A method for determining rock compressive strength based on deep learning rebound and audio

Abstract

The present application relates to a kind of rock compressive strength determination method based on deep learning rebound and audio, comprising the following steps: S1, rock rebound, rebound instrument audio recording and rock uniaxial compressive strength test result are collected by large sample test;S2, the rock rebound data collected, rebound audio set of percussion are processed;S3, based on the rock rebound value collected, the characteristic value of rebound audio spectrum of percussion and the data of corresponding rock uniaxial compressive strength, neural network calculation model is established;S4, by collecting the characteristic value of rock rebound value, rebound audio spectrum of percussion, based on the neural network calculation model established, the rock rebound value obtained by testing, the characteristic value of rebound audio spectrum of percussion is substituted into prediction model, and the rock uniaxial compressive strength of rock is calculated.The present application carries out deep learning training to obtain information and summarize rule to establish rock strength prediction model, and is suitable for various types of rock.

Description

TECHNICAL FIELD

[0001] The present application relates to the fields of geotechnical engineering, rock acoustics and artificial intelligence, and in particular to a method for determining rock compressive strength based on deep learning rebound and audio. BACKGROUND

[0002] Rock compressive strength is an important indicator for analyzing the mechanical properties of rock (body) and is also a basic basis for geotechnical engineering design and construction scheme development. Currently, the most commonly used method for testing rock compressive strength is rock uniaxial compressive strength test. However, this method has high requirements for sample selection, specimen size, surface smoothness and loading speed, and the detection process will cause damage to the specimen, making it impossible to perform secondary reinspection, which is a destructive test. In addition, this test requires a large number of drilling and coring, soft rock is difficult to form, and the press machine equipment is bulky, which cannot be used for on-site testing, and the sample needs to be transported from the field to the laboratory for sample preparation and testing, resulting in poor economic efficiency and timeliness. In contrast, the rebound hammer, with its small size, light weight and convenient operation, has been widely used in the fields of railway, highway, water conservancy and hydropower engineering, etc. for preliminary survey of rock mass strength.

[0003] Although the rebound method has obvious advantages, its accuracy in predicting rock compressive strength still has great limitations. First, due to the diversity of rock types, the differences in weathering degree, structure and other factors of different rocks are significant, and the operation methods of different testers are not uniform, resulting in large dispersion of rebound values, and there is currently no uniform standard. Second, although the correlation between rebound value and uniaxial compressive strength is good, most existing studies are based on linear equation fitting, which is only applicable to specific types of rock and cannot cover all types of rock. Finally, a single rebound value is easily affected by factors such as rock surface fissures, roughness and water content, and the prediction accuracy is limited. Studies have shown that if combined with multi-parameter prediction methods such as sound waves and audio frequency spectrum, the prediction accuracy of rock strength will be significantly improved.

[0004] In the prior art, patent application CN201811088260.3 discloses a method for measuring rock surface strength based on deep learning of acoustic spectrograms. This method obtains an audio spectrogram by striking the rock surface with a geological hammer and trains a neural network model using rebound hammer data to determine the rock surface strength. However, this method can only predict the surface strength of the rock and cannot assess the overall rock strength. Furthermore, due to the uncertainty of the striking force, angle, and position of the geological hammer, the accuracy of the audio spectrogram features is difficult to guarantee. In addition, patent application CN202110726582.1 discloses a method for measuring rebound values ​​using an L-shaped Schmidt hammer and predicting the uniaxial compressive strength of rock using a random forest model. However, this method relies only on a single rebound value parameter, resulting in low fitting correlation and insufficient prediction accuracy. It is clear that existing technologies for measuring rock compressive strength still suffer from low accuracy, limited applicability, and insufficient model reliability. There is an urgent need for a method for measuring rock compressive strength that combines multi-parameter analysis and possesses high prediction accuracy to meet the practical needs of engineering applications. Summary of the Invention

[0005] To address the problems of low accuracy, limited applicability, and insufficient model reliability in existing rock compressive strength determination methods, this invention provides a method for determining rock compressive strength based on deep learning rebound and audio. By combining the multi-parameter features of rock rebound value and impact audio, and using a neural network model for deep learning training, the method accurately predicts the uniaxial compressive strength of rocks. This overcomes the prediction errors caused by single parameters or unstable factors in existing technologies, and is applicable to non-destructive testing of various types of rocks, significantly improving the accuracy and applicability of the test.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A method for determining the compressive strength of rock based on deep learning rebound and audio, comprising the following steps:

[0008] S1, the results of rock rebound, rebound hammer impact audio recording and rock uniaxial compressive strength test were collected through large sample test;

[0009] S2, performs data processing on the collected rock rebound data and rebound impact audio set;

[0010] S3. Based on the collected rock rebound values, the characteristic values ​​of the rebound impact audio spectrum, and the corresponding data on the uniaxial compressive strength of the rock, a neural network calculation model is established.

[0011] S4. By collecting the rock rebound value and the characteristic value of the rebound impact audio spectrum, and based on the established neural network calculation model, the rock rebound value and the characteristic value of the rebound impact audio spectrum obtained from the test are substituted into the prediction model to calculate the uniaxial compressive strength of the rock.

[0012] Preferably, step S1 includes:

[0013] S11. When selecting the experimental target, three representative types of typical rocks were chosen, including igneous rocks, sedimentary rocks and metamorphic rocks. Rock samples were mainly core samples, with a total of 600 sets of core samples. Of these, 400 sets were used to train the neural network calculation model and 200 sets were used to verify the accuracy of the network calculation model.

[0014] S12, number the core samples, grind the upper and lower end faces of the core cylindrical specimen, divide the bottom surface into three concentric circles: the center, the inner circle, and the circumference of the axis. Arrange 4 measuring points in each concentric circle, and arrange a total of 12 measuring points for each sample for testing.

[0015] S13. Select a rock core sample in its natural state and fix it in a simple way to ensure that it will not move during the test. Perform a rebound test at the test point. The rebound hammer test angle is perpendicular to the rock surface.

[0016] S14, while striking, use a recorder with a bit rate greater than 1500kbps to record the striking audio file and store the audio file;

[0017] S15, uniaxial compressive strength test was performed on the core sample.

[0018] Preferably, in step S13, different types of rocks are selected using appropriate rebound hammers. For rebound values ​​below 40, an L-type rebound hammer with an impact energy of 0.735 Nm is used, and for rebound values ​​above 40, an N-type rebound hammer with an impact capacity of 2.207 Nm is used. The test results are recorded.

[0019] Preferably, in step S15, the rock uniaxial compressive strength testing machine is a hydraulic testing machine with a load of 2000kN and a minimum resolution of 0.01kN.

[0020] Preferably, step S2 includes:

[0021] S21. The rock rebound data is analyzed. Two critical values ​​are removed from the 12 rebound readings in the three layers of the survey area. The arithmetic mean of the remaining 10 rebound readings is the rebound value of the survey area, forming a rock rebound value dataset.

[0022] S22 performs digital processing on the rebound audio.

[0023] Preferably, step S22 includes:

[0024] S221, Use short-time Fourier transform to generate the spectrogram of the recorded rebound impact audio file;

[0025] S222, Analyze the acoustic spectrum obtained in step S221. In order to better reflect the decay rate of the acoustic signal, the time length is intercepted starting from the time point of the impact. 0 to 0.5s is defined as the large amplitude segment of the acoustic signal, and 0 to 0.4s is defined as the decay segment of the acoustic signal.

[0026] S223, calculate the integral area of ​​the segments intercepted at both ends in step S222. The area of ​​the large amplitude segment curve and the horizontal axis is a1, and the area of ​​the attenuation segment curve and the horizontal axis is a2. The ratio of a1 to a2 is defined as the amplitude attenuation coefficient i.

[0027] S224, the frequency of the acoustic spectrum obtained in step S221 is mainly concentrated in the range of 0 to 20000 Hz, with 0 to 5000 Hz as the low frequency region and 5000 to 20000 Hz as the high frequency region.

[0028] S225, perform integral area calculation on the segments intercepted at both ends in step S224, and use the integral method to find the area between the two curves and the horizontal axis. The area of ​​the low frequency region is b1, the area of ​​the high frequency region is b2, and the ratio of b1 to b2 is defined as the high-low frequency ratio j.

[0029] S226, record the amplitude attenuation coefficient i in step S223 and the high-low frequency ratio j in step S225. At this point, all the acoustic spectrum features of a rebound impact audio have been extracted and digitized to form a rebound audio feature value dataset. The audio dataset corresponds one-to-one with the rock rebound value and rock compressive strength value.

[0030] Preferably, the neural network computation model in step S3 includes an input layer, a hidden layer, and an output layer, each layer having several nodes. The connection state between nodes in different layers is represented by weights. Step S3 includes:

[0031] S31 accepts data input. The number of nodes in the input layer is 3, corresponding to the rock rebound value, the characteristic value of the rebound impact audio spectrum, and the uniaxial compressive strength of the rock.

[0032] S32, process the input data;

[0033] S33, output the calculation results.

[0034] Preferably, the number of nodes in the hidden layer is set to 3 by default. Each node contains a perceptron, which includes input, weights, biases, activation functions, and outputs. During forward propagation, the input data is processed by the activation function after being calculated by the perceptron nodes. During backward propagation, the output results are compared with the expected results, and the weights of each node on the network are continuously adjusted through multiple iterations to optimize the model performance.

[0035] Preferably, the rock sample selection in step S11 also includes boulders and natural rock outcrops. A relatively flat and fresh outcrop is selected as the test area. The test area is at least 5 cm away from the edge of the rock mass. 16 test points are taken in one test area, and the distance between each test point is greater than 2 cm.

[0036] Preferably, in step S21, for outcrops of boulders or natural rocks, the three maximum values ​​and three minimum values ​​are removed from the 16 rebound values ​​in the test area, and the arithmetic mean of the remaining 10 rebound readings is the rebound value of the test area.

[0037] Compared with the prior art, the advantages of the present invention are as follows:

[0038] (1) Traditional uniaxial compressive strength testing of rocks requires core drilling and sample preparation, which is complex and costly. Furthermore, samples must be transported long distances from the field to the laboratory for preparation and testing, severely impacting economic efficiency and timeliness. This invention utilizes a deep learning-based method for determining rock compressive strength based on springback and audio signals. This eliminates the need for complete sample preparation, allowing for rapid and convenient testing directly at the field geological survey site. This avoids destructive testing of specimens, significantly improves testing efficiency, and substantially reduces testing costs.

[0039] (2) When using the rebound method alone to estimate the compressive strength of rock, the test error may be large due to the differences or defects inside the object, which limits the prediction accuracy. This invention integrates two parameters, the rebound value and the acoustic spectrum characteristics of the impact audio, and rapidly determines the compressive strength of rock through multi-parameter collaborative analysis, which significantly improves the accuracy and reliability of the prediction.

[0040] (3) In the existing technology, linear equations cannot be applied to the strength prediction of all types of rocks. Different functional relationships need to be set separately for different rock types, and the differences in rock structure and properties need to be considered. This invention inputs rock rebound value, audio spectrum feature value and compressive strength test data into a BP neural network, trains the model through deep learning, and automatically summarizes the rules of rock strength prediction. It can adapt to the measurement needs of various types of rocks, and is especially suitable for actual field geological exploration work scenarios, with stronger applicability and versatility.

[0041] (4) This invention adopts a non-destructive testing method, which will not damage the rock sample and solves the problem that traditional testing methods cannot be retested. At the same time, it combines rebound value and audio characteristic value for multi-scenario measurement, without relying on laboratory conditions, which greatly expands the application scenarios of rock compressive strength measurement and provides a more efficient and accurate basis for engineering design and construction. Attached Figure Description

[0042] Figure 1 This is a flowchart of the rock compressive strength determination process according to the present invention;

[0043] Figure 2 This is a schematic diagram of the rebound audio feature value extraction of the present invention;

[0044] Figure 3 This is a flowchart of the BP neural network training process of the present invention;

[0045] Figure 4 This is a diagram of the neural network model of the present invention;

[0046] Figure 5 This is a waveform diagram of the rebound audio signal of the present invention;

[0047] Figure 6 This is a waveform diagram after the amplitude positiveization processing of the present invention;

[0048] Figure 7 This is the spectrum diagram of the present invention;

[0049] Figure 8 This is a comparison chart of the original results and predicted results of the neural network model. Detailed Implementation

[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention belong to the present invention.

[0051] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0052] like Figure 1As shown, this embodiment discloses a method for determining the compressive strength of rock based on deep learning rebound and audio, characterized by including the following steps:

[0053] S1, the results of rock rebound, rebound hammer impact audio recording and rock uniaxial compressive strength test were collected through large sample test;

[0054] S2, performs data processing on the collected rock rebound data and rebound impact audio set;

[0055] S3. Based on the collected rock rebound values, the characteristic values ​​of the rebound impact audio spectrum, and the corresponding data on the uniaxial compressive strength of the rock, a neural network calculation model is established.

[0056] S4. By collecting the rock rebound value and the characteristic value of the rebound impact audio spectrum, and based on the established neural network calculation model, the rock rebound value and the characteristic value of the rebound impact audio spectrum obtained from the test are substituted into the prediction model to calculate the uniaxial compressive strength of the rock.

[0057] Specifically, step S1 includes:

[0058] S11, when selecting the experimental target, three representative rock types were chosen: igneous rocks, sedimentary rocks, and metamorphic rocks. Rock samples were primarily core samples. Representative igneous rocks included granite, diorite, diabase, andesite, and basalt. Representative sedimentary rocks included sandstone, mudstone, shale, limestone, dolomite, and tuff. Representative metamorphic rocks included slate, schist, and gneiss. A total of 600 core samples were collected, with 400 samples used to train the neural network computational model and 200 samples used to verify the accuracy of the model. It should be noted that the total number of core samples is not fixed; approximately 600 samples is sufficient. The total number of samples can be adjusted appropriately according to actual needs. In some embodiments, boulders and natural rock outcrops can also be selected. A relatively flat and fresh outcrop surface is selected as the test area. If the surface is covered with moss or sediment, it needs to be polished with sandpaper or an angle grinder. Avoid cracks during testing. The test area should be at least 5 cm away from the edge of the rock mass. 16 test points are taken in one test area, and the distance between each test point is greater than 2 cm.

[0059] S12, number the core samples, grind the top and bottom ends of the cylindrical core specimens, divide the bottom surface into three concentric circles: the center, the inner circle, and the circumference of the axis. Arrange four measuring points in each concentric circle, and arrange a total of 12 measuring points for each sample for testing.

[0060] S13. Select a natural rock core sample and fix it in a simple manner to ensure it does not move during testing. Perform a rebound test at the test point, with the rebound hammer angle perpendicular to the rock surface. Different types of rocks require different rebound hammers. For soft rocks such as mudstone, sandstone, mudstone, and weathered altered rocks, use an L-type rebound hammer with an impact energy of 0.735 Nm for rebound values ​​below 40. For hard rocks such as granite, andesite, limestone, basalt, and shale, use an N-type rebound hammer with an impact capacity of 2.207 Nm for rebound values ​​above 40. Record the test results.

[0061] S14. While the ball is being struck, record an audio file of the impact using a recorder with a bit rate greater than 1500kbps and store the audio file. Once the rebound value is relatively stable during the rebound test, record an audio file of the ball while it is being struck, keeping the recording environment as quiet as possible and avoiding noise.

[0062] S15. Uniaxial compressive strength test is performed on the rock core sample. Specifically, the test procedure can be found in Section 2.7 of the "Standard for Test Methods of Engineering Rock Mass" GB / T50266-2013. A hydraulic testing machine with a load capacity of 2000kN and a minimum resolution of 0.01kN is selected for the uniaxial compressive strength testing of rock.

[0063] Specifically, step S2 includes:

[0064] S21. Analyze the rock rebound data. For core samples, remove two critical values ​​from the 12 rebound readings of the three concentric layers in the test area, and take the arithmetic mean of the remaining 10 rebound readings as the rebound value of the test area, forming a rock rebound value dataset. For boulders or natural rock outcrops, remove three maximum and three minimum values ​​from the 16 rebound values ​​in the test area, and take the arithmetic mean of the remaining 10 rebound readings as the rebound value of the test area. Statistically analyze the rebound values ​​to form a rock rebound value dataset.

[0065] S22 performs digital processing on the rebound audio.

[0066] like Figure 2 As shown, step S22 specifically includes:

[0067] S221, Use short-time Fourier transform to generate the spectrogram of the recorded rebound impact audio file;

[0068] S222, Analyze the acoustic spectrum obtained in step S221. In order to better reflect the decay rate of the acoustic signal, the time length is intercepted starting from the time point of the impact. 0 to 0.5s is defined as the large amplitude segment of the acoustic signal, and 0 to 0.4s is defined as the decay segment of the acoustic signal.

[0069] S223, calculate the integral area of ​​the segments intercepted at both ends in step S222. The area of ​​the large amplitude segment curve and the horizontal axis is a1, and the area of ​​the attenuation segment curve and the horizontal axis is a2. The ratio of a1 to a2 is defined as the amplitude attenuation coefficient i.

[0070] S224, the frequency of the acoustic spectrum obtained in step S221 is mainly concentrated in the range of 0 to 20000 Hz, with 0 to 5000 Hz as the low frequency region and 5000 to 20000 Hz as the high frequency region.

[0071] S225, perform integral area calculation on the segments intercepted at both ends in step S224, and use the integral method to find the area between the two curves and the horizontal axis. The area of ​​the low frequency region is b1, the area of ​​the high frequency region is b2, and the ratio of b1 to b2 is defined as the high-low frequency ratio j.

[0072] S226, record the amplitude attenuation coefficient i in step S223 and the high-low frequency ratio j in step S225. At this point, all the acoustic spectrum features of a rebound impact audio have been extracted and digitized to form a rebound audio feature value dataset. The audio dataset corresponds one-to-one with the rock rebound value and rock compressive strength value.

[0073] The neural network computation model in step S3 includes an input layer, a hidden layer, and an output layer. Each layer has several nodes, and the connection state between nodes in different layers is represented by weights. Step S3 specifically includes:

[0074] S31 accepts data input. The number of nodes in the input layer is 3, corresponding to the rock rebound value, the characteristic value of the rebound impact audio spectrum, and the uniaxial compressive strength of the rock.

[0075] S32 processes the input data.

[0076] like Figures 3-4 As shown, the number of nodes in the input layer equals the dimension of the input, the number of nodes in the output layer equals the dimension of the output, and the number of nodes in the hidden layer can be set according to the actual situation, with a default setting of 3. Each node contains a perceptron (i.e., a single neuron), which includes input terms, weights, biases, activation functions, and outputs. During forward propagation, the input data is processed by the activation function after being calculated by the perceptron nodes to obtain the output result. During backward propagation, the output result is compared with the expected result, and the weights of each node on the network are continuously adjusted through multiple iterations to optimize the model performance.

[0077] S33, output the calculation results.

[0078] In this embodiment, the learning of the neural network mainly involves solving a set of W and b to minimize the network's error function. The computation process of the neural network includes two stages: forward propagation and backward propagation. In the forward propagation stage, the input data enters the network through the input layer, and after computation by the hidden layers and output layers, the output result is obtained. There are connections between neurons in each layer, and the weights and biases of these connections can be learned and adjusted through training. During forward propagation, the input data is processed by the perceptron nodes and then processed by the activation function to obtain the output result. In the backward propagation stage, the output result is compared with the expected result, the error is calculated, and then the error is propagated backward layer by layer from the output layer to the hidden layers, updating the weights and biases of each neuron in the network to reduce the error. This process iterates continuously until the stopping criterion is met. During backward propagation, the weights and biases are updated using gradient descent and gradient search techniques.

[0079] The measurement method of this embodiment will be verified through an example below:

[0080] First, experimental data were obtained through testing: rebound tests were conducted on the selected rock types to obtain rebound values ​​and corresponding rebound impact audio spectra; uniaxial compression tests were also conducted to obtain the uniaxial compressive strength of the rocks. For soft rocks with rebound values ​​below 40 (such as mudstone, sandstone, mudstone, weathered and altered rocks), an L-type rebound hammer with an impact energy of 0.735 Nm was used; for hard rocks with rebound values ​​above 40 (such as granite, andesite, limestone, etc.), an N-type rebound hammer with an impact capacity of 2.207 Nm was used. A recorder with a bit rate greater than 1500 kbps was selected to record and sample the audio when each group of rocks was impacted by the rebound hammer. A hydraulic testing machine with a load capacity of 2000 kN and a minimum resolution of 0.01 kN was selected for the uniaxial compressive strength testing. The final valid test data are shown in Table 1.

[0081] Table 1. Test data of rock rebound value, rebound audio, and uniaxial compressive strength.

[0082]

[0083] Secondly, audio data is processed to obtain audio feature values: the rebound audio waveform obtained from the experiment is positiveized to fully present the change process of sound wave amplitude over time, and acoustic feature values—amplitude attenuation coefficient i and high-low frequency ratio j—are obtained from the audio time-domain signal. Taking the 7-64.wav audio as an example, the rebound audio waveform, the waveform after amplitude positiveization processing, and the spectrum are obtained through analysis as follows: Figures 5-7 As shown.

[0084] Next, all audio files are processed to obtain their amplitude attenuation coefficients and high and low frequencies, as shown in Table 2 below.

[0085] Table 2 shows the amplitude attenuation coefficient and high / low frequency ratio for each audio audio.

[0086]

[0087] Finally, using the rebound value, amplitude attenuation coefficient, high-low frequency ratio, and compressive strength values ​​obtained from the above experiments (Tables 1 and 2) as training samples, the neural network was trained to obtain a predictive model for rock compressive strength. The original model results were compared with the predicted results as follows: Figure 8 As shown.

[0088] The training results show that most of the prediction results in the sample are close to the original experimental results, and the prediction effect is good. Some prediction results differ from the original results. This difference can be further reduced by increasing the number of experimental samples and adjusting the model training parameters to meet the accuracy required for engineering.

[0089] The accuracy of the model was verified using the test results of rock samples numbered 10-75 as an example. The test results showed that the rebound value of the rock was 75, the amplitude attenuation coefficient was 0.80, the high-low frequency ratio was 0.78, and the compressive strength was 90.90 MPa. By inputting the rebound value, amplitude attenuation coefficient, and high-low frequency ratio into the model, the predicted strength value was 86.22 MPa, which differed from the test result of 90.90 MPa by 4.68 MPa, with an error of 5.1% and an accuracy rate of 94.9%. The model prediction results were relatively accurate.

[0090] In summary, this invention discloses a method for determining the compressive strength of rock based on deep learning rebound and audio. By collecting rock rebound values, impact audio, and uniaxial compressive strength test data, it extracts characteristic values ​​such as the amplitude attenuation coefficient and high / low frequency ratio of the audio. Then, using a neural network model for deep learning training, it constructs a compressive strength prediction model applicable to various rock types. This method introduces multi-parameter collaborative analysis and intelligent prediction technology into rock mechanical property analysis, enabling rapid and accurate prediction of the uniaxial compressive strength of rocks. Implementing the technical solution of this invention can significantly improve the prediction accuracy and reliability of rock compressive strength. Compared with traditional rock uniaxial compressive strength tests, this method does not require damaging the rock sample, eliminating cumbersome core sampling and sample preparation operations. It can directly and rapidly detect the compressive strength of rocks in the field, significantly improving detection efficiency and reducing detection costs. Furthermore, this invention combines rebound values ​​and audio spectral characteristics, overcoming the limitations of single-parameter prediction methods, resulting in a wider range of applications, significantly reduced prediction errors, and suitability for various rock types and practical engineering needs.

[0091] This invention is of great significance in the fields of geotechnical engineering, rock acoustics, and artificial intelligence. On the one hand, it overcomes the limitations of traditional methods in terms of economy, timeliness, and applicability, providing an efficient and reliable technical means for geotechnical engineering design, construction scheme formulation, and rapid on-site evaluation. On the other hand, the multi-parameter collaborative analysis method based on deep learning provides a new perspective and technical route for the study of rock mechanical properties, promoting the intelligent and modern development of rock mechanical testing technology, and has broad industry application prospects and promotional value.

[0092] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Under the concept of the present invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the present invention as described above. For the sake of brevity, they are not provided in detail. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for determining the compressive strength of rock based on deep learning rebound and audio, characterized in that, Includes the following steps: S1, the results of rock rebound, rebound hammer impact audio recording and rock uniaxial compressive strength test were collected through large sample test; S2, performs data processing on the collected rock rebound data and rebound impact audio set; S3. Based on the collected rock rebound values, the characteristic values ​​of the rebound impact audio spectrum, and the corresponding data on the uniaxial compressive strength of the rock, a neural network calculation model is established. S4. By collecting the rock rebound value and the characteristic value of the rebound impact audio spectrum, and based on the established neural network calculation model, the rock rebound value and the characteristic value of the rebound impact audio spectrum obtained from the test are substituted into the prediction model to calculate the uniaxial compressive strength of the rock. Step S2 includes: S21. Analyze the rock rebound data. For core samples, remove two critical values ​​from the 12 rebound readings across three layers in the test area, and take the arithmetic mean of the remaining 10 rebound readings as the rebound value for that test area, thus forming a rock rebound value dataset. For boulders or natural rock outcrops, remove three maximum and three minimum values ​​from the 16 rebound values ​​in the test area, and take the arithmetic mean of the remaining 10 rebound readings as the rebound value for that test area. Statistically analyze the rebound values ​​to form a rock rebound value dataset. S22, digitally processes the rebound audio; Step S22 includes: S221, Use short-time Fourier transform to generate the spectrogram of the recorded rebound impact audio file; S222, Analyze the acoustic spectrum obtained in step S221. In order to better reflect the decay rate of the acoustic signal, the time length is intercepted starting from the time point of the impact. 0~0.5s is defined as the large amplitude segment of the acoustic signal, and 0~0.4s is defined as the decay segment of the acoustic signal. S223, calculate the integral area of ​​the segments intercepted at both ends in step S222. The area of ​​the large amplitude segment curve and the horizontal axis is a1, and the area of ​​the attenuation segment curve and the horizontal axis is a2. The ratio of a1 to a2 is defined as the amplitude attenuation coefficient i. S224, the frequency of the acoustic spectrum obtained in step S221 is mainly concentrated in the range of 0~20000Hz, with 0~5000Hz as the low frequency region and 5000~20000Hz as the high frequency region. S225, perform integral area calculation on the segments intercepted at both ends in step S224, and use the integral method to find the area between the two curves and the horizontal axis. The area of ​​the low frequency region is b1, the area of ​​the high frequency region is b2, and the ratio of b1 to b2 is defined as the high-low frequency ratio j. S226, record the amplitude attenuation coefficient i in step S223 and the high-low frequency ratio j in step S225. At this point, all the acoustic spectrum features of a rebound impact audio have been extracted and digitized to form a rebound audio feature value dataset. The audio dataset corresponds one-to-one with the rock rebound value and rock compressive strength value.

2. The determination method according to claim 1, characterized in that, Step S1 includes: S11. When selecting the experimental target, three representative types of typical rocks were chosen, including igneous rocks, sedimentary rocks and metamorphic rocks. Rock samples were mainly core samples, with a total of 600 sets of core samples. Of these, 400 sets were used to train the neural network calculation model and 200 sets were used to verify the accuracy of the network calculation model. S12, number the core samples, grind the upper and lower end faces of the core cylindrical specimen, divide the bottom surface into three concentric circles: the center, the inner circle, and the circumference of the axis. Arrange 4 measuring points in each concentric circle, and arrange a total of 12 measuring points for each sample for testing. S13. Select a rock core sample in its natural state and fix it in a simple way to ensure that it will not move during the test. Perform a rebound test at the test point. The rebound hammer test angle is perpendicular to the rock surface. S14, while striking, use a recorder with a bit rate greater than 1500 kbps to record the striking audio file and store the audio file; S15, uniaxial compressive strength test was performed on the core sample.

3. The determination method according to claim 2, characterized in that, In step S13, different rebound hammers are selected for different rocks. For rebound values ​​below 40, an L-type rebound hammer with an impact energy of 0.735 Nm is used, and for rebound values ​​above 40, an N-type rebound hammer with an impact capacity of 2.207 Nm is used. The test results are recorded.

4. The determination method according to claim 3, characterized in that, In step S15, the rock uniaxial compressive strength testing machine selected is a hydraulic testing machine with a load of 2000kN and a minimum resolution of 0.01kN.

5. The determination method according to claim 4, characterized in that, The neural network computation model in step S3 includes an input layer, hidden layers, and an output layer. Each layer has several nodes, and the connection state between nodes in different layers is represented by weights. Step S3 includes: S31 accepts data input. The number of nodes in the input layer is 3, corresponding to the rock rebound value, the characteristic value of the rebound impact audio spectrum, and the uniaxial compressive strength of the rock. S32, process the input data; S33, output the calculation results.

6. The determination method according to claim 5, characterized in that, The number of nodes in the hidden layer is set to 3 by default. Each node contains a perceptron, which includes input, weights, biases, activation functions, and outputs. During forward propagation, the input data is processed by the activation function after being calculated by the perceptron nodes. During backward propagation, the output results are compared with the expected results, and the weights of each node on the network are continuously adjusted through multiple iterations to optimize the model performance.

7. The determination method according to claim 6, characterized in that, Step S11 also includes selecting rock samples from boulders and natural rock outcrops. A relatively flat and fresh outcrop is selected as the test area. The test area is at least 5 cm away from the edge of the rock mass. 16 test points are taken in one test area, and the distance between each test point is greater than 2 cm.

Images