CN-120742406-B - Mine earthquake intelligent identification method based on multi-mode deep learning and signal processing

CN120742406BCN 120742406 BCN120742406 BCN 120742406BCN-120742406-B

Abstract

The invention discloses an intelligent mining earthquake recognition method based on multi-mode deep learning and signal processing, and belongs to the technical field of mining earthquake recognition. Aiming at the problems that the traditional mining earthquake signal processing method cannot fully utilize the spatial relevance and frequency domain characteristics of signals in a sensor network, has low automation level, causes slow response speed and influences emergency processing effects, the method is used for constructing a dynamic graph recognition model of a fusion Graph Neural Network (GNN) and a transducer based on text semantic characteristics generated by a Large Language Model (LLMs), frequency domain characteristics extracted by Fourier transform and time sequence characteristics processed by obspy through inputting earthquake waveform data and multi-mode characteristics, dynamically updating graph structure edge weights through cross-correlation among signals, and adapting to structural changes of complex environments such as mine roadways.

Inventors

HAN ZHENHUA
LI BIN
LIU JIA
XIA SHUQI
ZHAO CAI
ZHANG XIAOYI
CAO JIARUI
Han Yangzhi

Assignees

太原理工大学

Dates

Publication Date: 20260505
Application Date: 20250715

Claims (5)

1. An intelligent mining earthquake recognition method based on multi-mode deep learning and signal processing is characterized by comprising the following steps: Step 1, deploying a plurality of vibration sensors in a mine environment to form a mine sensor network, and acquiring seismic waveform data in real time through the mine sensor network, wherein the seismic waveform data comprises acceleration signals of P waves and S waves in the X, Y, Z axial direction, and synchronizing time stamp information; Step 2, performing noise suppression processing on the acquired seismic waveform data, reading original acceleration waveform data by using a seismology open source tool package Obspy, calling a BandpassFilter module in Obspy, performing band-pass filtering processing on the data signals, and filtering out low-frequency topographic drift and high-frequency mechanical noise; Step 3, sectioning the continuous waveform data signal according to a fixed time window, marking the sectioning data by combining a historical event record or a manual marking mode, and constructing a supervised learning data set; Step 4, calling the STA/LTA algorithm in Obspy to rapidly scan the processed waveform data, and detecting candidate abnormal vibration events through analysis of the ratio of short-time average energy to long-time average energy to form an event candidate set; Step 5, constructing a graph structure for the seismic waveform data based on a graph neural network, extracting frequency domain statistical features and text feature representations of the seismic waveform data in the event candidate set, fusing the frequency domain statistical features and the text feature representations, and inputting the fused multi-modal feature representations into the graph structure; Step 5.1, taking each sensor as a graph node to construct a graph structure According to the physical distance between vibration sensors and the cross correlation between corresponding signals, the weight of the edge is calculated, the construction of a self-adaptive graph structure is realized, and the formula is as follows: , wherein, Representing edge weights between nodes i and j; the cross-correlation is represented by a correlation, As the distance between the nodes of the network, Is a weighting parameter; A feature vector representing node i; A feature vector representing node j; and 5.2, carrying out short-time Fourier transform on the seismic waveform data of each event candidate set to generate a corresponding time-frequency spectrogram, wherein the short-time Fourier transform formula is as follows: , wherein, Representing the spectral intensities at time t and frequency f; representing a time signal; A window function representing the center at time t; representing complex exponential bases; Representing the frequency; representing an integral variable; Step 5.3, extracting frequency domain statistical characteristics based on the time-frequency spectrogram Mean frequency The calculation formula of (2) is as follows: ; Dominant frequency The calculation formula of (2) is as follows: ; step 5.4, acquiring metadata information related to the vibration sensor, inputting a large language model, and generating semantic description of the current signal environment; Step 5.5, inputting the semantic description text into an embedded layer of the large language model to carry out vectorization coding so as to generate the semantic description text By embedding the model Obtaining 128-dimensional text feature representations The formula is: ; Step 5.6, frequency domain statistics feature With text feature representation Splicing to form a complete multi-modal characterization of each shock sensor node And inputting a graph structure, wherein the multi-modal characteristic expression formula is as follows: , wherein, The frequency domain statistical characteristic of the ith vibration sensor node is a characteristic representation extracted from frequency domain analysis of vibration signals and used for representing the vibration of the node in the aspect of frequency correlation characteristics; a text feature representation representing an ith shock sensor node; representing post-splice multimodal features Is a dimension of (2); Step 6, constructing a time-space joint feature matrix, wherein the time sequence embedded sequence of each vibration sensor captures long-time dependence by using a multi-head attention mechanism through 4 layers of convectors, and injecting time stamp information into position codes so as to obtain the time-space joint feature matrix; and 7, event classification and report generation, wherein the space-time joint feature matrix is pooled through a full connection layer, then a classifier is input, and a classification result and key features are input LLMs to generate a structured report.
2. The intelligent mining vibration identification method based on multi-mode deep learning and signal processing according to claim 1, wherein the raw acceleration waveform data signals acquired in the step 2 are as follows Band pass filtering The formula of the treatment is as follows: where h (t) represents the pooled feature vector.
3. The intelligent mining earthquake recognition method based on multi-mode deep learning and signal processing according to claim 2, wherein the short-time average in the step 4 is that Average over long periods Energy ratio The formula of (2) is: , , , wherein, Representation of The short-time energy of the moment is calculated, the signal energy in a short window near the current moment is calculated, and the instantaneous change of the signal is captured; Representation of The long-time energy of the moment, the signal energy in a long window near the current moment is calculated, and the average energy level of the background noise is reflected; Representing a short window length; Indicating the long window length.
4. The intelligent mining earthquake identification method based on multi-mode deep learning and signal processing according to claim 3, wherein the specific operation of the step 6 is as follows: And 6.1, forming a time sequence input sequence by representing the characteristics of each vibration sensor node in a continuous time window, wherein the formula is as follows: , wherein, Representing a time sequence feature sequence of node i; indicating that sensor node i is at time Features at the location; Representing a time window length; And 6.2, processing the sequence input by using a 4-layer transducer encoder, and adding position codes into each layer of transducer encoder, wherein the formula is as follows: , wherein, Representing the encoded features; representing a position code; the update process of each layer of the transducer encoder is as follows: , wherein, Representing the layer structure in the layer I transform encoder; Embedding a time stamp accurate to millisecond into a position code, participating in attention weight calculation together with input, and finally outputting a space-time joint feature matrix to provide time sequence feature support for subsequent event classification and report generation, wherein the formula of the output space-time joint feature matrix is as follows: , wherein, Representing a space-time joint feature matrix corresponding to the ith vibration sensor node; representing the final feature representation after processing by the L-layer transducer encoder.
5. The intelligent mining earthquake identification method based on multi-mode deep learning and signal processing according to claim 4, wherein the specific operation of the step 7 is as follows: the spatio-temporal joint feature matrix is pooled and then input into a classifier: , , wherein, Representing time-space joint feature matrix The pooling operation is carried out to obtain pooling feature vectors, The representation represents a pooling operation, Representing the predicted probability distribution of the classifier output, Representing the Softmax activation function, A weight matrix representing the linear transformation in the classifier, Bias terms representing linear transformations in the classifier; The classification result and the characteristics are spliced and then input : , wherein, Representing the predicted probability distribution output by the previous classifier; a complete multi-modal representation of the ith shock sensor node; i.e. the previously pooled feature vectors.

Description

Mine earthquake intelligent identification method based on multi-mode deep learning and signal processing Technical Field The invention belongs to the technical field of ore shock identification, and particularly relates to an ore shock intelligent identification method based on multi-mode deep learning and signal processing. Background Mine safety monitoring is an important link for guaranteeing mine production safety and personnel life safety. Along with the expansion of mining scale and the improvement of monitoring requirements, real-time detection and accurate identification of mine earthquake signals become important points of research. The traditional mine earthquake detection method mainly depends on threshold triggering or single-mode signal analysis, the mode is easy to be interfered by environmental noise, the false alarm rate is high, complex noise and real vibration events are difficult to effectively distinguish, and the accuracy and the reliability of monitoring are affected. In recent years, deep learning technology has been widely focused in the field of mine earthquake signal processing, but the existing schemes focus on a single time sequence model, such as a Convolutional Neural Network (CNN) or a long-short-term memory network (LSTM), so that the spatial correlation and frequency domain characteristics of signals in a sensor network cannot be fully utilized, and the performance improvement of a detection model is limited. In addition, interpretation and report generation of the mine earthquake event often depend on manual completion, and the automation level is low, so that the response speed is slow, and the emergency treatment effect is affected. Disclosure of Invention Aiming at the problems that the traditional ore-mining-vibration signal processing method cannot fully utilize the spatial relevance and frequency domain characteristics of signals in a sensor network, has low automation level, causes slow response speed and influences emergency processing effects, the invention provides an ore-mining-vibration intelligent recognition method based on multi-mode deep learning and signal processing, which is an ore-mining-vibration signal intelligent recognition method integrating a Large Language Model (LLMs), fourier transformation, a seismic data processing tool (obspy), a Graph Neural Network (GNN) and a transform technology. The method can realize the fusion processing of multi-mode data, combines the space-time characteristics and the frequency domain characteristics of signals, realizes the high-precision real-time detection, classification and early warning of the mine earthquake event, and effectively improves the detection accuracy and response efficiency of monitoring. In order to achieve the above purpose, the present invention adopts the following technical scheme: an intelligent mining earthquake recognition method based on multi-mode deep learning and signal processing comprises the following steps: Step 1, deploying a plurality of vibration sensors in a mine environment to form a mine sensor network, and acquiring seismic waveform data in real time through the mine sensor network, wherein the seismic waveform data comprises acceleration signals of P waves and S waves in the X, Y, Z axial direction, and synchronizing time stamp information; Step 2, performing noise suppression processing on the acquired seismic waveform data, reading original acceleration waveform data by using a seismology open source tool package Obspy, calling a BandpassFilter module in Obspy, performing band-pass filtering processing on the data signals, and filtering out low-frequency topographic drift and high-frequency mechanical noise; The original acceleration waveform data signals acquired in the step2 are Band pass filteringThe formula of the treatment is as follows: , wherein, Representing the pooled feature vectors. Step 3, sectioning the continuous waveform data signal according to a fixed time window, marking the sectioning data by combining a historical event record or a manual marking mode, and constructing a supervised learning data set; Step 4, calling the STA/LTA algorithm in Obspy to rapidly scan the processed waveform data, and detecting candidate abnormal vibration events through analysis of the ratio of short-time average energy to long-time average energy to form an event candidate set; Short time averaging in said step 4 Average over long periodsThe formula of the energy ratio is:,, , wherein, Representation ofThe short-time energy of the moment is calculated, the signal energy in a short window near the current moment is calculated, and the instantaneous change of the signal is captured; Representation of The long-time energy of the moment, the signal energy in a long window near the current moment is calculated, and the average energy level of the background noise is reflected; Representing a short window length; Indicating the long window length. Step 5, constructing a graph s