Aero-engine group health evaluation method based on multi-working-condition dynamic clustering

Abstract

The invention discloses an aero-engine group health evaluation method based on multi-working-condition dynamic clustering, and belongs to the field of aero-engine health state evaluation. The method comprises the following steps: firstly, carrying out clustering analysis on set parameters in engine operation data, and carrying out merging processing on small-scale abnormal clusters to obtain a working condition category division result; secondly, constructing a health baseline data set, carrying out standardized preprocessing on sample data in a working condition category division result, and carrying out nonlinear dimensionality reduction to obtain a low-dimensional feature data set; thirdly, performing clustering analysis on the low-dimensional feature data set by adopting a Gaussian mixture model, and calculating an average mahalanobis distance between a sample of each clustering category and a health reference center to obtain multi-level health levels corresponding to different clustering categories; and finally, through fusing the membership soft probability and the sample individual mahalanobis distance, constructing a continuous health score and obtaining a health grade determination interval. According to the method, health state characteristics under different working conditions can be effectively identified, individual difference modeling and group transverse comparison evaluation are supported, and the accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the field of aircraft engine health status assessment and relates to an aircraft engine group health evaluation method based on multi-operating condition dynamic clustering. Background Art

[0002] With the development of intelligent aviation engine operation and maintenance, condition-based maintenance (CBM), based on real-time health status, is gradually replacing traditional maintenance strategies based on fixed cycles or fault triggers. Traditional approaches have significant limitations: on the one hand, excessively long maintenance intervals can easily lead to sudden failures, jeopardizing flight safety; on the other hand, overly frequent maintenance results in significant waste of resources. Research indicates that approximately 30% of engine maintenance costs stem from unnecessary preventive maintenance, demonstrating the significant value of accurate health assessments and dynamic maintenance strategies in improving economic efficiency and safety.

[0003] In practice, the health status of individual engines exhibits significant heterogeneity due to differences in production batches, service life, operating environments, and other factors. This makes maintenance decisions based on a uniform threshold unsatisfactory: some engines are over-maintained, while others are at risk of under-maintenance, resulting in an imbalance in resource allocation. Therefore, there is an urgent need to establish a refined health assessment mechanism tailored to individual engine differences and supporting horizontal comparative analysis of health across engine groups.

[0004] However, existing health management technologies still face several key challenges when dealing with complex operating conditions and large-scale heterogeneous data:

[0005] (1) Lack of automated modeling and multi-source fusion capabilities: For example, Chinese invention patent CN210864807U discloses an augmented reality-based fleet status management system, which uses augmented reality to achieve visual configuration of flight parameters and support resources. However, its health assessment still relies heavily on manual experience judgment and fails to establish an effective fusion analysis mechanism for multi-source heterogeneous sensor data of the engine, making it difficult to achieve automated and accurate individual health modeling.

[0006] (2) Limited modeling capabilities and inability to adapt to complex degradation characteristics: Chinese invention patent CN110428063 discloses an architecture and construction method for an integrated health management system for a fleet of aircraft. Although it integrates multiple types of operating data such as flight parameters, quality control, and tasks, its health monitoring method is still mainly based on fixed thresholds and simple statistics. It is difficult to capture subtle changes in aircraft engines under complex operating conditions and nonlinear degradation processes, and there is a risk of false alarms and missed reports.

[0007] (3) Data challenges are significant and traditional methods are difficult to adapt: ​​the sensor data of aircraft engines exhibit strong non-stationary and nonlinear change characteristics in different flight phases and environmental conditions, making it difficult to use a single model for global modeling; the sampling frequency, dimensional unit and noise level between sensors vary greatly, and direct fusion processing will obscure key information; in addition, fault samples are extremely scarce in actual operation, which seriously limits the applicability and generalization ability of supervised learning methods.

[0008] (4) Lack of group knowledge transfer and collaborative evaluation mechanisms: In the context of fleet management, the performance degradation patterns of engines of different models and service stages vary from individual to individual. Existing methods usually model individual engines and lack the ability to mine correlations and transfer knowledge from the fleet, resulting in low evaluation efficiency, decentralized decision-making, and uncoordinated maintenance strategies.

[0009] In summary, due to the significant multi-condition characteristics of aircraft engines in different flight phases and operating environments, traditional health assessment methods based on fixed thresholds have difficulty accurately identifying state changes. This invention effectively solves the multi-condition adaptive matching problem by integrating unsupervised dynamic clustering with a probability-distance dual assessment mechanism, significantly improving the sensitivity and specificity of health state identification, and providing reliable technical support for condition-based maintenance of aircraft engines. Summary of the Invention

[0010] In response to the problems existing in the existing technology, the present invention provides an aviation engine group health assessment method based on multi-operating condition dynamic clustering. Through multi-operating condition unsupervised modeling and health level adaptive clustering evaluation, it can effectively identify the health status characteristics under different operating conditions, and can support individual difference modeling and group horizontal comparative evaluation, thereby improving the accuracy of health assessment and reducing the misjudgment rate.

[0011] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0012] The aircraft engine group health assessment method based on multi-operating condition dynamic clustering includes the following steps:

[0013] Step 1: Use the HDBSCAN algorithm to perform cluster analysis on the set parameters in the engine operation data to achieve adaptive classification of operating condition categories and merge small-scale abnormal clusters to obtain stable operating condition classification results; specifically:

[0014] Step 1.1: Pre-process the collected raw engine operation data to extract a subset of all setting parameters that can characterize the engine operating environment and control status in each engine operation cycle sample. ,in is the sample size, The parameter subset is normalized to obtain a normalized parameter subset, so that the parameter subsets with different dimensions and value ranges have the same weight in the clustering process, thus avoiding clustering bias caused by feature scale differences.

[0015] Step 1.2, based on the normalized set parameter subset obtained in step 1.1, use the unsupervised density-based hierarchical clustering algorithm (Hierarchical Density-Based Spatial Clustering of Applications with Noise), referred to as HDBSCAN algorithm, to perform clustering. This algorithm identifies the cluster structure layer by layer by constructing a density-reachable relationship between sample points in the feature space. Compared with traditional clustering methods, the HDBSCAN algorithm only needs to set two parameters: the minimum cluster size (min_cluster_size) and the minimum number of samples (min_samples). There is no need to pre-specify the number of categories, that is, the number of clusters, but it is automatically determined according to the data distribution; at the same time, it can adapt to non-spherical cluster structures and mark individual isolated points that do not meet the density requirements as noise points, thereby improving the robustness of the working condition classification. Clustering obtains a set of category labels consisting of integer numbers ,in is the number of automatically determined working condition categories, -1 represents a noise point, and each sample in the normalized setting parameter subset corresponds to a category label in the category label set, and the preliminary clustering results are obtained.

[0016] Step 1.3, based on the preliminary clustering results obtained in step 1.2, in order to ensure the clustering quality, all noise point samples in the preliminary clustering results corresponding to the samples in the original engine operation data in step 1.1 are eliminated to obtain valid engine operation data. Afterwards, all non-noise point samples in the preliminary clustering results are counted. If the number of samples under a certain category label is too small, the category is marked as an abnormal small category, and the remaining categories are valid large categories. For these abnormal small categories, first calculate the geometric centroid of all valid large categories in the set parameter space to form a centroid set; then, use the nearest centroid principle to reclassify the abnormal small categories, that is, the samples in the abnormal small categories are Classify into the nearest valid category , the allocation rule is expressed as:

[0017] (1)

[0018] in, Represents effective categories The center of mass coordinates of .

[0019] Thus, the merged clustering results are obtained.

[0020] In step 1.4, renumber the category labels of the merged clustering results obtained in step 1.3. All category labels retained after the merging in step 1.3 are remapped to a sequence of natural numbers that increases continuously, starting from 1. This ensures that the operating condition categories are scalable and adaptable to different sample sets. The renumbered category labels are written into the valid engine operating data obtained in step 1.3 as the operating condition numbers of the samples, ultimately resulting in the operating condition category classification results.

[0021] Step 2: Construct a healthy baseline dataset for each working condition category in the working condition classification results. Use the healthy baseline dataset to perform standardization preprocessing on all sample data in the working condition classification results. Use the UMAP algorithm to perform nonlinear dimensionality reduction on the standardized data to finally obtain a low-dimensional feature dataset. Specifically:

[0022] In step 2.1, a minimum cycle threshold is set to ensure sufficient healthy period length. The operating condition classification results obtained in step 1 are then filtered to identify individual engines with sufficient healthy operating period lengths. This prevents engines with short life cycles from interfering with the healthy period determination. Sample operating data for these engines with sufficient healthy operating period lengths is extracted and filed separately according to the operating condition number obtained in step 1.4. This creates a healthy baseline dataset for each operating condition category, which serves as a reference for subsequent standardization.

[0023] In step 2.2, all the operating cycle samples in the operating condition category division results are grouped by operating condition category according to the operating condition number. For the samples in each operating condition category group, invalid features containing a large number of missing values, stable and unchanged, or irrelevant to health assessment are excluded, and the rest are valid features. The mean and standard deviation of the effective features of the health baseline data of each operating condition category obtained in step 2.1 are calculated, and the effective features of all sample data in the operating condition category group are standardized accordingly. The standardized feature data has the characteristics of a mean of 0 and a variance of 1, so that different feature data in the same operating condition category are distributed consistently, ensuring that different features are comparable. This standardization process is performed independently within each operating condition category to avoid confusion between operating conditions. Finally, high-dimensional feature data is obtained.

[0024] In step 2.3, the high-dimensional feature data normalized in step 2.2 is fed into the UMAP (Uniform Manifold Approximation and Projection) algorithm. UMAP is a nonlinear dimensionality reduction method based on the manifold assumption. It attempts to preserve the local neighborhood structure of high-dimensional data in a low-dimensional space, improving feature representation and computational efficiency. The target dimensionality reduction range is set from 1 to N, and representative samples of the high-dimensional feature data are randomly sampled to avoid memory explosion. Multiple rounds of dimensionality reduction structure learning are performed on the randomly sampled samples according to different target dimensions, resulting in N different dimensionality reduction results.

[0025] In step 2.4, calculate the KNN trustworthiness metric for each target dimension for each of the N dimensionality reduction results obtained in step 2.3. This evaluates the fidelity of these N dimensionality reduction results to the high-dimensional feature data structure obtained in step 2.2. The KNN trustworthiness metric reflects whether the neighbors of a sample in the low-dimensional space are still similar to those in the high-dimensional space. The more similar, the higher the metric value (closer to 1.0). The target dimension with the highest KNN trustworthiness metric value, i.e., the highest fidelity, is selected as the final target dimension.

[0026] In step 2.5, based on the final target dimension obtained in step 2.4, all high-dimensional feature data obtained in step 2.2 are input into the UMAP algorithm to obtain a low-dimensional feature data set after dimensionality reduction, which provides input for subsequent clustering modeling and health score calculation.

[0027] Step 3: Use the Gaussian mixture model to perform cluster analysis on the low-dimensional feature dataset, and calculate the average Mahalanobis distance between the samples of each cluster category and the health reference center to obtain the multi-level health level corresponding to different cluster categories; specifically:

[0028] Step 3.1: Input the low-dimensional feature dataset obtained in step 2.5 into the Gaussian mixture model for unsupervised clustering, and preset the number of clusters representing the number of health level categories you want to obtain. and covariance structure, automatically learning the probability distribution characteristics of sample data under each working condition category. The Gaussian mixture model is a typical generative probability model that assigns each sample a soft probability of belonging to each category, and is suitable for clustering modeling of sample data. The Gaussian mixture model assumes that the sample is generated by a mixture of multiple Gaussian distributions. The expectation-maximization (EM) algorithm is used to iteratively estimate the mean, covariance, and mixing coefficient of each distribution to characterize the characteristic patterns under different health states, and output information such as the central mean vector, covariance matrix, and sample membership soft probability of each cluster. This process independently models each working condition category to ensure that the health level division is closely related to the working condition characteristics. The overall model can be expressed as:

[0029] (2)

[0030] in, is the sample vector in the low-dimensional feature dataset obtained in step 2.5, with a dimension of , for The probability density value under the entire model, that is, "the possibility of this point in the mixed distribution", is the number of Gaussian components, that is, the number of clusters, For the The weights of the Gaussian components satisfy ,and , For the The mean vector of a Gaussian distribution, For the The covariance matrix of a Gaussian distribution, is the multivariate Gaussian distribution density function.

[0031] Parameters of Gaussian mixture models Learning is performed using the EM algorithm. The EM iteration is divided into two steps:

[0032] (1) E-step (Expectation Step): Calculate the soft probability of each sample belonging to each Gaussian component;

[0033] (2) M-step (Maximization Step): Use the results of the E-step to update the weights , mean , covariance ;

[0034] The EM algorithm continuously iterates the above two steps until the log-likelihood of the model converges or the maximum number of iterations is reached.

[0035] Finally, the Gaussian mixture model parameters under each working condition category are obtained, including the mean vector and covariance matrix of each cluster center, as well as the cluster category and soft probability of membership of each sample in the low-dimensional feature data set obtained in step 2.5, which are used for subsequent grading and health scoring.

[0036] In step 3.2, under each operating condition category, the cluster center of the corresponding cluster in step 3.1 of the samples in the healthy baseline data set obtained in step 2.1 is used as the healthy reference center; in order to fully reflect the overall distribution characteristics of the cluster category and its overall deviation from the healthy state, thereby avoiding the measurement bias caused by the insufficient representativeness of a single cluster center point, a part of the samples are randomly selected from each operating condition category of the low-dimensional feature data set obtained in step 2.5, and the Mahalanobis distance between each sample under the operating condition category and the healthy reference center under the operating condition category is calculated, and the cluster categories to which these sampled samples belong in step 3.1 are counted. Finally, the average Mahalanobis distance between all the sampled samples under each cluster category and the healthy reference center is calculated to measure the operating status of each cluster category and the degree of deviation of its healthy state from the healthy reference center.

[0037] In step 3.3, the health levels of the cluster categories are sorted according to the average Mahalanobis distance of each cluster category obtained in step 3.2. The larger the average Mahalanobis distance, the lower the health level. The final sorting result is the multi-level health level.

[0038] Step 4: By integrating the membership soft probability obtained in step 3 with the Mahalanobis distance of the sample individuals, a continuous health score is constructed and the health level judgment interval is obtained. Specifically:

[0039] Step 4.1: Under each working condition category, use the Gaussian mixture model parameters obtained in step 3.1 to extract the cluster category and membership soft probability of each sample, and set the weight for each level according to the health level obtained in step 3.3 according to the cluster category. , for any sample, its membership soft probability is recorded as , the weighted initial health score formula obtained by combining the soft probability of membership with the weight is as follows:

[0040] (3)

[0041] To further enhance the interpretability and consistency of the scores, Perform min-max normalization so that it is mapped to The normalized initial health score obtained within the interval is recorded as .at this time, , which represents the risk level of individual health status reflected by the cluster categories and membership soft probabilities obtained in step 3.1.

[0042] Step 4.2: To reflect the degree of deviation of each engine operation cycle sample individual from the cluster center of the cluster category obtained in step 3.1, calculate the individual Mahalanobis distance from the sample to the cluster center of the category to which it belongs. , and normalize the individual Mahalanobis distance into a deviation index; the deviation index is adjusted according to the adjustable fusion coefficient The normalized initial health score obtained in step 4.1 is recorded as Perform weighted fusion and truncate to , and obtain a comprehensive continuous health score , the formula is as follows:

[0043] (4)

[0044] in, is the adjustable fusion coefficient, are the minimum and maximum Mahalanobis distances of the sample individuals, respectively.

[0045] In step 4.3, based on the health level obtained in step 3.3 and the comprehensive health score obtained in step 4.2, a health level determination interval can be obtained. Specifically, the samples of each cluster category obtained in step 3.1 are calculated in step 4.2 to obtain their corresponding comprehensive health scores. The obtained comprehensive health scores can be used as the health level determination interval corresponding to the cluster category. In other words, the set of comprehensive health scores for a certain health level of a certain working condition category can be directly used as a reference for the health level determination interval.

[0046] The comprehensive health score and health grading results can be visualized in the time series dimension to show the health evolution trends of individuals and groups.

[0047] Innovative analysis of the present invention:

[0048] (1) Multi-operating-condition unsupervised modeling and adaptive clustering assessment of health levels: Targeting the multi-operating-condition characteristics of engine operation, a soft clustering model layered by operating condition is constructed. GMM is used to model standardized sensor data and automatically identify health level categories. By mapping soft probability outputs to health levels, a continuous health score is generated, avoiding reliance on fixed thresholds or labeled samples and improving the model's generalization and practicality.

[0049] (2) Support individual difference modeling and group horizontal comparative evaluation: The clustering model is trained independently under each working condition, and the category sorting and health level identification are realized according to the Mahalanobis distance or cluster center characteristics. Through a unified health scoring framework, the individual health scores of each engine under different working conditions can be compared horizontally between the fleet, supporting subsequent differentiated maintenance and dynamic resource scheduling.

[0050] The beneficial effects of the present invention are:

[0051] The present invention adopts unsupervised learning methods throughout the process, and can achieve accurate learning and evaluation of the health status of aircraft engines without relying on fault samples or remaining life labels required for training models by traditional methods. It automatically divides different operating condition categories through multi-operating condition dynamic clustering methods and establishes corresponding multi-level health grade evaluation and continuous health score calculation methods. By integrating the clustering categories of the Gaussian mixture model and the individual Mahalanobis distance deviation index, the evaluation accuracy is significantly improved. When modeling based on normal operating data, it can still effectively identify early performance degradation characteristics and display the evolution trend of the engine health status, providing reliable preventive maintenance decision support for aircraft engine fleets that lack historical fault data. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of the method of the present invention;

[0053] Figure 2 This is a schematic diagram of the original sequence (not divided into working conditions) of a random sensor of the engine with the longest operating history cycle;

[0054] Figure 3 This is a schematic diagram of the operating sequence of the same sensor of the engine under working condition 1;

[0055] Figure 4 This is a schematic diagram of the operating sequence of the same sensor in the engine under working condition 2;

[0056] Figure 5 This is a schematic diagram of the operating sequence of the same sensor in the engine under working condition 3;

[0057] Figure 6 This is a schematic diagram of the operating sequence of the same sensor in the engine under working condition 4;

[0058] Figure 7 This is a schematic diagram of the operating sequence of the same sensor in the engine under working condition 5;

[0059] Figure 8 This is a schematic diagram of the operating sequence of the same sensor in the engine under working condition 6;

[0060] Figure 9 It is a schematic diagram of group health scoring and grade assessment of the present invention. DETAILED DESCRIPTION

[0061] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0062] like Figure 1 As shown, a method for evaluating the health of an aircraft engine population based on dynamic clustering of multiple operating conditions includes the following steps:

[0063] Step 1: Use the HDBSCAN algorithm to perform cluster analysis on the set parameters in the engine operation data to achieve adaptive classification of operating condition categories and merge small-scale abnormal clusters to obtain stable operating condition classification results; specifically:

[0064] Step 1.1: Pre-process the collected raw engine operation data to extract a subset of all setting parameters that can characterize the engine operating environment and control status in each engine operation cycle sample. ,in is the sample size, is the feature dimension. In this embodiment, the parameter feature subset includes flight altitude, Mach number, and throttle lever angle. This parameter subset is normalized to obtain a normalized parameter subset. This ensures that parameters with different dimensions and value ranges have the same weight during clustering, thus avoiding clustering bias caused by differences in feature scales.

[0065] Step 1.2, based on the normalized set parameter subset obtained in step 1.1, use the unsupervised HDBSCAN algorithm for clustering. This algorithm identifies the cluster structure layer by layer by constructing a density-reachable relationship between sample points in the feature space. Compared with traditional clustering methods, the HDBSCAN algorithm only needs to set two parameters: the minimum cluster size (min_cluster_size) and the minimum number of samples (min_samples). There is no need to pre-specify the number of categories, that is, the number of clusters, but it is automatically determined according to the data distribution; at the same time, it can adapt to non-spherical cluster structures and mark individual isolated points that do not meet the density requirements as noise points, thereby improving the robustness of the working condition classification. In this embodiment, min_cluster_size=50, min_samples=10 are set. Clustering obtains a set of category labels consisting of integer numbers ,in is the number of automatically determined working condition categories, -1 represents a noise point, and each sample in the normalized setting parameter subset corresponds to a category label in the category label set, and the preliminary clustering results are obtained.

[0066] Step 1.3, based on the preliminary clustering results obtained in step 1.2, in order to ensure the clustering quality, all noise point samples in the preliminary clustering results that correspond to the samples in the original engine operation data in step 1.1 are eliminated to obtain valid engine operation data. Afterwards, all non-noise point samples in the preliminary clustering results are counted. If the number of samples under a certain category label is too small, in this embodiment, "too small" is set to a proportion lower than 0.5% of the total number of non-noise point samples, then the category is marked as an abnormal small category, and the remaining categories are valid large categories. For these abnormal small categories, first calculate the geometric centroids of all valid large categories in the set parameter space to form a centroid set; then, use the nearest centroid principle to reclassify the abnormal small categories, that is, the samples in the abnormal small categories Classify into the nearest valid category , the allocation rule is expressed as:

[0067] (1)

[0068] in, Represents effective categories The center of mass coordinates of .

[0069] Thus, the merged clustering results are obtained.

[0070] Step 1.4, renumber the category labels of the merged clustering results obtained in step 1.3, and remap all the category labels retained after the merging in step 1.3 to a natural number sequence that increases continuously starting from 1, to ensure that the working condition category is scalable and adaptable to different sample set sizes. The renumbered new category labels are written into the valid engine operating data obtained in step 1.3 as the operating condition number of the sample, and finally the working condition category division result is obtained. Take the random sensor data of the engine with the longest historical cycle number of the original engine operating data in this embodiment, and compare it before and after the working condition category division. Figure 2-Figure 8 As shown in the figure, it can be clearly seen that after the operating condition categories are clustered, the engine's decay trend under each operating condition with the increase in the number of operating cycles is more clearly shown, and the data jumps and interference caused by changes in operating conditions are significantly eliminated.

[0071] Step 2: Construct a healthy baseline dataset for each working condition category in the working condition classification results. Use the healthy baseline dataset to perform standardization preprocessing on all sample data in the working condition classification results. Use the UMAP algorithm to perform nonlinear dimensionality reduction on the standardized data to finally obtain a low-dimensional feature dataset. Specifically:

[0072] In step 2.1, a minimum cycle threshold is first set to ensure sufficient healthy period cycle length (based on common practices in the field of aircraft engine remaining life prediction and the cycle distribution characteristics of the dataset used in this embodiment, the minimum cycle threshold in this embodiment is set to 125). The operating condition classification results obtained in step 1 are then screened to extract individual engines with sufficient healthy period operating cycle lengths to prevent interference in the healthy period determination caused by engines with too short life cycles. In this embodiment, the first 20% of the operating cycle history of these individual engines with sufficient healthy period operating cycle lengths is extracted and treated as healthy operating sample data. These data are then archived separately according to the operating condition number obtained in step 1.4 to obtain a healthy baseline dataset for each operating condition category, which serves as a reference for subsequent standardization.

[0073] In step 2.2, all the operating cycle samples in the operating condition category division results are grouped by operating condition category according to the operating condition number. For the samples in each operating condition category group, invalid features containing a large number of missing values, stable and unchanged, or irrelevant to health assessment are excluded, and the rest are valid features. The mean and standard deviation of the effective features of the health baseline data of each operating condition category obtained in step 2.1 are calculated, and the effective features of all sample data in the operating condition category group are standardized accordingly. The standardized feature data has the characteristics of a mean of 0 and a variance of 1, so that different feature data in the same operating condition category are distributed consistently, ensuring that different features are comparable. This standardization process is performed independently within each operating condition category to avoid confusion between operating conditions. Finally, high-dimensional feature data is obtained.

[0074] In step 2.3, the high-dimensional feature data after the standardization process in step 2.2 is input into the UMAP algorithm. UMAP is a nonlinear dimensionality reduction method based on the manifold hypothesis. It attempts to maintain the local neighborhood structure of high-dimensional data in low-dimensional space, thereby improving feature expression capabilities and computational efficiency. The dimensionality reduction target dimension range is set from 1 to N, and some representative samples of the high-dimensional feature data are taken by random sampling to avoid memory explosion. In this embodiment, considering that the dimension of the high-dimensional feature data is 14, N is set to 10, and the number of randomly sampled samples is set to 3000. According to different target dimensions, multiple rounds of dimensionality reduction structure learning are performed on the randomly sampled samples to obtain N dimensionality reduction results.

[0075] In step 2.4, the K-nearest neighbor trustworthiness metric (KNNTrustworthiness) is calculated for each target dimension for the N dimensionality reduction results obtained in step 2.3. This evaluates the fidelity of the N dimensionality reduction results to the high-dimensional feature data structure obtained in step 2.2. The K-nearest neighbor trustworthiness metric reflects whether the neighbors of a sample in the low-dimensional space are still similar to those in the high-dimensional space. The more similar, the higher the metric value (closer to 1.0). The target dimension with the highest K-nearest neighbor trustworthiness metric value, i.e., the highest fidelity, is selected as the final target dimension. In this embodiment, the final target dimension is 7.

[0076] In step 2.5, based on the final target dimension obtained in step 2.4, all high-dimensional feature data obtained in step 2.2 are input into the UMAP algorithm to obtain a low-dimensional feature data set after dimensionality reduction, which provides input for subsequent clustering modeling and health score calculation.

[0077] Step 3: Use the Gaussian mixture model to perform cluster analysis on the low-dimensional feature dataset, and calculate the average Mahalanobis distance between the samples of each cluster category and the health reference center to obtain the multi-level health level corresponding to different cluster categories; specifically:

[0078] Step 3.1: Input the low-dimensional feature dataset obtained in step 2.5 into the Gaussian mixture model for unsupervised clustering, and preset the number of clusters representing the number of health level categories you want to obtain. and covariance structure, automatically learn the probability distribution characteristics of sample data under each working condition category. In this embodiment, the preset , representing three levels of health, namely health level, monitoring level and warning level; the covariance structure is set to the full covariance matrix (full covariance) to capture the correlation between features. The Gaussian mixture model is a typical generative probability model that assigns each sample a soft probability of belonging to each category, which is suitable for clustering modeling of sample data. The Gaussian mixture model assumes that the sample is generated by a mixture of multiple Gaussian distributions. The EM algorithm is used to iteratively estimate the mean, covariance and mixing coefficient of each distribution to characterize the characteristic patterns under different health states, and output the central mean vector, covariance matrix and sample membership soft probability of each cluster. This process independently models each operating condition category to ensure that the health level division is closely related to the operating condition characteristics. The overall model can be expressed as:

[0079] (2)

[0080] in, is the sample vector in the low-dimensional feature dataset obtained in step 2.5, with a dimension of , for The probability density value under the entire model, that is, "the possibility of this point in the mixed distribution", is the number of Gaussian components, that is, the number of clusters, For the The weights of the Gaussian components satisfy ,and , For the The mean vector of a Gaussian distribution, For the The covariance matrix of a Gaussian distribution, is the multivariate Gaussian distribution density function.

[0081] Parameters of Gaussian mixture models Learning is performed using the EM algorithm. The EM iteration is divided into two steps:

[0082] (1) E-step (Expectation Step): Calculate the soft probability of each sample belonging to each Gaussian component;

[0083] (2) M-step (MaximizationStep): Use the results of the E-step to update the weights , mean , covariance ;

[0084] The EM algorithm continuously iterates the above two steps until the log-likelihood of the model converges or the maximum number of iterations is reached.

[0085] Finally, the Gaussian mixture model parameters under each working condition category are obtained, including the mean vector of each cluster center, the covariance matrix and the soft probability of sample membership, which are used for subsequent grading and health scoring.

[0086] In step 3.2, under each operating condition category, the cluster center of the corresponding cluster in step 3.1 of the samples in the healthy baseline data set obtained in step 2.1 is used as the healthy reference center; in order to fully reflect the overall distribution characteristics of the cluster category and its overall deviation from the healthy state, thereby avoiding the measurement deviation caused by the insufficient representativeness of a single cluster center point, a portion of samples are randomly selected from each operating condition category of the low-dimensional feature data set obtained in step 2.5. In this embodiment, the number of samples selected for each operating condition category is 1000. The Mahalanobis distance between each sample selected under the operating condition category and the healthy reference center under the operating condition category is calculated, and the cluster categories to which these selected samples belong in step 3.1 are counted. Finally, the average Mahalanobis distance between all the selected samples under each cluster category and the healthy reference center is calculated to measure the operating status of each cluster category and the degree of deviation of its healthy state from the healthy reference center.

[0087] In step 3.3, the cluster categories are ranked by health level based on the average Mahalanobis distance of each cluster category obtained in step 3.2. A larger average Mahalanobis distance indicates a lower health level, and the final ranking result is a multi-level health level. In this embodiment, the health levels of the final three health levels are healthy level, monitoring level, and warning level from high to low. The specific classification is as follows:

[0088] Health level (level 0): corresponds to the category with the smallest average Mahalanobis distance, representing the optimal operating state of the equipment;

[0089] Monitoring level (Level 1): corresponds to the category in the middle of the ranking, used to indicate an operating state where performance has deviated but not yet seriously deteriorated;

[0090] Warning level (Level 2): ​​corresponds to the category with the largest average Mahalanobis distance. This category is farthest from the health baseline and is considered a high-risk state.

[0091] Step 4: By integrating the membership soft probability obtained in step 3 with the Mahalanobis distance of the sample individuals, a continuous health score is constructed and the health level judgment interval is obtained. Specifically:

[0092] Step 4.1: Under each working condition category, use the Gaussian mixture model parameters obtained in step 3.1 to extract the cluster category and membership soft probability of each sample, and set the weight for each level according to the health level obtained in step 3.3 according to the cluster category. In this embodiment, the three levels of health, monitoring and warning are assigned weight values , for any sample, its membership soft probability is recorded as , the weighted initial health score formula obtained by combining the soft probability of membership with the weight is as follows:

[0093] (3)

[0094] To further enhance the interpretability and consistency of the scores, Perform min-max normalization so that it is mapped to The normalized initial health score obtained within the interval is recorded as .at this time, , represents the individual health status risk level reflected by the clustering categories and soft probabilities obtained in step 3.1. In this embodiment, it corresponds to the specific classification in step 3.3. Corresponding to pure health (completely at level 0), Corresponding to severe anomalies (entirely Level 2).

[0095] Step 4.2: To reflect the degree of deviation of each engine operation cycle sample individual from the cluster center of the cluster category obtained in step 3.1, calculate the individual Mahalanobis distance from the sample to the cluster center of the category to which it belongs. , and normalize the individual Mahalanobis distance into a deviation index; the deviation index is combined with the normalized initial health score obtained in step 4.1 according to the adjustable fusion coefficient and recorded as Perform weighted fusion and truncate to , and obtain a comprehensive continuous health score , the formula is as follows:

[0096] (4)

[0097] in, is an adjustable fusion coefficient, which is set to 0.8 in this embodiment. are the minimum and maximum Mahalanobis distances of the sample individuals, respectively.

[0098] In step 4.3, based on the health level obtained in step 3.3 and the comprehensive health score obtained in step 4.2, the health level determination interval can be obtained. Specifically: the samples of each cluster category obtained in step 3.1 are calculated in step 4.2 to obtain their corresponding comprehensive health scores, and the obtained comprehensive health scores can be used as the health level determination interval corresponding to the cluster category, that is, the set of comprehensive health scores for a certain health level of a certain working condition category can be directly used as a reference for the health level determination interval. In this embodiment, the monitoring threshold is approximately 0.4; the warning threshold is approximately 0.75. This score comprehensively considers the soft probability information and statistical distance characteristics in the cluster, which can not only reflect the health level attribution of the data in the cluster structure, but also quantify its relative deviation from the healthy sample, thereby more comprehensively and robustly evaluating the individual status, and finally realizing a closed loop from the cluster model to the score interpretation. The comprehensive health score and health grading results can be visualized in the time series dimension to show the health evolution trend of individuals and groups. In this embodiment, the historical operating data of two engines under a certain working condition are randomly taken for visualization, and the results are as follows. Figure 9 shown.

[0099] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made in accordance with the spirit of the present invention are intended to be covered by the scope of protection of the present invention.

Claims

1. An aircraft engine population health assessment method based on multi-operating condition dynamic clustering, characterized by: The method for evaluating the health of an aircraft engine population comprises the following steps: Step 1: Use an unsupervised density-based hierarchical clustering algorithm to perform cluster analysis on the set parameters in the engine operation data, achieve adaptive classification of operating condition categories, and merge small-scale abnormal clusters to obtain stable operating condition classification results; Step 2: construct a healthy baseline dataset for each working condition category in the working condition classification results, use the healthy baseline dataset to perform standardization preprocessing on all sample data in the working condition classification results, use the UMAP algorithm to perform nonlinear dimensionality reduction on the standardized data, and finally obtain a low-dimensional feature dataset; Step 3: Use the Gaussian mixture model to perform cluster analysis on the low-dimensional feature dataset obtained in step 2, and calculate the average Mahalanobis distance between the samples of each cluster category and the health reference center to obtain the multi-level health level corresponding to different cluster categories; Step 4: By fusing the membership soft probability and the Mahalanobis distance of the sample individuals, a continuous health score is constructed and the health level judgment interval is obtained.

2. The method for aircraft engine group health assessment based on multi-operating condition dynamic clustering according to claim 1 is characterized in that: The step 1 is specifically as follows: Step 1.1: Pre-process the collected raw engine operation data to extract a subset of all setting parameters that can characterize the engine operating environment and control status in each engine operation cycle sample. ,in is the sample size, is the feature dimension; Normalizing the setting parameter subset to finally obtain a normalized setting parameter subset; Step 1.2: Based on the set parameter subset, use the unsupervised density-based hierarchical clustering algorithm, referred to as HDBSCAN algorithm, to perform clustering and obtain a set of category labels consisting of integer numbers. ,in is the number of working condition categories automatically determined, -1 represents a noise point, and each sample in the normalized set parameter subset corresponds to a category label in the category label set to obtain the preliminary clustering result; Step 1.3: All noise point samples in the preliminary clustering results and the corresponding samples in the original engine operation data in step 1.1 are removed to obtain valid engine operation data; then, all non-noise point samples in the preliminary clustering results are counted to obtain abnormal small classes and valid large classes; first, the geometric centroids of all valid large classes in the set parameter space are calculated to form a centroid set; then, the abnormal small classes are reclassified using the nearest centroid principle, and the samples in the abnormal small classes are classified into Classify into the nearest valid category , the allocation rule is expressed as: （1） in, Represents effective categories The coordinates of the center of mass of Thus, the merged clustering results are obtained; In step 1.4, the category labels of the clustering results obtained in step 1.3 are renumbered, and all category labels are remapped into a sequence of natural numbers that increases continuously starting from 1; the renumbered new category labels are written into the valid engine operating data obtained in step 1.3 as the operating condition number of the sample, and finally the operating condition category division result is obtained.

3. The method for aircraft engine group health assessment based on multi-operating condition dynamic clustering according to claim 2 is characterized in that: The step 2 is specifically as follows: Step 2.1: First, set a minimum cycle threshold, filter the operating condition classification results obtained in step 1, and determine the engine individual with the longest healthy period operation cycle; Extract the running sample data of the individual engine under healthy state, and file them separately according to the operating condition number obtained in step 1.4 to obtain the healthy baseline data set under each operating condition category; Step 2.2: According to the operating condition number, all the operating cycle samples in the operating condition classification result are grouped by operating condition category and standardized to obtain high-dimensional feature data; Step 2.3: Input the high-dimensional feature data after the normalization process in step 2.2 into the UMAP algorithm; set the dimensionality reduction target dimension range from 1 to N, and randomly sample some representative samples from the high-dimensional feature data; perform multiple rounds of dimensionality reduction structure learning on the randomly sampled samples according to different target dimensions, and obtain N dimensionality reduction results; Step 2.4: Calculate the K-nearest neighbor credibility index for each target dimension for the N dimensionality reduction results obtained in step 2.3, evaluate the fidelity of the N dimensionality reduction results to the high-dimensional feature data structure obtained in step 2.2, and select the target dimension with the highest K-nearest neighbor credibility index value, that is, the target dimension with the highest fidelity, as the final target dimension; In step 2.5, based on the final target dimension obtained in step 2.4, all high-dimensional feature data obtained in step 2.2 are input into the UMAP algorithm to obtain a low-dimensional feature data set after dimensionality reduction.

4. The method for aircraft engine group health assessment based on multi-operating condition dynamic clustering according to claim 3 is characterized in that: The step 2.2 is specifically as follows: For the samples in each working condition category group, invalid features containing a large number of missing values, stable and unchanged, or irrelevant to health assessment are excluded, and the rest are valid features; the mean and standard deviation of the valid features of the health baseline data of each working condition category obtained in step 2.1 are calculated, and the valid features of all sample data in the working condition category group are standardized accordingly. This standardization process is performed independently in each working condition category, and finally high-dimensional feature data is obtained.

5. The method for aircraft engine group health assessment based on multi-operating condition dynamic clustering according to claim 4 is characterized in that: The step 3 is specifically as follows: Step 3.1: Input the low-dimensional feature dataset obtained in step 2.5 into the Gaussian mixture model for unsupervised clustering, and preset the number of clusters representing the number of health level categories you want to obtain. and covariance structure, automatically learning the probability distribution characteristics of sample data under each working condition category; Model each operating condition category independently to ensure that the health level classification is closely related to the operating condition characteristics; Using the EM algorithm to learn, we finally obtain the Gaussian mixture model parameters for each operating condition category, including the mean vector and covariance matrix of each cluster center, as well as the cluster category and soft probability of membership of each sample in the low-dimensional feature dataset obtained in step 2.5, which are used for subsequent grading and health scoring; In step 3.2, under each operating condition category, the cluster center of the samples in the healthy baseline dataset obtained in step 2.1 corresponding to the cluster in step 3.1 is used as the healthy reference center; based on the low-dimensional feature dataset obtained in step 2.5, a portion of samples is randomly selected from each operating condition category, and the Mahalanobis distance between each sample in that operating condition category and the healthy reference center in that operating condition category is calculated. The cluster category to which the sampled samples belong in step 3.1 is counted, and finally the average Mahalanobis distance between all the sampled samples in each cluster category and the healthy reference center is calculated to measure the operating status of each cluster category and the degree of deviation of its health status from the healthy reference center. In step 3.3, the health levels of the cluster categories are sorted according to the average Mahalanobis distance of each cluster category obtained in step 3.

2. The larger the average Mahalanobis distance, the lower the health level. The final sorting result is the multi-level health level.

6. The method for aircraft engine group health assessment based on multi-operating condition dynamic clustering according to claim 5 is characterized in that: In step 3.1: The Gaussian mixture model is expressed as: （2） in, is the sample vector in the low-dimensional feature dataset, with dimension ; for Probability density value under Gaussian mixture model; is the number of Gaussian components, indicating the number of clusters; For the The weights of the Gaussian components satisfy ,and ; For the The mean vector of the Gaussian distribution; For the The covariance matrix of a Gaussian distribution; is the multivariate Gaussian distribution density function; The EM algorithm is used to learn the parameters of the Gaussian mixture model. , EM iteration is divided into two steps: (1) Calculate the soft probability of each sample belonging to each Gaussian component; (2) Update weights using membership soft probabilities , mean , covariance ; The above two steps are iterated continuously until the log-likelihood of the Gaussian mixture model converges or the maximum number of iterations is reached.

7. The method for aircraft engine group health assessment based on multi-operating condition dynamic clustering according to claim 6, characterized in that: The step 4 is specifically as follows: Step 4.1: Under each working condition category, use the Gaussian mixture model parameters obtained in step 3.1 to extract the cluster category and membership soft probability of each sample, and set the weight for each level according to the health level obtained in step 3.3 according to the cluster category. , for any sample, its membership soft probability is recorded as , the weighted initial health score formula obtained by combining the soft probability of membership with the weight is as follows: （3） right Perform min-max normalization so that it is mapped to Within the interval, the normalized initial health score is recorded as ;at this time, , represents the risk level of individual health status reflected by the cluster categories and membership soft probabilities obtained in step 3.1; Step 4.2: To reflect the degree of deviation of each engine operation cycle sample individual from the cluster center of the cluster category obtained in step 3.1, calculate the individual Mahalanobis distance from the sample to the cluster center of the category to which it belongs. , and normalize the individual Mahalanobis distance into a deviation index; the deviation index is adjusted according to the adjustable fusion coefficient The normalized initial health score obtained in step 4.1 is recorded as Perform weighted fusion and truncate to , and obtain a comprehensive continuous health score , the formula is as follows: （4） in, is the adjustable fusion coefficient, are the minimum and maximum values ​​of the Mahalanobis distance of the sample individuals, respectively; Step 4.3, based on the health level obtained in step 3.3 and the comprehensive health score obtained in step 4.2, obtain the health level determination interval.

8. The method for aircraft engine group health assessment based on multi-operating condition dynamic clustering according to claim 7 is characterized in that: The step 4.3 is specifically as follows: The samples of each cluster category obtained in step 3.1 are used in step 4.2 to calculate their corresponding comprehensive health scores. The obtained comprehensive health scores are used as the health level judgment interval corresponding to the cluster category. That is, the set of comprehensive health scores for a certain health level of a certain working condition category can be directly used as a reference for the health level judgment interval. The comprehensive health score and health grading results are visualized in the time series dimension to show the health evolution trends of individuals and groups.

Images