A dam progress simulation parameter dynamic updating method based on incremental learning

Abstract

The application discloses a dam progress simulation parameter dynamic updating method based on incremental learning, mainly comprising the following steps: acquiring construction sensing data flow; adopting an isolated forest algorithm to detect abnormal values of the construction sensing data flow; adopting a cubic spline interpolation method to interpolate and fill in missing values in the construction sensing data flow; using Gaussian mixture distribution to model simulation parameter distribution of the time series data column after interpolation and filling; using an online kernel density estimation method to continuously learn new input construction sensing data flow data after Gaussian mixture distribution modeling, and update new Gaussian mixture distribution, and output dynamic updated simulation parameters; inputting the dynamic updated simulation parameters into a rock-fill dam construction simulation model, wherein the rock-fill dam construction simulation model changes with the dynamic change of the construction simulation parameters, and outputs dynamic changes conforming to actual situations, thereby supporting high-quality intelligent construction in engineering sites and providing decision support for dam progress management.

Description

Technical Field

[0001] The present invention belongs to the field of dam simulation, and in particular relates to a method for dynamically updating dam progress simulation parameters based on incremental learning. Background Art

[0002] Rockfill dams have the characteristics of being able to make full use of local materials, adapt to various terrain, geological and climatic conditions, strong earthquake resistance, low cost and simple structure. They have developed into one of the most widely used dam types in water conservancy and hydropower project construction at home and abroad. [1] .

[0003] Construction simulation technology is considered to be an effective tool for describing and analyzing the complex construction process of high core rockfill dams, predicting construction progress, and optimizing construction plans. Construction simulation parameters directly affect the accuracy of high core rockfill dam simulation results. [2] Updating the construction simulation parameter distribution refers to the integration, processing, and analysis of key data indicators perceived during the dam construction process. As an important component of the simulation model, simulation parameters are repeatedly called upon during the simulation clock process and are key to ensuring the accuracy of simulation results. This is essentially a dynamic change problem, and adopting appropriate modeling and updating methods is the core of the construction simulation parameter distribution problem.

[0004] In traditional high-core rockfill dam construction simulation models, simulation parameters are mostly selected based on similar project experience or fitted into a certain distribution based on monitoring data and randomly sampled during the simulation. This makes it difficult to dynamically and timely reflect changes in simulation parameters, which has a certain impact on the accuracy of the simulation results.

[0005] With the rapid development of science and technology, particularly real-time monitoring and deep learning, the construction of rockfill dams generates massive amounts of sensory data. These data include roller instantaneous speed, roller deflection angle, overlap width, stagger width, current compaction progress, surface boundary shape, surface flow unit division scheme, and other construction simulation parameters. Leveraging this data to update rockfill dam construction simulation parameters is crucial for improving the accuracy of rockfill dam construction simulation models. However, during actual rockfill dam construction, due to the influence of various factors and the occurrence of unexpected situations, the sensory data obtained often contains a large amount of noise, necessitating detection and processing. Furthermore, the massive and real-time sensory data stream poses challenges to the dynamic updating of rockfill dam construction simulation models.

[0006] Later, scholars [3] The Bayesian update method is used to update parameters, but it is still necessary to assume that the simulation parameters follow a certain form of distribution. [4] A modeling method based on the Dirichlet mixture model was proposed, and some scholars [5]Innovatively, Bayesian field theory is used for adaptive modeling. With the development of artificial intelligence, many scholars have begun to explore the use of machine learning methods to dynamically predict simulation parameters. Xiao et al. [6] An adaptive chaotic differential evolution support vector machine method is proposed to dynamically update the simulation parameters of the diversion tunnel construction; Wang et al. [7] An improved long short-term memory neural network is used to predict the cable machine operating time parameter in the high arch dam construction simulation model; Zhang et al. [8] A particle swarm optimization multilayer perceptron model is proposed to time series predict the dam area temperature and embed it into the warehouse surface refined simulation model. The above studies all use time series methods to predict simulation parameters, lacking consideration of the dynamic influence of weather, construction machinery and other factors. To address this issue, Lv et al. [9] Considering the influence of vehicle type, weather conditions and other factors, an improved grey wolf algorithm optimized XGBoost model is used to predict the driving speed in the earthwork transportation construction simulation model. Song et al.

[10] A probability prediction model of arch dam construction simulation parameters is established considering the influence of meteorological factors. However, these studies all use all historical data as the same database to build the prediction model, and do not consider the difference between new working conditions and historical working conditions, which is difficult to meet the dynamic change of simulation parameters.

[0007] In summary, the existing research on rockfill dam construction simulation lacks the detection and processing of abnormal values in the perception data, and still relies on manual assistance for real-time data stream detection, analysis and processing. At the same time, the existing rockfill dam construction simulation parameter updating method has the problems of needing to set hyperparameters based on experience, low calculation efficiency leading to difficulty in meeting the real-time demand of simulation calculation, etc. The existing research on rockfill dam construction simulation is mostly based on the results of the whole process of rockfill dam construction simulation, so it cannot be analyzed and guided during the rockfill dam construction simulation process, and it has no real-time feedback and control ability during the rockfill dam construction process.

[0008] [References]

[0009] [1] Zhong D H, Shi M N, Cui B, et al. Research Progress on Intelligent Construction of Dams[J]. Journal of Hydraulic Engineering, 2019, 50(01): 38-52.

[0010] [2] Lv F, Zhong D H, Yu J, et al. IGOA-MLP Dynamic Prediction Model of High Core Wall Rockfill Dam Construction Simulation Parameters Based on Transfer Learning Framework[J]. Journal of Hydraulic Engineering, 2023, 54(10): 1151-1162.

[0011] [3] Zhong D H, Guan T, Ren B Y. Simulation Research on High Arch Dam Construction Progress Based on Improved Resampling Method[J].

[0012] [4]HU W, ZHONG D, WU B, et al. Construction phase oriented dynamic simulation: taking RCC dam placement process as an example[J]. Journal of civil engineering and management, 2019, 25(7): 654-672.

[0013] [5]YU J. Construction simulation and optimization of complex long-distance diversion tunnel considering risk influence[D]. Tianjin University, 2019.

[0014] [6]XIAO Y, ZHONG D, WANG D, et al. Dynamic updating of simulation parameters for diversion tunnel construction based on ACDE-SVM[J].

[0015] [7]WANG G, YU J, WANG X, et al. Dynamic updating model of EMD-P-ILSTM for construction simulation parameters of high arch dam[J]. Journal of Hydroelectric Engineering, 2021, 40(12): 106-118.

[0016] [8]ZHANG J, YU J, REN B Y, et al. Construction simulation model of high core wall rockfill dam considering the influence of high-cold temperature[J]. Journal of Hydraulic Engineering, 2022, 53(02): 200-211.

[0017] [9]LV F, WANG J, CUI B, et al. An improved extreme gradient boosting approach to vehicle speed prediction for construction simulation of earthwork[J]. Automation in construction, 2020, 119: 103351.

[0018]

[10] SONG W S, GUAN T, REN B Y, et al. Multi-factor probability prediction method for construction simulation parameters of arch dam[J]. Journal of Hydroelectric Engineering, 2022, 41(09): 150-160. SUMMARY

[0019] The quality problems existing in the construction sensing data of the rockfill dam and the dynamic updating and incremental learning requirements existing in the distribution updating process of the rockfill dam construction simulation parameters cannot be met by the existing research results, and the present application provides a dam progress simulation parameter dynamic updating method based on incremental learning, which can process the rockfill dam construction sensing data and dynamically update the construction simulation parameter distribution, solve the problems and deficiencies existing in the existing research, and has very important theoretical and practical significance for the research on the rockfill dam construction sensing data processing and construction simulation parameter updating.

[0020] In order to solve the above technical problems, the present application provides a dam progress simulation parameter dynamic updating method based on incremental learning, mainly including:

[0021] S1, the construction sensing data stream obtained is stored in an excel file;

[0022] S2, the isolated forest algorithm is used for abnormal value detection of the construction sensing data stream; first, the excel file storing the construction sensing data stream is converted into a csv file, the speed time series data column of each type of construction machinery is extracted, then the isolated forest algorithm is used for abnormal value detection, wherein the contamination is set to 0.055; and the abnormal values are deleted, thereby forming a vacancy in the above time series data column;

[0023] S3, the cubic spline interpolation method is used to interpolate and fill the vacancy in the construction sensing data stream obtained in step S2, and the process is as follows:

[0024] 3-a) interval division: given a set of data points The part between adjacent data points on the x-axis is regarded as an interval, so that the entire range is divided into a intervals Wherein, i=0,1,…,a-1;

[0025] 3-b) construct a cubic polynomial: in each interval A cubic polynomial is constructed: Wherein, a i ,b i ,c i ,d i are the coefficients of the multiple top formula;

[0026] The cubic multiple top formula meets the following conditions:

[0027] Continuity condition: function value continuity: And First-order derivative continuity: Second-order derivative continuity:

[0028] Boundary conditions: and

[0029] According to the above continuity conditions and boundary conditions, a linear equation system is established to solve the coefficient a of the cubic polynomial. i ,b i ,c i ,d i ;

[0030] 3-c) Construct spline function: Use the coefficients obtained in step 3-b) to construct the cubic spline function for each interval Thus, the spline interpolation function on the entire domain is obtained;

[0031] 3-d) Interpolation evaluation: For any given x value, find the corresponding interval and the corresponding polynomial Calculate the value of the missing value, which is the interpolation result;

[0032] S4. Use Gaussian mixture distribution to simulate parameter distribution modeling of the time series data column after the gap value interpolation filling, as follows:

[0033] Suppose there is an n-dimensional data set with an arbitrary vector x. If the vector x obeys a multivariate Gaussian distribution, the probability density estimation function of the vector x is expressed as follows:

[0034]

[0035] In formula (1), μ is the n-dimensional mean vector, ∑ is an n-dimensional covariance matrix, and the two are summed up to φ(x|μ,∑);

[0036] When there are multiple multivariate Gaussian components, they are independent of each other and superimposed on each other, and the convex function they form is a Gaussian mixture distribution; the probability density estimation function of the Gaussian mixture distribution is as follows:

[0037]

[0038] In formula (2), k is the number of multivariate Gaussian components, and its domain is a positive integer; μ z is the mean vector of the z-th multivariate Gaussian component; ∑ z is the covariance matrix of z multivariate Gaussian components; α z is the weight of the z-th multivariate Gaussian component, and α z Satisfy α z >0,∑ z α z =1;

[0039] S5. Use the online kernel density estimation method to dynamically update the simulation parameters after the Gaussian mixture distribution modeling in step S4, that is, use the online kernel density estimation method to update the parameters in formula (2) in step S4, including adaptive update, local learning and dynamic model adjustment, so as to achieve continuous learning of the newly input construction perception data stream data after the interpolation and filling processing in step S3 after the Gaussian mixture distribution modeling in step S4, and update the new Gaussian mixture distribution to output the dynamically updated simulation parameters.

[0040] Furthermore, the method for dynamically updating dam progress simulation parameters of the present invention includes:

[0041] In step S1, the process of acquiring the construction perception data stream is: using the airborne monitoring terminal to obtain the positioning differential information of the construction machinery used through the differential base station and the satellite positioning information obtained by the observation positioning satellite to obtain the real-time spatial positioning information of the construction machinery, and send the real-time spatial positioning information to the relay station, and then send it to the server; the server calculates the speed data of various types of construction machinery used through the program to form a construction perception data stream, and stores the construction perception data stream in an Excel file.

[0042] The process of step S2 is as follows:

[0043] 2-a) Constructing an isolation tree: For a dataset containing time-series speed data for various types of construction machinery, randomly sample subsamples. Then, randomly select a feature and a random split value for that feature. Based on this split value, the subsample is divided into two subsets. This process is recursively repeated until each subsample is isolated into a single point or the preset tree depth is reached.

[0044] 2-b) Constructing a forest: Repeat the above 2-a) process to construct multiple isolated trees to form an isolation forest;

[0045] 2-c) Calculate path length: For each data point, calculate its path length in each isolated tree;

[0046] 2-d) Calculate anomaly score: Calculate the anomaly score of each point based on the path length obtained in step 2-c);

[0047] 2-e) Determine outliers: A point with an anomaly score higher than the set contamination threshold is determined as an outlier. The anomaly score corresponding to the outlier is the outlier value.

[0048] The process of step S5 is as follows:

[0049] 5-a) After the Gaussian mixture distribution modeling in step S4, the online kernel density estimation method automatically adjusts the mean vector, covariance matrix, and weight of each multivariate Gaussian component after the Gaussian mixture distribution modeling through an adaptive update strategy for the newly input construction perception data stream processed in step S3;

[0050] 5-b) For the parameters automatically adjusted in step 5-a), the online kernel density estimation method uses a local learning strategy to dynamically update only the Gaussian mixture distribution portion that is relevant to the newly input construction perception data stream processed in step S3. For other irrelevant Gaussian mixture distribution portions, the existing learning results are retained. Data correlation means that the Gaussian mixture distribution is less than or equal to the threshold YZ, and data irrelevant means that the Gaussian mixture distribution is greater than the threshold YZ. The formula for threshold YZ is as follows:

[0051]

[0052] In formula (3), X max Refers to the maximum value of the data set, X min It refers to the minimum value of the data set.

[0053] 5-c) In response to the difference between the newly input construction perception data stream data after processing in step S3 and the previously input construction perception data stream data, the online kernel density estimation method dynamically adjusts the number of multivariate Gaussian components of the Gaussian mixture distribution. The adjustment process is: by adding new multivariate Gaussian components to adapt to the data characteristics displayed in the newly input perception data stream, and by deleting old multivariate Gaussian components to remove components that cannot correctly display the characteristics of the existing sample data to avoid overfitting of the Gaussian mixture distribution.

[0054] Furthermore, the dynamically updated simulation parameters outputted in step S5 are inputted into the rockfill dam construction simulation model. The rockfill dam construction simulation model changes with the dynamic changes of the construction simulation parameters and outputs dynamic simulation results.

[0055] The simulation results output by the rockfill dam construction simulation model are based on the actual on-site conditions, using the construction perception data stream generated during construction and adopting dynamically updated simulation parameter inputs, so as to output dynamic changes that conform to the actual on-site conditions, support high-quality intelligent construction at the project site, and provide decision-making support for dam progress management.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] (1) The method of the present invention involves the detection and processing of anomaly in the perception data of rockfill dam construction simulation, wherein the isolation forest algorithm and cubic spline interpolation method are used to detect and process the outliers, which effectively improves the quality of the perception data stream.

[0058] (2) The method of the application involves a rockfill dam construction simulation parameter distribution modeling and dynamic updating incremental learning method (OKDE-GMM), which can ensure previous learning information and learn new information, and realize dynamic modeling and dynamic updating of rockfill dam construction simulation parameter distribution. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 is the simulation perception data anomaly detection and processing framework of the method of the application, which is used for Figure 1

[0060] Figure 2 is the OKDE-GMM framework of the simulation parameter distribution modeling and dynamic updating incremental learning method in the method of the application;

[0061] Figure 3 is the flowchart of the method of the application. DETAILED DESCRIPTION

[0062] The design concept of the dam progress simulation parameter dynamic updating method based on incremental learning proposed by the application is that the simulation perception data quality and the simulation parameter distribution updating efficiency are the keys to affect the accuracy of the rockfill dam simulation model, therefore, in order to further improve the dynamic updating rate and quality of the simulation model in the construction process, the influence of the two factors needs to be considered. However, in the traditional high core wall rockfill dam construction simulation model, the simulation parameters are usually selected according to similar engineering experience or monitoring data, fitted into a certain distribution, and randomly sampled in the simulation process. This method is usually determined in the design period, and therefore it is difficult to reflect the dynamic changes of the simulation parameters in real time, thereby affecting the accuracy of the simulation results. At the same time, the existing construction simulation parameter updating method needs manual expert experience to set the hyperparameters, and the low calculation efficiency leads to long simulation calculation time of the rockfill dam construction period, insufficient real-time performance, and inability to solve the contradiction between massive perception data and parameter dynamic updating. Therefore, obtaining high-quality construction simulation parameter perception data and applying it to the simulation parameter distribution modeling and dynamic updating are the key to solving the existing engineering problems and improving the simulation quality in the construction period.

[0063] The application proposes a dam progress simulation parameter dynamic updating method based on incremental learning, and the research framework of the construction simulation perception data anomaly detection and processing part is as shown in Figure 1 First, the original integrated unmanned roller multi-modal perception framework is used to obtain the construction perception data stream, then the isolation forest algorithm and the cubic spline interpolation method are used for anomaly value detection and processing of the perception data, and finally the construction perception data stream after anomaly value detection and processing is output. The research framework of the construction simulation parameter distribution modeling and dynamic updating method is as shown in Figure 2The abnormal value detection and processing of the roller compaction speed data stream and the transportation speed data stream are simulated by using the Gaussian mixture distribution in the dam progress simulation parameter dynamic updating method based on Newton interpolation incremental learning, the simulation parameter distribution modeling is performed by using the online kernel density estimation method, the simulation parameter distribution is dynamically updated, the simulation parameter is dynamically output, and the simulation parameter is input into the simulation model to obtain the changed simulation results.

[0064] The present application mainly includes two parts, the first part is a rockfill dam construction simulation perception data stream abnormality detection and processing method, and the second part is a real-time perception data stream construction simulation parameter distribution dynamic updating kernel density estimation algorithm.

[0065] The first part firstly inputs the real-time monitoring based on the used construction machinery (for example, the typical warehouse surface compaction speed of the compactor and the dam material transportation speed of the dump truck) data stream, secondly performs abnormal value detection on the perception data stream by using the isolation forest algorithm, then performs abnormal value processing on the perception data stream by using the cubic spline interpolation method, and finally outputs the processed perception data stream.

[0066] The second part introduces the concept of incremental learning based on the traditional parameter modeling method GMM, adopts the online kernel density estimation (OKDE) incremental learning framework, proposes a rockfill dam construction simulation parameter distribution modeling and dynamic updating method (OKDE-GMM), completes the modeling and dynamic updating of the rockfill dam construction simulation parameter distribution, and achieves incremental learning under complex construction environment.

[0067] The present application will be further described below in combination with the drawings and specific embodiments, but the following embodiments are by no means any limitation on the present application.

[0068] The present application proposes a dam progress simulation parameter dynamic updating method based on incremental learning, as shown in the accompanying drawings, which comprises the following steps: Figure 3

[0069] Step one, obtain the construction perception data stream;

[0070] The real-time spatial positioning information of the construction machinery is obtained by using the differential base station of the airborne monitoring terminal and the satellite positioning information obtained by the observation positioning satellite, and the real-time spatial positioning information is sent to the relay station and then to the server; the server calculates the speed data of various types of construction machinery by program to form the construction perception data stream.

[0071] ​The input simulation parameters of the current dynamic simulation model for rockfill dam construction mainly include two types: bin surface compaction simulation parameters and dam material transportation simulation parameters. Regarding bin surface compaction simulation parameters, the instantaneous travel speed of the roller is selected as a representative parameter in this embodiment; regarding dam material transportation simulation parameters, the present invention selects the travel speed of the dam material transportation dump truck as a representative parameter. The data acquisition of the present invention relies on the real-time monitoring system for rockfill dam construction developed by the National Key Laboratory of Intelligent Construction and Operation of Water Conservancy Engineering of Tianjin University to acquire the instantaneous travel speed data stream of the dam surface roller during the construction period, such as Figure 1 The obtained construction simulation perception data stream is stored in an Excel file for the next step of detecting and processing abnormalities in the rockfill dam construction simulation parameter perception data.

[0072] Step 2: For the acquired perception data stream, use the Isolation Forest algorithm to detect outliers. In this embodiment, Python 3.9 is used to write relevant code programs. First, the Excel file storing the construction perception data stream is converted into a CSV file, read, and the speed time series data columns of various types of construction machinery are extracted. Then, the Isolation Forest algorithm is used for outlier detection, where contamination is set to 0.055; running the code will identify abnormal data values ​​and mark them in red, then delete the abnormal values, leaving some vacant values ​​of the time series data, thereby forming vacant values ​​in the above-mentioned time series data columns.

[0073] The Isolation Forest algorithm is an algorithm used for anomaly detection. Its core idea is to use a tree structure to "isolate" observation points. Because outliers are few in number and different from normal points, they are easier to isolate. Its steps can be described as follows:

[0074] 2-a) Constructing an isolation tree: For a given data set, for example, a data set including speed time series data of various types of construction machinery in the present invention, random sampling is performed to form subsamples. Then, a feature and a random split value of the feature are randomly selected. The subsamples are divided into two subsets based on the split value. This process is recursively repeated until each subsample is isolated into a single point or the preset tree depth is reached.

[0075] 2-b) Constructing a forest: Repeat the above 2-a) process to construct multiple isolated trees to form an isolation forest;

[0076] 2-c) Calculate path length: For each data point, calculate its path length in each isolated tree;

[0077] 2-d) Calculate anomaly score: Calculate the anomaly score of each point based on the path length obtained in step 2-c);

[0078] 2-e) Determine outliers: Determine whether a data point is an outlier based on its anomaly score. Typically, points with scores above a certain threshold are considered outliers. Therefore, points with an anomaly score above the set contamination threshold (0.055) are considered outliers. The anomaly score corresponding to the outlier is the outlier value.

[0079] Step 3: Use cubic spline interpolation to fill missing values ​​in the simulation parameter time series data and mark them green, completing the detection and processing of anomalies in the rockfill dam simulation parameter perception data. The processed perception data shows significant quality improvements, reducing a large amount of noise and anomalies, which facilitates the modeling and dynamic updating of rockfill dam simulation parameter distribution.

[0080] The cubic spline interpolation method is used to interpolate and fill the missing values ​​in the construction perception data stream obtained in step S2. The processing process is as follows:

[0081] Cubic spline interpolation is a smooth interpolation method widely used in numerical analysis. It uses a series of cubic polynomials to approximate a given function, and these polynomials are smoothly connected between each data point. The cubic spline interpolation method approximates the given data points by constructing a series of cubic polynomials (each polynomial is a cubic equation) and inserting a polynomial between every two adjacent data points. These polynomials have continuous first-order and second-order derivatives at the endpoints of each segment, ensuring the smoothness of the entire function. Its steps can be described as follows:

[0082] 3-a) Interval partitioning: Given a set of data points The part between adjacent data points on the x-axis is considered as an interval, thereby dividing the entire range into a intervals Among them, i=0,1,…,a-1.

[0083] 3-b) Construct a cubic polynomial: In each interval On , construct a cubic polynomial: Among them, a i ,b i ,c i ,d i is the coefficient of the polynomial; it needs to be determined in a further step.

[0084] Application of continuity conditions: In order to ensure smooth transition of the function at the seams between regions, the following continuity conditions need to be met: Function value continuity: and First-order derivative is continuous: The second derivative is continuous:

[0085] Determination of boundary conditions: Select appropriate boundary conditions. In the present invention, natural boundary conditions are selected, namely: and Of course, other specific boundary conditions can also be selected.

[0086] Solve the coefficient equations: Based on the above continuity conditions and boundary conditions, establish a linear equation system to solve the coefficient a of the cubic polynomial i ,b i ,c i ,d i .

[0087] 3-c) Construct spline function: Use the coefficients obtained in step 3-b) to construct the cubic spline function for each interval Thus we get the spline interpolation function on the entire domain.

[0088] 3-d) Interpolation evaluation: For any given x value, find the corresponding interval and the corresponding polynomial Calculate the value of the missing value, which is the interpolation result;

[0089] Step 4: Use the Gaussian mixture model (GMM) module in the OKDE-GMM of the present invention to simulate parameter distribution modeling of the time series data column after the gap value interpolation is filled.

[0090] GMM is a typical probabilistic model used for clustering or density estimation of sample data. A single Gaussian distribution can only estimate simple probability density functions and fit ideal data samples. However, using multiple independent Gaussian distributions with different shape parameters can estimate more complex probability density functions and fit more realistic data samples. In theory, it can estimate the probability density distribution state of any situation.

[0091] First, let's define the probability density function of a multivariate Gaussian distribution (also known as a multivariate Gaussian component). Given an n-dimensional dataset containing an arbitrary vector x, if x follows a multivariate Gaussian distribution, the probability density function of x can be written as follows:

[0092]

[0093] In formula (1), μ is the n-dimensional mean vector, ∑ is an n-dimensional covariance matrix, and the two are summed up to φ(x|μ,∑);

[0094] When there are multiple multivariate Gaussian components, they are independent of each other and superimposed on each other, and the convex function they form is a Gaussian mixture distribution.

[0095] The probability density estimation function of the Gaussian mixture distribution is as follows:

[0096]

[0097] In formula (2), k is defined as the number of Gaussian mixture components. It is a hyper-parameter that needs to be manually input in the Gaussian mixture model. Its domain is a positive integer. The optimal value or better value can be selected by using optimization algorithms or manual multiple experiments based on engineering cases, such as the parameter update of rockfill dam simulation. z is the mean vector of the z-th multivariate Gaussian component; ∑ z is the covariance matrix of z multivariate Gaussian components; α z is the weight or mixture coefficient of the z-th multivariate Gaussian component, and α z Satisfy α z >0,∑ z α z =1.

[0098] The parameters of the Gaussian mixture model mentioned above are typically solved using the maximum likelihood function. One commonly used tool is the Expectation Maximization (EM) algorithm. The EM algorithm performs maximum likelihood estimation on latent variables (incomplete data) in an existing sample dataset. Through continuous iteration, it reduces complexity and converges to the optimal value.

[0099] In GMM solution, the basic idea of ​​EM algorithm is to assume that the existing sample set containing incomplete data is independent and identically distributed in the Gaussian mixture model, and then estimate the parameter distribution based on the obtained GMM parameters in each iteration, and then obtain the new GMM parameters under the maximum likelihood condition.

[0100] However, the EM algorithm has a significant drawback: its slow convergence speed. Sometimes it converges to a local minimum and fails to obtain the global optimal solution, which affects the estimation effect. It is precisely because of the real-time demand for fast convergence and the pursuit of the global optimal solution that the present invention introduces the Online Kernel Density Estimation (OKDE) method to update the simulation parameters.

[0101] Step 5: Use the online kernel density estimation method to dynamically update the simulation parameters after the Gaussian mixture distribution modeling in step 4.

[0102] Kernel Density Estimation (KDE) is a nonparametric statistical method for estimating probability density functions. The basic idea is to place a kernel (usually a probability density function, such as a Gaussian kernel) at each data point and then stack these kernels to form a smooth probability density estimate.

[0103] The mathematical principle of KDE involves the convolution of probability density functions, and its core formula is:

[0104]

[0105] in, is the probability density estimate at point m, T is the number of samples, h is the bandwidth (determines the width of the kernel), M t is a sample point, and Ker(·) is a kernel function.

[0106] The core mathematical formula of KDE involves the convolution of probability density functions, where the choice of bandwidth (kernel width) has a significant impact on the results. However, traditional KDE is mainly applied to static data sets. In the context of dynamic data streams, Online Kernel Density Estimation (OKDE) has emerged. OKDE is a powerful statistical method that dynamically adapts to changing data streams and updates probability density estimates in real time, providing an effective tool for real-time analysis and monitoring. The basic idea of ​​OKDE is rooted in the framework of kernel density estimation (KDE), and its key step is to process data streams through an incremental method to make it more adaptable to the characteristics of dynamic data.

[0107] In this paper, OKDE introduces new mechanisms and mathematical principles to adapt to dynamic data flows. The basic idea is to continuously update the kernel density estimate through an incremental approach. The mathematical principle involves incremental learning and calculation to achieve dynamic updates.

[0108] The parameters in equation (2) are updated using an online kernel density estimation method, including adaptive updating, local learning, and dynamic model adjustment. This allows continuous learning of the newly input construction perception data stream after interpolation and padding in step S3 after the Gaussian mixture distribution modeling in step S4, and updates the new Gaussian mixture distribution to output dynamically updated simulation parameters. The specific process is as follows:

[0109] 5-a) The initialization phase involves setting up the initial kernel density estimate. During this phase, it's crucial to select an appropriate kernel function and bandwidth, which directly impacts the accuracy and sensitivity of the estimate. After the Gaussian mixture distribution modeling in step 4, the online kernel density estimation method uses an adaptive update strategy to automatically adjust the mean vector, covariance matrix, and weights of each multivariate Gaussian component modeled using the Gaussian mixture distribution for the newly input construction perception data stream processed in step 3.

[0110] 5-b) For the parameters automatically adjusted in step 5-a), the online kernel density estimation method uses a local learning strategy to dynamically update only the Gaussian mixture distribution portion that is relevant to the newly input construction perception data stream processed in step S3. For other irrelevant Gaussian mixture distribution portions, the existing learning results are retained. Data correlation means that the Gaussian mixture distribution is less than or equal to the threshold YZ, and data irrelevant means that the Gaussian mixture distribution is greater than the threshold YZ. The formula for threshold YZ is as follows:

[0111]

[0112] Where, X max Refers to the maximum value of the data set, X min It refers to the minimum value of the data set.

[0113] Bandwidth adjustment is to better adapt to changes in data distribution. In a dynamic environment, the distribution of data may evolve over time, so dynamic bandwidth adjustment helps maintain the sensitivity of the estimation and avoid over-smoothing or over-fitting.

[0114] 5-c) Probability density estimation is the final output of OKDE. Based on the discrepancies between the newly input construction perception data stream processed in step 3 and the previously input construction perception data stream, the online kernel density estimation method dynamically adjusts the number of multivariate Gaussian components in the Gaussian mixture distribution. This adjustment process involves adding new multivariate Gaussian components to adapt to the data characteristics displayed in the newly input perception data stream and removing old multivariate Gaussian components to remove components that do not correctly reflect the characteristics of the existing sample data to avoid overfitting the Gaussian mixture distribution.

[0115] At any moment, the probability density estimate of the data can be obtained by calculating the current kernel density estimate. This provides users with real-time, dynamic data distribution information, providing strong support for decision-making. Through these steps, OKDE can adapt to changes in data streams in real time and provide dynamic probability density estimates, making it widely applicable in real-time analysis and monitoring.

[0116] Step six, OKDE-GMM realizes the function of incremental learning, and can perform rockfill dam construction simulation parameter distribution modeling and dynamic updating on the real-time sensing data stream obtained by the real-time monitoring system of the rockfill dam after abnormal value detection and processing.

[0117] The dynamically updated simulation parameters output in step five are input into the rockfill dam construction simulation model, and the rockfill dam construction simulation model changes with the dynamic change of the construction simulation parameters, and outputs dynamic simulation results. The simulation results output by the rockfill dam construction simulation model rely on the actual situation on site, use the construction sensing data stream generated during construction, adopt dynamically updated simulation parameters input, and thus output dynamic changes conforming to the actual situation on site, support high-quality intelligent construction on site, and provide decision support for dam progress management.

[0118] In summary, by using the OKDE-GMM method proposed in the present application, the processed sensing data stream is used, the kernel density estimation is continuously updated through the incremental method, dynamic data distribution information is provided to provide strong support for decision-making, and rockfill dam simulation parameter distribution modeling and dynamic updating are realized.

[0119] Although the present application is described above in conjunction with the drawings, the present application is not limited to the specific embodiments described above, and the specific embodiments described above are merely illustrative rather than limiting, and many improvements and changes can be made by those of ordinary skill in the art under the inspiration of the present application without departing from the purpose of the present application, and these all belong to the protection of the present application.

Claims

1. A method for dynamically updating dam progress simulation parameters based on incremental learning, characterized in that: The following steps are involved: S1. Store the acquired construction perception data stream into an Excel file; S2. Use the isolation forest algorithm to detect outliers in the construction perception data stream. First, convert the Excel file storing the construction perception data stream into a CSV file, extract the speed time series data column of various types of construction machinery, and then use the isolation forest algorithm to detect outliers. The contamination value is set to 0.

055. The outliers are deleted, thus forming vacancies in the time series data column. S3. Use the cubic spline interpolation method to interpolate and fill the missing values ​​in the construction perception data stream obtained in step S2. The processing process is as follows: 3-a) Interval partitioning: Given a set of data points The part between adjacent data points on the x-axis is considered as an interval, thereby dividing the entire range into a intervals Among them, i=0,1,…,a-1; 3-b) Construct a cubic polynomial: In each interval On , construct a cubic polynomial: Among them, a i ,b i ,c i ,d i is the coefficient of the polynomial; The cubic polynomial satisfies the following conditions simultaneously: Continuity condition: Function value is continuous: and First-order derivative is continuous: The second derivative is continuous: Boundary conditions: and According to the above continuity conditions and boundary conditions, a linear equation system is established to solve the coefficient a of the cubic polynomial. i ,b i ,c i ,d i ; 3-c) Construct spline function: Use the coefficients obtained in step 3-b) to construct the cubic spline function for each interval Thus, the spline interpolation function on the entire domain is obtained; 3-d) Interpolation evaluation: For any given x value, find the corresponding interval and the corresponding polynomial Calculate the value of the missing value, which is the interpolation result; S4. Use Gaussian mixture distribution to simulate parameter distribution modeling of the time series data column after the gap value interpolation filling, as follows: Suppose there is an n-dimensional data set with an arbitrary vector x. If the vector x obeys a multivariate Gaussian distribution, the probability density estimation function of the vector x is expressed as follows: In formula (1), μ is the n-dimensional mean vector, ∑ is an n-dimensional covariance matrix, and the two are summed up to φ(x|μ,∑); When there are multiple multivariate Gaussian components, they are independent of each other and superimposed on each other, and the convex function they form is a Gaussian mixture distribution; the probability density estimation function of the Gaussian mixture distribution is as follows: In formula (2), k is the number of multivariate Gaussian components, and its domain is a positive integer; μ z is the mean vector of the z-th multivariate Gaussian component; ∑ z is the covariance matrix of z multivariate Gaussian components; α z is the weight of the z-th multivariate Gaussian component, and α z Satisfy α z >0,∑ z α z =1; S5. Use the online kernel density estimation method to dynamically update the simulation parameters after the Gaussian mixture distribution modeling in step S4, that is, use the online kernel density estimation method to update the parameters in formula (2) in step S4, including adaptive update, local learning and dynamic model adjustment, so as to achieve continuous learning of the newly input construction perception data stream data after the interpolation and filling processing in step S3 after the Gaussian mixture distribution modeling in step S4, and update the new Gaussian mixture distribution to output the dynamically updated simulation parameters.

2. The method for dynamically updating dam progress simulation parameters according to claim 1, characterized in that: In step S1, the process of acquiring the construction perception data stream is as follows: An airborne monitoring terminal is used to obtain the positioning differential information of the construction machinery used through a differential base station and the satellite positioning information obtained by the observation positioning satellite to obtain the real-time spatial positioning information of the construction machinery, and the real-time spatial positioning information is sent to the relay station and then to the server; the server calculates the speed data of various types of construction machinery used through a program to form a construction perception data stream, and stores the construction perception data stream in an Excel file.

3. The method for dynamically updating dam progress simulation parameters according to claim 1, characterized in that: The process of step S2 is as follows: 2-a) Constructing an isolation tree: For a dataset containing time-series speed data for various types of construction machinery, randomly sample subsamples. Then, randomly select a feature and a random split value for that feature. Based on this split value, the subsample is divided into two subsets. This process is recursively repeated until each subsample is isolated into a single point or the preset tree depth is reached. 2-b) Constructing a forest: Repeat the above 2-a) process to construct multiple isolated trees to form an isolation forest; 2-c) Calculate path length: For each data point, calculate its path length in each isolated tree; 2-d) Calculate anomaly score: Calculate the anomaly score of each point based on the path length obtained in step 2-c); 2-e) Determine outliers: A point with an anomaly score higher than the set contamination threshold is determined as an outlier. The anomaly score corresponding to the outlier is the outlier value.

4. The method for dynamically updating dam progress simulation parameters according to claim 1, characterized in that: The process of step S5 is as follows: 5-a) After the Gaussian mixture distribution modeling in step S4, the online kernel density estimation method automatically adjusts the mean vector, covariance matrix, and weight of each multivariate Gaussian component after the Gaussian mixture distribution modeling through an adaptive update strategy for the newly input construction perception data stream processed in step S3; 5-b) For the parameters automatically adjusted in step 5-a), the online kernel density estimation method uses a local learning strategy to dynamically update only the Gaussian mixture distribution portion that is relevant to the newly input construction perception data stream processed in step S3. For other irrelevant Gaussian mixture distribution portions, the existing learning results are retained. Data correlation means that the Gaussian mixture distribution is less than or equal to the threshold YZ, and data irrelevant means that the Gaussian mixture distribution is greater than the threshold YZ. The formula for threshold YZ is as follows: In formula (3), X max Refers to the maximum value of the data set, X min Refers to the minimum value of the data set; 5-c) In response to the difference between the newly input construction perception data stream data after processing in step S3 and the previously input construction perception data stream data, the online kernel density estimation method dynamically adjusts the number of multivariate Gaussian components of the Gaussian mixture distribution. The adjustment process is: by adding new multivariate Gaussian components to adapt to the data characteristics displayed in the newly input perception data stream, and by deleting old multivariate Gaussian components to remove components that cannot correctly display the characteristics of the existing sample data to avoid overfitting of the Gaussian mixture distribution.

5. The method for dynamically updating dam progress simulation parameters according to claim 1, characterized in that: The dynamically updated simulation parameters outputted in step S5 are inputted into the rockfill dam construction simulation model. The rockfill dam construction simulation model changes with the dynamic changes of the construction simulation parameters and outputs dynamic simulation results.

6. The method for dynamically updating dam progress simulation parameters according to claim 5, characterized in that: The simulation results output by the rockfill dam construction simulation model are based on the actual on-site conditions, using the construction perception data stream generated during construction and adopting dynamically updated simulation parameter inputs, so as to output dynamic changes that conform to the actual on-site conditions, support high-quality intelligent construction at the project site, and provide decision-making support for dam progress management.

Images