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

Abstract

The method for evaluating the health of an aero-engine group based on multi-working condition dynamic clustering belongs to the field of aero-engine health state evaluation. Firstly, the set parameters in the engine operation data are subjected to clustering analysis, and small-scale abnormal clusters are subjected to merging processing to obtain working condition category division results. Secondly, a health baseline data set is constructed, the sample data in the working condition category division results are subjected to standardization preprocessing and nonlinear dimension reduction to obtain a low-dimensional feature data set. Thirdly, the low-dimensional feature data set is subjected to clustering analysis by using a Gaussian mixture model, the average Mahalanobis distance of the sample and the health reference center of each clustering category is calculated, and the multi-level health grades corresponding to different clustering categories are obtained. Finally, by fusing the membership soft probability and the sample individual Mahalanobis distance, a continuous health score is constructed, and a health grade judgment interval is obtained. The present application can effectively identify the health state characteristics under different working conditions, support individual difference modeling and group horizontal comparison evaluation, and improve the accuracy.

Description

Technical Field

[0001] This invention belongs to the field of aero-engine health status assessment and relates to a method for aero-engine group health assessment based on multi-condition dynamic clustering. Background Technology

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

[0003] In practical applications, the health status of individual engines exhibits significant heterogeneity due to differences in production batches, service life, and operating environments. This makes maintenance decisions based on uniform thresholds inadequate for practical needs: 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 construct a refined health assessment mechanism that addresses the individual differences of engines and supports cross-engine group health comparative analysis.

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

[0005] (1) Lack of automated modeling and multi-source fusion capabilities: For example, Chinese invention patent CN210864807U discloses a fleet status management system based on augmented reality, which realizes the display of flight parameters and the visualization configuration of support resources through augmented reality. However, its health assessment still relies heavily on human 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, unable to adapt to complex degradation characteristics: Chinese invention patent CN110428063 discloses an architecture and construction method for a fleet health integrated management system. Although it integrates multiple types of operational data such as flight parameters, quality control, and missions, its health monitoring method is still mainly based on fixed thresholds and simple statistics, which makes it difficult to capture the subtle changes of aero engines in complex operating conditions and nonlinear degradation processes, and there is a risk of false alarms and missed alarms.

[0007] (3) Significant data challenges and traditional methods are difficult to adapt: ​​The sensor data of aero engines exhibit strong non-stationary and nonlinear variation characteristics under different flight stages and environmental conditions, making it difficult to perform global modeling using a single model; There are large differences in sampling frequency, unit of measurement and noise level among sensors, 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 mechanism: Under the background of fleet management, the performance degradation patterns of engines of different models and service stages vary individually. Existing methods usually model on a single-engine basis, lacking 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 aero-engines under different flight phases and operating environments, traditional health assessment methods based on fixed thresholds are insufficient to accurately identify state changes. This invention effectively solves the multi-condition adaptive matching problem by integrating unsupervised dynamic clustering and a probability-distance dual assessment mechanism, significantly improving the sensitivity and specificity of health status identification and providing reliable technical support for condition-based maintenance of aero-engines. Summary of the Invention

[0010] To address the problems existing in the prior art, this invention provides a group health assessment method for aero-engines based on multi-condition dynamic clustering. Through multi-condition unsupervised modeling and adaptive clustering assessment of health levels, it can effectively identify health status characteristics under different conditions, and can support individual difference modeling and group cross-comparison assessment, thereby improving the accuracy of health assessment and reducing the misjudgment rate.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0012] A group health assessment method for aero-engines based on multi-condition dynamic clustering includes the following steps:

[0013] Step 1: Using the HDBSCAN algorithm, cluster analysis is performed on the set parameters in the engine operating data to achieve adaptive classification of operating condition categories, and small-scale abnormal clusters are merged to obtain stable operating condition category classification results; specifically:

[0014] Step 1.1: Preprocess the collected raw engine operating data to extract the subset of set parameters that characterize the engine operating environment and control state contained in each engine operating cycle sample. ,in For sample size, The feature dimension is defined as follows. This subset of parameters is then normalized to obtain a normalized subset of parameters, ensuring that parameters with different dimensions and value ranges have the same weight during clustering, thus avoiding clustering bias caused by differences in feature scale.

[0015] Step 1.2: Based on the normalized subset of parameters obtained in Step 1.1, an unsupervised, density-based hierarchical clustering algorithm (Hierarchical Density-Based Spatial Clustering of Applications with Noise), or HDBSCAN algorithm, is used for clustering. This algorithm identifies cluster structures layer by layer by constructing density reachability relationships between sample points in the feature space. Compared to traditional clustering methods, the HDBSCAN algorithm only requires setting two parameters: minimum cluster size (min_cluster_size) and minimum number of samples (min_samples). It does not require pre-specifying the number of categories (i.e., the number of clusters), but is automatically determined based on the data distribution. Furthermore, it can adapt to non-spherical cluster structures and marks isolated points that do not meet the density requirements as noise points, improving the robustness of application condition classification. The clustering results in a set of category labels composed of integer indices. ,in The number of working condition categories is automatically determined, and -1 represents noise points. Each sample in the normalized parameter subset corresponds to a category label in the category label set, thus obtaining the preliminary clustering results.

[0016] Step 1.3: Based on the preliminary clustering results obtained in Step 1.2, to ensure clustering quality, all noise point samples in the preliminary clustering results are removed from the original engine operation data in Step 1.1, resulting in valid engine operation data. Then, all non-noise point samples in the preliminary clustering results are statistically analyzed. If the number of samples under a certain category label is too small, that category is marked as an abnormal subclass, and the remaining categories are considered valid major classes. For these abnormal subclasses, the geometric centroids of all valid major classes in the set parameter space are first calculated, forming a centroid set. Subsequently, the nearest centroid principle is used to reclassify the abnormal subclasses, that is, the samples in the abnormal subclasses are reclassified. It is assigned to the nearest valid class. The allocation rule is expressed as follows:

[0017] (1)

[0018] in, Representative effective categories The coordinates of the centroid.

[0019] Thus, the clustering results after merging are obtained.

[0020] Step 1.4 involves re-numbering the category labels of the merged clustering results obtained in Step 1.3. All category labels retained after merging in Step 1.3 are remapped to a sequence of continuously increasing natural numbers starting from 1, ensuring the scalability of the operating condition categories and adaptability to different sample set sizes. The re-numbered category labels are then written into the valid engine operating data obtained in Step 1.3 as the operating condition number of the samples, ultimately yielding the operating condition category classification results.

[0021] Step 2: Construct a health baseline dataset for each operating condition category in the operating condition classification results. Use the health baseline dataset to standardize and preprocess all sample data from the operating condition classification results. Then, use the UMAP algorithm to perform non-linear dimensionality reduction on the standardized data to obtain a low-dimensional feature dataset. Specifically:

[0022] Step 2.1: First, set a minimum cycle threshold to ensure sufficient healthy operating period length. Filter the operating condition category classification results obtained in Step 1, extracting individual engines with sufficient healthy operating period length to avoid interference from engines with excessively short lifecycles in health period determination. Extract operating sample data of these engine individuals with sufficient healthy operating period length under healthy conditions, and archive them according to the operating condition number obtained in Step 1.4 to obtain the health baseline dataset for each operating condition category, which will serve as a reference for subsequent standardization.

[0023] Step 2.2: Based on the operating condition number, group all operating cycle samples from the operating condition category classification results according to the operating condition category. For samples in each operating condition category group, exclude invalid features containing a large number of missing values, stable and unchanging features, or features irrelevant to health assessment; the remaining features are valid. Calculate the mean and standard deviation of the valid features of the health baseline data obtained in Step 2.1 for each operating condition category, and standardize the valid features of all sample data within that operating condition category group accordingly. The standardized feature data has a mean of 0 and a variance of 1, ensuring a consistent distribution of different feature data within the same operating condition category and guaranteeing the comparability of different features. This standardization process is performed independently within each operating condition category, avoiding confusion between operating conditions. Finally, high-dimensional feature data is obtained.

[0024] Step 2.3: Input the standardized high-dimensional feature data from Step 2.2 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 is set from 1 to N, and a portion of representative samples from 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 dimensionality reduction results.

[0025] Step 2.4: For each of the N dimensionality reduction results obtained in Step 2.3, calculate the K-Nearest Neighbor Trustworthiness (KNN) index for each target dimension to evaluate the fidelity of the N dimensionality reduction results to the high-dimensional feature data structure obtained in Step 2.2. The KNN Trustworthiness index 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 they are, the higher the index value (closer to 1.0). Select the target dimension with the highest KNN Trustworthiness index value, i.e., the highest fidelity, as the final target dimension.

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

[0027] Step 3: A Gaussian mixture model is used to perform cluster analysis on the low-dimensional feature dataset. The average Mahalanobis distance between samples in each cluster and the health reference center is calculated to obtain the multi-level health grades 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, with a preset number of clusters representing the desired number of health level categories. The Gaussian Mixture Model (GMM) automatically learns the probability distribution characteristics of sample data under each operating condition category, using a covariance structure. GMM is a typical generative probabilistic model that assigns a soft probability of membership to each category to each sample, making it suitable for clustering modeling of sample data. GMM assumes that samples are generated by a mixture of multiple Gaussian distributions. It iteratively estimates the mean, covariance, and mixing coefficients of each distribution using the Expectation-Maximization (EM) algorithm, thereby characterizing the feature patterns under different health states and outputting the centroid mean vector, covariance matrix, and sample membership soft probability for each cluster. This process models each operating condition category independently, ensuring a close correlation between health level classification and operating condition characteristics. The overall model can be expressed as:

[0029] (2)

[0030] in, The sample vectors in the low-dimensional feature dataset obtained in step 2.5 have a dimension of . , for The probability density value across the entire model, i.e., "the probability of this point in the mixture distribution". This represents the number of Gaussian components, i.e., the number of clusters. For the first The weights of the Gaussian components satisfy... ,and , For the first a vector of mean values ​​of a Gaussian distribution. For the first The covariance matrix of a Gaussian distribution. It is the multivariate Gaussian distribution density function.

[0031] Parameters of Gaussian mixture model The EM algorithm is used for learning. The EM iteration consists of two steps:

[0032] (1) E-step (Expectation Step): Calculate the soft probability of membership of each sample 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 iterates through 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 for each work condition category are obtained, including the mean vector of each cluster center, the covariance matrix, and the cluster category and soft probability of each sample in the low-dimensional feature dataset obtained in step 2.5, which are used for subsequent level classification and health scoring.

[0036] Step 3.2: Under each operating condition category, the cluster center of the corresponding cluster in Step 3.1 of the samples in the health baseline dataset obtained in Step 2.1 is taken as the health reference center. In order to fully reflect the overall distribution characteristics of the cluster category and its overall deviation from the health status, and thus avoid the measurement bias 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 dataset obtained in Step 2.5. The Mahalanobis distance between each sample in the operating condition category and the health reference center in the operating condition category is calculated, and the cluster category to which these sampled samples belong in Step 3.1 is counted. Finally, the average Mahalanobis distance between all sampled samples in each cluster category and the health reference center is calculated to measure the operating status of each cluster category and the degree of deviation of its health status from the health reference center.

[0037] Step 3.3: Based on the average Mahalanobis distance of each cluster category obtained in Step 3.2, the health levels of the cluster categories are ranked. The larger the average Mahalanobis distance, the lower the health level. The final ranking result is the multi-level health level.

[0038] Step 4: By fusing the membership soft probabilities obtained in Step 3 with the Mahalanobis distances of individual samples, a continuous health score is constructed, and the health level determination interval is obtained. Specifically:

[0039] Step 4.1: Under each working condition category, using the Gaussian mixture model parameters obtained in Step 3.1, extract the cluster category and membership soft probability of each sample. Based on the health level obtained in Step 3.3, manually assign weights to each health level according to the cluster category. For any sample, its soft probability of membership is denoted as . The weighted initial health score formula, obtained by combining the soft probability of membership with the weights, is as follows:

[0040] (3)

[0041] To further enhance the interpretability and consistency of the scoring, Perform min-max normalization to map it to... Within the interval, the normalized initial health score is denoted as . .at this time, , which represents the degree of risk of an individual's health status as reflected by the cluster category and membership soft probability obtained in step 3.1.

[0042] Step 4.2: To reflect the deviation of each sample from the cluster center of its respective cluster category obtained in Step 3.1, calculate the Mahalanobis distance from the sample to the cluster center of its respective category. The individual Mahalanobis distances were standardized into a deviation index; the deviation index was then fused using an adjustable fusion coefficient. The normalized initial health score obtained in step 4.1 is denoted as Perform weighted fusion and truncate to A comprehensive continuous health score was obtained. The formula is as follows:

[0043] (4)

[0044] in, The fusion coefficient is adjustable. These are the minimum and maximum values ​​of the Mahalanobis distance for the sample individuals, respectively.

[0045] 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 comprehensive health score corresponding to the samples of each cluster category obtained in Step 3.1 is calculated in Step 4.2. The obtained comprehensive health score can be used as the health level determination interval corresponding to that 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.

[0046] The comprehensive health score and health classification results can be visualized in the time series dimension, showing the health evolution trend of individuals and groups.

[0047] Analysis of the innovativeness of this invention:

[0048] (1) Multi-condition unsupervised modeling and adaptive clustering assessment of health level: In view of the multi-condition characteristics of engine operation, a soft clustering model with hierarchical structure according to the condition is constructed. The GMM is used to model the standardized sensor data and automatically identify the health level category. By mapping the soft probability output with the health level, a continuous health score is generated, which avoids the dependence on fixed thresholds or labeled samples and improves the generalization and practicality of the model.

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

[0050] The beneficial effects of this invention are:

[0051] This invention employs an unsupervised learning method throughout, enabling accurate learning and assessment of aero-engine health status without relying on fault samples or remaining life labels required for training models using traditional methods. It automatically categorizes different operating conditions using a multi-condition dynamic clustering method and establishes corresponding multi-level health level assessment and continuous health score calculation methods. By integrating clustering categories from a Gaussian mixture model with individual Mahalanobis distance deviation indices, it significantly improves assessment accuracy. Even when modeling based on normal operating data, it can still effectively identify early performance degradation characteristics and demonstrate the evolution trend of engine health status, providing reliable preventive maintenance decision support for aero-engine fleets lacking historical fault data. Attached Figure Description

[0052] Figure 1 This is a flowchart of the method of the present invention;

[0053] Figure 2 This is a schematic diagram of the original sequence (without operating conditions) of a random sensor from an engine with the longest number of operating cycles.

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

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

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

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

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

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

[0060] Figure 9 This is a schematic diagram of the population health scoring and rating system of the present invention. Detailed Implementation

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

[0062] like Figure 1 As shown, a method for assessing the health of aero-engine populations based on multi-condition dynamic clustering includes the following steps:

[0063] Step 1: Using the HDBSCAN algorithm, cluster analysis is performed on the set parameters in the engine operating data to achieve adaptive classification of operating condition categories, and small-scale abnormal clusters are merged to obtain stable operating condition category classification results; specifically:

[0064] Step 1.1: Preprocess the collected raw engine operating data to extract the subset of set parameters that characterize the engine operating environment and control state contained in each engine operating cycle sample. ,in For the sample size, The feature dimension is defined as follows. In this embodiment, the feature subset of the set parameters includes flight altitude, Mach number, and throttle angle. This set of set parameters is normalized to obtain a normalized set of set parameters, ensuring that set parameters with different dimensions and value ranges have the same weight during clustering, thus avoiding clustering bias caused by differences in feature scale.

[0065] Step 1.2: Based on the normalized subset of parameters obtained in Step 1.1, clustering is performed using the unsupervised HDBSCAN algorithm. This algorithm identifies cluster structures layer by layer by constructing density reachability relationships between sample points in the feature space. Compared to traditional clustering methods, the HDBSCAN algorithm only requires setting two parameters: minimum cluster size (min_cluster_size) and minimum number of samples (min_samples). It does not require pre-specifying the number of categories, i.e., the number of clusters, but is automatically determined based on the data distribution. Furthermore, it can adapt to non-spherical cluster structures and marks isolated points that do not meet the density requirements as noise points, improving the robustness of condition category classification. In this embodiment, min_cluster_size = 50 and min_samples = 10 are set. The clustering yields a set of category labels composed of integer numbers. ,in The number of working condition categories is automatically determined, and -1 represents noise points. Each sample in the normalized parameter subset corresponds to a category label in the category label set, thus obtaining the preliminary clustering results.

[0066] Step 1.3: Based on the preliminary clustering results obtained in Step 1.2, to ensure clustering quality, all noise point samples in the preliminary clustering results are removed from the original engine operation data in Step 1.1, resulting in valid engine operation data. Then, all non-noise point samples in the preliminary clustering results are statistically analyzed. If the number of samples under a certain category label is too small (in this embodiment, "too small" is set to less than 0.5% of the total number of non-noise point samples), then that category is marked as an abnormal sub-category, and the remaining categories are valid major categories. For these abnormal sub-categories, the geometric centroids of all valid major categories in the set parameter space are first calculated to form a centroid set; then, the abnormal sub-categories are reclassified using the nearest centroid principle, i.e., the samples in the abnormal sub-categories are reclassified. It is assigned to the nearest valid class. The allocation rule is expressed as follows:

[0067] (1)

[0068] in, Representative effective categories The coordinates of the centroid.

[0069] Thus, the clustering results after merging are obtained.

[0070] Step 1.4 involves re-numbering the category labels of the merged clustering results obtained in Step 1.3. All category labels retained after merging in Step 1.3 are remapped to a sequence of continuously increasing natural numbers starting from 1, ensuring the scalability of the operating condition categories and adaptability to different sample set sizes. The re-numbered category labels are then written into the valid engine operating data obtained in Step 1.3 as the operating condition number of the samples, ultimately yielding the operating condition category classification result. In this embodiment, random sensor data from the engine with the longest historical number of original engine operating data cycles is taken, and a comparison is made before and after the operating condition category classification. Figures 2-8 As shown, it is clear that after clustering and classifying the operating conditions, the decline trend of the engine with the increase of the number of operating cycles under each operating condition is more clearly shown, significantly eliminating data jumps and interference caused by changes in operating conditions.

[0071] Step 2: Construct a health baseline dataset for each operating condition category in the operating condition classification results. Use the health baseline dataset to standardize and preprocess all sample data from the operating condition classification results. Then, use the UMAP algorithm to perform non-linear dimensionality reduction on the standardized data to obtain a low-dimensional feature dataset. Specifically:

[0072] Step 2.1: First, a minimum cycle threshold is set to ensure sufficient healthy operating period length (based on common practices in the field of aero-engine remaining life prediction and the cycle distribution characteristics of the dataset used in this embodiment, the minimum cycle threshold is set to 125). The operating condition category classification results obtained in Step 1 are then filtered to extract individual engines with sufficient healthy operating period length, avoiding interference from engines with excessively short lifecycles in the health period determination. In this embodiment, the first 20% of the operating period history of these individual engines with sufficient healthy operating period length is extracted and regarded as operating sample data under healthy conditions. These are then archived according to the operating condition number obtained in Step 1.4 to obtain the health baseline dataset for each operating condition category, which serves as a reference for subsequent standardization.

[0073] Step 2.2: Based on the operating condition number, group all operating cycle samples from the operating condition category classification results according to the operating condition category. For samples in each operating condition category group, exclude invalid features containing a large number of missing values, stable and unchanging features, or features irrelevant to health assessment; the remaining features are valid. Calculate the mean and standard deviation of the valid features of the health baseline data obtained in Step 2.1 for each operating condition category, and standardize the valid features of all sample data within that operating condition category group accordingly. The standardized feature data has a mean of 0 and a variance of 1, ensuring a consistent distribution of different feature data within the same operating condition category and guaranteeing the comparability of different features. This standardization process is performed independently within each operating condition category, avoiding confusion between operating conditions. Finally, high-dimensional feature data is obtained.

[0074] Step 2.3: Input the standardized high-dimensional feature data from Step 2.2 into the UMAP 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 for dimensionality reduction is set from 1 to N, and a portion of representative samples from the high-dimensional feature data are randomly sampled to avoid memory explosion. In this embodiment, considering the dimensionality of the high-dimensional feature data is 14, N is set to 10, and the number of randomly sampled samples is set to 3000. Multiple rounds of dimensionality reduction structure learning are performed on the randomly sampled samples according to different target dimensions to obtain N dimensionality reduction results.

[0075] Step 2.4: Calculate the K-Nearest Neighbors (KNN) trustworthiness index for each of the N dimensionality reduction results obtained in Step 2.3, evaluating the fidelity of the N dimensionality reduction results to the high-dimensional feature data structure obtained in Step 2.2. The KNN trustworthiness index 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 they are, the higher the index value (closer to 1.0). The target dimension with the highest KNN trustworthiness index value, i.e., the highest fidelity, is selected as the final target dimension. In this embodiment, the final target dimension is 7.

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

[0077] Step 3: A Gaussian mixture model is used to perform cluster analysis on the low-dimensional feature dataset. The average Mahalanobis distance between samples in each cluster and the health reference center is calculated to obtain the multi-level health grades 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, with a preset number of clusters representing the desired number of health level categories. The covariance structure automatically learns the probability distribution characteristics of sample data under each working condition category. In this embodiment, a preset structure is used. The three health levels are represented by three values: healthy, monitoring, and warning. The covariance structure is set as a full covariance matrix to capture the correlation between features. Gaussian mixture model (GM) is a typical generative probabilistic model that assigns a soft probability of membership to each category to each sample, suitable for clustering modeling of sample data. GM assumes that samples are generated by mixing multiple Gaussian distributions. It iteratively estimates the mean, covariance, and mixing coefficients of each distribution using the EM algorithm, thereby characterizing the feature patterns under different health states and outputting the mean vector, covariance matrix, and soft probability of membership for each cluster. This process models each operating condition category independently, ensuring a close correlation between health level classification and operating condition characteristics. The overall model can be expressed as:

[0079] (2)

[0080] in, The sample vectors in the low-dimensional feature dataset obtained in step 2.5 have a dimension of . , for The probability density value across the entire model, i.e., "the probability of this point in the mixture distribution". This represents the number of Gaussian components, i.e., the number of clusters. For the first The weights of the Gaussian components satisfy... ,and , For the first a vector of mean values ​​of a Gaussian distribution. For the first The covariance matrix of a Gaussian distribution. It is the multivariate Gaussian distribution density function.

[0081] Parameters of Gaussian mixture model The EM algorithm is used for learning. The EM iteration consists of two steps:

[0082] (1) E-step (ExpectationStep): Calculate the soft probability of membership of each sample 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 iterates through 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 for each work condition category are obtained, including the mean vector of each cluster center, the covariance matrix, and the sample membership soft probability, which are used for subsequent level classification and health scoring.

[0086] Step 3.2: Under each operating condition category, the cluster center of the corresponding cluster in Step 3.1 of the samples in the health baseline dataset obtained in Step 2.1 is taken as the health reference center. In order to fully reflect the overall distribution characteristics of the cluster category and its overall deviation from the health status, thereby avoiding the measurement bias 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 dataset 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 health reference center under the operating condition category is calculated, and the cluster category to which these sampled samples belong in Step 3.1 is counted. Finally, the average Mahalanobis distance between all sampled samples under each cluster category and the health reference center is calculated to measure the operating status of each cluster category and the degree of deviation of its health status from the health reference center.

[0087] Step 3.3: Based on the average Mahalanobis distance of each cluster category obtained in Step 3.2, the cluster categories are ranked by health level. A larger average Mahalanobis distance indicates a lower health level. The final ranking result is the multi-level health level. In this embodiment, the final three health levels, from highest to lowest, are healthy, monitoring, and warning, 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): This category corresponds to the middle position in the sorting and is used to represent an operating state where performance deviates but has not yet seriously deteriorated;

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

[0091] Step 4: By fusing the membership soft probabilities obtained in Step 3 with the Mahalanobis distances of individual samples, a continuous health score is constructed, and the health level determination interval is obtained. Specifically:

[0092] Step 4.1: Under each working condition category, using the Gaussian mixture model parameters obtained in Step 3.1, extract the cluster category and membership soft probability of each sample. Based on the health level obtained in Step 3.3, manually assign weights to each health level according to the cluster category. In this embodiment, weight values ​​are assigned to the three levels: health, monitoring, and early warning. For any sample, its soft probability of membership is denoted as . The weighted initial health score formula, obtained by combining the soft probability of membership with the weights, is as follows:

[0093] (3)

[0094] To further enhance the interpretability and consistency of the scoring, Perform min-max normalization to map it to... Within the interval, the normalized initial health score is denoted as . .at this time, This represents the degree of individual health status risk reflected by the clustering categories and membership soft probabilities obtained in step 3.1. In this embodiment, it corresponds to the specific classification in step 3.3. Corresponds to pure health (completely at level 0). This corresponds to a severe anomaly (entirely level 2).

[0095] Step 4.2: To reflect the deviation of each sample from the cluster center of its respective cluster category obtained in Step 3.1, calculate the Mahalanobis distance from the sample to the cluster center of its respective category. The individual Mahalanobis distance is standardized into a deviation index; the deviation index is then denoted by the normalized initial health score obtained in step 4.1, using an adjustable fusion coefficient. Perform weighted fusion and truncate to A comprehensive continuous health score was obtained. The formula is as follows:

[0096] (4)

[0097] in, The fusion coefficient is adjustable; in this embodiment, it is set to 0.8. These are the minimum and maximum values ​​of the Mahalanobis distance for the sample individuals, respectively.

[0098] 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 comprehensive health score corresponding to the samples of each cluster category obtained in Step 3.1 is calculated in Step 4.2. The obtained comprehensive health score can be used as the health level determination interval corresponding to that 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 features in the cluster, which can reflect the health level attribution of the data in the cluster structure and quantify its relative deviation from the healthy samples, thereby more comprehensively and robustly evaluating the individual status and ultimately realizing a closed loop from the cluster model to the score interpretation. The comprehensive health score and health classification 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 selected for visualization, and the results are as follows. Figure 9 As shown.

[0099] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An aero-engine population health evaluation method based on multi-condition dynamic clustering, characterized in that, The aero-engine population health evaluation method comprises the following steps: Step 1, using unsupervised density-based hierarchical clustering algorithm, clustering analysis is carried out on the set parameters in the engine operation data, the adaptive division of working condition categories is realized, and the small-scale abnormal clusters are merged, and the stable working condition category division result is obtained; Step 2, health baseline data set is constructed under each working condition category in the working condition category division result, all sample data in the working condition category division result are standardized pretreated by using the health baseline data set, and the UMAP algorithm is used for nonlinear dimension reduction of the standardized data, and finally the low-dimensional feature data set is obtained; Specifically: Step 2.1, first, set a minimum cycle threshold, filter the working condition category division result obtained in step 1, determine the engine individual with the length of health period operation cycle, and extract the running sample data of the engine individual in the health state, according to the running condition number, respectively archive, obtain the health baseline data set under each working condition category; Step 2.2, according to the running condition number, all running cycle samples in the working condition category division result are grouped according to the working condition category, standardized, and high-dimensional feature data are obtained; Step 2.3, the high-dimensional feature data standardized in step 2.2 are input into the UMAP algorithm; set the dimension reduction target dimension range from 1 to N, and take part of the representative samples in the high-dimensional feature data by random sampling; according to different target dimensions, the random sampling samples are subjected to multi-round dimension reduction structure learning, and N kinds of dimension reduction results are obtained; Step 2.4, the K nearest neighbor confidence index of each target dimension is calculated for the N kinds of dimension reduction results obtained in step 2.3, the fidelity of N kinds of dimension reduction results to the high-dimensional feature data structure obtained in step 2.2 is evaluated, and the target dimension with the highest K nearest neighbor confidence index value, i.e. the highest fidelity, is selected as the final target dimension; Step 2.5, according to the final target dimension obtained in step 2.4, the high-dimensional feature data obtained in step 2.2 are input into the UMAP algorithm, and the low-dimensional feature data set after dimension reduction is obtained; Step 3, Gaussian mixture model is used for clustering analysis of the low-dimensional feature data set obtained in step 2, and the average Mahalanobis distance between the samples of each clustering category and the health reference center is calculated, and the multi-level health grades corresponding to different clustering categories are obtained; Specifically: Step 3.1, input the low-dimensional feature dataset obtained in step 2.5 into a Gaussian mixture model for unsupervised clustering, and preset the number of clusters representing the number of health level categories desired and covariance structure, automatically learn the probability distribution characteristics of the sample data under each working condition category; independently model each working condition category, ensuring that the health level division is closely related to the working condition characteristics; EM algorithm is used for learning, and finally the Gaussian mixture model parameters under each working condition category are obtained, including the clustering center mean vector, the covariance matrix and the clustering category and membership soft probability of each sample in the low-dimensional feature data set obtained in step 2.5, which are used for subsequent grade division and health score. Step 3.2, in each working condition category, the cluster center corresponding to the cluster to which the sample in the health baseline data set obtained in step 2.1 belongs in step 3.1 is taken as a health reference center; based on the low-dimensional feature data set obtained in step 2.5, a part of samples in each working condition category are randomly extracted, the Mahalanobis distance between each extracted sample and the health reference center in the working condition category is calculated, and the cluster category to which the extracted sample belongs in step 3.1 is counted, and finally the average Mahalanobis distance between all extracted samples in each cluster category and the health reference center is calculated to measure the running state of each cluster category and the deviation degree of the health state of each cluster category from the health reference center; Step 3.3, according to the size of the average Mahalanobis distance of each cluster category obtained in step 3.2, the cluster categories are sorted in health level, and the larger the average Mahalanobis distance, the lower the health level, and finally the sorting result is obtained, that is, the multi-level health level; Step 4, a continuous health score is constructed by fusing the membership soft probability and the individual Mahalanobis distance of the sample, and a health level judgment interval is obtained; Specifically: Step 4.1, under each working condition category, using the Gaussian mixture model parameters obtained in step 3.1, extract the clustering category and membership soft probability of each sample, and according to the health level obtained in step 3.3, set the weight of each level artificially For any sample, the membership soft probability is denoted as The weighted initial health score formula obtained by combining the membership soft probability and the weight is as follows: （3） min-max normalization processing is performed, so as to be mapped into interval, and the normalized initial health score is recorded as ; at this time, , indicates the individual health state risk degree reflected by the clustering category and membership soft probability obtained in step 3.1.​ Step 4.2, to reflect the deviation degree of each engine running cycle sample individual from the cluster center of the belonging cluster category obtained in step 3.1, the individual Mahalanobis distance of the sample to the cluster center of the belonging category is calculated , and the individual Mahalanobis distance is standardized into a deviation degree index; the deviation degree index is weighted and fused with the normalized initial health score obtained in step 4.1 according to an adjustable fusion coefficient , and recorded as ; the weighted fusion is truncated to , and a comprehensive continuous health score is obtained, and the formula is as follows: （4） wherein, is an adjustable fusion coefficient, are the minimum and maximum values of the Mahalanobis distance of the sample individual, respectively. 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 judgment interval is obtained.

2. The method of claim 1, wherein, The step 1 is specifically: Step 1.1, pre-processing the original engine operation data collected, extracting a subset of all set parameters representing the engine operation environment and control state in each engine operation cycle sample wherein is the sample number, is the feature dimension; The set parameter subset is normalized to obtain a normalized set parameter subset; Step 1.2, based on the set parameter subset, using unsupervised density-based hierarchical clustering algorithm, referred to as HDBSCAN algorithm, clustering to obtain a class label set consisting of integer numbers wherein is the automatically determined number of working condition categories, -1 represents a noise point, and each sample in the normalized set parameter subset corresponds to a class label in the class label set, to obtain a preliminary clustering result; Step 1.3, all noise point samples in the preliminary clustering results are removed from the corresponding samples in the original engine operation data in step 1.1 to obtain effective engine operation data; then, all non-noise point samples in the preliminary clustering results are counted to obtain an abnormal small class and an effective large class; first, the geometric centroids of all effective large classes in the set parameter space are calculated to form a centroid set; then, the abnormal small class is reclassified by using the nearest centroid principle, and the samples in the abnormal small class are assigned to the effective large class closest to them The assignment rule is represented as:​ （1） wherein, representative effective broad class of the centroid coordinates; Thus, the merged clustering result is obtained; Step 1.4, the cluster result obtained in step 1.3 is renumbered, and all cluster labels are remapped to a natural number sequence starting from 1 and continuously increasing; the new cluster label after renumbering is written into the effective engine operation data obtained in step 1.3 as the operation condition number of the sample, and finally the working condition category division result is obtained.

3. The method of claim 1, wherein, The step 2.2 is specifically: For the samples in each working condition category group, the invalid features containing a large number of missing values, being stable and unchanged or irrelevant to health evaluation 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 based on this, and the standardization process is independently carried out in each working condition category, and finally the high-dimensional feature data is obtained.

4. The method of claim 1, wherein, In step 3.1: The Gaussian mixture model is expressed as: （2） wherein, is a sample vector within a low-dimensional feature dataset, with dimension ; is a probability density value under a Gaussian mixture model; is the number of Gaussian components, representing the number of clusters; is the weight of the th Gaussian component, satisfying , and ; ; is the mean vector of the th Gaussian distribution; is the covariance matrix of the th Gaussian distribution; is a multivariate Gaussian distribution density function; The learning by the EM algorithm is specifically: learning parameters of a Gaussian mixture model by the EM algorithm The EM iteration is divided into two steps: (1) calculate the membership soft probability of each sample belonging to each Gaussian component; (2) Using membership soft probabilities to update weights , mean , covariance ; The above two steps are iterated until the log-likelihood of the Gaussian mixture model converges or the maximum iteration number is reached.

5. The method of claim 1, wherein, The step 4.3 is specifically: The samples in each cluster category obtained in step 3.1 have corresponding comprehensive health scores calculated in step 4.2, and the comprehensive health score is taken as the health level judgment interval corresponding to the cluster category, that is, the comprehensive health score set of a health level in a working condition category can be directly taken as the reference of the health level judgment interval; The comprehensive health score and the health grading result are visualized in the time series dimension to show the health evolution trend of individuals and groups.

Images