1. Introduction
Acoustic emission in rocks (geoacoustic emission or GAE) occurs as a result of their deformation and is accompanied by the radiation of elastic sound waves, which are recorded by acoustic receivers [
1]. The increased interest in the study of GAE worldwide can be explained by several important tasks related to the study of dangerous man-made and natural phenomena, such as rock bursts in mines, petrophysical property prediction [
2,
3], geophysical exploration [
1], rock failure in geotechnical construction, landslides and seismic activity in seismically hazardous regions. Additionally, particular attention deserves tasks related to studying the influence of GAE on other geophysical fields [
4]. On the Kamchatka Peninsula, in the area of the Petropavlovsk-Kamchatsky geodynamic testing site, GAE signals are recorded in the high-frequency range (up to the first tens of kilohertz). The importance of recording such signals is associated with studying the stress-strain state of the geological medium, which causes anomalies in GAE. In most cases, these anomalies precede seismic events—earthquakes. For example, the article [
1] showed that in half of the cases, anomalies of high-frequency GAE in near-surface sedimentary rocks preceded strong seismic events within in the 1–3 day interval. Therefore, high-frequency GAE can be considered an operational precursor of earthquakes.
To effectively study the stressed-deformed state of geo-environments with high-frequency AES, it is necessary to develop its mathematical models. In the theoretical work [
5,
6] a mathematical model was proposed on the basis of two related linear oscillators with non-constant coefficients which described the linear interaction between two dislocation sources (Berlage pulses). The quantitative and qualitative properties of the mathematical model were investigated. However, the modelling results were not compared with experimental data.
In the present work we will construct a mathematical model (Berlage oscillator) for one dislocation source, parameters of which are clarified by means of experimental data obtained at the observation point IKIR FEB RAS in Kamchatka. For this purpose, we will solve the problem of multidimensional two-step optimization using differential evolution methods [
7] and Nelder-Mida [
8]. The methodological innovation of the work consists in the application of a hybrid two-stage optimization algorithm to solve the reverse problem of identification of the parameters of the source of GAE. Unlike standard approaches to inversion of acoustic emission data, which are often limited to manually selecting parameters or using local methods, the proposed algorithm provides a robust search for a global minimum in a multidimensional parameter space, which is critically important for noisy field data. This allows not only to improve the accuracy of the approximation, but also to increase the reliability of the physical interpretation of the found model parameters. The work shows that the proposed algorithms allow to efficiently refine the values of key parameters of the proposed model.
The structure of the study in the article consists of the following sections: the introduction gives a description of the object of research and indicates the relevance of the selected topic (
Section 1).
Section 2 describes the registration of GAE in Kamchatka and provides information on acoustic receivers.
Section 3 provides a mathematical description of the Berlage oscillator for high-frequency GAE research, as well as a description of the two-step multidimensional optimization method. In
Section 4, the results of the application of the multidimensional two-step optimization method are shown, the visualization of the simulation results is carried out, as well as the table of the updated values of the Berlage oscillator.
2. GAE Recording Methodology
Analysis of the dynamics of geoacoustic emission pulses in the final stages of preparation of seismic events is of great interest for the identification of earthquake precursors, which is confirmed by multi-year observations of IKIR FEB RAS conducted since 1999 [
9,
10,
11].
Geoacoustic signals of the kilohertz frequency range are generated by sources located in a rock volume limited to the semisphere radius
m [
5]. The diagram of the process for recording geoacoustic emission signals is shown in the figure.
The methodology for recording geoacoustic emissions in Kamchatka is based on the use of specialized acoustic receivers—piezoelectric sensors of various types. The main element of the registration system—broadband hydrophones with a frequency range of 20 Hz–20 kHz, installed on the bottom of natural and artificial water bodies (
Figure 1) [
12].
For recording low-frequency oscillations (0.2–400 Hz) a combined receiver is used, which consists of a three-component vertical receiver and hydrophone (
Figure 2b) [
13]. The three-component receiver is mounted on the end surface of the optical strainer. The feature of its installation is the spatial orientation, in which the Z-component coincides with the axis of the gauge measurement, and the X-component is directed vertically.
For high-frequency measurements, a system of four cylindrical membrane-type hydrophones with a diameter of 65 mm (
Figure 2c), placed in cubic artificial water bodies (
) with vertical orientation is used. The three hydrophones (No. 2–4) form an equilateral triangle with a length of 5 m around the reflector tensometry attachment, and the fourth one (No. 1) is located 50 m from the central point [
9,
14]. Constructively, each receiver is a metal body with piezoelement, consisting of 8 rings made of piezoceramic CTBS-3 (dimensions
mm), indented on the rod. On one side the rod is rigidly fixed in the body, on the other—connected to a membrane that receives acoustic pressure. This design provides high sensitivity and measurement reliability over a wide frequency range [
15,
16]. The signals are digitized with a sampling rate of 44.1 kHz. The output files have a flac format lasting 15 min.
GAE signal is characterized by a sequence of relaxation pulses of different amplitude and duration with frequency of filling from hundreds of hertz to tens of kilohertz. The frequency of repetition in calm periods is one per second, and during anomalies preceding seismic events it reaches tens or hundreds per second.
An example of recording of high-frequency geoacoustic emission data at the observation point <<Karimshina>> (52.83° N, 158.13° E) in Kamchatka located in the seismotectonic fault zone from 24 June to 27 June 2017 is on
Figure 3. 27 June 2017 02:23 UTC there was an earthquake with a magnitude of 5.3 depth 40.5 km in the Pacific (52.58° N, 160.8° E) of Kamchatka. An analysis of geoacoustic emission data recorded within 72 h prior to the seismic event revealed statistically significant changes in acoustic signal parameters indicating earthquake preparedness (
Figure 3). There is an increase of the average amplitude by 35–40% compared to the background level, a more dense time distribution (reduction of intervals between pulses).
On the
Figure 4a you can see that 2 days before the earthquake (25.06) pulse activity already exceeded background values, but had a more chaotic nature without a clear temporal structure. For 48 h the appearance of individual high-amplitude pulses is observed, and for 36 h—the frequency of pulses and the growth of their amplitude. In the 24 h prior to the earthquake, 259,747 geoacoustic emission pulses were analyzed and 63,301 pulses (24.4%) similar in shape on Berlage functions were identified.
Next in the article we will use the geo-acoustic impulses given in the
Figure 4b–e to specify the values of key parameters of the mathematical model (functions) of Berlage.
3. Mathematical Model of GAE
For the analytical description of seismic acoustic oscillations in seismic exploration, the formulas N.N. Bubbles (Gaussian pulses) and G.P. Berlage [
12,
17] are widely used. In the works [
12,
18] to improve the quality of approximations of geoacoustic signals, a methodology was proposed that includes both Berlage and Gauss functions. It has been shown that this technique allows to create compact, diluted representations of geoacoustic emission signals (5–6 non-zero expansion coefficients) with a permissible error of no more than
[
19].
In this paper we use the following mathematical model to describe a single GAE pulse [
5]:
where
—pulse amplitude, from. units;
—parameter responsible for signal envelope shape;
—maximum position;
—signal envelope slope;
—time, c;
—start time, s;
—pulse duration, c;
[200, 20,000]—fill frequency, Hz;
—initial phase;
;
;
.
It can be shown that the solution to the model (
1) is the Berlage function [
6]:
The mathematical model (
1), whose solution is the Berlage function (
2), we will call the Berlage oscillator.
The Berlage oscillator (
1) describes one dislocation source (cracking) emitting one pulse of a high-frequency GAE.
The parameters of the Berlage oscillator directly reflect the physical processes accompanying the formation of cracks and the release of energy in rocks. The parameter a describes the dynamics of crack opening. Large values of correspond to a slow, inertial increase in amplitude—the crack opens gradually, with acceleration. This is typical for viscous fracture. Small values of , on the contrary, indicate a sharp, almost instantaneous nature of the fracture—the amplitude quickly reaches a maximum, which is typical for brittle fracture. The damping coefficient, determined by parameter b, characterizes the dissipative properties of the medium: a small value of b corresponds to strong attenuation, which reflects the intense dissipation of elastic wave energy on rock inhomogeneities and friction at the crack boundaries. The oscillating member is related to the resonant properties of the fracture source—its frequency is determined by the size of the crack and the elastic modules of the rock, with higher frequencies corresponding to smaller dislocation sizes. The initial conditions simulate the kinematic parameters of the crack nucleation event: the initial phase reflects the moment of source activation relative to the start of registration, and the initial amplitude is related to the magnitude of the displacement jump in the source. Thus, the combination of these parameters describes the complete event cycle: rapid release of energy during crack formation (initial conditions), resonant vibrations of the material in the dislocation region (oscillatory term) and subsequent attenuation of vibrations due to energy absorption by the medium (damping coefficient).
The articles [
5,
6] described the interaction of two dislocation sources using two linearly connected Berlage oscillators. Taking into account more than two dislocation sources leads us to a chain of connected Berlage oscillators, which is equivalent to considering the wave process of GAE propagation.
In this article, we will refine the parameters of the Berlage oscillator (
1) using multidimensional optimization taking into account the experimental data of the GAE obtained at the observation point of the IKIR FEB RAS (
Figure 4b–e). The refined values were compared with parameter values according to work results [
10].
The multidimensional optimization method included two steps: the differential evolution method [
7,
20] and the Nelder-Mead method [
8].
Differential evolution occupies a special place among modern methods of machine learning as one of the most effective optimization algorithms. This population method, belonging to the class of evolutionary algorithms, demonstrates high performance in solving complex optimization tasks. For evolutionary algorithms of scaled disturbance, the method uses a specific recombination of vectors. The algorithm consistently compares the quality of new solutions with current ones, keeping only the most effective options in the population, which ensures its high consistency and reliability when working with multidimensional search spaces [
21].
The Nelder-Mead method (deformable polyhedron method, simplex method) is used for local tuning, that is, to find the local extremum of a function in several variables. The main essence of the method for optimizing functions is to successively move and deform a simplex around an extremum point. Simplex is a polyhedron of vertices in n measurements, where each point is a set of parameter values of the optimized function. The method does not use a derivative (gradients) function, so it is applicable to uneven and/or noisy functions.
In the first stage of solving the problem of multidimensional optimization, the method of differential evolution is used, and in the second stage the found values of parameters of the model (
1) are adjusted using the Nelder-Mead method.
4. Study Results
The algorithms of the multidimensional optimization methods described above, as well as the visualization of the results, were implemented in the Python programming language.
The initial parameter values for the Berlage function (
2) were first set according to the work [
10]:
,
,
,
,
, then refined using multidimensional two-stage optimization. The time interval for the pulses is selected according to the time interval of the experimental pulses (
Figure 4b–e).
The differential evolution method was implemented with a population size of 15 individuals, the best1bin strategy, and a limit of 1000 iterations for global search. At the fine-tuning stage, the Nelder-Meade algorithm used strict convergence criteria (), which ensured high accuracy of the final parameters of the model. To analyze the performance of the selected optimization methods, we calculated the following statistical parameters: is the mean square error, —MSE of the initial model, is the coefficient of determinism, is the improvement of errors.
Figure 5a for the pulse 27 June 2017 02:22:08 clearly shows that although the model as a whole reproduces the shape of the pulse, there are minor discrepancies: the phase of amplitude increase in the experiment occurs faster than the model predicts, and the attenuation process is accompanied by additional oscillations. Spectral analysis in
Figure 5b shows a clear dominant peak at 3059.8 Hz.
The graph (
Figure 5c) shows non-closed phase trajectories that have the shape of spirals. First, the spiral begins to unwind, while the amplitude of vibrations increases until some point in time (
Figure 5a). Then the spiral begins to twist and the amplitude of the oscillations slowly decays.
The optimized model (blue line) (
Figure 5d) replicates the experimental pulse form (black line) much better than the initial model. The standard error (MSE) decreased by 96.41%. The coefficient of determination
confirms that the model explains 66.23% dispersion of data.
The graphs (
Figure 6a,d) show that the optimized model (
) reproduces key pulse characteristics by amplitude, which corresponds to a reduction of the mean squared error by 97.65% compared to the initial model (
). The coefficient of determination
(example 2) indicates the explanation of 55.37% dispersion of data. The basic frequency 5145.4 Hz (
Figure 6b) is well reproduced by the model. The phase trajectories (
Figure 6c) have the same dynamics as in
Figure 5c.
The
Figure 7a,d shows that the model accurately reproduces both the build-up phase of the pulse and the pattern of fading. Of particular interest are the phase trajectories on
Figure 7c, where the optimized model shows a substantially better match with experimental data compared to examples 1–2. The spectral analysis (
Figure 7b) confirms the accurate reproduction of the basic frequency 5139.6 Hz and a significant improvement in the correspondence both in the low-frequency (up to 12%) and in the high-frequency ranges.
The model on
Figure 8 achieved the best fit to the experimental data with a coefficient of determination of
and a decrease in MSE of
. The graphs show the coincidence of amplitude characteristics (
Figure 8a) and phase trajectories (
Figure 8c) (asymmetry of only 0.08), especially in the region of the main frequency of 3044.7 Hz (
Figure 8b). At the same time, phase trajectories keep characteristic features of experimental data, which indicates the correctness of using the model (
1). The results obtained, especially against previous examples (comparison with
,
and
of MSE reduction) confirm the stable effectiveness of the two-step optimization method for different types of geoacoustic pulses.
The current model adequately describes the main characteristics of the pulse, but requires modification to accurately reproduce subtle dynamic features.
The analysis of the optimized parameters shows a steady trend: pulses with higher frequencies (parameter c > 30,000 Hz, Examples 2 and 3 from
Table 1 and
Table 2) are characterized by extremely low values of the attenuation parameter b (~0.0003). This indicates a very rapid energy dissipation, which may be due to the generation of these pulses by smaller cracks in a highly inhomogeneous medium. On the contrary, pulses at low frequencies (c ~ 19,000 Hz, Examples 1 and 4 from
Table 1 and
Table 2) exhibit less intense attenuation, possibly due to the larger size of the source or the better elastic properties of the host block.
The choice of MSE is due to its computational efficiency and sensitivity to large deviations, which ensures a good match of the overall waveform. MSE heavily penalizes outliers and large discrepancies between the model and data, which helps to achieve an overall match of the waveforms. However, MSE has limitations, including high sensitivity to outliers and noise in the data, as well as the inability to account for phase shifts between signals, which can lead to physically incorrect fitting of model parameters. Additional analysis was applied for the complex validation of the model (
Figure 2 and
Figure 3). A visual comparison of the time realizations of experimental and model data makes it possible to assess the correspondence not only of the amplitudes, but also of the phase dynamics of the pulses. The comparison of the amplitude spectra provides control over the reproduction of the frequency characteristics of the source, and the analysis of the phase trajectories makes it possible to verify the correctness of the description of the oscillatory dynamics of the system. This complex approach compensates for the limitations of MSE and increases the reliability of the physical interpretation of the found model parameters.
In all examples, a significant decrease in MSE was achieved (by 96.41–99.25%) compared to the initial model. Best result (Example 4): MSE reduced to
with an improvement of
. High values of
(0.5537–0.9755) confirm the adequacy of the model. The differential evolution method provided an initial gross reduction in MSE (seen by the difference of
and MSE from
Table 3). The method is especially effective in Examples 1–2 (complex cases with low
) (
Table 3). The Nelder-Mead method gave the final fine tuning (achieving an MSE of about
). Maximum efficiency in examples 3–4 (
).
5. Conclusions
The combined use of methods made it possible to achieve high quality approximation (MSE is reduced by 2 orders of magnitude), while the Nelder-Mead method provided accurate final adjustment of parameters. Double optimization significantly improved the model’s fit to experimental data. Optimizing the parameters of the model using a combination of the differential evolution method and the Nelder-Mead method significantly improved its accuracy. The optimized model closely follows the dynamics of geoacoustic emission, including the moments of pulse generation and their attenuation. Coincidence is observed not only in the main peaks, but also in the smaller amplitudes of the signals, which confirms the adequacy of the model. It was found that the geoacoustic emission pulses that are most similar to the Berlage functions have a coefficient of determination of and an improvement of . The results obtained indicate that the proposed model correctly describes the process of generating geoacoustic pulses during deformation of geological environments. The results confirm the prospects of using the Berlage oscillator for the analysis of high-frequency GAE.
To further develop the work, we plan to scale the application of experimental data and the considered optimization methods in the case of a dynamic system with two or more interconnected dislocation sources.
Another continuation of the GAE study is related to its comparison with other studies in various geophysical fields, for example, such as the radon field [
22] or the electromagnetic field [
23], as well as a number of others. Such an integrated approach, in our opinion, should bring us closer to developing more effective methods for monitoring and diagnosing natural environments.
The equipment of the Center of Collective Use North-Eastern Heliogeophysical Center CKP 558279 was used in this work.