Civil aircraft fault risk bidirectional reasoning evaluation method

Abstract

The application discloses a civil aviation aircraft fault risk bidirectional reasoning evaluation method. The method calculates three risk characterization parameters of occurrence degree, detection degree and severity degree, adopts a method of combining dynamic data with static data to perform bidirectional reasoning of fault occurrence risk, fuses evidence information given by multiple experts, fuzzily process language variables of the experts, respectively gives corresponding weights to the three risk characterization parameters, and performs weighted calculation and optimization on RPN values. The reverse FTF method is adopted, various influence factors of an aircraft flight control system are analyzed, and bottom events are deeply analyzed. The fault tree is converted into a Bayesian network, which is convenient for posterior probability calculation and bidirectional reasoning. Dynamic and static data are organically combined, and the prediction accuracy of the network is improved. The D-S evidence fusion method is optimized, and the conflict problem of evidence information given by experts is solved. Language variables given by experts are processed by fuzzification, so that the calculation result is more objective and accurate.

Description

Technical Field

[0001] This invention relates to the field of aviation transport safety risks, and in particular to a two-way reasoning assessment method for civil aircraft failure risks. Background Technology

[0002] With the development of the civil aviation industry, mechanical equipment is becoming increasingly complex, and the technological content and integration of aircraft are becoming more sophisticated. The increasing number of onboard devices leads to increasingly stringent requirements for aircraft safety and reliability. Therefore, it is essential to diagnose and analyze potential aircraft failure risks. Currently, methods for diagnosing and analyzing aircraft failure risks include fault tree analysis using expert systems and physical models of the aircraft, Bayesian network analysis, fault diagnosis methods based on historical experience and data mining techniques, and neural network analysis. However, methods for diagnosing aircraft control system failures are particularly lacking. Domestic airlines rely heavily on aircraft maintenance manuals, fault isolation manuals, and the experience of senior engineers for fault diagnosis of aircraft control systems, which introduces considerable uncertainty.

[0003] Existing fault risk assessment systems often rely solely on fault maintenance records or subjective expert judgments, resulting in a lack of comprehensive evaluation across multiple factors. Furthermore, Bayesian network models can only obtain the probability of occurrence for corresponding accident modes through data processing. The determination of Bayesian network nodes is typically based on expert experience, making the results highly susceptible to its influence. Moreover, different experts have varying degrees of familiarity with the fault modes of the target aircraft system, leading to significant differences in their findings. Additionally, traditional Bayesian networks lack the dimensionality required for data processing to calculate accident risks, thus failing to achieve adequate risk assessment.

[0004] Traditional fault tree analysis and FMECA methods cannot perform bidirectional reasoning about fault states due to their unidirectional logical connections. This invention combines fault trees with Bayesian networks, and uses the directed acyclic graph principle of Bayesian networks to fuse dynamic QAR data to achieve bidirectional reasoning that combines dynamic and static data. Summary of the Invention

[0005] In view of the problems existing in the prior art, this invention provides a two-way reasoning assessment method for civil aircraft failure risk. This method optimizes the traditional fuzzy RPN algorithm, employs a reverse FMECA+FTA approach to construct a failure assessment system for the aircraft flight control system, and combines dynamic QAR data to construct a Bayesian network. Reasoning is performed on prior reliability data originally derived solely from user feedback, combining dynamic and static data to obtain the posterior probability of the failure mode, and classifying levels based on the posterior probability. Simultaneously, relevant experts and maintenance personnel assess and score the detection difficulty, using an improved DS evidence theory to fuse the evidence information provided by the experts and determine the detectability level. Then, an uncertain ordered weighted average operator is used to assign corresponding weights to each expert, and a triangular membership function is used to determine the severity of the final multi-expert opinion combination based on the maximum membership degree. Finally, an AHP-entropy weighting method is used to assign weights to the three risk characterization parameters, fusing the three variables of the optimized fuzzy RPN algorithm to obtain the final risk level number.

[0006] The Fault Tree Analysis (FTA) method is complex in terms of fault tree construction, qualitative analysis, and quantitative calculation, and large fault trees are difficult to understand. The FMECA method is also prone to oversights in practical applications and is unsuitable for causal analysis of complex relationships. Therefore, to comprehensively analyze the failure risk mechanism of aircraft flight control systems, a reverse FTF method is adopted. First, the top event is determined based on the functional requirements of the aircraft flight control system. Then, FMECA is performed on the bottom events obtained from the FTA method. This not only allows for a focused analysis of the multiple factors affecting the system and a more detailed consideration of multiple failures, but also saves significant time. Due to the high similarity between fault trees and Bayesian network topologies, and the fact that Bayesian networks are used for reasoning in complex systems due to their computational advantages, the fault tree is further transformed into a Bayesian network. Posterior probabilities can then be calculated based on the QAR data detected in reality.

[0007] The technical solution adopted in this invention is as follows: A two-way reasoning assessment method for civil aircraft failure risk, which calculates the occurrence degree, detectability, and severity of risk characterization parameters, uses a combination of dynamic and static data to perform two-way reasoning on failure risk, and assigns corresponding weights to the three risk characterization parameters to optimize the RPN value through weighted calculation; including the following steps:

[0008] Step 1: Calculation of the occurrence degree of risk characterization parameters: Divide the structural hierarchy of the aircraft flight control system, analyze the cause of failure for each structural hierarchy, obtain static data, and construct a fault tree structure diagram; analyze the QAR dynamic data of the aircraft flight control system, perform dynamic and static data fusion to generate a Bayesian network, and calculate the posterior probability.

[0009] Step 2: Calculation of the detectability of the risk characterization parameter: Based on the detection status scoring table, the fault occurrence stage is judged, and the information on the fault occurrence stage given by different experts is integrated to resolve the problem of conflicting evidence on the fault occurrence stage.

[0010] Step 3: Calculation of the severity of the risk characterization parameter: Assign corresponding weights to the fuzzy linguistic variables of the severity of the failure as determined by different experts, and determine the severity of the failure using the maximum membership principle based on the triangular fuzzy number function.

[0011] Step 4: Determining the weights of risk characterization parameters: Using the AHP-entropy weight method, the importance of the three risk characterization parameters—occurrence, detectability, and severity—in the risk assessment of flight control system failure is distinguished, the weights of the three risk characterization parameters are determined, the role of different characterization parameters in risk assessment is measured, and the weighted calculation optimizes the RPN value.

[0012] This invention, considering both subjective and objective factors and fault maintenance records, employs a machine learning clustering algorithm to combine dynamic monitoring data of aircraft flight status with the inherent reliability of the aircraft flight control system. Regarding the difficulty of fault mode detection, an improved DS evidence theory is used, effectively handling situations where inconsistencies or even conflicts arise in the information provided by experts, resulting in more accurate detection results. For the severity calculation, considering the subjective differences in linguistic variable expression and its fuzzy nature, a fuzzy comprehensive evaluation method is used to score based on a scoring criterion table. An uncertain ordered weighted average operator is used to assign weights to experts, and finally, a triangular membership function is used to select the maximum membership value as the severity score. This approach considers both the fuzziness of linguistic variable expression and the comprehensive handling of differences in multiple expert opinions. Finally, the AHP-entropy weight method is used to obtain the weights of each risk characterization parameter. This method integrates expert subjective opinions while also considering the structural rationality of the risk characterization parameters themselves, making the results obtained from optimizing the Risk Priority Number (RPN) more valuable for practical reference. This provides maintenance personnel with more effective assessment results.

[0013] The beneficial effects of this invention are as follows:

[0014] 1. A reverse FTF method combining FMECA and FTA is adopted, which can focus on analyzing the influence of multiple factors on the aircraft flight control system and consider multiple failure problems in detail. FMECA analysis results can be used selectively based on FTA, and in-depth analysis of underlying events can be performed, saving significant analysis time.

[0015] 2. Based on the high similarity between the topological structures of Bayesian networks and fault trees, fault trees are transformed into Bayesian networks, which facilitates the calculation of posterior probabilities for complex systems and enables bidirectional reasoning.

[0016] 3. A sliding window algorithm is used to segment QAR data according to flight phases, and the K-Means clustering method from machine learning is used to identify outliers. Introducing outliers as dynamic data into the Bayesian network improves the network's prediction accuracy, organically combining dynamic and static data for more objective and accurate calculation results.

[0017] 4. The detection rate is calculated by using the DS evidence theory to fuse information based on the evaluation information of multiple experts, taking into full account the problem of conflicting evidence information and solving the problem of the confidence level between different pieces of information.

[0018] 5. For the severity calculation, an uncertain ordered weighted average operator was adopted to assign appropriate weights to the evaluation of each expert language variable. The triangular membership function was used to determine the final severity by the maximum membership degree, resulting in a more objective and scientific severity score.

[0019] 6. The AHP + entropy weight method is used to assign appropriate weights to the three parameters of the optimized RPN, determining the comprehensive weight of the risk characterization parameters, and the evaluation results have a certain degree of objectivity. Combining the risk characterization parameter weights obtained through these two methods yields a more scientific and reasonable comprehensive weight vector for the risk characterization parameters. Attached Figure Description

[0020] Figure 1 This is a flowchart of an embodiment of the present invention;

[0021] Figure 2 This is a block diagram illustrating the fault hierarchy of an aircraft flight control system in an embodiment of the present invention.

[0022] Figure 3 for Figure 2 Fault tree diagram of the steering control system in the middle;

[0023] Figure 4 for Figure 2 Fault tree diagram of the mid-wing control system;

[0024] Figure 5 for Figure 2 Fault tree diagram of the elevator control system;

[0025] Figure 6 for Figure 2 Fault tree diagrams for the central rudder trim system and aileron trim system;

[0026] Figure 7 for Figure 2 Fault tree structure diagram of a mid-level stabilizer balancing system;

[0027] Figure 8 for Figure 2Fault tree diagram of the mid-trailing edge flap control system;

[0028] Figure 9 for Figure 2 Fault tree diagram of the central spoiler control system;

[0029] Figure 10 This is a diagram showing the relationship between the fault tree and Bayesian network in this invention.

[0030] Figure 11 This is a Bayesian network diagram combining dynamic and static data in this invention;

[0031] Figure 12 This is a schematic diagram of the membership image of the severity triangular fuzz number in this invention. Detailed Implementation

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] Reference Figure 1 A two-way reasoning assessment method for civil aircraft failure risk, through the calculation of risk characterization parameters occurrence degree, risk characterization parameter detectability, and risk characterization parameter severity, employs a combination of dynamic and static data to perform two-way reasoning on failure occurrence risk, and assigns corresponding weights to the three risk characterization parameters to calculate and optimize the RPN value; including the following steps:

[0034] Step 1: Calculation of the occurrence degree of the risk characterization parameter: Divide the structural hierarchy of the aircraft flight control system, analyze the cause of failure for each structural hierarchy, obtain static data, and construct a fault tree structure diagram; analyze the QAR dynamic data of the aircraft flight control system, perform dynamic and static data fusion to generate a Bayesian network, and calculate the posterior probability; Step 2: Calculation of the detectability of the risk characterization parameter: Make a judgment on the failure occurrence stage based on the detection status scoring table, and integrate the failure occurrence stage information given by different experts to resolve the problem of conflicting evidence of failure occurrence stage information.

[0035] Step 3: Calculation of the severity of the risk characterization parameter: Assign corresponding weights to the fuzzy linguistic variables of the severity of the failure as determined by different experts, and determine the severity of the failure using the maximum membership principle based on the triangular fuzzy number function.

[0036] Step 4: Determining the weights of risk characterization parameters: Using the AHP-entropy weight method, the importance of the three risk characterization parameters—occurrence, detectability, and severity—in the risk assessment of flight control system failure is distinguished, the weights of the three risk characterization parameters are determined, the role of different characterization parameters in risk assessment is measured, and the weighted calculation optimizes the RPN value.

[0037] In step one, the hierarchical division of the aircraft flight control system structure is determined based on the aircraft maintenance manual and fault isolation manual. The inverse FTF analysis method of FMECA+FTA is applied to identify key nodes in the underlying events of flight control system failures, forming a set of important key underlying events. The QAR dynamic data of the aircraft flight control system is analyzed using the K-Means clustering algorithm to identify abnormal data, and multiple linear regression is used to classify the abnormal situations into categories such as human and environmental factors or equipment factors. A conditional probability table of a Bayesian network is constructed sequentially. A complete Bayesian network is built through the transformation relationship between the constructed fault tree and the Bayesian network, achieving an organic combination of dynamic and static data to obtain the posterior probability under actual flight conditions. The occurrence degree of the risk characterization parameter is obtained by comparing the posterior probability scoring table.

[0038] Constructing the conditional probability table for a Bayesian network includes the following steps:

[0039] Step A1: Using a dynamic sliding window algorithm, the sliding window width is set according to the rate of change of QAR dynamic data in each flight phase of the aircraft flight control system, and the flight phase is divided into five phases: takeoff, climb, cruise, approach, and landing; data analysis is performed according to the different data change characteristics of each of the five phases.

[0040] Step A2: For each flight phase, use the K-Means clustering algorithm to divide the data into two categories: abnormal data and normal data. Store the abnormal data in the abnormal database for multidimensional abnormal data analysis in the next step.

[0041] Step A3: Based on the principles of flight, select flight parameters relevant to the flight parameters being analyzed, and use the following formula to calculate the correlation between the relevant flight parameter variables x and y:

[0042]

[0043] In the formula: are the sample means of variables x and y, respectively; r represents the correlation between variables x and y; based on the correlation strength, flight parameters with strong correlation are selected for further multiple linear regression analysis.

[0044] Step A4: Analyze the QAR dynamic data in the anomaly database using the following multiple linear regression equation to obtain the QAR dynamic dataset:

[0045]

[0046] In the formula: Y is the time series vector of the objective function; x ij= (i = 1, 2, ... i; j = 1, 2, ... j) represents the value of flight parameter i at time j; [k1, k2, ..., k m ] T is the regression coefficient vector; b is the constant term.

[0047] The QAR dynamic dataset is trained to fit the target function. The values ​​of the predicted objective function at any time stamp are compared with the data in the actual normal database. The average distance between the two values ​​is used as a threshold. If the distance between the outlier and the predicted point is less than the threshold, it indicates that the current abnormal data is not caused by a sudden failure, but by a change in the objective flight state. If the distance between the outlier and the predicted point is greater than the threshold, it indicates that the current abnormal data is an isolated sudden state caused by a failure of the sensor or equipment that detects the flight parameters.

[0048] Step A5: Treat the irreversible degradation of aircraft flight control system components as nodes, collect and analyze the QAR dynamic data before each node, classify the causes of data anomalies as human and environmental factors or equipment failure, and use this statistical data as the conditional probability table of the Bayesian network; transform the fault tree and the Bayesian network to obtain the posterior probability of the combined dynamic and static data, and score the occurrence level according to the posterior probability scoring table.

[0049] In step two, to resolve the conflict of evidence information during the failure occurrence phase, the weighting of each expert's score during the failure occurrence phase is calculated and allocated, and the basic reliability allocation of the original expert scores is corrected. This includes the following steps:

[0050] Step B1: Perform orthogonal calculations on the basic reliability allocation probability functions of two or more experts, using... The expert assesses the probability of several rating levels for evaluating the difficulty of fault mode detection. Within the identification framework Ω, the basic confidence levels for two independent detection scenarios are assigned as m1 and m2, and the Dempster combination rule is defined. for:

[0051]

[0052] In the formula, X i Y represents the first expert's assessment of the detection performance for the i-th rating level. i The assessment of the detection situation for the second expert at the i-th rating level; 1 / (1-k) is the normalization factor; k is the conflict coefficient, representing X i With Y i The conflict size between them, k∈[0,1]; A is the component failure mode that needs to be assessed for detection.

[0053] Step B2: Next, the Pearson correlation coefficient is used to calculate and assign the weight proportions of each expert's score to resolve the problem of conflicting evidence when assessing the failure occurrence stage.

[0054] First, construct the Pearson correlation coefficient matrix, and then the basic confidence assignment made by the i-th expert is m. i The basic reliability assignment m made by the j-th expert j Pearson correlation coefficient p ij The calculation formula is:

[0055]

[0056] In the formula: Cov(m) i m j ) represents the reliability assignment made by the experts. i and m j covariance; and Assign m to the basic reliability respectively i and m j The standard deviation.

[0057] Step B3: Based on the Pearson correlation coefficient p ij To construct the Pearson correlation coefficient matrix P:

[0058]

[0059] Step B4: Next, the basic reliability assignment of the original expert ratings is corrected. Based on the weight ratio and reliability calculation of non-positively correlated expert ratings in the identification framework, the non-positively correlated values ​​are corrected to 0.001; the basic reliability assignments of other expert ratings to the experts' basic reliability assignments m i The scoring conflict is Con(m) i ):

[0060]

[0061] Step B5: After normalizing the conflicting evidence information at the fault occurrence stage in the expert rating, the basic reliability assignment m made by the experts is obtained. i The relative conflict is In(m) i ):

[0062]

[0063] Step B6: Utilize relative conflict In(m) i The original basic reliability allocation is corrected, and the corrected basic reliability allocation is as follows:

[0064]

[0065] Step B7: After obtaining the corrected basic reliability assignment, replace the items with a basic reliability assignment of 0 by reducing the highest probability value by 0.001, while ensuring that each expert rating satisfies:

[0066]

[0067] Step B8: Finally, the modified multi-expert rating information is used to calculate the basic reliability distribution of the target, i.e., the probability value m:

[0068]

[0069] Based on the calculated probability values, the score result with the highest probability is selected as the final risk characterization parameter detectivity result.

[0070] In step three, the severity of the fault is determined using the maximum membership principle based on the triangular fuzzy number function; the severity assessment values ​​given by experts based on the triangular fuzzy number reference are used to synthesize the assessment information of different experts using the uncertain ordered weighted average operator; and the final severity score, the risk characterization parameter, is determined using the maximum membership principle.

[0071] In step three, the method for applying the uncertain ordered weighted average operator to determine expert weights includes the following steps:

[0072] Step C1: Calculate the arithmetic mean of the n triangular fuzzy numbers. The calculation formula is as follows:

[0073]

[0074] in: Suppose there are n experts evaluating the severity of a certain type of failure mode, R. m Let be the average triangular fuzzy number transformed by k experts into the assessment level j of the i-th risk characterization parameter variable, where i represents the severity score for the required assessment failure mode, and a, b, and c represent the corresponding values ​​of the left, top, and right vertices of the triangular fuzzy number given by the experts, respectively.

[0075] Step C2: Calculate the triangular fuzzy number R k Arithmetic mean of triangular fuzzy numbers Distance measure between The calculation formula is as follows:

[0076]

[0077] In the formula R k This represents the triangular fuzzy number representing the assessment level j of the i-th risk characterization parameter variable transformed by the k-th expert, i.e.

[0078] Step C3: Calculate the triangular fuzzy number R k and the arithmetic mean of triangular fuzzy numbers similarity between Given n triangular fuzzy numbers R k Arithmetic mean of triangular fuzzy numbers The similarity is considered as:

[0079]

[0080] Step C4: Perform fuzzy number aggregation and calculate the comprehensive triangular fuzzy number R; use the aggregation method of UOWA operators to aggregate the fuzzy numbers of different expert evaluation values ​​to obtain the comprehensive triangular fuzzy number R as follows:

[0081]

[0082] In the formula, R1, R2, ..., R n ω represents the triangular fuzzy number evaluation results given by the 1st to nth experts. k This represents the weight of the evaluation result of the k-th expert obtained under the UOWA operator in the overall evaluation result.

[0083] Step C5: Take the x-coordinate value of the vertex of the comprehensive triangular fuzzy number R, and regard it as the severity assessment value of the failure mode under the specific risk characterization parameters.

[0084] Example: The flight control system of the CRJ200, a 50-seat twin-engine turbofan regional jet designed by Bombardier Aircraft Company of Canada, is used as an example for demonstration. It is a modified version of the Challenger business jet. Although it has some characteristics and defects of business jets, its flight control system is relatively typical and has a certain representativeness for civil airliners.

[0085] First, it is necessary to determine the parameters characterizing the occurrence of failure risk. Based on the aircraft maintenance manual, failure isolation manual, and expert system, the functional components of the system are defined, resulting in a principle block diagram and a mission reliability block diagram. A first failure cause analysis (FTA) is then performed, identifying the top event as a failure of the aircraft flight control system. Based on the system failure logic hierarchy, intermediate events and confirmed bottom events are then divided, such as... Figures 2 to 9 As shown in Table 1, a qualitative analysis was performed, and the basic events in the fault tree were listed in order of importance. The hierarchical structure of the flight control system of the CRJ200 aircraft is shown in Table 1.

[0086] Table 1. Hierarchical Structure of the Flight Control System of the CRJ200 Passenger Aircraft

[0087]

[0088]

[0089]

[0090] A second Failure Cause Analysis (FMECA) is conducted. At this time, a scoring criterion for a "Failure Cause Scoring Table" for the aircraft flight control system is established. Each item in the list of important underlying events in the fault tree analysis is scored and judged to determine the underlying events that require risk assessment. The scoring criteria are shown in Tables 2 and 3.

[0091] Table 2 Failure Cause Scoring Criteria (Reliability Required)

[0092]

[0093]

[0094] Table 3 Failure Cause Scoring Criteria (Predicted Reliability)

[0095]

[0096] In the scoring criteria of the "Failure Cause Scoring Table," each scoring factor is assigned a score value between 0 and 1.0, with an average grade difference of 0.2 between levels. After assessing the "Required Reliability Score (X)" and "Predicted Reliability Score (Y)" of a component, the score value is calculated using the formula shown in the table, and the two values ​​are compared. If X < Y, the component meets the reliability requirements; conversely, if X > Y, the component does not meet the reliability requirements and is judged as a critical node. When conducting failure risk analysis, unless there are special circumstances, analyzing only critical nodes can largely pinpoint critical failure modes, saving tedious discussions and analysis and a significant amount of time. The obtained critical nodes are shown in Table 4.

[0097] Table 4. Key Nodes in the Flight Control System

[0098]

[0099] At this point, it is necessary to determine the prior probability of the bottom-level events. This is calculated using data from aircraft users, i.e., airlines—that is, static data. Static data is measured using reliability, calculating the probability that equipment or components in the flight control system will perform their intended functions under specified conditions and within the expected timeframe. This reliability is then used as the initial prior probability of the bottom-level events in the input fault tree. Records of component failures after one year of use are collected and analyzed from each airline. The ratio of these failures to the total number of components in the batch is calculated, representing unreliability—that is, the static failure probability inherent in the component's endogeneity.

[0100] The following analysis will focus on QAR dynamic data. QAR dynamic data is crucial flight data, mirroring the data from the onboard black box. It is commonly used for post-flight technical analysis, engine health analysis, and flight safety incident investigations, serving as an important data repository for flight quality analysis, operational quality analysis, and aircraft health management. QAR records cover most aircraft flight parameters, such as latitude and longitude, altitude, wind speed, wind direction and angle of attack, fuel consumption, temperature, and air pressure. Since QAR dynamic data exists dynamically in every flight, its performance can be considered linked to equipment conditions and human and environmental factors during flight. This invention inputs this dynamic data into the fault tree for risk occurrence analysis and calculation, forming a conditional probability table. By merging static and dynamic data, the occurrence results are made more scientific and objective.

[0101] Single-dimensional QAR dynamic data anomaly detection determines anomalies based on the correlation of data within the same dimension under time series analysis. Since the temporal correlation of flight status data at different flight phases is weak, potentially leading to significant numerical differences, a sliding window algorithm is used to segment the flight time series into multiple subsequences. The data characteristics of each individual subsequence are analyzed, and correlation analysis is performed on the data characteristics of different subsequences to detect anomalies. Then, multiple linear regression is used to perform multi-dimensional anomaly detection on the anomaly subsequence data stored in the anomaly database, thereby classifying the causes of data anomalies as: equipment factors, human factors, and environmental factors. These are then input as conditional probabilities into a conditional probability table.

[0102] The dynamic sliding window method is used here. The entire flight phase is divided into takeoff, climb, cruise, approach and landing phases based on the parameters of each flight phase. The width of the sliding window is then adjusted according to the rate of change of flight data in each phase. The corresponding window widths for each phase are shown in Table 5.

[0103] Table 5. Sliding window width values ​​for each flight phase.

[0104] take off quick 20 Climb Faster 180 cruise slow 500 Approach Faster 180 landing quick 20

[0105] Segmentation can effectively extract local features for each specific flight phase, achieving dimensionality reduction of the time series and significantly reducing computational load. Based on flight principles, QAR dynamic data of relevant key components are selected, and the following data structure is defined for each data point: i = [time, value, d_means, is_outlier, type], where time is the timestamp of the current data point; value is the parameter value at the current moment; d_means is the Euclidean distance between the value of point i and the mean m(i0) of the sliding window centered on that point, with a default value of 0; is_outlier is a boolean parameter used to determine whether the current data point is an outlier, True indicates an outlier, False indicates a normal value, with a default value of False; type is the data type of the outlier, 0 indicates a component malfunction, 1 indicates other factors such as human and environmental factors, and the default value is -1, indicating non-outlier data.

[0106] The purpose of unidimensional anomaly detection is to determine the `is_outlier` attribute of each data point, establish an anomaly database, and store the detected anomaly data. The target flight parameters are scanned throughout the entire flight process, and the mean of each point and the mean of the corresponding centered window is calculated and stored in the `d_means` attribute of each data point. After the sliding window has scanned all data points, K-Means clustering analysis is performed on the `d_means` value of each data point in the sliding window segment, where the number of clusters k = 2, representing normal values ​​and anomalies respectively. The smaller cluster is considered an anomaly, its `is_outlier` attribute is set to `True`, and it is stored in the anomaly database for further multidimensional anomaly data analysis.

[0107] Multidimensional anomaly analysis requires spatial correlation based on QAR data. Therefore, it is necessary to select other flight parameters related to the flight parameters being analyzed according to flight principles and perform correlation analysis on them. The correlation between variables x and y is studied according to equation (1). r is the correlation between variables x and y, r∈[-1, 1]. The larger |r| is, the higher the correlation between x and y. |r| is weakly correlated in [0, 0.3], moderately correlated in (0.3, 0.6], and strongly correlated in (0.6, 1.0]. Based on the magnitude of the correlation, parameters with strong correlation are selected for further multiple linear regression analysis.

[0108] At this point, it is necessary to determine whether the occurrence of abnormal data is due to human and environmental factors or equipment and sensor malfunctions, and to judge whether the current data anomaly is caused by a single equipment failure or by changes in multiple parameters leading to changes in flight conditions. Since QAR dynamic data is a multivariate time series, multiple linear regression using equation (2) is used to analyze the QAR data.

[0109] In the QAR dynamic dataset, 80% of the data was randomly selected as the training set for training the model, and 20% of the data was used as the test set. The training yielded the multiple linear regression equation shown in Equation (2), which was used to fit the value of the predicted objective function at any time stamp. The predicted function value fitted by the regression equation was compared with the data in the actual normal database, and the average distance was calculated as a threshold to measure the degree of fit between the data in the abnormal database and the corresponding value of the predicted function. If the distance between the abnormal point and the predicted point is less than the threshold, it indicates that the current abnormal data is not caused by a sudden failure, but by a change in the objective flight state, and type is set to 1; if the distance between the abnormal point and the predicted point is greater than the threshold, it is considered that it cannot be fitted, indicating that the current abnormal data is an isolated sudden state, which is likely caused by a failure of the sensor or device that detects the parameter, and type is set to 0.

[0110] During the actual operation of an aircraft flight control system, component failures can alter its operational state, leading to changes in the distribution and relationships of related parameters. As the number of flight cycles increases, the performance of each functional component inevitably degrades, resulting in changes in its operational state and consequently affecting parameter variations. Functional degradation leading to changes in operational state can be categorized into two types: recoverable degradation and unrecoverable degradation. Recoverable degradation occurs when components gradually accumulate internal contaminants during operation, leading to a decrease in performance. Airlines may wash these components to remove internal contaminants and restore this performance. The other type of degradation cannot be recovered through washing, resulting in permanent degradation of component performance and different operational states. Therefore, this study uses the point at which components exhibit unrecoverable degradation after several washes as a node. Multidimensional anomaly analysis is performed on each previous QAR dynamic data point to classify the faults during each flight as being caused by human and environmental factors or equipment failure. This yields the parameter anomalies caused by human and environmental factors during actual flight, presented as a conditional probability table. The study also transforms fault trees into Bayesian networks, such as... Figure 10 As shown, the top-level event, the aircraft flight control system failure, is transformed into the parent node of the Bayesian network. The intermediate events of the second and third levels are transformed into intermediate nodes of the Bayesian network, and the bottom-level events, i.e., critical node failures, are transformed into child nodes of the Bayesian network. By assigning values ​​to these child nodes using relevant data, a complete Bayesian network can be obtained, as shown below. Figure 11 As shown, this implements bidirectional fault reasoning by combining dynamic and static data. The formula for calculating the posterior probability is as follows:

[0111]

[0112] In the formula, P(B) i ) is B i The probability of an event occurring; P(A|B) i) indicates that event A occurs in event B. i The probability of something happening given that it has already happened.

[0113] Based on the multiplication theorem and conditional probability, combined with the above formula, we can obtain:

[0114]

[0115] In the formula, P(B) j |A) represents event B j The probability that event A may occur given that event A has already occurred.

[0116] Finally, based on the posterior probability value obtained by combining the dynamic and static data, the occurrence level can be obtained according to Table 6.

[0117] Table 6 Occurrence Level Scoring Table

[0118]

[0119] Next, the components are rated based on the objective detection of the failures. This requires aircraft manufacturers and airline maintenance personnel to determine the stage of the failure mode. The rating criteria used are shown in Table 7.

[0120] Table 7 Detectability Level Rating Table

[0121]

[0122]

[0123] When applying expert experience, inconsistencies in conclusions are inevitable due to differences in the subjects' expressions. Therefore, when faced with probeness assessment information provided by different experts, the Dempster-Shafer evidence theory is used to fuse information on imprecise and uncertain data. This requires first establishing an identification framework Ω, which is a set of mutually exclusive and jointly exhaustive events. The identification framework is defined as: Ω = {almost certain, very high, high, moderately high, moderate, low, very low, tiny, very tiny, almost impossible}. Let m be from 2... Ω The quality function mapped to [0, 1] is called the basic confidence assignment, which satisfies the following equation:

[0124] and

[0125] In the formula, represents the empty set, i.e., a proposition that cannot occur; m(A) represents the body of evidence, i.e., the degree of support of an expert m for event A.

[0126] For components requiring evaluation, multiple experts assess their detectivity, providing the evidentiary information needed for information fusion. The DS evidence fusion method involves orthogonally calculating the probability functions of two or more experts, using... This indicates that experts need to make probabilistic assessments of several rating levels for the difficulty of detecting this failure mode. Within the Ω framework, two independent basic confidence levels are assigned as m1 and m2, defining the Dempster combination rule. Equation (3) is given.

[0127] Because experts have varying qualifications and experience, their scoring may differ significantly, even conflicting, leading to conclusions that contradict the correct results when using the DS combination rule for information fusion. When two or more experts evaluate the same failure mode, the evidence they provide shows some correlation and association. Therefore, the Pearson correlation coefficient is used to calculate and allocate the weight of each expert's scoring. First, a Pearson correlation coefficient matrix needs to be constructed, with m experts... i and m j Pearson correlation coefficient p ij The calculation formula is (4), then according to p ij The Pearson correlation coefficient matrix P is constructed as shown in equation (5). Next, the basic reliability assignment of the original expert ratings needs to be corrected. Due to the weighting and reliability calculation of non-positively correlated expert ratings in the identification framework, the non-positively correlated values ​​are corrected to 0.001. Other expert ratings for expert m... i The scoring conflict is Con(m) i As shown in equation (6), after normalization, we obtain the expert m. i The relative conflict is In(m) i As shown in equation (7). The original basic reliability assignment is corrected using relative conflict, and the corrected basic reliability assignment is... As shown in equation (8). After obtaining the new basic confidence level assignment, in order to avoid the problem of zero confidence in the traditional DS evidence theory, it is necessary to replace the items with a basic confidence level assignment of 0, reduce the highest probability value by 0.001, and at the same time ensure that each expert score satisfies equation (9).

[0128] Finally, the multi-expert scoring information is combined for calculation. The corrected expert scoring information is calculated, and the probability m of the target is obtained using equation (10). Based on the calculated probability value, the scoring result with the highest probability is selected as the final detection result.

[0129] Table 8 Severity Level Rating Table

[0130]

[0131]

[0132] As shown in Table 8, referring to this severity rating table, the severity assessment values ​​given by experts based on triangular fuzzy number references are used, and the concept of Uncertain Ordered Weighted Average (UOWA) operator is applied to synthesize the assessment information from different experts. This process determines the weight of each expert by comparing the degree of difference between the fuzzy number of each expert's assessment value and the average estimated fuzzy number obtained by synthesizing the opinions of different experts. This takes into account the fuzziness of expert language variables and assigns weight to each expert's response, enabling the determination of the accuracy of the severity assessment expressed by experts based on the differences between expert language variables.

[0133] A panel of n experts evaluates the severity of a certain failure mode using triangular fuzzy numbers. The evaluation level j of the i-th risk characterization parameter variable by the k-th expert is transformed into a triangular fuzzy number. For i = (1,2,3,4,5) and j = (1,2,...,n), the fuzzy number of evaluation values ​​from different experts is synthesized using the UOWA operator. The specific calculation steps are as follows:

[0134] 1. Calculate the arithmetic mean of n triangular fuzzy numbers. The calculation formula is shown in equation (11).

[0135] 2. Calculate R k and Distance measure between The calculation formula is shown in equation (12).

[0136] 3. Calculate R k and similarity between Given n triangular fuzzy numbers R k Its average The similarity is shown in equation (13).

[0137] 4. Fuzzy number aggregation and calculation of the comprehensive triangular fuzzy number R. The aggregation method of UOWA operator can be used to aggregate the fuzzy numbers of different expert evaluation values ​​to obtain the comprehensive triangular fuzzy number R as shown in equation (14).

[0138] After summarizing expert evaluation opinions using the UOWA operator, the resulting evaluation is still a triangular fuzzy number. By combining this result with the membership function, we can obtain fuzzy evaluation results for different risk characterization parameters under specific failure modes after synthesizing expert opinions.

[0139] Figure 12The thin line represents the triangular fuzzy number obtained after gathering expert opinions. Taking the x-coordinate value of its upper vertex gives the severity assessment value of the failure mode under specific risk characterization parameters.

[0140] Given that the parameters in the constructed aircraft flight control system fault risk characterization parameter system exhibit a relatively obvious hierarchical structure and differences in importance, this invention uses appropriate weighting coefficients to quantify the relative importance of different risk characterization parameters.

[0141] First, risk characterization parameters are determined using the Analytic Hierarchy Process (AHP). Based on the established risk characterization parameter system, pairwise comparisons are made between parameters to create a judgment matrix. This matrix represents the relative importance of each risk characterization parameter at this level compared to its superior counterparts. After passing a consistency test, the weight vector of each risk characterization parameter is determined. The weight vectors of different experts for each risk characterization parameter at the same level are calculated, and the arithmetic mean method is used to aggregate the experts' evaluations to determine the final weight vector for that level's risk characterization parameter. Finally, the weight value of this risk characterization parameter at the same level is weighted and calculated together with the weight value of its superior counterpart to obtain the overall weight value of that risk characterization parameter in the entire system.

[0142] In information theory, a smaller entropy value indicates a greater degree of discrepancy among the data covered by the indicator, meaning the indicator carries more information and thus a larger weight value. This invention uses the entropy weight method to determine the weight vectors of three risk characterization indicators: occurrence, detectability, and severity. Expert experience is required to score these indicators, with scores ranging from 1 to 10; higher scores indicate greater importance for the indicator.

[0143] First, we construct an initial assessment matrix. Assuming that m experts evaluate and score n risk characterization parameters, the resulting initial assessment matrix R is:

[0144]

[0145] In the formula: x ij Let be the evaluation value of the i-th expert for the j-th risk characterization parameter, i∈(1,2,...,m), j∈(1,2,...,n).

[0146] Next, to ensure the accuracy and simplicity of the calculation results, the initial evaluation matrix is ​​standardized according to the formula.

[0147]

[0148] In the formula: r ij (0≤r ij≤1) represents the evaluation standard value of the i-th expert for the j-th risk characterization parameter. The standardized evaluation matrix obtained after standardization is denoted as R. * =(r ij ) m×n .

[0149] To calculate the accurate information entropy, it is necessary to calculate, according to the formula, the proportion P of the score assigned by the i-th expert to the j-th risk characterization parameter. ij .

[0150]

[0151] Calculate the information entropy of each risk characterization parameter and the information utility value. The information entropy E of the j-th risk characterization parameter is... j The calculation formula is shown in the figure. Based on information entropy E... j The information utility value d can be calculated from the value. j =1-E j .

[0152]

[0153] The information utility value d calculated at this time j The larger the value, the more important the risk characterization parameter; therefore, its weight should also be larger when assigning weights to risk characterization parameters. Calculate the entropy weight ω of the risk characterization parameter. j The formula is shown in the figure.

[0154]

[0155] The risk characterization parameter weight vector calculated solely using the AHP (Analytic Hierarchy Process) relies heavily on the subjective preferences of the raters. While it can differentiate the importance of evaluation indicators to some extent, it suffers from significant subjectivity. In contrast, the entropy weight method, although also incorporating expert opinions in its risk characterization parameter weight vector calculation, approaches the problem from the perspective of information entropy, fully preserving the information inherent in the intrinsic structure of each indicator. Therefore, the evaluation results possess a degree of objectivity. The comprehensive risk characterization parameter weights obtained by combining both methods are more scientific and reasonable. The comprehensive risk characterization parameter weights W are determined using the AHP-entropy weight method. j The formula is:

[0156] W j =αω j (1-α)v j

[0157] In the formula: w j v jThe risk characterization parameter weights were obtained using the entropy weight method and the AHP method, respectively; α is the preference coefficient, and 0≤α≤1. After consulting with experts, the preference coefficient α was determined to be 0.6.

[0158] Therefore, the comprehensive weight vector of the risk characterization parameters can be obtained as W = (W1, W2, ..., W...). n ).

[0159] This invention distributes questionnaires to experts in the field, eliminates invalid samples, and further obtains a judgment matrix based on the questionnaire results. After passing the consistency test, the combined entropy weight method and AHP method are used to obtain the comprehensive weight of each risk characterization parameter, as shown in Table 9.

[0160] Table 9 Weights of Risk Characterization Parameters

[0161] Risk Occurrence Rate (O) 0.3 Risk detectability (D) 0.25 Risk severity (S) 0.45

[0162] Finally, linear weighting yields the optimized RPN value for the final result. Taking the control surface centering mechanism and the flap control handle as examples, the control surface centering mechanism has an occurrence degree of 4, a detection degree of 3, and a severity degree of 6, resulting in an optimized RPN value of 4.65; while the flap control handle has an occurrence degree of 3, a detection degree of 5, and a severity degree of 5, resulting in an optimized RPN value of 4.4. Therefore, it can be determined that the maintenance priority of the control surface centering mechanism is higher than that of the flap control handle.

Claims

1. A two-way reasoning assessment method for civil aircraft failure risk, characterized in that: The method calculates the occurrence degree, detectability, and severity of risk characterization parameters, and uses a combination of dynamic and static data to perform bidirectional reasoning on failure risk. It assigns corresponding weights to the three risk characterization parameters and calculates and optimizes the RPN value using a weighted average. The method includes the following steps: Step 1: Calculation of the occurrence degree of risk characterization parameters: Divide the structural hierarchy of the aircraft flight control system, analyze the cause of failure for each structural hierarchy, obtain static data, and construct a fault tree structure diagram; The QAR dynamic data of the aircraft flight control system is analyzed, and the dynamic and static data are fused to generate a Bayesian network to calculate the posterior probability. Step 2: Calculation of the detectability of the risk characterization parameter: Based on the detection status scoring table, the fault occurrence stage is judged, and the information on the fault occurrence stage given by different experts is integrated to resolve the problem of conflicting evidence on the fault occurrence stage. Step 3: Calculation of the severity of the risk characterization parameter: Assign corresponding weights to the fuzzy linguistic variables of the severity of the failure as determined by different experts, and determine the severity of the failure using the maximum membership principle based on the triangular fuzzy number function; Step 4: Determining the weights of risk characterization parameters: Using the AHP-entropy weight method, the importance of the three risk characterization parameters—occurrence, detectability, and severity—in the risk assessment of flight control system failure is distinguished, the weights of the three risk characterization parameters are determined, the role of different characterization parameters in risk assessment is measured, and the weighted calculation optimizes the RPN value. In step one, the division of the aircraft flight control system structure hierarchy is based on the aircraft maintenance manual and the fault isolation manual to determine the hierarchical structure of the civil aircraft flight control system. The reverse FTF analysis method of FMECA+FTA is applied to sort out the key nodes in the underlying events of the flight control system faults to form a set of important key underlying events. The analysis of QAR dynamic data of the aircraft flight control system involves using the K-Means clustering algorithm to identify abnormal data and using multiple linear regression to classify abnormal situations into human and environmental factors or equipment factors. A conditional probability table of a Bayesian network is then constructed. A complete Bayesian network is built through the transformation relationship between the constructed fault tree and the Bayesian network, achieving an organic combination of dynamic and static data to obtain the posterior probability under actual flight conditions. The occurrence degree of the risk characterization parameter is then obtained by comparing the posterior probability scoring table. Constructing the conditional probability table for a Bayesian network includes the following steps: Step A1: Using a dynamic sliding window algorithm, the sliding window width is set according to the rate of change of QAR dynamic data in each flight phase of the aircraft flight control system, and the flight phase is divided into five phases: takeoff, climb, cruise, approach, and landing; data analysis is performed according to the different data change characteristics of each of the five phases. Step A2: For each flight phase, use the K-Means clustering algorithm to divide the data into two categories: abnormal data and normal data. Store the abnormal data in the abnormal database for multidimensional abnormal data analysis in the next step. Step A3: Based on the principles of flight, select flight parameters relevant to the flight parameters being analyzed, and use the following formula to calculate the correlation between the relevant flight parameter variables x and y: ； In the formula: , are the sample means of variables x and y, respectively; r represents the correlation between variables x and y; based on the correlation, flight parameters with strong correlation are selected for further multiple linear regression analysis; Step A4: Analyze the QAR dynamic data in the anomaly database using the following multiple linear regression equation to obtain the QAR dynamic dataset: ； In the formula: Y is the time series vector of the objective function; This represents the value of flight parameter i at time j; b is the regression coefficient vector; b is the constant term; The QAR dynamic dataset is trained to fit the target function. The values ​​of the predicted objective function at any time stamp are compared with the data in the actual normal database. The average distance between the two values ​​is used as a threshold. If the distance between the outlier and the predicted point is less than the threshold, it indicates that the current abnormal data is not caused by a sudden failure, but by a change in the objective flight state. If the distance between the outlier and the predicted point is greater than the threshold, it indicates that the current abnormal data is an isolated sudden state caused by a failure of the sensor or equipment that detects the flight parameters. Step A5: Take the irreversible degradation of the aircraft flight control system components as a node, collect and analyze the QAR dynamic data before each node, classify the cause of the data anomaly as human and environmental factors or equipment failure, use this statistical data as the conditional probability table of the Bayesian network, and transform the fault tree and Bayesian network to obtain the posterior probability of the dynamic and static data, and score the occurrence level according to the posterior probability scoring table. In step two, to resolve the conflict of evidence information during the failure occurrence phase, the weighting of each expert's score during the failure occurrence phase is calculated and allocated, and the basic reliability allocation of the original expert scores is corrected. This includes the following steps: Step B1: Perform orthogonal calculations on the basic reliability allocation probability functions of two or more experts, using... The expert stated that they would make a probabilistic assessment of several rating levels for evaluating the difficulty of detecting failure modes within the identification framework. The two independent basic confidence levels are assigned to m1 and m2, and the Dempster combination rule is defined. for: ； In the formula, X i Y represents the first expert's assessment of the detection performance for the i-th rating level. i The assessment of the detection situation for the second expert at the i-th rating level; 1 / (1-k) is the normalization factor; k is the conflict coefficient, representing X. i With Y i The size of the conflict between them A represents the component failure mode that requires an assessment of the detection status. Step B2: Next, the Pearson correlation coefficient is used to calculate and assign the weight proportions of each expert's score in order to resolve the problem of conflicting evidence when assessing the failure occurrence stage. First, construct the Pearson correlation coefficient matrix, and then the basic confidence assignment made by the i-th expert is m. i The basic reliability assignment m made by the j-th expert j Pearson correlation coefficient p ij The calculation formula is: ； In the formula: The reliability assignment m made by the experts i and m j covariance; and Assign m to the basic reliability respectively i and m j Standard deviation; Step B3: Based on the Pearson correlation coefficient p ij To construct the Pearson correlation coefficient matrix P: ； Step B4: Next, the basic reliability assignment of the original expert ratings is corrected. Based on the weight ratio and reliability of non-positively correlated expert ratings in the identification framework, the non-positively correlated values ​​are corrected to 0.001; the basic reliability assignments m of other expert ratings for expert i are also corrected. i The scoring conflict is Con(m) i ): ； Step B5: After normalizing the conflicting evidence information at the fault occurrence stage in the expert scoring, the basic reliability assignment m made by expert i is obtained. i The relative conflict is In(m) i ): ； Step B6: Utilize relative conflict In(m) i The original basic reliability allocation is corrected, and the corrected basic reliability allocation is as follows: : ； Step B7: After obtaining the corrected basic reliability assignment, replace the items with a basic reliability assignment of 0 by reducing the highest probability value by 0.001, while ensuring that each expert rating satisfies: ； Step B8: Finally, the modified multi-expert rating information is used to calculate the basic reliability distribution of the target, i.e., the probability value m: ； Based on the calculated probability values, the score result with the highest probability is selected as the final risk characterization parameter detectivity result.

2. The two-way reasoning assessment method for civil aircraft failure risk according to claim 1, characterized in that: In step three, the severity of the fault is determined using the maximum membership principle based on the triangular fuzzy number function; the severity assessment values ​​given by experts based on the triangular fuzzy number reference are used to synthesize the assessment information of different experts using the uncertain ordered weighted average operator; and the final severity score, the risk characterization parameter, is determined using the maximum membership principle.

3. The two-way reasoning assessment method for civil aircraft failure risk according to claim 2, characterized in that: In step three, the method for applying the uncertain ordered weighted average operator to determine expert weights includes the following steps: Step C1: Calculate the arithmetic mean of the n triangular fuzzy numbers. The calculation formula is as follows: ； in: ; Suppose there are n experts evaluating the severity of a certain type of failure mode, R. m Let be the average triangular fuzzy number converted by k experts into the assessment level j of the i-th risk characterization parameter variable, where i represents the severity score for the required assessment failure mode, and a, b, and c represent the corresponding values ​​of the left, top, and right vertices of the triangular fuzzy number given by the experts, respectively. Step C2: Calculate the triangular fuzzy number R k Arithmetic mean of triangular fuzzy numbers Distance measure between The calculation formula is as follows: ； In the formula R k This represents the triangular fuzzy number representing the assessment level j of the i-th risk characterization parameter variable transformed by the k-th expert, i.e. ; Step C3: Calculate the triangular fuzzy number R k and the arithmetic mean of triangular fuzzy numbers similarity between , n triangular fuzzy numbers R k Arithmetic mean of triangular fuzzy numbers The similarity is considered as: ； Step C4: Perform fuzzy number aggregation and calculate the comprehensive triangular fuzzy number R; use the aggregation method of UOWA operators to aggregate the fuzzy numbers of different expert evaluation values ​​to obtain the comprehensive triangular fuzzy number R as follows: ； In the formula This represents the triangular fuzzy number evaluation results given by the 1st to nth experts. This represents the weight of the evaluation result of the k-th expert obtained under the UOWA operator in the overall evaluation result; Step C5: Take the x-coordinate value of the vertex of the comprehensive triangular fuzzy number R, and regard it as the severity assessment value of the failure mode under the specific risk characterization parameters.

Images