Intelligent inversion method of debris flow flow based on multi-scale infrasound signal fusion

Abstract

The present invention discloses a method for intelligent inversion of debris flow flow by fusing multi-scale infrasound signals. In view of the defect that the prior art cannot effectively fuse debris flow infrasound signals collected by two methods, namely, water flume simulation test and field observation, the present invention provides a method for fusing multi-scale debris flow infrasound signals, which eliminates the characteristic differences between the two types of original signals by improving the water flume experimental equipment and designing the signal processing flow and conditions, realizes signal fusion, and solves the problem of basic data expansion in the research of debris flow infrasound monitoring and early warning. The present invention provides a method for constructing a machine learning sample set, extracts energy, time domain, and frequency domain characteristic indicators from debris flow infrasound signals, and combines them with debris flow motion characteristic indicators to form a characteristic vector for machine learning. The method for intelligent inversion and measurement of debris flow flow provided by the present invention is a stack generalization integrated learning model based on debris flow infrasound signals, which can be applied to on-site debris flow monitoring and early warning. The present invention also provides application schemes of each technical scheme.

Description

Technical Field

[0001] The present invention relates to a method for processing and utilizing debris flow acoustic signals, and in particular to a method for fusing debris flow experiments with field multi-scale infrasound signals, and a method for intelligently inverting debris flow flow using machine learning using fused signals. The invention belongs to the fields of signal processing technology and geological environmental disaster prevention and control technology. Background Art

[0002] The infrasound (abbreviated as infrasound) signals released by debris flow processes contain rich information on geological environment and debris flow disasters, and are an important data resource in the field of debris flow prevention and control. In the field of debris flow acoustic wave monitoring and early warning, the inherent physical characteristics of infrasound signals have made researchers pay great attention to the research on the utilization of debris flow infrasound signals. In the research on debris flow acoustic wave monitoring and early warning, sufficient acoustic signal data is the basis of all research. With the introduction of machine learning in this field, each research project first needs to obtain sufficient and effective training data to construct a training sample set in order to meet the requirements of training set and test set division. In recent years, the value of ensemble learning has been more recognized and accepted. In order to support the algorithm ideal that each learner in the learner set is trained on a specific subset of training objects, the importance of training data scale has been further highlighted.

[0003] When developing a machine learning debris flow monitoring and early warning technology solution based on infrasound signals, there are at least three difficulties in solving the problem of training data scale: First, collecting original signals from the field. Collecting original signals from the field is the main source of original data in the current research on debris flow infrasound monitoring and early warning technology. Since infrasound signal collection requires a special infrasound collection device system, debris flow must occur in the debris flow channel where the monitoring station / research station has been built, or at least in the debris flow channel where the infrasound monitoring and collection system has been installed, and the monitoring system is always running normally, in order to effectively collect the original signal. This means that data generation and the debris flow disaster characteristics represented by the data are all limited by complex and comprehensive environmental factors. It is difficult for researchers to obtain some specific infrasound signals according to the experimental design in research and development. Second, collect original signals from indoor water tank experiments. Indoor water tank experiments have always been an important means to obtain various types of original data in debris flow research, and can provide a variety of debris flow characteristic data according to the research purpose. However, the debris flow disaster process is the result of the comprehensive changes of multiple environmental factors. Environmental factors are constantly changing during the movement of debris flow. In the analysis of various characteristics of debris flow, environmental factors have always been important variables. Therefore, in the study of debris flow disasters based on acoustic signals, the acoustic signals released by environmental changes are not simple background sounds or noises, but parts that can be used. In the indoor water tank experiment, it is impossible to simulate the comprehensive natural environment and its changes when the debris flow occurs, so it is impossible to simulate more realistic debris flow acoustic signals. Third, the fusion and utilization of the original signals in the field and the original signals in the indoor experiment. As can be seen from the above, although they are both called "debris flow acoustic signals", in terms of the real signal source, the field collection method is different from the indoor experimental collection method. At the same time, due to the low frequency and long wavelength of infrasound, it has the characteristics of being able to propagate a long distance in the atmosphere without almost attenuation. Therefore, the debris flow infrasound signals collected in the field are the signals released by the debris flow in the long-distance movement process within the mathematical axiom. In the indoor water tank experiment, it is impossible to simulate such a long distance, and the debris flow infrasound signal collected is only the signal released by the debris flow within a short distance of several meters. This further increases the difference in the real signal source between the two acquisition methods. Therefore, the original signals obtained by the two methods cannot be simply summarized and directly used for signal processing analysis.

[0004] The existing technology "Characteristics and type identification of debris flow acoustic wave signals" (Hu Zhihua, June 2020) discloses a scheme for waveform feature extraction, first arrival identification and machine learning classification algorithm for different types of debris flow infrasound signals through indoor debris flow infrasound simulation experiments and field field synchronous observation data. In the scheme, the acoustic signal generated by the debris flow in the water tank is collected by an infrasound sensor. In the experiment, the infrasound signals generated by debris flow flows of different properties are measured by changing the bulk density of the debris flow. Although the complete technical scheme from sample data collection to the construction of a machine learning classifier has been completed, the technical defects of the scheme are as summarized in the article-the study "although the type of debris flow is similar to that of natural debris flow, and the infrasound frequency is almost the same as that of natural debris flow, there is still a certain gap and difference with the infrasound of natural debris flow, which also hinders the promotion and application of research results. In the future, field debris flow properties and infrasound observation data can be widely collected, and indoor and field debris flow observation data samples can be collected to increase the number of samples, improve classifier performance and recognition accuracy." It can be seen that the defects of this technical solution are due to two points: first, the basic data comes from indoor simulation tests, and there is an irreconcilable gap between it and the field debris flow data; second, the number of samples is insufficient. In fact, as far as the current main debris flow infrasound monitoring and early warning research scheme is concerned, if the scale difference of the basic data obtained by the two types of collection methods and the technical problem of improving the integration level are not solved, and the technical route of obtaining the basic data required for research through a single collection method is continued, it will be difficult to achieve the technical goal of increasing the number of samples. Summary of the invention

[0005] The purpose of the present invention is to provide a multi-scale debris flow infrasound signal fusion method to address the deficiencies of the prior art, which can comprehensively utilize the original debris flow infrasound signals collected by field and indoor water tank experiments to provide sample data for the development of debris flow infrasound monitoring and early warning models. At the same time, the obtained data is used to construct a machine learning model for debris flow infrasound monitoring and early warning.

[0006] To achieve the above object, the present invention first provides a multi-scale debris flow infrasound signal fusion method, and its technical solution is as follows:

[0007] A multi-scale debris flow infrasound signal fusion method, characterized by:

[0008] The infrasound signals of different movement characteristics of debris flow were collected by water tank experiment, which were marked as experimental infrasound signals ISE. In the water tank experiment, cement and sand were mixed to coat the inner wall and inner bottom of the water tank, so that the inner wall and inner bottom of the water tank were similar to the natural channel. The infrasound signals of different movement characteristics of debris flow were collected by field observation, which were marked as field infrasound signals ISF. ISE and ISF were filtered by Chebyshev II low-pass filter, and the filtered signals were marked as ISE and ISF. ch 、ISFch ;

[0009] ISF ch With ISE ch The data is enhanced by translation, noise addition and filtering methods respectively to obtain the expanded data set ISE. au With ISE au , among which, ISE ch The translation time window is [-1s, +1s], ISF ch The translation time window is [-3s, +3s], ISE ch With ISF ch The noise addition and filtering conditions are the same;

[0010] The signal ISE au With ISF au The data volume ratio is not less than 1:1, and the debris flow infrasound fusion signal is obtained.

[0011] The basic principle of the above-mentioned multi-scale debris flow infrasound signal fusion method is that the experimental results designed based on various conditions in the early stage of the present invention found that the key to the difference between the debris flow infrasound signals collected by the flume experiment and the field observation is that the flume experiment cannot completely simulate and replicate the movement process of the debris flow in the natural environment. By building a flume experiment system under the natural open conditions of the debris flow channel and designing corresponding experiments, the present invention found that the key to the difference between the two collection methods is not the multiple sound signal backgrounds in the environment, but the fact that there are always some factors in the simulation experiment that cannot be controlled by experimental design but have an impact on the infrasound generation of the debris flow movement process. Debris flow infrasound is mainly generated from the inside of the debris flow fluid, as well as the interaction between the fluid and the environmental matrix such as the channel. In other words, the difference in the components of the debris flow fluid and the difference in the environmental matrix of the movement process will cause the scale differences of the infrasound signals collected by different means. By coating the inner wall and bottom of the flume with a mixture of cement and sand, the present invention improves the sound source composition of the infrasound signal in the traditional debris flow flume experimental system, which can partially reduce the difference between ISE and ISF. However, the experimental acquisition method still cannot control all influencing factors by improving the experimental "hardware" or "software". For example, the influencing factors that have not yet been verified may lie in the diverse interactions between the debris flow and the channel environment during the long-distance movement. Through the analysis of a large number of experimental results, the present invention found that the use of data enhancement can make the difference between the two types of signals easier to identify, and further, the purpose of eliminating the scale difference between the signals can be achieved through simple corrections in the signal time domain characteristics.

[0012] Under optimal conditions, the Chebyshev II low-pass filter conditions in the above debris flow infrasound signal fusion method can be set as follows: passband cutoff frequency 20Hz, stopband cutoff frequency 30Hz, passband maximum attenuation 0.15dB, stopband minimum attenuation 30dB.ch With ISF ch In the data enhancement, Gaussian noise perturbation can be used for noise addition, and Chebyshev II type low-pass filter can be used for filtering.

[0013] The debris flow infrasound signal fusion method of the present invention can realize the expansion of the debris flow infrasound signal data volume, so as to be further applied as basic data to the development of debris flow infrasound monitoring and early warning model. Therefore, the present invention provides the following technical solutions:

[0014] The above-mentioned multi-scale debris flow infrasound signal fusion method is applied in expanding the amount of debris flow infrasound signal data.

[0015] The application of the above-mentioned multi-scale debris flow infrasound signal fusion method in the development of debris flow infrasound monitoring and early warning model.

[0016] The multi-scale debris flow infrasound signal fusion method can improve the scale of debris flow infrasound signals and serve as training data for machine learning. Therefore, the present invention further provides a machine learning sample set construction method, and its technical solution is as follows:

[0017] A method for constructing a machine learning sample set using the multi-scale debris flow infrasound signal fusion method is characterized in that: the debris flow infrasound fusion signal is used as machine learning training data, marked as IS, and the energy features, time domain features, frequency domain features and corresponding motion feature indicators in the IS are extracted, and the energy features, time domain features and frequency domain features are used together as sample features x i , using debris flow motion characteristic index as sample label y i , forming a sample set [ISS] for machine learning;

[0018] The energy features include: using db5 wavelet basis to decompose IS into three layers, using soft threshold to denoise, and marking the denoised signal as IS st , calculate IS st Wavelet energy coefficients of each layer of signal components, a total of 4 energy coefficients ec i ;

[0019] The time domain features include: analyzing IS st Time domain characteristic values, calculate the maximum amplitude in the time domain, the average amplitude in the time domain, the effective value in the time domain, the variance in the time domain, the peak-to-peak value, the kurtosis index, the peak index, the margin index, the skewness index, the pulse index, the waveform index, the average short-time zero-crossing rate, a total of 12 time domain characteristic indexes td i ;

[0020] The frequency domain features include: IS with a frame shift of 100 sampling points, a frame length of 200 sampling points, and a semi-overlapping Hamming window stPerform frame division and windowing processing, analyze frequency domain characteristics, calculate peak frequency, center frequency, spectral line width, frequency root mean square, spectrum energy mean, power spectrum peak frequency, a total of 6 frequency domain characteristic indicators fd i .

[0021] Since the debris flow movement characteristics contained in the infrasound signal are reflected in the time, frequency and energy characteristics of the signal, three types of characteristic indicators and corresponding debris flow movement characteristic indicators are extracted from the training data IS to construct a sample set for debris flow movement feature machine learning.

[0022] The present invention also provides a machine learning solution for intelligently inverting debris flow flow using debris flow infrasound fusion signals. The details are as follows:

[0023] A method for infrasound inversion and calculation of debris flow flow using the machine learning sample set construction method of the present invention is characterized by:

[0024] S100, build sample set [ISS]

[0025] The sample set [ISS] is constructed using the machine learning sample set construction method, and the sample uses the debris flow flow as the sample label y i ;

[0026] S200, sample set [ISS] preprocessing

[0027] S300, divide the training set and test set

[0028] S400, build an integrated learning model using the training set

[0029] S410, taking the features x of each sample in the training set i The feature value is input, with sample label y i The debris flow flow value is output, and the random forest algorithm RF, k-nearest neighbor algorithm kNN, and extreme gradient boosting algorithm XGBoost are used to build base learners to obtain the outputs of three base learners;

[0030] S420, using the output of the three base learners as input and the sample label y i The debris flow flow value is output, and a gradient boosting decision tree GBDT meta-learner is constructed to combine each base learner to form a debris flow flow infrasound inversion model;

[0031] S500. Use the test set to complete the generalization ability evaluation of the debris flow infrasound inversion model.

[0032] The above-mentioned debris flow infrasound inversion estimation method is an integrated learning model based on the stack generalization idea. The model contains two layers of networks. In the first layer, three base learners are constructed using the random forest algorithm RF, the k-nearest neighbor algorithm kNN, and the extreme gradient boosting algorithm XGBoost. In the second layer, the gradient boosting decision tree GBDT is used as a meta-learner to combine the predictions of all base learners to obtain the infrasound inversion model of debris flow.

[0033] Based on the above-mentioned machine learning sample set construction method, the present invention provides an optimization scheme based on previous research: using the peak flow rate Q of the debris flow, one of the characteristic indicators of debris flow movement, as the sample label y i , with 11 sample features x i (1 energy characteristic ec i 、5 time domain characteristic indicators td i , 5 frequency domain characteristic indicators fd i ) is a set of feature vectors, and each set of feature vectors is extracted from the training data IS to construct a sample set [ISS].

[0034] The present invention also provides the following technical solutions:

[0035] The application of the above-mentioned machine learning sample set construction method in machine learning is to use the sample set [ISS] as the sample set of a machine learning regression model that uses debris flow geoacoustic signals to invert debris flow operation characteristic indicators (the preferred debris flow operation characteristic indicator is the debris flow peak flow Q).

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) In the two main original debris flow infrasound signal acquisition methods (experimental water tank experiment, field observation), the simulated experimental conditions cannot fully present the field environmental conditions, so that the signal source has essentially scale differences in the signal acquisition process, thereby restricting the common use of the two original signals by the prior art. The multi-scale debris flow infrasound signal fusion method of the present invention addresses this defect, determines the key difference between the characteristics of the two original signals through comparative analysis of a large amount of signal data, and amplifies the key difference by improving the water tank experimental equipment and designing a signal enhancement means with the same signal processing flow but different conditions, and finally eliminates the feature difference between the two original signals by using a simple signal processing operation, so that the two can be fused to form a signal set with consistent characteristics, which is suitable for further processing under the same analysis conditions in the subsequent signal analysis process. The method combines the technical advantages of the two signal acquisition methods, targets and effectively solves the problem of basic data expansion in debris flow infrasound monitoring and early warning research, and can provide great convenience for research and model development in this field. (2) The machine learning sample set construction method of the present invention provides a machine learning sample set scheme based on sufficient debris flow infrasound signals, by extracting the time, energy and frequency characteristics of the signal, and forming a 22-dimensional feature vector for debris flow movement feature analysis. Through screening, the above sample set construction method is optimized to a debris flow infrasound signal sample set for debris flow flow machine learning. (3) The debris flow flow infrasound inversion measurement method of the present invention is an intelligent inversion scheme for debris flow flow based on debris flow infrasound signals, and a stack generalization class integrated learning model constructed with debris flow infrasound signals as data basis. The generalization performance of the model can meet the requirements. The intelligent inversion model of debris flow flow based on infrasound monitoring can be applied to on-site debris flow monitoring and early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is the infrasound signal of the water tank experiment after fusion processing ISE au Time domain and frequency domain plots.

[0038] Figure 2 It is the field infrasound signal ISF after fusion processing au Time domain and frequency domain plots.

[0039] Figure 3 This is a time domain and frequency domain diagram of an original debris flow infrasound signal. DETAILED DESCRIPTION

[0040] The preferred embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

[0041] Embodiment 1

[0042] The method of the present invention constructs fusion data of debris flow infrasound signals.

[0043] 1. ISE acquisition of experimental infrasound signals in debris flow flume experiments

[0044] 1.1 Debris flow infrasound simulation water tank experimental system

[0045] The debris flow infrasound simulation water tank experimental system makes three improvements based on the existing technology ("Characteristics and Type Identification of Debris Flow Acoustic Wave Signals" (Hu Zhihua, June 2020)): (1) The inner wall and inner bottom of the water tank are coated with a mixture of cement and sand to simulate the natural channel matrix and ensure that the shape of the inner wall and inner bottom of the water tank is close to that of natural channels; (2) The shape and size of the tailings pool are changed and a connecting slope is added between the tailings pool and the water tank to avoid the ups and downs and impact movement of the fluid when it enters the tailings pool from the water tank; the inner wall, inner bottom and connecting slope of the tailings pool are coated with a mixture of cement and sand.

[0046] The main signal monitoring and acquisition include:

[0047] (1) Ultrasonic mud level meter: to obtain the mud depth of debris flow fluid;

[0048] (2) Infrasound sensor (frequency response range 0.5 Hz to 200 Hz, sensitivity 50 mv / Pa): collects sound wave signals including infrasound and some audible sound.

[0049] (3) High-speed camera recording + floating markers: record images of debris flow process and calculate debris flow velocity using image analysis method

[0050] 1.2 Experimental design

[0051] Particle gradation of debris flow materials: Referring to the gradation curve of Jiangjiagou debris flow samples, four groups of particle gradations were designed using particle size experiments, with upper limit particle sizes of 5mm, 10mm, 20mm, and 40mm respectively.

[0052] Debris flow material bulk density: 5 types of debris flow bulk density are designed, which are 1.6t / m 3 , 1.7t / m 3 、1.8t / m 3 , 1.9t / m 3 , 2.0t / m 3 .

[0053] Channel slope: Adjust the inclination of the water tank to design 5 channel slopes, namely 5°, 10°, 15°, 20°, and 25°

[0054] According to the experimental purpose, different experimental groups, control groups and repeated groups were designed to simulate different common debris flow movements in the wild.

[0055] 1.3 Experimental data

[0056] A total of 395 infrasound signals with different movement characteristics of debris flow were collected in the flume simulation experiment, marked as experimental infrasound signals ISE.

[0057] 2. ISF collection of field infrasound signals for field observation

[0058] 312 infrasound signals with different movement characteristics of debris flows were collected from multiple real debris flow records monitored in the field and marked as experimental infrasound signals ISF.

[0059] 2. Processing and fusion of original infrasound signals

[0060] 2.1 ISE processing of experimental infrasound signals

[0061] The ISE (395) was filtered by a Chebyshev II low-pass filter, and the filtered signals were marked as ISE ch Filtering conditions: passband cutoff frequency 20Hz, stopband cutoff frequency 30Hz, passband maximum attenuation 0.15dB, stopband minimum attenuation 30dB.

[0062] ISE ch Do data enhancement to get the expanded data set ISE au . In order, they include:

[0063] Translation: With the signal energy peak as the center, slide along the time axis within the time window [-1s, +1s] to fine-tune the phase.

[0064] Noise addition: Randomly disturb the debris flow infrasound signal. This embodiment adopts the method of adding Gaussian noise disturbance.

[0065] Filtering: A low-pass filter with a cutoff frequency of 20 Hz is used for the waveform segment. This embodiment uses a Chebyshev II low-pass filter. The passband cutoff frequency is 20 Hz, the stopband cutoff frequency is 30 Hz, the passband maximum attenuation is 0.15 dB, and the stopband minimum attenuation is 30 dB.

[0066] After the above three steps of data enhancement and conventional signal comparison screening, the expanded data set ISE is obtained. au 685 items. Figure 1 It is the fusion processed water tank simulation experimental infrasound signal ISE au .

[0067] 2.2 ISF processing of infrasound signals in the field

[0068] ISF ch Do data enhancement to get the expanded data set ISF au . In order, they include:

[0069] Translation: With the signal energy peak as the center, slide along the time axis within the time window [-3s, +3s] to fine-tune the phase;

[0070] Noise addition: same as ISE ch ;

[0071] Filtering: Same as ISE ch .

[0072] After the above three steps of data enhancement and conventional comparison and screening, the expanded data set ISF is obtained. au 515 items.

[0073] Figure 2 It is the field infrasound signal ISF after fusion processing au .

[0074] The signal ISE au With ISF au The debris flow infrasound fusion signals were merged to obtain a total of 1,200 signals.

[0075] Embodiment 2

[0076] The debris flow infrasound fusion signal obtained in Example 1 is used to construct a machine learning sample set [ISS].

[0077] The debris flow infrasound fusion signal obtained in Example 1 is marked as IS and used as training data for constructing machine learning.

[0078] Extract the energy characteristics of IS: Use the db5 wavelet basis to perform 3-layer wavelet decomposition on the signal to obtain 4 frequency bands, perform soft threshold processing on the wavelet coefficients, and mark the signal as IS after denoising st ; Perform wavelet energy coefficient analysis on each frequency band signal (use the square sum of wavelet coefficients to represent its energy, then add up the squares of the wavelet coefficients of each frequency band signal to get the total energy of the corresponding signal, and finally perform normalization to get the wavelet energy coefficient of each frequency band), and get 4 wavelet energy coefficients ec i .

[0079] Extracting time domain features of IS: Analyzing IS st Time domain characteristic values, calculate the maximum amplitude in the time domain, the average amplitude in the time domain, the effective value in the time domain, the variance in the time domain, the peak-to-peak value, the kurtosis index, the peak index, the margin index, the skewness index, the pulse index, the waveform index, the average short-time zero-crossing rate, a total of 12 time domain characteristic indexes td i .

[0080] Extract frequency domain features of IS: For IS st , with a frame shift of 100 sampling points, a frame length of 200 sampling points, and a half-overlapping Hamming window for IS stPerform frame division and windowing processing to calculate the peak frequency, center frequency, spectral line width, frequency root mean square, spectrum energy mean, and power spectrum peak frequency, a total of 6 frequency domain characteristic indicators fd i .

[0081] The above extraction yields 4 wavelet energy coefficients ec i 、12 time domain characteristic indicators td i , 6 frequency domain characteristic indicators fd i Together as sample features x i , with the corresponding debris flow motion characteristic index as the sample label y i , forming a sample set for machine learning [ISS].

[0082] In this embodiment, the peak flow rate Q of the debris flow in the debris flow motion characteristic index is used as the sample label y i The sample set [ISS] contains a total of 1,200 sample records.

[0083] Table 1 List of some data in the sample set [ISS] (showing 4 energy and 3 time domain feature indicators)

[0084]

[0085] Table 2 Partial data list in sample set [ISS] (showing 9 time domain feature indicators)

[0086]

[0087] Table 3 Partial data list in sample set [ISS] (showing 6 frequency domain characteristic indicators)

[0088]

[0089] Embodiment 3

[0090] Based on the sample set [ISS] constructed in Example 2, an integrated learning method is used to construct an inversion and calculation method for the peak flow rate Q of debris flow, so as to realize the calculation of the infrasound anomaly of the peak flow rate Q of debris flow.

[0091] 1. Sample set [ISS] preprocessing

[0092] The variance filtering method is used to remove sample features x whose variance of the sample set [ISS] is less than 0.05 i In this implementation, six features are filtered out, namely: wavelet energy coefficient 2, time domain average amplitude, time domain effective value, time domain variance, average short-time zero-crossing rate, and kurtosis index.

[0093] According to formula 1, sample feature x i With sample label y i The mutual information value MI of the sample x is removed, and the sample feature x with MI = 0 is removed.i In this embodiment, five features are filtered out, namely: wave energy coefficient 1, wavelet energy coefficient 3, maximum amplitude in time domain, peak-to-peak value, and power spectrum peak frequency.

[0094] The remaining sample features x i To clean the sample feature x m , a total of 11, including 1 energy feature (wavelet energy coefficient 4), 5 time domain feature indicators td i (peak index, margin index, skewness index, pulse index, waveform index), 5 frequency domain features fd i (peak frequency, center frequency, spectral line width, frequency root mean square, spectrum energy average). 11 clean sample features are retained in the sample set [ISS] x m and the corresponding sample label y m .

[0095] The cleaning sample features x m The characteristic values ​​of are standardized to make the sample data obey the normal distribution; the sample label y m The peak flow Q value of debris flow is normalized (Formula 2) to scale the flow data to the range of [0,1] and retain the data distribution characteristics. The processed sample set is marked as sample set [ISSp].

[0096]

[0097] In the formula, f' is the normalized data, f is the input data, and f min is the minimum value in the input data, f max is the maximum value in the input data.

[0098] The preprocessing established the association between the flume experimental data and the field measured data, and the eigenvalues ​​of the eigenvectors in the sample set achieved a complete transition and splicing from the flume scale to the watershed scale.

[0099] 2. Divide the training set and test set

[0100] The sample set [ISSp] was randomly divided into a training set and a test set in a ratio of 4:1 for subsequent machine learning.

[0101] 3. Use the training set to build an integrated learning model

[0102] Take the features x of each sample in the training set m The feature value is input, with sample label y mThe peak flow rate Q value of the debris flow is output. The random forest algorithm RF, the k-nearest neighbor algorithm kNN, and the extreme gradient boosting algorithm XGBoost are used to build base learners to obtain the prediction results (output) of the three base learners; the prediction results (output) of the three base learners are concatenated and merged as input, and the sample label y is used as the input. i The peak flow Q value of debris flow is output, and a gradient boosting decision tree GBDT meta-learner is constructed to combine each base learner. The effect of each learner is adjusted to the optimal value to form an intelligent inversion model of the peak flow Q of debris flow.

[0103] Through learning curve and grid search, the best combination of hyperparameters of each learner is obtained, as shown in Table 4.

[0104] Table 4. Ensemble learning hyperparameter combinations

[0105]

[0106] 4. Use the test set to complete the generalization ability evaluation of the intelligent inversion model of debris flow peak flow Q.

[0107] Take the features x of each sample in the test set m The feature value is input, with sample label y m The peak flow rate Q value of debris flow is used as the output test of the integrated model, and the mean square error, square absolute error, and determination coefficient are used as evaluation indicators of the integrated learning model. The results show that the mean square error is 0.024, the square absolute error is 0.011, and the determination coefficient R2 accuracy is 0.9021, indicating that the generalization performance of the model can meet the requirements.

[0108] Embodiment 4

[0109] Application of the infrasound inversion model of the peak flow rate Q of debris flow based on infrasound monitoring constructed in Example 3.

[0110] The debris flow intelligent inversion model based on infrasound monitoring constructed in Example 3 is applied to on-site debris flow monitoring and early warning, and the debris flow peak flow Q is inverted and calculated to obtain the current debris flow flow information in real time.

[0111] The infrasound monitoring equipment installed in the upper reaches of Jiangjiagou in Dongchuan, Yunnan, captured the infrasound signal generated by a debris flow. Figure 3 This is a time domain and frequency domain diagram of an original debris flow infrasound signal as an example to demonstrate the method of inverting and calculating the peak flow rate Q of the debris flow.

[0112] First, write a computational program to extract 11 sample features x from the time and frequency domains of the infrasound signal. i (1 energy characteristic ec i 、5 time domain characteristic indicators td i, 5 frequency domain characteristic indicators fd i ) function; then Figure 3 The infrasound signal is input into the above program to obtain the wavelet energy coefficient 4, peak index, margin index, skewness index, pulse index, waveform index, peak frequency, center frequency, spectral line width, frequency root mean square, and spectrum energy average value, which are 9.65, 59.26, 72.48, -2.14, 71.38, 13.81, 9.88, 7.92, 5.13, 10.52, and 6.35 respectively; finally, these characteristic values ​​are input into the debris flow peak flow infrasound inversion model, and the peak flow of the debris flow at this moment is obtained to be 4.52m 3 / s, which is consistent with the actual measured instantaneous flow rate of debris flow of 4.47m 3 / s is relatively consistent.

Claims

1. A method for constructing a debris flow multi-scale infrasound signal fusion sample set, characterized by: The multi-scale infrasound fusion signal of debris flow is used as machine learning training data, marked as IS; The method for acquiring the multi-scale infrasound fusion signal of the debris flow is as follows: using a water tank experiment to collect infrasound signals of different movement characteristics of the debris flow, which are marked as experimental infrasound signals ISE. In the water tank experiment, cement and sand are mixed to coat the inner wall and the inner bottom of the water tank, so that the inner wall and the inner bottom of the water tank are close to the natural channel. Using field observations, infrasound signals of different movement characteristics of the debris flow are collected, which are marked as field infrasound signals ISF; ISE and ISF are filtered by Chebyshev II low-pass filters respectively, and the filtered signals are marked as ISE respectively. ch 、ISF ch ; ISE ch With ISF ch The data is enhanced by using translation, noise addition and filtering methods in turn to obtain the expanded data set ISE. au With ISF au , where ISEc h The translation time window is [-1s, +1s], ISF ch The translation time window is [-3s, +3s], ISE ch With ISF ch The noise addition and filtering conditions are the same; the signal ISE au With ISF au The data volume ratio is not less than 1:1 to obtain the debris flow infrasound fusion signal Extract the energy features, time domain features, frequency domain features and corresponding debris flow motion characteristic indicators in the IS, and use the energy features, time domain features and frequency domain features as sample features x i , using debris flow motion characteristic index as sample label y i , forming a sample set [ISS] for machine learning; The energy features include: using db5 wavelet basis to decompose IS into three layers, using soft threshold to denoise, and marking the denoised signal as IS st , calculate IS st Wavelet energy coefficients of each frequency band, a total of 4 energy coefficients ec i ; The time domain features include: analyzing IS st Time domain characteristic values, calculate the maximum amplitude in the time domain, the average amplitude in the time domain, the effective value in the time domain, the variance in the time domain, the peak-to-peak value, the kurtosis index, the peak index, the margin index, the skewness index, the pulse index, the waveform index, the average short-time zero-crossing rate, a total of 12 time domain characteristic indexes td i ; The frequency domain features include: IS with a frame shift of 100 sampling points, a frame length of 200 sampling points, and a semi-overlapping Hamming window st Perform frame division and windowing processing, analyze frequency domain characteristics, calculate peak frequency, center frequency, spectral line width, frequency root mean square, spectrum energy average, power spectrum peak frequency, a total of 6 frequency domain characteristic indicators fd i .

2. The sample set construction method according to claim 1, characterized in that: The conditions of the Chebyshev II low-pass filter are: a passband cutoff frequency of 20 Hz, a stopband cutoff frequency of 30 Hz, a passband maximum attenuation of 0.15 dB, and a stopband minimum attenuation of 30 dB.

3. The sample set construction method according to claim 1, characterized in that: The ISE ch With ISF ch In the data enhancement, Gaussian noise perturbation is used for noise addition and Chebyshev II low-pass filter is used for filtering.

4. The method for constructing a sample set according to any one of claims 1 to 3, characterized in that: The debris flow motion characteristic index is the debris flow peak flow g, and the sample feature x extracted from the IS i Includes 1 energy signature ec i : Wavelet energy coefficient 4, 5 time domain features td i : Peak index, margin index, skewness index, pulse index, waveform index, 5 frequency domain features fd i : Peak frequency, center frequency, spectral line width, frequency root mean square, and spectrum energy average.

5. Application of the method for constructing a debris flow multi-scale infrasound signal fusion sample set according to any one of claims 1 to 3 in machine learning, characterized in that: The sample set [ISS] is used as the sample set of a machine learning regression model that uses debris flow infrasound signals to invert debris flow movement characteristic indicators or debris flow flow.

6. The use according to claim 5, characterized in that: The sample set [ISS] is constructed using the sample set construction method described in claim 4.

7. A method for inversely calculating debris flow flow using the method for constructing a debris flow multi-scale infrasound signal fusion sample set according to any one of claims 1 to 3, characterized in that: S100, build sample set [ISS] The sample set [ISS] is constructed using the machine learning sample set construction method, and the sample uses the debris flow flow as the sample label y i ; S200, sample set [ISS] preprocessing S300, divide the training set and test set S400, build an integrated learning model using the training set S410, taking the characteristic value of each sample feature in the training set as input and the debris flow flow value of the sample label as output, respectively using the random forest algorithm RF, the k-nearest neighbor algorithm kNN, and the extreme gradient boosting algorithm XGBoost to construct base learners, and obtaining the outputs of three base learners; S420, using the output of the three base learners as input and the sample label y i The debris flow flow value is output, and a gradient boosting decision tree GBDT meta-learner is constructed to combine each base learner to form a debris flow flow infrasound inversion model; S500. Use the test set to complete the generalization ability evaluation of the debris flow infrasound inversion model.

8. The inversion calculation method according to claim 7, characterized in that: The sample set [ISS] preprocessing in S200, if the sample set [ISS] is constructed according to the method of claim 5, execute S210 to S220; if the sample set [ISS] is constructed according to the method of claim 6, execute S220; S210, use variance filtering method to remove sample features x whose variance of sample set [ISS] is less than 0.05 i , use the mutual information method to remove the sample label y i Unrelated sample features x i , the rest are the cleaning sample features x m and the corresponding sample label y m ; S220, cleaning sample feature x m or x i The eigenvalues ​​of are standardized to make the values ​​obey the normal distribution, and the sample labels y m or i The data is normalized to scale the data within the range of [0, 1] and keep the data distribution characteristics unchanged. The obtained sample data constitutes the pre-processed sample set [ISSp] and proceeds to S300.

9. The inversion calculation method according to claim 7, characterized in that: The sample set [ISS] is constructed using the sample set construction method described in claim 4.

10. The inversion calculation method according to claim 8, characterized in that: The sample set [ISS] is constructed using the sample set construction method described in claim 4.

11. Application of the method for infrasound inversion calculation of debris flow flow as claimed in claim 7, characterized in that: It is used in debris flow monitoring and early warning.

12. Application of the method for infrasound inversion calculation of debris flow flow according to any one of claims 8 to 10, characterized in that: It is used in debris flow monitoring and early warning.

Images