The Use of Shape Accel Array for Deformation Monitoring and Parameter Inversion of a 300 m Ultrahigh Rockfill Dam

. Te global construction of rockfll dams has now surpassed the 300 m height level. Despite great achievements in dam design and construction, monitoring techniques have lagged behind the development of high rockfll dams. Existing deformation monitoring techniques are ill-suited to the high earth and water pressures, and extended monitoring periods are required for ultrahigh rockfll dams. Tis study introduces, for the frst time, the use of a shape accel array (SAA) to monitor internal displacement in a 300 m high earth core rockfll dam. Te SAA employs a rope-like array of capacitive MEMS accelerometers for deformation measurement. Compared to conventional monitoring techniques, SAA is a data-intensive monitoring technique. Based on the intensive data obtained from SAA, we employed a parameter inversion method, utilizing multiobjective optimization algorithm, the nondominated sorting genetic algorithm-III (NSGA-III), to inverse the constitutive model parameters of the rockfll dam. Te multiobjective parameter inversion method maximizes the use of multisource monitoring data for predicting rockfll dam deformation.


Introduction
Rockfll dams, one of the most promising types of dams, have seen remarkable technological advancements in recent years [1]. From 100 to 200 m dams, technology has successfully advanced to building 300 m ultrahigh dams [2]. Since the 1960s, many ultrahigh rockfll dams have been constructed, such as the Chicoasen dam (261 m, Mexico), the Nurek dam (300 m, USSR), the Shuangjiangkou dam (315 m, China), and the Rumei dam (315 m, China). Accurate and long-term monitoring data of these structures contain valuable space-time information that aids in studying the deformation mechanism of rockfll dams, confrming the accuracy of constitutive models, and assessing the safety of design, construction, and operation [3]. Table 1 outlines typical issues of conventional monitoring instruments caused by various factors.
Creating novel monitoring instruments that are appropriate for ultrahigh rockfll dams is crucial due to the difculties encountered with conventional monitoring instruments, such as sparse monitoring points and massive construction interference. Te monitoring instruments should be suitable for the complicated and challenging circumstances of ultrahigh rockfll dams, which involve high pressures and large shears.
Microelectromechanical systems (MEMS) have been applied in various aspects, because they are characterized by their compact size, light weight, and low cost. For example, by integrating gas detecting sensors into MEMS platforms, it is feasible to create miniaturized, portable sensing devices with extremely minimal power requirements [7]. Besides, it is possible to use multiple MEMS inertial measurement units for the robot's position estimation, maintaining real-time updates of its location and operating status, and thus further achieving autonomous driving functions [8]. MEMS has made remarkable achievements in the feld of biomedicines and has been developed to the extent of manufacturing MEMS microcoils to stimulate biological nerve tissues [9]. High-performance MEMS sensors are in signifcant demand for geophysical applications, including natural disasters such as earthquake monitoring, earth tides, and volcano activity monitoring, exploration of earthquake, gravity, and magnetic resources, and gravity-assisted navigation, drilling process monitoring [10]. Terefore, it is feasible to apply MEMS to the deformation measurement of rockfll dam engineering.
One of the representatives of deformation measurement using MEMS as the core component is shape accel array (SAA), which achieves precise, continuous, and steady deformation data acquisition. SAA was frst studied by Bennett et al. to monitor the deformation of an unstable slope [11] and evaluate the landslide by the monitoring data acquired [12]. Subsequently, Wang et al. successfully applied SAA to seabed landslide monitoring and successfully established a system that could record hydrodynamic data and monitor seabed landslide deformation [13]. Yan et al. used SAA techniques to capture temporal variations in distributed track settlement profles, analyze track safety issues, and inform track maintenance plans [14]. In our study, we applied SAA to monitor a 300-meter high rockfll dam to reduce the infuence of burial on construction and provide more deformation monitoring points. Te great information of spatio-temporal information contained support fnite element parameter inversion analysis, which works for more accurate prediction of dam deformation.
Sophisticated monitoring instruments and monitoring data are essential for evaluating dam safety and predicting dam deformation accurately. Tese monitoring data and prediction results can provide a precise assessment of dam safety.
Te prediction of dam deformation currently mainly includes data-driven prediction methods based on time series forecasting [15][16][17], as well as model-driven prediction methods basically based on the fnite element method (FEM). FEM plays a crucial role in dam deformation prediction. Many scholars have devoted themselves to studying advanced constitutive models to more accurately characterize the relationship between stress and strain of soil. For example, Guan et al. coupled machine learning with the fnite element method to directly learn the constitutive relationship from the data set [18]. In practical FEM calculations, parameters for rockfll dams are often obtained from laboratory experiments or feld trials. However, these parameters can be infuenced by experimental conditions, scale efects, random sampling, and other factors, which can result in inaccurate parameters that cause great error in FEM calculation [19]. Terefore, it is very important for the FEM calculation of the dam to identify the parameters accurately through appropriate parameter inversion analysis.
Many scholars have studied the parameter inversion analysis method for rockfll dams in recent years. For instance, Yu et al. proposed an efective Bayesian inversion method for reservoir parameters, which utilizes a datadriven error model and considers interactions among predictive performance metrics [20]. Li et al. adopts extreme learning machine (ELM) as a surrogate model to estimate the simulation results of a physical model, utilizing NSGA-II algorithm for multiobjective parameter inversion analysis [21]. Xu and Wu combine an unsaturated seepage model, principal component analysis (PCA), and NSGA-II to obtain more accurate parameters [22]. In addition, many scholars have conducted inverse analysis on various model parameters of concrete dams, including the formulation of objective functions, optimization of surrogate models, selection of multiobjective optimization algorithms, and development of multiobjective decision criteria [23][24][25][26][27]. Tese innovative works have also inspired the progress of our research. Efcient and accurate parameter inversion methods should be investigated to improve the evaluation of the safety status of rockfll dams. FEM results obtained using parameters acquired through parameter inversion, along with precise monitoring data for dam deformation, may signifcantly impact the assessment of dam safety.
Te main contributions and innovations of this paper are as follows: (1) We use the SAA to acquire crucial information of the deformation of a 300 m rockfll dam. Te installation of the SAA in an ultrahigh rockfll dam substantially decreases construction difculties and the infuence of instrument burial, greatly increases monitoring points for dam body deformation, and promotes dam safety monitoring. (2) We utilize the more efcient and dense deformation monitoring data obtained by the SAA. To adapt to the multimaterial zoning and multisource monitoring features at a 300 m rockfll dam, we perform an NSGA-III multiobjective optimization algorithm-based parameter inversion Table 1: Conventional instruments and monitoring difculties.

Instruments
Monitoring difculties Horizontal displacement gauge of tensional wire Steel wire creep, thermal expansion, cold contraction, and the infuence of the junction disc on the steel wire support [4] Hydraulic overfow settlement gauge Te resistance of liquid increases exponentially along a long pipeline and interferes with observations. Better burial technology is needed, such as 1% strictly refned burial gradient, and massive construction interference to provide the connected water-containing pipeline [5] Electromagnetic settlement gauge Easily demagnetized during long-term monitoring and large construction impact [6]. During the installation process, a protective tube with a diameter of about one meter is required to protect this instrument. Te diameter of the construction impact reaches about three meters 2 Structural Control and Health Monitoring analysis that uses SAA monitoring data as primary data and additional typical monitoring instrument data as supplementary data. Te overall structure of this paper is as follows: Te second section of this paper explains the structure and operating concept of the SAA, while the third section describes its application at the dam. Te fourth section details the specifc procedure of the parameter inversion method based on NSGA-III. Te ffth section applies this approach to multisource settlement monitoring data primarily based on SAAs, then analyzes and summarizes the fndings.

Structure and Monitoring Principle of the SAA
Te SAA uses a high-precision capacitive MEMS accelerometer for continuous acceleration monitoring and employs Kalman fltering to reduce observation system noise. Displacement of each monitoring point is determined by summing the displacements of each SAA part.

Structure and Monitoring
Principle of the SAA. One section of SAA consists of a high-precision capacitive MEMS accelerometer, a rigid pipe, a fexible hose, and cables. Te MEMS accelerometer is the most important component and is installed in the middle of the pipe, connected via an RS-485 bus. Data are transferred and stored in an external logger through the cable. Te SAA is user-friendly, providing easy access to dam monitoring information through integrated software. Other components make contributions to improving accuracy and simplifying layout. For example, the horizontally embedded SAA is protected by a PE pipe, while the vertically embedded SAA is buried in the inclinometer pipe. Te forced centring device is designed to ensure that the SAA can fexibly change position in the center of the inclinometer pipe. Table 2 and Figure 1 show the specifc structure and functions of SAA.

Calculation Method of Monitoring Point Displacement.
Te MEMS sensors in the SAA are arranged in a linear array. Te quaternion or Euler angle approach is used for MEMS accelerometer attitude analysis [28,29]. Te SAA employs only a MEMS accelerometer, making the solution simple. Each section of the SAA measures a point, and the multiple sections of the SAA are linked to monitor multipoints. After solving the data for each section of the SAA, the computed data value at each monitoring point may be acquired. Te specifc data calculation process is as follows: 2.2.1. Determining Shaft Inclination. Te MEMS accelerometer measures acceleration and outputs (accX i , accY i , accZ i ), the acceleration of i th SAA in three diferent orthogonal directions, (X, Y, Z). In equation (1), the corresponding inclinations can be obtained, where θ X i , θ X i , and θ X i denote the inclination angle formed by the current time position of the SAA and the initial position, respectively, A is the constant value bias of the MEMS accelerometer, K is the calibration factor of the MEMS accelerometer, and G is the gravitational acceleration.

Determining the Displacement Based on the Inclination
Angle. Te displacement of the nth node may be determined by summing the displacement outputs from each section, given that the length of each SAA is L. Te inclination angles of each SAA are calculated in the frst step. Ten, the displacement of the nth monitoring point can be obtained according to the following equation: Te two-axis MEMS accelerometer is often utilized in place of the three-axis MEMS accelerometer in practical engineering applications due to its lower cost. In terms of computation, SAA using two-axis and tri-axis MEMS accelerometers difers by one additional axis and one fewer axis.

Application of the SAA in an Ultrahigh Central Core Rockfill Dam Project
Te SAA in this research is manufactured by Beijing Shengkerui Technology Co., Ltd. For settlement monitoring, the SAA employs two-axis MEMS accelerometers, whereas three-axis MEMS accelerometers are used for measuring horizontal displacement.

Engineering Background of a Central Core Rockfll Dam.
Te dam in our research, situated on a river in China, is a rockfll core dam with a crest elevation of 2878.00 m, bottom elevation of 2582.00 m, riverbed dam height of 295.00 m, and core wall top width of 6.00 m. It comprises of an impervious body, flter layer, transition layer, and rockfll area. Te seepage-proof body has a straight core wall of gravel soil, while the shell is flled with rockfll material. Te core wall and dam shell rockfll include flter and transition layers. A 1.0 m thick concrete cover plate is placed on the bank slope between the core wall and dam Structural Control and Health Monitoring abutment on both sides, with a 4.0 m thick contact clay at the core wall-plate interface. Upstream and downstream of the core wall are two layers of flter protection dirt.
Te central core rockfll dam is equipped with various innovative monitoring equipment, including the SAA, pipe robots [30], and distributed optical fbre monitoring    [31,32], in addition to conventional monitoring instruments such as hydraulic overfow settlement gauge, horizontal displacement gauge of tensional wire, electromagnetic settlement gauge, and cross beam settlers. Tese instruments provide multisource data, which can be fused to create an integrated intelligent safety analysis platform for the dam body. Te SAA is the representative innovative internal monitoring device installed in the upstream and downstream rockfll areas and gravel soil core wall of the dam. Te project and dam material zoning map is shown in Figure 2.

Installation Process of the SAA in the Rockfll Dam.
Te SAA embedding process for this rockfll dam is signifcantly easier compared to the challenging construction technology required by horizontal displacement gauges of tensional wire and hydraulic overfow settlement gauges. Te embedding procedures of SAA are as follows: To monitor the uneven settling of the dam, SAAs are embedded at diferent elevations in the upstream rockfll region and the core wall of the rockfll dam, with a simpler installation process than traditional inclinometer tube displacement meters. First, we need to transport the SAA to the construction site and inspect the instruments to ensure they are in good condition. After confrming that the instruments are not damaged during manufacturing or transportation, we apply butter for lubrication on the SAA and place it into the PE pipe. We connect the SAA to the solar power generation device to ensure long-term automated monitoring. However, the hard retaining wall of the SAA is susceptible to brittle fracture, which can lead to failure of the SAA. To prevent such damage, the pipes can be flled with fne sand instead of cement. To embed the SAA in the dam, a trench of x m × 1.5 m × 1.3 m is excavated along the dam axis at the cushion layer location. After leveling the trench bed, a 23 cm thick layer of clay is backflled and compacted manually. Te SAA is then surrounded with a PE pipe, fxed to the cushion, and flled with concrete mortar, surrounding the sensor with 55 cm of fne sand and backflling with contact clay follows. After manually and uniformly compressing the trench with a vibrator, the dam flling material is used to fnish the embedding. Figure 3 illustrates the SAA embedding process.

Layout of the SAA in the Rockfll Dam.
Te rockfll dam includes 13 sets of SAAs that are organized in a total length of 1147 meters. Te SAA includes a two-axis MEMS accelerometer to reduce expenses. Figure 4 depicts the precise burial position and length of the SAA.
Te layout drawings illustrate the horizontal placement of SAA on the core wall section and the 3-3 monitoring section with a 1 m spacing, featuring closely spaced monitoring points. Te length of 1147 meters of SAA in the rockfll dam provides extensive monitoring data over an extended period and spatial range, surpassing the limitations of hydraulic overfow settlement gauges and electromagnetic settlement gauges with wider measurement intervals and vertical arrangement, respectively.

Analysis of SAA Data in the Rockfll Dam. Te IN-R11
and IN-R12 SAAs were installed vertically within the rockfll dam to measure horizontal displacement. IN-R12 was intended to be buried after flling the top of the dam, but its installation was delayed due to material shortages and a limited monitoring period. Te other SAAs were installed horizontally to measure settlement. Te position of each instrument's monitoring point is calculated based on the origin at the contact point between the instrument and the dam body. Monitoring data of IN-R1 to IN-R7, and IN-R11 are presented in Figure 5.
Based on the data collected from diferent SAAs, overall monitoring data conform to the deformation law of a rockfll dam. Te monitoring data from SAAs showed the following patterns: the settlement in the middle of the core wall is greater than that near the riverbank; there is not much diference in the settlement quantity between both sides of the dam with the central axis as the reference, which confrms the rationality of the SAA monitoring data; over time, the settlement amount of the dam body shows an increasing trend. As for horizontal displacement, only the IN-R11 data are available as a reference. Overall, the downstream displacement measured by SAA increases over time and the pattern presented spatially, due to the infuence of water storage, has certain reference value.
Certain embedded SAAs are damaged, with certain monitoring points exhibiting erratic variations. Earlier research indicates that the load distribution within rockfll dams is nonuniform, similar to that of the loads of train tracks, leading to rheological soil deformation. Tis may cause inconsistent pressure on embedded SAAs and continuing plastic deformation of the soil mass, potentially changing the observation data's point or damaging the instrument [14,32].

Multiobjective Parameter Inversion Method Based on Multisource Monitoring Data
Parameter inversion is a common approach to determine rockfll dam calculation parameters. Multiobjective parameter inversion is employed, utilizing deformation monitoring data from various instruments, multimaterial zoning, and the objective function value based on the degree of ftting between deformation monitoring and calculation values. A constitutive model, parameters, multiobjective functions, optimization algorithm, and machine learning surrogate model are required for this approach.

Constitutive Model Selection.
Stress and deformation calculations of ultrahigh rockfll dam and parameter inversion both require a suitable rockfll constitutive model. Zhang et al. [33] compared the rockfll constitutive model under the complex stress path of the rockfll dam and concluded that Duncan-Chang EB, one of the most wellknown nonlinear elastic models, can describe the nonlinear relation of rock deformation more accurately. Te parameters of the Duncan-Chang EB model have simple measurements and clear physical concepts. Tis model is widely applied in the design of rockfll dams and has many engineering reference bases [34]. As a result, the constitutive model for the FEM computation of this ultrahigh rockfll dam is the Duncan-Chang EB model.

Structural Control and Health Monitoring
where E t is the tangent elastic modulus; K is the tangent modulus; n is the tangent modulus index; E ur is the unloading rebound elastic modulus; K ur is the unloading modulus; n ur is the unloading modulus index; B t is the tangent bulk modulus; K b is the bulk modulus; m is the bulk modulus index and R f is the material failure ratio; c is the cohesion of the material; P a is the unit atmospheric pressure; and φ is the internal friction angle, considering the change with the confning pressure, which is shown as follows  A Merchant viscoelastic model based on exponential attenuation is also used to account for the rheological deformation of rockfll dams over time. Equations (7) to (9) can be used to express the rheological deformation, fnal volumetric rheology, and fnal shear rheology of a rockfll dam as follows: ε Vf � α σ 3 P a n 1 + β q P a where ε t is the rheological deformation; ε f is the fnal fow variable over time; ε Vf is the fnal volume fow variable; c f is the fnal shear rheology; and λ, α, β, δ, n 1 , n 2 , and n 3 are the seven calculation parameters of the rheological model, where λ refects the rheological rate and α, β and δ refect the fnal rheological deformation.

Multiobjective Function. Parameter inversion involves
fnding appropriate parameters to approximate the monitored displacement value and calculated displacement value. Relevant objective functions in diferent material zones of high rockfll dams are constructed, and deformation monitoring data from several monitoring sensors are introduced. Equation (10) shows the constructed multiobjective function as follows: where the function f(x) contains n optimization objective functions, and x i and x * i are the monitoring displacement value and FEM calculation value of m (n) internal monitoring points selected from the nth material partition of the dam body.
Deb and Jain [45] and Jain and Deb [46] updated the NSGA algorithm to NSGA-III in 2018. Te core foundation of the NSGA-III algorithm is GA. A rapid nondominated sorting algorithm, present and associated reference points, and other selection mechanisms are used to improve population individuals, iterate to the convergence algebra, and search for a Pareto optimal solution set in genetic operator selection. Using NSGA-III, we solve parameter inversion as a multiobjective optimization problem. Figure 6 depicts the generational fow of NSGA-III.

Machine Learning Surrogate Model.
Calculating objective function values for all parameter combination individuals for each target space through FEM demands signifcant computing space and time. Te machine learning regression algorithm's superior simulation capacity can replace FEM calculation in the iterative process and increase the efciency of parameter inversion calculation by mapping material parameters to deformation calculation results.
We train samples using Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boost (XGBoost), and Artifcial Neural Network (ANN). Besides, we use the mean absolute error ratio (MAPE), root mean square error (RMSE), and the determination coefcient (R 2 ) to evaluate each machine learning surrogate model and choose the best one. MAPE, RMSE, and R 2 are calculated as follows:

Structural Control and Health Monitoring
where x * i is the prediction results of the centralized surrogate model and x i is the actual result.

Multiobjective Parameter Inversion Process.
Based on the multiobjective function construction strategy described above, the NSGA-III multiobjective optimization algorithm and machine learning surrogate model are used to propose a multiobjective parameter inversion method that can adapt to the multimaterial zoning characteristics of ultrahigh rockfll dams and fully utilize multisource deformation monitoring data. Figure 7 shows the parameter inversion method.

Application of Inversion Based on SAA and Other Multisource Monitoring Data
It is worth mentioning that the FEM calculations in this paper were performed using the model that we developed based on ABAQUS. Te surrogate model and multiobjective optimization algorithm were implemented using Python and can be found in the open-source libraries Keras and Pymoo, respectively.  (1) , f (2) ..., f (i) ,..., f (I) }

Establishment of a Multiobjective Function Model and
Objective fuction sample {f (1) , f (2) ..., f (i) ,..., f (I) }  and SAA data are densely distributed in time and space, mainly SAA data are chosen when selecting the inversion monitoring points to ensure that instrument performance does not afect calculation accuracy. Inversion computation uses monitoring data from the core wall section and 3-3 section before the frst impoundment (December 2020), the second impoundment (June 2021), and the last monitoring period (September 2021). Table 3 shows the selected instrument zones and measuring locations.
Based on the multiobjective function construction method, the 2-norm of the diference between the monitoring settlement value and the calculated settlement value of the selected monitoring points in the four main material partitions of the rockfll area and the core wall area is used as the partition objective function to build a 4objective optimization model, as shown in the following equation: y � y (1) , y (2) , y (3) , y (4) T , where f (1) , f (2) , f (3) , and f (4) are the objective functions of rockfll areas I, II, and III and the core wall, respectively; m (i) is the number of monitoring points selected in the ith partition; n (i) is the characteristic time point selected for each monitoring point; U (i) * jk is the settlement calculation value of the jth monitoring point in the zone at the kth time point; and U (i) jk is the corresponding measured value. y (i) is the parameter to be back analysed in the partition, which is K, n, K b , and m in the Duncan-Chang E-B model and λ, α, β, and δ in the Merchant seven parameter model; y (i) min d and y (i) max d are the upper and lower limits of the parameters to be back analysed.

Sample Design and Surrogate Model Training.
Considering that it is difcult to ensure the accuracy of the inversion results in the target search space with too many dimensions, some parameters have little impact on the stress and deformation of the dam body, and it is difcult for some parameters to be accurately determined through indoor and on-site experiments. Tus, eight parameters, K, n, K b , m, λ, α, β, and δ of the Duncan-Chang EB model and Merchant viscoelastic model are fnally selected as the parameter set to be back analysed.
Te fnite element model illustrated in Figure 8 has 150,776 elements and 150,397 nodes. We obtained 520 sets of parameter samples using orthogonal test design and random sample design methods and performed 520 groups of FEM forward computations. Tese works were carried out using our self-developed secondary development model based on ABAQUS.
We utilize ANN, XGBoost, RF, and SVR for regression. Te parameter combinations are used as input and the FEM results are used as output. 75% of the data is used as the training set, while the remaining 25% is used as the testing set. Normalizing the fnal projected and measured values yields Figure 9. Te accuracy and generalization capabilities of the characterization training model increase when RMSE, MAE, and R 2 approach 100%. Figure 9 shows that the expected output from the ANN training model matches the predicted output best, and each partition has the maximum accuracy. Te ftting results of XGBoost and RF are inferior to those of ANN but comparable to SVR. When the sample size and kernel function mapping dimension are enormous, the calculation amount is large, and the ftting efect is low, making the SVR unsuitable. It should be noted that the type of ANN used in this study is the basic fully connected neural network. Based on various metrics shown in Figure 9, ANN is fully capable of ftting parameters to FEM results and outperforms the other three models in this regard.

Multiobjective Optimization.
We utilize NSGA-III to optimize the 4-objective model iteratively. Te algorithm's parameters are as follows: the initial population size is 520, the maximum evolutionary algebra is 400, the probability of variation is 0.1, and the cross-over and variation parameters are 30 and 20, respectively. Figure 10 shows the objective function values and mean values for each iteration.  Core wall  21  0  9  30  Rockfll area I  9  4  3  18  Rockfll area II  0  0  6  6  Rockfll area III  0  12  0  12  Sum  30  16 18 66

Core Wall Section
Core Wall Section

Structural Control and Health Monitoring
Te multiobjective evaluation index evaluates the convergence and dispersion of the optimal solution set. Te hypervolume indicator is consistent with the optimization solution set. Its physical meaning is the hypercube volume formed by the solution set individual and the reference point in the target space. Te larger the indicator value is, the more dominant the optimization solution set. Te reference point can be set as (1.1, 1.1, 1.1, 1.1) when the leading edge is unknown. Figure 11 shows the calculated hypervolume indicator.
After 100 generations of NSGA-III iteration, objective function value and the hypervolume indicator show that population evolution converges to a higher level. Figure 12 shows a Pareto frontier with 255 Pareto optimal options after optimization. Te linear weights for the goal functions are 0.25 for rockfll area I, 0.20 for rockfll II, 0.25 for rockfll area III, and 0.30 for the core wall. Te parameter combination with the minimal value of 0.25f 1 + 0.2f 2 + 0.25f 3 + 0.3f 4 is the fnal multiobjective optimization parameter, which is shown as the yellow line in Figure 12.

Evaluation of the Multiobjective Parameter Inversion
Efect. To demonstrate the signifcance of SAA monitoring data in parameter inversion and to validate the rationality of the parameter inversion method proposed, we similarly substitute the parameter inversion results without SAA data        Table 4 shows inversion values of the parameters. Table 4 shows that the inclusion of SAA monitoring data has a signifcant impact on the parameter inversion results, which confrms the importance of SAA in the analysis of parameter inversion. To further verify the efectiveness of the parameters obtained by the inversion method in evaluating the overall safety level of the dam with FEM calculations, we conducted calculations using these two sets of parameters in the fnite element model. Te result of settlement cloud map of the dam is shown in Figure 13, which displays the top view, the 3-3 monitoring section, and the core wall section.
Based on the calculation results, it can be observed that the inclusion of SAA data in the parameter inversion analysis resulted in lower maximum settlement calculated values during the various stages of dam construction. Prior to the frst impoundment in December 2020, the maximum settlement calculated with SAA data was 3.03 m, which accounted for 1.03% of the maximum dam height. Similarly, during the second impoundment in June 2021, the maximum settlement with SAA data was 3.49 m, accounting for 1.18% of the maximum dam height. Finally, at the completion of dam flling in December 2021, the maximum settlement calculated with SAA data was 3.69 m, accounting for 1.25% of the maximum dam height. In contrast, the maximum settlement values calculated without SAA data were higher for all stages of construction, with values of 5.27 m, 5.93 m, and 6.20 m, accounting for 1.79%, 2.01%, and 2.10% of the maximum dam height, respectively. Based on the parameters, the overall FEM result with SAA data of the dam is within a reasonable range, and the distributions and development trends of the FEM results follow the standard rockfll dam deformation laws.
We have extracted the calculated values of some single monitoring points, whose location can be found in the subgraph c in Figure 13, from the FEM calculation with or without SAA data, as well as the instrument monitoring values for comparison. Specifc data is illustrated in Figure 14. Figure 14 indicates that the FEM calculation results obtained by multiobjective optimization parameters conform to the general law in the distribution law of settlement and can achieve a good approximation of the numerical values between local monitoring points and values. However, when SAA data are missing, the parameter inversion results are bad, indicating the importance of SAA monitoring data.
Te results show that the multiobjective parameter inversion method can well realize the global approximation and local optimization in the multiobjective parameter inversion of a 300 m high rockfll dam. Te results also show that the optimization parameters can well refect the deformation characteristics of rockfll materials under actual engineering conditions and that the method is accurate and reliable.

Conclusions
Tis study was the frst to utilize SAA in a 300 m ultrahigh rockfll dam, which signifcantly enhanced the monitoring capabilities of the dam while minimizing construction-related interference and difculties. SAAs can be vertically buried in the inclinometer pipe for long-term monitoring and easily replaced. With numerous monitoring points, SAAs provide more comprehensive internal deformation monitoring than other instruments. Tis ultrahigh rockfll dam has 1147 efective monitoring points, enriching the monitoring data and enabling the use of multisource data for parameter inversion.
Using new settlement monitoring data from the SAA instrument, we proposed a multiobjective parameter inversion method that adapts to the multimaterial zoning and multisource monitoring layout of the 300 m rockfll dam. We used the NSGA-III algorithm to optimize and coordinate the inversion process and fully utilized the multisource monitoring data of each subarea to enhance the accuracy of parameter inversion. We optimize the overall parameters of the rockfll dam and coordinate with the multimaterial zones, so that the parameters of each material zone can be optimized at the same time. Furthermore, through parameter inversion, we have validated the signifcance of SAA data, demonstrating that SAA can contribute to both deformation monitoring and prediction of dam body simultaneously.
To improve the accuracy of monitoring data in the future, we will strive to investigate ways to properly process SAA data when they fuctuate signifcantly. To provide a more thorough result, we will simultaneously enrich the parameter inversion process and update and refne the surrogate model and optimization method utilized in parameter inversion.

Data Availability
Te data that support the fndings of this study are available from the corresponding author upon reasonable request.

Conflicts of Interest
Te authors declare that they have no conficts of interest.