Micro-seismic monitoring event classification method
By combining the single-channel classification model based on deep learning and the microseismic event classification model based on machine learning in the microseismic monitoring system, the problems of high complexity and high migration deployment cost in the existing technology are solved, and efficient microseismic event classification and rapid deployment are achieved.
Patent Information
- Application Number
- CN202411759084.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-03
- Publication Date
- 2025-03-11
AI Technical Summary
When existing microseismic monitoring systems process multi-channel data, the network model is complex, lacks modular flexibility, high migration and deployment costs, and it is difficult to achieve specific optimization through component modification or replacement.
A single-channel classification model based on deep learning is used to focus on the characteristics within the channel, and combined with a microseismic event classification model based on machine learning, the obtained single-channel information is integrated for classification. This method enhances the interpretability of the system and the decoupling between components by separating the two stages of single-channel feature learning and multi-channel information integration.
It realizes accurate classification of microseismic events, rapid learning deployment, and reduces migration and training costs, improves the robustness and classification accuracy of the system, and reduces hardware investment.
Smart Images
Figure CN119598350A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of microseismic monitoring, and in particular to a microseismic monitoring event classification method. Background Art
[0002] Microseismic events caused by engineering production such as mining, oil and gas extraction, groundwater extraction or natural gas injection usually endanger production safety, so they usually need to be continuously monitored to timely discover and prevent and resolve further risks. Taking mining operations as an example, it is usually necessary to reasonably arrange sensors in the monitoring area to collect elastic waves generated by tiny fractures in the rock mass in real time. In the actual monitoring and collection process, due to various complex mining and rock environments, blasting, transportation, anchoring and other production activities, the microseismic monitoring system will collect a large number of invalid interference signals, which will cause false alarms and missed alarms in the micro-monitoring system. On the other hand, if a large number of invalid signals are not screened, they will seriously affect the working efficiency of the microseismic monitoring system. Therefore, it is necessary to classify the event types of the signals collected by the microseismic monitoring system.
[0003] Microseismic event classification is a research focus in the field of microseismic monitoring. Its core is to identify signals of specific event types from the numerous signals recorded by the monitoring system. In the framework of two-classification tasks, "specific event types" only refer to "microseismic" events, and other types of signals can be considered as "noise" signals. In the more detailed multi-classification task framework, "specific event types" usually also include event types such as blasting, excavation, and knocking. The current microseismic event classification methods can be divided into two types, including manual recognition and classification and automatic recognition and classification. Manual recognition and classification have high accuracy, but they rely too much on the subjective experience of staff, and cannot process large-scale signal data in a timely and effective manner, which has obvious limitations. The automatic recognition and classification methods are usually based on data-driven, using machine learning or deep learning technology to learn the discriminative differences between different types of events through model training and other methods, and automatically recognize and classify the collected signals. This type of method has high recognition efficiency and can effectively deal with large-scale signal data that flood into the system in a short period of time.
[0004] In the automatic recognition and classification methods, machine learning or deep learning methods are mainly used. Classification methods based on machine learning usually select appropriate features first, and then input the features into the machine learning classification algorithm. The selected features generally include time domain features, frequency domain features, and time-frequency features. The machine learning algorithms used include: support vector machine algorithm, decision tree, and random forest. In current research, the Light Gradient Boosting Machine (LightGBM), eXtreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost) algorithms in ensemble learning are also commonly used. Classification methods based on deep learning usually integrate feature selection and classification in a unified network to achieve end-to-end recognition. This type of research avoids the complexity of multi-step processing. The model can automatically learn and extract features from the input data, and also reduces the dependence on domain knowledge. In addition, through the overall network structure, the model can be trained and predicted more efficiently, and the optimization process is more consistent. It usually performs well on large-scale data sets. However, the end-to-end model compresses the entire process into a network model. The various parts of the model are highly dependent on each other and difficult to optimize or replace independently. It lacks modular flexibility. In addition, the various layers and parameters of the model are interdependent and complex. When the model needs to be improved, it is difficult to modify or replace individual components.
[0005] Although traditional end-to-end models can directly learn features from raw data and complete classification tasks, their internal structures are highly coupled, making it difficult to independently optimize or replace components, and the migration cost is high when applied to different scenarios. In addition, such models usually require a large amount of labeled data to achieve good performance. Existing signal classification methods for microseismic monitoring systems are usually implemented based on deep learning methods. The network models used in such methods are highly complex and lack modular flexibility. It is difficult to achieve optimization and improvement in specific aspects by modifying or replacing components, and the interpretability and decoupling are poor. On the other hand, such methods are usually designed for specific operating scenarios, and have poor portability when migrated and deployed to other operating environments, and the migration and deployment costs are high. Summary of the invention
[0006] In view of the above problems, a microseismic monitoring event classification method is provided. The present invention focuses on the intra-channel features through a single-channel classification model based on deep learning, and uses a microseismic event classification model based on machine learning to integrate the obtained single-channel information to predict the classification results. This method has good interpretability and decoupling. The single-channel classification model only focuses on the characteristics of a single channel. Therefore, when migrating and deploying, compared with the end-to-end classification model that directly uses multiple channel data as input, the training time and hardware cost are significantly reduced, and the migration cost is low.
[0007] To achieve the above object, the technical solution adopted by the present invention is.
[0008] A microseismic monitoring event classification method comprises the following steps: Step 1: acquiring historical data of multiple channels through multiple sets of sensors constituting a sensor array, and marking the historical data with event categories , i ={1, 2, ..., n}, n is the number of channels; Step 2: Split the annotated historical data in step 1 by the number of channels, and annotate each single-channel historical data with the corresponding event category. y s Input them into the single-channel classification model to obtain the prediction results of the single-channel classification model for each single-channel historical data : in, x i For the i Historical data input from each channel, i ={1, 2, ..., n}, n is the number of channels; f 1 is the function mapping relationship corresponding to the single-channel classification model; the arg max function represents the index of obtaining the maximum value; P i is the output of the single-channel classification model i The probability distribution corresponding to each channel; P i1 For the i The probability that the channel data belongs to event category 1, j ={1,2, ..., m}; m is the number of event categories; Step 3: Set the threshold to adjust the probability distribution of the output in step 2 p i Threshold truncation is performed to remove redundant auxiliary information of event category attributes and obtain the probability value after threshold truncation. : in, θ To set the threshold; f 2 is the function mapping relationship corresponding to the threshold cutoff; P ij Indicates i Channel data belongs to the category j The probability of Step 4: After the threshold is cut off in step 3, n channels, m The probability value of the number of event categories Concatenate the vectors one by one to get a vector with a length of n × m A one-dimensional vector V: Step 5: Use the event category annotation in step 1 as the label of the one-dimensional vector V y s , and the 1-dimensional vector V and its label y s Input to the microseismic event classification model and output the prediction results of microseismic event categories : in, f 3 It is the function mapping relationship corresponding to the microseismic event classification model.
[0009] Preferably, the historical data of the multiple groups of channels are in the form of sgy files or ndarray arrays.
[0010] Preferably, the output of the single-channel classification model is i The probability distribution corresponding to the channel p i Event categories j The sum of the probabilities is 1, that is .
[0011] Preferably, the event categories include microseismic events, rock burst events, blasting events, percussion events, excavation events, drilling events, light source vibration events, oil hammer events and background noise events.
[0012] Preferably, the threshold value is set θ The value range is (0, 1].
[0013] Preferably, the threshold value is set θ is 0.3.
[0014] Preferably, the single-channel classification model uses the AST audio event classification model as a pre-trained model.
[0015] Preferably, the function mapping relationship corresponding to the single-channel classification model is f 1 Including logits output and activation function.
[0016] Preferably, the activation function includes a softmax function or a sigmoid function.
[0017] Preferably, the microseismic event classification model adopts a hard voting classifier composed of 5 base classifiers based on a machine learning model.
[0018] Preferably, the base classifiers include support vector machine SVW, random forest RF, lightweight gradient boosting machine LightGBM, extreme gradient boosting XGBoost and classification boosting CatBoost.
[0019] Due to the adoption of the above technical solution, the present invention has the following beneficial effects.
[0020] (1) The present invention solves the problems of high network model complexity, lack of modular flexibility, and high migration and deployment costs in the prior art by combining a single-channel classification model based on deep learning with a microseismic event classification model based on machine learning. It achieves the effects of accurately classifying microseismic events, quickly learning and deploying, and reducing migration and training costs.
[0021] (2) By separating the two stages of single-channel feature learning and multi-channel information integration, the present invention enhances the interpretability of the entire system and the decoupling between components, making it easier to optimize specific aspects. Since the single-channel classification model only focuses on the characteristics of a single channel, it only requires a small amount of adjustment when migrating to a different operating environment to adapt to the new environment, greatly reducing the migration cost.
[0022] (3) The present invention improves the overall accuracy of microseismic event classification by making full use of the detailed features in a single channel and cross-channel information. At the same time, compared with the method of directly processing multi-channel data, this solution significantly reduces the demand for GPU video memory and reduces hardware investment.
[0023] (4) The present invention collects historical data collected by the microseismic monitoring system, i.e., multiple groups of sensors, and labels the historical data for event categories, thereby ensuring sufficient and accurate data support in the subsequent model training process. The single-channel classification model based on deep learning is trained with the labeled historical data to focus on extracting features from the data of each individual channel. The pre-trained AST audio event classification model is used as the basis for pre-training the model, which can be effectively transferred to the microseismic event classification task, thereby reducing the amount of labeled data required and improving the model performance, thereby improving the efficiency and accuracy of microseismic time category recognition.
[0024] (5) The present invention processes the probability distribution output by the single-channel classification model by setting a threshold, and splices all processed single-channel probability distributions into a one-dimensional vector, which can effectively retain key features while reducing redundant information, and provide clearer data input for the training and analysis of subsequent microseismic event classification models. Using the one-dimensional vector and its label as input, the microseismic event classification model is trained through an ensemble learning method, i.e., a hard voting classifier composed of five base classifiers, and the final microseismic event classification is achieved. This method combines the advantages of multiple base classifiers and improves the robustness and classification accuracy of the overall system.
[0025] (6) The present invention focuses on the intra-channel features through a single-channel classification model based on deep learning, and uses a microseismic event classification model based on machine learning to integrate the obtained single-channel information to predict the classification results. This method has good interpretability and decoupling. The single-channel classification model only focuses on the characteristics of a single channel. Therefore, when migrating and deploying, compared with the end-to-end classification model that directly uses multiple channel data as input, the training time and hardware cost are significantly reduced, and the migration cost is low. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] The following is a detailed discussion of the making and application of the preferred embodiments of the present invention. However, it should be understood that the present invention provides many applicable inventive concepts that can be embodied in various specific environments. The specific embodiments discussed are only for the purpose of illustrating the specific ways of making and using the present invention, and do not limit the scope of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative work.
[0027] Figure 1 It is a flow chart of the present invention.
[0028] Figure 2 It is a framework diagram of the multi-channel event classification method for microseismic monitoring of the present invention.
[0029] Figure 3 This is a structural diagram of the hard voting classifier of the present invention. DETAILED DESCRIPTION
[0030] The following is a detailed discussion of the making and application of the preferred embodiments of the present invention. However, it should be understood that the present invention provides many applicable inventive concepts that can be embodied in various specific environments. The specific embodiments discussed are only intended to illustrate specific ways to make and use the present invention and do not limit the scope of the present invention.
[0031] The present invention solves the problems of high network model complexity, lack of modular flexibility and high migration and deployment costs in the prior art by combining a single-channel classification model based on deep learning with a microseismic event classification model based on machine learning, thereby achieving the effects of accurately classifying microseismic events, quickly learning and deploying, and reducing migration and training costs.
[0032] The following is combined with Figure 1-3 Further elaboration.
[0033] like Figure 1 , Figure 2 The method for classifying microseismic monitoring events shown in the figure comprises the following steps: Step 1: acquiring historical data of multiple groups of channels through multiple groups of sensors constituting a sensor array, and marking the historical data with event categories , i ={1, 2, ..., n}, n is the number of channels. The historical data of the multiple groups of channels are in the form of sgy files or ndarray arrays. The event categories include microseismic events, rock burst events, blasting events, knocking events, excavation events, drilling events, light source vibration events, oil hammer events and background noise events.
[0034] Step 2: Split the annotated historical data in step 1 by the number of channels, and annotate each single-channel historical data with the corresponding event category. y s Input them into the single-channel classification model to obtain the prediction results of the single-channel classification model for each single-channel historical data : in, x i For the i Historical data input from each channel, i ={1, 2, ..., n}, n is the number of channels; f 1 is the function mapping relationship corresponding to the single-channel classification model, and the function mapping relationship corresponding to the single-channel classification model f 1 It includes logits output and activation function; the activation function includes softmax function or sigmoid function. The arg max function indicates the index of obtaining the maximum value; P i is the output of the single-channel classification model i The probability distribution corresponding to each channel; P i1 For the iThe probability that the channel data belongs to event category 1, j ={1, 2, ..., m}; m is the number of event categories; the output of the single-channel classification model is i The probability distribution corresponding to the channel p i Event categories j The sum of the probabilities is 1, that is The single-channel classification model uses the AST audio event classification model as a pre-training model.
[0035] Step 3: Set the threshold to adjust the probability distribution of the output in step 2 p i Threshold truncation is performed to remove redundant auxiliary information of event category attributes and obtain the probability value after threshold truncation. : in, θ is to set a threshold value, the setting threshold value θ The value range is (0, 1], and the threshold value set in this embodiment θ is 0.3; f 2 is the function mapping relationship corresponding to the threshold cutoff; P ij Indicates i Channel data belongs to the category j probability; Step 4: After the threshold is cut off in step 3, n channels, m The probability value of the number of event categories Concatenate the vectors one by one to get a vector with a length of n × m A one-dimensional vector V: .
[0036] Step 5: Use the event category annotation in step 1 as the label of the one-dimensional vector V y s , and the 1-dimensional vector V and its label y s The input is sent to the microseismic event classification model, which uses a hard voting classifier composed of five base classifiers based on the machine learning model. Figure 3 As shown in the figure, the base classifiers include support vector machine SVW, random forest RF, lightweight gradient boosting machine LightGBM, extreme gradient boosting XGBoost and classification boosting CatBoost. The final output is the prediction result of the microseismic event category. : in, f 3 It is the function mapping relationship corresponding to the microseismic event classification model.
[0037] Further, combined with Figure 1-3 Give detailed explanation.
[0038] like Figure 1 A microseismic monitoring event classification method is shown in FIG. Step 1: Acquisition and annotation of historical data. The signals collected by the microseismic monitoring system are generally obtained through multiple sets of sensors deployed. These signals are processed by the hardware end in the form of "sgy files containing multiple sets of channels" as system input, and each set of channels corresponds to a set of sensors.
[0039] First, obtain multiple sets of channel historical data collected in the monitoring system as training data and then annotate them. The annotations required include: first annotate all channels of the sgy file with event categories , and then label the categories of all single channels in the sgy file, i ={1, 2,..., n}, n is the number of channels.
[0040] In the labeling process, first determine the overall event category of the sgy file for all channels, such as microseismic events, rock burst events, blasting events, percussion events, excavation events, drilling events, light source vibration events, oil hammer events and background noise events, that is, obtain the label for the entire file y s Then filter out the noise channels that do not contain event information in the data file, and mark the single channel event category of these channels as "noise". After completion, make the event category labeling of the remaining channels consistent with the category labeling of the overall event y s Keep it the same.
[0041] The present invention collects historical data collected by a microseismic monitoring system, i.e., multiple groups of sensors, and labels the historical data with event categories, thereby ensuring sufficient and accurate data support in the subsequent model training process.
[0042] Step 2: Training of single-channel classification model. After completing data labeling, use the data in step 1 to train the "single-channel classification model". Split the input sgy file by channel, and then label the split single-channel historical data together with the corresponding event category. y s Input them into the single-channel classification model to obtain the prediction results of the single-channel classification model for each single-channel historical data : in, x i For the i Historical data input from each channel, i ={1, 2, ..., n}, n is the number of channels; f 1 is the function mapping relationship corresponding to the single-channel classification model, and the function mapping relationship corresponding to the single-channel classification model f 1 It includes logits output and activation function; the activation function includes softmax function or sigmoid function. The arg max function indicates the index of obtaining the maximum value; P i is the output of the single-channel classification model i The probability distribution corresponding to each channel; P i1 For the i The probability that the channel data belongs to event category 1, j ={1, 2, ..., m}; m is the number of event categories; the output of the single-channel classification model is i The probability distribution corresponding to the channel p i Event categories j The sum of the probabilities is 1, that is The single-channel classification model uses the AST audio event classification model as a pre-training model.
[0043] After completing the training of the single-channel classification model, the model's logits output is output and normalized using an activation function to obtain the probability distribution of the model output for each event category for the input single-channel data.
[0044] The present invention focuses on extracting features from the data of each individual channel by training a single-channel classification model based on deep learning through labeled historical data. The pre-trained AST audio event classification model is used as the basis for model pre-training, which can be effectively transferred to the microseismic event classification task, thereby improving the model performance while reducing the amount of required labeled data, thereby improving the efficiency and accuracy of microseismic time category recognition.
[0045] Step 3: Threshold truncation and vector concatenation of probability distribution. The probability distribution output by the single-channel classification model is processed according to the set truncation threshold. The probability distribution output by the single-channel classification model in step 2 reflects the correspondence between the input data and each event category. Compared with the prediction result obtained by probability distribution, probability distribution can provide more auxiliary information about the data category attributes. For example, when p i1 =0.6 p i2 =0.3, it can be inferred that the model believes that the data of this channel belongs to category 1, but also contains some features of category 2. These auxiliary information are useful for subsequent further analysis, but when the probability of a certain category is too small (for example, the probability is 0.3), since the model has been trained and has good credibility, it can be considered that the correlation between the channel data and this category is small. In this invention, by setting a threshold, the probability less than the threshold is set to zero, and there is in, θ is to set a threshold value, the setting threshold value θ The value range is (0, 1], and the threshold value set in this embodiment θ is 0.3; f 2 is the function mapping relationship corresponding to the threshold cutoff; P ij Indicates i Channel data belongs to the category j probability.
[0046] Next, the threshold truncated n The probability distributions of the channels are concatenated in sequence to form a n × m A one-dimensional vector V: , the value of each element in this vector is the probability value.
[0047] Step 4: Training of multi-channel classification model. The event category annotation in step 1 is used as the label of the one-dimensional vector V y s , and the 1-dimensional vector V and its label y s Input into the microseismic event classification model, the one-dimensional vector V and its label y s As the input of the "microseismic event classification model" that integrates single-channel information, after the model training is completed, the application can be deployed and the prediction results can be output. in, f 3is the function mapping relationship corresponding to the microseismic event classification model. Finally, the prediction result is used as the final output of the classification method, indicating the corresponding event category predicted by the method for the input sgy file containing multiple groups of channels.
[0048] The present invention processes the probability distribution output by the single-channel classification model by setting a threshold, and splices all processed single-channel probability distributions into a one-dimensional vector, which can effectively retain key features while reducing redundant information, and provide clearer data input for the training and analysis of subsequent microseismic event classification models. Using a one-dimensional vector and its label as input, the microseismic event classification model is trained through an ensemble learning method, i.e., a hard voting classifier composed of five base classifiers, and the final microseismic event classification is achieved. This method combines the advantages of multiple base classifiers and improves the robustness and classification accuracy of the overall system.
[0049] like Figure 2 The framework diagram of the multi-channel event classification method for microseismic monitoring is shown in FIG. First, the signals collected by the microseismic monitoring system are processed by the hardware end and input as multi-channel historical signals in the form of sgy files of multiple channels. Secondly, the input sgy files are split by channel, and the signal data of each channel is input into the trained "single channel classification model" respectively, and the output results of the probability distribution of each channel are obtained as the output of the single channel classification model.
[0050] Then, the probability distribution is truncated according to the set threshold to remove redundant category auxiliary information. n The probability distributions of the channels are concatenated in sequence to form a one-dimensional vector V with a length of n×m. Finally, the one-dimensional vector V and its label y s It serves as the input of the "microseismic event classification model" that integrates single-channel information and outputs the prediction result, which is the final output of the microseismic event classification method, indicating the corresponding event category predicted by this method for the input signal data containing multiple groups of channels.
[0051] In addition, the basic idea of the present invention is to use a combination of models, taking into account single-channel information and inter-channel information, to achieve accurate classification of microseismic events. The "input data" in the aforementioned technical solution is generally in the form of an sgy file, which is a commonly used data format in the field of earthquakes and microseisms. In addition, there are also some forms using ndarray arrays, which can also be used as input of the method proposed in the present invention after splitting and necessary conversion.
[0052] The "single-channel classification model" in the above technical solution can be implemented by using a pre-trained audio event classification model. Specifically, the Audio Spectrogram Transformer (AST) model can be used as a pre-trained model, and on this basis, the training is completed through the obtained microseismic event data. The AST model is an audio event classification model, and the scale of the data used for its training is about 2 million. The data used for training includes a variety of audio events such as humans, animals, music, and natural sounds. With the help of transfer learning theory, using the pre-trained AST model, better recognition results can be achieved by training on less microseismic event data.
[0053] For the "probability distribution" in the above technical solution, in multi-classification tasks, the softmax function is used to convert the logits output into a probability distribution; in binary classification tasks, the sigmoid function is used to convert the logits output into a probability distribution, and softmax will make the sum of the probabilities that the input data belongs to all categories equal to 1, while the sigmoid function only makes the probability that the input data belongs to a certain category between 0 and 1, but does not make "the sum of the probabilities that the input data belongs to all categories equal to 1". The embodiment of the present invention defaults to a multi-classification task and uses the softmax function to obtain the probability distribution. In actual applications, after the probability distribution obtained by the sigmoid function is threshold-cut, it is used as the input of the "microseismic event classification model" that integrates single-channel information, which can provide better training effects. This is because the sigmoid function does not make the sum of the output equal to 1, which relatively reduces the difference between high-probability elements and low-probability elements, and amplifies the influence of auxiliary information on the microseismic event classification model. Therefore, in this embodiment, the sigmoid function is used to convert the logits output.
[0054] The "microseismic event classification model" that integrates single-channel information in the above technical solution can be implemented through a machine learning model. Compared with deep learning models, machine learning models require less data, and the amount of data on microseismic events can meet the requirements. The machine learning models used in the present invention include: Support Vector Machine (SVM), Random Forest (RF), Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (eXtreme Gradient Boosting, XGBoost) and Categorical Boosting (CatBoost). Except for SVM, the other four models are ensemble learning models. The present invention selects these five models as the base classifiers of the hard voting classifier, and uses the composed hard voting classifier as the "microseismic event classification model" that integrates single-channel information, that is, the hard voting classifier composed of five base classifiers is used as the implementation of the "microseismic event classification model" that integrates single-channel information. The structure of the hard voting classifier is as follows Figure 3 As shown, the random initial states of each base classifier are fixed to ensure the reproducibility of training.
[0055] Compared with the prior art, the present invention has the following advantages.
[0056] 1. The representation ability of machine learning is weaker than that of deep learning, but the demand for data scale is also less than that of deep learning. Therefore, the machine learning model is used to directly classify the sgy files containing multiple channels. Since it is difficult to learn features with good discriminability, the classification effect is not good. When the deep learning model is used to directly use the sgy files as input, it is limited by the scale of labeled data. The present invention uses a single-channel classification model based on deep learning to focus on the learning of intra-channel features. The amount of data at this stage is n times ( n is the number of channels), which helps to play the role of deep learning. The use of a microseismic event classification model based on machine learning reduces the requirements for feature characterization by integrating the obtained single-channel information, which helps to play the role of machine learning. The combination of the two can provide more accurate classification results.
[0057] 2. Taking mining operations as an example, when the monitoring system deployed in a coal mine is applied to another coal mine, due to the similarity of the features within the channel, the single-channel classification model only needs simple fine-tuning to be deployed and applied, and less labeled data is required. The time cost of training the microseismic event classification model affected by the relationship between channels is significantly lower than that of the "deep learning-based classification model that directly uses sgy files as input". From the perspective of hardware requirements, since only the features within the channel are concerned, the GPU video memory used for training the single-channel classification model proposed in this invention is only 1 / 4 of the video memory used for training the same type of multi-channel classification model. n . Therefore, the method proposed in the present invention is more tolerant of the hardware requirements for model training.
[0058] 3. The method proposed in this invention has better interpretability and decoupling, that is, the single-channel model focuses on the features within the channel, and the event classification model focuses on the information between channels. The end-to-end model used by other solutions is highly complex, lacks modular flexibility, and the various parts of the model are highly dependent. It is difficult to achieve optimization and improvement in specific aspects by modifying or replacing the components therein, and the interpretability and decoupling are poor.
[0059] The present invention enhances the interpretability of the entire system and the decoupling between components by separating the two stages of single-channel feature learning and multi-channel information integration, making it easier to optimize specific aspects. Since the single-channel classification model only focuses on the characteristics of a single channel, it only requires a small amount of adjustment to adapt to the new environment when migrating in different operating environments, greatly reducing the migration cost. The present invention improves the overall accuracy of microseismic event classification by making full use of the detailed features and cross-channel information within a single channel. At the same time, compared with the method of directly processing multi-channel data, this solution significantly reduces the demand for GPU video memory and reduces hardware investment.
[0060] The present invention focuses on the characteristics within the channel through a single-channel classification model based on deep learning, and uses a microseismic event classification model based on machine learning to integrate the obtained single-channel information to predict the classification results. This method has good interpretability and decoupling. The single-channel classification model only focuses on the characteristics of a single channel. Therefore, when migrating and deploying, compared with the end-to-end classification model that directly uses multiple channel data as input, the training time and hardware cost are significantly reduced, and the migration cost is low.
[0061] Although the specification has been described in detail, it should be understood that various changes, substitutions and modifications may be made without departing from the spirit and scope of the invention as defined by the appended claims. In addition, the specific embodiments described are not intended to limit the scope of the invention, and those of ordinary skill in the art can easily understand based on the present invention that currently existing or later to be developed processes, machines, manufactures, material compositions, means, methods, or steps can perform substantially the same functions as the embodiments of the present invention or obtain substantially the same results. Therefore, the appended claims are intended to include such processes, machines, manufactures, material compositions, means, methods or steps within their scope.
Claims
1. A microseismic monitoring event classification method, characterized by: The following steps are involved: Step 1: Obtain historical data of multiple channels through multiple sets of sensors that form a sensor array, and label the historical data with event categories , i ={1, 2, ..., n }, n is the number of channels; Step 2: Split the annotated historical data in step 1 by the number of channels, and annotate each single-channel historical data with the corresponding event category. y s Input them into the single-channel classification model to obtain the prediction results of the single-channel classification model for each single-channel historical data : in, x i For the i Historical data input from each channel, i ={1, 2, ..., n }, n is the number of channels; f 1 is the function mapping relationship corresponding to the single-channel classification model; the arg max function represents the index of obtaining the maximum value; P i is the output of the single-channel classification model i The probability distribution corresponding to each channel; P i1 For the i The probability that the channel data belongs to event category 1, j ={1, 2,..., m }; m is the number of event categories; Step 3: Set the threshold to adjust the probability distribution of the output in step 2 p i Threshold truncation is performed to remove redundant auxiliary information of event category attributes and obtain the probability value after threshold truncation. : in, θ To set the threshold; f 2 is the function mapping relationship corresponding to the threshold cutoff; P ij Indicates i Channel data belongs to the category j The probability of Step 4: After the threshold is cut off in step 3, n channels, m The probability value of the number of event categories Concatenate the vectors one by one to get a vector with a length of n × m A one-dimensional vector V: Step 5: Use the event category annotation in step 1 as the label of the one-dimensional vector V y s , and the 1-dimensional vector V and its label y s Input to the microseismic event classification model and output the prediction results of microseismic event categories : in, f 3 It is the function mapping relationship corresponding to the microseismic event classification model.
2. A microseismic monitoring event classification method according to claim 1, characterized in that: The historical data of the multiple groups of channels are in the form of sgy files or ndarray arrays.
3. A microseismic monitoring event classification method according to claim 1, characterized in that: The single-channel classification model outputs i The probability distribution corresponding to the channel p i Event categories j The sum of the probabilities is 1, that is .
4. A microseismic monitoring event classification method according to claim 3, characterized in that: The event categories include microseismic events, rock burst events, blasting events, knocking events, excavation events, drilling events, light source vibration events, oil hammer events and background noise events.
5. A microseismic monitoring event classification method according to claim 1, characterized in that: The set threshold θ The value range is (0, 1], the threshold value is set θ is 0.
3.
6. A microseismic monitoring event classification method according to claim 1, characterized in that: The single-channel classification model uses the AST audio event classification model as a pre-training model.
7. A microseismic monitoring event classification method according to claim 1, characterized in that: Function mapping relationship corresponding to the single-channel classification model f 1 Including logits output and activation function.
8. A microseismic monitoring event classification method according to claim 1, characterized in that: The activation function includes a softmax function or a sigmoid function.
9. A microseismic monitoring event classification method according to claim 1, characterized in that: The microseismic event classification model adopts a hard voting classifier composed of five base classifiers based on a machine learning model.
10. A microseismic monitoring event classification method according to claim 9, characterized in that: The base classifiers include support vector machine SVW, random forest RF, lightweight gradient boosting machine LightGBM, extreme gradient boosting XGBoost and classification boosting CatBoost.