A rock mass JRC measurement device and method based on friction sound and deep learning

Abstract

The present invention provides a rock mass JRC measurement device and method based on friction sound and deep learning, including a probe end, an audio collection end, a data processing end, a control end and a power module. The probe end is provided with a detachable paddle for sliding on the surface of rock mass joints; the audio collection end is equipped with a recording component and a storage device for recording and storing the sound signal generated when the probe end slides on the surface of the rock mass, and transmitting the sound data to the data processing end; the data processing end pre-processes and extracts features of the audio, and identifies the features through a trained deep neural network model to obtain the roughness coefficient of the corresponding rock mass joint; the control end is used to display the measured joint roughness coefficient, power, and perform file naming, exporting and other operations. The present invention can measure the roughness coefficient of rock mass joints at low cost, quickly and with high automation, and provides convenience for measuring the roughness coefficient of rock mass joints in the field and in engineering.

Description

Technical Field

[0001] The invention relates to the fields of geotechnical engineering and artificial intelligence, and belongs to an automatic measuring device for rock mass joint roughness coefficient. Background Art

[0002] In geotechnical engineering, the mechanical properties of rock mass depend on the rock material and its structural planes (joints, faults, fissures, etc.) that constitute the rock mass. The structural plane is also called the weak plane. Its shear failure usually occurs before the rock material is crushed and is the main factor affecting the strength, stability and seepage characteristics of the rock mass. At present, the most commonly used method for evaluating the shear strength of rock mass joints is the JRC-JCS empirical model proposed by Barton. Therefore, the measurement of JRC (Joint Roughness Coefficient) is of great significance to the calculation of rock mass strength.

[0003] The most commonly used method for evaluating rock joint roughness is the comparison method, which is to estimate the JRC of any joint by comparing the 10 standard joint profile curves given by Barton and Choubey. This method is simple and fast, but it has strong subjectivity and human errors, and requires high experience from engineers. In addition, there are other commonly used JRC measurement methods, such as the straight edge method, the roughness statistical parameter method, and the fractal dimension method. However, these methods are very inconvenient to use at engineering sites or field surveys due to the large measurement workload, complex calculation process, expensive and bulky equipment, and other reasons, and have certain technical application limitations.

[0004] Therefore, it is of great significance to propose a method for measuring the roughness coefficient of rock joints that is suitable for engineering and field exploration.

[0005] In recent years, with the development of computer science, deep learning has made great achievements in image recognition, voiceprint recognition, regression prediction, etc., and has been widely used in various industries. Deep learning can analyze and learn data and automatically identify data by learning the inherent laws and representation levels of sample data. It has a wide range of applications and has the advantages of high degree of automation and universality. Inspired by the success of deep learning, this paper uses its excellent ability in regression prediction and pattern recognition to calculate the roughness coefficient of rock joints. Summary of the invention

[0006] The present invention aims to provide a method and device for measuring the roughness coefficient of rock joints based on friction sound and deep learning. Based on deep learning technology, the method extracts features and recognizes the friction sound of a pick passing over the surface of rock joints, and corrects the results through a correction algorithm, thereby realizing automatic and intelligent measurement of the roughness coefficient of rock joints.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A rock mass JRC measurement device based on friction sound and deep learning, comprising a measurement device body, the measurement device body having a probe end, an audio acquisition end, a data processing end, a control end and a power module;

[0009] The probe end includes a fixture and a pick. The pick is installed on the fixture and is used to scratch the surface of the rock joint to generate a sound signal.

[0010] The audio collection end is close to the probe end, including a recording component, a data storage center, and a USB interface. The recording component is used to collect and transmit the sound signal when the pick passes through the surface of the rock joint. The recording component stores the sound signal in the data storage center and exports it through the USB interface.

[0011] The data processing end is located in the main body of the measuring device. The data processing end is connected to the data storage center and the control end respectively, and is used to process the sound signal of the data storage center and display the results on the control end. The data processing end includes the following core algorithms: audio feature extraction algorithm, neural network classification algorithm and correction algorithm. The audio feature extraction algorithm includes the HMFCC extraction algorithm of audio, the short-time Fourier transform feature extraction algorithm and the fusion algorithm of 6 sound features (chromaticity frequency, energy root mean square, spectrum centroid, bandwidth, spectrum attenuation, zero crossing rate), which are used for preprocessing and feature extraction of sound; the neural network classification algorithm is a trained multi-layer two-dimensional convolutional neural network and a fully linked neural network algorithm, which are used to analyze and identify the extracted features and calculate the corresponding joint roughness coefficient and probability; the correction algorithm is used to correct the results obtained by the above-mentioned deep neural network classification algorithm, and the corrected results are used as the joint roughness coefficient of the sample to be tested as output;

[0012] The control end is connected to the power module, the data storage center and the data processing end respectively, and is used to control the data processing end to process data, read the remaining power in the power module, name the files in the data storage center, and display the relevant results on the display screen of the control end. The control end is used to display the measured joint roughness coefficient, power, and perform file naming and export operations;

[0013] The power module includes a power management chip, a battery, and a charging interface. The charging interface is connected to the battery for charging. The power management chip is connected to the battery and the control end to read the battery status and transmit the information to the control end.

[0014] A measurement method of a rock mass JRC measurement device based on friction sound and deep learning, comprising the following steps:

[0015] Step 1: Press and hold the recording button, use the pick to slide across the surface of the joint to be measured to record, and release the recording button after recording is completed;

[0016] Step 2: Name the recording file on the control end display screen, and control the data processing end to process the selected recording file;

[0017] Step 3: The data processing end first uses an audio feature extraction algorithm to resample and extract features from the recording file. The extracted features include high-order Mel-frequency cepstral coefficients (HMFCC), short-time Fourier transform (STFT), and fusion features of six sound features (chrominance frequency, root mean square energy, spectrum centroid, bandwidth, spectral attenuation, and zero-crossing rate);

[0018] Step 4: After extracting the features, use the trained two-layer two-dimensional convolutional neural network and fully connected neural network algorithms to analyze and identify the features, and output their corresponding joint roughness coefficients and probabilities;

[0019] Step 5: Finally, the data processing end uses a correction algorithm to correct the result based on probability and transmits the corrected result to the control end; the correction algorithm judges the probability of the prediction results of the above three neural networks. If the maximum prediction probability is greater than 0.5, the prediction result corresponding to the maximum probability is output; otherwise, the prediction results corresponding to the highest two probabilities are taken and the average value is output;

[0020] Step 6: Read the calculation results on the control terminal display screen.

[0021] Beneficial Effects

[0022] 1) The present invention measures the roughness coefficient by using the friction sound of a pick across the surface of rock joints and a trained neural network. The method is simple, fast, and highly automated. The equipment is portable and low-cost, and is convenient for simple measurement of the roughness coefficient of rock joints in preliminary engineering surveys or in the field.

[0023] 2) The present invention extracts a variety of sound features and provides a variety of neural network algorithms and correction algorithms with high accuracy.

[0024] 3) The present invention is provided with a data storage center and a USB interface, which can upload the measured samples to the terminal, and continuously train and optimize the network by expanding the samples, thereby improving the accuracy of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a structural schematic diagram of the present invention;

[0026] Figure 2Part of the algorithm flow chart of the present invention, wherein (a) is a fully connected neural network algorithm diagram, (b) is a HMFCC extraction algorithm diagram, (c) is a spectrogram extraction algorithm diagram, and (d) is a convolutional neural network classification algorithm diagram;

[0027] Figure 3 It is a system framework diagram of the present invention. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0029] A rock mass joint roughness coefficient measuring device based on friction sound and deep learning, which uses a simple and easy-to-measure sound signal as an input signal, and analyzes and processes it through a deep learning model with a high degree of automation to measure the rock mass joint roughness coefficient. It has the advantages of low cost, high speed, and portable equipment, and is convenient for field and engineering measurements;

[0030] The measuring device comprises a main body 1, which is provided with a probe end, an audio collection end, a data processing end 7, a control end 8 and a power supply module;

[0031] The probe end includes a fixture 2 and a paddle 3. The paddle 3 is mounted on the fixture 2. The paddle 3 is used to scratch the surface of the rock joint to generate a sound signal. The paddle 3 is replaced after being deformed and damaged by repeated scratching.

[0032] You can buy any clamp on the market and install it yourself. The pick is a soft nylon yueqin pick. To improve the accuracy of sound recognition, other types of picks should be avoided.

[0033] The audio collection end is close to the probe end, and includes a recording component 4, a data storage center 5, a USB interface 6, and a recording button 11. The recording component 4 is used to collect and transmit the sound signal when the pick 3 passes through the rock joint surface. The recording component 4 stores the sound signal in the data storage center 5 and exports it through the USB interface 6. The recording button 11 is used to control the start and end of the recording;

[0034] There are no special requirements for the relevant components. Optionally, in order to improve the sound recording effect, a sound enhancement and noise reduction device can also be provided.

[0035] The data processing terminal 7 is located in the measuring device body 1. The data processing terminal extracts and analyzes the features of the collected audio through the following algorithm to achieve the measurement of the rock joint roughness coefficient. The data processing terminal 7 includes the following core algorithms: audio feature extraction algorithm, neural network classification algorithm and correction algorithm. The audio feature extraction algorithm includes the audio HMFCC extraction algorithm, short-time Fourier transform feature extraction algorithm and 6 sound feature (chromaticity frequency, energy root mean square, spectrum centroid, bandwidth, spectrum attenuation, zero crossing rate) fusion algorithm, which are used to pre-process and extract features of the sound; the neural network classification algorithm is a trained multi-layer two-dimensional convolutional neural network and a fully linked neural network algorithm, which are used to analyze and identify the extracted features and calculate the corresponding joint roughness coefficient and probability; the correction algorithm is used to correct the results obtained by the above-mentioned deep neural network classification algorithm, and the corrected results are used as the joint roughness coefficient of the sample to be tested as the output;

[0036] The data processing end is the most innovative and core part of the present invention, including three core algorithms: audio feature extraction algorithm, neural network classification algorithm and correction algorithm, which are used to identify and classify sound features. The specific steps and formulas of the above three algorithms are as follows:

[0037] 1) Audio feature - spectrogram extraction algorithm: resample the original audio signal with a sampling frequency of 44.1kHz and convert it into mono; divide the audio signal into frames with a frame length of 1024, shift the frame by 502, and use the Hann window (Formula 1.1) to window the framed signal; perform fast Fourier transform (FFT) on each frame of the signal to extract its frequency domain information (Formula 1.2); splice the signal after each FFT frame (short-time Fourier transform STFT), assign different colors according to their amplitudes, obtain the spectrogram and convert it into a grayscale image.

[0038]

[0039] 2) Audio features - HMFCC extraction algorithm: After obtaining the STFT spectrum in the spectrogram extraction algorithm, the power spectrum is obtained through a square operation and multiplied with the Mel filter group (Formula 2.1-2.3) to obtain the Mel spectrum; a discrete cosine transform is performed to obtain 13 Mel-frequency cepstral coefficients (Formula 2.4); first-order and second-order difference operations are performed, and the results are combined to obtain high-order Mel-frequency cepstral coefficients.

[0040]

[0041] 3) Audio features - feature fusion extraction algorithm: extract the chromatogram (Formula 3.1), energy root mean square (Formula 3.2), spectrum centroid (Formula 3.3), bandwidth (Formula 3.4), sound spectrum attenuation (Formula 3.5), and zero-crossing rate (Formula 3.6) respectively; each feature is averaged and normalized respectively, and the six features are combined.

[0042]

[0043]

[0044] 4) Neural network classification algorithm: The present invention uses a trained neural network, and the neural network models used are a two-dimensional convolutional neural network and a deep neural network. During the training process, three audio features are used as input, the ReLU function is used as the nonlinear activation function (Formula 4.1), and JRC is used as the output. The network performance is evaluated by the loss function multi-classification cross entropy (4.2), and the network weights are updated based on the Adam optimization algorithm (4.3-4.5).

[0045]

[0046] m t =β 1 ·m t-1 +(1-β 1 )·g t (4.3)

[0047]

[0048] 5) Correction algorithm: The three neural networks output three prediction results and their prediction probabilities respectively. The correction algorithm corrects and outputs the three prediction results based on probability. The idea is that if the maximum prediction probability is greater than 0.5, the prediction result corresponding to the maximum probability is output; otherwise, the prediction results corresponding to the highest two probabilities are taken and the average value is output (Formula 5.1).

[0049]

[0050] Control terminal 8, which is used to display the measured joint roughness coefficient, electrical quantity, and perform file naming and export operations;

[0051] The power module is used for charging and power supply, and includes a power management chip, a battery 9, and a charging interface 10. The power module is located at the top of the device.

[0052] The algorithm of the present invention comprises an audio feature extraction algorithm, a neural network classification algorithm and a correction algorithm.

[0053] 1) Audio feature extraction algorithm: used to preprocess and extract features of audio. The extracted features are used as input parameters of the neural network classification algorithm.

[0054] 2) Neural network classification algorithm: There are multiple trained neural networks to identify the extracted sound features and output the calculated rock joint roughness coefficient and corresponding probability.

[0055] 3) Correction algorithm: The results are corrected based on the rock joint roughness coefficient and probability calculated by the neural network classification algorithm to reduce the error of the results.

[0056] System framework (such as Figure 3 As shown) as follows:

[0057] 1) Preprocess the sound signal collected by the audio acquisition end, including digitization, time length unification, resampling, framing, windowing, etc. In this solution, the sampling frequency is 44.1kHz, the frame length is 1024 (23ms), and the frame shift is 1 / 2 of the frame length and the window is the Hann window.

[0058] 2) Extract features from the preprocessed sound. The features extracted in this solution include 6 feature fusions (chrominance frequency, energy root mean square, spectrum centroid, bandwidth, spectrum attenuation, zero crossing rate), HMFCC features, and spectrogram features. The feature extraction algorithm is as follows: Figure 2 shown.

[0059] 3) The extracted features are put into the trained neural network for learning, and the predicted JRC value and its probability are output. In this scheme, three sound features are extracted, so there are three corresponding trained neural networks, namely one fully connected neural network and two two-dimensional convolutional neural networks. The network model is as follows Figure 2 shown.

[0060] 4) The result is corrected according to the JRC value and its probability obtained by the neural network, and the result with the highest probability is selected as the final JRC prediction value and displayed on the control end.

[0061] Control terminal such as Figure 1 As shown, the control end is mainly used for displaying results, power and other related contents and for operations such as file naming.

[0062] Power module such as Figure 1 As shown, the power module includes a power management chip, a battery, a charging interface, and a switch. The charging interface can be a USB data line interface, a Type-C data line interface, etc.

[0063] In the embodiment of the present invention, the user should use the paddle at the end of the probe to slide on the surface of the rock joint to be measured at a speed of about 2 cm / s, and press and release the recording button 11 at the beginning and end of the sliding, respectively. After the recorded sound is processed by the system, the roughness coefficient of the joint to be measured can be viewed on the display screen of the control end. After the measurement is completed, the sound file can be named and exported, and compared with the results of the detailed survey, and the final result can be uploaded to the terminal database, thereby increasing the number of database samples, further improving the classification accuracy of the model, and facilitating subsequent research and model optimization.

[0064] Special note: When using this device to slide on the surface of rock joints, it should be kept as stable and uniform as possible, and the speed should be controlled at about 2cm / s, otherwise it may cause inaccurate measurement. The measured JRC can only be used for preliminary investigation and cannot be used for structural design.

[0065] This device can only measure joints with a length of about 10cm for the time being. Joints of different lengths can be superimposed by algorithms in the later stage. In a noisy construction environment, noise reduction equipment can be installed on the audio acquisition end for optimization. This device is suitable for preliminary surveys. When there are enough samples, the accuracy of the network can be increased.

[0066] In summary, the joint roughness coefficient measurement method and device provided by the embodiment of the present invention are more objective, accurate, and rapid than traditional methods, and are cheaper and more portable than scanning methods. They make up for the shortcomings of existing JRC measurement methods and devices in engineering applications and field explorations, and combine deep learning with engineering measurement, providing a new idea for JRC measurement, which is conducive to the development and application of smart construction and intelligent monitoring.

[0067] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A measurement method of a rock mass JRC measurement device based on friction sound and deep learning, the measurement device comprising a measurement device body (1), the measurement device body (1) having a probe end, an audio collection end, a data processing end (7), a control end (8) and a power module; The probe end comprises a clamp (2) and a paddle (3), wherein the paddle (3) is mounted on the clamp (2), and the paddle (3) is used to slide across the surface of the rock joint to generate a sound signal; The audio collection end is close to the probe end and includes a recording component (4), a data storage center (5), and a USB interface (6). The recording component (4) is used to collect and transmit the sound signal when the pick (3) passes through the surface of the rock joint. The recording component (4) stores the sound signal in the data storage center (5) and exports it through the USB interface (6). The data processing terminal (7) is located in the measuring device body (1), and is connected to the data storage center (5) and the control terminal (8) respectively, and is used to process the sound signal of the data storage center (5) and display the result on the control terminal (8); A control end (8), the control end (8) is connected to the power module, the data storage center (5) and the data processing end (7) respectively, and is used to control the data processing end (7) to perform data processing, read the remaining power in the power module, name the files in the data storage center (5), and display the relevant results on the display screen of the control end (8). The control end (8) is used to display the measured joint roughness coefficient and power, and perform file naming and export operations; The power module comprises a power management chip, a battery (9), and a charging interface (10); the charging interface (10) is connected to the battery (9) for charging; the power management chip is connected to the battery (9) and a control terminal (8) for reading the state of the battery (9) and transmitting the information to the control terminal (8); It is characterized in that The measuring method comprises the following steps: Step 1: Press and hold the recording button and use the pick (3) to slide across the surface of the joint to be measured to record. After recording is completed, release the recording button to complete the recording; Step 2: Name the recording file on the display screen of the control end (8), and control the data processing end to process the selected recording file; Step 3: The data processing end (7) first uses an audio feature extraction algorithm to resample and extract features from the recording file. The extracted features include high-order Mel-frequency cepstrum coefficients, short-time Fourier transform, and chrominance frequency sound features, energy root mean square sound features, spectrum centroid sound features, bandwidth sound features, spectrum attenuation sound features, and zero-crossing rate sound features. Step 4: After extracting the features, use the trained two-layer two-dimensional convolutional neural network and fully connected neural network algorithms to analyze and identify the features, and output their corresponding joint roughness coefficients and probabilities; Step 5: Finally, the data processing end (7) uses a correction algorithm to correct the result based on probability and transmits the corrected result to the control end; the correction algorithm judges the probability of the prediction results of the above three neural networks, and if the maximum prediction probability is greater than 0.5, the prediction result corresponding to the maximum probability is output; Otherwise, take the prediction results corresponding to the highest two probabilities and output the average value; Step 6: Read the calculation result on the display screen of the control terminal (8).

Images