Aviation Engine Assembly Process Optimization Method Based on Assembly Data

Abstract

The present invention discloses an optimization method for an aero-engine assembly process based on assembly data, comprising: when the aero-engine passes the test run, collecting the assembly data of each unit in the aero-engine, and screening the assembly indicators of the assembly data of the unit; calculating the correlation between the screened assembly indicators of the unit, and constructing a unit assembly network model according to the complex network theory; introducing the network dynamics theory, fitting the unit assembly network model with a neural network dynamics model to obtain a unit dynamics model; determining the fitting error of each assembly indicator based on the solution of the unit dynamics model; constructing a unit evaluation interval by using the mean field theory based on the fitting error of the assembly indicator; for the unit to be evaluated, determining the evaluation indicator of the unit by using the assembly indicator of the unit; and determining whether the unit to be evaluated is qualified based on the relationship between the evaluation indicator and the unit evaluation interval.

Description

Technical Field

[0001] The present invention relates to the field of optimization of aero-engine assembly processes, and particularly to a method for optimizing aero-engine assembly processes based on assembly data. Background Art

[0002] Aero-engines are characterized by high technology concentration and content, complex system integration, high added value, and strong driving force. As a complex high-precision thermal mechanical system, aero-engines have numerous components and complex assembly processes, and the assembly process of aero-engines can directly affect their performance, reliability, and service life. By linking assembly and commissioning, the assembly efficiency and quality can be improved, and resource allocation can be further optimized.

[0003] The assembly process and commissioning process of aero-engines are two crucial processes in the manufacturing process. The assembly of aero-engines directly affects their performance, reliability, and service life, and the assembled aero-engines must undergo strict commissioning tests before they can be put into actual use. However, in the manufacturing process, there are often some aero-engines whose manufacturing dimensions of all components meet the design requirements, and their corresponding assembly work is carried out strictly in accordance with the regulations, but they cannot pass the commissioning procedure, and the whole engine has to be disassembled, reassembled, retested until the commissioning requirements are met, which ultimately leads to an increase in the production cost of the enterprise.

[0004] In recent years, aero-engine manufacturers have accumulated a large amount of aero-engine assembly data and commissioning results. In theory, the assembly and commissioning of aero-engines should be a complementary process; however, in actual engineering applications, the assembly data and commissioning results are still in an artificially "divided" state, resulting in a certain degree of blindness in the current optimization of aero-engine assembly processes. Secondly, the assembly data includes the assembly process parameters of each component, the detection data before and after assembly, the error records during the assembly process, and various correction operations. This data is high-dimensional and heterogeneous, and the correlation between indicators is strong, making it difficult to accurately guide the optimization of the assembly process, improve the assembly efficiency and quality. Finally, aero-engines mostly adopt a unitized structural design in actual engineering to achieve the interchangeability of units and improve the maintainability of aero-engines. However, the existing technology is still unable to effectively evaluate at the unit level, and it is necessary to wait until the whole engine is assembled and then commissioned, and then judge whether the whole engine needs to be disassembled or delivered according to the commissioning results, which greatly limits the assembly efficiency and quality of aero-engines. Therefore, aiming at the above defects, studying how to use assembly data and commissioning data to improve the enterprise's assembly efficiency and commissioning pass rate in actual engineering is an urgent problem to be solved in the field of aero-engine manufacturing.

[0005] The prior art discloses an aero-engine performance digital twin model integrating assembly data and a building method thereof (patent application with publication number CN118378141A): First, collect the assembly test data of multiple aero-engines of the same model under different test procedures, and determine the sensor monitoring data and assembly data related to the target performance parameters; Second, construct a performance digital twin model integrating assembly data, and train the performance digital twin model using the sensor monitoring data and assembly data; Finally, calculate the prediction accuracy of the trained performance digital twin model. If the prediction accuracy meets the requirements, end; If the prediction accuracy does not meet the requirements, retrain the performance digital twin model. However, the deficiencies of this method are as follows: This method uses the multi-head attention mechanism to learn the component weight coefficients and establishes a performance digital twin model through a recurrent neural network, which means that the result generation mechanism of the model cannot be clarified; In addition, this method can only predict the performance parameters of aero-engine tests and cannot provide an effective guidance plan for the actual assembly of aero-engines. Summary of the Invention

[0006] The object of the present invention is to provide an aero-engine assembly process optimization method based on assembly data, which fits the unit dynamics model based on the neural network dynamics model to evaluate the unit state, so as to provide an actual operation plan for aero-engine assembly and improve the aero-engine assembly efficiency and test passing rate.

[0007] To achieve the above tasks, the present invention adopts the following technical solutions:

[0008] The aero-engine assembly process optimization method based on assembly data includes:

[0009] When the aero-engine passes the test, collect the assembly data of each unit in the aero-engine, and screen the assembly indicators of the unit assembly data;

[0010] Calculate the correlation between the screened assembly indicators of the unit, and construct a unit assembly network model based on the complex network theory;

[0011] Introduce the network dynamics theory, fit the unit assembly network model with a neural network dynamics model to obtain a unit dynamics model; Based on the solution of the unit dynamics model, determine the fitting error of each assembly indicator;

[0012] Based on the fitting error of the assembly indicator, use the mean field theory to construct a unit evaluation interval; For the unit to be evaluated, use the assembly indicators of the unit to determine the evaluation indicators of the unit; Based on the relationship between the evaluation indicators and the unit evaluation interval, determine whether the unit to be evaluated is qualified.

[0013] Furthermore, the method further includes:

[0014] Calculate the evaluation index of the unit body by using the assembly index after screening of the unit body when the aero-engine test run passes or fails, divide the state interval of the evaluation index of the unit body, and construct a Bayesian network model with the test run index of the aero-engine as the target variable and the state interval as the attribute variable; predict the posterior probability when the test run index of the to-be-evaluated aero-engine is qualified based on the Bayesian network model.

[0015] Furthermore, the assembly data of each unit body consists of multiple assembly indexes; perform an analysis of variance on all the assembly indexes in the assembly data of each unit body, and eliminate the assembly indexes with zero variance based on statistical principles, so as to obtain the assembly indexes after screening of each unit body.

[0016] Furthermore, calculate the correlation between the assembly indexes after screening of the unit body, and construct a unit body assembly network model according to the complex network theory, including:

[0017] Perform networked modeling based on the assembly data of the unit body, abstract the assembly indexes after screening in the unit body assembly data into nodes in the network, determine the mutual relationship between the assembly indexes based on the correlation between the assembly indexes, and abstract the mutual relationship into edges in the network to construct a unit body assembly network model; construct an adjacency matrix by using the mutual relationship, and determine the positive adjacency matrix and the negative adjacency matrix according to the magnitude of the correlation; the correlation is calculated by using the Pearson correlation coefficient formula.

[0018] Furthermore, the unit body dynamics model is expressed as:

[0019] ;

[0020] where t is the time parameter, and represent the i-th and j-th assembly indexes, is the assembly index at the initial state; is the reciprocal of the self-change rate of the assembly index , is the number of assembly indexes in the unit body assembly network model, a is the trigger threshold, n is the slope of the activation function, and are respectively the elements in the i-th row and j-th column of the positive adjacency matrix and the negative adjacency matrix , and are respectively the excitation intensity and the inhibition intensity of the assembly index .

[0021] Furthermore, the specific determination formulas for the excitation intensity and the inhibition intensity are:

[0022] , ;

[0023] Among them, and are respectively the positive correlation coefficient matrix and the negative correlation coefficient matrix of the element at the i-th row and j-th column in

[0024] Furthermore, based on the solution of the unit dynamics model, the fitting errors of each assembly index are determined, including:

[0025] Let the value of the unit dynamics model be zero; the calculated value of the assembly index obtained by solving the unit dynamics model is used as the simulated value; the difference between the simulated value and the actual value of the assembly index is calculated as the fitting error of the assembly index.

[0026] Furthermore, based on the fitting errors of the assembly indexes, the unit evaluation interval is constructed by using the mean field theory, including:

[0027] The evaluation index of the unit is obtained by calculating the expectation of all assembly indexes in the unit assembly network model:

[0028] ;

[0029] Among them, is the number of assembly indexes in the unit assembly network model, is the -th assembly index in the unit assembly network model; the simulated value of the -th assembly index is used as to substitute into the above formula, and based on the fitting error of the -th assembly index, the lower bound and upper bound of the unit evaluation interval are constructed by the following formula:

[0030] ;

[0031] Among them, is the middle value of the unit evaluation interval, and are respectively the lower bound and upper bound of the unit evaluation interval, then the unit evaluation interval is .

[0032] Further, using the actual values of the assembly indicators after screening when the aero-engine test run is passed and not passed, calculate the evaluation indicators of the unit body when the test run is passed and not passed, and count the range of all evaluation indicators; use the elbow method to determine the optimal number of clusters of the evaluation indicators, use the K-Means clustering algorithm to cluster the evaluation indicators, and divide the range of the evaluation indicators into state intervals; assign a unique code to each divided state interval, so as to obtain the state interval of the evaluation indicator and its corresponding code.

[0033] Calculate the combinations of each attribute variable when the target variable is qualified through the Bayesian network model, screen out the combinations of attribute variables greater than the probability threshold set by the enterprise, and use the codes of each attribute variable in the combination to construct an assembly suggestion table for providing assembly suggestions for the enterprise to assemble aero-engines.

[0034] An aero-engine assembly process optimization device includes a processor, a memory, and a computer program stored in the memory; when the processor is executed by a computer, the aero-engine assembly process optimization method based on assembly data is implemented.

[0035] A computer-readable storage medium stores a computer program; when the computer program is executed by a processor, the aero-engine assembly process optimization method based on assembly data is implemented.

[0036] Compared with the prior art, the present invention has the following technical characteristics:

[0037] 1. The aero-engine assembly process optimization method based on assembly data constructed by the present invention can effectively break the "barriers" existing in the assembly process and the test run process in engineering applications, can predict the qualified probability of the aero-engine test run indicators, and provide an operable assembly suggestion table for the assembly process.

[0038] 2. The evaluation indicators and their evaluation intervals of the fan unit body, compressor unit body, and turbine unit body constructed by the present invention can effectively evaluate the unit body, thereby effectively improving the assembly quality and efficiency, and greatly reducing the test run cost of the enterprise.

[0039] 3. The assembly network model and dynamic model of each unit body of the aero-engine proposed by the present invention can accurately depict the mutual relationship between each assembly indicator, and explain the relationship between the assembly indicator and the change of the system state, and has interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a schematic flow chart of the method of the present invention;

[0041] Figure 2 is a schematic structural diagram of each unit body of the aero-engine;

[0042] Figure 3a This is the fan assembly network model in the embodiment of the present invention;

[0043] Figure 3b This is the compressor assembly network model in the embodiment of the present invention;

[0044] Figure 3c This is the turbine assembly network model in the embodiment of the present invention;

[0045] Figure 4a This is the dynamic fitting error graph of the fan in the embodiment of the present invention;

[0046] Figure 4b This is the dynamic fitting error graph of the compressor in the embodiment of the present invention;

[0047] Figure 4c This is the dynamic fitting error graph of the turbine in the embodiment of the present invention;

[0048] Figure 5a This is the schematic diagram of the evaluation index and evaluation interval verification of the fan in the embodiment of the present invention;

[0049] Figure 5b This is the schematic diagram of the evaluation index and evaluation interval verification of the compressor in the embodiment of the present invention;

[0050] Figure 5c This is the schematic diagram of the evaluation index and evaluation interval verification of the turbine in the embodiment of the present invention;

[0051] Figure 6a This is the evaluation index clustering performance graph of the fan in the embodiment of the present invention;

[0052] Figure 6b This is the evaluation index clustering performance graph of the compressor in the embodiment of the present invention;

[0053] Figure 6c This is the evaluation index clustering performance graph of the turbine in the embodiment of the present invention;

[0054] Figure 7 This is the schematic diagram of the Bayesian network model;

[0055] Figure 8 This is the ROC curve graph used to characterize the accuracy rate of the Bayesian network model. Detailed implementation manners

[0056] In order to elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description in conjunction with embodiments.

[0057] For the assembly data and test run data of aero-engines, the present invention combines the advantages of complex network and Bayesian network theories, and proposes an aero-engine assembly process optimization method based on assembly data, which can establish a bridge between the assembly indexes and test run indexes of aero-engines, construct an evaluation standard for the qualification diagnosis of each unit of aero-engines, give the qualification probability of predicting the test run indexes of aero-engines, and obtain an aero-engine assembly suggestion list that can guide the actual assembly process.

[0058] See the appendix Figure 1 , the present invention provides an aero-engine assembly process optimization method based on assembly data, comprising the following steps:

[0059] Step 1, when the aero-engine passes the test run, collect the assembly data of each unit in the aero-engine, and screen the assembly indexes from the assembly data of the unit; the units include a fan, a compressor and a turbine.

[0060] Among them, in addition to the above-mentioned fan, compressor and turbine, other components of the aero-engine can also be selected in actual use for the unit.

[0061] The assembly data of each unit consists of multiple assembly indexes; in this solution, variance analysis is performed on all the assembly indexes in the assembly data of each unit, and the assembly indexes with zero variance are excluded based on statistical principles, so as to obtain the screened assembly indexes of each unit, so as to realize the scientific screening of assembly indexes.

[0062] Step 2, calculate the correlation between the screened assembly indexes of the unit, and construct a unit assembly network model according to the complex network theory.

[0063] According to the complex network theory, perform networked modeling based on the assembly data of the unit, abstract the screened assembly indexes in the unit assembly data into nodes in the network, determine the mutual relationship between the assembly indexes based on the correlation between the assembly indexes, and abstract the mutual relationship into edges in the network to construct a unit assembly network model; use the mutual relationship to construct an adjacency matrix, and determine the positive adjacency matrix and negative adjacency matrix according to the magnitude of the correlation.

[0064] (2-1) Determine the mutual relationship and adjacency matrix between the assembly indexes.

[0065] Based on statistical principles, use the Pearson correlation coefficient formula to calculate the correlation between the th screened assembly index of the unit :

[0066] (1);

[0067] Among them, is the covariance corresponding to the assembly index , and is the standard deviation corresponding to the assembly index .

[0068] The 0-1 adjacency matrix is used to characterize the mutual relationship between nodes in the unit assembly network model, which is expressed as follows:

[0069] (2);

[0070] Among them, is the element in the i-th row and j-th column of the 0-1 adjacency matrix , which characterizes the mutual relationship between the assembly indexes ; when the absolute value of the correlation coefficient between the assembly indexes is greater than the preset correlation coefficient threshold , it indicates that there is an edge between the nodes in the assembly network model corresponding to the assembly index ; otherwise, there is no edge; thus, the 0-1 adjacency matrix corresponding to the unit is constructed; in this embodiment, the correlation coefficient threshold is set to 0.25.

[0071] (2-2) Determine the positive adjacency matrix and the negative adjacency matrix.

[0072] Based on the correlation between the assembly indexes , the positive adjacency matrix and the negative adjacency matrix of the unit assembly network model are constructed; among them .

[0073] When is greater than or equal to , the assembly index is positively correlated, and the element in the i-th row and j-th column of the positive adjacency matrix is 1; when is less than or equal to - , the assembly index is negatively correlated, and the element in the i-th row and j-th column of the negative adjacency matrix is 1; specifically, it is expressed as follows:

[0074] (3);

[0075] (4);

[0076] At Figures 3a to 3cThe examples given show the corresponding fan assembly network models, compressor assembly network models, and turbine assembly network models constructed when the unit bodies are fans, compressors, and turbines respectively.

[0077] Step 3: Introduce the network dynamics theory, fit the neural network dynamics model to the unit body assembly network model to obtain the unit body dynamics model; based on the solution of the unit body dynamics model, determine the fitting errors of each assembly index.

[0078] (3-1) Construct the unit body dynamics model.

[0079] In the constructed unit body assembly network model, the positive and negative correlations between assembly indexes are prevalent; a positive correlation indicates that an increase in one assembly index will prompt a synchronous increase in another assembly index; while a negative correlation means that an increase in one assembly index will lead to a decrease in another assembly index.

[0080] To quantitatively describe these dynamic interactions, differential equations can be used to model the variation laws between assembly indexes, thereby simulating their interaction; as a mature modeling method, neural network dynamics can simulate the information transmission process between neurons based on the interconnection structure and connection strength of neurons, and effectively depict its positive and negative correlation interaction mechanism; the original equation form of this dynamics model is as follows:

[0081] (5);

[0082] In the formula, t is the time parameter, is the triggering threshold, is the activation function slope, I is the basal activity, R is the reciprocal of the neuron mortality rate, is the number of neurons in the neural network, and respectively represent the activity values of the th and th neurons in the neural network, and respectively represent the excitation and inhibition intensities, and are the excitatory adjacency matrix and inhibitory adjacency matrix of the neural network respectively, and are respectively and the elements in the i-th row and j-th column of.

[0083] Based on neural network dynamics, introduce the assembly indexes, the number of nodes , the positive adjacency matrix and the negative adjacency matrix in the unit body assembly network model, and at the same time use the correlation representation between assembly indexes and , a unit dynamics model is constructed to quantitatively describe the interaction mechanism among various assembly indexes in the unit assembly network model:

[0084] (6);

[0085] where t is the time parameter, and represent the i-th and j-th assembly indexes, is the initial state of the assembly index , that is, the specification value of the assembly index; is the reciprocal of the self-change rate of the assembly index , is the number of assembly indexes in the unit assembly network model, a is the trigger threshold, generally set to -1; n is the slope of the activation function, generally set to 2; and are the elements in the i-th row and j-th column of the positive adjacency matrix and the negative adjacency matrix respectively, and are the excitation intensity and inhibition intensity of the assembly index respectively, and can be calculated by the following formula:

[0086] (7);

[0087] (8);

[0088] wherein, and are the elements in the i-th row and j-th column of the positive correlation coefficient matrix and the negative correlation coefficient matrix of the unit assembly network model respectively, and can be calculated by the following formula:

[0089] (9);

[0090] (10);

[0091] (3 - 2) Determine the fitting error of each assembly index.

[0092] Let the value of the unit dynamics model be zero (i.e., at the steady state of the system), that is ; The solution value of the assembly index is obtained by solving Equation (6) of the unit dynamics model as the simulation value ; Calculate the simulation value and the assembly index The difference between the actual value is used as the assembly index The fitting error of

[0093] In an example of the present invention, the construction of the fan network dynamics model, the compressor network dynamics model, and the turbine network dynamics model and the calculation of the fitting error of the corresponding assembly index are completed according to the above method, as Figures 4a to 4c shown; from Figures 4a to 4c it can be seen that the difference between the actual value of each assembly index and the simulated value calculated by the network dynamics model is very small, indicating that the unit body dynamics model constructed by this solution can accurately describe each assembly index and the assembly state of the unit body.

[0094] Step 4, based on the fitting error of the assembly index, use the mean field theory to construct the unit body evaluation interval; for the unit body to be evaluated, use the assembly index of the unit body to determine the evaluation index of the unit body; based on the relationship between the evaluation index and the unit body evaluation interval, determine whether the unit body to be evaluated is qualified.

[0095] According to the mean field theory, the system average behavior value can characterize the average state of the system. Therefore, the system average behavior value is used as the evaluation index of the unit body to evaluate its assembly state; specifically, the system average behavior value of the unit body can be obtained by calculating the expectation of all nodes (i.e., assembly indexes) in the unit body assembly network model:

[0096] (11);

[0097] wherein is the evaluation index of the unit body, is the number of assembly indexes in the unit body assembly network model, is the th assembly index in the unit body assembly network model.

[0098] Based on the above principle, substitute the simulated values of each assembly index as into Equation (11), and based on the fitting error of each assembly index, construct the lower bound and upper bound of the unit body evaluation interval through the following formula:

[0099] (12);

[0100] wherein is the middle value of the unit body evaluation interval, and are the lower bound and upper bound of the unit body evaluation interval respectively, then the unit body evaluation interval is .

[0101] For the unit to be evaluated, after obtaining the assembly data of the unit, screen the assembly indicators, and use the actual values of the screened assembly indicators to calculate the evaluation indicators of the unit through Equation (11). , if falls within the evaluation interval , then the unit is considered qualified; otherwise, it is unqualified.

[0102] The above evaluation interval can be used to determine whether the unit to be evaluated is qualified:

[0103] Among them, the units to be evaluated include units that have not been assembled into an aeroengine, and units included in an aeroengine that has completed the overall assembly but has not been tested. When it is determined that the unit to be evaluated is unqualified through the above method, it needs to be replaced and then assembled into an aeroengine. If the unit to be evaluated has been assembled into an aeroengine but is evaluated as unqualified, there is no need to conduct subsequent aeroengine testing. After replacing it with other qualified units, the testing can be carried out again.

[0104] Through the above method, a fan evaluation interval, a compressor evaluation interval, and a turbine evaluation interval can be constructed to respectively evaluate the qualification of the fan, compressor, and turbine to be evaluated.

[0105] On the basis of the above solution, the process optimization method can further include the following steps:

[0106] Step 5, after obtaining the assembly data of each unit when the aeroengine test fails and screening the assembly indicators, use the screened assembly indicators of the unit when the aeroengine test passes and fails to calculate the evaluation indicators of the unit, divide the state interval of the evaluation indicators of the unit, and construct a Bayesian network model with the test indicators of the aeroengine as the target variable and the state interval as the attribute variable. Based on the Bayesian network model, predict the posterior probability when the test indicators of the aeroengine to be evaluated are qualified, so as to provide guidance for the production and assembly of the aeroengine.

[0107] When the aeroengine test fails, collect the assembly data of each unit and screen the assembly indicators; use the actual values of the screened assembly indicators when the aeroengine test passes and fails, and calculate the evaluation indicators of the unit when the test passes and fails based on Equation (11). , and count all the evaluation indicators range; use the elbow method to determine the optimal number of clusters of the evaluation indicators , and use the K-Means clustering algorithm to cluster the evaluation indicators , and for the evaluation indicators Divide the range of the evaluation index into state intervals, and assign a unique code to each divided state interval, so as to obtain the evaluation index The state intervals and their corresponding codes.

[0108] Based on the Bayesian theory, take the state intervals of the evaluation index as the attribute variables, and the test run indexes of the aero-engine as the target variables to construct a Bayesian network model; calculate the combinations of each attribute variable when the target variable is qualified through the Bayesian network model, screen out the combinations of attribute variables greater than the probability threshold set by the enterprise, and use the codes of each attribute variable in the combination to construct an assembly suggestion table for providing assembly suggestions for the enterprise to assemble the aero-engine.

[0109] In an example of the present invention, when the unit bodies are set as the fan, the compressor, and the turbine, the process of constructing the Bayesian network model is as follows:[[]]

[0110] First, calculate the joint probability distribution:

[0111] (13);

[0112] Among them, G is the target variable of the given category, that is, the test run index; the given category is qualified and unqualified; is the prior probability distribution of the target variable G, indicating the probability that G takes different values without any other information; , , are the attribute variables, corresponding to the evaluation indexes of the fan, the compressor, and the turbine respectively of the state intervals; indicates that when G is determined, takes the conditional probability of different values, that is, the conditional probability that the fan evaluation index takes different state intervals; similarly, , have the same meaning, corresponding to the compressor and the turbine respectively; is the joint probability distribution, describing the probability that the target variable G and the attribute variables , , take specific state intervals at the same time.

[0113] Secondly, for formula (13), construct a Bayesian network model and calculate the posterior probability of the target variable G:

[0114] (14);

[0115] Among them, indicates that given the attribute variables , , Take the posterior probability of the target variable G in the corresponding state interval, that is, the probabilities corresponding to the target variable being qualified and unqualified; Denote the attribute variable 、 、 Take the total probability of the corresponding state interval.

[0116] According to the above method, calculate the posterior probability of the target variable G being qualified under the combination of state intervals corresponding to different encodings of each attribute variable, so as to obtain the posterior probability table; in this posterior probability table, the first three columns are the corresponding encodings of the state intervals of the attribute variables 、 、 respectively, and the fourth column is the posterior probability of the target variable G being qualified under the encoding combination of the first three columns, that is, the posterior probability of the aero-engine test run performance index being qualified.

[0117] For a to-be-evaluated aero-engine, use the assembled indexes after screening of the assembly data of this aero-engine, calculate the evaluation indexes of each unit body through formula (11); determine the state interval it falls into according to the value of this evaluation index, so as to obtain the encoding combination of the state intervals into which the evaluation indexes of each unit body fall; based on this encoding combination, query the posterior probability of the target variable G being qualified from the posterior probability table. If this posterior probability is greater than the probability threshold, allow this aero-engine to be test run; otherwise, directly disassemble and reassemble it to reduce the test run cost of the enterprise.

[0118] On the basis of the above scheme, the following processing can also be carried out:

[0119] Based on the posterior probability table, select the combination of state intervals of each unit body with the target variable being qualified and the posterior probability being greater than the probability threshold, and use the encodings of each state interval in the combination as the assembly suggestion table; for the unit bodies determined to be qualified through step 4, they can be preferentially assembled into an aero-engine according to the combination method of the assembly suggestion table to ensure that it can probably pass the test run. Through this way, the assembly process is optimized and guidance is provided for the aero-engine assembly process.

[0120] The following gives an application description of an embodiment of the present invention.

[0121] This embodiment takes a certain aero-engine manufacturing plant as the research object, and has collected the assembly data and its test run data of 49 aero-engines of a certain model, including 37 aero-engines that passed the test run and 12 aero-engines that did not pass the test run; the aero-engine assembly data details the assembly indexes of each unit body during the assembly process. This data has the characteristics of high-dimensional heterogeneity, multi-source and multi-modal, and there is a strong correlation between the assembly indexes; for the structural diagrams of each unit body of the aero-engine, see Figure 2The comprehensive vibration value of an aircraft engine is one of the most critical test indicators in the test data, and can comprehensively reflect the overall performance of the aircraft engine to a certain extent. In actual engineering, the upper limit of the comprehensive vibration value set by the aircraft engine manufacturer is 30, that is, when the comprehensive vibration value exceeds 30, the engine is judged to have failed the test.

[0122] This embodiment aims to construct evaluation indicators of each unit body of the aircraft engine and calculate its qualified interval based on the above-mentioned real assembly data and test data for fault diagnosis; through the state interval combination of the unit body evaluation indicators, predict the probability of the unknown aircraft engine test indicator being qualified, that is, the probability that the comprehensive vibration value is less than 30, and provide an assembly suggestion table to achieve assembly process optimization; the specific steps are:

[0123] By screening the assembly indexes of the aircraft engine assembly data, 23 assembly indexes of the fan are extracted (corresponding to Figure 3a 1 to 23), the compressor extracts 46 assembly indicators (corresponding to Figure 3b 1 to 46), turbo extractor 59 assembly indicators (corresponding to Figure 3c 1 to 59); the assembly indexes of the fan include the initial quantity after adjustment, angular direction, coaxiality, etc.; the assembly indexes of the compressor include the outer circle runout of the drum shaft end stop, coaxiality, assembly clearance at different positions, etc.; the assembly indexes of the turbine include runout value, angular direction, unbalance, etc.; they can be set according to actual needs in the application.

[0124] In this embodiment, for the assembly data of 37 aircraft engines that have passed the test after screening, the fan assembly network model, the compressor assembly network model, and the turbine assembly network model of the qualified aircraft engines are respectively constructed according to the unit body.

[0125] The network dynamics theory is introduced to perform network dynamics fitting on the assembly network model of each unit and calculate the fitting error; the fitting error of the assembly index of the fan, compressor and turbine is The mean values ​​are 0.0412, 0.1206 and 0.0457 respectively, and the kinetic fitting error diagram is drawn, as shown in Figure 4a , Figure 4b and Figure 4c As shown in the figure, there is little difference between the actual value of each unit assembly index and the simulated value calculated by the dynamic model, indicating that the dynamic model of each unit of the aircraft engine that has passed the test run can accurately characterize each assembly index and unit assembly status.

[0126] In this embodiment, the evaluation interval of each unit body can be constructed by formula (12), specifically: First, the simulated value of the assembly index of each unit body in steady state is Substitute to obtain the middle value of the evaluation interval , which are 0.5933, 0.8246, and 0.7091 respectively; then, substitute the mean value of the fitting error of each unit assembly index to obtain the upper bound and the lower bound of the unit evaluation interval, and the evaluation intervals of the fan, compressor, and turbine are obtained as [0.5521, 0.6345], [0.704, 0.9452], and [0.6634, 0.7548] respectively.

[0127] For the unit to be evaluated, calculate the value of the evaluation index of the unit using Equation (11). When the value falls within the evaluation interval of this unit, it can be determined that this unit is qualified. See Figure 5a , Figure 5b and Figure 5c as shown; in the figure, green represents the units that fall within the unit evaluation interval, and orange represents the units that fall outside the unit evaluation interval. It can be found that this method can accurately distinguish qualified and unqualified units.

[0128] In this embodiment, for the evaluation indexes of each unit of the aero-engine that has passed or failed the test run, calculate the sum of squared distances under each number of clusters, and use the elbow method to select the number of clusters corresponding to the inflection point as the optimal number of clusters for each unit evaluation index; as Figure 6a , Figure 6b and Figure 6c shown, the optimal number of clusters for the evaluation indexes of the fan, compressor, and turbine is all 4; on this basis, use K-Means to cluster the data of each unit evaluation index and divide the state interval, and assign a unique code to each divided state interval, so as to obtain the state interval of each unit evaluation index and its corresponding code, as follows:

[0129] Table 1: Division of the state interval of each unit evaluation index and its corresponding code.

[0130]

[0131] According to the Bayesian theory, take the evaluation indexes of each unit as attribute variables, and the test run index of the aero-engine as the target variable, that is, the comprehensive vibration value as the target variable, and construct a Bayesian network model. The specific structure is shown in Figure 7 as shown.

[0132] In terms of model evaluation, measure the prediction performance of the model by drawing the receiver operating characteristic curve (ROC curve); the area under the ROC curve is an important index for evaluating the prediction accuracy of the model. The larger its value, the stronger the prediction ability of the model; the area under the ROC curve of this Bayesian network model is 71.61%, as Figure 8As shown, it indicates that the model has a high prediction accuracy.

[0133] Based on the Bayesian network model, the posterior probability of the target variable G being qualified under the combination of the state intervals of different attribute variables is calculated through Equation (14), that is, the probability that the comprehensive vibration value of the aero-engine meets the requirements under the combination of the evaluation index values of different unit bodies, so as to obtain the posterior probability table.

[0134] Specifically, list all possible combinations of the state intervals of the unit body evaluation indexes, and calculate the posterior probability of the target variable G being qualified under each combination to obtain the posterior probability table of the target variable G being qualified, as shown in Table 2. In this posterior probability table, the first three columns are the state intervals corresponding to the evaluation index values of the fan, compressor, and turbine respectively, and the fourth column is the posterior probability of the target variable G being qualified under this combination, that is, the probability that the comprehensive vibration value of the aero-engine during the test run meets the requirements.

[0135] Table 2: Posterior Probability Table.

[0136]

[0137] Based on this posterior probability table, the probability threshold can be set in combination with the actual production requirements of the enterprise. When the qualified probability of the target variable G is greater than this probability threshold, it is considered that this aero-engine can probably pass the test run; in this embodiment, the probability threshold is set to 90%.

[0138] On this basis, if there is an aero-engine to be evaluated, the evaluation indexes of its each unit body can be calculated through Equation (11) using the assembly data, and the codes corresponding to the state intervals into which the evaluation indexes of each unit body fall can be queried with reference to Table 1, and then substituted into Table 2 to query the qualified probability of the target variable G under this combination of codes. If it is greater than 90%, it is allowed to conduct the test run, otherwise it is directly disassembled and reassembled to reduce the test run cost of the enterprise.

[0139] Based on this posterior probability table, select the combination of the state intervals of the unit body when the posterior probability is greater than 90% as the assembly suggestion table, as shown in Table 3. During the actual assembly process of the enterprise, the evaluation indexes of each assembled unit body can be calculated through Equation (11). If the value falls within the unit body evaluation interval, it is considered that this unit body is qualified and can be used for the assembly of the aero-engine; for the qualified unit bodies, they can be preferentially assembled into an aero-engine according to the combination method of the assembly suggestion table to ensure that it can probably pass the test run. Through this method, the assembly process is optimized and guidance is provided for the aero-engine assembly process.

[0140] Table 3: Assembly Suggestion Table.

[0141]

[0142] The embodiments show that the method for optimizing the aero-engine assembly process based on assembly data constructed by the present invention can effectively break the "barriers" existing in the assembly process and the test run process in engineering applications, can predict the qualified probability of the aero-engine test run indicators, and provide an operable assembly suggestion form for the assembly process; the evaluation indicators and their evaluation intervals of the fan, compressor and turbine constructed by the present invention can effectively evaluate the unit body, thereby effectively improving the assembly quality and efficiency, and can greatly reduce the enterprise test run cost; the assembly network model and dynamic model of each unit body of the aero-engine proposed by the present invention can accurately describe the mutual relationship between each assembly indicator and explain the relationship between the assembly indicator and the change of the system state, and has interpretability.

[0143] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. An aircraft engine assembly process optimization method based on assembly data, characterized in that: include: When the aircraft engine passes the test run, the assembly data of each unit in the aircraft engine is collected, and the assembly index of the unit assembly data is screened; Calculate the correlation between the assembly indicators after unit body screening, and construct a unit body assembly network model based on complex network theory, including: network modeling based on unit body assembly data, abstract the assembly indicators after screening in the unit body assembly data into nodes in the network, determine the mutual relationship between the assembly indicators based on the correlation between the assembly indicators, and abstract the mutual relationship into edges in the network to construct the unit body assembly network model; construct an adjacency matrix using the mutual relationship, and determine the positive adjacency matrix and the negative adjacency matrix according to the size of the correlation; the correlation is calculated using the Pearson correlation coefficient formula; The network dynamics theory is introduced to fit the unit assembly network model with a neural network dynamics model to obtain the unit dynamics model, which is expressed as: ； Where t is the time parameter, , represents the i-th and j-th assembly indexes, Assembly index The initial state of Assembly index The inverse of its own rate of change, is the number of assembly indicators in the unit body assembly network model, a is the trigger threshold, n is the slope of the activation function, and The positive adjacency matrix and the negative adjacency matrix The element in row i and column j in and Assembly index The intensity of excitation and inhibition; Based on the solution of the unit body dynamics model, the fitting error of each assembly index is determined; Based on the fitting error of the assembly index, the unit cell evaluation interval is constructed using the mean field theory, including: Evaluation indicators of unit body By calculating the expectation of all assembly indicators in the unit body assembly network model, we can get: ； in, is the number of assembly indices in the unit cell assembly network model, Assemble the network model for the unit assembly index; Simulated values ​​of assembly indices As Substitute the value of into the above formula, and based on the The fitting error of the assembly index , the lower and upper bounds of the unit cell evaluation interval are constructed by the following formula: ； in, is the middle value of the unit cell evaluation interval, and are the lower and upper bounds of the unit evaluation interval, respectively. Then the unit evaluation interval is ; For a unit body to be evaluated, an evaluation index of the unit body is determined by using the assembly index of the unit body; based on the relationship between the evaluation index and the evaluation interval of the unit body, it is determined whether the unit body to be evaluated is qualified.

2. The method for optimizing the assembly process of an aircraft engine based on assembly data according to claim 1, characterized in that: The method further comprises: The evaluation index of the unit body is calculated by using the assembly index of the unit body after screening when the aircraft engine test passes or fails, and the evaluation index of the unit body is divided into state intervals. A Bayesian network model is constructed with the test index of the aircraft engine as the target variable and the state interval as the attribute variable. The posterior probability that the test index of the aircraft engine to be evaluated is qualified is predicted based on the Bayesian network model.

3. The method for optimizing the assembly process of an aircraft engine based on assembly data according to claim 1, characterized in that: The assembly data of each unit body is composed of multiple assembly indicators. A variance analysis is performed on all the assembly indicators in the assembly data of each unit body, and the assembly indicators with zero variance are eliminated based on statistical principles, so as to obtain the screened assembly indicators of each unit body.

4. The method for optimizing the assembly process of an aircraft engine based on assembly data according to claim 1, characterized in that: The specific formula for determining the excitation intensity and the inhibition intensity is: ， ； in, and They are the positive correlation coefficient matrices of the unit cell assembly network model And the negative correlation coefficient matrix The element in the i-th row and j-th column of .

5. The method for optimizing the assembly process of an aircraft engine based on assembly data according to claim 1, characterized in that: Based on the solution of the unit body dynamics model, the fitting error of each assembly index is determined, including: The value of the unit body dynamics model is set to zero; the solution value of the assembly index is obtained by solving the unit body dynamics model as the simulation value; and the difference between the simulation value and the actual value of the assembly index is calculated as the fitting error of the assembly index.

6. The method for optimizing the assembly process of an aircraft engine based on assembly data according to claim 2, characterized in that: The actual values ​​of the assembly indexes after screening when the aircraft engine passes or fails the test are used to calculate the evaluation indexes of the unit body when the test passes or fails, and the range of all evaluation indexes is counted; the elbow method is used to determine the optimal clustering number of the evaluation index, the K-Means clustering algorithm is used to cluster the evaluation index, and the range of the evaluation index is divided into state intervals; each divided state interval is given a unique code, so as to obtain the state interval of the evaluation index and its corresponding code; The Bayesian network model is used to calculate the combination of attribute variables when the target variable is qualified, from which the combination of attribute variables greater than the probability threshold set by the enterprise is screened out, and the encoding of each attribute variable in the combination is used to construct an assembly suggestion table to provide assembly suggestions for the enterprise for the assembly of aircraft engines.

7. An aircraft engine assembly process optimization device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that: When the processor is executed by a computer, the method for optimizing the assembly process of an aircraft engine based on assembly data according to any one of claims 1 to 6 is implemented.

Images