Dynamical state reconstruction method of missile-guided multi-body system based on model-assisted network

Abstract

The present invention discloses a method for reconstructing the dynamic state of a multi-body system of a missile based on a model-assisted network, which has faster reasoning speed, high accuracy, and high computational efficiency. The method comprises the following steps: (10) model establishment: using a multi-body system dynamics method to establish a model of a multi-body system of a missile; (20) parameter calibration: using a finite element method to perform harmonic response analysis on the multi-body system of the missile to obtain connection relationship parameters; (30) dual extended Kalman filter construction: based on the connection relationship parameters, a dual Kalman filter is constructed based on the least squares method; (40) short trajectory multi-input GRU neural network construction: based on the GRU type neural network structure, a mature short trajectory multi-input GRU neural network is constructed and trained; (50) system dynamic state reconstruction: obtaining a dynamic reconstruction signal of the multi-body system of the missile from the output of the mature short trajectory multi-input GRU neural network.

Description

Technical Field

[0001] The present invention relates to the fields of dynamics and machine learning, and in particular to a method for reconstructing the dynamic state of a projectile-guided multi-body system based on a model-assisted network. Background Art

[0002] A multibody system is a system composed of multiple objects (rigid, elastic, flexible, and point masses) interconnected by specific connections. A typical example is a hard target penetration guidance system. These systems typically include multiple objects, such as the cartridge case, charge, sensor housing, sensor, and base screw. These objects transmit impact energy and overload information through various interconnections. Multibody system dynamics methods allow for both manual and computer-automated derivation of overall transfer equations and matrices, offering high accuracy, robust operability, and low computational cost. Consequently, they have found widespread application in various engineering fields, including launch dynamics.

[0003] Extracting state information from a complex multibody dynamics system can be broadly categorized into direct and indirect methods. In the direct method, sensors can directly measure the system's state parameters. However, due to the system's high complexity, direct sensor deployment for testing is difficult, and this approach is also very costly. Therefore, direct implementation of this approach is often impractical. In the indirect method, unmeasurable state information is derived from the responses of measurable subsystems. Therefore, the core of the problem is how to reconstruct the state information of unmeasurable subsystems using known system characteristics and measurable information.

[0004] A more effective method is to use the dynamic response characteristics of mechanical system modeling for indirect measurement. This is because the dynamic response (primarily acceleration signals) can be easily measured directly by sensors, which also forms an important foundation for indirect methods. However, the nonlinear transmission characteristics of complex mechanical systems have become a major challenge for indirect measurement, especially in the field of penetration weapon systems. Due to the constraints of the internal structure volume of the projectile, the position of the onboard sensor and the warhead often pass through a complex projectile dynamics system. Therefore, the information measured by the sensor is significantly different from the state information of the warhead, which can cause the onboard control system to misjudge the current projectile state, thereby affecting the ultimate damage effect. Therefore, it is necessary to reconstruct the state of the projectile multi-body system.

[0005] Multibody system dynamics state reconstruction refers to the process of re-extracting the state information of a complex multibody dynamics system. In practical engineering applications, prior knowledge of the state information of the different subsystems acting on a multibody dynamics system is crucial for current state determination, fault analysis, and reliability analysis.

[0006] Currently, methods for reconstructing the dynamic state of missile-guided multi-body systems generally replace simple principled models with purely data-driven methods. These methods use large amounts of labeled datasets for training. For example, in the literature (J. Zhou, L. Dong, W. Guan, and J. Yan, “Impact load identification of nonlinear structures using deep Recurrent Neural Network,” Mech. Syst. Signal Process., vol. 133, p. 106292, Nov. 2019, doi: 10.1016 / j.ymssp.2019.106292.), a nonlinear structural impact load identification method based on a deep recurrent neural network (RNN) is used for training through a large number of dynamic responses and impact loads to learn the complex inverse mapping relationship between structural input and output, thereby obtaining the dynamic state information of the missile-guided multi-body system.

[0007] However, these methods still have certain limitations. First, this data-based approach requires a large dataset for training, and may not be effective when data samples are limited. Second, although hardware computing power has greatly improved compared to previous generations, there is still an irreconcilable contradiction between computing power and power consumption in current embedded systems. If large, complex convolutional neural networks (CNNs) and transform-type networks are used, it is not practical for embedded systems with strict real-time requirements.

[0008] In summary, the problems existing in the existing technology are: the reasoning speed of the multi-body system dynamic state reconstruction method is not fast enough, the accuracy is low, and the calculation efficiency is low. Summary of the Invention

[0009] The purpose of the present invention is to provide a method for reconstructing the dynamic state of a multi-body system based on a model-assisted network, which has faster reasoning speed, high accuracy and high computational efficiency.

[0010] The technical solutions for achieving the purpose of the present invention are:

[0011] A method for reconstructing the dynamic state of a multi-body system of missiles and homing missiles based on a model-assisted network comprises the following steps:

[0012] (10) Model establishment: Using the multi-body system dynamics method, the projectile multi-body system model is established;

[0013] (20) Parameter calibration: using the finite element method, performing harmonic response analysis on the projectile-guided multi-body system model to obtain connection relationship parameters, wherein the connection relationship parameters include the equivalent stiffness coefficient and the equivalent damping coefficient of the projectile-guided multi-body system;

[0014] (30) Dual extended Kalman filter construction: According to the connection relationship parameters, a dual Kalman filter based on the least squares method is constructed;

[0015] (40) Construction of short trajectory multi-input GRU neural network: Based on the GRU type neural network structure, a mature short trajectory multi-input GRU neural network is constructed and trained;

[0016] (50) Reconstruction of the system dynamics state: The original sensor data and the output data of the Kalman filter are used as the input of the mature short trajectory multi-input GRU neural network, and the dynamic reconstruction signal of the projectile multi-body system is obtained from its output.

[0017] Compared with the prior art, the present invention has the following significant advantages:

[0018] 1. Faster reasoning speed: The present invention adopts a model-assisted network as a new state reconstruction method for multi-body system dynamics, which can learn the input-output mapping relationship from the initial data and model data, and has a faster reasoning speed.

[0019] 2. High Accuracy: In the model component, this invention uses a dual Least Squares Kalman Filter (DLEKF) for state reconstruction, enabling simultaneous estimation of system state and model parameters. Data correction and state reconstruction are performed under the premise of system parameter calibration, ensuring accurate model estimation results. Furthermore, the dimensionality of the state reconstruction data can be expanded to further characterize information about unmeasurable components.

[0020] 3. High computational efficiency: In the network component, this invention uses a short-trajectory, multi-input GRU (Gated Recurrent Unit) neural network, which can implement a mapping method that uses both sensor data and model-based reconstructed data as input. Although this expands the input dimension, it reduces the input data and improves computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is the main flow chart of the model-assisted network-based method for reconstructing the dynamic state of a multi-body system.

[0022] Figure 2 It is a schematic diagram of the mechanical transmission relationship of the projectile multi-body system.

[0023] Figure 3 yes Figure 1 Flowchart of the steps for building the dual extended Kalman filter.

[0024] Figure 4 yes Figure 1 Flowchart of the steps for building a multi-input GRU neural network for short and medium trajectories.

[0025] Figure 5 It is the input layer structure unit (GRU) graph.

[0026] Figure 6 This is the GRU neural network architecture diagram.

[0027] Figure 7 This is a comparison chart of the variance of the reconstructed state using different methods at a speed of 800 m / s.

[0028] Figure 8 This is a comparison chart of the variance of the reconstructed state using different methods at a speed of 900 m / s. DETAILED DESCRIPTION

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] like Figure 1 As shown, the method for reconstructing the dynamic state of a multi-body system based on a model-assisted network of the present invention includes the following steps:

[0031] (10) Model establishment: The multi-body system dynamics method is used to establish the projectile multi-body system model.

[0032] The model building steps (10) are specifically as follows:

[0033] According to Newton's second law, the response equation of the projectile multi-body system is established to determine the mechanical transmission relationship.

[0034] The response equation of the projectile multi-body system is as follows:

[0035]

[0036] Among them, M A 、M B 、M C and M D Respectively represent the mass of the projectile, the mass of the medicine, the mass of the sensor shell and the mass of the bottom screw; and x A ,x B ,x C and x D , respectively represent the corresponding displacements; and Represent the corresponding speeds respectively; and and represent the corresponding accelerations respectively; K1, K2, K3, K4 and K5 are the corresponding equivalent stiffness coefficients respectively; C1, C2, C3, C4 and C5 are the corresponding equivalent damping coefficients respectively; F A For external overload;

[0037] The mechanical transmission relationship is shown as follows Figure 2 shown.

[0038] The above system response equation for determining the mechanical transmission relationship can also be expressed in the following matrix form:

[0039]

[0040] Where the system state x(t)=[x A (t),x B (t),x C (t),x D (t)] T are the displacements of different components at time t, K, C and M are the structural parameter matrices, where they are the equivalent stiffness matrix, equivalent damping matrix and equivalent mass matrix of the structure, η is the external force influence matrix, and f(t) is the external overload.

[0041] According to the above-established elastic multi-body dynamic model, different connection and contact modes have different equivalent stiffness and equivalent damping coefficients. The estimation method of the elements in the equivalent stiffness matrix K and the equivalent damping matrix C is related to the natural frequency f of the axial vibration of the elastic multi-body system. The axial natural frequency can be approximated by the finite element method, K i 、C i It can be derived from the following formula:

[0042]

[0043] Among them, K i (i=1,2...,n) represents the corresponding element in the equivalent stiffness matrix, C i (i=1,2...,n) represents the corresponding element in the equivalent stiffness matrix, n represents the number of elements in the equivalent stiffness matrix and the equivalent damping matrix; M j represents the corresponding element in the equivalent mass matrix, j represents the equivalent mass of the corresponding stiffness and damping connection; w i represents the natural frequency, ζ i represents the dimensionless damping ratio.

[0044] (20) Parameter calibration: Using the finite element method, the harmonic response analysis of the projectile multi-body system model is performed to obtain the connection relationship parameters, which include the equivalent stiffness coefficient and the equivalent damping coefficient of the projectile multi-body system.

[0045] To analyze the characteristics of a multibody system under overload, it's necessary to know the equivalent stiffness and damping coefficients. Therefore, a harmonic response analysis is performed using finite element simulation software, focusing on the multibody dynamics system. The results of this harmonic response analysis can be used to determine the numerical values ​​of the connection relationships.

[0046] The finite element simulation software for performing harmonic response analysis is a well-known prior art in the art and will not be described in detail herein.

[0047] In the embodiment, the natural frequency values ​​of different connection relationships are as follows: the natural frequency of the shell and the warhead is 2.84kHz, the amplification factor is 3.43×10 4 g, the dimensionless damping ratio is 0.063; the natural frequency of the warhead and sensor is 4.83 kHz, and the amplification factor is 3.99×10 4 g, the dimensionless damping ratio is 0.152; the natural frequency of the sensor and the bottom thread is 16.81 kHz, and the amplification factor is 2.01×10 4 g, the dimensionless damping ratio is 0.147; the natural frequency of the shell and sensor is 2.44 kHz, and the amplification factor is 0.68×10 4 g, the dimensionless damping ratio is 0.114; the natural frequency of the shell and sensor is 2.44kHz, and the amplification factor is 0.19×10 4 g, and the dimensionless damping ratio is 0.052. By introducing formula (6), the equivalent damping and equivalent stiffness can be calculated as: 179.86×10 3 KN / m, 429.251×10 3 kN / m, 519.940×10 3 kN / m, 3.812×10 3 kN / m, 1.132×10 3 kN / mm, and the damping are 19.253kN·s / m, 26.128kN·s / m, 9.094kN·s / m, 0.356kN·s / m, and 0.0482kN·s / m respectively. The outer casing weighs 22.3Kg, the warhead weighs 18.4Kg, the sensor weighs 1.8400Kg, and the bottom screw weighs 0.6499Kg.

[0048] Generally speaking, equivalent mass can be roughly estimated through system design, so the parameters in the equivalent mass matrix can be considered known. For equivalent stiffness and equivalent damping, since multibody systems do not involve significant damage, the parameters in this matrix can be considered linear and time-invariant, and the approximate parameter range can be calibrated using finite element simulation software. However, even with this approach, the response of the established model will inevitably deviate from the actual model, and it cannot fully represent the state of the original system.

[0049] (30) Dual extended Kalman filter construction: According to the connection relationship parameters, a dual Kalman filter based on the least squares method is constructed;

[0050] like Figure 3 As shown, the (30) dual extended Kalman filter construction steps include:

[0051] (31) Setting the initial state and initial structure parameters: Set the initial state and initial structure parameters as follows:

[0052]

[0053] Where Z0 and θ0 are the initial state and initial structure parameters of the system respectively, and are the initial state and initial structural parameter estimation of the system, and are the system initial state covariance matrix and the initial structural parameter covariance matrix respectively, and E() is for finding the mean, such as E(Z0) is for finding the mean of Z0.

[0054] (32) Update state estimate: Calculate the new state estimate as follows:

[0055] The state vector is formed into an extended state vector Then the structural parameter matrix is ​​designed as θ(t)=[θ1(t),θ2(t)...θ n (t)] T According to the system modeling characteristics, considering the existence of certain process noise w(t) and measurement error v(t), the state equation and observation equation of the system can be written as

[0056]

[0057] It is assumed that the process noise w(t) and the measurement noise v(t) are Gaussian white noise sequences with zero mean, and their covariance matrix can be described in detail as: E[w(i)w T (j)]=Q Z δ(ij),E[v(i)v T (j)]=R Z δ(ij), and E[w(i)v T (j)], where δ(ij) is the Kronecker product. If and only if i = j, δ(ij) = 1, otherwise δ(ij) = 0. For θ(t), there are similar process errors and measurement errors. The covariance matrices of process noise and measurement noise are Q θ and R θ ,

[0058] Let t = kΔt, where Δt is the sampling step, then the discretization expression of w(kΔt) and v(kΔt) can be written as w k and v k ; From formula (9), it can be concluded that the observation matrix also belongs to the nonlinear function of the extended state, and then it can be rewritten as follows after discretization:

[0059] y k =h(Z k )-B u f k +v k (10)

[0060] in:

[0061]

[0062] B u =-M -1 η (12)

[0063] set up and Z k and f k The estimated value of and Performing linear expansion at , we can get:

[0064]

[0065] in:

[0066]

[0067] Then Substituting into formula (8) and combining (13) and (14) we can obtain:

[0068]

[0069] In addition, substituting formula (10) into formula (9), we can get:

[0070] B u f k =h(Z k )-y k +v k (16)

[0071] Based on the sensor's observation value, the least square method can be used to solve f k The approximate value of is:

[0072]

[0073] Error e for external overload f It can be written as:

[0074]

[0075] Where I is the identity matrix; is the projection matrix. When the partial overload approaches its true value, formula (10) approaches zero and we can get:

[0076] Φy k =Φh(Z k )+Φv k (19)

[0077] in: Equation (19) can be regarded as another improved observation equation. Compared with Equation (9), after introducing the projection matrix, this improved observation equation no longer contains the term that explicitly includes the external overload. In order to discretize the state equation, the estimated value of the state vector at step t = (k+1)Δt is for:

[0078]

[0079] (33) Update the state estimation error covariance: Calculate the new state estimation error covariance as follows:

[0080] From formula (15) and formula (20), we can get:

[0081]

[0082] Then through h(Z(k)) in Linear expansion is performed at:

[0083]

[0084] in:

[0085]

[0086] Further, we can get the covariance matrix of the prior error as:

[0087]

[0088] in Γ1=I+ΔtU k|k +Δt[0-H k|k 0] T , I is the diagonal unit moment matrix;

[0089] (34) Update the state estimation posterior covariance matrix: Calculate the new state estimation posterior covariance matrix as follows:

[0090] Based on (19) and its Bayesian estimation principle, the posterior state estimate is obtained by estimating the prior information and weighting the observations:

[0091]

[0092] in, is the Kalman gain matrix of EKF, which can be obtained by the following formula:

[0093]

[0094] The covariance matrix of the posterior estimation error can be further solved:

[0095]

[0096] Then the posterior covariance matrix can be expressed as:

[0097]

[0098] in

[0099] (35) Update structural parameters: Calculate the new structural parameters as follows:

[0100]

[0101] in are the estimated values ​​of the structural parameters, including the values ​​of different damping and parameter matrices, Update values ​​for model parameter state variables;

[0102] (36) Update the structural parameter error covariance: Calculate the new structural parameter error covariance as follows:

[0103]

[0104] in is the covariance matrix of structural parameter errors;

[0105] (37) Update the posterior covariance matrix of the structural parameters: Calculate the new posterior covariance matrix of the structural parameters as follows:

[0106]

[0107] in, is the posterior covariance matrix; is the gain matrix.

[0108] (40) Construction of short trajectory multi-input GRU neural network: Based on the GRU type neural network structure, a mature short trajectory multi-input GRU neural network is constructed and trained;

[0109] like Figure 4 As shown, the (40) short trajectory multi-input GRU neural network construction steps include:

[0110] (41) Constructing the GRU neural network unit: Construct the GRU neural network unit according to the following formula:

[0111] H t =(1-S t )H t '+S t H t-1 (32)

[0112] Among them, H t Indicates the output result, S t represents the update gate, H t-1 represents the past hidden state, H t ' represents the candidate hidden state;

[0113] The candidate hidden state is calculated as follows,

[0114] H t '=tanh(WZ t +UR t H t-1 +B) (33)

[0115] Where W and U represent weights, B represents the bias of the update gate, and Z t Represents the network input, R t Reset gate.

[0116] Update gate S G (t) is calculated as follows,

[0117] S G (t)=σ(W S ·Z G (t)+U S ·H G (t-1)+B S ) (34)

[0118] Where W S and U S represents the information weight of the update gate, B S represents the information bias of the update gate, Z G (t) represents the network input, H G (t-1) is the past hidden state;

[0119] Reset Gate R G (t) is calculated as follows,

[0120] R G (t)=σ(W R ·Z G (t)+U R ·H G (t-1)+B R ) (35)

[0121] Where W R and U R represents the information weight of the update gate, B R represents the information bias of the update gate;

[0122] (42) Constructing BI-GRU neural network unit: Construct BI-GRU neural network unit according to the following formula:

[0123]

[0124] Where GRU() represents the GRU nonlinear transformation output; represents the hidden state at time t, Represents the forward hidden state at time t-1, Represents the reverse hidden state at time t-1, W t Represents the forward hidden state at time t The weight value, V t Represents the reverse hidden state The weight value, B t Represents the bias value at time t. The BI-GRU consists of two GRUs stacked on the top and bottom, and the output is determined by the states of these two GRUs.

[0125] (43) Constructing a short trajectory multi-input GRU neural network:

[0126] The short trajectory multi-input GRU neural network consists of an input layer, a GRU layer, a BI-GRU layer, a fully connected layer, and an output layer. The input layer is an m×n dimensional time series input vector, which includes the initial data and the reconstructed data of DLEKF. The GRU layer is a multi-input m×n dimensional GRU unit that matches the input layer and can realize parallel processing of multiple sets of data inputs. The BI-GRU layer integrates and processes the data of the GRU layer, and the BI-GRU layer further realizes the learning of time series information through a forward and backward transmission method. The fully connected layer realizes the estimation of the required prediction step and realizes the state reconstruction of the components that cannot be directly measured. The output layer outputs the result. The input layer structural unit (GRU) is as follows: Figure 5 As shown, the GRU neural network architecture is as follows Figure 6 shown.

[0127] (44) Training short trajectory multi-input GRU neural network:

[0128] Based on simulation and experimental data, sensor acceleration data is collected from measurable components in a multi-body system. The collected data is divided into a training set for training the model, a validation set for adjusting model parameters and preventing overfitting, and a test set for evaluating model performance. During the training process, all initial signals in the training set are first reconstructed using LDEKF information to obtain the reconstructed signal of the target component. The initial and reconstructed signals are then used as input, and the signal of the target component obtained through simulation or experiment is used as output to train the neural network, resulting in a well-trained short-trajectory multi-input GRU neural network.

[0129] (50) Reconstruction of the system dynamics state: The original sensor data and the output data of the Kalman filter are used as the input of the mature short trajectory multi-input GRU neural network, and the dynamic reconstruction signal of the projectile multi-body system is obtained from its output.

[0130] This improved solution of the present invention uses a multi-body dynamics state reconstruction method assisted by a GRU neural network using DLEKF to reconstruct the system state of unmeasurable components using sensor information from measurable components. This achieves faster inference speed, higher accuracy, and higher computational efficiency.

[0131] Whether from the theoretical perspective of signal processing or from the practical perspective of penetrating ammunition functionality, the accuracy of impact type identification is the most important performance indicator.

[0132] The typical working conditions of 800m / s and 900m / s penetrating 100mm were selected for analysis. GRU, DLEKF and the method proposed in this article were used for state reconstruction respectively. The DLEKF method can reduce the error of the original signal, and only requires data at the current moment, and the delay required by the algorithm is very low. The GRU method can effectively fit the changing trend of the sensor signal, but this method requires more time series trajectories as input, which will not only cause the execution delay of the algorithm to become longer, but also affect the actual use of stored data. The method proposed in this article only requires shorter time series trajectories as input, and more accurate state reconstruction results can be obtained through more reconstruction information. The variance comparison of different methods is shown below. Figure 7 and Figure 8 shown. Figure 7 The figure shows the variance comparison of the reconstructed state using different methods at a speed of 800 m / s. Figure 8 The figure shows the variance comparison of the reconstructed state using different methods at a speed of 900 m / s.

[0133] from Figure 7 and Figure 8As can be seen in the figure, the error reaches its maximum around 0.025ms, and the maximum variance of the proposed method is 9.25% and 16.34% of that of DLEKF and GRU, respectively. Furthermore, over the entire time period, the mean squared error (MSE) of the proposed method is only 5.44% and 8.38%, indicating that the proposed method has significantly improved in simulation.

[0134] In the air cannon experiment, three runs were conducted, two of which were used as training sets, and the last as the test set. By combining the strengths of both approaches, the proposed method achieved significantly lower state prediction errors than the other two methods. In test set 1, the MSE was reduced by 37.1% and 67.6% compared to the GRU and DLEKF, respectively; the mean absolute error (MAE) was reduced by 17.8% and 39.2% compared to the GRU and DLEKF, respectively; and the maximum error (MAXERROR) was reduced by 28.3% and 54.9% compared to the GRU and DLEKF, respectively. Due to the high similarity of the signals used in test set 1, the improvement was limited. In test set 2, the MSE was reduced by 79.9% and 67.7% compared to the GRU and DLEKF, respectively; the MAE was reduced by 60.1% and 61.3% compared to the GRU and DLEKF, respectively; and the MAXEROR was reduced by 73.9% and 45.9% compared to the GRU and DLEKF, respectively. However, the signal similarity for test set 2 is not high. In this case, the data-based approach alone has disadvantages in many parameters compared to the model-based approach, with some indicators showing a significant decline. The proposed method combines the advantages of both approaches, resulting in a significant improvement in accuracy.

Claims

1. A method for reconstructing the dynamic state of a multi-body system based on a model-assisted network, characterized in that: The steps include: (10) Model establishment: Using the multi-body system dynamics method, the projectile multi-body system model is established; (20) Parameter calibration: using the finite element method, performing harmonic response analysis on the projectile-guided multi-body system model to obtain connection relationship parameters, wherein the connection relationship parameters include the equivalent stiffness coefficient and the equivalent damping coefficient of the projectile-guided multi-body system; (30) Dual extended Kalman filter construction: According to the connection relationship parameters, a dual Kalman filter based on the least squares method is constructed; (40) Construction of short trajectory multi-input GRU neural network: Based on the GRU type neural network structure, a mature short trajectory multi-input GRU neural network is constructed and trained; (50) System dynamics state reconstruction: using the original sensor data and the output data of the Kalman filter as the input of the mature short trajectory multi-input GRU neural network, and obtaining the dynamic reconstruction signal of the projectile multi-body system from its output; The model building steps (10) are specifically as follows: According to Newton's second law, the response equation of the elastic multi-body system is established as shown below to determine the mechanical transmission relationship; Among them, M A 、M B 、M C and M D Respectively represent the mass of the projectile, the mass of the medicine, the mass of the sensor shell and the mass of the bottom screw, and x A ,x B ,x C and x D Represent the corresponding displacements, and Represent the corresponding speeds, and Represent the corresponding accelerations respectively, K1, K2, K3, K4 and K5 are the corresponding equivalent stiffness coefficients respectively, C1, C2, C3, C4 and C5 are the corresponding equivalent damping coefficients respectively, F A For external overload; The (30) dual extended Kalman filter construction step includes: (31) Set the initial state and structure parameters: Set the initial state and initial structure parameters as follows: Among them, Z0 and θ0 are the initial state and initial structure parameters of the system respectively, and are the initial state and initial structural parameter estimation of the system, and are the system initial state covariance matrix and initial structural parameter covariance matrix respectively, and E(Z0) is the mean value of Z0; (32) Update state estimate: Calculate the new state estimate as follows: The state vector is formed into an extended state vector Then the structural parameter matrix is ​​designed as θ(t)=[θ1(t),θ2(t)...θ n (t)] T According to the system modeling characteristics, considering the existence of certain process noise w(t) and measurement error v(t), the state equation and observation equation of the system can be written as It is assumed that the process noise w(t) and the measurement noise v(t) are Gaussian white noise sequences with zero mean, and their covariance matrix can be described in detail as: E[w(i)w T (j)]=Q Z δ(ij),E[v(i)v T (j)]=R Z δ(ij), and E[w(i)v T (j)], where δ(ij) is the Kronecker product. If and only if i = j, δ(ij) = 1, otherwise δ(ij) = 0. For θ(t), there are similar process errors and measurement errors. The covariance matrices of process noise and measurement noise are Q θ and R θ , Let t = kΔt, where Δt is the sampling step, then the discretization expression of w(kΔt) and v(kΔt) can be written as w k and v k ; From formula (4), it can be concluded that the observation matrix also belongs to the nonlinear function of the extended state, so it can be rewritten as follows after discretization: y k =h(Z k )-B u f k +v k (5) in: h(Z k )=-M -1 [Cx(t)+Kx(t)] (6) B u =-M -1 the (7) set up and Z k and f k The estimated value of and Performing linear expansion at , we can get: in: Then Substituting into formula (3) and combining (8) and (9) we can get: In addition, substituting formula (5) into formula (4), we can get: Based on the sensor's observation value, the least square method can be used to solve f k The approximate value of is: Error e for external overload f It can be written as: Where I is the identity matrix; is the projection matrix. When the external overload approaches its true value, Equation (13) approaches zero and we can obtain: Φy k =Φh(Z k )+Φv k (14) in: Equation (14) can be regarded as another improved observation equation. Compared with Equation (4), after introducing the projection matrix, this improved observation equation no longer contains the term that explicitly includes the external overload. In order to discretize the state equation, the estimated value of the state vector at step t = (k+1)Δt is for: (33) Update the state estimation error covariance: Calculate the new state estimation error covariance as follows: From formula (10) and formula (15), we can get: Then through h(Z(k)) in Linear expansion is performed at: in: Further, we can get the covariance matrix of the prior error as: in Γ1=I+ΔtU k|k +Δt[0-H k|k 0] T , I is the diagonal unit moment matrix; (34) Update the state estimation posterior covariance matrix: Calculate the new state estimation posterior covariance matrix as follows: Based on (14) and its Bayesian estimation principle, the posterior state estimate is obtained by estimating the prior information and weighting the observations: in, is the Kalman gain matrix of EKF, which can be obtained by the following formula: The covariance matrix of the posterior estimation error can be further solved: Then the posterior covariance matrix can be expressed as: in (35) Update structural parameters: Calculate the new structural parameters as follows: in are the estimated values ​​of the structural parameters, including the values ​​of different damping and parameter matrices, Update values ​​for model parameter state variables; (36) Update the structural parameter error covariance: Calculate the new structural parameter error covariance as follows: in is the covariance matrix of structural parameter errors; (37) Update the posterior covariance matrix of the structural parameters: Calculate the new posterior covariance matrix of the structural parameters as follows: in, is the posterior covariance matrix; is the gain matrix.

2. The dynamic state reconstruction method according to claim 1, characterized in that: The (40) short trajectory multi-input GRU neural network construction steps include: (41) Constructing the GRU neural network unit: Construct the GRU neural network unit according to the following formula: H t =(1-S t )H′ t +S t H t-1 (27) Among them, H t Indicates the output result, S t represents the update gate, H t-1 represents the past hidden state, H t ' represents the candidate hidden state; The candidate hidden state is calculated as follows, H′ t =tanh(WZ t +UR t H t-1 +B) (28) Where W and U represent weights, B represents the bias of the update gate, and Z t Represents the network input, R t Reset gate. Update gate S G (t) is calculated as follows, S G (t)=σ(W S ·Z G (t)+U S ·H G (t-1)+B S ) (29) Where W S and U S represents the information weight of the update gate, B S represents the information bias of the update gate, Z G (t) represents the network input, H G (t-1) is the past hidden state; Reset Gate R G (t) is calculated as follows, R G (t)=σ(W R ·Z G (t)+U R ·H G (t-1)+B R ) (30) Where W R and U R represents the information weight of the update gate, B R represents the information bias of the update gate; (42) Constructing BI-GRU neural network unit: Construct BI-GRU neural network unit according to the following formula: Where GRU() represents the GRU nonlinear transformation output; represents the hidden state at time t, Represents the forward hidden state at time t-1, Represents the reverse hidden state at time t-1, W t Represents the forward hidden state at time t The weight value, V t Represents the reverse hidden state The weight value, B t Represents the bias value at time t. The BI-GRU consists of two GRUs stacked on the top and bottom, and the output is determined by the states of these two GRUs. (43) Constructing a short trajectory multi-input GRU neural network: The short-trajectory multi-input GRU neural network consists of an input layer, a GRU layer, a BI-GRU layer, a fully connected layer, and an output layer. The input layer is an m×n-dimensional time series input vector, including the initial data and the reconstructed data from the DLEKF. The GRU layer is a multi-input m×n-dimensional GRU unit that matches the input layer and can achieve parallel processing of multiple sets of data inputs. The BI-GRU layer integrates and processes the data from the GRU layer and further learns the time series information through a forward and backward transmission method. The fully connected layer estimates the required prediction steps and reconstructs the state of components that cannot be directly measured. The output layer outputs the results. (44) Training short trajectory multi-input GRU neural network: Based on simulation and experimental data, sensor acceleration data is collected from measurable components in a multi-body system. The collected data is divided into a training set for training the model, a validation set for adjusting model parameters and preventing overfitting, and a test set for evaluating model performance. During the training process, all initial signals in the training set are first subjected to LDEKF information reconstruction processing to obtain the reconstructed signal of the target component. The initial signal and the reconstructed signal are then used as input, and the signal of the target component obtained through simulation or experiment is used as output to train the neural network, resulting in a mature short-trajectory multi-input GRU neural network.

Images