An aviation maintenance risk assessment and intelligent diagnosis method based on a multi-dimensional Bayesian network

Abstract

The application discloses a kind of aviation maintenance risk assessment and intelligent diagnosis method based on multidimensional bayesian network, belong to aviation maintenance safety and data analysis technical field.The existing risk analysis method is aimed at solving the problem that there is dependence on expert subjective experience, it is difficult to represent the nonlinear coupling relationship of multidimensional factor, cannot effectively extract unstructured text risk information, lack of dynamic early warning and accident cause accurate tracing ability.Proceeding with the pretreatment and inspection of maintenance text;Cascade network node of four dimensions of man, machine, environment and management is constructed;Risk Boolean matrix is generated using keyword matching;Under the constraint of forbidden edge, structure learning is carried out, and parameter estimation is executed;After using AUC to evaluate model, forward risk prediction and cause reverse diagnosis are executed.It can realize the mapping from unstructured text to quantitative causal model, effectively represent complex nonlinear coupling relationship, realize dynamic early warning and accurate tracing, improve the objectivity and reliability of aviation maintenance safety management.

Description

Technical Field

[0001] This invention belongs to the field of aviation maintenance safety and data analysis technology, specifically relating to an aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks. Background Technology

[0002] Aircraft maintenance is a crucial link in ensuring flight safety, and its quality directly impacts aircraft airworthiness and operational safety. The aircraft maintenance process involves numerous complex human operations, equipment conditions, and management factors, which are interconnected and form complex causal relationships. Currently, aircraft maintenance risk analysis primarily employs fault tree analysis, event tree analysis, failure mode and effects analysis, and hazard and operability studies. While these methods provide a structured analytical framework for risk management, they rely excessively on expert experience, and the extraction of risk factors largely depends on manual analysis, which is not only inefficient but also makes it difficult to guarantee the objectivity and consistency of the analysis results. Furthermore, these methods typically assume that risk factors are independent or have linear relationships, making it difficult to capture the complex nonlinear correlations and uncertainties in aircraft maintenance accidents. In addition, existing methods are mostly post-event analyses, lacking the ability to predict accident risks. When maintenance conditions change, it is difficult to quickly assess the dynamic changes in the system risk level, failing to provide timely early warning information for safety management personnel.

[0003] The "Man-Machine-Environment-Management" (HMI-MI) theory is a widely used systems analysis framework in the field of safety engineering, emphasizing that accidents are the result of the combined effects of multiple factors, including personnel, machinery and equipment, the working environment, and management systems. However, existing methods have failed to organically integrate the risk factors of these four dimensions into a unified probabilistic model, making it difficult to achieve multi-dimensional risk assessment and intelligent diagnosis. Furthermore, with the continuous accumulation of aviation accident databases and maintenance report data, how to automatically extract risk factors from various dimensions using text data and construct data-driven Bayesian network models has become a crucial research problem urgently needing to be solved in the field of aviation safety. Bayesian networks, as a probabilistic graphical model, can intuitively represent the dependencies between variables, quantify these relationships through conditional probability tables, and support forward prediction and backward diagnostic reasoning. However, the application of traditional Bayesian network methods in the field of aviation maintenance still faces challenges such as network structure dependence on expert experience, difficulty in handling text data, and the need for a large number of complete samples for parameter learning.

[0004] Based on this, the present invention proposes an aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks to solve the problems existing in the prior art. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks. This method addresses the issues of existing risk analysis methods, such as reliance on expert subjective experience, difficulty in characterizing the nonlinear coupling relationships of multi-dimensional factors, inability to effectively extract risk information from unstructured text, and lack of dynamic early warning and accurate traceability of accident causes.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for aircraft maintenance risk assessment and intelligent diagnosis based on multi-dimensional Bayesian networks includes: Step 1: Acquire and preprocess the data to generate a standardized text set; Step 2: Based on the "human-machine-environment-management" theory in the field of aviation maintenance, the risk factors and outcome variables of the standardized text set are abstracted into a four-layer cascaded Bayesian network node structure; Step 3: Based on the nodes defined in Step 2, construct a corresponding keyword set for each node; Step 4: Use keyword matching algorithms to scan standardized text to identify risk factors in accident records; Step 5: Define the set of forbidden edge rules for the Bayesian network structure; Step 6: Under the premise of satisfying causal constraints, perform Bayesian network structure learning to construct a Bayesian network model of aviation maintenance safety risks; Step 7: After the topology is determined, calculate the conditional probability distribution of each node in the Bayesian network using the parameter learning method; Step 8: Evaluate the predictive performance of the Bayesian network model using the 5-fold cross-validation method; Step 9: After completing model training and validation, perform dynamic risk prediction using probabilistic inference methods; Step 10: When the outcome of the accident is known, calculate the posterior probability and magnitude of the risk factors through Bayesian inference to diagnose the cause of the accident.

[0007] In a preferred embodiment, step 1, which involves acquiring and preprocessing data, includes: Obtain the accident record dataset: ; in, This represents the total number of accident samples. Indicates the first There are 10 accident records, and in each record... All include detailed accident description text. ; Accident description text Denoising is performed to generate a standardized set of accident texts: ; in, Indicates the first An accident record is a pre-processed accident text.

[0008] In a preferred embodiment, the four-layer cascaded Bayesian network topology described in step 2 is divided from top to bottom as follows: L1 Objective Facts Layer: As the root node layer, it contains basic attribute nodes for personnel, machinery, environment, and management dimensions; L2 Operational Behavior Layer: Contains nodes reflecting direct behavioral deviations during personnel maintenance operations; L3 intermediate state layer: contains nodes that characterize physical or system abnormal states that occur during maintenance; L4 Risk Outcome Layer: Contains risk level nodes classified according to civil aviation safety standards.

[0009] In a preferred embodiment, step 4, scanning the standardized text using a keyword matching algorithm, includes: When the Standardized accident text The occurrence of keywords in the set When the matching words are found, there are risk factors in the incident record. Each accident record is represented as a risk factor vector: ; in: ; x ij Indicates the first i Does the standardized accident text contain the first... j One risk factor, i For accident text indexing, j Index for risk factors; After this operation, each accident record is transformed into a high-dimensional Boolean vector, and all vectors converge to form a risk factor data matrix: .

[0010] In a preferred embodiment, the edge prohibition rule in step 5 includes: Lower-level nodes are prohibited from pointing to upper-level nodes, nodes at the same level are prohibited from connecting to each other, and nodes at the risk result level are prohibited from pointing to any other level of nodes.

[0011] In a preferred embodiment, the process of establishing the Bayesian network model in step 6 includes: Step 6.1: Constructing a Bayesian network can be represented as a directed acyclic graph: ; in, This represents a set of nodes, corresponding to various risk factors. Edges representing causal relationships between nodes; The joint probability distribution of a Bayesian network is expressed as: ; in, Represents a node The set of parent nodes; Step 6.2: A structure learning method based on the scoring function is adopted to search the network structure by adding, deleting, and reversing edges, and a hill climbing algorithm is used to find the network structure with the largest scoring function. The scoring function uses the BDeu score, which measures the quality of a structure by calculating the marginal likelihood of its occurrence in a dataset given the structure. Its basic form is as follows: ; in, In network structure Dataset under conditions The probability of occurrence; A random restart mechanism is introduced during the search process. After reaching a local peak, the current structure is randomly perturbed and the search is restarted. .

[0012] In a preferred embodiment, when calculating the conditional probability distribution of each node in the Bayesian network in step 7, parameter learning is performed using the expectation-maximization algorithm. This involves alternately executing expectation and maximization steps, estimating the distribution of latent variables using known observation data, and continuously updating the parameters until convergence. Specifically, this includes: By alternating the desired steps and maximizing steps To estimate model parameters; In the expected step In the middle, the expected sufficient statistic of the latent variables is calculated based on the current parameter estimates; in the maximization step... In this process, the expected log-likelihood function is maximized to update the parameter estimates; Repeat the above two steps until the parameters converge or the preset number of iterations is reached, and finally obtain the conditional probability table of each node in the Bayesian network.

[0013] In a preferred embodiment, step 8, which evaluates the predictive performance of the model using a 5-fold cross-validation method, includes: Step 8.1: Randomly divide the dataset into... A subset; Step 8.2: Select one subset at a time as the test set, and the rest... A subset is used as the training set; Step 8.3: Repeat step 8 on each training set to build the model; Step 8.4: Perform tests on the test set for each risk level node. Make a prediction and obtain the predicted probability. and real labels ; Step 8.5: Use As an evaluation metric, let the positive examples in the test set be... The set is , negative example for ,but The calculation formula is: ; in, , These represent the sample sizes of positive and negative examples, respectively. , These represent the indices of the positive and negative samples, respectively. , These represent the model's prediction of the first... , Each sample represents the probability of the risk occurring; This indicates an indicator function that takes the value 1 when the condition is true and 0 otherwise. The closer the value is to 1, the better the ranking performance of the model.

[0014] In a preferred embodiment, step 9, which uses probabilistic reasoning to perform dynamic risk prediction, includes: For a given combination of risk factors The probability of the accident occurring is: ; in, Indicates the L4 layer risk level node. This indicates that an accident has occurred; Indicates the corresponding given The values ​​of each risk factor; This represents the joint probability of a risk occurring simultaneously when all risk factors take given values. This represents the probability that all risk factors will occur simultaneously when each factor takes a given value.

[0015] In a preferred embodiment, the posterior probability of the risk factor in step 10 is: ; in, Indicates risk factors The prior probability, Indicates risk factors under the conditions of an accident. The posterior probability; By comparing the increase in posterior probability with that in prior probability: ; The key factors that lead to this risk and their respective dimensions can be identified.

[0016] Compared with existing technologies, this invention provides an aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks, which has the following beneficial effects: 1. This invention extracts risk features directly from massive unstructured maintenance reports by constructing a standardized text preprocessing process and a keyword matching algorithm based on a domain thesaurus. This replaces the traditional manual expert scoring mode, eliminates the influence of subjective bias on the assessment results, and makes the risk identification process highly repeatable and consistent.

[0017] 2. This invention is the first to organically integrate the objective elements of the four dimensions of "human-machine-environment-management" with operational behavior, intermediate state, and risk outcome through a four-layer Bayesian network. Compared with traditional single-dimensional analysis or simple logical weighted models, this invention can accurately characterize the nonlinear causal chain in aviation maintenance and greatly improve the accuracy of characterizing the causes of complex accidents.

[0018] 3. This invention introduces a set of forbidden edges during the structure learning process, forcing the model to follow the inherent physical laws of aviation maintenance. This method, which combines data-driven approaches with knowledge constraints, overcomes the "pseudo-correlation" problem that is easily generated by pure data mining models, ensuring that the generated risk assessment logic conforms to engineering practice.

[0019] 4. This invention not only utilizes forward probabilistic reasoning to provide real-time quantitative early warning of risk levels in specific maintenance scenarios, offering decision support for safety management personnel; it also leverages posterior probabilistic reasoning and probability enhancement indicators to quickly identify the core causes from dozens of potential factors after an accident occurs. This bidirectional reasoning capability provides a unified technical support platform for "prevention, control, and tracing" in aircraft maintenance.

[0020] 5. This invention addresses the common problem of missing data in maintenance records by employing the EM algorithm and introduces BDeu scoring and random restart hill climbing algorithm to optimize the network topology. Even in the face of real-world aviation maintenance environments with incomplete samples and high data noise, it can still maintain stable performance. The introduction of the AUC evaluation index provides a quantitative standard for the online evolution and self-iteration of the model, ensuring the reliability of the system in long-term operation.

[0021] 6. This invention, through the deep integration of technical means, transforms traditional qualitative experience analysis into quantitative probabilistic reasoning, providing a scientific, efficient, and highly interpretable intelligent tool for aviation maintenance safety management. It has significant engineering application value for improving the level of civil aviation maintenance quality control and preventing major safety accidents.

[0022] It solves the problems of existing risk analysis methods, such as reliance on expert subjective experience, difficulty in representing the nonlinear coupling relationship of multi-dimensional factors, inability to effectively extract risk information from unstructured text, and lack of dynamic early warning and accurate traceability of accident causes. Attached Figure Description

[0023] Figure 1 This is a flowchart of the aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks, which is based on the present invention.

[0024] Figure 2 This is a diagram of the Bayesian network structure of the present invention.

[0025] Figure 3 This is a graph showing the increase in the L1 factor when various risks of the present invention occur. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] As a complex and large-scale system engineering project, aircraft maintenance safety is influenced by multiple dimensions, including personnel competence, machine condition, operating environment, and management efficiency. To achieve accurate quantification and intelligent diagnosis of aircraft maintenance risks, this invention provides an aircraft maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks.

[0028] Please see Figure 1 The overall execution flow of this method covers the entire process from mining raw text data to building and training a Bayesian network model, and finally, inference prediction and reverse diagnosis. Specifically, it includes the following steps: Step 1: Obtain data and preprocess the data; Obtain the accident record dataset from the airline maintenance management system or the civil aviation safety information network: ; Among them, represents the total number of accident samples, represents the th accident record, and each record contains detailed accident description text , which records the background, process, components involved, and final consequences of the accident in natural language form; Due to the existence of a large amount of noise, non-standard expressions, and redundant information in the original accident description text , in this step, irrelevant special symbols are removed by performing text cleaning operations; different forms of the same term are eliminated by case unification; the text is split into semantic tokens using a tokenization algorithm. In particular, the system performs a negative word check operation. By identifying trigger words such as "not", "no", "without", etc., and combining dependency syntactic analysis to determine the modified object of the negative word, it is possible to avoid misjudging "no crack found" as a risk factor. After the above processing, a preprocessed standardized text set is generated: ; Among them, is the accident text after preprocessing for the th accident record.

[0029] Step 2: According to the "human - machine - environment - management" theory in the field of aviation maintenance, abstract the risk factors and outcome variables of the standardized text set into a four - layer cascaded Bayesian network node structure; Please refer to Figure 2 , and this topological structure is divided into: L1 objective fact layer, L2 operation behavior layer, L3 intermediate state layer, and L4 risk outcome layer from top to bottom.

[0030] L1 objective fact layer: As the root node layer, it contains 9 basic dimension nodes that affect maintenance safety, specifically including the working years of maintenance personnel (divided into greater than or equal to 3 years and less than 3 years) and continuous working hours (divided into within 8 hours and more than 8 hours) in the personnel dimension; aircraft type, task type (routine or non - routine), and aircraft age in the mechanical dimension; maintenance time (day or night) and weather conditions in the environment dimension; whether there is job pressure and whether there is cross - shift handover in the management dimension. As shown in Table 1.

[0031] Table 1: Node definitions of L1 layer

[0032] L2 Operational Behavior Layer: Reflects direct behavioral deviations during personnel operation, including five nodes: failure to operate according to the manual, improper use of tools and equipment, omission of steps, communication failure, and violation of operating procedures.

[0033] L3 intermediate state layer: Characterizes the physical or system abnormal state that occurs during the maintenance process, including 5 nodes: installation connection error, fastener problem, foreign object left behind, abnormal system parameters, and component damage.

[0034] L4 Risk Outcome Layer: The final risk outcome layer R, based on civil aviation safety standards, classifies risks into four levels: Category I (catastrophic), Category II (lethal), Category III (critical), and Category IV (mild). The final outcome is divided into four risk levels, and the classification criteria are shown in Table 2.

[0035] Table 2: Risk Outcome Level Classification Criteria

[0036] This step, through the hierarchical division method described above, effectively isolates the causal relationship between the preceding causes and the subsequent results using physical logic.

[0037] Step 3: Based on the 23 nodes defined in Step 2, each node in the L1 objective fact layer... Construct the corresponding keyword set K j ; The selection of keywords combines civil aviation maintenance technical manuals, standard operating procedures (SOPs), and expert experience. For example, for the "foreign object left behind" node, the keyword set includes terms such as "extraneous object," "fallen," and "forgotten in the cabin"; for the "operational pressure" node, keywords include "rushing," "tight schedule," and "delay pressure." These keyword sets constitute a bridge for the transformation from unstructured text to structured data.

[0038] For example: Let L1 have a total of Each risk factor node is denoted as... Each node Corresponding to a set of keywords:

[0039] These keywords are used to identify the presence of the risk factor from accident texts.

[0040] Step 4: Then, use a keyword matching algorithm to standardize the text. A scan is performed to identify risk factors present in the accident records; When the Standardized accident text The occurrence of keywords in the set If a matching word is found, then the accident record is considered to contain risk factors. Therefore, each accident record can be represented as a risk factor vector: ; in: (1); x ij Indicates the first i Does the standardized accident text contain the first... j One risk factor, i For accident text indexing, j Index for risk factors; After this operation, each accident record is transformed into a high-dimensional Boolean vector, and all vectors converge to form a risk factor data matrix: ; This step transforms fragmented textual descriptions into structured samples that can be processed by mathematical models.

[0041] Step 5: To ensure that the Bayesian network structure conforms to the physical logic of aircraft maintenance, define a set of prohibited edge rules; the prohibited edge rules include: (1) Lower-level nodes are prohibited from pointing to upper-level nodes. That is, L2 level nodes cannot point to L1 level nodes, L3 level nodes cannot point to L1 or L2 level nodes, and L4 level nodes cannot point to L1, L2 or L3 level nodes, so as to ensure that the causal direction is passed from the basic conditions to the downstream results. (2) Prohibit nodes at the same level from connecting to each other. Within the same dimension or the same level, assume that there are no direct causal edges between factors to avoid circular dependencies. (3) Nodes in the risk result layer L4 are prohibited from pointing to nodes in any other layer. The risk result is the final output and should not be used as the cause of other nodes. By imposing these constraints, the search space of Bayesian networks is limited to a physically meaningful topological range.

[0042] Step 6: Under the premise of satisfying causal constraints, perform Bayesian network structure learning to construct a Bayesian network model of aviation maintenance safety risks; Step 6.1: Constructing a Bayesian network can be represented as a directed acyclic graph: ; in, This represents a set of nodes, corresponding to various risk factors. Edges representing causal relationships between nodes; The joint probability distribution of a Bayesian network can be represented as: (2); in, Represents a node The set of parent nodes; Step 6.2: To determine the network structure, a structure learning method based on the scoring function is adopted. The network structure is searched by adding, deleting, and reversing edges, and a hill-climbing algorithm is used to find the network structure with the largest scoring function. The scoring function uses the BDeu score, which measures the quality of a structure by calculating the marginal likelihood of its occurrence in a dataset given the structure. Its basic form is as follows: (3); in, In network structure Dataset under conditions The probability of occurrence is used to measure the network structure's ability to interpret data.

[0043] During the search process, the algorithm adds, deletes, or reverses edges, and calculates score changes in real time. To prevent the algorithm from getting stuck in local optima, this step introduces a random restart mechanism. After reaching a local peak, the current structure is randomly perturbed and the search restarts. (4); The final determined topology is as follows Figure 2 As shown in the figure, the complex dependency paths between various risk factors are illustrated.

[0044] Step 7: After the topology is determined, calculate the conditional probability distribution of each node in the Bayesian network using the parameter learning method; Because aircraft maintenance reports may contain incomplete information, this step employs the Expectation Maximization (EM) algorithm for parameter learning. By alternately executing the expectation step and the maximization step, the distribution of latent variables is estimated using known observation data, and the parameters are continuously updated until convergence. This algorithm is an iterative optimization method that alternates between the expectation step and the maximization step. and maximizing steps To estimate model parameters; in the expectation step In the middle, the expected sufficient statistic of the latent variables is calculated based on the current parameter estimates; in the maximization step... In this process, the expected log-likelihood function is maximized to update the parameter estimates; the above two steps are repeated until the parameters converge or the preset number of iterations is reached, and finally the conditional probability table of each node in the Bayesian network is obtained.

[0045] To improve the robustness of the model, this step introduces Laplace smoothing in parameter estimation, which effectively solves the zero probability problem under small sample conditions.

[0046] Step 8: Use the 5-fold cross-validation method to comprehensively evaluate the predictive performance of the Bayesian network model; the specific steps are as follows: Step 8.1: Randomly divide the dataset into... A subset; Step 8.2: Select one subset at a time as the test set, and the rest... A subset is used as the training set; Step 8.3: Repeat step 8 on each training set to build the model; Step 8.4: Perform tests on the test set for each risk level node. Make a prediction and obtain the predicted probability. and real labels ; Step 8.5: Use As an evaluation metric, let the positive examples in the test set be... The set is , negative example for ,but The calculation formula is: (5); in, , These represent the sample sizes of positive and negative examples, respectively. , These represent the indices of the positive and negative samples, respectively. , These represent the model's prediction of the first... , Each sample represents the probability of the risk occurring; This indicates an indicator function that takes the value 1 when the condition is true and 0 otherwise. The closer the value is to 1, the better the ranking performance of the model.

[0047] Step 9: After completing model training and validation, perform dynamic risk prediction using probabilistic inference methods; For a given combination of risk factors The probability of the accident occurring is: ; in, Indicates the L4 layer risk level node. This indicates that an accident has occurred; Corresponding to the given The values ​​of each risk factor; This represents the joint probability of a risk occurring simultaneously when all risk factors take given values. This represents the probability that all risk factors will occur simultaneously when each factor takes a given value.

[0048] In this way, safety managers can reduce the overall risk level of the system before the task begins by adjusting controllable factors at the L1 level (such as replacing senior personnel or relieving operational pressure).

[0049] Step 10: When the outcome of the accident is known, calculate the posterior probability and magnitude of the risk factors using Bayesian inference to diagnose the cause of the accident. (6); in, Indicates risk factors The prior probability, Indicates risk factors under the conditions of an accident. The posterior probability; By comparing the increase in posterior probability with that in prior probability: ; It can identify the key factors leading to the risk and their respective dimensions. Through this quantitative diagnostic process, the main contradictions can be automatically identified from a number of potential factors.

[0050] Example 1: To verify the effectiveness of the aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks described in this invention, verification is performed in this embodiment. This includes: Step 1: Taking maintenance accident report data from a certain airline as an example, a total of 5,000 accident records were obtained. Each record contains an accident description text and information such as the accident consequences. The accident description text was preprocessed by text cleaning, case unification, negative word checking, and sentence segmentation to obtain a standardized text set.

[0051] Step 2: Define 23 nodes and their hierarchical structure, including 9 nodes in the L1 layer, 5 nodes in the L2 layer, 5 nodes in the L3 layer, and 4 nodes in the L4 layer. All nodes are binary variables.

[0052] Step 3: Based on the nodes defined in Step 2, construct a corresponding set of keywords for each risk factor.

[0053] Step 4: Using the keyword matching algorithm of formula (1), scan 5000 preprocessed accident texts to generate a risk factor data matrix. Based on the L4 level risk level labeling, the sample distribution of each risk level is statistically analyzed, as shown in Table 3.

[0054] Table 3: Statistics on Risk Level Distribution

[0055] Step 5: Set 403 prohibited edges according to the rules to ensure that the causal direction is unidirectionally transmitted from L1 to L4.

[0056] Step 6: Use the BDeu scoring formula (3) and hill-climbing algorithm to learn the network structure and set the equivalent sample size. With a maximum in-degree limit of 5 and a forbidden edge constraint applied, a directed acyclic graph with 53 directed edges is obtained, as shown below. Figure 2 As shown.

[0057] Step 7: Calculate the conditional probability table for each node using the EM algorithm. Taking the "Not operated according to the manual" node as an example, its parent nodes are "Task type" and "Whether there is job pressure", and the conditional probabilities are shown in Table 4.

[0058] Table 4: Conditional Probability Table of "Failure to Follow Manual" Node

[0059] The above quantitative results profoundly reveal the coupled influence of management factors and task characteristics on personnel behavioral deviations.

[0060] Step 8: Using 5-fold cross-validation, the 5000 data points are randomly divided into 5 parts. One part is taken as the test set and the remaining 4 parts are taken as the training set. Step 7 is repeated. The AUC value is calculated for each risk level node according to formula (5). The results are shown in Table 5.

[0061] Table 5: AUC Assessment Results for Each Risk Level

[0062] The model demonstrates good predictive ability across the four risk levels, effectively identifying safe and high-risk states.

[0063] Step 9: Set different combinations of L1 layer nodes and predict the probability of each risk level in L4 layer according to formula (6). Set the following six different scenarios and consider them from the four dimensions of "human-machine-environment-management". The prediction results are shown in Table 6.

[0064] Table 6: Risk Level Probability in Different Scenarios

[0065] The results above show that when only human factors are at high risk, the probability of Type I risk is 14.6%; however, when all four dimensions—human, machine, environment, and management—are simultaneously in an unfavorable state, the probability of Type I risk surges. This indicates that aviation maintenance risk has a typical multi-factor cumulative effect, and this invention accurately characterizes this cumulative risk through the product form of a Bayesian network.

[0066] Step 10: Given “the occurrence of a Class I catastrophic risk” as evidence, calculate the posterior probability of each L1 factor according to formulas (7) and (8), and calculate the increase. The results are shown in Table 7.

[0067] Table 7: Posterior probability and magnitude of increase of L1 factor when Type I risk occurs

[0068] The results show that, in addition to seniority and stress, task type (increase of 0.083) and continuous working hours (increase of 0.070) also exhibited high correlations. In contrast, weather conditions and whether shifts were cross-teams showed smaller increases. This finding provides direct evidence for airlines to optimize safety resource allocation: prioritizing training new employees and optimizing shift scheduling to alleviate rush work pressure is more effective in reducing the risk of major accidents than simply improving lighting or weather monitoring.

[0069] Compare the increase in each L1 factor under different risk levels, such as Figure 3 As shown.

[0070] The increase in the magnitude of the same factor increases with the risk level. Under Level IV (slight risk), all factors are negative, and the actual occurrence frequency of various risk factors is lower than the average level, which is in line with the expected safety status. As the risk level increases, the increase in maintenance personnel seniority, work pressure, and task type increases significantly, all exceeding 0.08 under Category I disaster risk, while the increase in factors such as continuous working hours and aircraft age is relatively small.

[0071] From the perspective of human, machine, environment and management, personnel and management factors are the main driving factors for the occurrence of risks, while equipment and environmental factors have a relatively weak impact.

[0072] Based on the above results, it can be seen that: This invention constructs a closed-loop evaluation system from underlying textual features to high-level causal logic. The system eliminates the uncertainty of manual risk factor identification by performing automated text preprocessing and keyword matching; ensures the scientific validity of risk propagation paths by performing structure learning with physical constraints; and achieves the organic unity of risk prevention and accident diagnosis by performing bidirectional probabilistic reasoning.

[0073] In actual operation, this system can be integrated into the airline's safety management platform. Once a maintenance plan is released, the system automatically retrieves L1-level information such as personnel qualifications, aircraft status, weather forecasts, and planned schedules, calculating and displaying a risk heatmap in real time. If the predicted probability of a Class I or Class II risk exceeds the warning threshold, the system automatically triggers the warning module, recommending management intervention. During the accident investigation phase, the investigation team only needs to input the observed phenomena into the model, and the system can automatically generate a diagnostic report, identifying the most likely causal chain, thereby significantly improving the intelligence level of safety management.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for aircraft maintenance risk assessment and intelligent diagnosis based on multi-dimensional Bayesian networks, characterized in that, include: Step 1: Acquire and preprocess the data to generate a standardized text set; Step 2: Based on the "human-machine-environment-management" theory in the field of aviation maintenance, the risk factors and outcome variables of the standardized text set are abstracted into a four-layer cascaded Bayesian network node structure; Step 3: Based on the nodes defined in Step 2, construct a corresponding keyword set for each node; Step 4: Use keyword matching algorithms to scan standardized text to identify risk factors in accident records; Step 5: Define the set of forbidden edge rules for the Bayesian network structure; Step 6: Under the premise of satisfying causal constraints, perform Bayesian network structure learning to construct a Bayesian network model of aviation maintenance safety risks; Step 7: After the topology is determined, calculate the conditional probability distribution of each node in the Bayesian network using the parameter learning method; Step 8: Evaluate the predictive performance of the Bayesian network model using the 5-fold cross-validation method; Step 9: After completing model training and validation, perform dynamic risk prediction using probabilistic inference methods; Step 10: When the outcome of the accident is known, calculate the posterior probability and magnitude of the risk factors through Bayesian inference to diagnose the cause of the accident.

2. The aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks as described in claim 1, characterized in that, Step 1, the process of acquiring and preprocessing data, includes: Obtain the accident record dataset: ； in, This represents the total number of accident samples. Indicates the first There are 10 accident records, and in each record... All include detailed accident description text. ; Accident description text Denoising is performed to generate a standardized set of accident texts: ； in, Indicates the first An accident record is a pre-processed accident text.

3. The aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks as described in claim 1, characterized in that, In the four-layer cascaded Bayesian network topology described in step 2, it is divided from top to bottom as follows: L1 Objective Facts Layer: As the root node layer, it contains basic attribute nodes for personnel, machinery, environment, and management dimensions; L2 Operational Behavior Layer: Contains nodes reflecting direct behavioral deviations during personnel maintenance operations; L3 intermediate state layer: contains nodes that characterize physical or system abnormal states that occur during maintenance; L4 Risk Outcome Layer: Contains risk level nodes classified according to civil aviation safety standards.

4. The aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks as described in claim 1, characterized in that, Step 4 involves scanning the standardized text using a keyword matching algorithm, which includes: When the Standardized accident text The occurrence of keywords in the set When the matching words are found, there are risk factors in the incident record. Each accident record is represented as a risk factor vector: ； in: ； x ij Indicates the first i Does the standardized accident text contain the first... j One risk factor, i For accident text indexing, j Index for risk factors; After this operation, each accident record is transformed into a high-dimensional Boolean vector, and all vectors converge to form a risk factor data matrix: 。 5. The aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks as described in claim 1, characterized in that, The prohibited edge rules mentioned in step 5 include: Lower-level nodes are prohibited from pointing to upper-level nodes, nodes at the same level are prohibited from connecting to each other, and nodes at the risk result level are prohibited from pointing to any other level of nodes.

6. The aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks as described in claim 1, characterized in that, Step 6 describes the process of establishing the Bayesian network model, which includes: Step 6.1: Constructing a Bayesian network can be represented as a directed acyclic graph: ； in, This represents a set of nodes, corresponding to various risk factors. Edges representing causal relationships between nodes; The joint probability distribution of a Bayesian network is expressed as: ； in, Represents a node The set of parent nodes; Step 6.2: A structure learning method based on the scoring function is adopted to search the network structure by adding, deleting, and reversing edges, and a hill climbing algorithm is used to find the network structure with the largest scoring function. The scoring function uses the BDeu score, which measures the quality of a structure by calculating the marginal likelihood of its occurrence in a dataset given the structure. Its basic form is as follows: ； in, In network structure Dataset under conditions The probability of occurrence; A random restart mechanism is introduced during the search process. After reaching a local peak, the current structure is randomly perturbed and the search is restarted. 。 7. The aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks as described in claim 1, characterized in that, Step 7 calculates the conditional probability distribution of each node in the Bayesian network. The expectation-maximization algorithm is used for parameter learning. By alternately executing the expectation step and the maximization step, the distribution of latent variables is estimated using known observation data, and the parameters are continuously updated until convergence. Specifically, this includes: By alternating the desired steps and maximizing steps To estimate model parameters; In the expected step In the middle, the expected sufficient statistic of the latent variables is calculated based on the current parameter estimates; in the maximization step... In this process, the expected log-likelihood function is maximized to update the parameter estimates; Repeat the above two steps until the parameters converge or the preset number of iterations is reached, and finally obtain the conditional probability table of each node in the Bayesian network.

8. The aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks as described in claim 1, characterized in that, Step 8, which uses 5-fold cross-validation to evaluate the model's predictive performance, includes: Step 8.1: Randomly divide the dataset into... A subset; Step 8.2: Select one subset at a time as the test set, and the rest... A subset is used as the training set; Step 8.3: Repeat step 8 on each training set to build the model; Step 8.4: Perform tests on the test set for each risk level node. Make a prediction and obtain the predicted probability. and real labels ; Step 8.5: Use As an evaluation metric, let the positive examples in the test set be... The set is , negative example for ,but The calculation formula is: ； in, , These represent the sample sizes of positive and negative examples, respectively. , These represent the indices of the positive and negative samples, respectively. , These represent the model's prediction of the first... , Each sample represents the probability of the risk occurring; This indicates an indicator function that takes the value 1 when the condition is true and 0 otherwise. The closer the value is to 1, the better the ranking performance of the model.

9. The aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks as described in claim 1, characterized in that, Step 9, the process of performing dynamic risk prediction using probabilistic reasoning methods, includes: For a given combination of risk factors The probability of the accident occurring is: ； in, Indicates the L4 layer risk level node. This indicates that an accident has occurred; Indicates the corresponding given The values ​​of each risk factor; This represents the joint probability of a risk occurring simultaneously when all risk factors take given values. This represents the probability that all risk factors will occur simultaneously when each factor takes a given value.

10. The aviation maintenance risk assessment and intelligent diagnosis method based on multi-dimensional Bayesian networks as described in claim 9, characterized in that, The posterior probability of the risk factor mentioned in step 10 is: ； in, Indicates risk factors The prior probability, Indicates risk factors under the conditions of an accident. The posterior probability; By comparing the increase in posterior probability with that in prior probability: ； The key factors that lead to this risk and their respective dimensions can be identified.

Images