Microseismic data avalanche automatic detection method and system based on autoencoder

Abstract

The application relates to the field of natural disaster monitoring and early warning and artificial intelligence technology, and discloses a microseismic data avalanche automatic detection method based on a self-encoder, which comprises the following steps: continuously collecting microseismic waveform data through a microseismic sensor array and performing pretreatment; performing event picking and labeling, constructing a labeled data set containing avalanche events and non-avalanche noise events, and dividing a training set and a test set according to a time dimension; constructing a dual-domain coupled avalanche physical constraint self-encoder, obtaining an avalanche exclusive feature extractor through unsupervised training of a composite loss function; constructing a random forest classifier with an avalanche feature saliency weighting mechanism, training and hyperparameter optimizing the random forest classifier, and obtaining a trained avalanche classification model; and outputting an avalanche event detection result through two-stage post-processing of array space-time coherence aggregation and evolution continuity judgment. The application scheme has the advantages of strong anti-interference capability, high detection precision, good generalization and a low false negative rate.

Description

Technical Field

[0001] This invention relates to the fields of natural disaster monitoring and early warning and artificial intelligence technology, specifically to an automatic avalanche detection method and system based on microseismic data using an autoencoder. Background Technology

[0002] Avalanches are a major geological disaster that frequently occurs in high-altitude and cold mountainous areas. They are characterized by their suddenness, destructive power, and short warning window, seriously threatening the lives and property of mountain residents and the operational safety of transportation arteries and major engineering facilities. Achieving all-weather, high-precision, and low-missibility automatic detection of avalanche events is a core prerequisite for avalanche disaster early warning and emergency response.

[0003] Microseismic monitoring technology detects avalanches by collecting microseismic wave signals generated by friction and impact with the ground during avalanche movement. It is unaffected by weather and sunlight and is a promising technology for real-time avalanche monitoring. However, existing avalanche detection technologies based on microseismic data still have some shortcomings: (1) Traditional threshold-based methods rely on manually set fixed thresholds, which are greatly affected by environmental interference such as wind noise, human activities, aircraft passing, and natural microseismic waves, and cannot effectively distinguish avalanche signals from various interference signals; (2) Existing machine learning-based detection methods mostly use general feature engineering or standardized deep learning models, without combining the physical characteristics of avalanche microseismic signals. The features learned by the model have insufficient discrimination and poor generalization ability, and the detection performance decreases in different monitoring areas and different snow conditions; (3) The existing training process does not incorporate avalanche physical constraints, and the model is prone to learning noise features, resulting in insufficient targeting of feature extraction.

[0004] To address these issues, we have developed an automatic avalanche detection method and system based on autoencoders for microseismic data, which solves the aforementioned technical problems. Summary of the Invention

[0005] This invention provides an automatic avalanche detection method and system based on autoencoder microseismic data that combines the physical characteristics of avalanches, has strong anti-interference ability, high detection accuracy, good generalization, and low false negative rate.

[0006] Therefore, the present invention provides the following technical solution: An automatic avalanche detection method based on autoencoder microseismic data, the method comprising: Step 1: Collect continuous micro-vibration waveform data through a micro-vibration sensor array, and preprocess the collected micro-vibration waveform data; Step 2: Perform event picking and annotation on the preprocessed microseismic waveform data to construct an annotated dataset containing avalanche events and non-avalanche noise events, and divide the dataset into training and testing sets according to the time dimension. Step 3: Construct a dual-domain coupled avalanche physics-constrained autoencoder, and use the training set to obtain an avalanche-specific feature extractor through unsupervised training using the avalanche physics-constrained composite loss function, outputting a low-dimensional fused feature vector. Step 4: Construct a random forest classifier with a weighted mechanism for avalanche feature saliency, train and optimize the hyperparameters of the random forest classifier to obtain the trained avalanche classification model. Step 5: Output the avalanche event detection results through two-stage post-processing: array spatiotemporal coherence aggregation and evolution continuity determination.

[0007] Optionally, in step 1, the microseismic sensor array consists of at least 5 single-component or three-component microseismometers with a sampling frequency of not less than 200Hz, a deployment range covering the target avalanche monitoring area, and a monitoring radius of not less than 3km. The preprocessing includes dimensional transformation, linear detrending, bandpass filtering, and avalanche band energy weighting pre-enhancement processing to specifically enhance the signal energy of the core advantageous frequency band of avalanche and suppress non-avalanche interference frequency bands; the bandpass filtering is a 4th order Butterworth bandpass filter with a passband of 1-10Hz. The operator expression for the avalanche frequency band energy weighted pre-enhancement processing is: ; in, This is the time-domain microseismic signal after bandpass filtering. for The frequency domain signal obtained by Fast Fourier Transform For the pre-enhanced time-domain signal, The weighting function for the avalanche-specific frequency band is expressed as follows: .

[0008] Optionally, in step 2, the event picking and labeling specifically involves: combining synchronously acquired radar and camera verification data, determining the start and end times of microseismic events through manual visual picking, and performing binary labeling on the events, classifying them into avalanche events and non-avalanche noise events; among which, non-avalanche noise events include natural microseismic events, aircraft overflight, wind noise, and microseismic signals generated by human activities; simultaneously, dividing the complete avalanche event into sub-stages, labeling the time-series labels of the entire evolution stage of the avalanche event, including initiation, development, peak, and decay, to obtain the labeled dataset; performing sample augmentation on the labeled dataset through differential sliding window slicing, completing sample normalization processing, and then dividing the training set and the test set according to the time dimension to avoid time-series data leakage.

[0009] Optionally, in step 2, the differentiated sliding window slicing is specifically as follows: for labeled avalanche event samples, a sliding window of 10 seconds in length is used, with a window overlap rate of not less than 70%; for non-avalanche noise event samples, a sliding window of 10 seconds in length is used, with a window overlap rate of 50%. The normalization process is as follows: for each signal within a sliding window, the signal is independently normalized according to its own maximum absolute amplitude, and the signal amplitude is mapped to the interval [-1,1]. The training set and the test set are divided according to the time dimension, with the training set accounting for 70% and the test set accounting for 30%, and the model hyperparameters are optimized by three-fold cross-validation.

[0010] Optionally, in step 3, the dual-domain coupled avalanche physics-constrained autoencoder includes: The temporal coding branch consists of three one-dimensional convolutional layers and one bidirectional LSTM layer stacked sequentially. It is used to extract the temporal evolution features of the avalanche signal and output a 32-dimensional temporal feature vector. The convolutional kernel size is 20, the stride is 10, the number of convolutional kernels in the first layer is 32, and the number of convolutional kernels in each subsequent layer doubles. The spectral domain coding branch consists of three fully connected layers and one one-dimensional convolutional layer. The input is the amplitude spectrum obtained by fast Fourier transform of microseismic waveform data, which is used to extract the spectral features and frequency shift law of avalanche signals and outputs a 16-dimensional spectral domain feature vector. The coupled feature fusion layer dynamically allocates attention weights for temporal and spectral features through feature saliency weights, fusing temporal and spectral feature vectors into a 48-dimensional low-dimensional fusion feature vector. The dual-domain reconstruction decoder consists of a time-domain reconstruction branch and a spectral-domain reconstruction branch. It reconstructs the input time-domain signal and spectral signal respectively, and outputs a reconstructed time-domain signal and a reconstructed spectrum with the same dimensions as the input.

[0011] Optionally, in step 3, the dual-domain coupled avalanche physics-constrained autoencoder is trained unsupervised using the training set. During training, the avalanche physics-constrained composite loss function is used for iterative optimization. After training, the encoder part is retained as the avalanche-specific feature extractor. The expression of the avalanche physics-constrained composite loss function is: ; in, Preset weighting coefficients; The loss for bi-domain reconstruction is expressed as: ; In the formula, This is the time-domain microseismic signal after bandpass filtering. For the pre-enhanced time-domain signal, The amplitude spectrum of the input signal. To reconstruct the amplitude spectrum, MSE is the mean square error; The avalanche time-series evolution constraint loss is used to constrain the model to learn the energy evolution law specific to avalanches, and its expression is: ; In the formula, The energy envelope curve is obtained by applying the Hilbert transform and moving average to the signal. The Pearson correlation coefficient; The spatial coherence constraint loss for the microsensor array is used to constrain the model's learning of the specific spatial coherence characteristics of the avalanche extension surface source. Its expression is: ; In the formula, Let be the avalanche spatial coherence coefficient of the input signals from two sensors in the same window of the microsensor array. The reconstructed avalanche spatial coherence coefficient is calculated from the reconstructed time-domain signal output by the dual-domain coupled avalanche physical constraint autoencoder, corresponding to the time-domain input signals from both sensors within the same window. The expression for the avalanche spatial coherence coefficient is as follows: ; ; in, , The spectrum of the signals from the two sensors within the same window. for conjugate, For mathematical expectation; During the training process of the dual-domain coupled avalanche physics-constrained autoencoder, the avalanche feature separation index combined with the dual-domain reconstruction loss is used as the model optimization index to select the autoencoder model with the maximum avalanche feature separation and the minimum reconstruction error; the expression for the avalanche feature separation index is: ; in, , These are the mean values ​​of the feature vectors output by the avalanche-specific feature extractor for avalanche samples and non-avalanche noise samples in the training set, respectively. , These are the standard deviations of the feature vectors output by the avalanche-specific feature extractor for avalanche samples and non-avalanche noise samples in the training set, respectively.

[0012] Optionally, in step 4, the low-dimensional fused feature vector and avalanche-specific engineering features are used as inputs, the avalanche or non-avalanche binary label of the sample is used as the classification label, and the avalanche event-level F1 score weighted by the missed detection rate is used as the optimization objective to train and optimize the hyperparameters of the random forest classifier to obtain the trained avalanche classification model; the avalanche-specific engineering features include waveform time-domain features, spectral features, time-spectral features, and avalanche-specific physical features; the avalanche-specific physical features include avalanche spatial coherence coefficient, energy envelope fitting degree, and frequency domain directional frequency shift index.

[0013] Optionally, in step 5, the microseismic waveform data to be detected, after preprocessing and window slicing, is input into the trained avalanche-specific feature extractor and the avalanche classification model to obtain the avalanche prediction probability at the single-sensor window level; through a two-level post-processing strategy of array spatiotemporal coherence probability aggregation and avalanche event evolution continuity determination, the final avalanche event detection result is output; the two-level post-processing strategy specifically includes: Step 51: Within the same time window, fuse the single-window avalanche prediction probabilities of each sensor in the microseismic sensor array to obtain the array-level window avalanche confidence score. The expression is: ; in, This represents the number of sensors in the microseismic sensor array. For the first m Single-window avalanche prediction probability for a single sensor is the average avalanche spatial coherence coefficient of adjacent sensor pairs within the microseismic sensor array; If and only if and At that time, the window was determined to be a positive window for avalanche. Step 52: Perform event-level determination on the continuous sliding window detection results. When the continuous avalanche positive window sequence meets the following two conditions at the same time, it is determined as a valid avalanche event: the total event duration corresponding to the continuous avalanche positive window is not less than 10s; and the energy envelope of the event-corresponding signal has a Pearson correlation coefficient with the preset avalanche standard evolution template envelope not less than 0.6. After the determination is completed, output the start and end time of the avalanche event and the detection confidence.

[0014] An automatic avalanche detection system based on self-encoder microseismic data, the system comprising: The data acquisition and preprocessing unit acquires continuous micro-seismic waveform data through a micro-seismic sensor array and preprocesses the acquired micro-seismic waveform data. The labeled dataset construction unit performs event picking and labeling on the preprocessed microseismic waveform data to construct a labeled dataset containing avalanche events and non-avalanche noise events, and divides the training set and test set according to the time dimension; The avalanche-specific feature extractor construction unit constructs a dual-domain coupled avalanche physical constraint autoencoder. Using the training set, the avalanche-specific feature extractor is obtained through unsupervised training with the avalanche physical constraint composite loss function, and outputs a low-dimensional fused feature vector. The avalanche classification model building unit constructs a random forest classifier with a weighted mechanism for avalanche feature saliency, trains and optimizes the hyperparameters of the random forest classifier to obtain the trained avalanche classification model. The two-stage post-processing unit outputs avalanche event detection results through array spatiotemporal coherence aggregation and evolution continuity determination.

[0015] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to perform the steps of the autoencoder-based microseismic data avalanche automatic detection method.

[0016] The present invention provides an automatic avalanche detection method and system based on autoencoder microseismic data. First, continuous microseismic waveform data is acquired through a microseismic sensor array. Standardized signals are obtained through preprocessing, such as energy weighting pre-enhancement in the avalanche frequency band. A labeled dataset containing labels for the entire avalanche evolution stage is constructed, and the dataset is divided according to the time dimension to avoid temporal leakage. A dual-domain coupled avalanche physical constraint autoencoder is constructed, and an avalanche-specific feature extractor is obtained through unsupervised training using a composite loss function. Sample classification is completed using a feature saliency-weighted random forest classifier. Finally, the avalanche event detection result is output through two-stage post-processing: array spatiotemporal coherence aggregation and evolution continuity determination. This invention has strong anti-interference capabilities, high detection accuracy, and low false negative rate, enabling all-weather, real-time, and high-precision automatic monitoring and early warning of avalanche disasters. It is suitable for avalanche disaster monitoring and early warning scenarios in high-altitude and cold mountainous areas. Compared with existing technologies, this invention has the following technical advantages: (1) This invention suppresses non-target frequency band interference by energy weighting pre-enhancement processing of avalanche-specific frequency band, improves the signal-to-noise ratio from the data source, and effectively distinguishes avalanche surface source signals from point source interference and random noise by combining array spatial coherence constraints, thus maintaining detection accuracy in low signal-to-noise ratio scenarios.

[0017] (2) The present invention constructs a dual-domain coupled avalanche physical constraint autoencoder, extracts the temporal evolution features and spectral features of avalanche signals, embeds the avalanche-specific physical characteristics into the model training process, avoids the model learning noise features, and improves the separation of avalanche and interference events by the extracted features; the model does not need to make large-scale parameter adjustments for different monitoring areas, and its generalization ability is better than that of general machine learning models.

[0018] (3) The present invention optimizes hyperparameters through three-fold cross-validation, which improves the performance stability of the model in actual field applications and the laboratory test results are highly consistent with the field application results.

[0019] (4) This invention effectively eliminates single-window misjudgments and reduces the false alarm rate by using a two-level post-processing strategy of array spatiotemporal coherence probability aggregation and avalanche event evolution continuity determination, thus meeting the needs of disaster early warning. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of an automatic avalanche detection method based on autoencoder for microseismic data in a specific embodiment of the present invention; Figure 2 This is a schematic diagram of avalanche micro-vibration waveforms in a specific embodiment of the present invention; Figure 3 This is a schematic diagram of a natural microseismic waveform in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the micro-vibration waveform of an aircraft during flight, according to a specific embodiment of the present invention. Figure 5 This is a schematic diagram of the structure of an automatic microseismic data avalanche detection system based on an autoencoder in a specific embodiment of the present invention. Detailed Implementation

[0022] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] like Figure 1 As shown, Figure 1This is a flowchart of an automatic avalanche detection method based on autoencoder microseismic data according to a specific embodiment of the present invention. The automatic avalanche detection method based on autoencoder microseismic data includes: Step 1: Continuous microseismic waveform data is acquired through a deployed microseismic sensor array. The raw microseismic data is preprocessed to obtain a standardized signal to be processed. The preprocessing includes dimension conversion, linear detrending, bandpass filtering, and avalanche frequency band energy weighting pre-enhancement processing to specifically enhance the signal energy of the core avalanche frequency band and suppress non-avalanche interference frequency bands. The microseismic sensor array consists of at least 5 single-component or three-component microseismometers with a sampling frequency of not less than 200Hz, a deployment range covering the target avalanche monitoring area, and a monitoring radius of not less than 3km. The bandpass filter is a fourth-order Butterworth bandpass filter with a passband of 1-10Hz. The operator expression for the avalanche frequency band energy weighted pre-enhancement processing is: ; in, This is the time-domain microseismic signal after bandpass filtering. for The frequency domain signal obtained by Fast Fourier Transform For the pre-enhanced time-domain signal, The weighting function for the avalanche-specific frequency band is expressed as follows: .

[0025] Step 2: Event picking and labeling are performed on the preprocessed signal to construct a labeled dataset containing avalanche events and non-avalanche noise events. The labeled dataset is labeled with time-series labels for the entire evolution stage of avalanche events, from initiation to development, peak, and decay. The labeled dataset is augmented with samples through differential sliding window slicing. After sample normalization, the training set and test set are divided according to the time dimension to avoid time-series data leakage. In this embodiment, event picking and labeling specifically involves: combining synchronously acquired radar and camera verification data, determining the start and end times of microseismic events through manual visual picking, such as... Figure 2 The yellow vertical line section provides a binary label for the events, categorizing them into avalanche events and non-avalanche noise events; among them, non-avalanche noise events include natural micro-earthquakes (such as...). Figure 3 ), aircraft transit (such as Figure 4 The data includes wind noise, microseismic signals generated by human activities, and sub-stages of the complete avalanche event, with time-series labels for the entire evolution stage from initiation to development, peak, and decay.

[0026] In this embodiment, the differentiated sliding window slicing specifically refers to: for the labeled avalanche event samples, a sliding window of 10 seconds in length is used, such as... Figure 2For the purple-red vertical line section, the window overlap rate is no less than 70%; for non-avalanche noise event samples, a 10-second sliding window is used with a window overlap rate of 50%. The normalization process is as follows: for each signal within a sliding window, the signal is independently normalized according to its own maximum absolute amplitude, and the signal amplitude is mapped to the [-1,1] interval. The training and test sets are divided according to the time dimension, with the training set accounting for 70% and the test set accounting for 30%, and the model hyperparameters are optimized through three-fold cross-validation.

[0027] Step 3: Construct a dual-domain coupled avalanche physics-constrained autoencoder, which includes a temporal coding branch, a spectral coding branch, a coupled feature fusion layer, and a dual-domain reconstruction decoder; perform unsupervised training on the autoencoder using the training set data, and iteratively optimize the avalanche physics-constrained composite loss function during training; after training, retain the encoder part as an avalanche-specific feature extractor and output a low-dimensional fused feature vector. The specific structure of the dual-domain coupled avalanche physics-constrained autoencoder is as follows: The temporal coding branch consists of three one-dimensional convolutional layers and one bidirectional LSTM layer stacked sequentially. It is used to extract the temporal evolution features of the avalanche signal and output a 32-dimensional temporal feature vector. The convolutional kernel size is 20, the stride is 10, the number of convolutional kernels in the first layer is 32, and the number of convolutional kernels in each subsequent layer doubles. The spectral domain coding branch consists of three fully connected layers and one one-dimensional convolutional layer. The input is the amplitude spectrum obtained by fast Fourier transform of microseismic waveform data, which is used to extract the spectral features and frequency shift law of avalanche signals and outputs a 16-dimensional spectral domain feature vector. The coupled feature fusion layer dynamically allocates attention weights for temporal and spectral features through feature saliency weights, fusing temporal and spectral feature vectors into a 48-dimensional low-dimensional fusion feature vector. The dual-domain reconstruction decoder consists of a time-domain reconstruction branch and a spectral-domain reconstruction branch. It reconstructs the input time-domain signal and spectral signal respectively, and outputs a reconstructed time-domain signal and a reconstructed spectrum with the same dimensions as the input.

[0028] The expression for the composite loss function under avalanche physics constraints is as follows: ; in, Preset weighting coefficients; The loss for bi-domain reconstruction is expressed as: ; In the formula, This is the time-domain microseismic signal after bandpass filtering. For the pre-enhanced time-domain signal, The amplitude spectrum of the input signal. To reconstruct the amplitude spectrum, MSE is the mean square error; The avalanche time-series evolution constraint loss is used to constrain the model to learn the energy evolution law specific to avalanches, and its expression is: ; In the formula, The energy envelope curve is obtained by applying the Hilbert transform and moving average to the signal. Pearson correlation coefficient; The spatial coherence constraint loss for the microsensor array is used to constrain the model's learning of the specific spatial coherence characteristics of the avalanche extension surface source. Its expression is: ; In the formula, Let be the avalanche spatial coherence coefficient of the input signals from two sensors in the same window of the microsensor array. The reconstructed avalanche spatial coherence coefficient is the reconstructed time-domain signal output by the autoencoder, corresponding to the time-domain input signals of the two sensors mentioned above within the same window. The expression for the avalanche spatial coherence coefficient is: ; ; in, , The spectrum of the signals from the two sensors within the same window. for conjugate, For mathematical expectation; During autoencoder training, an avalanche feature separation index combined with a dual-domain reconstruction loss is used as the model optimization metric to select the autoencoder model with the highest avalanche feature separation and the lowest reconstruction error. The expression for the avalanche feature separation index is as follows: ; in, , These are the mean values ​​of the feature vectors output by the avalanche-specific feature extractor for avalanche samples and non-avalanche noise samples in the training set, respectively. , These are the standard deviations of the feature vectors output by the avalanche-specific feature extractor for avalanche samples and non-avalanche noise samples in the training set, respectively.

[0029] Step 4: Construct a random forest classifier with a saliency weighting mechanism for avalanche features. Use the low-dimensional fused feature vector output by the avalanche-specific feature extractor and the avalanche-specific engineering features as model inputs. Use the avalanche / non-avalanche binary label of the sample as the classification label. Use the avalanche event-level F1 score weighted by the false negative rate as the optimization objective. Train and optimize the hyperparameters of the random forest classifier to obtain the trained avalanche classification model. The avalanche-specific engineering features include waveform time-domain features, spectral features, time-spectral features, and avalanche-specific physical features; the avalanche-specific physical features include avalanche spatial coherence coefficient, energy envelope fitting degree, and frequency domain directional frequency shift index.

[0030] Step 5: After preprocessing and window slicing in Step 1, the microseismic data to be detected is input into the trained avalanche-specific feature extractor and avalanche classification model to obtain the avalanche prediction probability at the single sensor window level. Through a two-level post-processing strategy of array spatiotemporal coherence probability aggregation and avalanche event evolution continuity determination, the final avalanche event detection result is output.

[0031] The two-stage post-processing strategy specifically includes: Step 51: Fuse the single-window avalanche prediction probabilities of each sensor in the microseismic array within the same time window to obtain the array-level windowed avalanche confidence score, expressed as: ; in, This represents the number of sensors in the microseismic array. For the first m Single-window avalanche prediction probability for a single sensor The average avalanche spatial coherence coefficient of adjacent sensor pairs within the array; If and only if and At that time, the window was determined to be a positive window for avalanche. Step 52: Perform event-level determination on the continuous sliding window detection results. When the continuous positive window sequence meets the following two conditions at the same time, it is determined as a valid avalanche event: the total event duration corresponding to the continuous positive windows is not less than 10s; and the energy envelope of the event-corresponding signal has a Pearson correlation coefficient with the preset avalanche standard evolution template envelope of not less than 0.6. After the determination is completed, output the start and end time of the avalanche event and the detection confidence.

[0032] Accordingly, embodiments of the present invention also provide an automatic avalanche detection system based on self-encoder microseismic data, such as... Figure 5 The image shown is a structural schematic of the system. This automatic microseismic data avalanche detection system based on a self-encoder includes the following modules: The data acquisition and preprocessing unit 501 acquires continuous micro-vibration waveform data through a micro-vibration sensor array and preprocesses the acquired micro-vibration waveform data. The labeled dataset construction unit 502 performs event picking and labeling on the preprocessed microseismic waveform data, constructs a labeled dataset containing avalanche events and non-avalanche noise events, and divides the training set and test set according to the time dimension; Avalanche-specific feature extractor construction unit 503 constructs a dual-domain coupled avalanche physical constraint autoencoder, and obtains an avalanche-specific feature extractor through unsupervised training using the training set and a composite loss function, outputting a low-dimensional fused feature vector. Avalanche classification model building unit 504 constructs a random forest classifier with a saliency weighting mechanism for avalanche features. The random forest classifier is trained and its hyperparameters are optimized to obtain the trained avalanche classification model. The two-stage post-processing unit 505 outputs avalanche event detection results through two-stage post-processing of array spatiotemporal coherence aggregation and evolution continuity determination.

[0033] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0034] The present invention also provides a storage medium, which is a computer-readable storage medium storing a computer program thereon, the computer program being executable when it runs. Figure 1 The method shown may include some or all of the steps. The storage medium may include read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc. The storage medium may also include non-volatile memory or non-transitory memory, etc.

[0035] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data provider to another website, computer, server, or data provider via wired or wireless means.

[0036] The embodiments of the present invention have been described in detail above. Specific implementation methods have been used to illustrate the present invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method of the present invention, and are merely some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention, and the content of this specification should not be construed as a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An automatic avalanche detection method based on autoencoder microseismic data, characterized in that, The method includes: Step 1: Collect continuous micro-vibration waveform data through a micro-vibration sensor array, and preprocess the collected micro-vibration waveform data; Step 2: Perform event picking and annotation on the preprocessed microseismic waveform data to construct an annotated dataset containing avalanche events and non-avalanche noise events, and divide the dataset into training and testing sets according to the time dimension. Step 3: Construct a dual-domain coupled avalanche physics-constrained autoencoder, and use the training set to obtain an avalanche-specific feature extractor through unsupervised training using the avalanche physics-constrained composite loss function, outputting a low-dimensional fused feature vector. Step 4: Construct a random forest classifier with a weighted mechanism for avalanche feature saliency, train and optimize the hyperparameters of the random forest classifier to obtain the trained avalanche classification model. Step 5: Output the avalanche event detection results through two-stage post-processing: array spatiotemporal coherence aggregation and evolution continuity determination.

2. The automatic avalanche detection method based on autoencoder for microseismic data according to claim 1, characterized in that, In step 1, the microseismic sensor array consists of at least 5 single-component or three-component microseismometers with a sampling frequency of not less than 200Hz, and the deployment range covers the target avalanche monitoring area with a monitoring radius of not less than 3km. The preprocessing includes dimensional transformation, linear detrending, bandpass filtering, and avalanche band energy weighting pre-enhancement processing to specifically enhance the signal energy of the core advantageous frequency band of avalanche and suppress non-avalanche interference frequency bands; the bandpass filtering is a 4th order Butterworth bandpass filter with a passband of 1-10Hz. The operator expression for the avalanche frequency band energy weighted pre-enhancement processing is: ; in, This is the time-domain microseismic signal after bandpass filtering. for The frequency domain signal obtained by Fast Fourier Transform For the pre-enhanced time-domain signal, The weighting function for the avalanche-specific frequency band is expressed as follows: 。 3. The automatic avalanche detection method based on autoencoder for microseismic data according to claim 1, characterized in that, In step 2, the event picking and labeling specifically involves: combining synchronously acquired radar and camera verification data, determining the start and end times of microseismic events through manual visual picking, and performing binary labeling on the events, classifying them into avalanche events and non-avalanche noise events; among which, non-avalanche noise events include natural microseismic events, aircraft overflight, wind noise, and microseismic signals generated by human activities; simultaneously, dividing the complete avalanche event into sub-stages, labeling the time series labels of the entire evolution stage of the avalanche event, including initiation, development, peak, and decay, to obtain the labeled dataset; performing sample augmentation on the labeled dataset through differential sliding window slicing, completing sample normalization processing, and then dividing the training set and the test set according to the time dimension to avoid time series data leakage.

4. The automatic avalanche detection method based on autoencoder for microseismic data according to claim 3, characterized in that, In step 2, the differentiated sliding window slicing is specifically as follows: for labeled avalanche event samples, a sliding window of 10 seconds is used, with a window overlap rate of no less than 70%; for non-avalanche noise event samples, a sliding window of 10 seconds is used, with a window overlap rate of 50%. The normalization process is as follows: for each signal within a sliding window, the signal is independently normalized according to its own maximum absolute amplitude, and the signal amplitude is mapped to the interval [-1, 1]. The training set and the test set are divided according to the time dimension, with the training set accounting for 70% and the test set accounting for 30%, and the model hyperparameters are optimized by three-fold cross-validation.

5. The automatic avalanche detection method based on autoencoder for microseismic data according to claim 1, characterized in that, In step 3, the dual-domain coupled avalanche physics-constrained autoencoder includes: The temporal coding branch consists of three one-dimensional convolutional layers and one bidirectional LSTM layer stacked sequentially. It is used to extract the temporal evolution features of the avalanche signal and output a 32-dimensional temporal feature vector. The convolutional kernel size is 20, the stride is 10, the number of convolutional kernels in the first layer is 32, and the number of convolutional kernels in each subsequent layer doubles. The spectral domain coding branch consists of three fully connected layers and one one-dimensional convolutional layer. The input is the amplitude spectrum obtained by fast Fourier transform of microseismic waveform data, which is used to extract the spectral features and frequency shift law of avalanche signals and outputs a 16-dimensional spectral domain feature vector. The coupled feature fusion layer dynamically allocates attention weights for temporal and spectral features through feature saliency weights, fusing temporal and spectral feature vectors into a 48-dimensional low-dimensional fusion feature vector. The dual-domain reconstruction decoder consists of a time-domain reconstruction branch and a spectral-domain reconstruction branch. It reconstructs the input time-domain signal and spectral signal respectively, and outputs a reconstructed time-domain signal and a reconstructed spectrum with the same dimensions as the input.

6. The automatic avalanche detection method based on autoencoder for microseismic data according to claim 5, characterized in that, In step 3, the dual-domain coupled avalanche physics-constrained autoencoder is trained unsupervised using the training set. During training, the avalanche physics-constrained composite loss function is used for iterative optimization. After training, the encoder part is retained as the avalanche-specific feature extractor. The expression of the avalanche physics-constrained composite loss function is: ； in, Preset weighting coefficients; The loss for bi-domain reconstruction is expressed as: ； In the formula, This is the time-domain microseismic signal after bandpass filtering. For the pre-enhanced time-domain signal, The amplitude spectrum of the input signal. To reconstruct the amplitude spectrum, MSE is the mean square error; The avalanche time-series evolution constraint loss is used to constrain the model to learn the energy evolution law specific to avalanches, and its expression is: ； In the formula, The energy envelope curve is obtained by applying the Hilbert transform and moving average to the signal. The Pearson correlation coefficient; The spatial coherence constraint loss for the microsensor array is used to constrain the model's learning of the specific spatial coherence characteristics of the avalanche extension surface source. Its expression is: ； In the formula, Let be the avalanche spatial coherence coefficient of the input signals from two sensors in the same window of the microsensor array. The reconstructed avalanche spatial coherence coefficient is calculated from the reconstructed time-domain signal output by the dual-domain coupled avalanche physical constraint autoencoder, corresponding to the time-domain input signals from the two sensors within the same window. The expression for the avalanche spatial coherence coefficient is: ； ； in, , The spectrum of the signals from the two sensors within the same window. for conjugate, For mathematical expectation; During the training process of the dual-domain coupled avalanche physics-constrained autoencoder, the avalanche feature separation index combined with the dual-domain reconstruction loss is used as the model optimization index to select the autoencoder model with the maximum avalanche feature separation and the minimum reconstruction error; the expression for the avalanche feature separation index is: ； in, , These are the mean values ​​of the feature vectors output by the avalanche-specific feature extractor for avalanche samples and non-avalanche noise samples in the training set, respectively. , These are the standard deviations of the feature vectors output by the avalanche-specific feature extractor for avalanche samples and non-avalanche noise samples in the training set, respectively.

7. The automatic avalanche detection method based on autoencoder for microseismic data according to claim 1, characterized in that, In step 4, the low-dimensional fused feature vector and avalanche-specific engineering features are used as inputs, the avalanche or non-avalanche binary label of the sample is used as the classification label, and the avalanche event-level F1 score weighted by the false negative rate is used as the optimization objective to train and optimize the hyperparameters of the random forest classifier to obtain the trained avalanche classification model; the avalanche-specific engineering features include waveform time-domain features, spectral features, time-spectral features, and avalanche-specific physical features; The avalanche-specific physical characteristics include the avalanche spatial coherence coefficient, energy envelope fit, and frequency domain directional frequency shift index.

8. The automatic avalanche detection method based on autoencoder for microseismic data according to claim 1, characterized in that, In step 5, the microseismic waveform data to be detected, after preprocessing and window slicing, is input into the trained avalanche-specific feature extractor and the avalanche classification model to obtain the avalanche prediction probability at the single-sensor window level. Through a two-stage post-processing strategy of array spatiotemporal coherence probability aggregation and avalanche event evolution continuity determination, the final avalanche event detection result is output. The two-stage post-processing strategy specifically includes: Step 51: Within the same time window, fuse the single-window avalanche prediction probabilities of each sensor in the microseismic sensor array to obtain the array-level window avalanche confidence score. The expression is: ; in, This represents the number of sensors in the microseismic sensor array. For the first m Single-window avalanche prediction probability for a single sensor is the average avalanche spatial coherence coefficient of adjacent sensor pairs within the microseismic sensor array; If and only if and At that time, the window was determined to be a positive window for avalanche. Step 52: Perform event-level determination on the continuous sliding window detection results. When the continuous avalanche positive window sequence meets the following two conditions at the same time, it is determined as a valid avalanche event: the total event duration corresponding to the continuous avalanche positive window is not less than 10s; and the energy envelope of the event-corresponding signal has a Pearson correlation coefficient with the preset avalanche standard evolution template envelope not less than 0.

6. After the determination is completed, output the start and end time of the avalanche event and the detection confidence.

9. An automatic avalanche detection system based on microseismic data using an autoencoder, characterized in that, The system includes: The data acquisition and preprocessing unit acquires continuous micro-seismic waveform data through a micro-seismic sensor array and preprocesses the acquired micro-seismic waveform data. The labeled dataset construction unit performs event picking and labeling on the preprocessed microseismic waveform data to construct a labeled dataset containing avalanche events and non-avalanche noise events, and divides the training set and test set according to the time dimension; The avalanche-specific feature extractor construction unit constructs a dual-domain coupled avalanche physical constraint autoencoder. Using the training set, the avalanche-specific feature extractor is obtained through unsupervised training with the avalanche physical constraint composite loss function, and outputs a low-dimensional fused feature vector. The avalanche classification model building unit constructs a random forest classifier with a weighted mechanism for avalanche feature saliency, trains and optimizes the hyperparameters of the random forest classifier to obtain the trained avalanche classification model. The two-stage post-processing unit outputs avalanche event detection results through array spatiotemporal coherence aggregation and evolution continuity determination.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program, when run by the processor, executes the steps of the automatic avalanche detection method based on autoencoder for microseismic data as described in any one of claims 1 to 8.

Images