Method for operating data assimilation in elastic-plastic finite element technique
The data assimilation method addresses the inefficiencies in estimating nonlinear structure states by calculating sensitivity and optimizing updates using 'number of variables', effectively reducing computational burden and ensuring accurate estimation of internal properties.
Patent Information
- Application Number
- JP2023091968
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-12
- Publication Date
- 2024-11-22
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing data assimilation methods struggle to efficiently estimate the internal state of highly nonlinear structures, such as those exhibiting elasto-plastic behavior, due to challenges in calculating sensitivity to analysis errors and the increased computational burden of brute-force searches, especially when dealing with a large number of unknown variables.
A data assimilation method that calculates sensitivity to observation errors using a sensitivity calculation unit and an optimal solution search unit, employing a gradient method with constrained updates and reduced calculations by utilizing the concept of 'number of variables' to track changes in state variables over time, thereby minimizing the number of required updates.
Enables efficient estimation of internal physical properties like hardness and damage in nonlinear structures by reducing computational load and ensuring convergence to the optimal solution, even with a large number of unknown variables.
Smart Images

Figure 2024163826000001_ABST
Abstract
Description
Detailed Description of the Invention [Technical field]
[0001] The present invention relates to a data assimilation calculation method for nonlinear finite element analysis that is primarily intended for vibration problems. [Background technology]
[0002] Data assimilation has been used to estimate the current internal stiffness of embankments from the results of measuring embankment deformation as shown in Non-Patent Document 1, and to estimate the internal state of a bridge from the bridge acceleration response as shown in Non-Patent Document 2. When estimating the internal state of an object using data assimilation, appropriate variables are assumed, the finite element method is executed, and the variables are repeatedly updated to search for optimal variables so that the error between the values calculated by the finite element method and the measured values is minimized.
[0003] Representative data assimilation methods include the extended Kalman filter (Non-Patent Document 3), the ensemble Kalman filter (Non-Patent Document 4), and the particle filter (Non-Patent Document 5). These methods estimate the forecast error and find the pair of unknowns that minimizes the analytical error (error between the measured value and the calculated value).
[0004] In the Kalman filter and ensemble Kalman filter, the sensitivity of the unknowns to analytical errors (how much the analytical error changes when the unknowns change slightly) is calculated, and the unknown variables are updated to find the optimal solution in order to minimize the error between the measured values and the calculated values.
[0005] In the particle filter, unknown variables are randomly estimated, and a brute force search is performed to find the combination of unknown variables that minimizes the analysis error.
[0006] In a method such as the Kalman filter, which determines the sensitivity of unknowns to analytical errors and obtains the optimal solution by the gradient method, the optimal solution can be reached with the same number of searches even if the number of unknowns increases. However, when a highly nonlinear structure is the target, the Kalman filter may not be able to reach the optimal solution. In addition, in the elastic-plastic finite element method, the time evolution of state variables such as stress, displacement, and damage variables is complex, and when a highly nonlinear object is numerically analyzed, the behavior of the state evolution becomes highly chaotic. Therefore, when performing finite element analysis of a nonlinear object, it is not easy to determine the sensitivity of unknowns to analytical errors, and until now, methods for calculating sensitivity have only been used in situations where differential calculations are easy, such as the sensitivity of the hydraulic conductivity to pressure head and the sensitivity of the stiffness matrix to nodal displacement.
[0007] In methods such as particle filters that search for optimal solutions by exhaustive search, the realized values obtained by analysis are used as a solution space to search for optimal solutions, so data assimilation can be performed on highly nonlinear objects and implementation is easy. However, because it is an exhaustive search, the amount of calculation increases, and when there are a large number of unknowns to be identified, the number of searches increases, and it may not be possible to reach the optimal solution. Therefore, it is difficult to apply this method to problems that require unknown variables for each element of the finite element method, such as estimating the distribution of hydraulic conductivity. [Prior art documents] [Non-patent literature]
[0008] [Non-Patent Document 1] Makoto Suzuko and Kiyoshi Ishii: Estimation of spatial distribution of soil constants using extended Kalman filter, Journal of Japan Society of Civil Engineers, Vol.406, 3-11, pp.71-78, 1989. [Non-Patent Document 2] Hirohiro Matsuoka, Kiyoyuki Kaito, Munemasa Tokunaga, Tsutomu Watanabe, Masamichi Sogabe: Estimation of Displacement Response Based on Bridge Acceleration Response During Train Passage Using Sequential Data Assimilation, Journal of Japan Society of Civil Engineers, Proceedings of the Society of Civil Engineers, Vol. 69, No. 3, 2013. [Non-Patent Document 3] Toru Katayama: Applied Kalman Filter, Asakura Publishing, pp.73-78, 1983. [Non-Patent Document 4] Hiromasa Yamaguchi and Eiichi Nakakita: Development of an assimilation method for Doppler radar information using EnKF and rainfall forecast using a conceptual model, Disaster Prevention Research Institute, Kyoto University, No. 50B, pp. 569-577, 2007. [Non-Patent Document 5] Nasrellah, HA and Manohar, CS: Particle filter for structural system identification using multiple test and sensor data: A computational combined and experimental study, Structural Control and Health Monitoring, Vol.18, PP.99-120, 2011. Summary of the Invention DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0009] The present invention has been made in view of the above-mentioned points, and an object of the present invention is to provide a calculation method for measuring the vibration of a vibrating nonlinear structure and estimating the internal state.
[0010] In order to solve the above-mentioned problems and realize data assimilation, the present invention provides a method for estimating the internal physical properties of a highly nonlinear structure by setting the internal state of the object as an unknown variable from the observation data and searching for it in a scene where the structure is vibrating. By using this method, even if the number of unknowns increases in a nonlinear vibration problem, it becomes possible to arrive at an optimal solution with a limited number of updates of the unknowns.
[0011] The present invention can be used in situations where the internal state of an object can be estimated by applying an external force to an object to vibrate it and observing the time-series data of the vibration, or where the internal physical properties of an object, such as hardness, density, and degree of damage, can be estimated by measuring the vibrations of an object that is already vibrating due to the influence of wind, traffic load, etc.
[0012] The data assimilation method of the present invention is characterized by having a sensitivity calculation unit that determines the sensitivity of unknowns to observation errors, and an optimal solution search unit that estimates unknown variables such as physical properties and internal states that match the observation results using the sensitivities of the unknown variables.
[0013] Here, the observation error means the difference between the measurement result and the result estimated by numerical analysis, and the sensitivity of the unknown to the analytical error means the ratio of the amount of variation in the analytical error to the amount of variation in the unknown variable.
[0014] In the data assimilation method of the present invention, the sensitivity of the unknowns to the analytical errors obtained in the sensitivity calculation section is passed on to the optimal solution search section, and a calculation is performed to search for unknown variables that minimize the observation error. [Brief description of the drawings]
[0015] [Figure 1] 2 is a calculation flow of a data assimilation method according to the present invention. [Diagram 2] 1 is a flowchart showing a process in which a change in physical properties brings about a change in an internal state quantity. [Diagram 3] 1 is a computation graph showing the path of error propagation of a function f. [Figure 4] This is an image of how the effects of changes in physical properties in a vibration problem propagate to the surrounding area. [Diagram 5] 1 is a calculation graph showing a path of error propagation within an element in an elastic-plastic analysis. [Figure 6] FIG. 1 is a conceptual diagram of an operation for adjusting the scale of unknown variables and its effect. [Figure 7] 1 is a calculation flow of data assimilation according to the present invention in elastic-plastic analysis. [Figure 8] This is a block universe model that simulates the computational graph of the elastic-plastic finite element method. [Figure 9] This is a simplified model of the calculation graph for the elastic-plastic finite element method. [Figure 10] FIG. 13 is a conceptual diagram showing a range in which a variable number is propagated. [Figure 11]This is a spatial distribution diagram of the omission criteria for the number of variables. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0016] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of a data assimilation calculation method according to the present invention will be described below with reference to the drawings.
[0017] Fig. 1 shows a series of flows from setting analysis variables in the data assimilation calculation method in this embodiment through data assimilation calculation to outputting the results. As shown in Fig. 1, the data assimilation calculation in this embodiment uses acquired observation data and performs processing calculations in the order of sensitivity calculation section → optimal solution search section, and outputs the calculation results.
[0018] <Calculation in sensitivity calculation section> In elastoplastic analysis, the physical properties of an object, such as its density, hardness, and strength, are used as input variables to reproduce all kinds of nonlinear phenomena, such as the way an object vibrates, through numerical analysis. In numerical analysis, state variables, such as stress, strain, and the degree of damage to an object, are calculated from the values of the input variables, and then the effects of these values on the surrounding state variables are calculated. When a specific state variable among the surrounding state variables is observed, an observation error occurs between the state variable calculated by elastoplastic analysis and the observed data. Therefore, when a small change is made to the value of the unknown variable, the state variable changes under its influence, and as time passes, the surrounding state variables also change, and the amount of change in the observation error is obtained. In the sensitivity calculation section, the physical properties are used as input variables, and the amount of change in the state variable when the variable changes slightly is stored in the number of variables. Then, as shown in Figure 2, the number of variables is propagated according to the calculation graph of the numerical analysis to obtain how the changes in the state variables, such as the physical properties, affect the surroundings as they evolve over time. As a result, the change in the observation error can be calculated from the change in the state variables. The sensitivity of the unknown variable to the state variable or the observation error is approximately obtained by dividing the fluctuation of the state variable or the observation error by the fluctuation of the unknown variable. The concept and definition of the number of variables are shown in Figures 1 to 2. Also, an example of a calculation graph to explain the propagation of the number of variables is summarized in Figure 3. In particular, in a vibration problem, as shown in Figure 4, points other than the vibration propagation path or points where the vibration is attenuated and not transmitted are hardly affected by the change in physical properties, so there is no need to track the time evolution of the change amount, and the amount of calculation can be reduced. In an elastic-plastic analysis of a vibration problem, the number of variables is propagated within the element according to the calculation graph shown in Figure 5, and the fluctuation amount of each state variable is obtained. The method of calculating the sensitivity of the state variable using the number of variables when performing an elastic-plastic analysis of a vibration problem is summarized in Figures 76 to 83. This method of the number of variables is an invention in the sensitivity calculation part. The calculation graph created based on the block space model of the elastic-plastic finite element method shown in Figure 8 is summarized in Figure 9. Based on this calculation graph, the propagation range of the number of variables is determined as shown in Figure 10. When propagating the number of variables along a specific path between the fluctuation point and the observation point, the spatial distribution of the criterion for omitting the calculation of the number of variables is determined using the gravitational potential shown in FIG.
[0019] <Calculations in the optimal solution search section> In the optimum solution search section, the gradient of the objective function based on the tangent linear model shown in Equation 48 is calculated by Equation 49 to Equation 52 using the sensitivity of the unknown variable to the observation error calculated by the sensitivity calculation section, and the unknown variable is updated by the gradient method to search for unknown variables that fit the observed data. The first feature of the present invention is a calculation method in which the scale of the unknown variables is aligned as shown in FIG. 6 by Equation 49 when updating the unknown variables by the gradient method, and the magnitude of the update for each unknown variable is adjusted. The second feature of the present invention is to realize a stable search for unknown variables by taking the mean square of the current and past objective function gradients by Equation 64 and Equation 65 when updating the unknown variables. The third feature of the present invention is to realize a stable search for unknown variables by controlling the spatial variation of the unknown variables by applying a constraint condition to the objective function that smooths the spatially adjacent unknown variables shown in Equation 56 to Equation 62. The present invention has these features in the optimum solution search section.
[0020] Hereinafter, the concepts and definitions of variables will be summarized, followed by a case in which elastic-plastic analysis of a vibration problem is performed to estimate unknown variables such as physical properties that fit observed data, and an embodiment of the present invention for this case will be summarized.
[0021] As shown in the embodiment, a specific calculation process when the calculation method of the sensitivity calculation unit and the optimum solution search unit according to the present invention is applied to the elastic-plastic analysis of a vibration problem is shown in FIG.
[0022] <Concept and definition of variables> The number of variables is a number expressed by the symbol δ to indicate how much the value of a function with ω as a variable changes when the value of an unknown variable changes.
[0023]
number
[0024] Define a vector ω that combines np unknown variables. The pth component of this unknown variable is δω pWe want to know the amount of change in function f with ω as a variable when the p-th component of ω changes by δ p Write f and calculate as follows:
[0025]
number
[0026] In this way, the basic operation of variables is to find the amount of change in the unknown variable by substituting the unknown variable before and after the change into the function you want to find the amount of change in, and calculating the difference. As special cases, when performing addition, subtraction, or multiplication, the variable numbers can be calculated independently. Below are patterns in which the variable numbers can be calculated independently.
[0027] <Pattern 1: Addition> The function f is expressed as follows: p Consider the case where after multiplying by , we add a constant bp and then add each of these.
[0028]
number
[0029] The pth component of the unknown variable is δω p When the value of ω changes, the amount of change in function f with ω as a variable is
[0030]
number
[0031] For example, when the function f is defined as follows, the variation of the function f is
[0032]
number
[0033] This can be calculated by multiplying the coefficients of the unknown variables directly by the variation of each unknown variable.
[0034] <Pattern 2: Multiplication> The function f is expressed as follows: p Let us consider the case where each term is multiplied after multiplying by .
[0035]
number
[0036] The pth component of the unknown variable is δω p When the value of ω changes, the amount of change in function f with ω as a variable is
[0037]
number
[0038] For example, when a function f is defined in the form of a product with different unknowns, as in the following equation, the amount of change in function f can be found by calculating the product with other terms.
[0039]
number
[0040] In this way, in the case of addition / subtraction format or multiplication format, calculations can be performed independently using only the variable components.
[0041] <An example of tracing computational graphs using variable numbers> We will look at how to use the number of variables using a function f, which has a structure similar to the computational graph in the finite element method. The greatest feature of the number of variables is that in a complex computation process, the effect of the fluctuation of an unknown variable on each variable can be traced in a small number of computation steps. As an example, the ne-dimensional function f at time t t e is the function at time t-1, and two components ft-1 e-1~e Consider a situation that evolves over time under the influence of
[0042]
number
[0043] Here, e=1, ..., np. The initial condition at time t=1 is
[0044]
number
[0045] In this situation, the unknown variable ω p We want to investigate the effect of the fluctuation of on the function f at each time. The computation graph of this function is shown in Figure 3, where the fluctuation amount branches into two branches and propagates with time. In this function, δω p The amount of variation δ given to the k-th component of the function at each time when the variation of p f t e Let's check f t e The vector volume of unknown variables that directly affect t fe It is defined as:
[0046]
number
[0047] Next, the unknown variable vector θ is assigned the pth unknown variable ω p The amount of change due to the change in δ p Gives.
[0048]
number
[0049] Unknown variable ω p The function f due to the variation ofe The amount of change in is
[0050]
number
[0051] However, δ p θ t fe When the set of is close to 0, δ p f t e It is certain that δ is 0. In other words, if the absolute value of the maximum value of the variable component included in θ is smaller than the infinitesimal amount ε, then δ p f t e becomes 0, the calculation can be reduced by the following formula.
[0052]
number
[0053] Ψ(·) is an operator that extracts the maximum absolute value from the vector or matrix in the parentheses, normalizes it, and returns the maximum value among them. e Let us calculate the effect of e=p using the function of Equation 9. First, let us look at time t=2.
[0054]
number
[0055] When e=p+1,
[0056]
number
[0057] For other e,
[0058]
number
[0059] If we express this range of influence with the Ψ(·) operator, we get
[0060]
number
[0061] As time progresses to t=3, 4, …, at time t, the range of the fluctuation input to the function f expands to e≧pt, so
[0062]
number
[0063] If we calculate the recurrence formula shown in Equation 9 only for components where Ψ() is a small amount or more,
[0064]
number
[0065] The function f can be updated over time. In this way, by using the number of variables, calculations can be omitted when the change in the intermediate variable θ is small, reducing the computational cost. How to use the number of variables in data assimilation
[0066] We will explain how to use variables when estimating physical properties through data assimilation by taking a simple problem as an example. t p is obtained, and ω p A function f with variables t (ω) is used to estimate the observed quantity, and the difference is
[0067]
number
[0068] and the unknown variable ω to be minimized p Let us consider a case where we want to search for the following. The calculation procedure is to first use the number of variables to calculate the fluctuation of the observation estimate due to the fluctuation of the unknown variable as δ p f is found, and the following minimization problem is set up with a penalty term P(δω̂p) imposed on the prior information on the unknown variables.
[0069]
number
[0070] By updating the solution several times using the gradient descent method, we obtain the update amount δω^p of the unknown variables that minimizes the observation error,
[0071]
number
[0072] Using the updated unknown variables, we again calculate the function f and its variation δ p Find f and solve the minimization problem. By repeating this series of calculations, the unknown variable ωp that minimizes the observation error is found. If the propagation range of the variable number by the Ψ(·) operator calculated in the first update is stored and used in the second and subsequent updates, the amount of calculations can be reduced.
[0073] <Data assimilation for dynamic oscillation problems using number of variables> Consider a scenario in which a shock wave is applied to an object as shown in Figure 4, and the shock waveform is observed at multiple points to estimate the physical property information for each point, such as the hardness and density of the object. At this time, if we can determine how a slight change in the physical property at each point affects the observed data, and obtain the sensitivity by dividing the amount of change in the observed data by the amount of change in the physical property, we can determine the physical property that minimizes the observation error using a gradient method or other method. In the number of variables, the variables to be estimated (hardness, density, etc.) are slightly changed at each point, and how state variables such as stress, strain, and the progress of plasticization of the object have changed due to the change in the variable is determined, and the information on the change is boxed for each variable with the δ symbol. After that, as shown in Figure 2, the effect of the change in the variable at each point on the state variable is calculated, and the boxed information is propagated to predict how the surrounding state variables will be affected over time. As a result, we can see to what extent the predicted value of the observation amount is affected by the fluctuation of the unknown variable. Using the number of variables, we can track the time evolution of the information on the amount of change independently for each variable. Therefore, for example, it is not necessary to track the time evolution of the change amount at points other than the propagation path of the shock wave or at points where the shock wave is attenuated and does not propagate. By utilizing the spatiotemporal locality of the wave propagation effect in this way, it is possible to track the effect of the change with the minimum necessary amount of calculation.
[0074] <Propagation of variables in vibration problems> The calculation procedure for data assimilation using the number of variables in vibration problems is shown. The equation of motion for a dynamic elastic-plastic body at time t is given by the mass tensor M j , the damping tensor C ji (t), stiffness tensor K ji Using (t), it can be expressed as follows:
[0075]
number
[0076] e is the element number, i and j are the node numbers. In addition, nt is the number of analysis times, and ni is the number of free contact points. i (t), F j(t) is the external force nodal load and the free nodal displacement, and the superscript · means the increment per step. j (t), v i (t) are the velocity and acceleration of the free node, respectively, and · means increment. Also, please note that only the diagonal components of the mass tensor have values. Since an elastic-plastic problem is assumed in this example, the nodal displacements are expressed in incremental form. From this point on, the superscript · will be used as a symbol to represent the increment per step. The increments in velocity and acceleration can be expressed using central difference approximation as follows:
[0077]
number
[0078] If we use superscripts to represent time t as t step, t+Δt as t+1 step, and t-Δt as t-1 step, and approximate the equation of motion using central difference, we get
[0079]
number
[0080] In the Rayleigh-type damping model, the damping term Cji is calculated by multiplying the mass and stiffness by a coefficient and adding them together, as shown in the following equation.
[0081]
number
[0082] By replacing the damping term in the equation of motion with the mass matrix and stiffness matrix,
[0083]
number
[0084] The time update equation for the nodal displacement is
[0085]
number
[0086] Here, we define symbols that summarize the main matrix and vector components as follows:
[0087]
number
[0088] In order to investigate the sensitivity of unknown variables to the free node displacement of the structure, the relationship between the number of variables from the unknown variables to the node displacement is obtained. To do this, the main variables of the time update equation that are affected by the variation are expressed as functions to visualize the input-output relationship.
[0089]
number
[0090] e j is a unit vector whose j-th component is 1 and the other components are 0. Let p be the number of unknown variables, and ω be the unknown variable to be identified element by element. p and the unknown variables of the elements are the variables δω p is given, the fluctuation of the time update equation for the nodal displacement is
[0091]
number
[0092] U jis a nodal displacement vector that collects only the displacements of the nodes that make up the element that is in contact with node j. Because a large number of input variables are required to calculate the amount of variation in variables R and u, a variable θ is defined, a set of input variables is substituted, and then a variation is given collectively with δ. Ψ(·) is an operator that extracts the component with the maximum absolute value from the vector or matrix in the parentheses, divides it back by the real value of the original vector or matrix, and returns the standardized value. When there are multiple vectors or matrices in the parentheses, as in this case, the maximum values of each are extracted and standardized as shown in the following equation, and then the maximum component among them is returned.
[0093]
number
[0094] In this way, we can see that the number of variables is propagated to the nodal displacement only when there is a change in the nodal displacement from the past two steps to the present. All that remains is to find the amount of variation in the global stiffness matrix. To do this, we will examine the effect that each variable receives from a small change in the unknown variable while sorting out the relationships between the variables. First, in order to explicitly express the superposition of elements, we will create the coupling matrix L e and the element nodal displacement vector u e t and the nodal displacement vector U t Associate.
[0095]
number
[0096] connection matrix L e is a matrix in which the i-th component of the j-th row is 1 when the j-th component of element e corresponds to the i-th component of the global nodal displacement vector, and the other j-th row components are 0. In the finite element method, the elastic-plastic matrix D e t to B Matrix B e The element stiffness matrix is calculated by multiplying by and integrating the elements as follows:
[0097]
number
[0098] The overall composite matrix K is obtained by superposing this element stiffness matrix with the coupling matrix.
[0099]
number
[0100] From the above, the amount of change in the global stiffness matrix caused by a small change in the elastoplastic matrix is given by
[0101]
number
[0102] From the above equation, the amount of change in the global stiffness matrix can be calculated from the amount of change in the elastoplastic matrix. Next, we investigate the amount of change in the elastoplastic matrix due to small changes in the unknown variables. The strain increment and the nodal displacement increment are related by the following equation using the B matrix.
[0103]
number
[0104] D t e is the elastoplastic matrix for element e at time t, which relates the stress increment and strain increment by
[0105]
number
[0106] The stress σ at time t-1 is the sum of the stress increments from time 1 to t-1.
[0107]
number
[0108] Each stress increment is affected by the variation of the elastic-plastic matrix. The calculation graph within the element in the elastic-plastic analysis is shown in Figure 5. As shown in this calculation graph, the stress at time t-1 changes depending on the elastic-plastic matrix and the stress state before that, so the unknown variable ω p The sensitivity of to the stress increment is expressed as the differential linkage rate:
[0109]
number
[0110] μ is a variable that represents the progress of damage and plasticity in the elastic-plastic matrix. In models where plasticity progresses due to deviatoric stress, such as in metal materials, the variable μ can be defined as the equivalent plastic strain. In models where damage progresses due to the development of microcracks, such as in concrete and soil, the variable μ is a state variable that represents the progress of damage. Equation 41 is the total differential dD of the elastic-plastic matrix. t e / dω e It can be expressed as a recurrence formula with variables.
[0111]
number
[0112] The coefficients in the brackets on the first term of the right-hand side of Equation 42 are summarized as follows:
[0113]
number
[0114] This recurrence formula can be solved by calculating the differential chain rate in the following order:
[0115]
number
[0116] The method for determining the sensitivity of the state variable μ to unknown variables varies depending on the elastic-plastic model used. For example, in a thermodynamic damage progression model, the increment of the damage variable μ is determined by past stress and strain values. In general, the elastic-plastic matrix is updated very frequently. Therefore, if the update of the tangent gradient using the above differential chain rate is repeated many times using a tangent linear model such as a double number, gradient vanishing or gradient explosion, which is also seen in artificial intelligence, may occur when the elastic-plastic matrix changes suddenly in the process of time evolution. Therefore, the unknown variable ω p The vector quantity δ, which is the element-by-element value of the fluctuation amount of p ω e Define and solve the recurrence formula using the number of variables,
[0117]
number
[0118] ε is a very small quantity, and the fluctuations in the elastoplastic matrix and state variables are calculated only when the fluctuation in the value stored in the intermediate variable θ exceeds ε; otherwise, the calculation is omitted, thereby reducing the calculation load. The number of variables in the elastoplastic matrix in the third line is updated when there is a fluctuation in the input variables to the elastoplastic matrix or when the incremental strain fluctuation is greater than the reference value. In this way, by calculating the fluctuations in the main variables of the differential chain rate using the number of variables δ, it is possible to easily calculate the sensitivity of the unknown variables to the elastoplastic matrix and state variables such as stress and strain.
[0119] <Search method for optimal unknown variables> We summarize the method of searching for unknown variables that are consistent with the observed results. The time change in displacement u at ns points among the free nodes of the structure is s t Consider a situation where all free nodes are U t and the free nodes where the displacement is observed are expressed as the observation operator H t To ensure consistency between the observed and analytical values, the observation error vector r t is defined as follows:
[0120]
number
[0121] Observation operator H t is a matrix in which the components corresponding to the displacement observation nodes are 1 at the time when the displacement is observed, and the other components are 0. Yt is an np-dimensional vector that lists the observed displacement values at the time when the displacement is observed, and all components are 0 at the time when the displacement is not observed. The sensitivity of the unknown variables to the observation error can be approximately calculated using the number of variables with the following formula.
[0122]
number
[0123] Using this sensitivity, the optimal unknown variables that minimize the observation error are found by performing a two-stage update calculation, an outer loop and an inner loop. First, the inner loop performs a tangent linear approximation of the objective function space, and finds the update amount of the unknown variables when the observation error approaches the minimum value when the unknown variables are updated in the sensitivity direction. However, since the update amount of the unknown variables may become too large if a solution search is simply performed in the tangent linear approximation objective function space, prior information on the unknown variables is given as a constraint. Next, the unknown variables are actually updated in the outer loop, and the finite element method is then performed to evaluate how much the observation error has actually been reduced. This inner loop is repeated several times, and then the outer loop is executed, and the calculation is repeated to search for the optimal solution.
[0124] In order to search for the optimal unknown variable update amount in the inner loop, the sensitivity to the observation error is approximated by a tangent linear approximation using the variation of the observation error relative to the variation of the unknown variable. The optimal unknown variable variation amount δω^ that can minimize the observation error vector is calculated using the gradient of this tangent linear approximation. p The goal is to find the optimal amount of change in the unknown variables. The amount of change in the observation error vector corresponding to the optimal amount of change in the unknown variables is called δr^ t Then, in the inner loop, the following minimization problem is solved:
[0125]
number
[0126] LS(δω^ p ) is the constraint imposed on the unknown variables, and ρ is a diagonal matrix that lists the weights for the observation errors. In this case, the sensitivity obtained by differentiating the objective function LL with respect to the variation of the optimal unknown variables is
[0127]
number
[0128] When searching for an optimal solution using the gradient method, if the unknown variables to be updated have different magnitudes, only the unknown variables that are highly sensitive to the objective function will be significantly updated. To prevent this phenomenon, the expected maximum value ω pmax and the minimum value ω pmin is defined, and the unknown variables are normalized to have the same size by the following equation.
[0129]
number
[0130] As shown in Figure 6, the scale of the unknown variable and the sensitivity to the objective function are inversely proportional to each other. For example, ▲s f If you reduce the scale of the unknown variables by halving or reducing the size of ▼ by a factor of three, the sensitivity to the objective function will increase by a factor of two or three. f The larger the value of , the smaller the sensitivity to the objective function, and therefore the smaller the update amount of the entire variable by the gradient method. f If the value of is reduced, the entire variable is updated significantly. The number of variables before and after standardization satisfies the following relationship,
[0131]
number
[0132] The sensitivity of the objective function to the unknown variables after standardization is calculated from the differential chain rate using the following equation:
[0133]
number
[0134] In order to narrow the search range for a solution when updating the solution using the tangent linear model given by equation (52), we will summarize a method for setting constraint conditions using information known in advance about the unknown variables.The mean sum of squares of the observation error is defined by the following equation.
[0135]
number
[0136] When the mean sum of squares of the observation error is large, the influence of the constraints is strengthened to prioritize prior information, and when the mean sum of squares approaches the optimal solution and becomes small, the constraints are weakened and the sensitivity to the observation error estimated using the number of variables is prioritized, thereby performing an efficient solution search.To achieve this, the total number of updates τ of the inner loop is used as a variable, and the weight ρ for the constraints is defined as the ratio between the first update value of the mean sum of squares of the observation error and the current value, using the following equation.
[0137]
number
[0138] The constraint term LS imposes a constraint so that the values of the unknown variables of adjacent elements are smoothed during the update process of the unknown variables. In the finite element method, the unknown variable numbers that exist in positions adjacent to the unknown variable p are defined as a subscript set P p The index set P p The average value of each type of unknown variable is calculated using the formula below, and the difference between the value of the unknown variable of element e is minimized to smooth the unknown variables between adjacent elements. Here, the p-th component of the unknown variable ω or z is included in the set P bThe np-dimensional defining function, where the b-th component is 1 when the element belongs to the np-dimensional ...
[0139]
number
[0140] These definition functions are hereafter referred to as Pb The number of indices in a set is expressed as From TIFF2024163826000057.tif10169, the updated unknown variables are predicted in each inner loop.
[0141]
number
[0142] Using this unknown variable vector, set P p The average value of the unknown variables belonging to, i.e., the average value of the unknown variables adjacent to unknown variable p, can be calculated using the following formula.
[0143]
number
[0144] The deviation of the unknown variable from the mean value is given by the following formula:
[0145]
number
[0146] The sample variance of the unknown variable in the neighboring elements around the unknown variable p is
[0147]
number
[0148] Using these variables, we propose a method for setting constraints for two cases: when the statistical variation between adjacent elements of the unknown variables is known and when it is unknown. When the variation of the physical property that is the unknown variable is unknown, the following formula is used to set P around the average value obtained in each update process of the unknown variable. p It is possible to minimize the sample variance for each
[0149]
number
[0150] This makes it possible to smooth the values of the unknown variables between adjacent elements. By differentiating this constraint term with respect to the unknown variable vector, the sensitivity of the unknown variable update amount δZ^(τ) to the constraint term LS is obtained.
[0151]
number
[0152] Information about the spatial distribution of the variation of the unknown property, i.e., the true value of the sample variance▲s e ▼If it can be predicted in advance, the deviation between the predicted value of the sample variance and the true value is measured using an index that extends the Itakura-Saito pseudo distance (IS-Divergence) so that it does not diverge even if the variance estimate becomes 0 with a small value of ε, and the sum is calculated.
[0153]
number
[0154] By minimizing as a constraint term, the search range for a solution is narrowed down. By using this constraint term, when the observation error is large and the solution is not approaching the optimal solution, the sample variance can be brought closer to a value predicted in advance, and the convergence to the optimal solution can be accelerated. The sensitivity to the constraint term LS can be calculated using the following formula.
[0155]
number
[0156] τ is the total number of updates of the inner loop in each outer loop, and by adding it after the variable in parentheses, the number of updates of δz^ is specified. p In order to update, we define the gradient vector g(τ) that collects the sensitivity of the objective function at the τth variable update as g(τ) (64). Using equation 65, we define the unknown variable δz^ so that the variables can be updated without causing gradient divergence even if the gradient changes suddenly. p The update amount of is adjusted by the mean square of the current gradient and past gradients G(τ), and the optimal solution is searched for using the following equation.
[0157]
number
[0158]
number
[0159]
number
[0160] γ is set to about 0.9. η(τ) is the learning rate for each time step, and is set to about 1 / 10 of the unknown variable fluctuation range using the mean square sum of the observation error, so that the update width becomes smaller when the observation error becomes smaller.
[0161]
number
[0162] The ratio of the mean square sum of the observation error before the update (τ = 0) to the current value is adjusted. After determining the optimal update amount by repeating the update in the inner loop several times, the unknown variable z p is updated, the unknown variables before normalization are updated with the updated results, and the finite element method is executed again.
[0163]
number
[0164] Solution update direction δω^ p and the variation of the unknown variable δω p The error in numerical differentiation is smaller when these values match. Therefore, efficient solution search can be achieved by aligning the signs of the increase / decrease in the unknown variables and the fluctuations in the unknown variables when updating the outer loop using the following formula:
[0165]
number
[0166] In the elastic region, the elastoplastic matrix and stiffness matrix do not change with time. Therefore, the propagation of variables at the nodes can be expressed in a form in which the time symbol is removed from the stiffness matrix in Equation 32.
[0167]
number
[0168] In particular, the propagation of variables from other points can be simply expressed as in equation 70. Within an element, the variables are updated using the following equation. Specifically, the amount of change in the damage amount is carried over from the previous time, and the amount of change in the elastic matrix needs to be calculated only for cases where the unknown variables are within the element.
[0169]
number
[0170] On the other hand, in a region that does not become plastic within the analysis time and always remains in the elastic region, the variables do not need to be propagated within the element, so the calculation of Equation (71) can be omitted entirely.
[0171] In this way, in the elastic region, only the nodal displacement changes due to the propagation of variables from other points. Therefore, if the amount of variation in the nodal displacement increment due to unknown variables at other points is 0, there is no influence from the amount of variation at other points. Also, in an elastic state, the number of variables propagated within an element no longer affects other points. Therefore, in an elastic state, the calculation of the propagation of variables within an element in equation (71) can be omitted. From the above, the propagation of variables from other points in an elastic region can be determined only by the presence or absence of the amount of variation δp in the nodal displacement increment.
[0172] <Reduction of computational load by propagation of definition functions> We have used the Ψ(·) symbol to determine whether or not to calculate the number of variables. However, in actual finite element analysis, using the Ψ symbol to determine whether to update each time is not necessarily efficient. Therefore, we summarize a method to automatically determine whether or not to update state variables by determining at key points whether the number of variables and the size of state variables propagated along the branches of each computation graph of the finite element method meet the update criteria for the number of variables and propagating the judgment result itself along the computation graph. Using this method, it is possible to automatically omit the calculation of the number of variables that has little effect on future computation results at each time. The determination of whether or not to calculate the number of variables is made using a definition function that returns a value of 1 or more when it is necessary to calculate the number of variables because the number of variables is below a certain value, etc., and returns 0 when no calculation is necessary.
[0173]
number
[0174]
number
[0175] Initial values are set for the definition functions corresponding to the element quantities and nodal quantities, and the definition functions themselves are evolved over time along with the state variables. Whether or not to calculate the number of variables of the element quantities shown in Equation 45 is determined by the value of the definition function of the strain increment and elastoplastic matrix shown in the following equation.
[0176]
number
[0177] The definition functions in the third column return 1 if the number of variables corresponding to the variables required to calculate the elastoplastic matrix is equal to or greater than a certain value, 2 if no element strain was observed in the previous time step and there was no change in the elastoplastic matrix, and 0 otherwise. Other definition functions return 0 if the variable or number of variables in parentheses is equal to or less than a certain value. If these definition functions are 1 or greater, calculation of the number of variables is performed, and if they are 0, the calculation of the number of variables is omitted.
[0178] When updating the nodal displacement time shown in Equations 31 and 32, the determination as to whether or not to perform a variable number calculation is made using the definition function of the following equation.
[0179]
number
[0180] This definition function returns a value of 1 or more if the displacement increment of node j at time t exceeds the reference value and the number of variables needs to be calculated, and if it is below the reference value, it returns 0. These definition functions are mainly used to determine whether to calculate the number of variables.
[0181] First, we summarize the method of using the definition function to determine whether or not to omit updating the number of variables in an element when evolving the element amount over time. First, we set the initial value of the definition function. The unknown variable ω p The vector quantity δ, which is the element-by-element value of the fluctuation amount of p ω e Using the above, the initial value of the definition function for determining whether the elastic-plastic matrix is updated is determined by the following equation.
[0182]
number
[0183] Also, a definition function for determining whether the number of variations in the nodal displacement increment exceeds a reference value is set to 0 (less than or equal to the reference value).
[0184]
number
[0185] The update (time evolution) of the element amount shown by M in Fig. 8 is performed in the order of the number of variables → indicator function. First, the update of the number of variables shown in Equation 45 is performed by the following equation.
[0186]
number
[0187] When calculating the number of variables for the elastoplastic matrix in the second column and the damage variable in the sixth column, if the definition function is 2 because the elements have not deformed, the number of variables from the previous time step is inherited as is. When the indication value is 1, the elastoplastic matrix and the number of variables for the damage variable are updated using the state variables from the previous time, but only at times when the indication function changes value from 0 to 1 or 2 to 1, repeated calculations using the return mapping method are performed several times to update the calculations.
[0188] Next, the definition function of the element amount, which determines whether or not to perform a time update of the number of variables, is updated by the following equation.
[0189]
number
[0190] The definition functions for strain increments (number of variables) in columns 3 and 4 are updated by superimposing the definition functions for nodal displacement increments (number of variables) for each element. The definition function for the elastoplastic matrix shown in column 6 calculates the variation in the elastoplastic matrix as 1 if the number of variables or strain increment corresponding to the variables required for calculating the elastoplastic matrix is above a certain value, and when no element strain is observed in the previous time step, there is no change in the elastoplastic matrix, so it is set to 2 and the variation in the elastoplastic matrix from the previous time is inherited. Otherwise, it returns 0. The other definition functions return 0 if the variable or number of variables in parentheses is below a certain value. If these definition functions are 0, the calculation of the number of variables is omitted.
[0191] The state variables that have evolved over time within an element are superimposed on adjacent elements in the process of being converted to nodal quantities. During this superimposition process, the number of variables diffuses in the spatial direction, as shown in the information transmission process in Figure 8. The propagation process of the number of variables and the definition function during this superimposition of element quantities is summarized below. The number of variables of the elastoplastic matrix calculated as the element quantity is superimposed for each node using the following equation, and the definition function is also superimposed using a similar information transmission path.
[0192]
number
[0193] L e j is a matrix in which the component of the global nodal displacement vector number corresponding to the j-th component of element e is 1 and the other components are 0, and by taking the transpose, the element quantities can be converted into nodal quantities. The definition function of the global stiffness matrix in the third line has a value of 1 or more when node j and node i are adjacent to the same element, and is 0 otherwise.
[0194] The definition function returns 0 if the value in the parentheses is less than a certain value, so it is easy to decide whether to omit the calculation of variables required in the variable number calculation. The definition function is used to improve the efficiency of the calculation of the nodal displacement update formula using the explicit finite difference method. The update formula for the nodal displacement increment using the explicit finite difference method can be expressed as follows.
[0195]
number
[0196] When updating the number of variables with this nodal displacement update formula, if the definition function is updated at the same time, the calculation of the number of variables can be automatically omitted. The nodal quantity is evolved over time in the order of definition function → number of variables → definition function. First, the time update of the definition function and the number of variables shown in Equations 31 and 32 is performed using the following equations.
[0197]
number
[0198] The definition functions in the first and second lines of equation 82 are judged to be 0 if the nodal displacement increment and its number of variables have no values in any of the past two time steps. The definition function in the third line is judged to be 0 if both the global stiffness matrix and the nodal displacement increment have no values. These three types of definition functions are used to determine whether to omit part of the time update calculation of the nodal displacement increment shown in line 5. Using the update result of the nodal displacement increment and the number of variables, the definition function is updated by the following equation, and the time step of the finite element method is advanced.
[0199]
number
[0200] This definition function returns 0 if the fluctuation of the nodal displacement is below a certain value, and returns 1 if it is above the certain value.
[0201] <Transmission process and locality of spatiotemporal information> The flow of changes in state variables in the elastic-plastic finite element method based on the differential chain rate derived so far is expressed as information transmission in the block universe space-time. To this end, the set of stress and damage progression state in the plastic region is defined as follows.
[0202]
number
[0203] In the elastic region, damage does not progress, so the damage state from the previous time is inherited as it is. If we define the variables described in the incremental form of the finite element method as incremental amounts (superscript dots) and the rest as cumulative amounts, and describe the information transmission process of the block universe space-time in a discrete form, we get
[0204]
number
[0205] Function set F Y (·) is a function set that collects the functions required to calculate the subscript set Y, so that the values of the subscript set Y can be calculated all at once by inputting the necessary variables in (). C e is a set of time-independent constant matrices.
[0206] In the first column of equation 85, the current nodal displacement increment is calculated from the nodal displacement up to two steps ago and the state quantity M from one step ago, and in the second column, the element nodal displacement increment u e t is diffused, and the element nodal displacements are updated in the third column.
[0207] When propagating the number of variables, the update formulas in the first and fourth columns of Equation 85 correspond to Equation 82 and Equation 78, respectively. On the other hand, when propagating the definition function, the update formulas in the first and fourth columns of Equation 85 correspond to Equation 83 and Equation 79, respectively.
[0208] In the update equation in the fourth column, the time update method of the state quantity M changes depending on whether the element in the space is in a plastic or elastic state. In a plastic state, the increment of the current state quantity M from the diffused element node displacement increment and the previous state quantity M (accumulated quantity) is obtained by solving a nonlinear relational equation. In this way, the state is updated by performing a nonlinear transformation on the state quantity M, which is the accumulated value of the past, so nonlinear materials such as elastoplastic bodies exhibit chaotic properties. On the other hand, in an elastic state, the current state is determined by multiplying the previous element node displacement increment by a constant matrix consisting of an elastic matrix D and an identity matrix I. As a result, in an elastic state, the stress increment is determined by a linear transformation of the displacement increment, and the past damage state is inherited as it is. In order to visualize these phenomena in space-time, we consider a block universe model in which the space at each time t=1, ..., nt is divided into elements e=1, ..., ne in the same way in the time direction as in Figure 8, and a calculation graph simplified based on this model using a set of variables and functions is shown in Figure 9. In the world we live in, space itself evolves continuously over time, and the way that time evolves is governed by incremental quantities. However, we can only exist at a certain moment when space-time is sliced in the time direction, so we can only perceive cumulative quantities. As a result, we cannot fully perceive the causal relationships in the phenomena occurring before our eyes, and perceive them as chaotic phenomena. If we express the number 85 in variable numbers, we can see that it propagates along the same path as the original information path, as shown in the following equation.
[0209]
number
[0210] Information in the time direction that is always propagated to elements that remain in the elastic region does not affect future times. Therefore, variables from other points are not propagated to elements that always remain in the elastic region, and the fluctuation update formula in lines 4 and 5 of equation 86 can be omitted. In this way, by propagating variables along the information transmission path in the block universe, the sensitivity from unknown variables can be calculated.
[0211] The concept of the propagation range of variables is summarized in Figure 10. At a certain time t after excitation, the region where the physical properties change slightly is defined as A, and the region where the element node displacement changes due to the influence of region A, that is, the region where the variables of region A propagate with respect to the node displacement, is defined as region B. t The area that has become plastic due to vibration is area C. t In addition, the region including the node surrounding the outer element of each region is represented by the superscript "+". Then, the following proposition is derived from number 86.
[0212] Proposition 1: The nodes for which the fluctuation amount of the nodal displacement is updated using Equation 32 are up to the nodes surrounding one outer element of the element affected by the unknown variable in the time steps up to the second previous to the current time step.
[0213]
number
[0214] Proposition 2: The elements for propagating the variable number and updating the variation amount using equation 45 need only be elements that contain the unknown variable to be varied or elements that have been plasticized among the elements affected by the unknown variable.
[0215]
number
[0216] In the region that always remains in the elastic region during the analysis period, only the stiffness matrix of region A where the unknown variables are varied is affected by the number of variables, so in region A, it is sufficient to find only the amount of variation in the stiffness matrix using equation 37. On the other hand, in the common subset of elasto-plastic region C and propagation range B of the number of variables, the element state quantities are affected by the stress and damage state quantities at the previous time, so the amount of variation within the element is updated using equation 45.
[0217] <How to set the omission criteria for variable number calculation> We will summarize how to set the criteria for omitting calculations of variables using the Ψ operator. We will consider the Ψ operator, which determines whether or not to omit calculations of a function whose variable is the amount of variation in the unknown variable θ.
[0218]
number
[0219] In the Ψ operator, the fluctuation amount δ p θ is the assumed value δ p When θ^ is sufficiently smaller than the number of variables, the calculation of the number of variables is omitted. Therefore, the expected value δ of the state variable fluctuation p The decision to skip calculations is determined by θ^. In a nonlinear dynamical system, a small change in the unknown variable causes a large change in the observation point. However, in a vibration problem, the unknown variable ω p The path where the variation of the unknown variable is amplified shows low dimensionality, which leads to low dimensionality of the variable propagation. If the variation of the unknown variable is amplified as its influence value propagates to the observation point, then the path C p Identify the expected value δ only on the route. p If the value of θ̂ can be reduced, it will be possible to give priority to calculating the number of variables near the path and omit the calculation of the number of variables only at points away from the path.
[0220] The potential of a gravitational field decreases locally around an object with mass, and converges to a constant value when moving away from the object. By applying this property, the assumed value of an unknown variable is calculated as the gravitational potential, and the value is reduced only in a specific area. g mass μ(x g ) mass point exists,
[0221]
number
[0222] The gravitational potential at point x is calculated by mass μ(x g ), the larger the gravitational potential is, the wider the range in which it is reduced. In consideration of this property, the mass is set small in areas where the path Cp through which the number of variables is amplified is uniquely determined, and the mass is set large in areas where the path is difficult to identify.
[0223] Mass distribution is along path C pWhen the gravitational potential at point x exists continuously,
[0224]
number
[0225] This gravitational potential is calculated by path integral. p is equivalent to the gravitational potential due to the variable attractor when the mass point is moved along the state variable δ p The assumed value of θ at position x is found by multiplying the assumed value by the value obtained by exponentially mapping the gravitational potential.
[0226]
number
[0227] In this way, by using the gravitational potential when there is a mass distribution between the observation point and the fluctuation point in the distribution of the calculation omission criteria, route C as shown in Figure 11 p A region of low gravitational potential is formed in the vicinity, making it difficult for variables to disappear. Therefore, by assuming mass density in the variable amplification path, it becomes possible to extract only the variables that have a large effect on the observation point and perform efficient calculations.
[0228] In a nonlinear dynamical system, minute fluctuations in the initial stage of the state variables bring about large changes over time. In order to keep this chaotic phenomenon and the assumed value in balance, the magnitude of the assumed value also increases over time using a function T(t), and the following formula is used to determine whether or not to omit calculations by adding a time axis to the assumed value.
[0229]
number
[0230] The cumulative error function shown in the following formula can be used as a time function that increases with the passage of time t. When using the cumulative error function, the initial value is sufficiently small, so a small amount can be set to ε=1 to determine whether to omit calculations.
[0231]
number
[0232] Using the assumed values of the variable θ shown in Equation 93 that vary with position x and time t=1,...,nt, the Ψ operator is expressed as a four-dimensional space-time distribution function
[0233]
number
[0234] This makes it possible to set the calculation omission criteria separately for each space-time direction.
[0235] <Data assimilation method for multiple variables in linear elastic bodies> When estimating physical properties from vibration velocity or acceleration data of a linear elastic body, data assimilation can be performed quickly using the adjoint equation shown below based on the Adjoint method.
[0236]
number
[0237] LS is a constraint condition for unknown variables. This Lagrange equation is solved using the Adjoint method. The adjoint equation obtained by differentiating the Lagrange multiplier with respect to the nodal displacement is
[0238]
number
[0239] Terminal condition (λ nt+1=0) and solving in reverse time direction, the Lagrange multipliers are found. p The gradient with respect to is
[0240]
number
[0241] By using this gradient to update the unknown variables, it is possible to find unknown variables that fit the observed data.
[0242] In this way, in a spatial domain that always maintains an elastic state, the time evolution of the variable set M changes depending only on the nodal displacement and element strain, so the time evolution of the number of variables in the variable set M can be ignored. In an elastic state, the number of variables flowing through the element is the change in the unknown variable δω p and the change in element strain δ p ε e Therefore, the sensitivity can be easily obtained by adding the nodal displacement U to the unknown variables and tracing in the reverse propagation direction using the adjoint equation based on the Adjoint method. On the other hand, in the plastic region, the information emitted in the time direction toward the increment of the variable set M also depends on the variable set M at the previous time, so the number of variables flows in the time direction within the element. In order to trace the variables flowing in the time direction within the element in the reverse time direction using the adjoint equation, it is necessary to differentiate the elastoplastic matrix with respect to the nodal displacement U, which results in inefficient calculations. This makes it difficult to use the adjoint equation for sensitivity analysis in elastoplastic problems.
[0243] In this way, by summarizing the time evolution process of the state variables in the finite element method as an information transmission pathway in the block cosmology model, it is possible to visualize how the state variables relate to each other and evolve over time. [Industrial Applicability]
[0244] Ultrasonic testing is used when ultrasonic waves are propagated through structural components to non-destructively examine the internal soundness. In ultrasonic testing, scattered waves from defects are measured to estimate rough information such as the presence and location of defects. By applying the data assimilation method of the present invention to ultrasonic testing, detailed information such as the size, shape, and material of the defects can be estimated with a small amount of calculation. Even in cases where the internal state of a component is estimated using ultrasonic waves in this way, if it falls under any of the items described in the claims, such as sensitivity calculation using the number of variables or reduction in the amount of calculation using a defined function, it is within the scope of the present invention. [Explanation of symbols]
[0245] 1 Analysis variable setting section 2. Data assimilation calculation section 3 Calculation result display area
Claims
1. Set the analysis variables used in the elastic-plastic analysis. Among the analysis variables, variables to be fitted to the observed data are defined as unknown variables; When the variables obtained from the elastic-plastic analysis, such as stress, strain, nodal displacement, and damage progression, are defined as state variables, Analyzing the observation data by estimating the state quantity at each time by assuming the unknown variables; A data assimilation calculation method for updating the unknown variables with the goal of minimizing the difference between the actual observed data value and the analytically estimated value, i.e., the observation error, comprising: When the sensitivity is defined as the value obtained by dividing the magnitude of the change in another variable due to the change in the unknown variable by the magnitude of the change in the unknown variable, a sensitivity calculation unit that calculates the sensitivity of the unknown variable to the state variable to calculate the sensitivity to the measurement error; and an optimal solution search unit that uses the sensitivities inherited from the sensitivity calculation unit to determine the amount of update of the unknown variables that is suitable for the observation data. Data assimilation algorithm.
2. A method for calculating sensitivities of the unknown variables to the measurement errors, the method being performed by the sensitivity calculation unit, comprising: calculating a magnitude of change in the state variable at an adjacent position due to a change in the unknown variable; Calculating how the amount of variation of the state variable at adjacent positions changes over time; By calculating how the state variables at other points change due to the influence of the fluctuation amount of the unknown variable and the fluctuation amount of the state variables at adjacent positions, calculating the sensitivity of the unknown variable to the measurement error; Sensitivity calculation method.
3. A method for determining an update range of the unknown variables, which is performed by the optimal solution search unit, an np-dimensional gradient vector g(τ) in which the sensitivities to the measurement errors taken over from the sensitivity calculation unit are arranged is defined using the number of updates τ of the unknown variables; The mean square of the current gradient vector and the past gradient vectors G(τ) [0.65] By updating while adjusting the value by the constant γ, The update amount of the unknown variable is adjusted. A method for determining an update range for the unknown variables.
4. A method for unifying the scale of the unknown variables, which is performed by the optimal solution search unit, The unknown variable ω p The upper limit of the value of ω pmax and the lower limit ω pmin Assume that there is a constant s f Using [Number 50] and the size of the unknown variables is made uniform. A method for scaling the unknown variables.
5. A sensitivity condition determination method for determining whether the sensitivity of the unknown variable to the state variable calculated by the sensitivity calculation unit satisfies a certain condition, comprising: A definition function is used that returns a natural number n when the sensitivity value satisfies the n-th condition among N conditions, The definition function is propagated on a computation graph in a numerical analysis such as a finite element method while evolving with the spatiotemporal evolution of the state variables, The condition to which the state variables at all positions at each time step apply is automatically determined from the value of the definition function. The sensitivity condition determination method
6. The unknown variable ω p The observation error r t A method for calculating sensitivity to Mass Mj, damping term C ji t , stiffness K ji t When determining the physical properties as the unknown variables from the vibration data of the linear elastic body, Nodal displacement u by explicit finite difference method t j The time update equation of is used as a constraint term, and the Lagrange multiplier λ t j Lagrangian function multiplied by [Number 96] and the Lagrangian function is multiplied by the Lagrangian multiplier λ t j The adjoint equation differentiated by [Number 97] The end condition (λ nt+1 = 0) and solve in the reverse time direction to obtain the Lagrange multipliers, The unknown variable ω p Sensitivity to [Number 98] and differentiating the adjoint equation with respect to the unknown variables to calculate the sensitivity. A method for calculating the sensitivity of the unknown variables to the measurement errors.
