Micro-seismic signal spectrum analysis system and method based on DVEC

Abstract

The invention belongs to the technical field of seismic signal monitoring, and discloses a micro-seismic signal spectrum analysis system and method based on DVEC, and the system comprises a data collection device, a data preprocessing module, and a depth variation code-classifier. Acquiring a plurality of original vibration signals through a data acquisition device; performing fast Fourier transform (FFT) processing on the plurality of original vibration signals by using a data preprocessing module to obtain corresponding frequency domain sequence signals; splicing the original micro-seismic signals on a plurality of time domains and the corresponding frequency domain sequence signals in a dual-channel mode to form dual-channel micro-seismic signal input data; and inputting the dual-channel micro-seismic signal input data into the depth variational coding-classifier to obtain the classification category of each original micro-seismic signal. According to the method, different types of seismic signal features can be clearly distinguished, unknown types of seismic signals can be identified, and the method is expected to be an important technical means in the field of seismic signal identification.

Description

Technical Field

[0001] The present invention belongs to the technical field of seismic signal monitoring, and relates to a microseismic signal spectrum analysis system and method based on DVEC, mainly to the recognition and application of seismic signal analysis through a deep variational coding-classifier network (DVEC). Background Art

[0002] In the field of seismic signal recognition, accurate signal recognition is of great significance for earthquake monitoring, early warning and disaster reduction. Although traditional seismic signal analysis methods, such as Fourier transform and wavelet transform, can reveal the frequency characteristics of signals to a certain extent, they have obvious limitations in processing nonlinear and non-stationary signals. These methods often rely on linear transformations or fixed basis functions, which makes it difficult to effectively capture the nonlinear and dynamic characteristics of seismic signals. In addition, traditional methods usually require pre-setting parameters, such as window length and threshold. The selection of these parameters has a significant impact on the analysis results, but often lacks theoretical guidance, which easily leads to subjectivity and uncertainty in the results.

[0003] In recent years, with the development of deep learning technology, methods based on deep neural networks have made significant progress in the field of seismic signal recognition. Deep learning models, such as convolutional neural networks (CNN) and recurrent neural networks (RNN), can achieve high-precision classification of seismic signals by automatically extracting deep features of signals. These models can learn more abstract and rich feature representations from raw signals without relying on manually designed feature extractors. Therefore, compared with traditional methods, deep learning models have greater flexibility and generalization ability when processing seismic signals, and can better adapt to the complexity and diversity of signals.

[0004] However, most of the existing deep learning models are based on the closed set assumption, that is, all categories are known during the training phase. In practical applications, due to the diversity and uncertainty of seismic signals, new or unknown signal types may appear, which do not appear in the training data, resulting in limited recognition capabilities of existing models. When faced with unknown categories, existing models often show a tendency to be overconfident and misclassify unknown signals as known categories, thus affecting the accuracy and robustness of recognition. Therefore, for the problem of seismic signal recognition in open environments, further research on open set recognition algorithms is needed to reduce the impact of unknown categories on the microseismic signal spectrum analysis system, so as to better cope with the challenges in practical applications.

[0005] Therefore, it is urgent to propose a new seismic signal analysis method to solve the limitations, instability and subjectivity of traditional seismic signal analysis methods in seismic signal identification and analysis, as well as the inability of simple closed-set deep learning models to accurately classify unknown signals. Summary of the invention

[0006] The purpose of the present invention is to provide a microseismic signal spectrum analysis system based on DVEC (Deep Variational Encoder-Classifier) ​​to address the problems existing in the above-mentioned prior art, which can realize effective analysis of microseismic signals.

[0007] The invention also provides a microseismic signal spectrum analysis method based on DVEC.

[0008] The present invention introduces DVEC in seismic signal analysis, combining deep learning and time series analysis. It can effectively process non-stationary signals and extract deep features reflecting the state of seismic activity from the signals. Unlike traditional closed-set deep learning models, the DVEC algorithm can accurately classify known categories, thereby improving the robustness and adaptability of seismic signal recognition and improving the shortcomings of closed-set models.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions.

[0010] The DVEC-based microseismic signal spectrum analysis system provided by the present invention comprises:

[0011] A data acquisition device for collecting microseismic signals;

[0012] The data preprocessing module is used to perform fast Fourier transform (FFT) processing on the microseismic signals to obtain the corresponding frequency domain sequence signals; and to splice several original microseismic signals in the time domain and the corresponding frequency domain sequence signals in a dual-channel manner to form dual-channel microseismic signal input data;

[0013] A deep variational coding-classifier is used to extract features from dual-channel microseismic signal input data and classify the signal based on the extracted features; the deep variational coding-classifier includes an encoder and a classifier; the encoder is used to obtain latent features based on the dual-channel microseismic signal input data; the classifier is used to obtain the classification category of each original microseismic signal based on the latent features;

[0014] The above-mentioned data acquisition device can collect any microseismic signals from the ground, for example, it can be used to collect microseismic signals generated in industrial and engineering production processes such as deep geological energy development and geotechnical construction; the original microseismic signals can be collected through conventional vibration sensors.

[0015] In the above-mentioned deep variational coding-classifier, the encoder is used to obtain the expected value μ and the standard deviation σ according to the dual-channel microseismic signal input data, and from N(μ,σ 2) to obtain hidden features z; the encoder includes several convolution modules, several convolution layers and two fully connected layers; the convolution module includes a convolution layer and a pooling layer; the dual-channel microseismic signal spectrum input is first mapped to a low-dimensional latent space by several convolution modules using a probabilistic trapezoidal structure, and then further extracts features through several convolution layers, and finally is mapped to the mean and standard deviation of a Gaussian distribution through two fully connected layers.

[0016] In the above-mentioned deep variational coding-classifier, the classifier is used to obtain the classification category of each original microseismic signal in a known class based on latent features; the classifier is composed of a number of fully connected layers.

[0017] In the above deep variational encoding-classifier, the total loss function used is:

[0018] L=L CE -λ(x)D KL ;

[0019] Among them, L CE is the classification loss, and the cross entropy loss function is used; λ(x) represents the introduced adaptive weight, x represents the dual-channel microseismic signal spectrum input data, D KL represents the KL divergence regularization term.

[0020] Among them, the classification loss L CE Specifically expressed as:

[0021]

[0022] represents the expected value; p ξ (y|z) represents the category probability, which is the mapping from the latent feature z of the microseismic signal to the category label y of the microseismic signal, that is, the probability of obtaining category y when the latent feature z is given; the expected value The calculation is done by the method of samples, starting from We then use these samples to calculate logp θ The average value of (x|z).

[0023] KL divergence regularization term D KL It is expressed as:

[0024]

[0025] Among them, q φ (z|x) represents the posterior distribution, p θ (z) represents the hidden characteristic distribution of microseismic signals, and L represents the length of z.

[0026] Through the design of this loss function, the deep variational encoder-classifier (DVEC model) can balance the classification error and the regularization of feature distribution during the learning process, thereby achieving more accurate and robust classification of microseismic signals. This not only improves the recognition accuracy of known microseismic signal types, but also enhances the system's adaptability to unknown types, making up for the interference and classification of unknown classes on known classes in the original simple neural network, providing additional protection for the safety of dams and hydropower stations.

[0027] The above deep variational encoder-classifier also includes an adaptive weight adjustment module, which is expressed as:

[0028]

[0029] Among them, ω2 represents the predefined weight parameter; y represents the sample x l The true label of Represents the deep variational encoding-classifier for sample x l The predicted label of ; ω(x) represents the cumulative density function (CDF) of the Weibull distribution;

[0030] ω(x) is expressed as:

[0031]

[0032] Among them, κ represents the shape parameter in the Weibull distribution; s represents the variable used in the Weibull distribution; γ represents the scale parameter of the Weibull distribution.

[0033] The adaptive weight function regards the characteristic posterior distribution of each microseismic signal class as a sphere, and the KL distance from the center of the sphere to the boundary is regarded as the distance from the data feature to the center of the distribution. For the latent feature distribution space of the known class, the misclassified data is outside the sphere. The adaptive weight function λ(x) automatically sets a larger distribution convergence loss weight for extreme data to accelerate the distribution constraint of these extreme microseismic signal samples. In each training cycle, adaptive weights are set for new extreme data. At the same time, for the misclassified data, its KL loss weight is adaptively set to zero to avoid misleading distribution parameters during training.

[0034] The present invention also provides a training method for the above-mentioned DVEC-based microseismic signal spectrum analysis system, wherein a plurality of sample data for training are divided into a plurality of batches, and each batch is executed according to the following steps:

[0035] S1 uses the data preprocessing module to perform fast Fourier transform (FFT) processing on the original microseismic signals in the batch of sample data to obtain corresponding frequency domain sequence signals; and splices the original microseismic signals in the sample data and the corresponding frequency domain sequence signals to form dual-channel microseismic signal input data;

[0036] S2 inputs the dual-channel microseismic signal input data into the deep variational coding-classifier to obtain the KL divergence regularization term, and combines the label data in the sample data to obtain the classification loss;

[0037] S3 calculates the total loss by combining the classification loss with the KL divergence loss;

[0038] S4 optimizes the encoder and classifier parameters through the total loss;

[0039] Return to step S1, repeat steps S1-S4, and train the system through each batch of sample data until the system converges.

[0040] In the above step S4, the encoder and classifier parameters can be optimized by a gradient descent optimization algorithm.

[0041] Whether the system has converged can be determined by whether the total loss value reaches the set threshold range or whether the upper limit of the number of iterations is reached.

[0042] The present invention also provides a DVEC-based microseismic signal spectrum analysis method, which uses the above-mentioned system and is performed according to the following steps:

[0043] L1 collects microseismic signals through a data acquisition device;

[0044] L2 uses the data preprocessing module to perform fast Fourier transform (FFT) processing on the microseismic signal to obtain the corresponding frequency domain sequence signal; and splices the original microseismic signal and the corresponding frequency domain sequence signal to form dual-channel microseismic signal input data;

[0045] L3 inputs the dual-channel microseismic signal input data into the deep variational coder-classifier to obtain the classification result of the original microseismic signal.

[0046] Compared with the prior art, the DVEC-based microseismic signal spectrum analysis method provided by the present invention has the following outstanding advantages:

[0047] 1) The present invention combines multi-source data (time domain sequence data and frequency domain training data) to construct dual-channel microseismic signal input data, and then extracts and classifies features based on the dual-channel microseismic signal input data through the constructed deep variational coding-classifier (DVEC model), so as to realize effective analysis of microseismic signals;

[0048] 2) The dual-channel microseismic signal input data of the present invention fuses and fully utilizes the time domain and frequency domain information of the signal, which is the key to transferring from non-sequential data tasks to sequential signal tasks and also helps to extract robust features;

[0049] 3) The loss function design of the DVEC model in the present invention optimizes the feature extraction process, which helps to accurately extract key information from microseismic signals; in the monitoring of microseismic facilities, this means that microseismic activities that may affect the stability of buildings can be identified more quickly, so that necessary preventive measures can be taken in advance;

[0050] 4) The present invention strengthens the focus on samples near the classification boundary by applying adaptive weights to the DVEC model, which is of great significance for predicting and preventing rare but potentially harmful microseismic events; in the water conservancy and hydropower industry, this ability is crucial to ensure the safe operation of dams and hydropower stations;

[0051] 5) The DVEC model adds variational processing and prediction category constraints based on the supervised learning paradigm to the features extracted by the encoder, so that the posterior distribution of the features output by the encoder can approximate the true prior distribution; this mechanism not only improves the contribution of features to known class classification tasks, but also strengthens the feature aggregation of input data in the hidden layer of the network; by forcing different latent features to approximate different Gaussian models, the DVEC model can provide a latent feature space with strong distribution and high cohesion, thereby ensuring that the features within each category are highly aggregated.

[0052] 6) The DVEC model constructed by the present invention can clearly distinguish the characteristics of microseismic signals of different categories. In addition, the model uses KL divergence constraints to make the feature distribution more compact, thereby improving the intra-class aggregation; the high aggregation of these features also provides reliability and robustness in the subsequent recognition of unknown class data.

[0053] In practical applications, the DVEC-based microseismic signal spectrum analysis method can be applied to multiple fields such as microseismic monitoring, microseismic early warning, and microseismic disaster assessment. Compared with traditional methods, the DVEC model has shown significant advantages in processing complex microseismic signals, identifying unknown signal types, and adapting to signal variations, and is expected to become an important technical means in the field of seismic signal recognition. With the deepening of seismological research and the continuous development of artificial intelligence technology, the DVEC-based microseismic signal spectrum analysis method will play an increasingly important role in the field of microseismic signal recognition in the future. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a schematic diagram of the data preprocessing module principle;

[0055] Figure 2 Schematic diagram of the deep variational coding-classifier principle;

[0056] Figure 3 This is a schematic diagram of the training process of the microseismic signal spectrum analysis system based on DVEC;

[0057] Figure 4 Schematic diagram of the working principle of adaptive weight;

[0058] Figure 5 Schematic diagram of the process of microseismic signal spectrum analysis method based on DVEC. DETAILED DESCRIPTION

[0059] The present invention is described in detail below by embodiments and application examples. It is necessary to point out that the present embodiments are only used to further illustrate the present invention and cannot be construed as limiting the scope of the present invention. Those skilled in the art in this field can make some non-essential improvements and adjustments according to the above-mentioned content of the present invention. If specific conditions are not specified in the embodiments, they are carried out according to normal conditions or conditions recommended by the manufacturer. If the manufacturer is not specified for the reagents or instruments used, they are all conventional products that can be purchased commercially.

[0060] Example 1

[0061] The DVEC-based microseismic signal spectrum analysis system provided by the present invention provided in this embodiment includes: a data acquisition device, a data preprocessing module, and a deep variational coder-classifier.

[0062] (1) Data acquisition device

[0063] The data acquisition device is used to collect any microseismic signals from the ground, such as microseismic signals generated in industrial and engineering production processes such as deep geological energy development and geotechnical construction. The data acquisition device can be a conventional vibration sensor, which collects the original microseismic signals encountered in the above-mentioned various industrial productions. These original microseismic signals can reflect the geological conditions of industrial production areas such as dams, mines, and deep geology.

[0064] (2) Data preprocessing module

[0065] like Figure 1 As shown, the data preprocessing module is used to perform fast Fourier transform (FFT) processing on the microseismic signal to obtain the corresponding frequency domain sequence signal; and to splice several original microseismic signals and the corresponding frequency domain sequence signals to form dual-channel microseismic signal input data.

[0066] These original microseismic signals are time-domain sequence signals. By converting time-domain signals into frequency-domain signals, the different frequency fluctuations generated by microseismic activities can be observed more clearly.

[0067] In most methods, only one signal feature is used as the input of the diagnostic model, which is insufficient to reflect the diverse and variable microseismic signals and difficult to meet the requirements of signal complexity in the natural environment. Therefore, one-dimensional time domain or frequency domain analysis cannot fully reflect the characteristics of the signal, and different time-frequency patterns are specific to different microseismic signal conditions, which requires a combination of time-frequency analysis.

[0068] In the application problem faced by this embodiment, the original microseismic signal data of the sensor is a time domain sequence signal, and a frequency domain sequence signal is obtained after FFT. Then, the data in the frequency domain and time domain (i.e., the original microseismic signal and the frequency domain sequence signal) are combined in a dual-channel manner and used as the input of the subsequent DVEC network model. Through such preprocessing, the time domain and frequency domain information of the signal are integrated and fully utilized. This is the key to applying the open set recognition model to the microseismic analysis problem, transferring from non-sequential data tasks to sequential signal tasks; it is also one of the keys to overcome the defect that existing methods are difficult to extract robust features when applied to signal data sets. In water conservancy projects, this conversion helps to more accurately monitor and analyze abnormal vibrations that may be caused by microseisms, so as to take preventive measures.

[0069] (3) Deep Variational Encoder-Classifier (DVEC)

[0070] The deep variational coding-classifier is used to extract features from dual-channel microseismic signal input data and classify the signals based on the extracted features.

[0071] like Figure 2 As shown, the deep variational coding-classifier includes an encoder and a classifier. The encoder is used to obtain the expected value μ and the standard deviation σ according to the dual-channel microseismic signal input data, and 2 ) is sampled to obtain the latent feature z. The classifier is used to obtain the classification category of each original microseismic signal based on the latent feature.

[0072] like Figure 2 As shown in the figure, the encoder includes three convolution modules, three convolution layers and two fully connected layers; the convolution module includes a convolution layer and a pooling layer; the dual-channel microseismic signal spectrum input is first mapped to a low-dimensional latent space by three convolution modules with a probabilistic trapezoidal structure, and then further extracts features through three convolution layers, and finally mapped to the mean and standard deviation of the Gaussian distribution through two fully connected layers. Then, by 2 ) to obtain the hidden feature z.

[0073] like Figure 2 The classifier is composed of three fully connected layers. The classifier takes the latent feature z obtained by the encoder as input and outputs the classification category of the microseismic signal.

[0074] The deep variational coding-classifier also includes an adaptive weight adjustment module for dynamically adjusting the weight of the KL divergence loss in the loss function during system training.

[0075] The DVEC provided in this embodiment introduces a novel loss function based on the variational autoencoder (VAE). With the help of the strong correlation between the input x and the label y, the loss function replaces the first part of the loss function in the original VAE (i.e., the reconstruction error p(x|z)) with the sample classification error pξ(y|z). The new loss function enables the network to pay more attention to the known microseismic signal categories subject to variational constraints, learn the features subject to variational constraints, thereby obtaining a more accurate decision boundary, and ultimately improving the accuracy of the microseismic signal spectrum analysis system, which is particularly important for the real-time monitoring and early warning system of water conservancy and hydropower facilities.

[0076] In this embodiment, the total loss function in the training phase of the DVEC network model is:

[0077] L=L CE -λ(x)D KL ;

[0078] Among them, L CE is the classification loss, and the cross entropy loss function is used; λ(x) represents the introduced adaptive weight, x represents the dual-channel microseismic signal input data, and D KL represents the KL divergence regularization term.

[0079] ①Classification loss

[0080] Classification loss L CE Specifically expressed as:

[0081]

[0082] The expected value of the representation is achieved by the sample method, from We then use these samples to calculate logp ξ The average value of (x|z); p ξ (y|z) represents the probability distribution of each category obtained by the classifier.

[0083] The cross entropy loss is used to measure the accuracy of the classifier in classifying samples of known classes. This is to address the problem that VAE does not have sufficient ability to separate unknown classes from known classes, and the generation model decoder requires a large number of training samples. In addition, this can also add a supervised learning paradigm to the variational Bayesian algorithm to extract features that can well separate unknown classes from known classes.

[0084] ②KL divergence regularization term

[0085] KL divergence regularization term D KL It is expressed as:

[0086]

[0087] Among them, q φ (z|x) represents the posterior distribution, p θ (z) represents the hidden characteristic distribution of microseismic signals, and L represents the length of z.

[0088] This part measures the hidden feature distribution p through KL divergence θ The difference between (z) and the posterior distribution helps to reduce the intra-class diffusion in the feature space and improve the generalization ability of the network model. Among them, the expected μ and standard deviation σ of the approximate posterior are calculated by the neurons in the fully connected layer. The hidden feature z is defined as:

[0089] z=μ+σ·ε;ε∈N(0,I)

[0090] This is equivalent to sampling from the standard normal distribution to obtain ε, adding the dot product of σ and μ, and then obtaining N(μ,σ) after parameter transformation. 2 ). After this processing, the sampling operation does not need to participate in gradient descent, making the network model trainable.

[0091] Through the design of the above loss function, the deep variational encoding-classifier (DVEC model) can balance the classification error and the regularization of feature distribution during the learning process, thereby achieving more accurate and robust seismic signal recognition in the water conservancy and hydropower industry. This not only improves the recognition accuracy of known seismic signal types, but also enhances the system's adaptability to unknown or emerging types, providing additional protection for the safety of dams and hydropower stations.

[0092] ③Adaptive weight adjustment

[0093] The adaptive weight λ(x) is expressed as:

[0094]

[0095] Among them, ω2 represents the predefined weight parameter; y represents the sample x l The true label of Represents the deep variational encoding-classifier for sample x l The predicted label of ; ω(x) represents the cumulative density function (CDF) of the Weibull distribution;

[0096] ω(x) is expressed as:

[0097]

[0098] Among them, κ represents the shape parameter in the Weibull distribution; s represents the variable used in the Weibull distribution; γ represents the scale parameter of the Weibull distribution.

[0099] The role of the adaptive weight function in the deep variational coding-classifier is to regard the characteristic posterior distribution of each microseismic signal class as a sphere, with the center of the sphere representing the center of the distribution, and the Kullback-Leibler (KL) distance from the center of the sphere to the boundary is regarded as the distance from the data feature to the center of the distribution. In this latent feature distribution space, the data of the known class should be located inside the sphere, while the misclassified data is located outside the sphere. For the newly emerging extreme value data, the adaptive weight is automatically set, and for the misclassified data, its KL loss weight is adaptively set to zero to avoid the distribution parameters being misled during the training process. The key role of the adaptive weight function λ(x) is that it can automatically assign a larger distribution convergence loss weight to the extreme value data, which can strengthen the distribution constraints on these extreme microseismic signal samples, accelerate the convergence process of the model to these abnormal data, and thus improve the stability and accuracy of the model. Through this mechanism, the model can not only adapt to extreme value data faster, but also effectively avoid the negative impact of misclassified data on the overall distribution parameters, ensuring the reliability and efficiency of the training process.

[0100] The core advantage of the DVEC network model is that it can automatically learn the features related to seismic activity from the original seismic signal through an end-to-end learning mechanism without manually designing feature extractors. Therefore, it can effectively avoid the problem that traditional signal recognition methods such as fast Fourier transform cannot accurately extract nonlinear features and rely on artificial parameter settings and are inaccurate. In addition, the DVEC network model can also capture the dynamic change trend of the signal through time series analysis, thereby providing more comprehensive and accurate information for seismic signal recognition. By comprehensively analyzing the spectral characteristics of the signal, the DVEC network model can reveal the seismic activity patterns hidden in the signal and provide a scientific basis for earthquake monitoring and early warning.

[0101] In summary, the DVEC network model not only improves the recognition accuracy of known seismic signal types in the seismic signal recognition task in the water conservancy and hydropower industry, but also significantly enhances the system's adaptability to unknown or emerging types and reduces the impact of unknown categories on known categories. The realization of these technical effects provides a new and efficient technical means for the earthquake monitoring and early warning system of water conservancy and hydropower facilities, which helps to improve the safety level and risk management capabilities of the entire industry.

[0102] Example 2

[0103] This embodiment provides a training method for the microseismic signal spectrum analysis system based on DVEC in Embodiment 1. The sample data for training are divided into several batches; Figure 3 As shown, each batch is executed according to the following steps:

[0104] S1 uses the data preprocessing module to perform fast Fourier transform (FFT) processing on the original microseismic signals in the batch of sample data to obtain corresponding frequency domain sequence signals; and splices the original microseismic signals in the sample data and the corresponding frequency domain sequence signals to form dual-channel microseismic signal input data;

[0105] S2 inputs the dual-channel microseismic signal input data into the deep variational coding-classifier to obtain the KL divergence regularization term, and combines the label data in the sample data to obtain the classification loss;

[0106] S3 calculates the total loss based on the classification loss and KL divergence loss;

[0107] S4 optimizes the encoder and classifier parameters through the total loss; the optimization process can be performed through the gradient descent optimization algorithm.

[0108] like Figure 4 As shown in the figure, (a) is the classification result of a batch of samples obtained by the deep variational encoding-classifier. According to the classification accuracy of the batch and the distance from the distribution center, the size of the adaptive weight λ(x) is adjusted through the above operation steps; the classification result after the next batch of samples is trained is shown in (b). It can be seen that through adaptive weight updating, the network model can pay more attention to samples near the classification boundary.

[0109] After completing the training of a batch of samples, return to step S1 and repeat steps S1-S4 to train the system through each batch of sample data until the system converges.

[0110] Whether the system has converged can be determined by whether the total loss value reaches the set threshold range or whether the upper limit of the number of iterations is reached.

[0111] This embodiment effectively improves the classification accuracy of known categories through the design of preprocessing strategy, DVEC model optimization objectives and adaptive weights, reduces the impact of subsequent unknown class data on the recognition of known class data, and ensures the goodness and robustness of feature distribution.

[0112] Example 3

[0113] This embodiment provides a method for analyzing the spectrum of microseismic signals based on DVEC, using the system provided in Embodiment 1, and performing the following steps:

[0114] L1 collects microseismic signals through a data acquisition device;

[0115] L2 uses the data preprocessing module to perform fast Fourier transform (FFT) processing on the microseismic signal to obtain the corresponding frequency domain sequence signal; and splices the original microseismic signal and the corresponding frequency domain sequence signal to form dual-channel microseismic signal input data;

[0116] L3 inputs the dual-channel microseismic signal input data into the deep variational coder-classifier to obtain the classification result of the original microseismic signal.

[0117] The DVEC-based microseismic signal spectrum analysis system and method provided by the present invention can process and analyze complex vibration signals from sensors, which is of great significance for ensuring the safety of water conservancy and hydropower facilities. During the implementation process, the vibration signals generated by the water conservancy and hydropower facilities during operation are first captured by a high-precision sensor array. These signals may contain subtle changes caused by seismic activity, and accurate analysis of these signals is the key to preventing potential disasters. Then, the time domain signal is converted into a frequency domain signal using the designed preprocessing strategy, and the dual-channel data in the time domain and frequency domain are combined to enhance the model's learning and recognition capabilities for the input signal. The key to the DVEC network model lies in the design of its loss function, which combines variational Bayesian reasoning and supervised learning to approximate the true posterior distribution of the data by maximizing the lower bound of the log-likelihood of the data. This design allows the model to optimize the distribution of latent features through the KL divergence regularization term while reconstructing the input signal, thereby improving the model's recognition ability for known and unknown seismic signal types. In addition, the DVEC network model further optimizes the loss function through an adaptive weight strategy, which dynamically adjusts the weights according to the classification results of the samples, so that the network model pays more attention to those samples near the classification boundary. This method is particularly suitable for the water conservancy and hydropower industry, because the safety of dams and hydropower stations requires a high degree of vigilance for all possible seismic activities, even those uncommon or unknown earthquake patterns. The implementation of the method of the present invention in water conservancy and hydropower facilities can significantly improve the monitoring and early warning capabilities of seismic activities. Through real-time analysis of vibration signals, the DVEC model can quickly identify seismic events that may pose a threat to the stability of the dam, thereby providing a scientific basis for taking emergency response measures. In addition, the high robustness and adaptability of the method also make it an effective tool for long-term and stable monitoring of seismic signals in a constantly changing environment.

Claims

1. A microseismic signal spectrum analysis system based on DVEC, characterized in that: include: A data acquisition device for collecting microseismic signals; The data preprocessing module is used to perform fast Fourier transform (FFT) processing on the microseismic signals to obtain the corresponding frequency domain sequence signals; and to splice several original microseismic signals in the time domain and the corresponding frequency domain sequence signals in a dual-channel manner to form dual-channel microseismic signal input data; A deep variational coding-classifier is used to extract features from dual-channel microseismic signal input data and classify the signal based on the extracted features; the deep variational coding-classifier includes an encoder and a classifier; the encoder is used to obtain latent features based on the dual-channel microseismic signal input data; The classifier is used to obtain the classification category of each original microseismic signal according to the latent features; 2. The DVEC-based microseismic signal spectrum analysis system according to claim 1, characterized in that: The encoder includes several convolution modules, several convolution layers and two fully connected layers; the convolution module includes a convolution layer and a pooling layer.

3. The DVEC-based microseismic signal spectrum analysis system according to claim 1 or 2, characterized in that: In the deep variational coding-classifier, the classifier is composed of several fully connected layers.

4. The DVEC-based microseismic signal spectrum analysis system according to claim 3, characterized in that: In the deep variational coding-classifier, the total loss function is: L=L CE -λ(x)D KL ; Among them, L CE is the classification loss, and the cross entropy loss function is used; λ(x) represents the introduced adaptive weight, x represents the dual-channel microseismic signal spectrum input data, D KL represents the KL divergence regularization term.

5. The DVEC-based microseismic signal spectrum analysis system according to claim 4, characterized in that: in, Classification loss L CE Specifically expressed as: The expected value of p ξ (y|z) represents the category probability obtained by the classifier.

6. The DVEC-based microseismic signal spectrum analysis system according to claim 4, characterized in that: KL divergence regularization term D KL It is expressed as: Among them, q φ (z|x) represents the posterior distribution, p θ (z) represents the hidden characteristic distribution of microseismic signals, and L represents the length of z.

7. The DVEC-based microseismic signal spectrum analysis system according to claim 4, characterized in that: The deep variational coding-classifier also includes an adaptive weight adjustment module, which is expressed as: Among them, ω0 represents the predefined weight parameter; y represents the sample x l The true label of Represents the deep variational encoding-classifier for sample x l The predicted label of ; ω(x) represents the cumulative density function of the Weibull distribution.

8. The training method of the DVEC-based microseismic signal spectrum analysis system according to any one of claims 1 to 7, characterized in that: Divide the sample data for training into several batches, and execute each batch according to the following steps: S1 uses the data preprocessing module to perform fast Fourier transform (FFT) processing on the original microseismic signals in the batch of sample data to obtain corresponding frequency domain sequence signals; and splices the original microseismic signals in the sample data and the corresponding frequency domain sequence signals to form dual-channel microseismic signal input data; S2 inputs the dual-channel microseismic signal input data into the deep variational coding-classifier to obtain the KL divergence regularization term, and combines the label data in the sample data to obtain the classification loss; S3 calculates the total loss based on the classification loss and KL divergence loss; S4 optimizes the encoder and classifier parameters through the total loss; Return to step S1, repeat steps S1-S4, and train the system through each batch of sample data until the system converges.

9. The DVEC-based microseismic signal spectrum analysis system according to claim 8, characterized in that: In step S4, the encoder and classifier parameters are optimized by a gradient descent optimization algorithm.

10. A method for analyzing the spectrum of microseismic signals based on DVEC, characterized in that: Using the system according to any one of claims 1 to 7, the following steps are performed: L1 collects microseismic signals through a data acquisition device; L2 uses the data preprocessing module to perform fast Fourier transform (FFT) processing on the microseismic signal to obtain the corresponding frequency domain sequence signal; and splices the original microseismic signal and the corresponding frequency domain sequence signal to form dual-channel microseismic signal input data; L3 inputs the dual-channel microseismic signal input data into the deep variational coder-classifier to obtain the classification result of the original microseismic signal.

Images