Ice avalanche infrasound feature extraction and classification identification method and medium
Through the time-frequency analysis and feature extraction of infrasound signals, combined with machine learning algorithms, the problem of real-time monitoring and identification of ice collapse events is solved, accurate prediction and early warning of ice collapse disasters is achieved, and the effect of glacier safety monitoring is improved.
Patent Information
- Application Number
- CN202210423964.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-21
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2042-04-21
AI Technical Summary
The prior art is difficult to monitor and accurately identify ice collapse events in real time, especially in remote areas, which lead to difficulties in monitoring and predicting ice collapse disasters.
The time-frequency analysis, wavelet transformation and Hilbert yellow transformation of infrasound signals are used to extract features, combine support vector machines and artificial neural networks to extract and classify the infrasound infrasound, and build an infrasound feature database to realize the ice collapse Real-time monitoring and prediction of events.
By directly monitoring the infrasound signal, extracting features and identifying them, real-time and accurate prediction of ice collapse events is achieved, monitoring and early warning capabilities for ice collapse disasters are improved, and the safety of local residents is protected.
Smart Images

Figure CN114707558B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of infrasound processing, and in particular to a method and medium for extracting and classifying ice avalanche infrasound features. Background Art
[0002] Ice avalanches are instantaneous collapses of ice at the front of a glacier under the action of gravity. They often occur at the end of a glacier, and sometimes also in the middle and upper parts of a glacier. In recent years, the trend of global warming has intensified due to human activities and climate change, and the cryosphere has been seriously affected. In particular, the Qinghai-Tibet Plateau, as a sensitive area to global climate change, has seen large-scale glacier melting and retreat, leading to a gradual increase in glacier instability. At the same time, the probability of ice avalanches is gradually increasing due to the influence of internal and external forces such as topography, earthquakes, and precipitation.
[0003] Since ice avalanches mostly occur in remote areas and do not cause serious economic losses and casualties, and due to the limitations of monitoring technology, they have not received widespread attention and research in the early days. In recent years, frequent ice avalanche disasters and the disaster chain they have formed have seriously threatened the safety of life and property of local residents, and also brought great hidden dangers to the ecological environment of the plateau area. In addition, with the development of remote sensing technology, it provides powerful monitoring data for ice avalanche disasters, which has led to a gradual increase in research on ice avalanche disasters. At present, the research on ice avalanche disasters is mainly based on the data collected from the previous field survey, with the help of remote sensing images, seismographs, and numerical simulations. Since ice avalanche disasters often occur in remote mountainous areas with rugged terrain, field surveys are difficult and dangerous, so the actual data collected is very limited; the development of remote sensing technology has promoted the progress of ice avalanche disaster research. Remote sensing images can monitor glaciers in all directions, but remote sensing images rely on the temporal and spatial resolution of satellites and cannot be monitored in real time; seismographs are currently mainly used to record the frequency of ice avalanches, and do not record the entire dynamic process of ice avalanches; numerical simulations can simulate the main influencing factors of ice avalanches, but the simulation of its mechanism is still difficult. Summary of the invention
[0004] The present invention provides a method and medium for extracting and classifying ice avalanche infrasound features, and intends to carry out ice avalanche infrasound monitoring to build an infrasound feature database, thereby providing a technical foundation for identifying ice avalanche events based on infrasound.
[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:
[0006] According to a first aspect of the present invention, a method for extracting ice avalanche infrasound features is provided, the method comprising: cleaning an original signal to obtain an infrasound signal; analyzing the time domain and frequency domain of the infrasound signal to obtain a sound pressure range, kurtosis, skewness and power spectrum; based on the infrasound signal, determining the energy distribution of the infrasound signal at each scale by discrete wavelet transform, and extracting eigenvalues to obtain three types of features, namely wavelet energy entropy, scale entropy and singular entropy; based on the infrasound signal, obtaining a Hilbert marginal spectrum by Hilbert-Huang transform.
[0007] According to a second aspect of the present invention, a method for classifying and identifying ice avalanche infrasound is provided, characterized in that, based on the method for extracting ice avalanche infrasound features described in any embodiment of the present invention, the method comprises: forming a feature vector from the extracted sound pressure range, kurtosis, skewness, power spectrum, wavelet energy entropy, scale entropy, singular entropy and / or Hilbert marginal spectrum, and judging the occurrence of an ice avalanche event based on the feature vector.
[0008] According to a third aspect of the present invention, there is provided a computer-readable storage medium having computer-readable instructions stored thereon, which, when executed by a processor of a computer, causes the computer to execute the method described in any embodiment of the present invention.
[0009] Compared with the prior art, the present invention has the following beneficial effects:
[0010] The present invention adopts an infrasound monitoring method to directly monitor the infrasound generated by ice avalanches, extracts event samples and performs filtering and noise reduction on the collected infrasound signals, adopts time-frequency analysis, wavelet and wavelet packet analysis, and Hilbert-Huang transform to extract the infrasound characteristics of each event, and analyzes the characteristics of different events. Subsequently, a support vector machine and an artificial neural network method are used to identify and classify different events, and a classification and recognition model is derived to predict the category of unknown events, so as to achieve real-time and accurate prediction of ice avalanche events and protect the personal and property safety of local residents. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] Figure 1 The present invention is a flow chart of an ice avalanche infrasound feature extraction method.
[0012] Figure 2a is an original signal diagram according to an embodiment of the present invention.
[0013] Figure 2b 4 is a diagram of a DC offset removal signal according to an embodiment of the present invention.
[0014] Figure 3a is an amplitude response diagram of the Kaiser window according to an embodiment of the present invention.
[0015] Figure 3bis a phase response diagram of the Kaiser window according to an embodiment of the present invention.
[0016] Figure 4 3 is a time domain and frequency domain response diagram of the Kaiser window according to an embodiment of the present invention.
[0017] Figure 5 Schematic diagram of the sym4 wavelet basis according to an embodiment of the present invention.
[0018] Figure 6 It is a comparison diagram of the time domain and frequency domain of the wavelets sym2, sym4 and sym8 according to an embodiment of the present invention.
[0019] Figure 7 4 is a diagram showing the effect of wavelet hard threshold denoising according to an embodiment of the present invention.
[0020] Figure 8 4 is a wavelet soft threshold denoising effect diagram according to an embodiment of the present invention.
[0021] Fig. 9 is a time-frequency analysis diagram according to an embodiment of the present invention.
[0022] Fig.10 4 is a schematic diagram of five layers of wavelet decomposition according to an embodiment of the present invention.
[0023] Fig.11 4 is a schematic diagram of three layers of wavelet packet decomposition according to an embodiment of the present invention.
[0024] Fig.12 It is a waveform diagram of each layer of wavelet decomposition according to an embodiment of the present invention.
[0025] Fig.13 is a diagram showing the energy proportion of each layer according to an embodiment of the present invention.
[0026] Fig.14 is a graph showing the energy proportion of each node in a wavelet packet according to an embodiment of the present invention.
[0027] Fig.15 is a singular value decomposition ratio diagram according to an embodiment of the present invention.
[0028] Fig.16 is a Hilbert spectrum according to an embodiment of the present invention.
[0029] Fig.17 is a Hilbert marginal spectrum according to an embodiment of the present invention.
[0030] Fig.18 0-7 Hz Hilbert marginal spectrum according to an embodiment of the present invention.
[0031] Fig.19 Schematic diagram of a 21-dimensional feature vector of an ice avalanche event according to an embodiment of the present invention.
[0032] Fig. 20 21-dimensional feature vector diagram of background events according to an embodiment of the present invention.
[0033] Fig.21 This is a signal diagram of the ice avalanche event at 2021-8-1-08:24:00 according to an embodiment of the present invention.
[0034] Fig. 22 This is a signal diagram of the ice avalanche event at 2021-6-10-12:56:00 according to an embodiment of the present invention.
[0035] Fig.23 This is a signal diagram of the ice avalanche event at 2021-8-1-10:23:00 according to an embodiment of the present invention.
[0036] Fig.24 It is a technical roadmap of an ice avalanche infrasound classification and identification method according to an embodiment of the present invention.
[0037] Fig.25 4 is a diagram of a classification and recognition system for infrasound signals according to an embodiment of the present invention. DETAILED DESCRIPTION
[0038] The present invention will now be further described with reference to the accompanying drawings.
[0039] The embodiment of the present invention provides a method for extracting ice avalanche infrasound features. Figure 1 As shown, the method includes:
[0040] Step S100, cleaning the original signal to obtain the infrasound signal. The collected original signal usually contains a lot of useless information and high-frequency noise, so the original signal needs to be cleaned before data analysis, that is, through DC removal, filtering, noise reduction and other processing to improve the signal-to-noise ratio of the signal and remove useless signal components.
[0041] Due to the influence of the relationship between the electronic components inside the collector, the system will have DC drift or DC component. The DC component is the average value of the signal, or the signal component with zero frequency. Its characteristic is that the amplitude of the signal does not change with time. Therefore, this part of the signal does not carry information, but it carries energy, which will cause energy errors in subsequent data processing. Therefore, this part of the signal should be removed before signal processing. Figure 2a and Figure 2b As shown, the original signal has a DC drift of about 0.2 Pascal, so the DC =y(x)-mean[y(x)] can remove the DC drift signal.
[0042] In actual signal acquisition, in order to obtain more accurate infrasound signals, the sampling frequency of the sensor is set to 100 Hz, that is, the collected sound wave signal has an audible sound segment greater than 20 Hz. Since only the infrasound band is considered during analysis, a low-pass filter needs to be set before processing the signal to filter out signals greater than 20 Hz.
[0043] A suitable filter can reduce signal distortion as much as possible. According to the form of signal processing, the filter can be divided into analog filters and digital filters. In the embodiment of the present invention, a digital filter is selected, that is, the signal is filtered by digital filter software; according to the passband, it can be divided into high-pass filter, low-pass filter, band-pass filter, all-pass filter and band-stop filter. Since the infrasound wave is analyzed, a low-pass filter is selected, that is, the sound waves below the 20Hz band can pass, and the frequency band above 20Hz is blocked; according to the filter network structure (operation structure), it can be divided into infinite unit impulse response filter (Infinite impulse response, IIR) and finite unit impulse response filter (Finite impulse response, FIR). The present invention uses a FIR type filter. FIR filter has a strict linear phase compared to IIR filter, which can avoid phase distortion of the signal during signal processing. At the same time, because its unit impulse response is finite length and has no recursive structure, its system is more stable.
[0044] The filter adopts a relatively simple and widely used time domain design method, namely the window function design method, and selects the Kaiser window with strong adaptability. The form of its window function is shown in formula (1):
[0045]
[0046] Where N is the length of the Kaiser window, n is the filter order, I0 is the first kind of zero-order modified Bessel function, R N (n) is a rectangular window function, β is a parameter for adjusting the window shape, and can adjust the main lobe width and side lobe attenuation at the same time. The larger β is, the narrower the ω(n) window is, and the greater the side lobe attenuation of the window spectrum is, but the main lobe width increases accordingly. Therefore, the larger β is, the better. Usually, β is between 4 and 9, as shown in Table 1. The main lobe width mainly affects the signal energy distribution and frequency resolution, that is, the wider the main lobe, the lower the frequency resolution under the same conditions. The side lobe attenuation rate reflects the degree of energy leakage (spectrum smearing effect). The slower the attenuation, the more serious the energy leakage and spectrum smearing.
[0047] Table 1. Effect of Kaiser window parameters on filter performance
[0048]
[0049] The low-pass filter is designed by using the built-in Filter Designer of MATLAB2020b to obtain the corresponding filter parameters. For example, when selecting an FIR low-pass filter, the filter order (i.e., the number of harmonic filtering) is set to 200. The higher the order, the better the performance, but the calculation time will be longer. Select the Kaiser window, the corresponding β value is 9, the original signal sampling frequency is 100Hz, and the cutoff frequency is set to 20Hz.
[0050] Depend on Figure 3a and Figure 3b It can be seen that the amplitude response and phase response of the Kaiser window are close to the ideal linear phase filter to a certain extent, that is, the amplitude response is rectangular and the phase response is linear. Figure 4 As shown, in the time domain response, its leakage factor is zero, in the frequency domain response, its main lobe width is 0.017, and the relative side lobe attenuation is 66.2dB, that is, the signal energy is mainly concentrated in the main lobe, the closer to the real spectrum. Therefore, the filter designed by the embodiment of the present invention can effectively filter out useless signal frequency bands.
[0051] In order to further improve the signal-to-noise ratio of the signal and increase the weight of the required signal, the filtered signal is further subjected to denoising to remove part of the high-frequency part. The embodiment of the present invention uses the wavelet soft threshold denoising method, and calls the MATLAB2020b built-in wavelet denoising toolbox Wavelet Signal Denoiser to perform wavelet denoising. The key to denoising is the selection of wavelet basis, threshold, and threshold function. In the embodiment of the present invention, the sym4 wavelet basis is selected for 5-layer wavelet decomposition, and the unbiased risk estimation (SURE) soft threshold denoising is adopted.
[0052] The symlets wavelet has good orthogonality, regularity, compact support and approximate symmetry. Its support length is 2N-1 and its vanishing moment is N. The sym4 wavelet basis used in this study is as follows: Figure 5 , its support length is 7, and its vanishing moment is 4. Usually, the support length is between 5 and 9. If the support length is too long and the vanishing moment is too high, the wavelet function will be too oscillating and the boundary energy will be too concentrated. If the support length is too short and the vanishing moment is too low, it will be unfavorable for energy concentration. Figure 6 shown.
[0053] Commonly used thresholds include unbiased risk estimation threshold (SURE), minimum and maximum threshold (Minimaxi), universal threshold (Universal threshold), Bayesian threshold (Bayes), etc. Since filtering has been performed before denoising to filter out most of the high-frequency noise, the unbiased risk estimation threshold (SURE) is used in the embodiment of the present invention. Its denoising is relatively conservative and is suitable for the case where the noise is less distributed in the high-frequency band to reduce the removal of useful signals.
[0054] The threshold function is divided into hard threshold denoising and soft threshold denoising. Figure 7 and Figure 8 As shown, hard threshold denoising is to give a threshold. When the absolute value of the wavelet coefficient is less than the threshold, it is set to zero, and when it is greater than the threshold, it remains unchanged. Soft threshold denoising is to subtract the threshold when it is greater than the threshold. It can be seen that hard threshold denoising will cause the denoised wavelet coefficient to have a step, which will cause the reconstructed signal to have a certain oscillation, but it more accurately retains the original signal, while soft threshold denoising can relatively reduce the appearance of steps, and the reconstructed signal is smoother, but it causes a certain compression to the original signal, and there is a certain signal distortion. Therefore, the hard threshold denoising method is selected in the embodiment of the present invention.
[0055] After the original signal is processed, feature extraction is performed. The feature extraction part is the core content of the present invention. Feature extraction is the process of extracting effective information from a large amount of complex data. The extraction of features must meet two characteristics. One is that information of the same category has a high degree of similarity, and the other is that information of different categories has great differences. Feature extraction in the field of signal processing is mainly divided into time domain analysis, frequency domain analysis, time-frequency joint domain analysis and some other nonlinear analyses. By extracting the features of infrasound signals, a corresponding infrasound feature database can be constructed to provide a data basis for the infrasound research of ice avalanches.
[0056] Therefore, in step S200, the time domain and frequency domain of the infrasound signal are analyzed to obtain the sound pressure range, kurtosis, skewness and power spectrum.
[0057] The time domain analysis of the signal is mainly used to extract the characteristic information of the signal, such as the sound pressure, waveform, and event duration. In this study, the sound pressure range, kurtosis, and skewness of the signal are mainly extracted. The sound pressure range reflects the amplitude of the signal in the time domain; the kurtosis reflects the flatness of the waveform. The higher the kurtosis, the steeper the waveform; the skewness reflects the degree of skewness of the waveform. A positive skewness indicates a right skewness, and a negative skewness indicates a left skewness.
[0058] Sound pressure range indicates:
[0059] Pk=max(x)-min(x) Formula (2)
[0060] Kurtosis means:
[0061]
[0062] Skewness means:
[0063]
[0064] In equations (2)-(4), x represents the input signal, μ represents the mean value of the input signal, σ represents the standard deviation of the input signal, and E represents the expected value of the input signal.
[0065] The frequency domain analysis of a signal mainly obtains the spectrum of the signal through Fourier transform (FT), and then obtains the power spectrum and other characteristic information of the signal through the ratio of the square of the Fourier transform to the signal length. The spectrum reflects the distribution of the signal in the frequency domain, and the power spectrum is the power spectrum density function, which characterizes the power distribution of the signal within a unit frequency. In the embodiment of the present invention, the fast Fourier transform (FFT) method with more convenient calculation is mainly used to obtain the power spectrum of the signal. And extract the power spectrum range of the signal.
[0066]
[0067] Where X(f) represents the fast Fourier transform of the signal and N represents the length of the signal.
[0068] According to the above extracted related signal features, we can get Fig. 9 The time-frequency analysis diagram is shown.
[0069] In step S300, based on the infrasound signal, the energy distribution of the infrasound signal at each scale is determined by discrete wavelet transform, and characteristic values are extracted to obtain three types of features: wavelet energy entropy, scale entropy, and singular entropy.
[0070] Wavelet analysis is a kind of time-frequency analysis, which can change both the time window and the frequency window, and can well characterize the local characteristics of the signal. In the specific analysis, both the time window and the frequency window can be adjusted as needed. When analyzing the low-frequency part of the signal, the wavelet transform can sacrifice the time resolution in exchange for a higher frequency resolution: when analyzing the high-frequency part of the signal, the wavelet transform can sacrifice the frequency resolution to improve the time resolution. In the embodiment of the present invention, the discrete wavelet (packet) transform is mainly used to extract the characteristics of the signal. The principle of wavelet packet transform is the same as that of wavelet transform. The wavelet transform only decomposes the approximate part of each layer, and the wavelet packet transform decomposes the approximate part and the high-frequency part of each layer at the same time to determine whether there are signal components in the high-frequency part. Fig.10 Schematic diagram of wavelet decomposition. Fig.11 Schematic diagram of wavelet packet decomposition.
[0071] In the embodiment of the present invention, the energy distribution of the infrasound signal at each scale is determined by the following steps:
[0072] Step 1: Assume that the function ψ(t) satisfies ψ(t)∈L 2 (R), that is, square integrable, and its Fourier transform If the condition shown in equation (6) is satisfied, then ψ(t) is a mother wavelet.
[0073]
[0074] Step 2: Scale and translate the mother wavelet to obtain a set of wavelet basis functions ψ a,b (t), as shown in the following equation (7), where a, b∈R, a>0, a is the scale factor, and b is the translation factor, both of which are continuously changing values.
[0075]
[0076] Step 3: Discretize the scale parameter a and translation parameter b, that is, a = 2 j , b j,k =2 j k, j, k∈Z, and obtain the discrete wavelet basis function ψ j,k (t).
[0077]
[0078] Step 4: Preprocess the raw data χ(t) to obtain the preprocessed infrasound signal S(t). This step is to process the relevant data collected by the sensor, and the processing can be performed with reference to the method described in step S100. If step S100 has been executed, the infrasound signal processed in step S100 can be directly used as the infrasound signal S(t).
[0079] Step 5: Perform discrete wavelet transform on the infrasound signal S(t).
[0080]
[0081] Step 6: Calculate the energy and energy proportion at different scales.
[0082] As an example, the original signal is decomposed into 5 layers using wavelet transform and 3 layers using wavelet packet transform. Fig.12 Decompose the waveform of each layer into wavelet. Fig.13 is the energy proportion of each layer, and the frequency bands corresponding to 1, 2, 3, 4, 5, and 6 are 0-0.79Hz, 0.79-1.57Hz, 1.57-3.13Hz, 3.13-6.25Hz, 6.25Hz-12.5Hz, and 12.5-25Hz. Fig.14is the energy proportion of each node in the wavelet packet, Fig.15 is the proportion of singular value decomposition. The frequency bands corresponding to 1, 2, 3, 4, 5, 6, 7, and 8 are 0-3.13Hz, 3.13-6.25Hz, 6.25-9.38Hz, 9.38-12.5Hz, 12.5-15.63Hz, 15.63-18.75Hz, 18.75-21.88Hz, and 21.88-25Hz.
[0083] After obtaining the energy distribution of each scale of wavelet decomposition through the above steps, the eigenvalues will be extracted. In the embodiment of the present invention, three types of features, namely wavelet energy entropy, scale entropy and singular entropy, are mainly extracted.
[0084] After the wavelet decomposition has been determined, the energy E of each layer i And energy proportion P i , then the energy entropy is
[0085]
[0086] in,
[0087] Similarly, the wavelet scale entropy is the wavelet coefficients C of all nodes decomposed into wavelet packets. n Shannon entropy
[0088]
[0089] in,
[0090] Before solving the singular entropy, the signal needs to be decomposed into a singular spectrum, which decomposes and reconstructs the signal according to its time series to extract different components of the signal.
[0091] X(t)=U m×m ∑ m×n V n×n T Formula (12)
[0092] X(t) represents the matrix composed of the original signal. In this study, the singular value S is calculated for each node of the wavelet packet decomposition. j , and then find the singular entropy of the signal:
[0093]
[0094] in,
[0095] It can be seen from the energy proportion diagram of wavelet decomposition and wavelet packet decomposition and the singular value proportion diagram that the signal is concentrated in the low frequency band and the distribution difference is large. Therefore, when selecting features, not only the total energy entropy, total scale entropy and total singular entropy of the wavelet are extracted, but also 1 / 4 energy entropy, 1 / 4 scale entropy, 1 / 4 singular entropy and 1 / 2 energy entropy, 1 / 2 scale entropy and 1 / 2 singular entropy are taken as signal features for the first two nodes and the first four nodes of the wavelet packet decomposition respectively.
[0096] Finally, in step S400, based on the infrasound signal, the Hilbert marginal spectrum is obtained by Hilbert Huang transform. After the EMD, the signal is decomposed into a number of single-frequency intrinsic mode functions (IMF), and then each IMF component is Hilbert transformed to obtain its corresponding instantaneous frequency. These instantaneous frequencies reflect the time-frequency energy distribution of the original signal, that is, the Hilbert spectrum. The spectrum can accurately reflect the distribution law of the signal, including the distribution law of the signal energy in time and frequency.
[0097] It specifically includes the following steps:
[0098] Step 1: preprocess the original data χ(t) to obtain the preprocessed infrasound signal S(t);
[0099] Step 2: Find all the maximum points, minimum points and endpoints of the infrasound signal S(t);
[0100] Step 3: Interpolate all marked extreme points through the cubic spline function, and then obtain the upper and lower envelope sequence E of the infrasound signal S(t) up 、E down , and calculate the mean M1(t) of the upper and lower envelopes;
[0101] M1(t)=(E up +E down ) / 2 Formula (14)
[0102] Step 4: Calculate the IMF signal I(t) of the infrasound signal S(t);
[0103] I1(t)=S(t)-M1(t) Formula (15)
[0104] Step 5: Determine whether I(t) meets the two conditions of IMF, that is, the difference between the total number of zero crossing points of the signal waveform and the number of all extreme points is less than or equal to 1, and the local mean of the upper and lower envelopes of the signal is zero. If the above two conditions are met, IMF1 = I1(t), and the first IMF component is obtained; if the conditions are not met, I1(t) is treated as the original signal and the above steps from the second to the fourth are repeated until the two conditions of IMF are met to obtain IMF1;
[0105] Step 6: Remove the first IMF component and obtain the remaining component r1;
[0106] r1=S(t)-IMF1 Formula (16)
[0107] Step 7: Determine whether r1 meets the three conditions of EMD, namely, first, the number of extreme points of the signal is greater than or equal to 2, that is, there is at least one maximum and one minimum point; second, the time scale of the extreme point can and uniquely determine the local time domain characteristics of the original signal; third, if a signal has no extreme point, there must be a singular point or inflection point that can be obtained by several differentials. If the above three conditions are met, the EMD decomposition ends. If not, repeat the above second step 6 with r1 as the new original signal to obtain other IMFn components;
[0108] Step 8: When r n When the termination condition is met, the EMD decomposition ends, and the original signal S(t) can be expressed as the sum of multiple IMF components;
[0109]
[0110] Step 9: Perform Hilbert-Huang transform on each IMF component to obtain the Hilbert spectrum H(w,t), where Re represents the real part of the IMF component;
[0111]
[0112] Step 10: Integrate the Hilbert spectrum H(w,t) to obtain the marginal spectrum h(w);
[0113]
[0114] Based on the above steps, the Hilbert spectrum and Hilbert marginal spectrum can be obtained, such as Fig.16 and Fig.17 As shown, the present invention mainly uses the Hilbert marginal spectrum as the characteristic of different sample signals. The Hilbert marginal spectrum can reflect the local characteristics of the signal at a lower frequency, so the marginal spectrum of the signal in the range of 0 to 7 Hz is used as the characteristic value of the signal. Fig.18 shown.
[0115] Based on the above time domain and frequency domain analysis, wavelet (packet) transform, and Hilbert-Huang transform, a total of 141 dimensions of features are currently extracted, of which the time domain and frequency domain features are 8 dimensions, the wavelet (packet) transform extracts 21 dimensions of features, and the Hilbert-Huang transform extracts 112 dimensions of features. Fig.19 and Fig. 20 is a feature vector composed of the features extracted based on wavelet (packet) transform of ice avalanche events and background events. Fig.19and Fig. 20 It can be seen that there are obvious differences between the feature vectors of ice avalanche events and background events. The feature vector composed of features extracted based on wavelet (packet) transform can be used as a feature vector to determine whether it is an ice avalanche event.
[0116] In some embodiments, the embodiments of the present invention obtain the original signal through the design and construction research of field monitoring points. Field monitoring should take into account the adaptability of the equipment to the environment, the interference of background noise, the probability of occurrence of the event to be monitored and the harm to the downstream, the real-time nature of data transmission, and the convenience of equipment transportation and installation. Based on the above requirements, the equipment is tested for environmental adaptability to ensure that the equipment can adapt to an ambient temperature of -30°C. According to the needs of equipment installation, issues such as solar energy replenishment, lightning protection, and waterproofing should be considered. The designed installation poles must ensure that they have the above-mentioned modules or functions, while also considering the height of the poles and the balance of force. Finally, the transmission method of the equipment is improved to ensure that the collected infrasound data can be transmitted online.
[0117] The data collected in the embodiment of the present invention is from the Qinghai-Tibet Plateau. The preliminary work in the early stage of data collection includes power supply environment adaptability test, equipment installation pole design, infrasound collector online transmission test and equipment installation site selection.
[0118] Power supply environmental adaptability test: The environment in the Qinghai-Tibet Plateau is harsh, with temperatures reaching -30°C in winter. This puts a great test on the environmental adaptability of the equipment. Therefore, before the actual equipment installation, the equipment is tested first. The power supply is placed in a -30°C refrigerator and frozen for a week before being tested to see if it can work normally. The test results show that the power supply equipment has good antifreeze properties and can adapt to the harsh environment of the Qinghai-Tibet Plateau.
[0119] Equipment installation pole design: When designing equipment installation poles, it is necessary to comprehensively consider the impact of the environment on the poles. Therefore, the poles cannot be designed too high or too low, the power lines cannot be exposed to the air, and the different components must be clamped on the poles to ensure the stability of the entire circumference. After the design plan is determined, go to the processing plant for processing, assembly and testing to ensure the stability of the poles.
[0120] Online transmission test of infrasound collector: Since data needs to be transmitted in real time, an online transmission test is performed on the selected infrasound collector to ensure that it can be transmitted online when the network is stable and can be saved locally when the network is abnormal.
[0121] Equipment installation site selection: Equipment installation is affected by many factors. (1) Since the infrasound monitoring method is used, it is necessary to consider the interference of as little background noise as possible. The glacier at the upper end of Jialong Co is located on the northern slope of the middle section of the Himalayas. It is a continental glacier and is less affected by the Indian Ocean monsoon climate. The local climate is relatively stable, which reduces the interference to the infrasound signal. (2) In 2002, there were two ice avalanches at Jialong Co, which caused the avalanche body to enter the lake and trigger the glacial lake burst, forming mudslides and floods, which had a certain impact on the downstream Nyalam County and even the China-Nepal Friendship Bridge. (3) Since the infrasound sampling frequency is set to 100 Hz and the number is large, Beidou communication cannot meet the requirements of large-scale data transmission. The network coverage near Jialong Co is conducive to online data transmission. (4) Since the equipment installation involves heavier equipment such as poles, solar panels, and power supplies, Jialong Co can be reached by car, reducing the transportation costs of manpower and animal power. After comprehensive consideration, the final site was selected at a gentle ridge on the side of Jialong Co in Nyalam County, Shigatse City, Tibet Autonomous Region. Jialong Co is located at 28°12′ 24.55″~28°12′54.13″ north latitude and 85°51′00.63″~85°51′11.48″ east longitude, with an altitude of about 4,334 meters.
[0122] After the above preliminary preparations, the corresponding sound wave signal data set is collected. In order to ensure that the infrasound signal is not missed, the sampling frequency is usually set below 100Hz, and the data is collected in real time, which will result in a large amount of data. Traditional manual recognition for sample extraction will take a lot of time. Therefore, this study uses the ratio of Short time average (STA) to Long time average (LTA), that is, the STA / LTA time window ratio method to obtain the initial time of event arrival, and judge whether it is an event based on the initial time obtained in combination with video surveillance, and extract event samples.
[0123] The STA / LTA time window ratio method was first used to identify earthquake events. By setting a long-time window and a short-time window, the window moves with the time series. While moving, the signal amplitudes in the long-time window and the short-time window are averaged. When the ratio of the two exceeds the set threshold, it is determined that an event has occurred.
[0124]
[0125] Where W lta , W staare the lengths of the long window and short window, respectively, and A(i) is the amplitude of the signal. The long window is used to measure the intensity of environmental noise, and the short window reflects the sensitivity to event signals. Therefore, the long window cannot be set too long, and the short window cannot be set too short. The size of the threshold also reflects the sensitivity to events. The lower the threshold, the more sensitive the event recognition is, and there will be misrecognition of events. Therefore, when setting the threshold, the surrounding environment should be considered. In this study, the lengths of the long window and short window are set to 30s and 0.5s, and the threshold is set to 2. Figure 21-23 Three ice avalanche events were identified.
[0126] The present invention also provides a method for classifying and identifying ice avalanche infrasound. Fig.24 As shown, the method can be used to determine the occurrence of ice avalanche events based on the infrasound feature database constructed by the method described above. The method includes: constructing a feature vector based on the features extracted by time and frequency domain analysis, wavelet transform, and Hilbert-Huang transform, constructing a classification model, and improving the model parameters through continuous training to achieve the best classification effect; using the trained classification model for the recognition of unknown signals, determining the category of the signal, and outputting the classification result.
[0127] The classification recognition used in the embodiment of the present invention is supervised pattern recognition, that is, the signal classification training is performed when the classification result is known, and the classification recognition of unknown signals is realized through continuous learning and testing. A complete classification recognition system consists of two parts: training and recognition. Fig.25 , build a recognition model by training sample data, and continuously improve the model to achieve the best classification effect. Common classification and recognition algorithms include artificial neural networks, support vector machines, Bayesian classifiers, K-nearest neighbor method, decision trees, etc.
[0128] The embodiment of the present invention adopts a supervised pattern recognition method to classify infrasound signals, and selects support vector machine (SVM) and artificial neural network (ANN) classification and recognition algorithms to divide infrasound signals into training samples and test samples, pre-process the training samples, construct feature vectors based on features extracted by time and frequency domain analysis, wavelet transform, and Hilbert-Huang transform, build a classification model, and improve model parameters through continuous training to achieve the best classification effect; the classification model obtained by training is used for identifying unknown signals, and the signal is classified and determined, and the classification result is output.
[0129] The classification method of support vector machine is to map the vector to high-dimensional space and find a hyperplane in the high-dimensional space so that the distance between the samples closest to the plane on both sides of the plane is the largest. This algorithm uses the idea of kernel function, so it can solve the classification of nonlinear problems. At the same time, it is suitable for small sample classification, that is, it is suitable for classifying disaster types.
[0130] Artificial neural networks simulate the neural network of the human brain through mathematical, physical and information processing methods, and establish a simplified model consisting of an input layer, a hidden layer, and an output layer. The hidden layer can have several layers, and each layer is connected by a large number of simple processing elements to form a highly parallel nonlinear system, which can realize parallel processing of information. At the same time, its highly nonlinear system makes the entire network have good fault tolerance, that is, damage to some neurons in the network will not affect the entire network, and it can adapt to environmental changes according to the performance of its own system.
[0131] The feature extraction conclusion is based on the comparison of some features of time domain, frequency domain, and wavelet packet transform extracted from 79 groups of ice avalanche samples and 71 groups of background noise samples. The classification and recognition conclusion is based on the wavelet packet features extracted from 70 groups of ice avalanche samples and 70 groups of background noise samples, including 9 groups of wavelet energy entropy, wavelet scale entropy, and wavelet singular entropy, as well as 12 groups of power spectrum features of the first four nodes of the third layer, that is, a total of 21 groups of feature vectors for classification and recognition. Finally, the following conclusions are drawn:
[0132] 1) The mean sound pressure range of ice avalanche events is 0.0042 Pascal, and the sound pressure range of background events is 0.00074 Pascal. The sound pressure range of ice avalanche events is about 5.66 times that of background events, that is, the sound pressure of ice avalanche is wider than that of background events. The mean power spectrum range of ice avalanche events is 58.12 dB, and the mean power spectrum range of background events is 69.54 dB. The power spectrum range of ice avalanche is narrower than that of background.
[0133] 2) The mean kurtosis and skewness of the ice avalanche event are 6.52 and -0.14, while the mean kurtosis and skewness of the background event are 8.59 and -0.53, which means that the background event has a steeper waveform with a longer left tail.
[0134] 3) Based on the wavelet energy distribution, it can be seen that the frequency band of ice avalanche infrasound is mainly distributed in the range of 0 to 7 Hz, and the frequency band of background infrasound is mainly distributed in the range of 0 to 3 Hz, that is, there is an overlap between the two types of infrasound frequency bands. This also shows that the background noise cannot be directly removed by filtering and noise reduction processing.
[0135] 4) Based on the wavelet energy entropy E, scale entropy SE, and singular entropy SVD, as shown in Table 2, the mean entropy value of the ice avalanche event is basically higher than that of the background event, indicating that the information contained in the ice avalanche event is more complex, and it also proves again that there are overlapping frequency bands between the two types of infrasound signals.
[0136] Table 2. Comparison of ice avalanche characteristics and background characteristics based on wavelet entropy
[0137]
[0138] 5) The correct recognition rate of the classification and recognition algorithm based on the support vector machine is 98.6%, of which the correct recognition rate of ice avalanche events is 98.6%, and the correct recognition rate of background events is 98.6%. The total correct recognition rate based on the artificial neural network algorithm is 99.3%, of which the correct recognition rates of the training set and the validation set are both 100%, and the correct recognition rate of the test set is 95.2%, of which the correct recognition rate of ice avalanche events is 90.9%, and the correct recognition rate of background events is 100%. From the conclusion of feature extraction, it can be seen that there are large differences between the two types of samples at present, and the features are more obvious.
[0139] The above description is intended to be illustrative rather than restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. For example, a person of ordinary skill in the art can use other embodiments when reading the above description. In addition, in the above-mentioned specific embodiments, various features can be grouped together to simplify the present disclosure. This should not be interpreted as an intention that a disclosed feature that is not required to be protected is necessary for any claim. On the contrary, the subject matter of the present invention may be less than all the features of a specific disclosed embodiment. Thus, the following claims are incorporated into the specific embodiments as examples or embodiments, wherein each claim is independently used as a separate embodiment, and it is considered that these embodiments can be combined with each other in various combinations or arrangements. The scope of the present invention should be determined with reference to the attached claims and the full scope of equivalent forms granted by these claims.
Claims
1. A method for extracting ice avalanche infrasound features, characterized in that: The method comprises: Clean the original signal to obtain the infrasound signal; Analyze the time domain and frequency domain of the infrasound signal to obtain the sound pressure range, kurtosis, skewness and power spectrum; Based on the infrasound signal, the energy distribution of the infrasound signal at each scale is determined by discrete wavelet transform, and characteristic values are extracted to obtain three types of characteristics: wavelet energy entropy, scale entropy, and singular entropy; Based on the infrasound signal, a Hilbert marginal spectrum is obtained by Hilbert-Huang transform; Based on the infrasound signal, the energy distribution of the infrasound signal at each scale is determined by discrete wavelet transform, and characteristic values are extracted to obtain three types of features: wavelet energy entropy, scale entropy, and singular entropy, including: Assume that the function ψ(t) satisfies ψ(t)∈L 2 (R), that is, square integrable, and its Fourier transform If the condition shown in formula (6) is satisfied, then ψ(t) is a mother wavelet; The mother wavelet is scaled and translated to obtain a set of wavelet basis functions ψ as shown in formula (7): a,b (t); Among them, a, b∈R, a>0, a is the scale factor, b is the translation factor, both are continuously changing values Discretize the scale parameter a and translation parameter b, that is, a = 2 j , b j,k =2 j k, j, k∈Z, and the discrete wavelet basis function ψ is obtained as shown in formula (8): j,k (t): According to formula (9), the infrasound signal S(t) is subjected to discrete wavelet transform; Calculate energy and energy proportion at different scales; According to the energy distribution of each scale, the eigenvalues are extracted to obtain three types of features: wavelet energy entropy, scale entropy, and singular entropy. According to the energy distribution of each scale, the characteristic value is extracted to obtain three types of features: wavelet energy entropy, scale entropy, and singular entropy, including: According to the energy E of each layer after wavelet decomposition i And energy proportion P i , the energy entropy is calculated by formula (10): in, The wavelet scale entropy SE(p) is calculated by formula (11): in, C n are the wavelet coefficients of all nodes; Decompose and reconstruct the signal according to its time series to extract different components: X(t)=U m×m ∑ m×n V n×n T Formula (12) Among them, X(t) represents the matrix composed of the original signal; Find the singular value S for each node of the wavelet packet decomposition j , and calculate the singular entropy of the signal according to formula (13): in, 2. The method according to claim 1, characterized in that The cleaning of the original signal to obtain the infrasound signal comprises: A signal with a DC offset removed is obtained by removing the DC drift from the original signal; Filtering the DC offset-removed signal by a FIR filter; The FIR filter is designed by the window function shown in formula (1): Where N is the length of the Kaiser window, n is the filter order, I0 is the first kind of zero-order modified Bessel function, R N (n) is a rectangular window function, β is a parameter for adjusting the window shape, which is used to adjust the main lobe width and side lobe attenuation at the same time. The larger β is, the narrower the ω(n) window is, the greater the side lobe attenuation of the window spectrum is, but the main lobe width increases accordingly. The value of β is between 4 and 9; The filtered signal is further subjected to denoising by a wavelet soft threshold denoising method to remove part of the high frequency part; the threshold of the wavelet soft threshold denoising method is selected as an unbiased risk estimation threshold.
3. The method according to claim 1, characterized in that: The time domain and frequency domain of the infrasound signal are analyzed to obtain the sound pressure range, kurtosis, skewness and power spectrum, including: The sound pressure range, kurtosis, skewness and power spectrum are calculated by the following equations (2)-(4): Pk=max(x)-min(x) Formula (2) In equations (2)-(4), Pk is the sound pressure range, Ku is the kurtosis, Sk is the skewness, x represents the input signal, μ represents the mean of the input signal, σ represents the standard deviation of the input signal, and E represents the expected value of the input signal.
4. The method according to claim 1, characterized in that: After obtaining the Hilbert marginal spectrum based on the infrasound signal through Hilbert-Huang transform, the method further comprises: using the marginal spectrum of the infrasound signal in the range of 0 to 7 Hz as the characteristic value of the infrasound signal.
5. A method for classifying and identifying ice avalanche infrasound, characterized in that: Based on the ice avalanche infrasound feature extraction method according to any one of claims 1 to 4, the method comprises: The extracted sound pressure range, kurtosis, skewness, power spectrum, wavelet energy entropy, scale entropy, singular entropy and / or Hilbert marginal spectrum are combined into a feature vector, and the occurrence of an ice avalanche event is determined based on the feature vector.
6. The method according to claim 5, characterized in that The extracted sound pressure range, kurtosis, skewness, power spectrum, wavelet energy entropy, scale entropy, singular entropy and / or Hilbert marginal spectrum are used to form a feature vector, and the occurrence of an ice avalanche event is determined according to the feature vector, including: According to the feature vector, a classification model is constructed, and the model parameters are continuously trained and improved through measured data sets; The trained classification model is used to identify unknown signals, classify the signals, and output the classification results.
7. The method according to claim 6, characterized in that The measured data set is obtained by: Based on the signals collected in real time on the glacier, a long-time window and a short-time window are set. The long-time window and the short-time window move with the time series. The signal amplitudes in the long-time window and the short-time window are averaged while moving. When the ratio of the two exceeds the set threshold, it is determined to be an ice avalanche event. When the ratio of the two does not exceed the set threshold, it is determined to be a background event.
8. A computer-readable storage medium, characterized in that: Computer readable instructions are stored thereon, and when the computer readable instructions are executed by a processor of a computer, the computer is caused to execute the method of any one of claims 1 to 4 or 5 to 7.
Citation Information
Patent Citations
Multi-resolution singular-spectrum entropy and SVM based leakage acoustic emission signal identification method
CN105488520A
Landslide infrasound signal recognizing method
CN110196448A
Infrasound signal identification method, system and apparatus based on MFCC and HMM
CN113705418A
