4 and S5 mainly involve extracting reconstructed signals through the classification model, and further training the classification model by comparing the reconstructed signals with the target process of the operation.

[0030] After steps S1 to S5, the ground sound signal can be acquired on-site, identified and extracted, thereby determining the corresponding work procedure.

[0031] The present invention proposes a method for identifying and extracting ground sound signals under different work processes based on CNN. By establishing a feature library of ground sound signals under different work processes, and combining it with convolutional neural network algorithm, the method realizes the identification of ground sound signals under different work processes and the extraction of ground sound signals under specific work processes, thereby further promoting the intelligent monitoring of ground sound signals in coal and gas outbursts.

[0032] For any parts not mentioned above, existing technologies can be adopted or referenced.

[0033] The above descriptions are merely preferred embodiments of the present invention. The present invention is not limited to the embodiments listed above. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A method for identifying and extracting ground sound signals from different work processes based on CNN, characterized in that... Includes the following steps: S1. Collect ground sound signals from different work processes, perform preprocessing, and obtain time-frequency images of the ground sound signals; S2. Establish a CNN-based classification model for ground sound signals; S3. Based on the time-frequency images of ground sound signals from different work processes obtained in S1, train the CNN-based ground sound signal classification model established in S2. S4. Based on the convolutional and pooling layers in the classification model trained in S3, extract multi-scale feature maps of mixed signals under known multiple work processes, and reconstruct the temporal signal of the target process using deconvolutional layers. S5. Input the time-domain signal obtained in S4 into the CNN-based ground sound signal classification model, obtain the output result of the classification model, and complete the ground sound signal recognition and extraction.

2. The method for identifying and extracting ground sound signals based on different work processes according to claim 1, characterized in that: In S1, sound signals from coal mining, tunneling, blasting, belt conveying, coal removal, support, and no-operation are collected respectively. The sound signals from different operation procedures are converted into time and frequency using fast Fourier transform to obtain time and frequency images of the sound signals from different operation procedures.

3. The method for identifying and extracting ground sound signals based on different work processes according to claim 2, characterized in that: In S2, the established CNN-based ground sound signal classification model adopts a multi-branch convolutional network. The main branch extracts the local time-frequency features of the ground sound signal time-frequency image through a 3×3 convolution kernel, while the parallel branch expands the receptive field through dilated convolution to capture long-term signal patterns. The output layer is designed as a multi-label classifier.

4. The method for identifying and extracting ground sound signals based on different work processes according to claim 3, characterized in that: In step S3, the time-frequency images of the ground sound signals corresponding to different work procedures are marked, divided into training set and test set according to the proportion, and input into the classification model built in step S2 for training.

5. The method for identifying and extracting ground sound signals based on different work processes according to claim 4, characterized in that: In step S4, the convolutional and pooling layers of the classification model are used as encoders to extract multi-scale feature maps of the mixed signal. The feature maps are upsampled by the deconvolutional layer and combined with the shallow detail features and deep semantic features of the fusion encoder by skip connections to reconstruct the temporal signal of the target process.

6. The method for identifying and extracting ground sound signals based on different work processes according to claim 5, characterized in that: In step S5, the time-domain signal of the reconstructed target process is input into the classification model. If the output result of the classification model does not match the operation process contained in the mixed signal, the classification model is returned for re-optimization.

Images