Physical-data collaborative air bag vibration isolation system shafting centering state evaluation prediction model optimization method

Abstract

The invention discloses a physics-data collaborative optimization method for an air bag vibration isolation system shafting centering state evaluation prediction model, and the method comprises the steps: constructing an air bag vibration isolation system shafting centering state evaluation mathematical model, and carrying out the initialization according to design parameters; constructing a physical residual neural network and a centering state evaluation neural network at the same time; adding the collected data into a data set D; generating a physical residual network data set Dc according to the data set D; performing physical residual neural network training on the physical residual network data set Dc to obtain a centering state physical residual prediction model; generating a centering state evaluation data set Dr by using the centering state evaluation mathematical model and the centering state physical residual prediction model; and performing centering state evaluation neural network training on the centering state evaluation data set Dr to obtain a centering state evaluation prediction model. According to the method, physical knowledge and actually measured data are taken into consideration, dependency on collected data is reduced, and the method has high applicability.

Description

Technical Field

[0001] The present invention relates to the technical field of vibration reduction and noise reduction, and particularly to an optimization method for an evaluation and prediction model of the alignment state of the shafting of a physical-data collaborative airbag vibration isolation system. Background Art

[0002] At present, the airbag vibration isolation system has become a key technology for isolating mechanical vibration and noise in ship propulsion systems. The airbag vibration isolators in the system have a low natural frequency and good vibration isolation performance, and the alignment state of the shafting can be adjusted by controlling the airbag pressure to ensure the safe operation of the propulsion system. In engineering practice, the airbag vibration isolation system usually uses equipment such as eddy current sensors and laser alignment instruments to monitor the alignment state of the shafting in real time, and deploys pressure sensors to monitor the pressure state, which is used to detect the health state of the airbag and assist in completing control decisions. However, to carry out shafting alignment control, establishing the mapping relationship between the control object (airbag pressure) and the control target (shafting alignment state) is the prerequisite and key.

[0003] Existing technologies mostly construct mathematical models based on the assumptions of rigid body dynamics, and rely on prior knowledge of physical parameters such as the centroid position and stiffness coefficient of the bearing equipment. However, there are non-linear time-varying factors such as raft frame deformation, multi-condition coupling (such as attitude disturbance, load fluctuation) and equipment aging in the actual system, resulting in difficult accurate acquisition of model parameters and a significant decline in model prediction accuracy. For this reason, Patent CN2024108330624 proposes a pure data-driven method for evaluating the alignment state of the shafting. Although this method can avoid the problem of mechanism modeling through the state space mapping relationship, the model that completely relies on data training often lacks the utilization of physical laws, resulting in the following problems: 1) poor generalization ability under untrained conditions and susceptibility to sensor noise interference; 2) difficult to explain the physical meaning of the evaluation results for guiding the optimization of control strategies.

[0004] Therefore, the mechanism model is mismatched with the actual working conditions due to relying on accurate parameters, and the data-driven model ignores physical laws and lacks interpretability. There is an urgent need for a method for evaluating the alignment state of the shafting that integrates weak physical knowledge and data driving, while reducing the model dependence, ensuring the physical rationality and working condition adaptability of the evaluation results. Summary of the Invention

[0005] The purpose of the present invention is to address the above-mentioned deficiencies in the technology and provide a method for evaluating the alignment state of the shafting of an airbag vibration isolation system based on weak physical knowledge, while reducing the model dependence and ensuring the physical rationality and working condition adaptability of the evaluation results.

[0006] To achieve the above object, the present invention provides an optimization method for an axial alignment state evaluation prediction model of a physical-data collaborative airbag vibration isolation system. The airbag vibration isolation system includes a group of airbag vibration isolators. A main coordinate system is established with the center of gravity position of the main engine under ideal alignment conditions as the origin: the axial direction is y, the horizontal direction perpendicular to the axial direction y is x, and the direction perpendicular to the axial direction y and upward is z. γ represents the skew of the axial direction y, β represents the skew of the direction x, and α represents the skew of the direction z. A local coordinate system is established for each airbag, with the upward direction of the main elastic axis as the vertical direction R and the axial direction as Q. At the same time, the radial direction T is perpendicular to the axial direction Q and the vertical direction R. It is characterized in that the optimization method for the axial alignment state evaluation prediction model is specifically as follows:

[0007] 1) Construct a mathematical model for evaluating the axial alignment state of the airbag vibration isolation system and initialize it according to the design parameters; at the same time, construct a physical residual neural network and a neural network for evaluating the alignment state;

[0008] 2) Collect data and add it to the data set D;

[0009] 3) Generate a physical residual network data set D c ;

[0010] 4) Train the physical residual neural network on the physical residual network data set D c to obtain a physical residual prediction model for the alignment state;

[0011] 5) Use the mathematical model for evaluating the alignment state and the physical residual prediction model for the alignment state to generate an evaluation data set D r ;

[0012] 6) Train the neural network for evaluating the alignment state on the evaluation data set D r to obtain an evaluation prediction model for the alignment state.

[0013] Furthermore, the specific calculation process of the mathematical model for evaluating the alignment state in step 1) is as follows: Assuming that the raft frame is a rigid body structure, a simplified dynamic equation for the shafting state - airbag pressure is constructed as follows:

[0014] K(P)X g =F ext

[0015] where X g =[x g ,y g ,z g ,β,γ,α] is the translation and rotation motion matrix of the center of gravity of the airbag vibration isolation system in three directions of the main coordinate system. y g represents the offset in the axial direction y, x g represents the offset in the direction x, and z g represents the offset in the direction z; Fext is external interference, such as rolling and pitching moments, changes in equipment loads, etc.; K(P) is the system stiffness matrix, and the system stiffness matrix is related to the airbag pressure P distribution P = {p1,..., p N}, where p i represents the pressure of the i-th airbag, N represents the number of airbags in the system, and the system stiffness matrix K(p) is expressed as:

[0016]

[0017] where, G i is the position matrix, represents the position of the i-th airbag isolator in the main coordinate system, and the position matrix G i is expressed as:

[0018]

[0019] Γ i represents the transpose matrix and is expressed as:

[0020]

[0021] where the element represents the cosine value of the angle between the main elastic axis of the i-th airbag isolator and the main coordinate system of the system, a = T, Q, R, b = x, y, z; k(p i ) represents the stiffness, and the natural stiffnesses of the i-th airbag in the radial direction T, axial direction Q, and vertical direction R are respectively The pressure correction term is proportional to the current airbag pressure p i , and the pressure correction coefficients in the three directions are η p , η q , η r , then:

[0022]

[0023] When the pressure of the i-th airbag changes, only the change in bearing capacity in the vertical direction is considered, and the change amount is Δf i r = Δp i A e , where A e represents the effective area of the airbag and is a constant, and Δp i is the change amount of the pressure of the i-th airbag; for the pressure state p0 and the centering state corresponding to the pressure state p0 under known initial conditions calculate the centering state under the target pressure state That is, the centering state evaluation mathematical model is as follows:

[0024]

[0025] Among them, G c represents the position transformation matrix, and its purpose is to transform the coordinates at the center of gravity to the center position of the output axis system, expressed as:

[0026]

[0027] In the formula represents the three components of the vector from the origin of the main coordinate system (the center of gravity of the bearing device) to the center of the output axis system on the main coordinate system.

[0028] Furthermore, the physical residual neural network in the step 1) is a backpropagation neural network (BP-NN), including 1 input layer, 1 output layer and H r hidden layers, and activation layers are added between the output layer and the hidden layers, and between the hidden layers. The ReLU function is used as the activation function; the data input by the input layer is the airbag pressure value and the working condition data, and the data output by the output layer is the difference between the calculated value of the centering state obtained by the centering state evaluation mathematical model and the measured data of the centering state.

[0029] Furthermore, the centering state evaluation neural network in the step 1) is a backpropagation neural network (BP-NN), including 1 input layer, 1 output layer and H e hidden layers, and activation layers are added between the output layer and the hidden layers, and between the hidden layers. The ReLU function is used as the activation function; the data input by the input layer is the airbag pressure value and the working condition data, and the data output by the output layer is the centering state.

[0030] Furthermore, in the step 2), the airbag pressure value P = {p1, p2, p i ,..., p N} of the airbag pressure sensor is collected, where p i represents the airbag pressure value of the i-th airbag isolator; the working condition data C = {c1, c2, c i ,..., c d} is collected, where c i represents the i-th type of working condition data; the measured data of the shaft alignment state is collected Each set of collected data forms a record and is added to the data set D.

[0031] Furthermore, in the step 3), the data set D is traversed. For each record According to the airbag pressure value P in the data set D, the calculated value X g ' of the centering state corresponding to the airbag pressure value P is calculated using the centering state evaluation mathematical model, and a new data record R c = (P, C, ΔX g ) is generated, where Add new data records to the physical residual network dataset D c .

[0032] Furthermore, in step 4), the physical residual network dataset D c is randomly divided into a training set and a test set, where the training set accounts for 80% and the test set accounts for 20%. The training error is denoted as L c , and the root mean square is used as the evaluation criterion for the error L c ;

[0033] When training reaches , training is stopped, being the training target threshold;

[0034] After training, a centering state physical residual prediction model is obtained, denoted as:

[0035]

[0036] Furthermore, in step 5), the generation process of the centering state evaluation dataset D r is specifically as follows:

[0037] 51) Use the Latin hypercube sampling method to sample within the working pressure range of each airbag isolator to obtain a set of pressure value sampling point combinations P s ={P1,...,P M}; Sample different working conditions to obtain a set of working condition sampling point combinations C s ={C1,...,C K};

[0038] 52) Use the centering state evaluation mathematical model to calculate the centering state X i corresponding to the pressure value of the pressure value sampling point P g (1≤i≤M);

[0039] 53) Use the centering state physical residual prediction model to calculate the physical residuals under different working condition sampling points C j ;

[0040] 54) Traverse all airbag pressures and working condition sampling points, and add the calculated records to the centering state evaluation dataset D r ;

[0041] 55) Add all records in the dataset D to the dataset D r ;

[0042] Furthermore, in step 6), the dataset D rRandomly divided into a training set and a test set, where the training set accounts for 80% and the test set accounts for 20%. The training error is denoted as L, and the root mean square is used as the error evaluation criterion;

[0043] When training until L ≥ L T then stop training, and L T is the training target threshold.

[0044] After training, obtain the alignment state evaluation prediction model, denoted as:

[0045] X g = h(P, C)

[0046] If L < L T after training for N rounds, then jump to step 2).

[0047] Finally, save the parameters of the alignment state evaluation network, which can be deployed offline or online for use.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows: 1) Taking into account physical knowledge and measured data, reducing the dependence on collected data, making the method itself have strong applicability; 2) Making full use of the collected data, allowing deviations in the physical model, and realizing the self-learning of the model; 3) Providing another means of evaluating the alignment state other than directly measuring the alignment state, which can effectively improve the system reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic diagram of the composition of the airbag vibration isolation system and the alignment coordinates of the shafting;

[0050] Figure 2 It is a schematic diagram of the alignment state evaluation process;

[0051] Figure 3 It is a schematic diagram of the training process of the pure data method and the solution proposed in this patent. DETAILED DESCRIPTION OF THE INVENTION

[0052] To facilitate the understanding of the present invention, the present invention will be described comprehensively below with reference to the relevant drawings. The preferred embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0053] Such as Figure 1As shown, a typical airbag vibration isolation system for a propulsion motor usually uses multiple airbag vibration isolators to carry the equipment. The upper surface of the airbag vibration isolator is in contact with and installed on the main machine, and the lower surface is fixed on the base. The airbag vibration isolation system includes a control system that can measure the height changes at different positions through displacement sensors to achieve the purpose of monitoring the alignment state. It can also change the airbag pressure by inflating or deflating the airbag vibration isolator, thereby adjusting the alignment state of the shafting. Taking the center of gravity position of the main machine under ideal alignment conditions as the origin, a main coordinate system is established: the axial direction is y, the horizontal direction perpendicular to the axial direction y is x, and the upward direction perpendicular to the axial direction y is z. γ represents the skew in the axial direction y, β represents the skew in the direction x, and α represents the skew in the direction z. A local coordinate system is established for each airbag. The upward direction of the main elastic axis is the vertical direction R, the axial direction is Q, and at the same time, the radial direction T is perpendicular to the axial direction Q and the vertical direction R.

[0054] Assume that the airbag vibration isolation system of the present invention includes N = 8 airbag vibration isolators, which are evenly deployed on both sides of the main machine, and the airbags are inclined at 60 degrees, that is, the included angle with the x-axis is 60 degrees. Each airbag vibration isolator is equipped with an airbag pressure sensor. The method for evaluating the alignment state of the shafting of the physical-data-driven airbag vibration isolation system provided by the present invention is as Figure 2 shown, specifically:

[0055] 1) Initialization: Construct a mathematical model for evaluating the alignment state of the shafting of the airbag vibration isolation system and initialize it according to the design parameters; at the same time, construct a physical residual neural network and an alignment state evaluation neural network.

[0056] Furthermore, the specific calculation process of the alignment state evaluation mathematical model in step 1) is as follows: Assume that the raft is a rigid body structure, and at this time, a simplified dynamic equation of the shafting state - airbag pressure is constructed as follows:

[0057] K(P)X g =F ext

[0058] where X g =[x g ,y g ,z g ,α,β,γ] is the translation and rotation motion matrix of the center of gravity of the vibration isolation system in the three directions (radial, axial, and vertical, and the three directions are perpendicular to each other) of the coordinate axis. y g represents the offset in the axial direction y, x g represents the offset in the direction x, z g represents the offset in the direction z; F ext is an external disturbance, such as transverse and longitudinal tilting moments, changes in equipment loads, etc.; K(P) is the system stiffness matrix, and the system stiffness matrix is related to the airbag pressure P distribution P = {p1,..., p N}, where p iDenote the pressure of the $i$-th airbag, $N$ represents the number of airbags in the system, and the system stiffness matrix $K(P)$ is expressed as:

[0059]

[0060] Among them, $G$ i is the position matrix, and use to represent the position of the $i$-th airbag isolator in the coordinate system. The position matrix can be expressed as:

[0061]

[0062] Assume that the position of the first airbag isolator in the coordinate system is $(x_1,y_1,z_1)=(900,1000, - 600)$ in mm, then its position matrix can be expressed as:

[0063]

[0064] $\Gamma$ i represents the transpose matrix, which is expressed as:

[0065]

[0066] The element among them represents the cosine value of the angle between the main elastic axis of the $i$-th airbag isolator and the overall coordinate axis of the system. Considering that the inclination angle of the airbag isolator is 60 degrees and the first airbag sensor is on the left side of the axis system output direction, as Figure 1 shown, then the included angles between its local coordinate system $T$, $Q$, $R$ and the main coordinate system $x$, $y$, $z$ are respectively The transpose matrix $\Gamma$ i is:

[0067]

[0068] $k(p$ i ) represents the stiffness. Assume that the natural stiffness of the $i$-th airbag in the radial direction $T$, axial direction $Q$, and vertical direction $R$ are respectively The pressure correction term is proportional to the current airbag pressure $p$ i . Let the pressure correction coefficients in the three directions be $\eta$ p , $\eta$ q , $\eta$ r , which are:

[0069]

[0070] Assume that the stiffness parameters of all airbags are the same, and the stiffness coefficient $k$ i r $ = 1.5$ in the main axis direction $R$, and the stiffness coefficients in the axial direction $Q$ and radial direction $T$ are the same, which are: The correction coefficients are respectively: $\eta$p = η q = 1.2, η r = 0.6, at this time:

[0071]

[0072] When the pressure of the i-th airbag changes, only consider the change in bearing capacity in the vertical direction, and the change amount is Δf i r = Δp i A e , where A e represents the effective area of the airbag, and assume it is a constant A e = 7800mm 2 . For the pressure state P0 = {0.72, 0.73, 0.73, 0.74, 0.86, 0.84, 0.84, 0.84} and the centering state under known initial conditions Calculate the centering state under the target pressure state That is, the centering state evaluation mathematical model is as follows:

[0073]

[0074] Among them, G c represents the position transformation matrix, and its purpose is to transform the coordinates at the center of gravity to the center point position of the output axis system, and can be expressed as:

[0075]

[0076] In the formula, I is the identity matrix, represents the three components of the vector from the coordinate origin (the center of gravity of the bearing device) to the center of the output axis system on the coordinate axes. Assume that the coordinates of the center of the output axis system in the main coordinate system are (x, y, z) = (0, 1500, 0), then:

[0077]

[0078] So far, the parameters for calculating the centering state using the airbag pressure have been determined, and the centering state values under different airbag pressures can be calculated. Note that due to the incorporation of measured data, the mathematical model in the present invention does not require particularly accurate parameters. Note that considering that the axial offset and rotation have little impact on the shaft alignment, they are generally not considered in engineering. Here, the measured values only monitor X g = [x g , z g , α, β], and the output quantities of the physical residual neural network and the state evaluation network also only consider these four variables.

[0079] The physical residual neural network is a backpropagation neural network (BP-NN), including 1 input layer (with 9 input variables, including 8 airbag pressure values and 1 operating condition information), 1 output layer (with 4 output variables, including vertical offset, vertical skew, horizontal offset, and horizontal skew), and 2 hidden layers. An activation layer is added between the output layer and the hidden layer, and between the hidden layers. The ReLU function is used as the activation function. The data input by the input layer is the airbag pressure value and the operating condition data, and the data output by the output layer is the difference between the calculated centering state value obtained from the centering state evaluation mathematical model and the measured centering state data.

[0080] The centering state evaluation neural network is also a BP-NN, including 1 (with 9 input variables, including 8 airbag pressure values and 1 operating condition information), 1 output layer (with 4 output variables, including vertical offset, vertical skew, horizontal offset, and horizontal skew), and 2 hidden layers. An activation layer is added between the output layer and the hidden layer, and between the hidden layers. The ReLU function is used as the activation function.

[0081] 2) Collect data and add it to the dataset D: Collect the airbag pressure values P = {p1, p2, p i ,..., p N} of the airbag pressure sensor, where p i represents the airbag pressure value of the i-th airbag isolator; collect the operating condition data C = {c1, c2, c i ,..., c d}, where c i represents the i-th operating condition data; collect the measured centering state data of the shafting Note that if the axial offset y g and the axial skew γ cannot be collected, they can be directly set to zero for recording, that is, set y g = 0, γ = 0. Each collected data forms a record and add it to the dataset D.

[0082] 3) Generate the physical residual network dataset D c . Traverse the dataset D. For each record Calculate the calculated centering state value X g ' corresponding to the airbag pressure value P according to the airbag pressure value P in the dataset D using the centering state evaluation mathematical model, and generate a new data record R c = (P, C, ΔX g ), where Add the new data record to the physical residual network dataset D c .

[0083] 4) For the physical residual network dataset D cTrain a physical residual neural network to obtain a centering state physical residual prediction model. Divide the physical residual network dataset D c randomly into a training set and a test set, where the training set accounts for 80% and the test set accounts for 20%. Denote the training error as L c , and use the root mean square as the error evaluation criterion;

[0084] When training reaches , then stop training, which is the training target threshold.

[0085] After training, obtain the centering state physical residual prediction model, denoted as:

[0086] ΔX g = g(P, C)

[0087] 5) Use the centering state evaluation mathematical model and the centering state physical residual prediction model to generate the centering state evaluation dataset D r The generation process is as follows:

[0088] 51) Use the Latin hypercube sampling method to sample within the working pressure range of each airbag isolator to obtain a set of pressure value sampling point combinations P s = {P1,..., P M}; Sample different working conditions to obtain a set of working condition sampling point combinations C s = {C1,..., C K}. Assume that the working pressure range of each airbag isolator is 0 - 2.0 Mpa, and use the Latin hypercube sampling method to divide it into 4 equal parts. For example, the first equal part is 0 - 0.5 Mpa, and randomly sample a point within this equal part. The sampling point within the first equal part may be 0.38 Mpa. All the pressure value combinations form the sampling point combination P s , and there are 65536 = 4 8 elements in the combination; Assume there is only one rotational speed working condition here, and the rotational speed range is 0 - 200 rpm, which can be divided into 4 equal parts, and also randomly sample within each equal part. There are 5 elements in C s ;

[0089] 52) Use the centering state evaluation mathematical model to calculate the centering state X i (i) corresponding to the pressure value of P g (1 ≤ i ≤ M);

[0090] 53) Use the centering state physical residual prediction model to calculate the physical residuals under different working condition sampling points C j (1 ≤ j ≤ K);

[0091] 54) Traverse all the airbag pressures and working condition sampling points, and record the calculated Add to the in - state evaluation dataset D r ;

[0092] 55) Add all records in dataset D to the in - state evaluation dataset D r ;

[0093] 6) Train the in - state evaluation network. Randomly divide dataset D r into a training set and a test set, where the training set accounts for 80% and the test set accounts for 20%. Denote the training error as L, and use the root mean square as the error evaluation criterion;

[0094] When training until L≥L T , stop training, where L T is the training target threshold

[0095] After training, obtain the in - state evaluation prediction model, denoted as:

[0096] X g = h(P, C)

[0097] If after training for N rounds L < L T , then jump to step 2).

[0098] 7) Deploy and use. Save the parameters of the in - state evaluation network, which can be deployed and used offline or online.

[0099] After testing with a physical prototype, as Figure 3 shown, the proposed scheme of the present invention can converge faster than the pure data model; Table 1 shows the error comparison of the mathematical model, the pure data model and the scheme of the present invention. The present invention has achieved better prediction effects in all four components of the in - state.

[0100] Table 1

[0101]

Claims

1. An optimization method for the assessment and prediction model of the alignment state of the shafting in a physical-data collaborative airbag vibration isolation system. The airbag vibration isolation system includes a group of airbag vibration isolators. A main coordinate system is established with the center of gravity position of the main engine under ideal alignment conditions as the origin: the axial direction is y, the horizontal direction perpendicular to the axial direction y is x, and the upward direction perpendicular to the axial direction y is z. γ represents the skew of the axial direction y, β represents the skew of the horizontal direction x perpendicular to the axial direction y, and α represents the skew of the upward direction z perpendicular to the axial direction y. A local coordinate system is established for each airbag, with the upward direction of the main elastic axis as the vertical direction R and the axial direction as Q. At the same time, the direction perpendicular to the axial direction Q and the vertical direction R is the radial direction T. It is characterized in that: The optimization method for the shaft alignment status evaluation and prediction model is as follows: 1) Construct a mathematical model for evaluating the shaft alignment status of the airbag vibration isolation system and initialize it according to the design parameters; at the same time, construct a physical residual neural network and a neural network for evaluating the alignment status; 2) Collect data and add it to the dataset D; 3) Generate the physical residual network dataset D based on the dataset D c ; 4) For the physical residual network dataset D c Perform physical residual neural network training to obtain a centering state physical residual prediction model; 5) Generate the alignment status evaluation dataset D using the alignment status evaluation mathematical model and the alignment status physical residual prediction model r ; 6) Perform centering state evaluation neural network training on the centering state evaluation data set D r to obtain a centering state evaluation prediction model.

2. The optimization method for the shaft alignment state evaluation and prediction model of the physical-data collaborative airbag vibration isolation system according to claim 1, wherein: The specific calculation process of the mathematical model for evaluating the alignment status in step 1) is as follows: assuming that the raft frame is a rigid body structure, a simplified dynamic equation for the shaft system state - airbag pressure is constructed as follows: K(P)X g = F ext Among them, X g = [x g , y g , z g , β, γ, α] is the translation and rotation motion matrix of the bearing center of gravity of the airbag vibration isolation system in three directions of the main coordinate system. y g represents the axial y offset, and x g represents the offset in the x direction, and z g represents the offset in the z direction; F ext is the external disturbance; K(P) is the system stiffness matrix, and the system stiffness matrix is related to the airbag pressure P distribution P = {p1,..., p N}, where p i represents the pressure of the i-th airbag, N represents the number of airbags in the system, and the system stiffness matrix K(P) is expressed as: Among them, G i is the position matrix, indicating the position of the i-th airbag vibration isolator in the main coordinate system. The position matrix G i is expressed as: Γ i represents the transposed matrix, denoted as: Among them, the element represents the cosine value of the angle between the main elastic axis of the i-th airbag isolator and the main coordinate system of the system, a = T, Q, R, b = x, y, z; k(p i ) represents the stiffness. The natural stiffnesses of the i-th airbag in the radial direction T, the axial direction Q, and the vertical direction R are respectively The pressure correction term is proportional to the current airbag pressure p i and the pressure correction coefficients in the three directions are η p , η q , η r . Then: When the pressure of the i-th airbag changes, only the change in bearing capacity in the vertical direction is considered, and the change amount is Δf i r = Δp i A e , where A e represents the effective area of the airbag and is a constant, and Δp i is the pressure change amount of the i-th airbag; for the pressure state p0 and the centering state corresponding to the pressure state p0 under known initial conditions calculate the centering state under the target pressure state That is, the mathematical model for centering state evaluation is as follows: where G c represents a position transformation matrix, expressed as: where represent the three components of the vector from the origin of the main coordinate system to the center of the output axis system on the main coordinate system.

3. The optimization method for the shaft alignment state evaluation and prediction model of the physical-data collaborative airbag vibration isolation system according to claim 1, characterized in that: In step 1), the physical residual neural network is a backpropagation neural network, including 1 input layer, 1 output layer and H r hidden layers. An activation layer is added between the output layer and the hidden layer, and between the hidden layers. The ReLU function is used as the activation function. The data input by the input layer is the airbag pressure value and the working condition data, and the data output by the output layer is the difference between the calculated centering state value obtained by the centering state evaluation mathematical model and the measured centering state data.

4. The optimization method for the shaft alignment state evaluation and prediction model of the physical-data collaborative airbag vibration isolation system according to claim 1, wherein: The centering state evaluation neural network in step 1) is a backpropagation neural network, including 1 input layer, 1 output layer and H e hidden layers, and activation layers are added between the output layer and the hidden layer, and between the hidden layers. The ReLU function is used as the activation function; the data input by the input layer is the airbag pressure value and the working condition data, and the data output by the output layer is the centering state.

5. The optimization method for the shaft alignment state evaluation and prediction model of the physical-data collaborative airbag vibration isolation system according to claim 1, wherein: In step 2), the airbag pressure value P of the airbag pressure sensor is collected, where P = {p1, p2, p i ,..., p N}, and p i represents the airbag pressure value of the i-th airbag isolator; the working condition data C = {c1, c2, c i ,..., c d} is collected, and c i represents the i-th type of working condition data; the measured data of the shaft alignment state is collected Each set of collected data forms a record and is added to the data set D.

6. The optimization method for the shaft alignment state evaluation and prediction model of the physical-data collaborative airbag vibration isolation system according to claim 5, wherein: In step 3), traverse the dataset D, and for each record According to the airbag pressure value p in the dataset D, use the centering state evaluation mathematical model to calculate the centering state calculated value X corresponding to the airbag pressure value P g ', and generate a new data record R c =(P, C, ΔX g ), where Add the new data record to the physical residual network dataset D c .

7. The optimization method for the shaft alignment state evaluation and prediction model of the physical-data collaborative airbag vibration isolation system according to claim 6, characterized in that: In the step 4), the physical residual network dataset D c is randomly divided into a training set and a test set, and the training error is denoted as L c , and the root mean square is used as the evaluation criterion for the error L c ; When training reaches then stop training, which is the training target threshold; After training, a physical residual prediction model for the alignment status is obtained, denoted as:

8. The optimization method for the shaft alignment state evaluation and prediction model of the physical-data collaborative airbag vibration isolation system according to claim 7, wherein: The generation process of the alignment state evaluation data set D r in step 5) is specifically as follows: 51) Use the hyper-Latin sampling method to sample within the working pressure range of each airbag vibration isolator to obtain a set of pressure value sampling point combinations P s ={P1,...,P M}; Sample different working conditions to obtain a set of working condition sampling point combinations C s ={C1,...,C K}; 52) Calculate the pressure value sampling point P using the centering state evaluation mathematical model i (1 ≤ i ≤ M) The centering state X corresponding to the pressure value g (i); 53) Calculate the physical residuals at different working condition sampling points C using the physical residual prediction model for the centering state j of the physical residuals 54) Traverse all airbag pressure and working condition sampling points, and add the calculated records to the alignment state evaluation dataset D r ; 55) Add all the records in dataset D to dataset D r therein.

9. The optimization method for the shaft alignment state evaluation and prediction model of the physical-data collaborative airbag vibration isolation system according to claim 1, wherein: In step 6), the dataset D r is randomly divided into a training set and a test set. The training error is denoted as L, and the root mean square is used as the error evaluation criterion; When the training reaches L ≥ L T , the training is stopped, where L T is the training target threshold value. After training, an evaluation prediction model for the alignment status is obtained, denoted as: X g = h(P, C) If after N rounds of training, L < L T , then jump to step 2).

Images