Disclosure of Invention
In order to solve the problems in the prior art and overcome the defects of poor assembly coaxiality and poor assembly precision of the conventional low-pressure rotor, the invention provides an assembly process-oriented low-pressure rotor coaxiality prediction method, which mainly comprises the steps of establishing an elastic interaction model under the action of multiple bolts and realizing assembly process-oriented coaxiality prediction.
(1) And establishing an elastic interaction model under the action of multiple bolts. Due to the elastic interaction among the bolts, the bolts in front of the tightening process are influenced by the bolts behind, so that the actual pretightening force is far away from the expected pretightening force, and finally, the relative spatial poses of all the parts of the rotor are deflected due to the uneven distribution of the pretightening force of the bolt groups, so that the coaxiality error of the rotor is caused. The invention firstly simulates the real tightening process (realized by using a finite element method in the embodiment), establishes an elastic interaction model by fitting the pretightening force change value and realizes the dynamic calculation of the pretightening force in the assembling process.
(2) The coaxiality facing the assembly process is predicted. Aiming at the complex coupling and time sequence action relationship among the unbalance amount, the pretightening force and other assembly factors of the low-pressure rotor, the low-pressure fan rotor coaxiality prediction model based on the GRU network is constructed on the basis that the elastic interaction model realizes the dynamic calculation of the pretightening force and on the basis that the Gated Recovery Unit (GRU) processes the advantages of time sequence correlation, and the side-mounting and side-mounting prediction of the low-pressure rotor coaxiality is realized in the assembly process.
The technical scheme of the invention is as follows:
the low-pressure rotor coaxiality prediction method for the assembly process comprises the following steps:
step1: in the low-voltage rotor assembling process, carrying out numerical simulation on a multi-bolt tightening process under a specific assembling process, establishing an elastic interaction model by fitting a pretightening force variation value, and resolving to obtain an elastic interaction matrix [ A ] under the current assembling process;
step2: establishing a low-voltage rotor coaxiality prediction model based on a GRU network, and training the model to obtain a trained coaxiality prediction model based on the GRU network;
the input during the coaxiality prediction model training comprises the following input steps: time step T of the assembly process step The rotor assembly influencing factor F in the time step and the coaxiality low-pressure rotor Coa in the time step;
the rotor assembly influencing factor F comprises an assembly process F gy And assembly process F gy Elastic interaction matrix [ A ] obtained by step1]Converting the bolt pretightening force into a bolt pretightening force in the assembling process, and inputting the bolt pretightening force into a coaxiality prediction model;
and 3, step3: in the actual assembly process, rotor assembly influence factors are used as input and input into the coaxiality prediction model trained in the step2, and the low-pressure rotor coaxiality prediction method in the assembly process is achieved.
Further, in step1, the elastic interaction model is:
Prlod=A(T lj ,T sx ,T lc ,T qd )
wherein Prlod is the pretightening force of the bolt group; a is an elastic interaction matrix; t is lj Tightening torque for the bolt; t is sx The bolt tightening sequence is adopted; t is bs The bolt tightening turns are carried out; t is qd Is the starting point for bolt tightening.
Still further, in step1, the bolt tightening torque T is determined lj Bolt tightening sequence T sx Bolt tightening round T bs Bolt tightening starting point T qd Simulating the bolt tightening process in the assembly process, and according to the simulation result, utilizing a formula
Solving to obtain an elastic interaction matrix [ A ] under the current assembly process](ii) a Wherein a is ij Representing the elastic interaction matrix coefficient, f ij Representing the residual pretightening force of the ith bolt after the bolt j is screwed; Δ f j Indicating the variation of the pretension force when the jth bolt is tightened.
Further, in step2, the assembly process F gy The method comprises the steps of bolt tightening torque, tightening sequence, tightening round, tightening starting point and disk drum installation phase; wherein bolt tightening torqueTightening sequence, tightening round, tightening start through the elastic interaction matrix [ A ] obtained in step1]And converting the bolt pretightening force into the bolt pretightening force in the assembling process, and inputting the bolt pretightening force and the disc drum installation phase into a coaxiality prediction model.
Further, the bolt tightening torque, the tightening sequence, the tightening turns and the tightening starting point are utilized according to an elastic interaction formula
{F 0 }+[A]{ΔF}={F f }
After each loading step, pre-tightening each bolt; wherein { F 0 The bolt pretightening force (nx1) before the loading step is adopted; [ A ]]Is an elastic interaction matrix (n × n) corresponding to the current assembly process; { delta F } represents the increment of the bolt pretightening force (n multiplied by 1) after the loading step; { F f The bolt pretightening force (nx1) is obtained after the loading step; n is the number of bolts in the bolt group; one loading step for each bolt tightening.
Furthermore, the influencing factors in the rotor assembly process also comprise the processing quality F jg Material property F d And assembly accuracy F jd And an environmental parameter F hj 。
Still further, the process quality F jg The vector form expression is carried out as follows:
F jg =(Q pm ,Q td ,Q cc ,Q zx ,Q y ,Q cz ,Q tz )
in the formula Q pm Matching flatness of upper and lower rabbets of the disc drum; q td The disk drum is matched with the jumping degree of the seam allowance relative to the rotating shaft; q cc The surface roughness of the matching part of the disc drum seam allowance; q zx Is the straightness of the central axis of the disc drum; q y Roundness of the outer circle and the inner circle of the disk drum; q cz The perpendicularity between the end surface of the disk drum and the axis position is obtained; q tz The coaxiality of the disc drum is self;
said material property F d The vector form expression is carried out as follows:
F cl =(E,τ,ρ,μ)
wherein E is Young's modulus; τ is the poisson's ratio; rho is density; mu is a friction coefficient;
the above-mentionedAssembly accuracy F jd The vector form expression is carried out as follows:
F jd =(Paral,Intf)
in the formula, paral is the assembly parallelism of the upper end surface and the lower end surface of the disc drum seam allowance; intf is the interference of assembling the disc drum spigot;
the environmental parameter F hj The vector form expression is carried out as follows:
F hj =(T,RH)
wherein T is the ambient temperature during rotor assembly; and RH is the ambient humidity when the rotor is assembled.
And further, the influence factor data in the rotor assembling process is subjected to standardization processing by adopting a z-score standardization method, so that all the influence factor data are in the same order of magnitude.
The invention also provides a computer device comprising at least one connected processor, a memory and a transceiver; wherein the memory is configured to store program code and the processor is configured to call the program code in the memory to perform the steps of the assembly process oriented low pressure rotor concentricity prediction method.
The present invention also provides a computer storage medium, comprising: instructions which, when executed on a computer, cause the computer to perform the steps of the assembly process oriented low pressure rotor coaxiality prediction method.
Advantageous effects
The invention comprehensively considers the assembly factors influencing the change of the coaxiality, and proposes to establish an elastic interaction matrix by using a finite element simulation mode to realize the dynamic calculation of the bolt pretightening force aiming at the dynamic variability and the complex coupling action relation of the bolt pretightening force and other assembly factors; in addition, the regularity of the assembly factor change and the coaxiality change is considered, and the association relation between the pretightening force, other assembly factors and the coaxiality change is obtained by means of the learning and training with the advantages of the GRU network in the aspect of solving the related problems of the time sequence in a targeted mode, so that the applicability and the accuracy of the prediction model can be greatly improved.
Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.
Detailed Description
Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the accompanying drawings are illustrative and intended to explain the present invention and should not be construed as limiting the present invention.
In order to overcome the defects of the conventional method in the aspect of solving the coaxiality prediction problem during assembly and considering that bolt connection is the most representative connection mode in aircraft engine assembly, the embodiment provides the low-pressure rotor coaxiality prediction method for the assembly process, which is used for predicting the coaxiality of the assembled low-pressure rotor. The method mainly comprises three aspects: firstly, uniformly expressing the problem of influence factor diversity in the rotor assembly process; secondly, establishing an elastic interaction model aiming at the variation of pretightening force in the bolt tightening process; and then, aiming at the time sequence correlation problem between the pretightening force and the coaxiality, a coaxiality prediction model based on a GRU network is constructed, and the accurate prediction of the coaxiality is further realized by utilizing the long-term memory capacity of the model. The prediction implementation process is shown in fig. 1, and specifically includes the following processes:
1. acquiring influence factors in the rotor assembling process, and uniformly expressing the influence factors:
the influence factors influencing the coaxiality in the rotor assembling process are numerous and can be divided into five categories, and the five categories of factors are uniformly expressed as shown in a formula 1.
F=(F gy ,F jg ,F cl ,F jd ,F hj ) (1)
In the formula: f is an influencing factor; f gy Is an assembly process; f jg The quality of the processing; f d Is a material property; f jd The assembly precision is achieved; f hj Is an environmental parameter.
(1) Assembly process
Different assembly process parameters in the rotor assembly process can cause the change of bolt pretightening force and further cause the change of coaxiality, main process parameters in the process comprise bolt tightening torque, a tightening sequence, tightening turns, a tightening starting point and an installation phase, and the five assembly process parameters are expressed in a vector form and then are shown as a formula 2:
F gy =(T lj ,T sx ,T lc ,T qd ,P az ) (2)
in the formula: f gy Is an assembly process; t is a unit of lj The bolt tightening torque is set; t is sx The bolt tightening sequence is adopted; t is bs The bolt tightening turns are carried out; t is qd Is a starting point for screwing the bolt; p az And installing the phase for the disk drum.
(2) Quality of processing
The initial coaxiality is different due to different processing qualities of the assembled two-stage disc drum, so that the assembled coaxiality is changed. These process qualities mainly include: flatness, run-out, roughness, straightness, roundness, perpendicularity, and coaxiality, and the coaxiality described herein refers to the coaxiality of the single-stage disc drum itself, and does not refer to the overall coaxiality after assembly. The seven processing qualities are expressed in a vector form and then are shown as a formula 3:
F jg =(Q pm ,Q td ,Q cc ,Q zx ,Q y ,Q cz ,Q tz ) (3)
in the formula: f jg The quality of the processing is shown; q pm Matching flatness of upper and lower rabbets of the disc drum; q td For disc drums to cooperate with spigotsRun out with respect to the axis of rotation; q cc The surface roughness of the matching part of the disc drum seam allowance; q zx Is the straightness of the central axis of the disc drum; q y The roundness of the outer circle and the inner circle of the disc drum; q cz The disc drum end surface is perpendicular to the axis position; q tz Is the coaxiality of the disc drum.
(3) Material properties
The material property parameters of the disc drum can influence the friction force and the deformation degree, and further influence the change of the coaxiality. The material properties are mainly as follows: young's modulus, poisson's ratio, density, coefficient of friction. The four material properties are expressed in a vector form and then are shown in formula 4:
F cl =(E,τ,ρ,μ) (4)
in the formula: f cl Is a material property; e is Young's modulus; tau is the Poisson's ratio; rho is density; μ is the coefficient of friction.
(4) Assembly accuracy
When the parallelism error exists between the upper and lower disc drums, due to the matching of the seam allowances, the deflection of the upper and lower disc drums is caused by uneven pretightening force after the two contact surfaces are attached, and the coaxiality of the assembled rotor is influenced finally. The assembly precision parameters comprise: parallelism and interference. The two assembling accuracies are expressed in a vector form and then are shown as a formula 5:
F jd =(Paral,Intf) (5)
in the formula: f jd The assembly precision is achieved; paral is the assembly parallelism of the upper end face and the lower end face of the disc drum seam allowance; intf is the interference of the assembly of the disk drum seam allowance.
(5) Environmental parameter
The environmental parameters during assembly also affect the assembly coaxiality to some extent. The main environmental parameters include: temperature, humidity. The two main parameters are expressed in a vector form and then are shown in formula 6:
F hj =(T,RH) (6)
in the formula: f hj Is an environmental parameter; t is the ambient temperature when the rotor is assembled; during rotor assembly with RHThe ambient humidity.
Data preprocessing is carried out on influencing factors in rotor assembly process
The coaxiality influence factors comprise data such as the assembly process, the machining quality and the material attribute of the part, the data have different dimensions and orders of magnitude, and when the level difference between the data is large, the accuracy and the reliability of a subsequently established prediction model can be influenced, so that the original data needs to be standardized, and all the influence factor data are in the same order of magnitude.
The coaxiality influencing factors have design requirements in rotor design and process planning, so that the coaxiality influencing factors fluctuate within a theoretical range, and the distribution of the coaxiality influencing factors is in a Gaussian distribution manner, so that the assembly data can be subjected to standard processing by adopting a z-score standardization method, wherein the method comprises the following steps:
in the formula: x' is normalized data; mu is the mean value of the influence factor sequence; σ is the standard deviation of the sequence of influencers.
2. Constructing elastic interaction models
During the assembly process of the low-pressure fan rotor, elastic interaction exists between the bolt sets. The tightening torque, the tightening sequence, the number of tightening rounds and the tightening starting point in the assembly process are inconvenient to input in the model, so the invention uniformly converts the assembly processes into the change of the bolt pretightening force by establishing an elastic interaction model, and the process can be expressed as formula 8. According to the invention, a three-dimensional model with the same structure as the rotor is established, the variation condition of the pretightening force of the bolt group under a specific assembly process is simulated by using a finite element method, the obtained data is processed and calculated to obtain an elastic interaction matrix, and the elastic interaction model is established.
Prlod=A(T lj ,T sx ,T lc ,T qd ) (8)
In the formula: prlod is the pretightening force of the bolt group; a is an elastic phaseAn interaction matrix; t is lj The bolt tightening torque is set; t is sx The bolt tightening sequence is adopted; t is a unit of bs The bolt tightening turns are carried out; t is qd Is the starting point for bolt tightening.
In the embodiment, a simulation experiment is performed on the bolt tightening process in ABAQUS software, and an elastic interaction matrix is established according to the residual pretightening force of the bolt. In the embodiment, an elastic interaction matrix is established by taking 8-bolt connection and a cross-tightening low-pressure rotor as an example, and the bolt pretightening force is 2500N during simulation calculation.
The elastic interaction of the bolt set when loaded is shown as formula 9:
{F 0 }+[A]{ΔF}={F f } (9)
in the formula: { F 0 The pretightening force (n multiplied by 1) of the bolt before each loading step is used; [ A ]]Is an elastic interaction matrix (n × n); { delta F } represents the bolt pretightening force increment (n multiplied by 1) of each loading step; { F f The residual pretightening force (n multiplied by 1) of the bolt after each loading step is adopted; and n is the number of bolts in the bolt group.
Each vector in the above equation is represented as follows:
in the formula: f. of i 0 ,Δf i And f i f The residual pretension at the end of the previous loading, the pretension increment of the bolt i and the final residual pretension after the end of the loading are respectively represented.
The elastic interaction matrix [ A ] is represented by formula 11:
in the formula: a is ij Representing the elastic interaction matrix coefficient, f ij Representing the residual pretightening force of the ith bolt after the bolt j is screwed; Δ f j Indicating the variation of the pretension force when the jth bolt is tightened.
The variation of bolt pretension in a cross loading manner simulated in finite element software is shown in table 1:
TABLE 1 residual Pretightening force (N) of bolts at different loading steps
The elastic interaction matrix [ A ] can be calculated using equation 12 as follows:
3. constructing coaxiality prediction model
Constructing a coaxiality prediction model y based on the content of the steps tz :
Time step T using loading step contained in assembly process as model step In the assembly process F within the time step T gy Quality of processing F zp Material property F cl And assembly accuracy F jd Environmental parameter F hj And as an input, outputting the coaxiality Coa in the time step as a target, inputting the target output into the established prediction model, and training to obtain the coaxiality prediction model. The establishment of the coaxiality prediction model is shown as formula 13:
in the formula: y is
tz Assembling a coaxiality prediction model for the rotor;
training parameters representing a predictive model; t is
step Inputting a time step; coa is the coaxiality in the time step; and theta is a parameter set obtained by model training.
In this embodiment, the calculation of the change in the pre-tightening force is performed with 8 bolts for connection and an actual bolt tightening torque of 1.5N · m. The tightening torque is first converted to a pretension by equation 14.
T=0.2dF f (14)
In the formula, T is a tightening torque; f f Is pre-tightening force; d is the outer diameter of the bolt, M4 bolt is adopted in the text, and the outer diameter is 4mm, then F f =1875N。
The actual pretightening force in the tightening process can be obtained by performing corresponding processing calculation on the bolt pretightening force change by using the formula 9 and the elastic interaction matrix established in the front edge, wherein the calculation process is as follows:
step1: while tightening the first bolt, F 0 =[0,0,0,0,0,0,0,0] T ,ΔF=[1875,0,0,0,0,0,0,0] T Then, after tightening the first bolt, the pretension of the other bolts can be calculated by equation 9, as shown in the first column of table 2.
Step2: while tightening the second bolt, F 0 The pre-tightening force for tightening the first bolt is F 0 =[1875,78.8,97.5,97.5,0,91.9,91.9,0] T ,ΔF=[0,1875,0,0,0,0,0,0] T The pretension of the other bolts can be calculated by equation 9, as shown in the second column of table 2.
Step3: by analogy, the actual pretightening force of the bolts after all the bolts are screwed can be calculated. As shown in table 2:
TABLE 2 actual Pretightening force (N) of bolt during assembly
Training a coaxiality prediction model:
the whole assembling process of the low-pressure fan rotor is divided into n loading steps according to the number of the bolts, and each loading step comprises assembling factors, bolt group pre-tightening force values and coaxiality values in the period of time. The network training mainly comprises a forward computing part and a reverse tuning part of the prediction model.
(1) A prediction model forward calculation part: the step mainly takes the assembling influence factors, the bolt group pretightening force and the coaxiality value in the time period as input, then carries out calculation according to a GRU network forward calculation formula, and outputs the predicted value of the coaxiality of the rotor at the next moment after forward calculation through the GRU network.
The forward calculation formula of the GRU network is as follows:
r t =σ(W r [S t-1 ,x t ]) (15)
z t =σ(W z [S t-1 ,x t ]) (16)
y t =σ(W o S t ) (19)
wherein, W r Represents the weight of the reset gate, W z Represents the weight of the update gate, W h Representing hidden layer information S at time t-1 t-1 As a weight matrix at input, W o Represents the output layer weight, y t And the predicted value of the GRU network at the time t is shown.
The forward calculation formula shows W r 、W z 、W h 、W o The weight parameters are weight parameters required by training and learning of the GRU network, and the first three parameters can be divided into the following parameters according to the components:
W r =[W rx ,W rs ]W z =[W zx ,W zs ]W h =[W hx ,W hs ] (20)
suppose the actual value of the sample data at time t is y' t The output value at time t obtained by using the GRU network is y t Then the transmission loss amount (root mean square error MSE as loss function) of the network at time t is:
therefore, the amount of transmission loss of the network over the entire time period T is:
then, the following intermediate error term in the network at time t can be obtained:
δ y,t =(y' t -y t )·σ' (23)
δ S,t =δ y,t W o +δ z,t+1 W zS +δ t+1 W hS ·r t+1 +δ S,t+1 W rs +δ· S,t+1 ·(1-z t+1 ) (24)
δ t =δ S,t ·z t ·tanh' (26)
δ r,t =S t-1 ·((δ S,t ·z t ·tanh')W hS )·σ' (27)
since the derivative of the loss amount to the weight matrix is the gradient of the weight matrix, the sum of the gradients at each time is the final gradient in the total duration. Therefore, each weight parameter matrix W in the time period T o 、W zx 、W zs 、W hx 、W hs 、W rx 、W rs The gradient of (a) is calculated as:
finally, iterative updating of each parameter of the GRU network model can be achieved by using the formula and adopting a proper learning rate eta.
(2) A prediction model reverse optimization part: the part is mainly based on the error back propagation principle, the weight and the bias of each neuron in the GRU network are subjected to iterative updating, and then the accuracy of a prediction model is improved. The present embodiment treats the mean square error as a loss function of this part and then optimizes the gradient descent using the Adam algorithm.
The Adam method calculates a coefficient tau with dynamic change by using the moment of the current gradient, and then realizes the dynamic adjustment of the learning rate. τ is calculated as follows:
m t =p×m t-1 +(1-p)×g t (30)
n t =q×n t-1 +(1-q)×g t 2 (31)
wherein, g t The current gradient value of the parameter; m is t 、n t Each represents g t First order moment estimation and second order moment estimation; p and q are real numbers between 0 and 1; c is a non-zero constant.
We performed the coaxiality prediction model test as follows:
and randomly selecting 12 groups of experimental samples, and testing by using a coaxiality prediction model trained by a 4-fold cross-validation method, wherein the test result is shown in the attached figure 2.
In order to better evaluate the prediction effect of the coaxiality prediction model, the nash efficiency coefficient (NSEC) values of 12 groups of test samples are calculated, and the calculation formula is shown as formula 32. In general, the smaller the error between the predicted and true coaxiality values, the closer the NSEC value is to 1.
The NSEC calculation results for the 12 test samples are shown in table 3 below.
TABLE 3 comparison of predicted values and true values of test samples
As can be seen from fig. 2, the coaxiality prediction model test results in the 12 groups of test samples show that the difference between the coaxiality prediction value and the true value is not large, and the prediction accuracy is relatively good. From the NSEC values of the 12 test samples shown in table 3, it can be seen that the prediction model can realize the low-pressure fan rotor coaxiality prediction under the multi-bolt tightening effect.
In conclusion, the predicted value and the actual value of the coaxiality of the tightened low-pressure rotor have the same variation trend, the predicted result is relatively accurate, and the coaxiality prediction method for the assembling process is feasible and effective.
Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made in the above embodiments by those of ordinary skill in the art without departing from the principle and spirit of the present invention.