A method for extracting degradation features of an aircraft engine based on physical constraints

Abstract

The application discloses a physical constraint-based aircraft engine degradation feature extraction method, and relates to an aircraft engine degradation feature extraction method.The application is used for solving the problem that the wing monitoring signal aliasing leads to feature distortion.The steps of the application comprise the following steps: step 1, acquiring a multivariate sensor time series signal of an aircraft engine during in-flight operation; step 2, constructing a mapping network based on a Transformer encoder; step 3, defining a physical constraint loss function; step 4, constructing a joint objective function to ensure that the feature retains the essential information of the original signal; and step 5, applying the trained encoder to in-flight monitoring data to extract and output decoupled features with physical consistency in real time.The features eliminate severe sawtooth noise and present clear performance degradation trajectories.The application belongs to the technical field of aero-engine maintenance.

Description

Technical Field

[0001] This invention relates to a method for extracting degradation features of aircraft engines, belonging to the field of aero-engine maintenance technology. Background Technology

[0002] With the rapid development of aircraft health management technology, the extraction of degradation features from on-wing monitoring signals of aircraft engines has become a key prerequisite for predictive maintenance. However, in actual operation, aircraft engines are affected by complex operating conditions, atmospheric environment, and sensor measurement errors, resulting in highly nonlinear and noisy multivariate monitoring data (such as temperature, pressure, and speed).

[0003] Existing methods for extracting degradation features from aircraft engines have the following drawbacks:

[0004] Feature overlap and lack of discriminativeness: Existing data-driven feature extraction methods (such as conventional autoencoders) often fail to distinguish between operating condition fluctuations and actual performance degradation when processing variable operating condition data. This results in overlapping features in the spatial distribution of different degradation stages, directly leading to the failure of subsequent models.

[0005] Evolutionary trajectory violates physical laws: Traditional models lack constraints on the physical degradation process during feature extraction, which may lead to non-monotonic random fluctuations in the extracted degradation features, or even trends that are opposite to the actual performance degradation of the engine.

[0006] Lack of stable decoupling mechanism: Due to the lack of constraints on the stability of feature evolution, the extracted feature trajectories exhibit violent non-physical sawtooth fluctuations, which cannot provide stable performance benchmark data for subsequent accurate prediction, resulting in reduced credibility of maintenance decisions. Summary of the Invention

[0007] To address the problem of feature distortion caused by aliasing of on-wing monitoring signals, this invention proposes a method for extracting aircraft engine degradation features based on physical constraints.

[0008] The technical solution adopted by the present invention to solve the above problems is as follows: The steps of the present invention include: Step 1: Acquire timing signals from multiple sensors during wing-mounted aircraft engine operation; Step 2: Construct a mapping network based on the Transformer encoder; Step 3: Define the physical constraint loss function; Step 4: Construct a joint objective function to ensure that the features retain the essential information of the original signal; Step 5: Apply the trained encoder to the on-wing monitoring data to extract and output physically consistent decoupling features in real time; these features eliminate severe jagged noise and present a clear performance degradation trajectory.

[0009] Furthermore, in step 1, the time-series signals from the independent sensor include compressor temperature, low-pressure turbine temperature, fan speed, and engine pressure ratio; a sliding window technique is used to sample and slice the time-series signals to construct an input sample matrix, and then... Standardization methods eliminate dimensional differences between different sensor parameters, ensuring the balance of feature weights during subsequent decoupling.

[0010] Furthermore, step 2 specifically includes: Step 201: Use a linear embedding layer to perform dimensional transformation on the standardized data and introduce position coding to preserve the temporal position information of the sensor signal; Step 202: Utilize multiple attention heads to calculate the spatiotemporal dependencies between multi-sensor parameters in parallel, and automatically identify key components sensitive to engine degradation by calculating the Query, Key, and Value matrices. Step 203: Map the high-dimensional features encoded by the attention mechanism to a low-dimensional latent space feature manifold through a fully connected layer; remove redundant information related to operating condition fluctuations from the original monitoring signal, and retain only the core manifold features that can characterize performance degradation.

[0011] Furthermore, step 3 specifically includes: Step 301, Trend Monotonicity Constraint: Introduce a monotonicity index. By comparing the amplitude of feature points in adjacent time steps, if the feature evolution trend contradicts the known physical laws of engine wear, a penalty value is applied. Step 302, Manifold Smoothing Constraint: Calculate the Euclidean distance between feature points in the latent space, and filter out non-stationary impulse interference caused by sudden changes in the environment such as atmospheric temperature and altitude by limiting the rate of manifold change; Step 303, Temporal Correlation Weight: Combining attention weight information, the low-dimensional feature space is divided into degradation-sensitive region and noise-sensitive region. By manifold regularization, degradation features are forced to decouple from operating condition interference signals, and a smooth and discriminative performance degradation trajectory is output.

[0012] Furthermore, in step 4, the Adan optimizer is used to update the network parameters. Through multiple iterations, the model can strictly adhere to physical constraints while satisfying the data representation requirements.

[0013] The beneficial effects of this invention are: 1. This invention utilizes attention mechanisms and low-dimensional mapping to achieve feature purification, solving the problems of feature overlap and lack of discriminability. Addressing the deficiency of existing technologies in distinguishing between operating condition fluctuations and performance degradation, this invention introduces a multi-head self-attention mechanism in step ② to automatically identify and enhance key components sensitive to degradation, while simultaneously removing redundant operating condition interference through low-dimensional spatial projection. This mechanism ensures that the extracted features focus on the essential performance evolution, giving features at different degradation stages clear discriminative boundaries in the latent space, thus avoiding model failure caused by feature overlap.

[0014] 2. This invention introduces a physical monotonicity penalty mechanism to correct the phenomenon of trajectories violating physical laws. Addressing the problem that traditional models often exhibit non-monotonic fluctuations or even contradict actual decay patterns when extracting features, this invention constructs a physical monotonicity constraint function in step ③. This constraint term can impose real-time penalties on feature evolution that violates the physical logic of engine degradation, forcibly guiding the feature trajectory back to a monotonic evolution path, ensuring a high degree of consistency between feature evolution and the actual physical decay process of the engine.

[0015] 3. This invention achieves deep decoupling through manifold consistency constraints, eliminating sawtooth fluctuations and improving decision-making reliability. Addressing the problem of existing technologies lacking stability constraints, leading to sawtooth-like non-physical fluctuations in feature trajectories, this invention utilizes manifold distance consistency calculations in step ③ to limit the evolution rate of features in the latent space. This decoupling mechanism effectively filters out non-stationary environmental noise and sensor zero-point drift interference, outputting smooth and stable performance benchmark data, providing highly reliable input for subsequent accurate lifetime prediction and maintenance decisions. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention; Figure 2 This is an application effect diagram of the present invention. Detailed Implementation

[0017] Specific implementation method one: as follows Figure 1 and Figure 2 As shown, the steps of the physical constraint-based aircraft engine degradation feature extraction method described in this embodiment include: Step 1: Acquire multi-sensor timing signals of the aircraft engine during wing operation. This includes parameters such as compressor temperature, low-pressure turbine temperature, fan speed, and engine pressure ratio. A sliding window technique is used to sample and slice the time-series signal, with a window length of [missing information]. Step size is Construct the input sample matrix; utilize Standardization methods eliminate dimensional differences between different sensor parameters, ensuring the balance of feature weights during subsequent decoupling. Step 2: Construct a mapping network based on a Transformer encoder. First, use a linear embedding layer to transform the dimensionality of the standardized data and introduce position encoding to retain the temporal position information of the sensor signals. Then, use multiple attention heads to compute the spatiotemporal dependencies between multivariate sensor parameters in parallel. By calculating the Query, Key, and Value matrices, key components sensitive to engine degradation are automatically identified. Finally, map the high-dimensional features encoded by the attention mechanism to a low-dimensional latent space feature manifold through a fully connected layer. This step aims to remove redundant information related to operating condition fluctuations from the original monitoring signals and retain only the core manifold features that can characterize performance degradation. Step 3: To decouple the true degradation trend from the aliased signal, define a physical constraint loss function. The specific algorithm is as follows: Step 301, Trend Monotonicity Constraint: Introduce a monotonicity index by comparing feature points at adjacent time steps. and If the amplitude of the characteristic evolution trend contradicts the known physical laws of engine wear, a penalty value is applied. Step 302, Manifold Smoothing Constraint: Calculate the Euclidean distance between feature points in the latent space. By limiting the rate of manifold change, non-stationary pulse interference caused by sudden changes in the environment such as atmospheric temperature and altitude can be filtered out. Step 303, Temporal Correlation Weight: Combining attention weight information, the low-dimensional feature space is divided into degradation-sensitive region and noise-sensitive region. By manifold regularization, degradation features are forced to decouple from operating condition interference signals, and a smooth and discriminative performance degradation trajectory is output. Step 4: Construct the joint objective function ,in, The data reconstruction loss ensures that the features retain the essential information of the original signal. and The weights represent the weights of the loss function, which consists of the MSE function (data loss function) and the physical loss function. The system is composed of components that optimize the weight ratio of data loss and physical loss through network training to obtain the best output. At the same time, the Adan optimizer is used to update the network parameters, and through multiple rounds of iteration, the model can strictly adhere to physical constraints while satisfying the data representation. Step 5: Apply the trained encoder to the on-wing monitoring data to extract and output physically consistent decoupling features in real time. These features eliminate severe jagged noise and reveal a clear performance degradation trajectory.

[0018] Example Step 1: Data Access and Standardization Processing Step 101: Obtain sensor data streams throughout the engine's entire life cycle and select 14 key performance parameters, including compressor temperature, low-pressure turbine temperature, fan speed, and engine pressure ratio. Step 102: Set the length of the sliding window sampling points, step size This transforms a one-dimensional time series into a feature matrix. Step 103: Perform Z-Score standardization. The calculation formula is as follows: , This is the original data. The mean, Standard deviation This is the standardized value. This is the formula for standardizing data. Its function is to eliminate the dimensional differences between different sensor data and map all sensor data into a space with a mean of 0 and a variance of 1. Step 2: Feature-aware encoding and dimensionality compression The standardized matrix is ​​input into the Transformer encoder; through a multi-head self-attention module, the model assigns different attention weights to each parameter; the attention mechanism calculates... It automatically identifies and enhances features that are sensitive to engine wear, while also identifying and suppressing high-frequency components affected by ambient atmospheric temperature. The encoded features are compressed from high-dimensional to 3-dimensional latent space through a fully connected bottleneck layer, thereby eliminating redundancy in operating conditions. Step 3: Application of physical constraint functions and feature decoupling Calculating the manifold distance of feature points at consecutive time steps in the latent space Perform a physical monotonicity check: Define a decision function; if the extracted feature value is within the range of... Always display performance is better than At the moment when irreversible degradation physical laws are violated, the deviation value is calculated and used as the loss function. A portion of it propagates backward, using physical constraints to force the latent space manifold to evolve in a single direction; Step 4: Set model training parameters The learning rate was set to 0.001, and the physical constraint weight coefficient β = 0.5. After 300 iterations, the network converged.

[0019] Performance Verification: This dataset contains 6 complex operating conditions and 2 failure modes. The performance after feature extraction is as follows: Figure 2As shown, as the remaining life decreases, the feature trajectories extracted by this invention can clearly distinguish between the two failure modes, enabling the features to truly reflect the performance degradation of the engine; while without the application of this invention, it is difficult to distinguish the failure modes, which will lead to inaccurate model predictions.

[0020] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A method for extracting degradation features of aircraft engines based on physical constraints, characterized in that, The specific steps include: Step 1: Acquire timing signals from multiple sensors during wing-mounted aircraft engine operation; Step 2: Construct a mapping network based on the Transformer encoder; Step 3: Define the physical constraint loss function; Step 4: Construct a joint objective function to ensure that the features retain the essential information of the original signal; Step 5: Apply the trained encoder to the on-wing monitoring data to extract and output physically consistent decoupling features in real time; these features eliminate severe jagged noise and present a clear performance degradation trajectory.

2. The method for extracting degradation features of aircraft engines based on physical constraints according to claim 1, characterized in that, Step 1 involves using the time-series signals from the independent sensor, including compressor temperature, low-pressure turbine temperature, fan speed, and engine pressure ratio. A sliding window technique is employed to sample and slice the time-series signals, constructing an input sample matrix. Standardization methods eliminate dimensional differences between different sensor parameters, ensuring the balance of feature weights during subsequent decoupling.

3. The method for extracting degradation features of aircraft engines based on physical constraints according to claim 1, characterized in that, Step 2 specifically includes: Step 201: Use a linear embedding layer to perform dimensional transformation on the standardized data and introduce position coding to preserve the temporal position information of the sensor signal; Step 202: Utilize multiple attention heads to calculate the spatiotemporal dependencies between multi-sensor parameters in parallel, and automatically identify key components sensitive to engine degradation by calculating the Query, Key, and Value matrices. Step 203: Map the high-dimensional features encoded by the attention mechanism to a low-dimensional latent space feature manifold through a fully connected layer; remove redundant information related to operating condition fluctuations from the original monitoring signal, and retain only the core manifold features that can characterize performance degradation.

4. The method for extracting degradation features of aircraft engines based on physical constraints according to claim 1, characterized in that, Step 3 specifically includes: Step 301, Trend Monotonicity Constraint: Introduce a monotonicity index. By comparing the amplitude of feature points in adjacent time steps, if the feature evolution trend contradicts the known physical laws of engine wear, a penalty value is applied. Step 302, Manifold Smoothing Constraint: Calculate the Euclidean distance between feature points in the latent space, and filter out non-stationary impulse interference caused by sudden changes in the environment such as atmospheric temperature and altitude by limiting the rate of manifold change; Step 303, Temporal Correlation Weight: Combining attention weight information, the low-dimensional feature space is divided into degradation-sensitive region and noise-sensitive region. By manifold regularization, degradation features are forced to decouple from operating condition interference signals, and a smooth and discriminative performance degradation trajectory is output.

5. The method for extracting degradation features of aircraft engines based on physical constraints according to claim 1, characterized in that, In step 4, the Adan optimizer is used to update the network parameters. Through multiple iterations, the model can meet the data representation requirements while strictly adhering to physical constraints.

Images