Deep learning based 3d microseismic location method for deep concave empty area of open pit mine
By constructing a dual-channel convolutional neural network with a time difference matrix and a mask channel using deep learning methods, the problems of uncertainty in the propagation path of microseismic waves and signal loss in the complex environment of open-pit mines were solved. This enabled high-precision, low-latency three-dimensional microseismic positioning, improving the reliability and real-time performance of safe production in open-pit mines.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHEASTERN UNIV CHINA
- Filing Date
- 2026-03-10
- Publication Date
- 2026-06-05
AI Technical Summary
Traditional microseismic positioning technology struggles to maintain sufficient accuracy in the complex environment of open-pit mines, especially in open areas and under complex slope geometry, where the uncertainty of microseismic wave propagation paths and signal loss are difficult to resolve.
A deep learning-based approach is adopted to collect waveform data through multi-channel sensors, construct a time difference matrix and mask channels, train a dual-channel convolutional neural network, and combine fusion dilated convolution, residual connection and lightweight multi-head self-attention mechanism to achieve three-dimensional localization of the seismic source.
It significantly improves the adaptability and reliability of positioning in open-pit environments, and can maintain accurate error positioning when sensors are offline or data is missing. It has high real-time performance and low latency, meeting the accuracy requirements for safe production in open-pit mines.
Smart Images
Figure CN122151180A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of open-pit mine microseismic positioning technology, specifically a deep learning-based three-dimensional microseismic positioning method for deep hollow areas in open-pit mines. Background Technology
[0002] Microseismic monitoring has become one of the core technologies for safe production in open-pit mines. It can accurately capture weak vibration signals caused by the continuous expansion of micro-fractures within the rock mass, enabling early warning of typical disasters such as slope instability, stope collapse, and mine tremors. However, the complex slope geometry and working environment of open-pit mines bring many challenges to the application of microseismic monitoring technology. The irregular shape of stope voids and slopes causes significant detours of microseismic waves, making it difficult to accurately model the actual propagation path and velocity of microseismic waves. Traditional positioning techniques often fail to maintain sufficient accuracy in such complex areas, making it difficult to meet the stringent accuracy requirements for microseismic monitoring in open-pit mine safety production. To address this challenge, machine learning-based methods no longer rely on precise propagation models, providing a new approach to source location in voids or complex slope geometries. Furthermore, due to the potentially long distance between the source and the sensor, energy attenuates significantly along the propagation path, weakening the amplitude of the microseismic signal. This signal is easily drowned out by various background noises in the mining environment, and some sensors fail to trigger effective identification, resulting in widespread data loss, which traditional methods struggle to resolve.
[0003] Therefore, there is an urgent need to propose a positioning method that can effectively overcome the uncertainty of propagation path caused by the complex geometry of empty areas and slopes, and can cope with the lack of microseismic signals, so as to solve the problem of insufficient accuracy of traditional wave velocity models in the complex environment of open-pit mines. Summary of the Invention
[0004] To address the shortcomings of existing technologies, the present invention aims to propose a deep learning-based microseismic three-dimensional localization method for deep hollow areas in open-pit mines, comprising:
[0005] Determine the locations of multiple candidate seismic sources in the deep hollow area of an open-pit mine;
[0006] Multiple geophones are placed in the deep recessed area of the open-pit mine. Based on the candidate source locations and the microseismic waveforms collected by the geophones, a time difference matrix and a mask channel are constructed, and then a training sample set is built.
[0007] The dual-channel convolutional neural network is trained based on the training sample set to obtain the trained dual-channel convolutional neural network.
[0008] The microseismic waveforms acquired by the detector are obtained in real time. A time difference matrix and a mask channel are constructed. The time difference matrix and the mask channel are input into a trained dual-channel convolutional neural network to obtain the inverted three-dimensional coordinates of the seismic source.
[0009] Optionally, determine the locations of multiple candidate seismic sources in the deep hollow area of the open-pit mine, including:
[0010] The deep recessed area of the open-pit mine is divided into multiple cubic blocks according to a regular grid. The cubic blocks located on the slope of the deep recessed area of the open-pit mine are fine-tuned to obtain fine-tuned three-dimensional blocks. The normalized real coordinates of each fine-tuned three-dimensional block and each cubic block in three-dimensional space are used as candidate earthquake source locations to obtain multiple candidate earthquake source locations.
[0011] Optionally, the cubic blocks at the slope location can be fine-tuned to obtain fine-tuned three-dimensional blocks, including:
[0012] According to the boundary angle of the slope, the slope is divided into multiple line segments, and the endpoints of each line segment are connected to obtain a finely adjusted three-dimensional block.
[0013] Optionally, based on the candidate source locations and the microseismic waveforms acquired by the detectors, a time difference matrix and a mask channel are constructed, thereby building a training sample set, including:
[0014] For each candidate seismic source location, an earthquake is initiated at the candidate seismic source location. The time it takes for the microseismic waveform to reach each detector is calculated based on the wave velocity, and a time difference matrix is constructed. Each row and column of the time difference matrix corresponds to a detector. Each element in the time difference matrix represents the time difference value between the two detectors corresponding to that element.
[0015] Randomly mark multiple detectors as having not acquired microseismic waveforms to construct mask channels; repeat the above operation to generate multiple mask channels;
[0016] The mask channel is a mask matrix, with each row and column corresponding to a detector; when all detectors in the row and column acquire micro-vibration waveforms, the value of the corresponding element in the mask matrix is 1; when at least one detector in the row and column fails to acquire a micro-vibration waveform, the value of the corresponding element is 0.
[0017] The constructed time difference matrix is copied to obtain multiple time difference matrices; one time difference matrix and one mask channel constitute the input sample, the candidate source location is used as the output sample, the input sample and the output sample constitute the training sample, and multiple training samples constitute the training sample set.
[0018] Optionally, the dual-channel convolutional neural network is trained based on the training sample set to obtain a trained dual-channel convolutional neural network, including:
[0019] Input samples from the training samples are fed into a dual-channel convolutional neural network to obtain the inverted three-dimensional coordinates of the seismic source.
[0020] Based on the inverted three-dimensional coordinates of the seismic source and the corresponding output samples in the training samples, the spatially weighted loss function value is calculated. ;
[0021] Based on loss function value The parameters in the dual-channel convolutional neural network are updated to obtain the trained dual-channel convolutional neural network.
[0022] Optionally, the input samples from the training samples are input into a dual-channel convolutional neural network to obtain the inverted three-dimensional coordinates of the seismic source, including:
[0023] The time difference matrix and mask channels in the input samples are used to form a feature tensor. The feature tensor is then subjected to three sets of dilation convolutions with different dilation rates to obtain the first feature, the second feature, and the third feature, respectively. The first feature, the second feature, and the third feature are then fused to obtain the fused feature. The fused feature and the feature tensor are then added element by element to obtain the fourth feature. The fourth feature is then processed through a multi-head self-attention mechanism and passed through a fully connected layer to obtain the three-dimensional coordinates of the inverted source.
[0024] Optionally, based on the inverted source's three-dimensional coordinates and the corresponding output samples from the training samples, a spatially weighted loss function value is calculated. Specifically, this is achieved through the following formula:
[0025] ;
[0026] in, This represents the number of training samples in a single training batch, which is divided from the training sample set. This represents the normalized 3D coordinates of the inverted source corresponding to the i-th training sample. Indicates the location of the candidate earthquake source. express The corresponding coordinates at the real physical scale, express The corresponding coordinates at the real physical scale; This represents the Smooth L1 loss function, where d represents the spatial coordinate dimension. This represents the normalized inverse coordinate of the i-th training sample in the d-th spatial coordinate dimension. This represents the normalized true coordinates of the i-th training sample in the d-th spatial coordinate dimension.
[0027] The beneficial effects of adopting the above technical solution are as follows:
[0028] This invention, based on waveform data acquired by multi-channel sensors, constructs a time difference matrix by calculating the arrival time difference between any two channels. This enables accurate regression of the seismic source location and initiation time without relying on a complete velocity field, significantly improving the adaptability and reliability of positioning in empty environments. Furthermore, addressing the common problem in mining areas where some sensors are far away and difficult to capture microseismic signals, this invention introduces a binary virtual masking mechanism. Missing or low signal-to-noise ratio channels (those that did not acquire microseismic waveforms) are explicitly marked. The time difference matrix and the masked channels are input together into a dual-channel convolutional neural network that integrates dilated convolution, residual connections, and a lightweight multi-head self-attention mechanism, enhancing the model's ability to perceive missing information. Through a dynamic missing data simulation training strategy, the model is continuously exposed to different missing data modes during training, thus possessing stronger generalization ability and robustness. In contrast, while traditional machine learning methods can address velocity model uncertainty to some extent, they often struggle to effectively handle incomplete channel data and are prone to performance degradation in missing environments. It achieves accurate error positioning even when sensors are largely offline, and the single-frame inference time after edge deployment is extremely short, which can meet the second-level early warning requirement; thus, it is superior to existing microseismic positioning methods in terms of reliability, real-time performance, cost control and system integration convenience. Attached Figure Description
[0029] Figure 1 This is one of the flowcharts illustrating the deep learning-based microseismic three-dimensional localization method for deep void areas in open-pit mines according to an embodiment of the present invention.
[0030] Figure 2 This is the second flowchart illustrating the deep learning-based microseismic three-dimensional localization method for deep hollow areas in open-pit mines, as described in this embodiment of the invention.
[0031] Figure 3 This is a schematic diagram of the slope location in an embodiment of the present invention;
[0032] Figure 4 This is a micro-vibration waveform diagram from an embodiment of the present invention;
[0033] Figure 5 This is a schematic diagram of the model lightweighting process in an embodiment of the present invention;
[0034] Figure 6 This is a schematic diagram of the inverted three-dimensional coordinates of the seismic source in an embodiment of the present invention. Detailed Implementation
[0035] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.
[0036] To address the problems of existing technologies, this invention provides a deep learning-based three-dimensional microseismic localization method for deep hollow areas in open-pit mines. This method does not rely on a complete velocity field and can directly complete the three-dimensional inversion of the seismic source based on observation data. Compared with traditional methods using manual data acquisition, it has advantages such as high real-time performance, stable positioning accuracy, and low maintenance workload. Even in conditions with low sensor sensitivity and numerous hollow areas and complex slopes, it can still continuously and quickly output reliable results, thereby comprehensively improving the universality and reliability of open-pit mine microseismic early warning systems. Specifically, this method eliminates the dependence on a complete velocity model. It constructs a feature matrix by calculating the arrival time difference between sensor pairs and uses a dual-channel convolutional neural network that integrates dilated convolution, residual connections, and a lightweight multi-head self-attention mechanism to achieve end-to-end regression of the three-dimensional coordinates of the seismic source and the initiation time. Furthermore, to address situations such as offline sensors or missing data, this invention further introduces a binary masking mechanism to explicitly label missing channels and combines this with a dynamic missing data simulation training strategy to enhance the robustness of the model in scenarios with incomplete channels. During the training phase of the localization model, the monitoring area is divided into regular cubic blocks, and the boundary blocks near the open-pit slope are further subdivided into quasi-trapezoidal units according to the slope angle. Simultaneously, geometric perturbations are applied to each training scenario to generate a source dataset covering diverse scenarios. This method possesses comprehensive advantages such as high accuracy, low latency, and low cost, providing strong technical support for safe production in the complex environment of open-pit mines.
[0037] Based on this, the present invention provides a deep learning-based microseismic three-dimensional localization method for deep hollow areas in open-pit mines, combined with... Figure 1 and Figure 2 This may include the following steps:
[0038] Step 1: Determine the locations of multiple candidate seismic sources in the deep hollow area of the open-pit mine;
[0039] The deep recessed area of the open-pit mine is divided into multiple cubic blocks according to a regular grid. The cubic blocks located on the slope of the deep recessed area of the open-pit mine are fine-tuned to obtain fine-tuned three-dimensional blocks. The normalized real coordinates of each fine-tuned three-dimensional block and each cubic block in three-dimensional space are used as candidate earthquake source locations to obtain multiple candidate earthquake source locations.
[0040] In this process, the cubic blocks at the slope location are fine-tuned to obtain the finely-tuned three-dimensional blocks, combined with... Figure 3 ,include:
[0041] Based on the slope's boundary angles, the slope is divided into multiple line segments. The endpoints of these segments are connected, and the cubic blocks are further subdivided into quasi-trapezoidal or irregular polyhedral units, combined with boundary weighting. Simultaneously, random perturbations are applied to the slope angle and step height to obtain fine-tuned 3D blocks. This fine-tuning makes the block geometry closer to the real slope and void structure, thus obtaining a virtual seismic source dataset with higher consistency with the actual environment.
[0042] Step 2: Place multiple geophones in the deep recessed area of the open-pit mine. Based on the candidate source location and the microseismic waveforms collected by the geophones, construct a time difference matrix and a mask channel, and then construct a training sample set.
[0043] Multiple geophones are placed in the deep recessed area of an open-pit mine. For each candidate seismic source location, an earthquake is initiated at the candidate seismic source location. The time it takes for the microseismic waveform to reach each geophone is calculated based on the wave velocity, and a time difference matrix is constructed. Each row and column of the time difference matrix corresponds to a geophone. Each element in the time difference matrix represents the time difference value between the two geophones corresponding to that element.
[0044] Each detector maintains millisecond-level synchronization with the server and samples micro-vibration waveform signals in real time, such as... Figure 4 In the specific implementation process, after each complete waveform is acquired, the arrival time difference between any two sensors is immediately calculated to construct a time difference matrix.
[0045] Randomly mark multiple detectors as having not acquired microseismic waveforms to construct mask channels; repeat the above operation to generate multiple mask channels;
[0046] The mask channel is a mask matrix, with each row and column corresponding to a detector; when all detectors in the row and column acquire micro-vibration waveforms, the value of the corresponding element in the mask matrix is 1; when at least one detector in the row and column fails to acquire a micro-vibration waveform, the value of the corresponding element is 0.
[0047] The constructed time difference matrix is copied to obtain multiple time difference matrices; one time difference matrix and one mask channel constitute the input sample, the candidate source location is used as the output sample, the input sample and the output sample constitute the training sample, and multiple training samples constitute the training sample set.
[0048] In each training round, multiple sets of input samples are copied at the beginning, and several detector channels are randomly masked. New masks are generated synchronously and discarded after training ends, then regenerated in the next round to ensure that sample combinations are not repeated and cover different missing data conditions. The validation set is only randomly masked once during the initial construction and remains fixed to evaluate model performance. Meanwhile, a fixed validation set is constructed with three times the size of the original dataset, and channels are randomly masked once and remain unchanged throughout the training process to periodically evaluate model performance and avoid evaluation bias. Early termination is triggered when the validation error does not decrease significantly within a consecutive preset period, thus obtaining a lightweight localization model that maintains high accuracy under different missing data ratios.
[0049] Step 3: Train the dual-channel convolutional neural network based on the training sample set to obtain the trained dual-channel convolutional neural network;
[0050] Step 3.1: Input the input samples from the training samples into a dual-channel convolutional neural network to obtain the inverted three-dimensional coordinates of the seismic source;
[0051] The time difference matrix and mask channels in the input samples are used to form a feature tensor. The feature tensor is then subjected to three sets of dilation convolutions with different dilation rates to obtain the first feature, the second feature, and the third feature, respectively. The first feature, the second feature, and the third feature are then fused to obtain the fused feature. The fused feature and the feature tensor are then added element by element to obtain the fourth feature. The fourth feature is then processed through a multi-head self-attention mechanism and passed through a fully connected layer to obtain the inverted three-dimensional coordinates of the seismic source.
[0052] In the specific implementation process, the three dilation rates are 1, 2 and 4 respectively, and the dilation convolution kernel is 3×1, so as to expand the effective receptive field without increasing the number of parameters. In addition to capturing local phase differences, it also integrates cross-channel long-scale time difference models to provide sufficient contextual information for microseismic location.
[0053] Introducing a lightweight multi-head self-attention mechanism can dynamically enhance effective observations and reduce the interference of missing channels on overall characteristics, thereby maintaining the accurate expression of long-range correlation across detectors.
[0054] This invention proposes a loss function that assigns a large weight to 3D coordinate errors to suppress gradient fluctuations caused by missing samples. Specifically, during the training phase, a spatially weighted loss function is used, assigning a baseline weight to 3D coordinate errors and a lower weight to the earthquake initiation time error. The overall loss coefficient is adaptively adjusted according to the proportion of missing channels in the batch to suppress the adverse effects of missing data on the gradient. The specific location information of microseisms is calculated based on known arrival time difference data. Model training is then performed, using the collected microseismic arrival time difference data to train an initial network model in *.pth format.
[0055] Step 3.2: Based on the inverted source 3D coordinates and the corresponding output samples in the training samples, calculate the spatially weighted loss function value. Specifically, this is achieved through the following formula:
[0056] ;
[0057] in, This represents the number of training samples in a single training batch, which is divided from the training sample set. This represents the normalized 3D coordinates of the inverted source corresponding to the i-th training sample. Indicates the location of the candidate earthquake source. express The corresponding coordinates at the real physical scale, express The corresponding coordinates at the real physical scale; This represents the Smooth L1 loss function, where d represents the spatial coordinate dimension. This represents the normalized inverse coordinate of the i-th training sample in the d-th spatial coordinate dimension. This represents the normalized true coordinate of the i-th training sample in the d-th spatial coordinate dimension;
[0058] Step 3.3: Based on the loss function value The parameters in the dual-channel convolutional neural network are updated to obtain the trained dual-channel convolutional neural network.
[0059] like Figure 5 As shown, to enhance the model's portability and inference speed, this invention first uses the PyTorch export interface to convert the *.pt weights obtained from training the self-developed network into *.onnx format. Then, it uses OpenVINO to generate an intermediate representation (IR) for various hardware, resulting in a file set including a weight file (.bin, storing weights and biases), a structure file (.xml, describing the network topology), and a configuration file (.json, recording data interaction information). The IR model acts as a unified bridge, enabling seamless migration between different runtime environments and allowing for on-demand quantization to FP16 or even FP8 versions to reduce storage bandwidth and improve inference speed, thus maintaining high performance and flexible configuration when deployed on edge devices.
[0060] Step 4: Acquire the microseismic waveforms from the detector, construct the time difference matrix and mask channel, and input the time difference matrix and mask channel into the trained dual-channel convolutional neural network to obtain the inverted three-dimensional coordinates of the seismic source, such as... Figure 6 .
[0061] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.
Claims
1. A deep learning-based microseismic three-dimensional localization method for deep hollow areas in open-pit mines, characterized in that, include: Determine the locations of multiple candidate seismic sources in the deep hollow area of an open-pit mine; Multiple geophones are placed in the deep recessed area of the open-pit mine. Based on the candidate source locations and the microseismic waveforms collected by the geophones, a time difference matrix and a mask channel are constructed, and then a training sample set is built. The dual-channel convolutional neural network is trained based on the training sample set to obtain the trained dual-channel convolutional neural network. The microseismic waveforms acquired by the detector are obtained in real time. A time difference matrix and a mask channel are constructed. The time difference matrix and the mask channel are input into a trained dual-channel convolutional neural network to obtain the inverted three-dimensional coordinates of the seismic source.
2. The method for three-dimensional microseismic localization of deep hollow areas in open-pit mines based on deep learning according to claim 1, characterized in that, The locations of multiple candidate seismic sources in the deep hollow area of an open-pit mine were determined, including: The deep recessed area of the open-pit mine is divided into multiple cubic blocks according to a regular grid. The cubic blocks located on the slope of the deep recessed area of the open-pit mine are fine-tuned to obtain fine-tuned three-dimensional blocks. The normalized real coordinates of each fine-tuned three-dimensional block and each cubic block in three-dimensional space are used as candidate earthquake source locations to obtain multiple candidate earthquake source locations.
3. The method for three-dimensional microseismic localization of deep hollow areas in open-pit mines based on deep learning according to claim 2, characterized in that, Fine-tuning the cubic blocks at the slope location yields the fine-tuned three-dimensional blocks, including: According to the boundary angle of the slope, the slope is divided into multiple line segments, and the endpoints of each line segment are connected to obtain a finely adjusted three-dimensional block.
4. The method for three-dimensional microseismic localization of deep hollow areas in open-pit mines based on deep learning according to claim 1, characterized in that, Based on the candidate source locations and microseismic waveforms acquired by the detectors, a time difference matrix and mask channels are constructed, and then a training sample set is built, including: For each candidate seismic source location, an earthquake is initiated at the candidate seismic source location. The time it takes for the microseismic waveform to reach each detector is calculated based on the wave velocity, and a time difference matrix is constructed. Each row and column of the time difference matrix corresponds to a detector. Each element in the time difference matrix represents the time difference value between the two detectors corresponding to that element. Randomly mark multiple detectors as having not acquired microseismic waveforms to construct mask channels; repeat the above operation to generate multiple mask channels; The mask channel is a mask matrix, with each row and column corresponding to a detector; when all detectors in the row and column acquire micro-vibration waveforms, the value of the corresponding element in the mask matrix is 1; when at least one detector in the row and column fails to acquire a micro-vibration waveform, the value of the corresponding element is 0. The constructed time difference matrix is copied to obtain multiple time difference matrices; one time difference matrix and one mask channel constitute the input sample, the candidate source location is used as the output sample, the input sample and the output sample constitute the training sample, and multiple training samples constitute the training sample set.
5. The method for three-dimensional microseismic localization of deep hollow areas in open-pit mines based on deep learning according to claim 4, characterized in that, The dual-channel convolutional neural network is trained based on the training sample set to obtain the trained dual-channel convolutional neural network, including: Input samples from the training samples are fed into a dual-channel convolutional neural network to obtain the inverted three-dimensional coordinates of the seismic source. Based on the inverted three-dimensional coordinates of the seismic source and the corresponding output samples in the training samples, the spatially weighted loss function value is calculated. ; Based on loss function value The parameters in the dual-channel convolutional neural network are updated to obtain the trained dual-channel convolutional neural network.
6. The method for three-dimensional microseismic localization of deep hollow areas in open-pit mines based on deep learning according to claim 5, characterized in that, The input samples from the training samples are fed into a dual-channel convolutional neural network to obtain the inverted three-dimensional coordinates of the seismic source, including: The time difference matrix and mask channels in the input samples are used to form a feature tensor. The feature tensor is then subjected to three sets of dilation convolutions with different dilation rates to obtain the first feature, the second feature, and the third feature, respectively. The first feature, the second feature, and the third feature are then fused to obtain the fused feature. The fused feature and the feature tensor are then added element by element to obtain the fourth feature. The fourth feature is then processed through a multi-head self-attention mechanism and passed through a fully connected layer to obtain the inverted three-dimensional coordinates of the seismic source.
7. The method for three-dimensional microseismic localization of deep hollow areas in open-pit mines based on deep learning according to claim 5, characterized in that, Based on the inverted three-dimensional coordinates of the seismic source and the corresponding output samples in the training samples, the spatially weighted loss function value is calculated. Specifically, this is achieved through the following formula: ； in, This represents the number of training samples in a single training batch, which is divided from the training sample set. This represents the normalized 3D coordinates of the inverted source corresponding to the i-th training sample. Indicates the location of the candidate earthquake source. express The corresponding coordinates at the real physical scale, express The corresponding coordinates at the real physical scale; This represents the SmoothL1 loss function, where d represents the spatial coordinate dimension. This represents the normalized inverse coordinate of the i-th training sample in the d-th spatial coordinate dimension. This represents the normalized true coordinates of the i-th training sample in the d-th spatial coordinate dimension.