CN-122020396-A - Engine composite material damage state evaluation method and system based on machine learning and considering service environment

CN122020396ACN 122020396 ACN122020396 ACN 122020396ACN-122020396-A

Abstract

The invention discloses an engine composite material damage state evaluation method and system based on machine learning and considering service environment, and belongs to the technical field of structural health monitoring, data processing and machine learning. According to the method, operation use, damage maintenance and service environment data of a target key piece are obtained, a multi-source data object association relation is established, characteristic parameters such as damage frequency and damage area are extracted, service environment influence factors are analyzed, an evaluation index system comprising the service environment influence factors, namely the service life, the number of use cycles, precipitation, temperature, monthly air temperature which is greater than 20 months, humidity, frost period and salt spray exposure attribute is constructed, after sample pretreatment and label set generation, an evaluation index is taken as an input characteristic, a damage grading result is taken as an output label, a machine learning damage state evaluation model is trained, a mapping relation between service environment factors and damage states is established, and the method can be used for overall process state evaluation, maintenance detection, risk judgment, decision support and the like of the key piece of an engine composite material.

Inventors

SUN SHUANG
YANG JINCHENG
ZHANG CHUNXIAO
DING SHUITING
WANG ZHIPING
LU PENGCHENG

Assignees

中国民航大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (13)

1. An engine composite damage state evaluation method based on machine learning and considering service environment is characterized by at least comprising the following steps: The method comprises the steps of SS1, obtaining operation and use information, damage overhaul information and service environment data of an engine composite material target key piece, and establishing a corresponding association relation based on a key piece identifier, an airplane identifier, maintenance approach time and overnight airport information to form an original data set facing the target key piece; SS2, extracting damage characteristic parameters of the target key piece based on an original data set, carrying out association analysis on damage records and corresponding flight environments, overnight airports and regional environments, and determining service environment influence factors and corresponding relations of the service environment influence factors for influencing the damage state of the target key piece; SS3, mapping service environment data to a target key piece according to a correlation analysis result, extracting evaluation indexes reflecting service states of the key piece and external environment actions, wherein the evaluation indexes at least comprise the service life of the key piece, the number of using cycles, precipitation, temperature, monthly air temperature which is more than 20 degrees for months, humidity, frost period and salt fog exposure attribute, and establishing a damage state evaluation index system; SS4, carrying out at least missing data processing, dimension unification, class variable coding and sample balancing processing on the damage state evaluation index system, and generating a training sample set with labels corresponding to each evaluation index according to the damage grading result in the maintenance record; SS5, training a machine learning damage state evaluation model by taking an evaluation index in a training set with a marked sample as an input characteristic and a damage grading result as an output label, and establishing a mapping relation between service environment factors and damage states of key parts of the engine composite material; And SS6, inputting the current evaluation index of the target key piece to be evaluated into a machine learning damage state evaluation model, and outputting the damage state category and the corresponding probability value.
2. The method of claim 1, wherein in step SS1, the operation usage information at least includes a first installation time, an age of the part, a number of flight cycles, a number of flight hours, and a number of flight take-off and landing times of the target key, wherein the age of the part is determined based on the first installation time and a maintenance approach time, the damage maintenance information at least includes a maintenance approach time, a damage position, a damage type, a damage size, and a damage classification result of the target key, the service environment data at least includes weather observation data corresponding to an airport area where the target key is subjected to service aircraft over night, including annual average precipitation, annual average air temperature, annual average relative humidity, monthly air temperature greater than 20 months, annual average frost day, and airport coastal salt spray area attribution information, and the source data are collected, cleaned, and aligned according to a unified time reference and a unified object identifier to form a multi-source raw data set for evaluating a damage state of the target key.
3. The method according to claim 1 or 2, wherein in step SS1, the establishment of the correspondence relationship at least includes: S11, object identification association, namely, based on the object key piece identification and the aircraft identification, combining the installation time, the disassembly time and the maintenance approach time, matching the installation correspondence between the object key piece and the aircraft, and establishing installation attribution relation data of the object key piece in different service stages; S12, period data collection, namely dividing service maintenance periods of the target key piece according to installation time, removal time and adjacent maintenance approach time, and extracting flight task records and operation use records of corresponding aircraft in each service maintenance period to form a use data set taking the service maintenance period as a unit; s13, environment data mapping, namely calling environment parameter data of a region of a corresponding airport in a service environment database in a corresponding time interval based on overnight airport information corresponding to each flight task, and matching and summarizing according to a service maintenance period to form an environment exposure data set; And S14, carrying out consistency association on the multisource data, namely associating the installed attribution relation data, the use data set, the environment exposure data set and the damage overhaul information according to the target key piece identification, the service overhaul period and the maintenance approach time to form target key piece sample data.
4. The method according to claim 1, wherein in the step SS2, the damage characteristic parameters at least comprise damage frequency and damage area, wherein the damage frequency is obtained by dividing damage frequency of the target key piece in a preset statistical period by a fleet size corresponding to the service environment, the fleet size is the number of aircraft under the same service environment condition, the damage area is used for representing the size of the damaged range of the target key piece, the damage area is determined by reading area data of the corresponding damage area in a maintenance record, a detection record or a maintenance measurement result, and when no direct area value exists in the original record, area calculation is performed according to the boundary size of the damage area and/or the detection labeling result, so as to form the area parameter capable of quantitatively representing the damage degree.
5. The method according to claim 1 or 4, wherein in step SS2, the correlation analysis and determination of service environment influence factors influencing the damage status of the target key member and their corresponding relations at least include: S21, classifying service environment, namely constructing an area environment characteristic data table based on rainfall, air temperature, monthly air temperature which is more than 20 months, humidity, frost period and salt spray exposure attribute of an area where an overnight airport belongs to which a target key piece corresponds, performing cluster analysis on the service environment by adopting a systematic clustering method, and classifying the service environment corresponding to the target key piece into cold dry, high-temperature and high-humidity salt spray and high-rainfall frost areas; s22, service feature merging and damage frequency statistics, namely performing correlation test on the number of cycles and the age of a part of a target key part, and selecting the age of the part as a service feature variable for damage frequency influence analysis when the linear relation between the number of cycles and the age of the part passes the significance test; S23, performing analysis of variance and significance test, namely taking the damage frequency as a dependent variable and taking the service life and service environment type as independent variables, establishing an analysis of variance model comprising service life, service environment type and interaction items of the service life and the service environment type, and performing significance test on the influence relationship between the damage state of the target key piece and the service environment; S24, screening the service environment influence factors and service characteristics which have obvious influence on the damage state of the target key piece according to the saliency test result, and determining the corresponding relation among the age, precipitation, temperature, humidity, frost period, salt fog exposure attribute and the damage state of the target key piece.
6. The method of claim 1, wherein in step SS3, the damage status assessment index system at least comprises a part usage cycle number and a piece age, wherein the part usage cycle number and the piece age are used for representing working parameters, the precipitation, the temperature, the monthly air temperature is larger than 20 degrees, the humidity, the frost period and the salt spray exposure attribute are represented by the service environment, the piece age is determined by the accumulated service duration of a target key piece from the first installation date to the current maintenance approach date, the usage cycle number is determined by the number of flight cycles of aircraft accumulated completion in the service time interval of the target key piece, the precipitation, the temperature, the monthly air temperature is larger than 20 degrees, the month record of the corresponding overnight airport of the target key piece in the service time interval, the humidity, the frost period and the salt spray exposure attribute are calculated based on the regional environment data, the frost period is used for representing the continuous condition of the target key piece in the low-temperature icing related environment, and the salt spray exposure attribute is used for representing the exposure level of the target key piece in the high-salt environment.
7. The method of claim 1, wherein in step SS4, the missing data processing includes completing or eliminating missing values of continuous variable evaluation indexes, merging and marking missing items of component type variable evaluation indexes, dimension unification includes normalizing or normalizing continuous variable evaluation indexes of different dimensions, class variable encoding includes vectorizing airport, region and damage type related class information by adopting a single thermal encoding mode to convert the information into numerical characteristics which can be recognized by a machine learning model, sample balancing processing includes linear interpolation synthesis of minority class samples in a feature space by adopting a synthetic minority class oversampling technology aiming at unbalanced number of different damage state class samples in damage data of a target key piece, and the number of the damage state class samples reaches a relatively balanced distribution state.
8. The method according to claim 1 or 7, wherein in step SS4, when the labeled training sample set is generated according to the injury classification result, the service state evaluation index of each target key before the corresponding maintenance approach time is paired with the injury classification result corresponding to the maintenance record to form a single labeled sample, and for the case that the same target key has multiple maintenance records, multiple labeled samples are respectively constructed according to time sequence, and the evolution information of the injury states of the same target key in different service stages is reserved.
9. The method of claim 1, wherein in step SS5, the machine learning impairment state assessment model is constructed based on a random forest model, comprising at least: s510, sample set division and parameter setting, namely dividing a sample set with labels into a training set and a test set according to a preset proportion, taking each evaluation index in the training set and the test set as a model input characteristic, and taking a damage grading result as a model output label; S511, constructing sub-sample sets and training decision trees, namely, based on a Bootstrap sampling mode, putting back extracted samples from the training sets, constructing a plurality of training sub-sample sets, and respectively building decision trees for each training sub-sample set; S512, integrating classification and probability output, namely inputting a test set into a trained random forest model, integrating classification results of all decision trees in a voting mode, determining a damage state evaluation result of a target key piece according to the damage category with the largest number of votes, generating a probability value of a corresponding damage category according to the voting duty ratio obtained by each damage category, and establishing a random forest mapping relation between service environment factors and damage states; S513, model evaluation and parameter correction, namely calculating model accuracy, recall rate, precision and F1 evaluation indexes based on confusion matrix results of the training set and the testing set, and correcting the number of decision trees and node splitting feature numbers according to the evaluation results to obtain a random forest model meeting the damage state evaluation requirement.
10. The method of claim 1, wherein in step SS5, the machine learning impairment state assessment model is constructed based on a BP neural network model, comprising at least: S520, defining network input and output, namely dividing a sample set with labels into a training set and a testing set according to a preset proportion, taking each evaluation index as network input, taking a damage grading result as network output, carrying out normalization processing on input data, and eliminating the influence of different evaluation index dimension differences on a network training process; S521, establishing a network structure, namely establishing a BP neural network structure comprising an input layer, an hidden layer and an output layer, wherein the number of nodes of the input layer corresponds to the number of evaluation indexes, the number of nodes of the output layer corresponds to the number of damage classification categories, and the number of nodes of the hidden layer is set according to the scale of training samples and the complexity of evaluation tasks; S522, performing error back propagation training, namely inputting a training sample into a BP neural network model, performing forward propagation calculation to obtain output results of each damage state type, performing back propagation according to errors between the output results and a real label, iteratively updating connection weights and threshold parameters between layers, and ending network training when the output errors are smaller than a preset error limit value or the training times reach a preset upper limit; and S523, inputting the test set into the BP neural network model, outputting the damage category corresponding to the target key piece and the probability value of each damage category, calculating the accuracy, recall rate, precision and F1 evaluation index based on the prediction results of the training set and the test set and the confusion matrix result, and verifying the identification capability of the BP neural network model on the damage state of the target key piece.
11. The method of claim 4, wherein in step SS4 and step SS5, the labeled sample set and the machine learning damage state assessment model employ a two-dimensional modeling method combining damage probability prediction and damage state identification, and the method at least comprises: Calculating damage frequency according to damage frequency of a target key piece in a preset statistical period and a fleet scale under a corresponding service environment, taking a continuous probability value or probability level corresponding to the damage frequency as a first output label, constructing a first training sample set facing damage probability prediction by taking each evaluation index as a first input characteristic, and establishing a first evaluation model for outputting the damage occurrence probability value or probability level; According to the damage grading result in the maintenance record, taking damage categories corresponding to the slight damage, the moderate damage and the severe damage as second output labels, taking each evaluation index as a second input characteristic to construct a second training sample set facing damage state identification, and establishing a second evaluation model for outputting the damage state categories; When the target key piece to be evaluated is input, the damage occurrence probability value or probability level and the damage state category are respectively output, so that the combined prediction evaluation of the damage risk level and the damage state level is realized.
12. The method of claim 1, wherein step SS6 is followed by a step SS7 for iterative model updating, wherein the step SS7 is to write back the repair and overhaul result, the damage classification result and the corresponding service environment data of the target key element, which are newly added, to the training sample set, periodically correct the damage state evaluation index system, and retrain the machine learning damage state evaluation model according to a preset iteration period.
13. An engine composite damage state evaluation system for realizing the engine composite damage state evaluation method based on machine learning and considering service environment according to any one of claims 1-12.

Description

Engine composite material damage state evaluation method and system based on machine learning and considering service environment Technical Field The invention belongs to the technical field of structural health monitoring, data processing and machine learning, relates to material damage state evaluation by combining service environment information and a machine learning model, and particularly relates to an engine composite material damage state evaluation method and system based on machine learning and considering service environment. Background The composite material structure has good high toughness, fatigue resistance, temperature resistance and corrosion resistance, and can be widely applied to large civil aircraft structures, and particularly plays an important role in engine nacelle and fan blade parts. Compared with the traditional metal material, the composite material does not generate electrochemical corrosion in the service process, has higher specific strength, specific rigidity and damage tolerance, and can still maintain better structural integrity in a complex load environment. However, the interlayer bonding strength of the composite material is relatively low, and when subjected to factors such as foreign object impact, fatigue load, environmental aging and the like, hidden damages such as delamination, matrix cracking, interface debonding and the like are easily generated. If the damage cannot be found and accurately estimated in time, the structural bearing capacity is gradually expanded and weakened, and the flight safety is further affected. With the great application of composite materials in aircraft structures, the damage and repair problems of the composite material structures are more and more prominent, and damage detection and repair decisions of the composite material structures are key problems to be solved in the field of aviation repair. Aiming at damage prediction and state evaluation of a composite material structure, the existing research is mainly focused on theoretical analysis of damage evolution mechanism, simulation modeling based on finite element simulation and research on performance degradation rule under laboratory conditions. The above research is to simulate the damage of the composite material through experiments, record the damage data of the composite material, analyze the damage development rule, further predict the performance degradation, damage growth and residual life of the composite material, and have relatively independent performance prediction and simulation for the composite material and single environment. However, in the actual running process of civil aircraft, structural members not only bear complex alternating load, but also are exposed to interaction of multiple environmental factors such as temperature change, humidity alternation, salt spray corrosion and the like for a long time, and the occurrence and development of damage are nonlinear and uncertain. The existing method does not fully utilize overhaul data accumulated in the actual operation and maintenance process of an airline company, so that a data model capable of evaluating the damage state of the composite material according to limited external environment information is needed. However, in the research of the aircraft damage state evaluation, the research of the damage state evaluation based on the actual overhaul data of the airlines is blank. In recent years, various machine learning algorithms show remarkable advantages in the fields of multidimensional data feature extraction, nonlinear classification and prediction, and a technical basis is provided for constructing a data-driven damage state evaluation model. However, when the algorithm is applied to the damage state evaluation of the engine composite material, a plurality of outstanding technical problems are faced, such as how to scientifically construct an index system covering the multi-dimensional service environment characteristics and the working parameters from the real operation and maintenance data, how to treat the class imbalance problem caused by the fact that the probability difference of each damage level is great in the real overhaul data, and how to reasonably quantify and encode the geographic climate characteristics of different service areas, so that the geographic climate characteristics are effectively incorporated into a modeling framework of a machine learning model. The problems restrict the effective popularization and application of the data driving method in engineering practice in the field. In summary, in the prior art, the damage state of the key component of the engine composite material is evaluated, and the problems of insufficient consideration of real service environment factors, low utilization degree of multi-source operation and maintenance data, insufficient characterization of complex nonlinear damage influence relationship and the like still exist. Therefore, how to construct an