Audio coal gangue identification method

Abstract

The invention discloses an audio coal gangue identification method. The method comprises the following steps: S1, collecting audio signal data in a coal caving process of a related field of a fully-mechanized face; s2, preprocessing the audio signal data to obtain the audio signal data after noise reduction; s3, performing data calibration on the audio signal features; s4, constructing an audio coal gangue recognition algorithm model; and S5, outputting a coal and gangue recognition result by using the audio coal and gangue recognition algorithm model. According to the invention, a novel audio identification system mainly based on the audio acquisition sensor is utilized, automation of top coal caving of the fully-mechanized caving face is realized through voice identification of coal caving, gangue caving and non-coal caving of the fully-mechanized caving face, and the system has the characteristics of high-precision coal caving and gangue identification and improvement of coal caving efficiency.

Description

Technical Field

[0001] The invention belongs to the field of top coal caving working face mining, and in particular, relates to an audio coal gangue identification method. Background Art

[0002] Coal resources are in a dominant position in my country's energy and occupy a pivotal position in the entire national economy. Comprehensive top coal caving can achieve high yield, high efficiency and low consumption. It is one of the important mining processes in my country's coal mining. Therefore, the research and promotion of top coal caving technology is of particular significance to the development of my country's coal industry. In the process of caving, the control of caving generally adopts the principle of "closing the window when seeing gangue" to control the control valve to close the coal caving window and stop caving. The coal loss of this caving process accounts for 64% of the total loss. In order to improve the top coal caving rate, the caving time is continued for a period of time after the gangue is seen, which is called over-caving. This increases the gangue content and reduces the coal quality. In order to ensure the quality of coal, the caving is stopped immediately after the gangue is seen, which is called under-caving. This results in a lower recovery rate and a waste of coal resources. At present, the caving process of comprehensive top coal caving still relies on manual control. Through experience, the gangue flow is divided into two stages: top coal caving and coal gangue mixed caving. It is difficult to control the top coal caving rate. The coal quality and production efficiency are largely affected by human factors.

[0003] In addition, the labor intensity of top coal caving workers is high and the working environment is harsh, so it is urgent to implement unmanned automatic top coal caving. The first problem to be solved in automatic caving is how to predict coal gangue when top coal is mixed.

[0004] In the Chinese invention patent with the patent number CN202011447649.X, a coal gangue identification system based on heat detection is disclosed, including: A. A coal gangue identification system based on heat detection: a1. N groups of top coal caving hydraulic supports arranged side by side in the same direction, discharging a coal-rock mixture through the coal discharging ports of the top coal caving hydraulic supports, and a transportation passage is provided below each of the top coal caving hydraulic supports; N≥3; a2. A scraper conveyor, located on the transportation passage, for receiving the coal-rock mixture discharged from the (n - 2)-th coal discharging port and transporting it, and successively passing through the coal discharging ports of each top coal caving hydraulic support; 3≤n≤N; a3. N - 1 chemical spraying devices, each chemical spraying device is respectively located on each top coal caving hydraulic support except the first one, and the (n - 2)-th chemical spraying device is used to spray a chemical reagent on the coal-rock mixture transported by the scraper conveyor at the current position; a4. N - 1 infrared heat detection devices, each infrared heat detection device is respectively located on each top coal caving hydraulic support except the first one, and the (n - 2)-th infrared heat detection device is used to detect the heat released after the chemical reaction between the coal-rock mixture and the chemical reagent, obtain the (n - 2)-th heat value, generate a heat signal according to the heat value and convert it into an electrical signal; a5. A controller, respectively connected to each of the chemical spraying devices, each of the infrared heat detection devices and each of the top coal caving hydraulic supports, for controlling the (n - 2)-th coal discharging port, the (n - 2)-th chemical spraying device and the (n - 2)-th infrared heat detection device to close according to the electrical signal, and the (n - 1)-th coal discharging port, the (n - 1)-th chemical spraying device and the (n - 1)-th infrared heat detection device to open. B. A coal gangue identification method based on heat detection: b1. Control the first coal discharging port to open and start coal discharging through the controller; when 3≤n<N, when the coal-rock mixture is transported to the lower part of the (n - 1)-th top coal caving hydraulic support by the scraper conveyor, detect the heat released by the coal-rock mixture and the chemical reagent through the (n - 2)-th infrared heat detection device, obtain the (n - 2)-th heat value, generate a heat signal according to the heat value and convert it into an electrical signal; b2. The controller controls the (n - 2)-th coal discharging port, the (n - 2)-th infrared heat detection device and the (n - 2)-th chemical spraying device to close according to the electrical signal, and controls the coal discharging port of the (n - 1)-th top coal caving hydraulic support to open, the (n - 1)-th infrared heat detection device to open, and the (n - 1)-th chemical spraying device to open; b3. When n = N, the N-th coal discharging port does not discharge coal, the controller controls the (N - 1)-th coal discharging port, the (N - 1)-th infrared heat detection device and the (N - 1)-th chemical spraying device to close, and the coal discharging ends.

[0005] The defect in the existing patent is that although the detection of coal gangue identification is achieved, the identification speed is slow and the accuracy is poor, and the quality of coal production and the recovery rate of top coal cannot be improved. Summary of the Invention

[0006] Aiming at the problem that the existing coal gangue identification method has slow identification speed and poor accuracy and cannot improve the coal production quality and top coal recovery rate, the present invention provides an audio coal gangue identification method.

[0007] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is as follows:

[0008] An audio coal gangue recognition method comprises the following steps:

[0009] S1. Collect audio signal data during coal caving at the fully mechanized mining face;

[0010] S2, preprocessing the audio signal data to obtain the audio signal data after noise reduction;

[0011] S3, performing data calibration on audio signal characteristics;

[0012] S4, constructing an audio coal gangue recognition algorithm model;

[0013] S5, outputting coal gangue recognition results using the audio coal gangue recognition algorithm model;

[0014] Furthermore, the detailed steps of collecting the audio signal data during the on-site coal caving process of the fully mechanized mining face in step S1 include:

[0015] S11, collecting sound signals of coal gangue flows with different gangue contents hitting the tail beam of the hydraulic support during the complete top coal caving process through the sound pickup module, and processing the sound signals through the audio coding chip to obtain digital audio signals as training sets;

[0016] The whole top coal caving process is divided into multiple coal caving stages, and the gangue content of the coal gangue flow in each coal caving stage changes continuously;

[0017] Furthermore, step S2 pre-processes the audio signal data, and the detailed steps of obtaining the audio signal data after noise reduction include:

[0018] S21, performing noise reduction processing on the digital audio signal in the training set through a generative adversarial network, then performing feature extraction to obtain a feature matrix, storing the feature matrix, and transmitting the feature matrix corresponding to each type of audio to the STM32 embedded module;

[0019] S22. In the STM32 embedded module, the feature matrix is ​​input into the constructed neural network model for training;

[0020] S23, STM32 embedded module sends control signal through wireless communication module to control the coal placing window of hydraulic support to open and start the automatic top coal placing process;

[0021] The S24 and STM32 embedded modules input the feature matrix of the sound signal picked up during the automatic top coal caving process into the trained model, recognize the sound signal in the field environment, and output the recognition result to the handheld terminal;

[0022] S25, monitoring the recognition result, when the gangue content corresponding to the sound of the hydraulic support tail beam of the coal gangue flow loader reaches the set ratio, the STM32 embedded module sends a control signal through the wireless communication module to control the operation of the operating mechanism and close the coal window; when the gangue content corresponding to the recognized sound does not reach the set ratio, the STM32 embedded module does not send a control signal and returns to S24;

[0023] In S21, feature extraction includes the following sub-steps:

[0024] S201, calculating the frequency spectrum contrast feature of the audio signal;

[0025] The audio signal is framed and each frame is short-time Fourier transformed. The window size of the short-time Fourier transform is 1024 and the frame shift is 512. The transform result is then divided into 6 sub-bands. For each sub-band, the energy contrast is estimated by comparing the average energy of the top quantile and the bottom quantile, and the log mean is taken as the peak value. k ; Then calculate the valley value Valley k , then calculate the difference SC k =Peak k -Valley k , and finally {SC k , Valley k} as the original spectrum contrast feature;

[0026] S202, calculating the chromaticity characteristics of the audio spectrum;

[0027] First, the audio signal is framed and each frame is short-time Fourier transformed. The window size of the short-time Fourier transform is 1024 and the frame shift is 512, so that the amplitude spectrum of the signal is obtained. Then, the amplitude spectrum is peak detected, that is, the largest first npeaks peaks within the fixed frequency range are found, the difference between any two peaks is calculated, and then the base frequency f of the original signal is obtained by combining the standard frequency specified in the twelve-tone equal temperament. ref ; Finally, the chromaticity feature vector is obtained by mapping each half-order tone class through the spectrum peak energy. The calculation formula for the element value of each dimension in the vector is:

[0028]

[0029] Among them, a i and f iare the amplitude and frequency of the i-th peak, wn, f i The frequency is f i The weight of the signal for the half-order tone class n is calculated as:

[0030] First determine the central reference frequency f of each semitone class n :

[0031] f n =f ref 2 n / 12 , n=1,2,……12

[0032] Define each signal peak frequency f i and the center reference frequency f of the semi-tone class n The interval distance between them is:

[0033]

[0034] Among them, m is an integer adjustment factor that minimizes the value of |d|. The weight value is calculated according to the following formula:

[0035]

[0036] Where, l is the preset weighted window length;

[0037] S203, using the spectrum contrast feature of the sound signal obtained in S201 to extract features of the original audio signal, to obtain a corresponding spectrum contrast feature matrix;

[0038] The chromaticity feature of the sound spectrum obtained in S202 is used to extract features of the original audio signal to obtain a corresponding chromaticity feature matrix;

[0039] The extracted spectrum contrast feature matrix and the chromaticity feature matrix of the sound spectrum are spliced ​​in the vertical direction, that is, spliced ​​in the time dimension, and the time dimension, frequency dimension, energy dimension and timbre dimension of the audio signal are fused to obtain the original audio signal fusion feature matrix, and then the fusion feature matrix of the adjacent two-frame signal after smoothing and the fusion feature matrix of the adjacent three-frame signal after smoothing are calculated based on the fusion feature matrix, and the neural network model is trained using the fusion feature matrix to obtain the coal gangue recognition model after training;

[0040] In S22, the neural network model is an MBCNN-A model of a three-branch convolutional neural network that integrates an attention mechanism, and the model includes an input layer, a hidden layer, a concatenate layer, a BatchNormalization layer, a multi-head attention layer, a fully connected layer, and an output layer in sequence. The hidden layer includes three parallel branches, and each branch includes a convolution layer, a BatchNormalization layer, a global maximum pooling layer, and a Dropout layer structure in sequence; the stride of the convolution layer in the three branches is 1, the convolution kernel sizes are 3, 4, and 5 respectively, and the number of filters is 128;

[0041] During model training, the original fusion feature matrix, the smoothed fusion feature matrix of two adjacent frames, and the smoothed fusion feature matrix of three adjacent frames are output to the branches corresponding to the convolution kernels 3, 4, and 5 respectively. The outputs of the three branches are concatenated and fused in the concatenate layer, and then pass through the multi-head attention layer. Finally, the output features of the multi-head attention layer are concatenated with the output features of the concatenate layer in the form of residual branches and output to the output layer through the fully connected layer.

[0042] Furthermore, the detailed steps of performing data calibration on the audio signal features in step S3 include:

[0043] S301, retaining useful information and discarding useless information from the algorithm unit state through the forget gate layer;

[0044] S302, determine the value to be updated through the input gate layer;

[0045] S303, creating a new candidate value in the tanh layer and adding it to the algorithm unit state;

[0046] S304, determining the output value of the algorithm unit through the output gate and the unit state.

[0047] Furthermore, the detailed steps of constructing the audio gangue recognition algorithm model in step S4 and outputting the gangue recognition result using the audio information include:

[0048] S401, inputting the sum of the audio data within the time Δt into the algorithm network;

[0049] S402, the input data is mapped into a multi-dimensional vector through the input embedding layer;

[0050] S403, the multi-dimensional vector is input into multiple layers of algorithm units for processing, and the output of each layer of algorithm units is used as the input of the next layer of algorithm units.

[0051] S404, the output of the algorithm unit of the last layer is mapped into a one-dimensional data through the fully connected layer, which is the predicted value of the current working state.

[0052] Furthermore, step S5 can identify the target results of the hydraulic support: coal releasing state, gangue releasing state, and neither coal releasing nor gangue releasing state. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is an overall flow chart of an audio coal gangue identification method of the present invention;

[0054] Figure 2 This is a flow chart of the noise reduction process in the present invention;

[0055] Figure 3 Schematic diagram of the MBCNN-A model structure in the present invention;

[0056] Figure 4 This is a comparison chart of coal gangue identification accuracy using different models in the present invention;

[0057] Figure 5 This is a comparison chart of loss values ​​using different models in the present invention. DETAILED DESCRIPTION

[0058] In order to facilitate the understanding of those skilled in the art, the present invention is further described below in conjunction with embodiments and drawings. The contents mentioned in the implementation modes are not intended to limit the present invention.

[0059] like Figure 1 As shown, this embodiment provides an audio coal gangue recognition method, comprising the steps of:

[0060] S1. Collect audio signal data during the coal caving process at the fully mechanized mining face; S2. Preprocess the audio signal data to obtain the audio signal data after noise reduction; S3. Calibrate the audio signal features; S4. Construct an audio gangue recognition algorithm model and output the gangue recognition results using audio information; S5. Calculate the hydraulic support target recognition results;

[0061] In this embodiment, a pickup is installed on the hydraulic support side of the fully mechanized caving working face to collect the sound of coal and gangue caving on site.

[0062] Furthermore, the detailed steps of collecting the audio signal data during the on-site coal caving process of the fully mechanized mining face in step S1 include:

[0063] S11, collecting sound signals of coal gangue flows with different gangue contents hitting the tail beam of the hydraulic support during the complete top coal caving process through the sound pickup module, and processing the sound signals through the audio coding chip to obtain digital audio signals as training sets;

[0064] The whole top coal caving process is divided into multiple coal caving stages, and the gangue content of the coal gangue flow in each coal caving stage changes continuously;

[0065] Furthermore, step S2 pre-processes the audio signal data, and the detailed steps of obtaining the audio signal data after noise reduction include:

[0066] S21, performing noise reduction processing on the digital audio signal in the training set through a generative adversarial network, then performing feature extraction to obtain a feature matrix, storing the feature matrix, and transmitting the feature matrix corresponding to each type of audio to the STM32 embedded module;

[0067] S22. In the STM32 embedded module, the feature matrix is ​​input into the constructed neural network model for training;

[0068] S23, STM32 embedded module sends control signal through wireless communication module to control the coal placing window of hydraulic support to open and start the automatic top coal placing process;

[0069] The S24 and STM32 embedded modules input the feature matrix of the sound signal picked up during the automatic top coal caving process into the trained model, recognize the sound signal in the field environment, and output the recognition result to the handheld terminal;

[0070] S25, monitoring the recognition result, when the gangue content corresponding to the sound of the hydraulic support tail beam of the coal gangue flow loader reaches the set ratio, the STM32 embedded module sends a control signal through the wireless communication module to control the operation of the operating mechanism and close the coal window; when the gangue content corresponding to the recognized sound does not reach the set ratio, the STM32 embedded module does not send a control signal and returns to S24;

[0071] In S21, feature extraction includes the following sub-steps:

[0072] S201, calculating the frequency spectrum contrast feature of the audio signal;

[0073] The audio signal is framed and each frame is short-time Fourier transformed. The window size of the short-time Fourier transform is 1024 and the frame shift is 512. The transform result is then divided into 6 sub-bands. For each sub-band, the energy contrast is estimated by comparing the average energy of the top quantile and the bottom quantile, and the log mean is taken as the peak value. k ; Then calculate the valley value Valley k , then calculate the difference SC k =Peak k -Valley k , and finally {SC k , Valley k} as the original spectrum contrast feature;

[0074] S202, calculating the chromaticity characteristics of the audio spectrum;

[0075] First, the audio signal is framed and each frame is short-time Fourier transformed. The window size of the short-time Fourier transform is 1024 and the frame shift is 512, so the amplitude spectrum of the signal is obtained. Then, the amplitude spectrum is peak detected, that is, the largest first npeaks peaks within the fixed frequency range are found, the difference between any two peaks is calculated, and then the base frequency f of the original signal is obtained by combining the standard frequency specified in the twelve-tone equal temperament. ref ; Finally, the chromaticity feature vector is obtained by mapping each half-order tone class through the spectrum peak energy. The calculation formula for the element value of each dimension in the vector is:

[0076]

[0077] Among them, a i and f i are the amplitude and frequency of the i-th peak, wn, f i The frequency is f i The weight of the signal for the half-order tone class n is calculated as:

[0078] First determine the central reference frequency f of each semitone class n :

[0079] f n =f ref 2 n / 12 , n=1,2,……12

[0080] Define each signal peak frequency f i and the center reference frequency f of the semi-tone class n The interval distance between them is:

[0081]

[0082] Among them, m is an integer adjustment factor that minimizes the value of |d|. The weight value is calculated according to the following formula:

[0083]

[0084] Where, l is the preset weighted window length;

[0085] S203, using the spectrum contrast feature of the sound signal obtained in S201 to extract features of the original audio signal, to obtain a corresponding spectrum contrast feature matrix;

[0086] The chromaticity feature of the sound spectrum obtained in S202 is used to extract features of the original audio signal to obtain a corresponding chromaticity feature matrix;

[0087] The extracted spectrum contrast feature matrix and the chromaticity feature matrix of the sound spectrum are spliced ​​in the vertical direction, that is, spliced ​​in the time dimension, and the time dimension, frequency dimension, energy dimension and timbre dimension of the audio signal are fused to obtain the original audio signal fusion feature matrix, and then the fusion feature matrix of the adjacent two-frame signal after smoothing and the fusion feature matrix of the adjacent three-frame signal after smoothing are calculated based on the fusion feature matrix, and the neural network model is trained using the fusion feature matrix to obtain the coal gangue recognition model after training;

[0088] In S22, the neural network model is an MBCNN-A model of a three-branch convolutional neural network that integrates an attention mechanism, and the model includes an input layer, a hidden layer, a concatenate layer, a BatchNormalization layer, a multi-head attention layer, a fully connected layer, and an output layer in sequence. The hidden layer includes three parallel branches, and each branch includes a convolution layer, a BatchNormalization layer, a global maximum pooling layer, and a Dropout layer structure in sequence; the stride of the convolution layer in the three branches is 1, the convolution kernel sizes are 3, 4, and 5 respectively, and the number of filters is 128;

[0089] During model training, the original fusion feature matrix, the smoothed fusion feature matrix of two adjacent frames, and the smoothed fusion feature matrix of three adjacent frames are output to the branches corresponding to the convolution kernels 3, 4, and 5 respectively. The outputs of the three branches are concatenated and fused in the concatenate layer, and then pass through the multi-head attention layer. Finally, the output features of the multi-head attention layer are concatenated with the output features of the concatenate layer in the form of residual branches and output to the output layer through the fully connected layer.

[0090] The S22 and STM32 embedded modules send control signals through the wireless communication module to control the coal placing window of the hydraulic support to open and start the automatic top coal placing process;

[0091] S23, the STM32 embedded module inputs the feature matrix of the sound signal picked up during the automatic top coal caving process into the trained model, recognizes the sound signal in the field environment, and outputs the recognition result to the handheld terminal;

[0092] The MBCNN-A model can also output the coal release rate, so by judging the coal release rate, problems such as coal jamming during coal release can be eliminated.

[0093] S24, monitoring the recognition result, when the gangue content corresponding to the sound of the hydraulic support tail beam of the coal gangue flow loader reaches the set ratio, the STM32 embedded module sends a control signal through the wireless communication module to control the operation of the operating mechanism and close the coal window; when the gangue content corresponding to the recognized sound does not reach the set ratio, the STM32 embedded module does not send a control signal and returns to S253;

[0094] like Figure 4 and 5 As shown, models of the same type such as 1DCNN, 1DCNN-MA, 2DCNN, and 2DCNN-SA are used as comparison models, and accuracy and loss value are selected as evaluation indicators. The predicted distribution output by the model proposed in the present invention is closer to the true distribution of real samples, the error loss between the predicted value and the true value is smaller, and the model converges faster, with better accuracy.

[0095] Compared with the prior art, the present invention has the following beneficial effects:

[0096] Through filtering and denoising using technologies such as microphone array, wavelet transform, and MFCC matrix, effective sound is obtained. The classification model of coal caving status (coal caving, gangue caving, and no coal caving) is trained using the artificial intelligence-based LSTM algorithm, and linked with the electro-hydraulic control device, combined with the on-site coal caving process, coal caving automation is achieved, thereby achieving the goal of reducing manpower and increasing efficiency in the fully-mechanized caving face and reducing the gangue content of the coal output.

[0097] The above is a detailed introduction to an audio coal gangue identification method provided by the present application. The description of the specific embodiment is only used to help understand the method and its core idea of ​​the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. An audio coal gangue identification method, characterized in that: Includes steps: S1. Collect audio signal data during coal caving at the fully mechanized mining face; S2, preprocessing the audio signal data to obtain the audio signal data after noise reduction; S3, performing data calibration on audio signal characteristics; S4, constructing an audio coal gangue recognition algorithm model; S5. Use the audio gangue recognition algorithm model to output the gangue recognition result.

2. The audio coal gangue identification method according to claim 1, characterized in that: The detailed steps of step S1 for collecting audio signal data during the on-site coal caving process of the fully mechanized mining face include: S11, collecting sound signals of coal gangue flows with different gangue contents hitting the tail beam of the hydraulic support during the complete top coal caving process through the sound pickup module, and processing the sound signals through the audio coding chip to obtain digital audio signals as training sets; The entire top coal caving process is divided into multiple coal caving stages, and the gangue content of the coal gangue flow in each coal caving stage changes continuously.

3. The audio coal gangue identification method according to claim 1, characterized in that: Step S2 pre-processes the audio signal data, and the detailed steps of obtaining the audio signal data after noise reduction include: S21, performing noise reduction processing on the digital audio signal in the training set through a generative adversarial network, then performing feature extraction to obtain a feature matrix, storing the feature matrix, and transmitting the feature matrix corresponding to each type of audio to the STM32 embedded module; S22. In the STM32 embedded module, the feature matrix is ​​input into the constructed neural network model for training; S23, STM32 embedded module sends control signal through wireless communication module to control the coal placing window of hydraulic support to open and start the automatic top coal placing process; The S24 and STM32 embedded modules input the feature matrix of the sound signal picked up during the automatic top coal caving process into the trained model, recognize the sound signal in the field environment, and output the recognition result to the handheld terminal; S25. Monitor the recognition results. When the gangue content corresponding to the sound of the hydraulic support tail beam of the coal gangue flow loader reaches the set ratio, the STM32 embedded module sends a control signal through the wireless communication module to control the action of the operating mechanism and close the coal window; when the gangue content corresponding to the recognized sound does not reach the set ratio, the STM32 embedded module does not send a control signal and returns to S24.

4. The audio coal gangue identification method according to claim 1, characterized in that: Step S3 includes the following detailed steps for data calibration of audio signal features: S301, retaining useful information and discarding useless information from the algorithm unit state through the forget gate layer; S302, determine the value to be updated through the input gate layer; S303, creating a new candidate value in the tanh layer and adding it to the algorithm unit state; S304, determining the output value of the algorithm unit through the output gate and the unit state.

5. The audio coal gangue identification method according to claim 1, characterized in that: The detailed steps of constructing the audio gangue recognition algorithm model in step S4 include: S401, inputting the sum of the audio data within the time Δt into the algorithm network; S402, the input data is mapped into a multi-dimensional vector through the input embedding layer; S403, the multi-dimensional vector is input into a multi-layer algorithm unit for processing, and the output of each layer of algorithm unit is used as the input of the next layer of algorithm unit; S404, the output of the algorithm unit of the last layer is mapped into a one-dimensional data through the fully connected layer, which is the predicted value of the current working state.

6. According to the audio coal gangue recognition method of claim 1, step S5 can identify the target results of the hydraulic support: coal release state, gangue release state, and neither coal release nor gangue release state.

Images