Multi-classification method, device, terminal and storage medium for underground acoustic emission sources

Abstract

The present invention discloses a multi-classification method, device, terminal and storage medium for downhole acoustic emission sources, wherein the method comprises: obtaining a time series of downhole acoustic emission waveforms to be detected; performing Markov transfer field encoding on the time series of downhole acoustic emission waveforms to be detected to obtain a two-dimensional image Markov transfer field; inputting the two-dimensional image Markov transfer field into a pre-trained downhole acoustic emission source identification and classification model, and outputting the classification result of the time series of downhole acoustic emission waveforms to be detected. The one-dimensional time series of downhole acoustic emission is converted into a two-dimensional image with high recognition by using Markov transfer field encoding. This two-dimensional image not only retains the trend of change on the time scale, but also reflects the transfer probability of different time spans. Combined with the advantages of convolutional neural network, a representative algorithm in deep learning, in the recognition and classification of two-dimensional images, it is significantly superior to traditional waveform recognition and classification methods in terms of recognition efficiency, stability and accuracy.

Description

Technical Field

[0001] The invention relates to the field of acoustic emission ground pressure monitoring in mining engineering, and in particular to a multi-classification method, device, terminal and storage medium for underground acoustic emission sources. Background Art

[0002] As one of the effective means of ground pressure monitoring, acoustic emission monitoring has been widely used in mines at home and abroad. Correctly and timely identifying AE events in surrounding rock mass is an important part of the prediction and early warning of acoustic emission monitoring system.

[0003] The monitoring environment in underground mines is relatively complex, generally involving rock drilling operations, loading and unloading machinery operations such as shovel loaders, blasting operations, other manual operations, and random interference signals. In order to eliminate the interference of these non-surrounding rock AE events, frequency domain filtering is often used to filter out some non-surrounding rock AE signals. However, some noise source events have cross-frequency bands with surrounding rock AE events, so frequency domain filtering cannot effectively eliminate underground noise signals. Excessive noise events will cause false alarms in the AE early warning system, which will have a significant adverse impact on mine production safety and the safety of operators. At present, the more accurate method for identifying surrounding rock AE events and underground noise events is the manual identification method. However, the efficiency of manual identification is too low, and the identification results rely on the experience of technicians and have poor stability.

[0004] The generalization ability of the traditional waveform recognition and classification model to extract and identify sensitive features is weak, and it is difficult to adapt to the recognition and classification of acoustic emission signals in complex underground environments. In addition, there are subjective differences in the feature extraction process, and too many or too few feature parameters affect the accuracy and computational efficiency of classification. In recent years, the rapid development of intelligent algorithms such as deep learning has provided a strong guarantee and foundation for the realization of intelligence in other disciplines and fields. As a representative deep learning algorithm, convolutional neural network has great advantages and wide applications in image classification. In the document "Automatic Classification and Recognition of Seismic Waveforms Based on Deep Learning Convolutional Neural Network", a one-dimensional waveform time series is used as 1D-CNN input, and 1D-CNN is used to identify seismic waveforms and noise waveforms. Due to the complexity of seismic waveforms, and their waveforms often present multiple forms, the difficulty of 1D-CNN recognition on the time scale is increased, so the recognition result accuracy is low. For most CNN structures, the input of two-dimensional images can better exert their performance. The above methods all have defects in the recognition of underground acoustic emission events. The present invention overcomes the above shortcomings and provides a new method for the recognition and classification of underground acoustic emission events. Summary of the invention

[0005] The present invention provides a method, device, terminal and storage medium for multi-classification of underground acoustic emission sources, so as to solve the problems of low accuracy and poor stability of traditional acoustic emission source classification methods based on acoustic emission waveform images and acoustic emission waveform time series.

[0006] In a first aspect, a multi-classification method for downhole acoustic emission sources is provided, comprising:

[0007] Obtaining the time series of downhole acoustic emission waveforms to be detected;

[0008] Markov transfer field encoding is performed on the time series of acoustic emission waveforms of the downhole to be detected to obtain a two-dimensional image Markov transfer field;

[0009] The two-dimensional image Markov transfer field is input into a pre-trained downhole acoustic emission source recognition and classification model, and the classification result of the downhole acoustic emission waveform time series to be detected is output; wherein the downhole acoustic emission source recognition and classification model is obtained by training a convolutional neural network based on the historical two-dimensional image Markov transfer field corresponding to the historical downhole acoustic emission waveform time series.

[0010] In order to give full play to the advantages of convolutional neural networks in image recognition and classification, optimize the data input format of convolutional neural networks, and overcome the difficulty of traditional time-frequency analysis methods in comprehensively extracting waveform features, the technical solution of the present invention proposes a Markov transition field encoding based on the Markov transfer matrix related theory in probability theory and mathematical statistics, and encodes the one-dimensional time series of acoustic emission signals into a two-dimensional image. This two-dimensional image not only retains the trend of change on the time scale, but also reflects the transfer probability of different time spans. Then, the convolutional neural network is used to automatically extract features from the two-dimensional image and identify and classify it. It is significantly superior to traditional waveform recognition and classification methods in terms of recognition efficiency, stability and accuracy, and can correctly and timely identify acoustic emission signals to provide reliable data support for mine ground pressure monitoring and timely warning.

[0011] Furthermore, the Markov transfer field encoding is performed on the time series of the downhole acoustic emission waveform to be detected to obtain the Markov transfer field of the two-dimensional image, which specifically includes:

[0012] For the downhole acoustic emission waveform time series X = [x 1 ,x 2 ,…,x n ], determine Q quantiles, which means that the downhole acoustic emission waveform to be detected is divided into Q intervals in the vertical direction, and each x i Assigned to the corresponding interval q j , j∈[1,Q];

[0013] The transfer between each quantile interval of the downhole acoustic emission waveform time series to be detected is calculated along the time axis in the manner of a first-order Markov chain to construct a weighted adjacency matrix Q×Q; the weighted Markov transfer matrix W is obtained after normalization, which is insensitive to the distribution and time scale dependence of the waveform time series X. Getting rid of time dependence will lead to excessive information loss in W;

[0014] By considering the time position, the weighted Markov transfer matrix W containing the transition probabilities on the amplitude axis is expanded into the Markov transfer field matrix, resulting in an n×n Markov transfer field M:

[0015]

[0016] Conditional probability P{X(a+h)=q j |X(a)=q i} means that the time series is in interval q at time a i Under the premise of j The conditional probability is w ij|xa∈qi,|xa+h∈qj .

[0017] Furthermore, after obtaining the n×n Markov transition field M, it also includes:

[0018] Grid the Markov transition field M;

[0019] The sub-image in each grid is replaced by its average value to generate a fuzzy Markov transition field.

[0020] By converting the n×n Markov transition field M into a fuzzy Markov transition field, the size is reduced, the computational cost is reduced, and the computational efficiency is improved.

[0021] Furthermore, the downhole acoustic emission source identification and classification model is obtained by training a convolutional neural network based on the historical two-dimensional image Markov transfer field corresponding to the historical downhole acoustic emission waveform time series, and specifically includes:

[0022] Obtain several historical downhole acoustic emission waveform time series and their corresponding acoustic emission source categories;

[0023] The Markov transfer field encoding is performed on several historical downhole acoustic emission waveform time series to obtain the corresponding historical two-dimensional image Markov transfer field;

[0024] Based on the obtained historical two-dimensional image Markov transfer field and its corresponding acoustic emission category label, a training sample set and a test sample set are constructed;

[0025] Taking the Markov transfer field of historical two-dimensional images as input and the acoustic emission category as output, the convolutional neural network is trained based on the training sample set and the test sample set to obtain the downhole acoustic emission source recognition and classification model.

[0026] Furthermore, the categories of acoustic emission sources include surrounding rock acoustic emission signals, blasting operation signals, scraper operation signals, rock drilling operation signals and other types of signals; the training sample set and the test sample set both contain samples of the five categories.

[0027] Furthermore, when training the downhole acoustic emission source identification and classification model, a stochastic gradient descent optimization algorithm is used to train the convolutional neural network.

[0028] In a second aspect, a multi-classification device for downhole acoustic emission sources is provided, comprising:

[0029] A data acquisition module, used to acquire the time series of the downhole acoustic emission waveform to be detected;

[0030] An encoding module is used to perform Markov transfer field encoding on the time series of the downhole acoustic emission waveform to be detected, so as to obtain a two-dimensional image Markov transfer field;

[0031] The classification module is used to input the two-dimensional image Markov transfer field into a pre-trained downhole acoustic emission source identification classification model, and output the classification result of the downhole acoustic emission waveform time series to be detected; wherein the downhole acoustic emission source identification classification model is obtained by training a convolutional neural network based on the historical two-dimensional image Markov transfer field corresponding to the historical downhole acoustic emission waveform time series.

[0032] In a third aspect, a downhole acoustic emission source multi-classification terminal is provided, comprising at least one memory and a processor;

[0033] The memory stores a computer program that can be run on the processor, and when the computer program is executed by the processor, the multi-classification method for downhole acoustic emission sources as described above is implemented.

[0034] In a fourth aspect, a computer-readable storage medium is provided, which stores a computer program, and when the computer program is executed by a processor, the multi-classification method for downhole acoustic emission sources as described above is implemented.

[0035] Beneficial Effects

[0036] The present invention proposes a multi-classification method, device, terminal, and storage medium for underground acoustic emission sources. The Markov transition field coding is used to convert the one-dimensional time series of underground acoustic emission into a highly recognizable two-dimensional image. This two-dimensional image not only retains the trend of change on the time scale, but also reflects the transition probability of different time spans. Combined with the advantages of the representative algorithm convolutional neural network in deep learning for two-dimensional image recognition and classification, it is significantly superior to the traditional waveform recognition and classification method in terms of recognition efficiency, stability, and accuracy. Compared with the recognition method using the waveform original image method, the method of the present invention continues to improve the accuracy and stability of recognition and classification. The ability to correctly and timely identify acoustic emission signals provides reliable data support for timely warning of mine ground pressure monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0038] Figure 1 It is a flow chart of a multi-classification method of downhole acoustic emission sources provided by an embodiment of the present invention;

[0039] Figure 2 is a schematic diagram of a Markov coding process provided by an embodiment of the present invention;

[0040] Figure 3 It is a flow chart of building a downhole acoustic emission source identification and classification model provided by an embodiment of the present invention;

[0041] Figure 4 It is a flow chart of downhole acoustic emission source identification and classification model training provided by an embodiment of the present invention;

[0042] Figure 5 These are four typical signal waveform diagrams provided by the embodiments of the present invention;

[0043] Figure 6 These are some other types of waveform diagrams provided by the embodiments of the present invention;

[0044] Figure 7 It is a fuzzy Markov transition field after encoding of a partial waveform time series provided by an embodiment of the present invention;

[0045] Figure 8 is a curve diagram of accuracy during training in an engineering example provided by an embodiment of the present invention;

[0046] Fig. 9 It is a diagram showing the change of the Loss function during training in the engineering example provided in the embodiment of the present invention;

[0047] Fig.10 is a schematic diagram of a waveform classification confusion matrix provided by an embodiment of the present invention;

[0048] Fig.11 It is the recognition accuracy of the test set under different signal-to-noise ratios in the engineering example provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0049] To make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be described in detail below. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other implementation methods obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention.

[0050] The purpose of the present invention is to encode the one-dimensional time series of underground acoustic emission into a two-dimensional image with obvious characteristics based on the Markov transfer field, and combine the advantages of convolutional neural network (CNN) in image classification to propose a new scheme for multi-classification of underground surrounding rock acoustic emission events and mining operation noise events, so as to solve the problems of low accuracy and poor stability of the traditional acoustic emission waveform recognition and classification methods based on acoustic emission waveform images and acoustic emission waveform time series, realize the recognition and classification of multiple acoustic emission sources in the mine, and provide reliable data support for the multi-parameter early warning of the mine ground pressure monitoring system. Based on this, the present invention provides the following embodiments.

[0051] Before performing multi-classification of downhole acoustic emission sources, it is necessary to pre-train the downhole acoustic emission source identification and classification model, such as Figure 3 As shown, the specific process includes:

[0052] Step 1: Acoustic emission data collection and sample set division

[0053] The historical downhole acoustic emission time series data is generally collected by the acoustic emission monitoring system and exported to Excel format, as shown in Table 1. The specific data length depends on the parameter setting of the monitoring system, and the data length in the same sample set should be equal. The sample set is divided into a test sample set and a training sample set. Labels are added to each type of downhole acoustic emission waveform collected in sequence, such as 0, 1, 2, 3, 4, ..., and each number represents a category. For example, in this embodiment, the categories include surrounding rock acoustic emission signals, blasting operation signals, scraper operation signals, rock drilling operation signals, and other types of signals.

[0054] Table 1 Waveform sample format

[0055] Sampling points (N) <![CDATA[Amplitude (m / s 2 )]]> 1 -1 2 -3 3 -8 4 -7 5 -14 ... ... 1023 -6 1024 3

[0056] Step 2: Markov transfer field coding of acoustic emission waveform time series

[0057] (1) For each collected historical acoustic emission waveform time series X = [x 1 ,x 2 ,…,x n ], n represents the length of the time series (total number of sampling points), and Q quantiles are determined. Q quantiles represent that the historical acoustic emission waveform is divided into Q intervals in the vertical direction. For details, see Figure 2As shown, each x i Assigned to the corresponding interval q j (j∈[1,Q]).

[0058] (2) A weighted adjacency matrix Q×Q is constructed by calculating the transfer between the quantile intervals of the historical acoustic emission waveform time series along the time axis in a first-order Markov chain manner; the weighted Markov transfer matrix W is obtained after normalization. It is insensitive to the distribution and time scale dependence of the waveform time series X. Getting rid of time dependence will lead to excessive information loss in W.

[0059] (3) Constructing n×n Markov transfer field. In order to overcome the disadvantage of weighted Markov transfer matrix W, the Markov transfer matrix is ​​expanded and the Markov transfer field matrix M is proposed. First, a QxQ Markov transfer matrix is ​​constructed. Time nodes i and j are located in the quantile intervals qi and qj respectively. That is to say, by considering the time position, the weighted Markov transfer matrix W containing the transfer probability on the amplitude axis is expanded into the Markov transfer field matrix, and the n×n Markov transfer field M is obtained, which is a two-dimensional image. The expression of Markov transfer field M is as follows:

[0060]

[0061] Conditional probability P{X(a+h)=q j |X(a)=q i} means that the time series is in interval q at time a i Under the premise of j The conditional probability is w ij|xa∈qi,|xa+h∈qj .

[0062] (4) By ij The probability of assigning the quantile interval from time step i to time step j is obtained. The Markov transition field M actually encodes the multi-span transition probability of the acoustic emission waveform time series. i,j||i-j|＝k M represents the transition probability between points with a time interval of k. i,j|i-j|＝1 It represents the transition probability with an interval of 1 along the time axis, and the main diagonal represents the transition probability from each quantile interval to itself (self-transition probability). For the downhole acoustic emission waveform time series data with a length of n, the converted Markov transfer field M is a matrix of size [n,n]. In order to improve the computational efficiency, the size of the Markov transfer field M is reduced. The Markov transfer field M is gridded, and then the sub-graph in each grid is replaced by its average value, that is, a fuzzy Markov transfer field is generated. Figure 2 In the process of Markov transition field encoding of acoustic emission waveform time series, the scale of gridding is determined according to the specific instance.

[0063] Step 3: Establish a classification model for underground acoustic emission source identification

[0064] Convolutional neural network is a type of feedforward neural network with convolution calculation and deep structure, and is one of the representative algorithms of deep learning. Its structure generally includes convolution layer, pooling layer, activation layer, and fully connected layer. CNN (convolutional neural network) has the characteristics of automatically extracting high-dimensional features of raw data and having translation invariance, avoiding the tedious feature extraction and data reconstruction process of traditional neural networks.

[0065] (1) Convolutional layer

[0066] The convolution layer extracts high-dimensional features of the blurred Markov transfer field of the input two-dimensional image. It not only extracts pixel features, but also generally contains multiple different convolution kernels. Each element of the convolution kernel corresponds to a weight coefficient and a bias.

[0067] (2) Activation function

[0068] The activation function is usually after the convolution operation, and the excitation function is mainly used to re-express complex features. The excitation function commonly used in convolutional neural networks is the linear rectified unit (ReLU), which is a ramp function in linear algebra. ReLU can effectively prevent gradient disappearance and enhance features. In this embodiment, ReLU is used as the activation function.

[0069] (3) Pooling layer

[0070] The pooling layer is a process of reducing the dimensionality of high-dimensional data. The pooling layer can reduce the redundancy of information and prevent overfitting. The pooling layer is mainly controlled by the pooling size, stride and padding. The main pooling methods are Lp pooling, random / mixed pooling, and spectral pooling.

[0071] (4) Fully connected layer and Softmax classifier

[0072] In the CNN structure, one or more fully connected layers are usually connected after the convolutional layer and the pooling layer. The fully connected layer can realize the linear combination of the features extracted by the convolutional layer and the high-dimensional nonlinear transformation of the data. The commonly used classifier of CNN is the Softmax classifier, which uses the Softmax logistic regression to classify after the linear combination of the features through the fully connected layer. The result obtained by the Softmax classifier is the probability distribution of the Markov transition field encoding map of each sample waveform belonging to different labels.

[0073] (5) CNN network construction principles

[0074] The impact of depth on the network is greater than the size of the convolution kernel and the pooling kernel. In a CNN structure, the convolution layer has the greatest impact on the performance of the network. The more layers the network structure has, the stronger the model learning ability is. The deeper the CNN network structure is, the better the expression ability is. If the CNN structure is too simple, the model learning ability is poor and cannot effectively extract useful features of the input data. As the depth of the network structure increases, model training takes more time, and it becomes more difficult to build an efficient and suitable model. For the design of the convolution kernel, generally speaking, the larger the convolution kernel, the larger the receptive field is, so that the deep structure can obtain more information. The smaller the convolution kernel is, the more effective it is in extracting local features.

[0075] (6) Network model training

[0076] This embodiment selects the stochastic gradient descent optimization algorithm (SGD) to train the network model and determine the network parameter weights and biases. SGD effectively avoids redundant calculations of similar samples and usually runs faster. The high oscillation of SGD makes iterative jump out of the current local minimum to find a better local minimum. The high oscillation also makes convergence difficult. Reasonable selection of the learning rate can speed up the convergence.

[0077] According to the general output format of the waveform of the acoustic emission ground pressure monitoring system, this embodiment constructs a main network structure including a convolutional layer, a pooling layer, a fully connected layer, etc. according to the design principles of the convolutional neural network structure, targeting the application scenarios and classification difficulty of the present invention. At the same time, Dropout is used to prevent overfitting and improve the generalization ability of the model, and batch normalization (Batch Normalization, BN) is used to improve the network operation speed, and the expression ability before and after data processing is retained as much as possible.

[0078] The process of training the convolutional neural network model based on the constructed training sample set and test sample set to obtain the final downhole acoustic emission source identification and classification model is shown in the attached Figure 4 .

[0079] Based on the pre-built downhole acoustic emission source identification and classification model, downhole acoustic emission sources can be classified and identified. The present invention provides the following embodiments.

[0080] Example 1

[0081] like Figure 1 As shown, this embodiment provides a multi-classification method for downhole acoustic emission sources, including:

[0082] S1: Obtain the time series of downhole acoustic emission waveform to be detected;

[0083] S2: Markov transfer field encoding is performed on the time series of the downhole acoustic emission waveform to be detected to obtain a two-dimensional image Markov transfer field;

[0084] S3: Input the Markov transfer field of the two-dimensional image into a pre-trained downhole acoustic emission source identification and classification model, and output the classification result of the downhole acoustic emission waveform time series to be detected.

[0085] In order to give full play to the advantages of convolutional neural networks in image recognition and classification, optimize the data input format of convolutional neural networks, and overcome the difficulty of traditional time-frequency analysis methods in comprehensively extracting waveform features, the technical solution of the present invention proposes a Markov transition field encoding based on the Markov transfer matrix related theory in probability theory and mathematical statistics, and encodes the one-dimensional time series of acoustic emission signals into a two-dimensional image. This two-dimensional image not only retains the trend of change on the time scale, but also reflects the transfer probability of different time spans. Then, the convolutional neural network is used to automatically extract features from the two-dimensional image and identify and classify it. It is significantly superior to traditional waveform recognition and classification methods in terms of recognition efficiency, stability and accuracy, and can correctly and timely identify acoustic emission signals to provide reliable data support for mine ground pressure monitoring and timely warning.

[0086] Among them, the specific process of performing Markov transfer field encoding on the time series of the downhole acoustic emission waveform to be detected to obtain the Markov transfer field of the two-dimensional image can be referred to step 2 in the aforementioned process of building the downhole acoustic emission source identification and classification model: Markov transfer field encoding of the acoustic emission waveform time series, which will not be repeated here.

[0087] Example 2

[0088] This embodiment provides a multi-classification device for downhole acoustic emission sources, including:

[0089] A data acquisition module, used to acquire the time series of the downhole acoustic emission waveform to be detected;

[0090] An encoding module is used to perform Markov transfer field encoding on the time series of the downhole acoustic emission waveform to be detected, so as to obtain a two-dimensional image Markov transfer field;

[0091] The classification module is used to input the two-dimensional image Markov transfer field into a pre-trained downhole acoustic emission source identification classification model, and output the classification result of the downhole acoustic emission waveform time series to be detected; wherein the downhole acoustic emission source identification classification model is obtained by training a convolutional neural network based on the historical two-dimensional image Markov transfer field corresponding to the historical downhole acoustic emission waveform time series.

[0092] Example 3

[0093] This embodiment provides a downhole acoustic emission source multi-classification terminal, including at least one memory and a processor;

[0094] The memory stores a computer program that can be run on the processor, and when the computer program is executed by the processor, the multi-classification method for downhole acoustic emission sources as described in Example 1 is implemented.

[0095] Example 4

[0096] This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the multi-classification method for downhole acoustic emission sources as described in Embodiment 1 is implemented.

[0097] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0098] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0099] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0100] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1The steps for the functions specified in one or more boxes.

[0101] It can be understood that the same or similar parts of the above embodiments can be referenced to each other, and the contents not described in detail in some embodiments can refer to the same or similar contents in other embodiments.

[0102] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention belong.

[0103] In order to further understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with a specific engineering example.

[0104] Project Examples

[0105] Step 1: Data collection and sample set division

[0106] This example is based on the Panlong Lead-Zinc Mine in Guangxi Zhongjin Lingnan. The Panlong Mine is equipped with an STL-24 remote monitoring system with 24 channels. The monitoring data of the Panlong Lead-Zinc Mine on September 14, 2020 was manually identified, and four types of typical surrounding rock acoustic emission signals, blasting operation signals, scraper operation signals, rock drilling operation signals and other types of signals were identified. Some of the typical signals are as follows Figure 5 As shown in Figure 1, other types of signals are mainly generated by human knocking, water drops, electric locomotive operation, etc., with a small number and large uncertainty. Some other types of waveforms are shown in Figure 1. Figure 6 shown.

[0107] This time, 5000 waveform data were selected to form a sample set for experimental research, including 500 waveforms of acoustic emission events in surrounding rock bodies, 1500 waveforms of rock drilling operations, 1500 waveforms of scraper operations, 1500 waveforms of blasting operations, and 500 waveforms of other categories. Each waveform consists of 1024 sampling points. The training set and test set are divided according to the ratio of 7:3. One-hot encoding is used to add labels to the five types of signals, including acoustic emission waveforms of surrounding rock bodies, blasting operation waveforms, rock drilling operation waveforms, scraper operation waveforms, and other types of waveform labels are 0, 1, 2, 3, and 4 respectively.

[0108] Step 2: Markov transition field coding of acoustic emission waveform time series

[0109] The acoustic emission waveform time series is encoded into a two-dimensional image. According to the Markov transfer field principle, each waveform time series of the sample set is encoded with the Markov transfer field to generate a 1024×1024-dimensional Markov transfer field. The grid operation is performed on it to generate a 256×256-dimensional fuzzy Markov transfer field. The fuzzy Markov transfer field after the conversion of some waveform time series is shown in the figure. Figure 7 shown.

[0110] Step 3: Establish an acoustic emission waveform recognition and classification model

[0111] (1) This example is based on the network design principles and the classic network structure LeNet-5. Through personal experience and repeated experiments, the relevant hyperparameters are adjusted, and the parameters of the model structure are finally obtained as shown in Table 2. The network structure mainly includes convolution layer, pooling layer, dropout, Batch Normalization, and fully connected layer. In this example, the convolutional neural network structure mainly includes convolution layers C1, C2, and C3, and the convolution kernel size is 3×3. Dropout layers D1, D2, D3, and D4, and fully connected layers FC1 and FC2. Dropout is used to reduce overfitting and improve the generalization ability of the model. Batch Normalization is performed after each dropout to increase the training speed of the network. The fully connected layer FC1 linearly combines the features extracted by the convolution layer, and the fully connected layer FC2 realizes the high-dimensional nonlinear transformation of the input data.

[0112] Table 2 Main structural parameters of the model

[0113]

[0114]

[0115] (2) Results Analysis

[0116] The training set and test set are input into the above network structure respectively. After 150 iterations, the accuracy curve and loss curve are obtained as follows: Figure 8-Figure 9 shown.

[0117] The results show that when the iteration reaches 70 times, the accuracy of the training set reaches 92.56%, the accuracy of the test set reaches 91.51%, and the Loss value is reduced to between 0.1 and 0.25. The model has fully converged, and then the accuracy and Loss loss functions tend to be stable. It can be proved that the Markov transfer field-CNN model proposed in this paper is effective and feasible in identifying and classifying underground surrounding rock mass acoustic emission signals and mining operation noise signals.

[0118] Fig.10The waveform classification confusion matrix shown in the figure is used to further analyze the feature extraction and waveform classification capabilities of CNN for Markov transfer field coding images. Fig.10 The middle coordinates 0-4 represent blasting waveforms, rock drilling waveforms, scraper operation waveforms, acoustic emission waveforms, and other types of waveforms. The values ​​in the dark area on the diagonal represent the accuracy of each category, and the values ​​in the light area represent the classification error rate.

[0119] (3) Verification of model generalization ability

[0120] In actual underground mining engineering applications, due to the complex underground environment, the data measured by the remote monitoring system will inevitably be interfered by background noise, so the study of the model's adaptability to noise is also important. This section uses the original data without adding noise as the training set, and adds Gaussian white noise with different signal-to-noise ratios to the original waveform to form a test set. Adding white noise can simulate the impact of various types of waveform classification affected by noise interference. The signal-to-noise ratio (SNR) is defined as the ratio of signal power to noise power. The smaller the signal-to-noise ratio, the greater the noise power. The unit of measurement is dB, and the calculation method is:

[0121] SNR = 101g (P s / P n )

[0122] Among them, Ps and Pn represent the signal power and noise power respectively.

[0123] This section will discuss the waveform recognition performance of the dataset in a noisy environment. The SNR is set to -4-8dB. The accuracy of the test set under different signal-to-noise ratios is as follows: Fig.11 As shown in the figure, without reinforcement learning, when SNR is less than 1dB, the current model has weak signal recognition ability. When SNR is 6, the accuracy reaches 90.36%. Then, as the signal-to-noise ratio increases, the recognition accuracy approaches 86%.

[0124] In order to further improve the generalization ability and noise resistance of the model, 25 samples with SNR = 0dB in each type of waveform are randomly selected, a total of 100 samples, and the samples are added to the original training set for reinforcement learning. The parameters are fine-tuned to obtain the reinforcement learning accuracy curve as shown in the figure. Fig.11 By comparing the accuracy curves before and after enhancement, it can be concluded that after adding noise components to part of the original data, the model can learn the data features under noise interference, reduce the overfitting problem, and greatly improve the model recognition accuracy in a noisy environment.

[0125] (4) Comparative analysis

[0126] In order to show the advancedness of the present invention based on Markov transfer field-CNN, the recognition and classification of the original waveform image is compared with the traditional machine learning method support vector machine (SVM), artificial neural network (ANN) and CNN. The environment in which the model runs is Spyder, the hardware environment is Intel (R) Core (TM) i7-9750H, and the graphics card is GTX1650. The SVM parameters are set as follows: the kernel function is a Gaussian radial odd function, the penalty factor is 8, the kernel function radius is 0.4, and other parameters are defaulted. The ANN parameters are set as follows: the hidden layer contains 100 neurons, the learning rate is 0.2, the momentum factor is 0.05, and the maximum number of iterations is 100 times. The recognition and classification parameters of the original image are set as shown in Table 3.

[0127] Table 3 CNN structure parameters for original waveform recognition and classification

[0128]

[0129] In order to evaluate the performance of the above classification model, three indicators are used for evaluation: accuracy, precision, and recall. The calculation formulas of each indicator are shown in the formula:

[0130]

[0131]

[0132]

[0133] In the formula, a is the number of samples of a certain category (assuming the sample label is 1) that are correctly identified; b is the number of other samples that are identified as category 1; c is the number of samples of category 1 that are incorrectly identified; and d is the number of samples of other categories that are correctly identified.

[0134] In order to reduce the accidental error, four methods were used for 10 tests each, and the average values ​​of the indicators of the 10 test results were used as model evaluation indicators. The values ​​of the indicators are shown in Table 4. The method proposed in the present invention accurately identified various waveforms, and the indicators reached more than 91%. The indicators of the Markov transfer field-CNN method in this paper are significantly better than the traditional methods SVM and ANN. The comparison between Markov transfer field-convolutional neural network and CNN shows that the waveform time series is encoded as a two-dimensional image as the recognition object of CNN, which has a higher accuracy rate. CNN is more sensitive to the feature extraction and recognition classification of two-dimensional images.

[0135] Table 4 Performance comparison of Markov transfer field-CNN and SVM, ANN, CNN (%)

[0136]

[0137]

[0138] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A multi-classification method for downhole acoustic emission sources, characterized in that, it includes: Obtain the time series of the downhole acoustic emission waveform to be detected; Perform Markov transfer field encoding on the time series of the downhole acoustic emission waveform to be detected to obtain a two-dimensional image Markov transfer field; Input the two-dimensional image Markov transfer field into a pre-trained downhole acoustic emission source recognition and classification model, and output the classification result of the time series of the downhole acoustic emission waveform to be detected; wherein, the downhole acoustic emission source recognition and classification model is trained based on the historical two-dimensional image Markov transfer field corresponding to the historical downhole acoustic emission waveform time series; The performing Markov transfer field encoding on the time series of the downhole acoustic emission waveform to be detected to obtain a two-dimensional image Markov transfer field specifically includes: For the downhole acoustic emission waveform time series X to be detected x 1 ,x 2 ,…,xn ], determine Q quantiles, which means that the downhole acoustic emission waveform to be detected is divided into Q intervals in the vertical direction, and each xi Assign to the corresponding interval qj , j ∈[1,Q]; Calculate the transfer between the quantile intervals of the time series of the downhole acoustic emission waveform to be detected in the manner of a first-order Markov chain along the time axis to construct a Q×Q weighted adjacency matrix; obtain the weighted Markov transfer matrix W after normalization; By considering the time position, expand the weighted Markov transfer matrix W containing the transfer probability on the amplitude axis into a Markov transfer field matrix to obtain an n×n Markov transfer field M: ； Conditional Probability Indicates that the time series is a When in the interval qi Under the premise of a+h Transfer to interval qj The conditional probability is .

2. The multi-classification method for downhole acoustic emission sources according to claim 1, characterized in that, after obtaining the n×n Markov transfer field M, it further includes: Grid the Markov transfer field M; Replace the subgraph in each grid with its average value to generate a fuzzy Markov transfer field.

3. The multi-classification method for downhole acoustic emission sources according to any one of claims 1 to 2, characterized in that, the downhole acoustic emission source recognition and classification model is trained based on the historical two-dimensional image Markov transfer field corresponding to the historical downhole acoustic emission waveform time series, specifically including: Obtain a number of historical downhole acoustic emission waveform time series and their respective corresponding acoustic emission source categories; Perform Markov transfer field encoding on the number of historical downhole acoustic emission waveform time series to obtain the corresponding historical two-dimensional image Markov transfer field; Based on the obtained historical two-dimensional image Markov transfer field and its corresponding acoustic emission category label, construct a training sample set and a test sample set; Taking the historical two-dimensional image Markov transfer field as the input and the acoustic emission category as the output, train the convolutional neural network based on the training sample set and the test sample set to obtain a downhole acoustic emission source recognition and classification model.

4. The multi-classification method for downhole acoustic emission sources according to claim 3, characterized in that, the acoustic emission source categories include surrounding rock mass acoustic emission signals, blasting operation signals, load-haul-dump operation signals, rock drilling operation signals and other category signals; both the training sample set and the test sample set contain samples of these five categories.

5. The multi-classification method for downhole acoustic emission sources according to claim 3, characterized in that, when training the downhole acoustic emission source recognition and classification model, use the stochastic gradient descent optimization algorithm to train the convolutional neural network.

6. A multi-classification device for downhole acoustic emission sources, characterized in that, it includes: A data acquisition module, used to acquire the time series of the downhole acoustic emission waveform to be detected; An encoding module is used to perform Markov transfer field encoding on the time series of the downhole acoustic emission waveform to be detected, so as to obtain a two-dimensional image Markov transfer field; A classification module is used to input the two-dimensional image Markov transfer field into a pre-trained downhole acoustic emission source identification classification model, and output the classification result of the downhole acoustic emission waveform time series to be detected; wherein the downhole acoustic emission source identification classification model is obtained by training a convolutional neural network based on the historical two-dimensional image Markov transfer field corresponding to the historical downhole acoustic emission waveform time series; The Markov transfer field encoding of the downhole acoustic emission waveform time series to be detected to obtain the Markov transfer field of the two-dimensional image specifically includes: For the downhole acoustic emission waveform time series X to be detected x 1 ,x 2 ,…,xn ], determine Q quantiles, which means that the downhole acoustic emission waveform to be detected is divided into Q intervals in the vertical direction, and each xi Assign to the corresponding interval qj , j ∈[1,Q]; The transfer between each quantile interval of the downhole acoustic emission waveform time series to be detected is calculated along the time axis in the manner of a first-order Markov chain to construct a Q×Q weighted adjacency matrix; the weighted Markov transfer matrix W is obtained after normalization; By considering the time position, the weighted Markov transfer matrix W containing the transition probabilities on the amplitude axis is expanded into the Markov transfer field matrix, resulting in an n×n Markov transfer field M: ； Conditional Probability Indicates that the time series is a When in the interval qi Under the premise of a+h Transfer to interval qj The conditional probability is .

7. A multi-classification terminal for underground acoustic emission sources, It is characterized in that comprising at least one memory and a processor; The memory stores a computer program that can be run on the processor, and when the computer program is executed by the processor, the multi-classification method for downhole acoustic emission sources according to any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium storing a computer program, It is characterized in that When the computer program is executed by a processor, the multi-classification method for downhole acoustic emission sources as claimed in any one of claims 1 to 5 is implemented.

Images