Method for self-avoidance of drum cutting support roof beam
By installing sound sensors on the coal mining machine drum, and using preprocessing and convolutional neural networks to identify whether the drum is cutting into the support beam, the problem of drum wear was solved, automatic avoidance in complex environments was achieved, and the equipment life was extended.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI TIANDI MINING EQUIP TECH CO LTD
- Filing Date
- 2023-08-15
- Publication Date
- 2026-07-10
Smart Images
Figure CN117307158B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for automatically avoiding the top beam of a support by a drum, belonging to the field of coal mining technology. Background Technology
[0002] In fully mechanized mining faces, geological conditions and other factors can easily cause the scraper conveyor to fail to push the conveyor into place. In such cases, the coal mining machine operator may accidentally cut into the top beam of the support frame while operating the drum to cut coal, leading to wear on both the drum and the top beam, and premature failure of both components. This phenomenon of the drum cutting into the top beam is quite common in fully mechanized mining faces.
[0003] The existing method to solve the above problems is to use a visual sensor to identify the positional relationship between the roller and the support beam and control the roller to avoid the support beam. However, this method requires a large amount of video data, large data processing volume, and high latency. In addition, in the underground coal cutting environment, there is a lot of dust, and the video shooting is often unclear and cannot be identified. Therefore, the reliability is not high. Summary of the Invention
[0004] The present invention aims to provide a method for automatic avoidance of the top beam of the drum cutting support. It uses sound to determine whether the drum has cut the top beam of the support, so as to control the drum to reduce its height in time to achieve automatic avoidance. It has high recognition in the complex environment of high dust and high water mist in coal mine working face, and can greatly extend the service life of the drum and coal mining machine.
[0005] The main technical solutions of this invention are as follows:
[0006] A method for self-avoidance of the top beam of a drum cutting support is provided. The method involves collecting the sound signal of the drum cutting coal and rock by a sound sensor installed on the drum of the coal mining machine, and determining whether the sound signal is the sound signal emitted when the drum cuts the top beam of the support by sound recognition. If so, the drum is immediately controlled to lower its height to avoid the collision.
[0007] The step of determining whether the sound signal is the sound signal emitted when the roller cuts the top beam of the support by sound recognition can be as follows:
[0008] Preprocessing: The sound signal is preprocessed to remove noise;
[0009] Feature extraction: Mel spectrum coefficients are extracted from the preprocessed sound signal, and the subband power distribution feature map algorithm is applied to the Mel spectrum coefficients to obtain the subband power distribution grayscale spectrum.
[0010] Feature comparison: The grayscale spectrum of the sub-band power distribution is enhanced using a convolutional neural network, and the useful signal is compared with the standard signal of the roller hitting the top beam in the sample library. When the recognition rate is greater than or equal to the set threshold, the corresponding sound signal is identified as the sound signal emitted when the roller cuts the top beam of the support.
[0011] The sound sensor is preferably installed on the blades of the coal mining machine drum.
[0012] When collecting the sound signal of the drum cutting coal and rock, the sampling is performed at the same frequency.
[0013] While controlling the rollers to lower their height to avoid obstacles, the system also issues warning signals.
[0014] The warning reminder may include a voice reminder.
[0015] The preprocessing may include the following steps:
[0016] a) Perform frame segmentation and windowing on the audio signal to obtain a signal sequence consisting of several frames of noisy audio signals;
[0017] b) Perform a short-time Fourier transform on each frame of the noisy audio signal to obtain the amplitude spectrum of the noisy audio signal;
[0018] c) Calculate the continuous prior signal-to-noise ratio using the IMCRA algorithm;
[0019] d) Calculate the gain function based on the prior signal-to-noise ratio;
[0020] e) Taking the square of the magnitude of the amplitude spectrum yields the power spectrum of the noisy audio signal;
[0021] f) Multiply the gain function by the power spectrum of the noisy sound to obtain the power spectrum of the enhanced pure sound;
[0022] g) Perform a short-time inverse Fourier transform on the power spectrum of the enhanced pure sound to obtain the enhanced pure sound signal.
[0023] The steps for extracting Mel spectral coefficients from the preprocessed audio signal can be as follows:
[0024] 1) The preprocessed audio signal is framed and windowed;
[0025] 2) For each short-time analysis window, the corresponding power spectrum is obtained through fast Fourier transform and modulus calculation;
[0026] 3) The power spectrum is converted into a Mel spectrum and then logarithmically processed. The conversion of the power spectrum into a Mel spectrum is achieved through a Mel filter bank. Specifically, from low frequency to high frequency, a bandpass filter with a density from dense to sparse is set through the corresponding center frequency to filter the power spectrum. The energy obtained by adding the output signals of each individual filter is used as the basic feature of the corresponding sound signal, i.e., the Mel spectrum. Then, the logarithm of the basic feature is taken to obtain the Mel spectrum coefficients.
[0027] The convolutional neural network adopts a lightweight deep convolutional separable neural network model, ShuffleNet V2. In Unit 1 of ShuffleNet V2, a convolutional attention module is added to the branch that originally did not perform any operation on the branch input to perform channel attention weighting and spatial attention weighting.
[0028] The convolutional neural network consists of a 3×3 convolution, a max pooling layer, three ShuffleNet modules, a Dropout module, a global pooling layer, a fully connected layer, and a softmax activation function. The first and third ShuffleNet modules each contain one unit 2 and three units 1, respectively, while the second ShuffleNet module contains one unit 2 and seven units 1. The input of unit 1 is split into two branches by channel separation. One branch houses the convolutional attention module, and the other branch consists of three convolutions. The three convolutions are a 1×1 convolution, a 3×3 depthwise separable convolution, and a 1×1 convolution, respectively. The outputs of the two branches are concatenated and shuffled to obtain the output of Unit 1. The input of Unit 2 is separated into two branches by channel separation. One branch consists of two convolutions, which are a 3×3 depthwise separable convolution and a 1×1 convolution, respectively. The other branch consists of three convolutions, which are a 1×1 convolution, a 3×3 depthwise separable convolution, and a 1×1 convolution, respectively. The outputs of the two branches are concatenated and shuffled to obtain the output of Unit 2.
[0029] The beneficial effects of this invention are:
[0030] By collecting and identifying the sound signals emitted when the drum cuts coal and rock, it determines whether the signal indicates that the drum has struck the top beam of the support structure. Based on this, the drum height is adjusted in a timely manner to avoid cutting the top beam, thus eliminating the phenomenon of the drum cutting the top beam. The entire process is automated, improving the safety of underground coal mining equipment.
[0031] Using sound recognition technology to automatically avoid the top beam of the support by the drum has a higher recognition rate in the complex environment of high dust and high water mist in coal mine working faces compared with video image recognition, thus greatly extending the life of the drum and even the coal mining machine. Attached Figure Description
[0032] Figure 1 The flowchart shows the self-avoidance method of the top beam of the drum cutting support according to the present invention;
[0033] Figure 2 The flowchart for noise removal through preprocessing is as follows:
[0034] Figure 3 Here is a flowchart of the Mel spectrum coefficient extraction algorithm;
[0035] Figure 4 This is a basic structural diagram of a convolutional neural network classifier;
[0036] Figure 5 This is a structural diagram of Unit 1 of the ShuffleNet module;
[0037] Figure 6 This is a structural diagram of Unit 2 of the ShuffleNet module. Detailed Implementation
[0038] This invention discloses a method for self-avoidance when a drum cuts the top beam of a support. A sound sensor mounted on the coal mining machine drum collects the sound signal (i.e., the original noisy sound signal) of the drum cutting coal and rock. Sound recognition is then used to determine if the sound signal is emitted when the drum cuts the top beam of the support. If so, the drum is immediately controlled to lower its height to avoid the impact. This method employs sound recognition technology. Because the sharp noise generated when the drum cuts the top beam is not altered by the complex environment of high dust and water mist in coal mine working faces, it has better recognition accuracy compared to video / image recognition, which is easily affected by the environment and prone to blind spots. Therefore, the corresponding avoidance method is more reliable and can significantly extend the lifespan of the drum and even the coal mining machine.
[0039] The acquisition and recognition of sound signals are carried out simultaneously with the cutting operation of the coal mining machine. Once it is determined that the drum has cut the top beam of the support, the module used for sound acquisition and recognition immediately transmits information to the electrical control system of the coal mining machine. Then, the electrical control system executes the command to lower the rocker arm to achieve automatic avoidance.
[0040] like Figure 1 As shown, the preferred step of determining whether the sound signal is emitted when the roller cuts the top beam of the support through sound recognition is as follows:
[0041] 1. Preprocessing: The sound signal is preprocessed to remove noise;
[0042] 2. Feature Extraction: Mel spectral coefficients (MFSC, log mel-frequency spectral coefficients) are extracted from the preprocessed audio signal. Then, the subband power distribution (SPD) feature map algorithm is applied to the MFSC to obtain the grayscale spectrum of the subband power distribution (see Jonathan Dennis, HuyDat Tran and Eng Siong Chng, “Image Feature Representation of the Subband Power Distribution for RobustSound Event Classification”, IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGEPROCESSING, VOL.21, NO.2, pp.367-377, FEBRUARY 2013).
[0043] Extracting Mel spectrum coefficients is equivalent to simulating the characteristics of human auditory perception and processing, which helps to improve the sound recognition rate.
[0044] The SPD feature map algorithm is based on the power spectrum distribution of each frequency band over time, which converts the useful sound signal into a continuous region, thereby separating the signal from noise.
[0045] The SPD feature map algorithm is used to highlight the mid-to-low frequency components of sound, and it effectively represents the time, frequency, and energy distribution characteristics of sound signals. The SPD feature map algorithm can extract distribution statistics over long time periods, giving it high temporal resolution and enabling better extraction of distribution features. It better aligns with human auditory characteristics, as it does not compromise between time and frequency resolution, and it outperforms the traditional short-time Fourier transform in representing sound signal characteristics.
[0046] 3. Feature comparison: The grayscale spectrum of the sub-band power distribution is enhanced using a convolutional neural network, and the useful signal is compared with the standard signal of the roller hitting the top beam in the sample library. When the recognition rate is greater than or equal to the set threshold, the corresponding sound signal is identified as the sound signal emitted when the roller cuts the top beam of the support.
[0047] The sound sensor can be installed on the blades of the coal mining machine drum to collect sound signals when the drum cuts coal and rock at closer range.
[0048] When collecting sound signals from the drum cutting coal and rock, it is preferable to collect them at the same frequency.
[0049] While controlling the drum to lower its height to avoid obstacles, it can also issue warning signals. All controls can be completed through the coal mining machine's electrical control system.
[0050] The warning reminders may include voice reminders to facilitate staff receiving warning information.
[0051] The purpose of preprocessing the original noisy audio signal is to remove obvious noise and enhance the sound, resulting in a relatively clean audio signal, such as... Figure 2 As shown, its main steps are:
[0052] a) The original noisy audio signal is framed and windowed to obtain a signal sequence f(n) consisting of several frames of noisy audio signal, where n is the time sequence number and n is a natural number;
[0053] b) Perform a short-time Fourier transform on each frame of the noisy audio signal to obtain the amplitude spectrum F(k) of the noisy audio signal;
[0054] c) Calculate the continuous prior signal-to-noise ratio SNRk(i) using the IMCRA algorithm;
[0055] d) Calculate the gain function based on the prior signal-to-noise ratio, where the gain function is a function of the prior signal-to-noise ratio;
[0056] e) Taking the square of the magnitude of the amplitude spectrum F(k) yields the power spectrum |F(k)| of the noisy audio signal. 2 ;
[0057] f) Using the gain function and the power spectrum |F(k)| of the noisy sound 2 Multiplying them yields the power spectrum of the enhanced, pure sound;
[0058] g) Perform a short-time inverse Fourier transform on the power spectrum of the enhanced pure sound to obtain the enhanced pure sound signal f^(n).
[0059] like Figure 3 As shown, the steps for extracting Mel spectral coefficients from the preprocessed audio signal are as follows:
[0060] 1) The preprocessed audio signal is framed and windowed;
[0061] 2) For each short-time analysis window, the corresponding power spectrum is obtained through fast Fourier transform and modulus calculation;
[0062] 3) Convert the power spectrum to a Mel spectrum and then perform logarithmic operations. The conversion is achieved using a Mel filter bank, specifically by setting bandpass filters with increasingly dense and sparse bandpasses at corresponding center frequencies from low to high frequencies. The power spectrum is filtered by summing the output signals of each individual filter, and the resulting energy is used as the basic characteristic of the corresponding sound signal, i.e., the Mel spectrum. The logarithm of this basic characteristic is then taken to obtain the Mel spectrum coefficients.
[0063] The convolutional neural network employed is a lightweight deep convolutional separable neural network based on the ShuffleNet V2 model. A convolutional attention module is added to the branch in Unit 1 of the ShuffleNet V2 that originally did not perform any operation on the branch input, performing channel attention weighting and spatial attention weighting. This module makes the entire model more focused on the acoustic features during the cutting of the top beam by the drum and improves the information interaction between different channels during training. Therefore, the model selectively strengthens feature information and weakens useless information, enhancing its feature comparison ability on the audio samples of the cut top beam, thus improving model performance while meeting real-time requirements.
[0064] like Figure 4 , 5 As shown in Figure 6, the convolutional neural network sequentially includes a 3×3 convolution, a max pooling layer, three ShuffleNet modules, a Dropout module, a global pooling layer, a fully connected layer, and a normalized exponential softmax activation function. The first and third ShuffleNet modules contain one unit 2 and three units 1, respectively, while the second ShuffleNet module contains one unit 2 and seven units 1. The input of unit 1 is divided into two branches after channel separation. One branch houses the convolutional attention module, and the other branch consists of three units 2 and seven units 1. The input of Unit 2 is composed of three convolutions, namely a 1×1 convolution, a 3×3 depthwise separable convolution, and a 1×1 convolution. The outputs of the two branches are concatenated and shuffled to obtain the output of Unit 1. The input of Unit 2 is divided into two branches after channel separation. One branch consists of two convolutions, namely a 3×3 depthwise separable convolution and a 1×1 convolution. The other branch consists of three convolutions, namely a 1×1 convolution, a 3×3 depthwise separable convolution, and a 1×1 convolution. The outputs of the two branches are concatenated and shuffled to obtain the output of Unit 2.
[0065] In Unit 1, the number of channels in the two branches is half the total number of channels. The channels of the two branches are added together through channel concatenation to fuse the corresponding features, and finally, channel shuffling is performed. Channel shuffling is an operation that reduces the tight connections between channels. The steps are: first, the input image is divided into corresponding subgroups according to channels; then, different convolutional kernels are used to perform group convolutions on each subgroup. This avoids the stacking of group convolutions and prevents the input of a few channels from affecting the output of the entire image.
[0066] In Unit 2, the input feature map is divided into two branches. The input to each branch is the same as the original input map. The left branch uses a 3×3 depthwise separable convolution and a 1×1 regular convolution. The right branch is the same as in Unit 1, consisting of a 3×3 depthwise separable convolution and two 1×1 regular convolutions. Finally, the outputs of the two branches are channel-wise superimposed. Because the number of input channels for each branch is the same as the number of channels in the original input, the sum is twice the number of channels in the original input. This operation expands the number of layers in the network, making Unit 2 extract feature information more effectively than Unit 1. Finally, channel shuffling is used to achieve the exchange of feature information.
[0067] The convolutional neural network acts as a classifier to distinguish between useful and useless signal categories and calculates the probability of the corresponding sound signal category, thereby enabling the detection of whether the drum is cutting the top beam based on sound.
[0068] The convolutional neural network is a trained convolutional neural network.
[0069] The original noisy sound signal corresponding to the standard signal of the roller hitting the top beam in the sample library is the sound signal of the roller hitting the top beam collected on site. In this way, the judgment benchmark of whether the roller has cut the top beam of the support is closer to reality, and the judgment is more reliable.
Claims
1. A method for self-avoidance of the top beam of a roller cutting support, characterized in that: Sound sensors mounted on the coal mining machine drum collect the sound signals of the drum cutting coal and rock. Sound recognition is then used to determine if the sound signal is emitted when the drum cuts the top beam of the support. If so, the drum is immediately lowered to avoid the impact. The sound sensors are installed on the blades of the coal mining machine drum and collect the sound signals of the drum cutting coal and rock at a constant frequency. The steps for determining whether the sound signal is emitted when the drum cuts the top beam of the support are as follows: Preprocessing: The sound signal is preprocessed to remove noise; Feature extraction: Mel spectrum coefficients are extracted from the preprocessed sound signal, and the subband power distribution feature map algorithm is applied to the Mel spectrum coefficients to obtain the subband power distribution grayscale spectrum. Feature comparison: The grayscale spectrum of the sub-band power distribution is enhanced using a convolutional neural network, and the useful signal is compared with the standard signal of the roller hitting the top beam in the sample library. When the recognition rate is greater than or equal to the set threshold, the corresponding sound signal is identified as the sound signal emitted when the roller cuts the top beam of the support. The preprocessing includes the following steps: a) Perform frame segmentation and windowing on the audio signal to obtain a signal sequence consisting of several frames of noisy audio signals; b) Perform a short-time Fourier transform on each frame of the noisy audio signal to obtain the amplitude spectrum of the noisy audio signal; c) Calculate the continuous prior signal-to-noise ratio using the IMCRA algorithm; d) Calculate the gain function based on the prior signal-to-noise ratio; e) Taking the square of the magnitude of the amplitude spectrum yields the power spectrum of the noisy audio signal; f) Multiply the gain function by the power spectrum of the noisy sound to obtain the power spectrum of the enhanced pure sound; g) Perform a short-time inverse Fourier transform on the power spectrum of the enhanced pure sound to obtain the enhanced pure sound signal; The steps for extracting Mel spectral coefficients from the preprocessed audio signal are as follows: 1) The preprocessed audio signal is framed and windowed; 2) For each short-time analysis window, the corresponding power spectrum is obtained through fast Fourier transform and modulus calculation; 3) Convert the power spectrum into a Mel spectrum and then perform logarithmic operations; the conversion of the power spectrum into a Mel spectrum is achieved through a Mel filter bank, specifically from low frequency to high frequency, by setting a bandpass filter from dense to sparse at the corresponding center frequency to filter the power spectrum, and the energy obtained by adding the output signals of each individual filter is used as the basic feature of the corresponding sound signal, i.e., the Mel spectrum, and then taking the logarithm of the basic feature to obtain the Mel spectrum coefficients. The convolutional neural network adopts a lightweight deep convolutional separable neural network model of ShuffleNet V2, and adds a convolutional attention module to the branch that originally did not perform any operation on the branch input in unit 1 of ShuffleNet V2 to perform channel attention weighting and spatial attention weighting. The convolutional neural network consists of a 3×3 convolution, a max pooling layer, three ShuffleNet modules, a Dropout module, a global pooling layer, a fully connected layer, and a softmax activation function. The first and third ShuffleNet modules each contain one unit 2 and three units 1, respectively, while the second ShuffleNet module contains one unit 2 and seven units 1. The input of unit 1 is split into two branches by channel separation. One branch houses the convolutional attention module, and the other branch consists of three convolutions. The three convolutions are a 1×1 convolution, a 3×3 depthwise separable convolution, and a 1×1 convolution, respectively. The outputs of the two branches are concatenated and shuffled to obtain the output of Unit 1. The input of Unit 2 is separated into two branches by channel separation. One branch consists of two convolutions, which are a 3×3 depthwise separable convolution and a 1×1 convolution, respectively. The other branch consists of three convolutions, which are a 1×1 convolution, a 3×3 depthwise separable convolution, and a 1×1 convolution, respectively. The outputs of the two branches are concatenated and shuffled to obtain the output of Unit 2.
2. The method for self-avoidance of the top beam of the drum cutting support as described in claim 1, characterized in that: While controlling the rollers to lower their height to avoid obstacles, the system also issues warning signals.
3. The method for self-avoidance of the top beam of the drum cutting support as described in claim 2, characterized in that: The warnings and alerts include voice prompts.