Signal processing method and device based on XGBoost hybrid multi-step prediction

Abstract

The application provides a signal processing method and device based on XGBoost hybrid multi-step prediction. The method comprises the following steps: extracting a first signal before acoustic emission arrival and a second signal superimposed by acoustic emission and noise from a monitoring signal; constructing a hybrid multi-step prediction model comprising a plurality of XGBoost models and trained by the first signal based on the duration of the second signal; inputting the first signal into the trained model to predict noise, obtaining predicted noise; and finally removing the predicted noise from the second signal to obtain an acoustic emission signal. The whole process predicts the noise signal after acoustic emission arrival by the characteristics of the noise signal before acoustic emission arrival, and removes the noise signal to achieve the purpose of noise reduction. The noise prediction is relatively accurate, does not need to obtain acoustic emission characteristics in advance, can avoid damage to information such as frequency and amplitude of the acoustic emission signal, has excellent noise reduction effect, and is helpful for accurate ground pressure monitoring and analysis.

Description

Technical Field

[0001] This application relates to the field of ground pressure monitoring technology, and in particular to a signal processing method and apparatus based on XGBoost hybrid multi-step prediction. Background Technology

[0002] In underground mining, the excavation of ore bodies causes a redistribution of stress in the original rock, leading to stress concentration and subsequent rock failure. This results in instability in the goaf, causing ground pressure hazards. As mining activities progress deeper, goafs become increasingly complex, and the resulting ground pressure hazards become more severe. Disastrous ground pressure in complex goafs is a typical major mine hazard, seriously affecting normal mine production and surrounding safety, and is also a major type of mine disaster impacting social stability. Monitoring, predicting, forecasting, and controlling disastrous ground pressure in complex goafs has always been a research hotspot and cutting-edge topic in the fields of safety engineering and mining engineering both domestically and internationally.

[0003] Currently, real-time ground pressure monitoring primarily employs acoustic emission (AE) signal monitoring. Acoustic emission refers to the phenomenon where a localized source in a material rapidly releases energy to generate transient elastic waves. However, in underground mines, the ground pressure monitoring environment is complex and variable. Acoustic emission sensors are affected by numerous interference sources, and the monitoring signals obtained during AE cannot accurately reflect the damage state of the surrounding rock, easily leading to misjudgments of the ground pressure situation. Therefore, effective noise reduction of the monitoring signals is essential. Traditional noise reduction methods mainly include amplitude filtering, frequency domain filtering, wavelet filtering, probabilistic filtering, and empirical mode decomposition (EMD). These methods are essentially based on the existence of extractable features in the acoustic emission signal (also known as the useful signal) that are clearly distinct from noise (also known as the useless signal). They involve deconstructing and reconstructing the monitoring signal, predicting the acoustic emission signal, and using the differences in the deconstructed components to limit noise, thereby achieving noise reduction.

[0004] However, acoustic emission is inherently unstable due to the fracturing of the surrounding rock, making it difficult to extract stable features. Neither frequency nor amplitude can serve as effective features to distinguish acoustic emission signals from noise. Furthermore, the acoustic emission in the monitoring signal is itself contaminated and mixed with noise, making it difficult to reflect its original appearance. Consequently, the extraction of effective features becomes even more difficult. Therefore, existing noise reduction methods for monitoring signals are ineffective, resulting in poor accuracy in ground pressure monitoring. Summary of the Invention

[0005] This application provides a signal processing method and apparatus based on XGBoost hybrid multi-step prediction, which can be used to solve the technical problems of poor noise reduction effect and poor accuracy of ground pressure monitoring in existing monitoring signals.

[0006] In a first aspect, embodiments of this application provide a signal processing method based on XGBoost hybrid multi-step prediction, the method comprising:

[0007] The first signal and the second signal are extracted from the monitoring signal. The first signal is the pure noise signal before the acoustic emission arrives, and the second signal is the signal of the acoustic emission and the noise superimposed. The duration of the second signal is k time steps.

[0008] A hybrid multi-step prediction model is constructed. The hybrid multi-step prediction model is built based on k XGBoost models and trained using the first signal. The hybrid multi-step prediction model is used to obtain noise prediction data for each prediction time step based on the first signal. Each prediction time step is a time step in the second signal.

[0009] After processing the first signal, it is input into the hybrid multi-step prediction model to obtain noise prediction data for each prediction time step.

[0010] A noise prediction signal is generated based on the noise prediction data at each prediction time step, and the noise prediction signal is removed from the second signal to obtain the acoustic emission signal.

[0011] In conjunction with the first aspect, in one possible implementation of the first aspect, extracting the first signal and the second signal from the monitoring signal includes:

[0012] The arrival and end times of acoustic emissions are determined from the monitoring signals;

[0013] The monitoring signal located between the arrival and end times of the acoustic emission is identified as the second signal;

[0014] The first signal is extracted from the monitoring signal preceding the second signal. The duration of the first signal is n time steps, and the first signal and the second signal constitute a signal that is continuous in the time domain for n+k time steps, where n is the second multiple of k.

[0015] In conjunction with the first aspect, in one feasible implementation of the first aspect, a hybrid multi-step prediction model is constructed, including:

[0016] A multi-output prediction model consisting of k XGBoost models is constructed. The multi-output prediction model is used to generate a k-row, k-column output matrix, and the average value of each prediction element corresponding to the same prediction time step in the output matrix is ​​used as the final prediction value of the prediction time step. Each XGBoost model is used to predict each prediction element in a corresponding column vector of the output matrix. The j-th column vector of the output matrix is ​​composed of the prediction elements of the corresponding XGBoost model based on the input matrix, which are j-lag time steps behind. j is a positive integer less than or equal to k. The input matrix is ​​a k-row, m+1-column matrix constructed by converting n time series data into supervised learning data based on each time step in the second signal. m is the first multiple of k.

[0017] The training set is obtained from the first signal, and the hyperparameters of the multi-output prediction model are tuned using the training set to obtain the tuned multi-output prediction model.

[0018] The optimized multi-output prediction model was determined to be a hybrid multi-step prediction model.

[0019] In conjunction with the first aspect, in one possible implementation of the first aspect, the input matrix is ​​constructed as follows:

[0020] Based on each time step in the second signal, the m+1 elements between the (nk-m+1)th time step and the (n-k+1)th time step are determined as the input feature vector of the first row;

[0021] Based on the first signal and the input feature vector of the first row, the input feature vectors of the remaining rows are constructed using the hysteresis method.

[0022] The k rows of input feature vectors together constitute the input matrix.

[0023] In conjunction with the first aspect, in one possible implementation of the first aspect, the output matrix is ​​represented as follows:

[0024]

[0025] Where Pred is the output matrix, x n-k+2 For the prediction element at time step (n-k+2), x n+k Let n be the prediction element at the (n+k)th time step, where n is the duration of the first signal and k is the duration of the second signal.

[0026] In conjunction with the first aspect, in one possible implementation of the first aspect, the prediction elements corresponding to the same prediction time step in the output matrix satisfy the following condition:

[0027] i+j=g, g∈[k+1, 2k], i, j∈[1, k]

[0028] Where i is the row number of the predicted element corresponding to the prediction time step, j is the column number of the predicted element corresponding to the prediction time step, k is the total number of rows and columns of the output matrix, and the prediction time step is greater than or equal to n+1 and less than or equal to n+k.

[0029] In conjunction with the first aspect, in one possible implementation of the first aspect, a training set is obtained from the first signal, and the hyperparameters of the multi-output prediction model are tuned using the training set to obtain the tuned multi-output prediction model, including:

[0030] The training set is obtained from the first signal, and the hyperparameters of each XGBoost model are optimized individually using the training set to obtain the optimal hyperparameters of each XGBoost model.

[0031] The hyperparameters of each XGBoost model are updated to the optimal hyperparameters to obtain the tuned multi-output prediction model.

[0032] In conjunction with the first aspect, in one possible implementation of the first aspect, the hyperparameters of each XGBoost model are individually optimized using the training set to obtain the optimal hyperparameters for each XGBoost model, including:

[0033] For each XGBoost model, the optimal solution is obtained by optimizing each hyperparameter of the XGBoost model based on a preset genetic algorithm.

[0034] The optimal solution is determined as the optimal hyperparameter of the XGBoost model.

[0035] In conjunction with the first aspect, in one possible implementation of the first aspect, the hyperparameters of the XGBoost model are optimized based on a pre-defined genetic algorithm, including:

[0036] The hyperparameters of the XGBoost model are converted into the operation units of the preset genetic algorithm through binary encoding.

[0037] The sum of the mean squared error (MSE) values ​​from each forward validation is used as the fitness function of the genetic algorithm. The genetic algorithm performs a preset number of genetic iterations on each operational unit. The training and test data used in each forward validation are obtained from the first signal. The MSE value of each forward validation is the predicted value obtained by the XGBoost model based on the corresponding training data and the MSE value of the corresponding test data.

[0038] Secondly, embodiments of this application provide a signal processing apparatus based on XGBoost hybrid multi-step prediction, the apparatus comprising:

[0039] The extraction unit is used to extract a first signal and a second signal from the monitoring signal. The first signal is a pure noise signal before the arrival of acoustic emission, and the second signal is a signal superimposed on the acoustic emission and noise. The duration of the second signal is k time steps.

[0040] The modeling unit is used to construct a hybrid multi-step prediction model. The hybrid multi-step prediction model is built based on k XGBoost models and trained using the first signal. The hybrid multi-step prediction model is used to obtain noise prediction data for each prediction time step based on the first signal. Each prediction time step is a time step in the second signal.

[0041] The prediction unit is used to process the first signal and input it into the hybrid multi-step prediction model to obtain noise prediction data for each prediction time step.

[0042] The noise reduction unit is used to generate a noise prediction signal based on the noise prediction data at each prediction time step, and remove the noise prediction signal from the second signal to obtain the acoustic emission signal.

[0043] Thus, this application provides a signal processing method and apparatus based on XGBoost hybrid multi-step prediction. In the signal processing method, a first signal before the arrival of acoustic emission and a second signal superimposed with noise are extracted from the monitoring signal. A hybrid multi-step prediction model including k XGBoost models is constructed based on the duration k of the second signal. This multi-output prediction model is trained using the first signal. After data processing, the first signal is input into the hybrid multi-step prediction model for prediction, obtaining noise prediction data for each prediction time step, thus achieving noise prediction within the time range of the second signal. Finally, the predicted noise signal is removed from the second signal to obtain the acoustic emission signal. The entire method predicts the noise signal after the arrival of acoustic emission based on the characteristics of the noise signal before the arrival of acoustic emission, and then eliminates the predicted noise signal from the signal after the arrival of acoustic emission, thereby achieving the noise reduction purpose of the second signal. The noise signal prediction is relatively accurate, does not require prior acquisition of acoustic emission characteristics, and avoids damage to the frequency, amplitude, and other information of the acoustic emission signal, resulting in excellent noise reduction effect and contributing to more accurate ground pressure monitoring and analysis. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the workflow of a signal processing method based on XGBoost hybrid multi-step prediction provided in an embodiment of this application;

[0045] Figure 2 This is a schematic diagram illustrating an example of a monitoring signal for acoustic emission;

[0046] Figure 3 This is a schematic diagram of the hysteresis method used to construct the input matrix in the embodiments of this application;

[0047] Figure 4 This is a schematic diagram of the multi-output prediction model performing multi-output prediction in the embodiments of this application;

[0048] Figure 5 This is a schematic diagram illustrating the specific process of hyperparameter optimization based on genetic algorithms in the embodiments of this application;

[0049] Figure 6 This is a schematic diagram of the forward verification process in an embodiment of this application;

[0050] Figure 7 This is a schematic diagram of the specific process of the signal processing method based on XGBoost hybrid multi-step prediction provided in the embodiments of this application;

[0051] Figure 8 This is a comparison chart of useful signal denoising under noise with a signal-to-noise ratio of 3dB.

[0052] Figure 9 This is a comparison chart of useful signal denoising under 5dB noise level.

[0053] Figure 10 This is a comparison chart of useful signal denoising under noise with a signal-to-noise ratio of 10dB.

[0054] Figure 11 This is a comparison chart of useful signal denoising under noise conditions with a signal-to-noise ratio of 20dB.

[0055] Figure 12 This is a schematic diagram illustrating an example of the spectrum of the original measured signal.

[0056] Figure 13 yes Figure 12 A schematic diagram of the spectrum of a pure noise signal;

[0057] Figure 14 These are schematic diagrams of the signals before and after noise reduction using a hybrid multi-step prediction model;

[0058] Figure 15 This is a schematic diagram of the signal spectrum after noise reduction using a hybrid multi-step prediction model;

[0059] Figure 16 This is a schematic diagram of the signal before and after wavelet denoising;

[0060] Figure 17 This is a schematic diagram of the signal spectrum after wavelet denoising;

[0061] Figure 18 This is a schematic diagram of the signal before and after EMD noise reduction;

[0062] Figure 19 This is a schematic diagram of the signal spectrum after EMD noise reduction;

[0063] Figure 20 This is a schematic diagram of a signal processing device based on XGBoost hybrid multistep prediction provided in an embodiment of this application. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0065] Traditional noise reduction methods for monitoring signals mainly include amplitude filtering, frequency domain filtering, wavelet filtering, probabilistic filtering, and EMD. Amplitude filtering involves statistically predicting changes in signal amplitude, and includes methods such as amplitude limiting filtering, median filtering, and mean filtering. However, in actual mines, background noise is complex, variable, and non-stationary, making amplitude filtering less flexible and accurate, and thus rarely used in practical acoustic emission monitoring. Frequency domain filtering is a commonly used noise reduction method in acoustic emission monitoring. It primarily uses Fourier transform to convert the time-domain signal into a frequency-domain signal, and then performs filtering in the frequency domain. Classic filters include Butterworth filters and Chebyshev filters. However, this method requires that the signal and noise spectra not overlap to completely separate the signal and noise. In actual mine monitoring, however, the noise signal spectrum and noise spectrum often overlap, making it difficult to effectively eliminate the influence of noise in the frequency domain. Wavelet filtering, employing a variable wavelet basis, addresses the shortcomings of frequency domain filtering in processing non-stationary signals. However, wavelet filtering requires selecting a suitable wavelet basis based on the noise environment during noise reduction, and different wavelet bases yield varying noise reduction effects. Therefore, it lacks sufficient adaptability in mine acoustic emission monitoring. Probabilistic filtering, a widely used noise reduction method in communication signal processing in recent years, primarily predicts future signal points through mathematical statistics on past signal data, thus achieving noise reduction. Methods such as particle filtering, Kalman filtering, and Wiener filtering possess excellent adaptability. However, in mines, acoustic emission, as a useful signal, exhibits uncertainties in its morphology and timing, making it difficult to establish a complete state equation. Consequently, it struggles to achieve effective noise reduction in actual acoustic emission monitoring. EMD (Effective Mode Decomposition) is a widely used denoising method in the fields of seismic, microseismic, and acoustic emission signal processing in recent years. It continuously separates noise and useful signals according to different frequency bands based on the characteristics of the signal itself, solving the difficulties in wavelet basis selection and the limitations of the uncertainty principle in wavelet filtering. However, EMD decomposition suffers from mode aliasing, meaning that an IMF (Intrinsic Mode Function) component cannot guarantee that it has only one time-scale characteristic component. This is a problem inherent to the signal itself, but also to the EMD algorithm. Furthermore, EMD decomposition requires many iterations, and there is no standard for stopping the iterations. Therefore, different stopping conditions result in different IMFs, leading to varying denoising effects. In summary, existing denoising methods require sufficient prior knowledge of the useful signal (i.e., the acoustic emission signal), and the useful signal needs to possess extractable features that are clearly distinguishable from noise.However, acoustic emission is inherently unstable due to the fracturing of the surrounding rock, making it difficult to extract stable features. Neither frequency nor amplitude can serve as effective features to distinguish acoustic emission signals from noise. Furthermore, the acoustic emission in the monitoring signal is itself contaminated and mixed with noise, making it difficult to reflect its original appearance. Consequently, the extraction of effective features becomes even more difficult, resulting in poor noise reduction.

[0066] Based on the above introduction, the applicant considers noise to be relative to useful signals, while useful signals are "useless signals" of a noise system relative to noise. Although acoustic emission occurs for a short time in actual monitoring signals, noise exists constantly, so its studyable characteristics are far more numerous than acoustic emission. Therefore, to address the problems of poor noise reduction effects and low accuracy in ground pressure monitoring of existing monitoring signals, this application discloses a signal processing method based on XGBoost hybrid multi-step prediction through the following embodiments. The signal processing method provided in this application extracts the local region where acoustic emission occurs from the monitoring signal. In the acoustic emission region, acoustic emission and noise are superimposed. The noise in this region can be considered unknown and can be predicted using the data characteristics of the pure noise region. This predicted noise is then eliminated to achieve a noise reduction effect, providing a basis for subsequent parameter analysis and ground pressure early warning, and significantly improving the noise reduction effect and the accuracy of ground pressure monitoring. Specifically, the prediction of noise in the acoustic emission region of the monitoring signal can be achieved by establishing a noise prediction model. (Refer to...) Figure 1 The flowchart shown illustrates the signal processing method based on XGBoost hybrid multi-step prediction provided in this application embodiment, which specifically includes the following steps:

[0067] 101: Extract the first signal and the second signal from the monitoring signal.

[0068] The first signal is a pure noise signal before the arrival of acoustic emission, the second signal is a signal of acoustic emission and noise superimposed, and the duration of the second signal is k time steps.

[0069] In some implementations, the first and second signals can be extracted from the monitoring signals in the following ways:

[0070] Step 1: Determine the arrival and end times of acoustic emission from the monitoring signals.

[0071] Step two: The monitoring signal located between the arrival and end times of the acoustic emission is identified as the second signal.

[0072] Step 3: Extract the first signal from the monitoring signal located before the second signal.

[0073] The duration of the first signal is n time steps, and the first and second signals constitute a time-domain continuous signal for n+k time steps, where n is a multiple of k. Since the first signal is used for subsequent prediction, n should focus on the correlation and periodicity between data, while also taking into account the efficiency of the algorithm. For example, n can be 5-7 times the duration of acoustic emission.

[0074] Figure 2 This is a schematic diagram illustrating an example of acoustic emission monitoring signals. (Refer to...) Figure 2 As shown, in the monitoring signal of acoustic emission from rock mass in underground mines, A is the first signal, B is the second signal, the duration is k time steps, t1 is the arrival time of acoustic emission from rock mass, t2 is the end time of acoustic emission from rock mass, and a is the upper envelope of the monitoring signal.

[0075] 102: Construct a hybrid multi-step prediction model.

[0076] The hybrid multi-step prediction model is built upon k XGBoost models and trained using the first signal. It is used to obtain noise prediction data for each prediction time step based on the first signal, where each prediction time step is a time step within the second signal.

[0077] In some implementations, a hybrid multi-step prediction model can be constructed using the following steps:

[0078] Step 1: Build a multi-output prediction model that includes k XGBoost models.

[0079] The multi-output prediction model generates a k-row, k-column output matrix and uses the average of the predicted elements at the same prediction time step as the final predicted value for that time step. Each XGBoost model predicts only one predicted element in a corresponding column vector of the output matrix. The j-th column vector of the output matrix is ​​composed of the predicted elements j time steps behind, obtained by the corresponding XGBoost model based on the input matrix, where j is a positive integer less than or equal to k. The input matrix is ​​a k-row, m+1-column matrix constructed by converting n time series data into supervised learning data based on each time step in the second signal, where m is a multiple of k. For example, m can be 2-3 times the acoustic emission duration.

[0080] A common method in signal prediction is to establish a regression model. The XGBoost model in this embodiment uses the XGBoost regression algorithm, which is currently the fastest and best open-source Boosted tree algorithm. It adds a regularization term to the loss function and performs a second-order Taylor expansion on the loss function, thereby improving the calculation accuracy and effectively avoiding overfitting. Moreover, it uses feature column sampling, which can not only further improve the fitting effect but also reduce the computational load of the algorithm.

[0081] The input matrix and data matrix are described below.

[0082] Since the prediction of acoustic emission noise signals is essentially a time series prediction, transforming the time series prediction problem into a supervised learning problem has always been a common method for machine learning in time series prediction. In constructing a supervised learning problem for time series, the input features are past data, and by utilizing the inherent relationships between the order of data, the output is the predicted data for the corresponding future steps. In this embodiment, a lag method is used to construct the input feature vector for acoustic emission noise signal prediction, and then multiple input feature vectors are combined to form an input matrix. Specifically, for a monitoring signal sequence X of length n+k composed of the first and second signals, the values ​​V(n+1), V(n+2), ..., V(n+k) of any time step from n+1 to n+k can be predicted by lagging the corresponding time step s (s∈[n+1, n+k]) to construct input vectors (Var(nm), Var(n-m+1), ..., Var(n)).

[0083] In some implementations, the method for constructing the input matrix may include the following steps:

[0084] First, based on each time step in the second signal, m+1 elements from the (nk-m+1)th time step to the (n-k+1)th time step are determined as the input feature vector of the first row. Then, based on the first signal and the input feature vector of the first row, the input feature vectors of the remaining rows are constructed using the lag method. Finally, the k rows of input feature vectors together constitute the input matrix. In other words, the embodiments of this application can construct the input matrix by successively lagging each element in the monitoring signal sequence X by one step.

[0085] Figure 3 This is a schematic diagram of the lag method used to construct the input matrix in the embodiments of this application. (Refer to...) Figure 3 As shown, the input feature vector in the first row of the input matrix is ​​[x n-k+1-m x n-k+2-m , ..., x n-k-1 x n-k x n-k+1The input feature vector for the second row is obtained by delaying the input vector by one step. n-k+2-m x n-k+3-m , ..., x n-k x n-k+1 x n-k+2 The input feature vector of the k-th row is [x], ..., ... n-m x n-m+1 , ..., x n-2 x n-1 x n ]. Where, x n-k+1-m To monitor the element corresponding to the (nk-m+1)th time step in the signal sequence X, x n To monitor the elements corresponding to the nth time step in the signal sequence X, V(tm), ..., V(t-3), V(t-2), V(t-1), and V(t) represent the column vectors arranged in chronological order. The elements obtained by lag time step s of the input feature vector in each row are x. n-k+1+s x n-k+2+s x n-k+3+s ... x n+s These elements together form the column vector V(s).

[0086] In some implementations, the output matrix obtained by predicting the input matrix is ​​represented by the following formula (1):

[0087]

[0088] In formula (1), Pred is the output matrix, x n-k+2 To monitor the prediction element at the (n-k+2)th time step in the signal sequence X, x n+k The prediction element is for the (n+k)th time step of the monitoring signal sequence X, where n is the duration of the first signal and k is the duration of the second signal.

[0089] Specifically, the prediction elements corresponding to the same prediction time step in the output matrix satisfy the following conditions:

[0090] i+j=g, g∈[k+1, 2k], i、j∈[1,k] Formula (2)

[0091] In formula (2), i is the row number of the predicted element corresponding to the prediction time step, j is the column number of the predicted element corresponding to the prediction time step, and k is the total number of rows and columns of the output matrix. The prediction time step is greater than or equal to n+1 and less than or equal to n+k. It can be seen from formulas (1) and (2) that the predicted element corresponding to the prediction time step is located on the lower right diagonal of the output matrix.

[0092] In this embodiment, noise prediction is a multi-step prediction method, which uses past data to predict data multiple steps into the future. Each XGBoost model is used to predict only one predicted element in a corresponding column vector of the output matrix. For example, the first XGBoost model predicts only the predicted element in the column vector with a lag of one time step, the second XGBoost model predicts only the predicted element in the column vector with a lag of two time steps, and the kth XGBoost model predicts only the predicted element in the column vector with a lag of k time steps. Since each XGBoost model only predicts one noisy data point within k steps, k different XGBoost models are needed for the (n+1)th to (n+k)th prediction time steps (a total of k). These k different XGBoost models together constitute the entire multi-output prediction model. In this way, the advantages of traditional direct prediction strategies and multi-output strategies can be combined to create a new prediction method that differs from traditional prediction methods (e.g., direct multi-step prediction strategies, recursive multi-step prediction strategies, direct-recursive hybrid multi-step prediction strategies, and multi-output prediction strategies), resulting in more accurate and effective predictions.

[0093] Figure 4 This is a schematic diagram illustrating the multi-output prediction model performing multi-output prediction in an embodiment of this application. (Refer to...) Figure 4 As shown, model1, model2, ..., modelk-1, and modelk represent k XGBoost models in the multi-output prediction model. Each of the k XGBoost models corresponds one-to-one with a column vector of the output matrix. The j-th XGBoost model can predict the j-th column vector V(t+j) of the output matrix based on the input matrix. Each predicted element in the j-th column vector V(t+j) is an element of the input matrix predicted by the corresponding j-th XGBoost model, lagged by j time steps, where j is a positive integer less than or equal to k. For example, the first column vector V(t+1) is a predicted element of the input matrix predicted by the first XGBoost model, lagged by 1 time step; the second column vector V(t+2) is a predicted element of the input matrix predicted by the second XGBoost model, lagged by 2 time steps; ..., the k-th column vector V(t+k) is a predicted element of the input matrix predicted by the k-th XGBoost model, lagged by k time steps. After generating the output matrix, the XGBoost model may output different values ​​for each prediction time step because the elements on the lower right diagonal of the output matrix are predicted values ​​for the same prediction time step. These predicted values ​​correspond to input feature vectors in different rows. In order to ensure the accuracy of multi-step prediction and the continuity between data, this embodiment takes the average value of the elements on the lower right diagonal of the output matrix as the predicted value for each prediction time step.

[0094] Step 2: Obtain the training set from the first signal and use the training set to perform hyperparameter tuning on the multi-output prediction model to obtain the tuned multi-output prediction model.

[0095] Specifically, in the embodiments of this application, the multi-output prediction model mainly relies on the XGBoost algorithm to predict data. XGBoost is a type of machine learning algorithm. In order to ensure the accuracy of prediction, it needs to be trained using existing data before prediction to achieve hyperparameter tuning.

[0096] In some implementations, step two can involve hyperparameter tuning of the multi-output prediction model using the following steps:

[0097] The first step is to obtain the training set from the first signal and use the training set to optimize the hyperparameters of each XGBoost model individually to obtain the optimal hyperparameters for each XGBoost model.

[0098] Specifically, XGBoost is an ensemble algorithm based on gradient boosting trees. Its hyperparameters can be controlled in three core ways: the ensemble algorithm itself, the weak evaluator used for ensemble, and other processes used in the application. The first two parts contain the core principles and mathematical processes of XGBoost, while the last part is mainly used in its application. XGBoost's regression prediction capability is primarily adjusted by the first two parts; therefore, this embodiment focuses on optimizing the hyperparameters of the first two parts. Specific hyperparameters are shown in Appendix Table 1.

[0099] Table 1: Hyperparameters of XGBoost

[0100] Hyperparameter name illustrate Optimization range accuracy n_estimators Number of weak classifiers [10，1500] 10 learning_rate Learning rate [0，1] 0.01 max_child_weight Leaf nodes require minimum sample weights and sums. [0，15] 0.1 max_depth Maximum depth of tree [1，50] 1 gamma Minimum loss function decrease required for node splitting (0，20] 0.01 subsample The sample ratio used to construct each tree (0，1] 0.01 reg_lambda L1 regularization penalty coefficient [0，5] 0.1 Reg_alpha L2 regularization penalty coefficient [0，5] 0.1

[0101] In XGBoost regression, eight hyperparameters as shown in Table 1 need to be adjusted. Different hyperparameters have different functions, which makes the optimization of XGBoost hyperparameters an extremely complex problem.

[0102] In some implementations, the hyperparameters of each XGBoost model are individually optimized using the training set to obtain the optimal hyperparameters for each XGBoost model. This can be done in the following ways:

[0103] For each XGBoost model, a pre-defined genetic algorithm (GA) is used to optimize each hyperparameter to obtain the optimal solution. This optimal solution is then determined as the optimal hyperparameters of the XGBoost model.

[0104] Traditional XGBoost hyperparameter tuning often employs grid search, but this method suffers from poor optimization, is prone to getting stuck in local optimization, and requires numerous iterations when there are many parameters. Genetic algorithms, computer simulation algorithms based on evolutionary theory and genetics, are widely used for hyperparameter optimization because their global search optimization speed is unmatched by grid search. This application employs a genetic algorithm to optimize the hyperparameters of XGBoost, thereby ensuring the accuracy of the multi-output prediction model based on XGBoost.

[0105] Specifically, in this embodiment, the hyperparameters of the XGBoost model are first converted into preset genetic algorithm operation units using binary encoding. Then, the sum of the mean squared error (MSE) values ​​from each forward validation is used as the fitness function of the genetic algorithm, and the operation units of the genetic algorithm undergo a preset number of genetic iterations. The training and test data used in each forward validation are obtained from the first signal, and the MSE value for each forward validation is the sum of the predicted value obtained by the XGBoost model based on the corresponding training data and the MSE value of the corresponding test data. For example, the preset number of iterations can be set to 100-200.

[0106] The following is a detailed, exemplary introduction to genetic algorithms.

[0107] Genetic algorithms use genes as the unit of operation. In this embodiment, during hyperparameter optimization, the eight hyperparameters of XGBoost are converted into eight genes using binary encoding. These eight genes constitute a chromosome (hyperparameter combination), and 20 chromosomes constitute a population. In this embodiment, the total MSE (Mean Sequence Size) during XGBoost model training is used as the fitness function (optimization objective). The smaller the MSE of the prediction set, the higher the fitness, and the better the corresponding hyperparameter combination. Chromosome crossover mutation is allowed in this embodiment.

[0108] Figure 5 This is a schematic diagram illustrating the specific process of hyperparameter optimization based on genetic algorithms in an embodiment of this application. (Refer to...) Figure 5As shown, in one example, the generation number GEN starts from 0, and an initial population is generated based on the initial hyperparameters. It is determined whether GEN is greater than 200. If GEN is less than or equal to 200, the fitness of each individual is calculated (i.e., the total MSE is calculated). The number of generations starts from 0, and it is determined whether the number of generations is equal to 20. If the number of generations is not equal to 20, a genetic operation (roulette wheel) is selected based on probability. Genetic operations include replication, crossover, and mutation. During replication, an individual is selected to replicate, then the individual is added to the new population, and the generation number is incremented by one. During crossover, two individuals are selected, their generation number is incremented by one, and then crossbreeding is performed. The two offspring are inserted into the new population, and the generation number is incremented again. During mutation, an individual is selected to mutate, then the individual is inserted into the new population, and the generation number is incremented by one. The process returns to the step of determining whether the generation number is equal to 20, and the above genetic operation steps are repeated until the generation number equals 20. GEN is then incremented by one, and the process returns to the step of determining whether GEN is greater than 200, repeating the above process until GEN is greater than 200, thus obtaining the optimal hyperparameters.

[0109] The following is an exemplary description of forward verification.

[0110] Hyperparameter optimization requires model evaluation, but methods like k-fold cross-validation with random dataset splitting are not suitable for time series prediction. Therefore, this application employs an alternative validation method: forward validation. Forward validation involves specifying a test length and progressively advancing the training data for testing and validation, scoring each validation attempt until the last portion of data is validated. For example, validation can be performed 2-3 times, with each validation using a training set length of 2-3 acoustic emission durations and a test set length of 1 acoustic emission duration. Each validation uses the data processing method (i.e., the input matrix construction method for converting the time series into supervised learning data as described above, which will not be repeated here) to process the training set. The MSE value between each predicted value and the test set is used as the score for each validation, and the sum of the MSE values ​​from all validations is used as the final evaluation metric (fitness function). Here, the MSE value refers to the expected value of the squared difference between the parameter estimate and the true parameter value.

[0111] Figure 6 This is a schematic diagram of the forward verification process in an embodiment of this application. (Refer to...) Figure 6As shown in the example, for the time series data in the first signal, three training sets and three corresponding test sets are divided. First, the XGBoost model is used to predict based on each training data 1 in the first training set, determining the predicted value and its corresponding MSE1 for each test data in the test set. Then, the XGBoost model is used to predict based on each training data 2 in the second training set, determining the predicted value and its corresponding MSE2 for each test data in the test set. Finally, the XGBoost model is used to predict based on each training data 3 in the third training set, determining the predicted value and its corresponding MSE3 for each test data in the test set. The sum of MSE1, MSE2, and MSE3 is then determined as the fitness function. Generally, the smaller the fitness function, the better the corresponding hyperparameter combination.

[0112] The second step is to update the hyperparameters of each XGBoost model to the optimal hyperparameters to obtain the tuned multi-output prediction model.

[0113] It should be noted that in the embodiments of this application, each XGBoost model is trained independently.

[0114] Step 3: Determine the optimized multi-output prediction model as a hybrid multi-step prediction model.

[0115] It should be noted that, since the duration of the second signal varies in different monitoring signals, the structure of the constructed multi-output prediction model also varies accordingly. After each construction, it is necessary to perform hyperparameter tuning using the time series signal of the first signal before obtaining the final hybrid multi-step prediction model used to predict the predicted values ​​of each time step in the second signal.

[0116] 103: After processing the first signal, input it into the hybrid multi-step prediction model to obtain noise prediction data for each prediction time step.

[0117] Specifically, data processing of the first signal refers to converting the time series data of the first signal into supervised learning data according to the structure of the input matrix. That is, filling in the corresponding elements according to the time step order of the input matrix. For example, if the input feature vector of the first row of the input matrix consists of m+1 elements between the (nk-m+1)th and (n-k+1)th time steps, then the m+1 elements in the first signal located between the (nk-m+1)th and (n-k+1)th time steps are determined as the input feature vector of the first row of the input matrix. For details, please refer to the relevant content above; it will not be repeated here.

[0118] 104: Generate a noise prediction signal based on the noise prediction data at each prediction time step, and remove the noise prediction signal from the second signal to obtain the acoustic emission signal.

[0119] To more clearly illustrate the overall flow of the signal processing method provided in the embodiments of this application, Figure 7 This is a schematic flowchart illustrating the signal processing method based on XGBoost hybrid multi-step prediction provided in this application embodiment. (Refer to...) Figure 7 As shown, firstly, the pure noise signal (i.e., the first signal) and the signal of the superposition of acoustic emission and noise (i.e., the second signal) are extracted from the monitoring signals collected by the acoustic emission monitoring system. Then, a hybrid multi-step prediction model architecture based on XGBoost multi-output is constructed. Next, the pure noise signal is divided into training sets, and the training sets are processed. Hyperparameter genetic algorithm tuning and forward validation are then performed to obtain the final hyperparameters. During tuning, the sum of the MSE (mean squared error) values ​​from the forward validation is used as the fitness function of the genetic algorithm, and a total of 100-200 genetic iterations are performed. During each forward validation, data processing is used to convert the model into supervised learning, resulting in a k-row, m+1-column input matrix. Then, the hybrid multi-step prediction model updated with the final hyperparameters is used to predict the noise in the acoustic emission region. Finally, the predicted noise is removed from the signal of the superposition of acoustic emission and noise to obtain the denoised acoustic emission signal.

[0120] Thus, in the signal processing method based on XGBoost hybrid multi-step prediction provided in this application embodiment, a first signal before the arrival of acoustic emission and a second signal superimposed with noise are extracted from the monitoring signal. A hybrid multi-step prediction model including k XGBoost models is constructed based on the duration k of the second signal. This multi-output prediction model is trained using the first signal. After data processing, the first signal is input into the hybrid multi-step prediction model for prediction, obtaining noise prediction data for each prediction time step. This achieves the prediction of noise within the time range of the second signal. Finally, the predicted noise signal is removed from the second signal to obtain the acoustic emission signal. The entire method predicts the noise signal after the arrival of acoustic emission based on the characteristics of the noise signal before the arrival of acoustic emission, and then eliminates the predicted noise signal from the signal after the arrival of acoustic emission, thereby achieving the noise reduction purpose of the second signal. The noise signal prediction is relatively accurate, does not require prior acquisition of acoustic emission characteristics, and avoids damage to the frequency, amplitude, and other information of the acoustic emission signal, resulting in excellent noise reduction effect and contributing to more accurate ground pressure monitoring and analysis.

[0121] To make the objectives, contents, and advantages of this invention clearer, the following description will be further combined with simulation experiments and practical application cases.

[0122] (1) Simulation test verification

[0123] To verify the noise reduction performance of the signal processing method provided in this application under different noise environments, the following dynamic simulation signal was generated using MATLAB 2017b to simulate the acoustic emission signal of the rock mass:

[0124]

[0125] In formula (3), Ay1 represents simulated background noise, the intensity and dominant frequency of which can be controlled by A and ω; y2 represents simulated acoustic emission signal, and the corresponding t g The acoustic emission arrival time is 700s in the simulation, the duration is 200s, rand is the white noise function, the sampling interval is 1s, and the sampling length is 1024s in the simulation.

[0126] In noisy environments, the signal-to-noise ratio (SNR) is typically used to reflect the relative intensity of the useful signal to the background noise. In simulations, different SNR values ​​are obtained mainly by adjusting the amplitudes A and ω. The SNR can usually be calculated using the following formula:

[0127]

[0128] In formula (4), x i x is the useful signal value. j Let M be the noise signal value, and M and N be the lengths of the useful signal and the noise signal, respectively.

[0129] In the simulation, four different noise environments (extremely strong noise environment, strong noise environment, general noise environment, and weak noise environment) were generated by controlling the signal-to-noise ratio (SNR) intensity, with SNRs of 3dB, 5dB, 10dB, and 20dB, respectively. Each noise environment was further divided into three types of noise with different dominant frequencies, resulting in a total of 12 noise environments. Within each noise environment, five different sets of random simulated signals were generated using the rand function, resulting in a total of 60 sets of signals. MSE and R² were used as evaluation metrics. MSE refers to the expected value of the squared difference between the denoised signal and the useful signal, which can be calculated using the following formula:

[0130]

[0131] In the formula, M is the length of the useful signal, and y 2t For useful signal values, This is the useful signal after noise reduction.

[0132] R 2 This represents the degree of fit between the noise reduction model and the real signal. The optimal value is 1, and the closer it is to 1, the better the noise reduction effect. The calculation formula is as follows:

[0133]

[0134] In formula (6), This represents the average value of the useful signal.

[0135] Based on the above simulation signal, different methods were used to denoise the useful signal, and the MSE and R0 of the denoised signals were compared. 2 For the statistical results of the average values, please refer to Appendix 2 and Appendix 3.

[0136] Table 2: MSE Statistics of Useful Signals Before and After Noise Reduction under Different Noise Environments

[0137]

[0138] Table 3: RA before and after noise reduction of useful signals under different noise environments 2 Average Statistics Table

[0139]

[0140]

[0141] Tables 2 and 3 show the MSE and R after denoising the useful signal using different methods. 2 The average statistical results show that wavelet denoising and EMD denoising are commonly used methods for denoising non-stationary signals. For wavelet denoising, db2 is chosen as the wavelet basis, a 5-level decomposition is performed, and soft thresholding is used for denoising. For EMD denoising, spurious components are removed based on prior knowledge of the useful signal and its correlation coefficient. Tables 2 and 3 show that under different noise environments, denoising through hybrid multi-step prediction can effectively reduce the MSE of the useful signal and reduce noise interference. The hybrid multi-step prediction denoising model performs better than the direct multi-step prediction denoising model and the multi-output multi-step prediction denoising model. The R values ​​before and after denoising are... 2 The values ​​show that the hybrid multi-step prediction denoising model, the direct multi-step prediction denoising model, and the multi-output multi-step prediction denoising model all improve the fit between the signal and the useful signal, although the R values ​​of the three models are relatively low. 2 Although the values ​​are relatively similar, it is clear that the hybrid multi-step prediction denoising model has a better fit. Compared to wavelet denoising and EMD denoising, the multi-step prediction denoising model shows higher MSE and RE values ​​after denoising. 2 The metrics are significantly better, and the MSE and R after wavelet denoising are also significantly better. 2 The indicator scores are significantly poor.

[0142] For a comparison of the noise reduction effect of useful signals under different noise environments, please refer to Appendix Table 4.

[0143] Table 4: Statistical Table of Noise Reduction Effect of Useful Signals under Different Noise Environments

[0144]

[0145]

[0146] As shown in Appendix Table 4, in the MSE evaluation index, the hybrid multi-step prediction model achieved a denoising effect of over 90%, with an average of 92.16%. While wavelet denoising and EMD denoising showed some denoising effect in low signal-to-noise ratio environments, they increased the MSE index in high signal-to-noise ratio environments (10dB, 20dB), resulting in negative denoising values ​​and greater signal deviation. In R... 2 In the metrics, because the useful and noisy signals are quite similar in high signal-to-noise ratio (SNR) environments (10dB, 20dB), the R² index before denoising is close to 1. Therefore, the hybrid multi-step prediction model denoising produces poor denoising performance in high SNR environments (10dB, 20dB), while the denoising effect is more obvious in low SNR environments (3dB, 5dB). Furthermore, in high SNR environments (10dB, 20dB), the R² index of wavelet denoising and EMD denoising is relatively high. 2 The metric exhibited the same negative noise reduction effect as the MSE metric. In summary, from MSE and R... 2 The metrics show that the hybrid multi-step prediction denoising model in this invention significantly outperforms the other two prediction denoising models, and also significantly outperforms traditional wavelet denoising and EMD denoising in terms of denoising effect. (Denoising effect (MSE index) = (MSE before denoising - MSE after denoising) / MSE before denoising, Denoising effect (R) 2 Indicator) = ((R after noise reduction) 2 - R before noise reduction 2 ) / R before noise reduction 2 )

[0147] Figure 8 This is a comparison chart of useful signal denoising under noise with a signal-to-noise ratio of 3dB. Figure 9 This is a comparison chart of useful signal denoising under 5dB noise level. Figure 10 This is a comparison chart of useful signal denoising under noise with a signal-to-noise ratio of 10dB. Figure 11 This is a comparison chart of useful signal denoising under 20dB noise conditions. (Refer to...) Figures 8 to 11 As shown, the signal denoised using the hybrid multi-step prediction model provided in this application has a small deviation from the useful signal, and the signal curves before and after denoising basically overlap. Compared to the attached... Figure 8 c-curve graph, appendix Figure 9 c-curve graph, appendix Figure 10 c-curve graph, appendix Figure 11The C-curve graph shows that, under different signal-to-noise ratio environments, the signals after wavelet and EMD denoising deviate significantly from the useful signal, especially in the low-amplitude portion. Wavelet denoising relies on the wavelet basis, the number of decomposition levels, and the choice of soft or hard thresholds, all of which depend on having sufficient knowledge of signal-to-noise signal differentiation (such as frequency or amplitude). When the difference between the useful and noise signals is small, wavelet denoising performs poorly. Figure 10 a-curve graph, appendix Figure 11 In the a-curve graph, the low-amplitude portion of the useful signal and the noise signal have little difference in amplitude, leading to... Figure 10 c-curve graph, appendix Figure 11 In the c-curve graph, the low-amplitude portion of the wavelet-denoised signal deviates significantly from the useful signal, which explains why the denoising effect is negative when the signal-to-noise ratio is high. When the noise and useful signals are similar in frequency and amplitude, EMD is prone to producing IMF components containing different time-scale components during signal sieving. Furthermore, the selection, sieving, and combination of IMF components in EMD denoising lack definitive conclusions and rely on prior knowledge, lacking sufficient flexibility and adaptability. This explains the negative and negative denoising effects observed in Tables 2 and 3 for EMD denoising in high signal-to-noise ratio environments. Figure 10 c-curve graph, appendix Figure 11 The significant deviation of the low-amplitude portion in the C-curve graph after noise reduction is due to the small difference in frequency and amplitude between the low-amplitude signal and the noise signal. This causes the useful signal in the IMF component to overlap with the noise signal, making them difficult to distinguish and resulting in poor noise reduction. In summary, the hybrid multi-step prediction model, which predicts noise in the useful signal based on prior knowledge of the noise signal, has stronger adaptability and provides a wider range of criteria for distinguishing between the useful and noise signals, thus achieving better noise reduction results.

[0148] (2) Verification of measured signals

[0149] To verify the performance of the noise reduction method of the present invention in practical applications, the measured raw signal obtained from the daily monitoring of an acoustic emission monitoring system in a lead-zinc mine is selected for noise reduction research. Figure 12 This is a schematic diagram illustrating an example of the spectrum of the original measured signal. Figure 13 yes Figure 12 A schematic diagram of the spectrum of a pure noise signal. Figure 14 This is a schematic diagram of the signal before and after noise reduction using a hybrid multi-step prediction model. Figure 15 This is a schematic diagram of the signal spectrum after noise reduction using a hybrid multi-step prediction model. Figure 16 This is a schematic diagram of the signal before and after wavelet denoising. Figure 17 This is a schematic diagram of the signal spectrum after wavelet denoising. Figure 18 This is a schematic diagram of the signal before and after EMD noise reduction. Figure 19This is a schematic diagram of the signal spectrum after EMD noise reduction.

[0150] Reference Figures 12 to 19 As shown, the arrival and termination times of the acoustic emission signal in the monitoring signal are 764×200μs and 896×200μs, respectively. Fourier transforms were performed on the pure noise signal before the arrival time and the signal in the acoustic emission region to obtain their spectra (see attached). Figure 13 With appendix Figure 15 Comparing the spectra of the two, it can be found that their frequency bands overlap, with no clearly distinguishable frequency range, making noise removal difficult in the frequency domain. (See attached...) Figure 13 It can be observed that the high-amplitude frequencies of the pure noise signal are mainly concentrated in the 800–1300 Hz range, with a maximum component amplitude of 861, corresponding to a frequency of 879.2 Hz. In the acoustic emission region, the signal frequency amplitude distribution is more uniform, with high-amplitude components mainly concentrated in the high-frequency range, with a maximum amplitude of 1029, corresponding to a frequency of 2462 Hz. Comparing the frequency distribution characteristics of the two, it can be found that noise interference is mainly concentrated in the 800–1300 Hz range, and the frequencies corresponding to the high-amplitude components are inconsistent. Furthermore, the amplitude of the high-frequency portion of the pure noise signal is smaller, while the amplitude of the high-frequency portion of the signal in the acoustic emission region is larger. Therefore, it can be considered that noise has a relatively small impact on the maximum amplitude of acoustic emission. (See attached diagram.) Figure 14 Appendix Figure 16 and appendix Figure 18 It can be observed that all three denoising methods produce some denoising effect, but the signals denoised by wavelet and EMD show a significant shift, falling into the trap of time-shifting defects, and have a significantly greater impact on the original frequency distribution of acoustic emission. (See Appendix) Figure 15 Appendix Figure 17 and appendix Figure 19 The frequency distributions before and after denoising using the three methods reveal that the hybrid multi-step predictive denoising model primarily weakens the amplitude of noise components within the 800–1300 Hz range, particularly significantly reducing the amplitude in the high-amplitude frequency range of the pure noise spectrum, while having less impact on components in other frequency ranges. Wavelet denoising weakens the amplitude of components in all frequency ranges to some extent, especially the amplitude of components corresponding to the higher frequencies where noise has less influence, thus its negative denoising effect is significantly greater. EMD denoising also weakens the amplitude of components in all frequency ranges to some extent, but the weakening of high-frequency components is smaller. The above comparative analysis shows that the hybrid multi-step predictive denoising model is more targeted in the denoising process, has less impact on the original frequency distribution of acoustic emissions, retains more useful information, and avoids the dilemma of time-shift defects.

[0151] Thus, the signal processing method based on XGBoost hybrid multi-step prediction provided in this application predicts the noise signal after the arrival of acoustic emission by using the characteristics of the noise signal before the arrival of acoustic emission, and then eliminates the noise prediction signal from the signal after the arrival of acoustic emission to achieve the purpose of noise reduction of the second signal, avoiding the problem of excessive noise reduction. Moreover, the noise signal prediction is relatively accurate, and it does not require obtaining the acoustic emission characteristics in advance. It can also avoid destroying the frequency, amplitude and other information of the acoustic emission signal, so the noise reduction effect is excellent and helps to perform more accurate ground pressure monitoring and analysis.

[0152] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0153] Figure 20 This is a schematic diagram of a signal processing device based on XGBoost hybrid multi-step prediction provided in an embodiment of this application. Figure 20 As shown, the device has the function of implementing the above-described signal processing method. This function can be implemented in hardware or by hardware executing corresponding software. The device may include: an extraction unit 2001, a modeling unit 2002, a prediction unit 2003, and a noise reduction unit 2004.

[0154] Extraction unit 2001 is used to extract a first signal and a second signal from the monitoring signal. The first signal is a pure noise signal before the arrival of acoustic emission, and the second signal is a signal in which acoustic emission and noise are superimposed. The duration of the second signal is k time steps.

[0155] Modeling unit 2002 is used to construct a hybrid multi-step prediction model. The hybrid multi-step prediction model is built based on k XGBoost models and trained using the first signal. The hybrid multi-step prediction model is used to obtain noise prediction data for each prediction time step based on the first signal. Each prediction time step is a time step in the second signal.

[0156] Prediction unit 2003 is used to process the first signal and input it into the hybrid multi-step prediction model to obtain noise prediction data for each prediction time step.

[0157] The noise reduction unit 2004 is used to generate a noise prediction signal based on the noise prediction data of each prediction time step, and remove the noise prediction signal from the second signal to obtain an acoustic emission signal.

[0158] In one possible implementation, the extraction unit 2001 is specifically used for:

[0159] The arrival and end times of acoustic emissions are determined from the monitoring signals;

[0160] The monitoring signal located between the arrival and end times of the acoustic emission is identified as the second signal;

[0161] The first signal is extracted from the monitoring signal preceding the second signal. The duration of the first signal is n time steps, and the first signal and the second signal constitute a signal that is continuous in the time domain for n+k time steps, where n is the second multiple of k.

[0162] In one possible implementation, the modeling unit 2002 is specifically used for:

[0163] A multi-output prediction model consisting of k XGBoost models is constructed. The multi-output prediction model is used to generate a k-row, k-column output matrix, and the average value of each prediction element corresponding to the same prediction time step in the output matrix is ​​used as the final prediction value of the prediction time step. Each XGBoost model is used to predict each prediction element in a corresponding column vector of the output matrix. The j-th column vector of the output matrix is ​​composed of the prediction elements of the corresponding XGBoost model based on the input matrix, which are j-lag time steps behind. j is a positive integer less than or equal to k. The input matrix is ​​a k-row, m+1-column matrix constructed by converting n time series data into supervised learning data based on each time step in the second signal. m is the first multiple of k.

[0164] The training set is obtained from the first signal, and the hyperparameters of the multi-output prediction model are tuned using the training set to obtain the tuned multi-output prediction model.

[0165] The optimized multi-output prediction model was determined to be a hybrid multi-step prediction model.

[0166] In one possible implementation, the input matrix is ​​constructed as follows:

[0167] Based on each time step in the second signal, the m+1 elements between the (nk-m+1)th time step and the (n-k+1)th time step are determined as the input feature vector of the first row;

[0168] Based on the first signal and the input feature vector of the first row, the input feature vectors of the remaining rows are constructed using the hysteresis method.

[0169] The k rows of input feature vectors together constitute the input matrix.

[0170] In one possible implementation, the output matrix is ​​represented as follows:

[0171]

[0172] Where Pred is the output matrix, x n-k+2For the prediction element at time step (n-k+2), x n+k This is the prediction element at the (n+k)th time step, where n is the duration of the first signal and k is the duration of the second signal.

[0173] In one feasible approach, the prediction elements at the same prediction time step in the output matrix satisfy the following condition:

[0174] i+j=g, g∈[k+1, 2k], i, j∈[1, k]

[0175] Where i is the row number of the predicted element corresponding to the prediction time step, j is the column number of the predicted element corresponding to the prediction time step, k is the total number of rows and columns of the output matrix, and the prediction time step is greater than or equal to n+1 and less than or equal to n+k.

[0176] In one possible implementation, the modeling unit 2002 is configured to acquire a training set from the first signal and use the training set to perform hyperparameter tuning on the multi-output prediction model to obtain a tuned multi-output prediction model, including:

[0177] The training set is obtained from the first signal, and the hyperparameters of each XGBoost model are optimized individually using the training set to obtain the optimal hyperparameters of each XGBoost model.

[0178] The hyperparameters of each XGBoost model are updated to the optimal hyperparameters to obtain the tuned multi-output prediction model.

[0179] In one possible implementation, the modeling unit 2002 is configured to perform hyperparameter optimization on each XGBoost model individually using the training set to obtain the optimal hyperparameters for each XGBoost model, including:

[0180] For each XGBoost model, the optimal solution is obtained by optimizing each hyperparameter of the XGBoost model based on a preset genetic algorithm.

[0181] The optimal solution is determined as the optimal hyperparameter of the XGBoost model.

[0182] In one possible implementation, the modeling unit 2002 is configured to optimize the various hyperparameters of the XGBoost model based on a preset genetic algorithm, including:

[0183] The hyperparameters of the XGBoost model are converted into the operation units of the preset genetic algorithm through binary encoding.

[0184] The sum of the mean squared error (MSE) values ​​from each forward validation is used as the fitness function of the genetic algorithm. The genetic algorithm performs a preset number of genetic iterations on each operational unit. The training and test data used in each forward validation are obtained from the first signal. The MSE value of each forward validation is the predicted value obtained by the XGBoost model based on the corresponding training data and the MSE value of the corresponding test data.

[0185] Thus, this application provides a signal processing device based on XGBoost hybrid multi-step prediction, which predicts the noise signal after the arrival of acoustic emission by using the characteristics of the noise signal before the arrival of acoustic emission, and then eliminates the noise prediction signal from the signal after the arrival of acoustic emission to achieve the noise reduction purpose of the second signal. The noise signal prediction is relatively accurate, it does not require obtaining the acoustic emission characteristics in advance, and it can also avoid destroying the frequency, amplitude and other information of the acoustic emission signal, so the noise reduction effect is excellent and it helps to perform more accurate ground pressure monitoring and analysis.

[0186] The present application has been described in detail above with reference to specific embodiments and exemplary examples; however, these descriptions should not be construed as limiting the present application. Those skilled in the art will understand that various equivalent substitutions, modifications, or improvements can be made to the technical solutions and implementation methods of the present application without departing from the spirit and scope of the present application, and all such modifications and improvements fall within the scope of the present application. The scope of protection of the present application is determined by the appended claims.

Claims

1. A signal processing method based on XGBoost hybrid multi-step prediction, characterized in that, The method comprises: extracting a first signal and a second signal from a monitoring signal, the first signal being a pure noise signal before the arrival of acoustic emission, and the second signal being a signal with acoustic emission superimposed on noise, the second signal lasting for k time steps; constructing a hybrid multi-step prediction model, the hybrid multi-step prediction model being used to obtain noise prediction data of each prediction time step according to the first signal, each prediction time step being each time step in the second signal; inputting the first signal after data processing into the hybrid multi-step prediction model to obtain noise prediction data of each prediction time step; generating a noise prediction signal according to the noise prediction data of each prediction time step, and removing the noise prediction signal from the second signal to obtain an acoustic emission signal; the construction of the hybrid multi-step prediction model comprises: building a multi-output prediction model comprising k XGBoost models, the multi-output prediction model being used to generate an output matrix of k rows and k columns, and taking the average of each prediction element corresponding to the same prediction time step in the output matrix as the final prediction value of the prediction time step, wherein each XGBoost model is used to predict each prediction element in a corresponding column vector of the output matrix, the j-th column vector of the output matrix is composed of prediction elements of lag j time steps predicted by the corresponding XGBoost model based on an input matrix, j is a positive integer less than or equal to k, and the input matrix is a matrix of k rows and m+1 columns constructed by converting n time series data of each time step in the second signal into supervised learning data, m is the first multiple of k; obtaining a training set from the first signal, and using the training set to perform hyperparameter tuning on the multi-output prediction model to obtain a tuned multi-output prediction model; determining the tuned multi-output prediction model as the hybrid multi-step prediction model.

2. The method of claim 1, wherein, The extraction of the first signal and the second signal from the monitoring signal comprises: determining the arrival time and the end time of acoustic emission from the monitoring signal; determining the monitoring signal between the arrival time and the end time of acoustic emission as the second signal; cutting the first signal from the monitoring signal before the second signal, wherein the first signal lasts for n time steps, and the first signal and the second signal constitute a signal of n+k time steps in time domain, and n is the second multiple of k.

3. The method of claim 2, wherein, The input matrix is constructed by: determining m+1 elements between the n-k-m+1 time step and the n-k+1 time step in the second signal as the input feature vector of the first row based on each time step in the second signal; constructing the input feature vectors of the remaining rows by using the lag method based on the first signal and the input feature vector of the first row; the k rows of input feature vectors together constitute the input matrix.

4. The method of claim 3, wherein, The output matrix is represented by: where Pred is the output matrix, x n-k+2 is the predicted element at the n-k+2 time step, x n+k is the predicted element at the n+k time step, n is the duration of the first signal, and k is the duration of the second signal.

5. The method of claim 4, wherein, each prediction element corresponding to the same prediction time step in the output matrix satisfies the following conditions: i+j=g, g∈[k+1, 2k], i, j∈[1, k]; wherein, i is a row number of a prediction element corresponding to a prediction time step, j is a column number of the prediction element corresponding to the prediction time step, k is a total row number and a total column number of the output matrix, the prediction time step is greater than or equal to n+1 and less than or equal to n+k.

6. The method of claim 1, wherein, The method further includes: obtaining a training set from the first signal, and performing hyperparameter optimization on each XGBoost model based on the training set to obtain optimal hyperparameters of each XGBoost model. The method further includes: updating the hyperparameters of each XGBoost model to the optimal hyperparameters to obtain an optimized multi-output prediction model. The method further includes: performing hyperparameter optimization on each XGBoost model based on a preset genetic algorithm to obtain an optimal solution.

7. The method of claim 6, wherein, The method further includes: updating the optimal solution to the optimal hyperparameters of the XGBoost model. The method further includes: converting each hyperparameter of the XGBoost model into an operation unit of the preset genetic algorithm through binary coding. The method further includes: performing genetic iteration on the operation unit of the genetic algorithm for a preset number of times by taking a sum of mean square error (MSE) values of each forward validation as a fitness function of the genetic algorithm, wherein the training data and the test data used in each forward validation are obtained from the first signal, and the MSE value of each forward validation is an MSE value of a prediction value of the XGBoost model based on corresponding training data and corresponding test data.

8. The method of claim 7, wherein, The device includes: The extraction unit is configured to extract a first signal and a second signal from a monitoring signal, the first signal being a pure noise signal before arrival of acoustic emission, and the second signal being a signal in which acoustic emission and noise are superimposed, and a duration of the second signal being k time steps. The modeling unit is configured to construct a hybrid multi-step prediction model, the hybrid multi-step prediction model being configured to obtain noise prediction data of each prediction time step according to the first signal, each prediction time step being each time step in the second signal.

9. A signal processing apparatus based on XGBoost hybrid multi-step prediction, characterized in that, The prediction unit is configured to input the first signal after data processing into the hybrid multi-step prediction model to obtain noise prediction data of each prediction time step. The noise reduction unit is configured to generate a noise prediction signal according to the noise prediction data of each prediction time step, and remove the noise prediction signal from the second signal to obtain an acoustic emission signal. ​ ​ ​ The modeling unit is specifically configured to: build a multi-output prediction model comprising k XGBoost models, the multi-output prediction model being configured to generate an output matrix of k rows and k columns, and to take an average value of each prediction element corresponding to a same prediction time step in the output matrix as a final prediction value of the prediction time step, wherein each XGBoost model is configured to predict each prediction element in a corresponding column vector of the output matrix, a jth column vector of the output matrix being composed of prediction elements of a lagged j time steps predicted by a corresponding XGBoost model based on an input matrix, j being a positive integer less than or equal to k, the input matrix being a matrix of k rows and m+1 columns constructed by converting n time series data into supervised learning data based on each time step in the second signal, m being a first multiple of k; obtain a training set from the first signal, and perform hyperparameter tuning on the multi-output prediction model by using the training set to obtain a tuned multi-output prediction model; determine the tuned multi-output prediction model as a hybrid multi-step prediction model.

Images