Aero-engine working condition invariant degradation detection method based on physical constraint causal decoupling
By constructing a thermodynamic mapping network to decouple operating condition effects from degradation signals, and establishing a dual-channel health index that remains unchanged under operating conditions, the problem of high false alarm rate in degradation detection of aero-engines under multiple operating conditions is solved. This enables accurate identification of degradation signals and cross-engine application without relying on a complete physical model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XI AN JIAOTONG UNIV
- Filing Date
- 2026-04-29
- Publication Date
- 2026-07-31
AI Technical Summary
Existing methods for monitoring the health of aero-engines are unable to effectively decouple operating conditions from degradation signals under multiple operating conditions, resulting in a high false alarm rate for degradation detection, and there is a lack of effective methods that do not rely on a complete physical model.
By constructing a multilayer perceptron network as a thermodynamic mapping network, the causal mapping relationship between operating conditions and physical coupling parameters is learned, a dual-channel health index with unchanged operating conditions is established, and physical residual and environmental anomaly information are integrated to achieve causal decoupling between operating condition effects and degradation signals.
It achieves accurate identification of degradation signals under multiple operating conditions, reduces false alarm rate, and the lightweight model supports cross-engine migration applications.
Smart Images
Figure CN122490194A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of aero-engine fault diagnosis and health monitoring technology, specifically relating to an aero-engine condition invariant degradation detection method based on physical constraint causal decoupling. Background Technology
[0002] Aero engines are the core power units of modern aircraft, and their reliability is directly related to flight safety. As the cumulative operating hours of an engine increase, degradation mechanisms such as turbine blade oxidation and corrosion, increased sealing clearances, and combustion chamber carbon buildup cause engine performance to gradually deviate from the design baseline. Accurately detecting the onset and progression of degradation is of significant engineering importance for developing condition-based maintenance strategies and reducing the risk of unplanned downtime. Currently, aero engine health monitoring mainly relies on sensor parameters recorded by the electronic control unit, including physically coupled parameters such as turbine exhaust temperature, engine speed, and fuel pump displacement. However, these parameters have a strong thermodynamic coupling relationship with thrust conditions: when the pilot adjusts the thrust setting, the normal parameter changes are much larger than the slow drift caused by degradation, causing degradation signals to be masked by changes in operating conditions.
[0003] Traditional methods typically discretize the operating space into several intervals, independently establishing parameter baselines and calculating deviations within each interval. This approach is highly sensitive to the granularity of discretization: coarse granularity can lead to the mixing of different operating points within the same interval, resulting in blurred baselines; while fine granularity can result in insufficient samples in some intervals, leading to unstable statistical estimations. In recent years, machine learning methods, such as isolated forests and long short-term memory network autoencoders, have been introduced into the field of aero-engine anomaly detection. Although these methods have strong data fitting capabilities, their black-box nature prevents the models from explicitly modeling the physical causal structure from thrust conditions to sensor parameters, resulting in a high false alarm rate in real-world operating scenarios with frequent operating condition changes.
[0004] Physical model-driven methods can calculate nominal parameter values under various operating conditions based on thermodynamic cycle theory, theoretically allowing for precise isolation of operating condition effects. However, establishing high-fidelity thermodynamic cycle models for aero-engines requires detailed component characteristic diagrams, cooling gas path models, and combustion efficiency curves, among other parameters. These parameters are often proprietary technologies of engine manufacturers and difficult to obtain. Therefore, effectively decoupling operating condition effects from electronic control unit data without relying on complete physical model parameters has become a critical problem that urgently needs to be solved in the field of aero-engine health monitoring.
[0005] In recent years, data-driven methods combined with physical prior knowledge have become the forefront of research in this field. Hanachi H et al. systematically reviewed gas turbine health monitoring methods based on performance parameters in the paper "Performance-Based Gas Turbine Health Monitoring, Diagnostics, and Prognostics: A Survey" (IEEE Transactions on Reliability, 2018, 67(3): 1340-1363), pointing out the core challenge of operating condition normalization. Cheng K et al. proposed a health assessment method that combines physical prior knowledge and spatial-temporal graph neural network in the paper "Research on gas turbine health assessment method based on physical prior knowledge and spatial-temporal graph neural network" (Applied Energy, 2024, 367: 123419). Yue Y et al., in their paper "An anomaly detection method for gas turbines based on single-condition training with zero-fault sample" (Mechanical Systems and Signal Processing, 2025, 225: 112168), explored anomaly detection methods based on single-condition training and zero-fault samples. Jiang X et al., in their paper "A thermodynamics-informed deep learning approach for lightweight modeling of gas turbine performance" (Engineering Applications of Artificial Intelligence, 2025, 142: 110122), proposed a lightweight deep learning modeling method guided by thermodynamic information. These studies demonstrate that embedding physical knowledge into data-driven models is an effective way to improve the robustness of degradation detection, but no method has yet systematically achieved degradation detection with invariant operating conditions from the perspective of causal decoupling. Therefore, there is an urgent need for a method that can embed thermodynamic mapping as a causal mechanism into the degradation detection framework and achieve complete decoupling of operating condition effects and degradation signals. Summary of the Invention
[0006] To overcome the shortcomings of the prior art, the purpose of this invention is to provide a method for detecting degradation of aero-engines under invariant operating conditions based on physical constraint causal decoupling. By learning the thermodynamic mapping relationship between the aero-engine operating condition input and the physical coupling parameters of the electronic control unit, the causal decoupling between operating condition effects and degradation signals is achieved, and a dual-channel health index with invariant operating conditions is constructed. This solves the technical problem that existing methods are unable to effectively identify degradation signals under multiple operating conditions.
[0007] To achieve the above objectives, the technical solution provided by this invention is as follows: A method for detecting condition-invariant degradation of aero-engines based on causal decoupling of physical constraints includes the following steps: Step 1: Collect data from the aircraft engine electronic control unit, including physical coupling parameters, environmental parameters, and operating condition vectors; Step 2: Extract health status data from the entire lifecycle data of the aero-engine as a training set; Step 3: Construct a multilayer perceptron network as a thermodynamic mapping network, with the operating condition vector as input and the physical coupling parameters as output. Train this thermodynamic mapping network only on health state data to learn the causal mapping relationship from operating conditions to physical coupling parameters. Step 4: Use the trained thermodynamic mapping network to predict the full life cycle data, calculate the residual between the predicted value and the actual observed value, and convert it into a physical residual score; Step 5: Calculate the anomaly score of environmental parameters relative to the baseline of health status; Step Six: Merge the physical residual score and the environmental anomaly score according to preset weights to form a dual-channel health indicator; Step 7: Perform exponential moving average smoothing on the merged health indicators to output the final condition-invariant health indicators.
[0008] In step one, the physical coupling parameters include seven parameters: turbine exhaust temperature, engine speed, fuel pump displacement, fuel pump speed, actuator current, engine volumetric flow rate, and combustion chamber pressure; the environmental parameters include three parameters: thermal control unit temperature, electronic control unit housing temperature, and fuel temperature; and the operating condition vector includes two thrust command signals.
[0009] In step three, the thermodynamic mapping network is a four-layer multilayer perceptron network with hidden layer dimensions of 64, 128, and 64 respectively, and the activation function is the modified linear unit function; the training loss function of the thermodynamic mapping network is the mean squared error loss function.
[0010] In step three, the Adam optimizer and cosine annealing learning rate scheduling strategy are used for training. The initial learning rate is set to 0.001, and the training rounds are 500. After training, the determination coefficient of the thermodynamic mapping network reaches 0.97.
[0011] Step four specifically includes: for the observed values at each moment in the full life cycle data, calculating the predicted values using a thermodynamic mapping network, and taking the absolute value of the difference between the predicted and observed values as the residual vector; calculating the z-score for each feature dimension of the residual vector, with the z-score based on the mean and standard deviation of the residuals under healthy conditions; and weighting the physical residual score by the mean and maximum value of the positive parts of the z-score. and The weighted fusion is obtained.
[0012] The calculation method for the environmental anomaly score in step five is as follows: calculate the absolute value of the difference between each environmental parameter and the mean of the healthy state, divide it by the standard deviation of the healthy state to obtain the z-score, and take the mean of the z-scores of all environmental parameters as the environmental anomaly score.
[0013] The fusion formula for the dual-channel health indicators in step six is the sum of the physical residual score multiplied by the physical channel weight and the environmental anomaly score multiplied by the environmental channel weight, where the physical channel weight... Environmental channel weight .
[0014] The smoothing coefficient of the exponential moving average in step seven The smoothed health index is equal to the smoothing coefficient multiplied by the original health index at the current moment, plus one minus the smoothing coefficient multiplied by the smoothed health index at the previous moment.
[0015] The aforementioned method for detecting invariant degradation of aero-engine operating conditions based on physical constraint causal decoupling is established based on a structural causal model. The structural causal model represents the physical coupling parameters as the sum of the thermodynamic mapping network output, degradation deviation, and measurement noise. The thermodynamic mapping network captures the deterministic causal relationship between operating conditions and parameters, the degradation deviation characterizes parameter drift caused by equipment aging, and the measurement noise characterizes random interference.
[0016] The thermodynamic mapping network has a total of no more than 18,000 parameters and supports cross-engine migration, meaning that after being trained on the health data of one engine, it can be directly applied to the degradation detection of other engines of the same model without retraining.
[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention is the first to embed thermodynamic mapping as a structural causal mechanism into the degradation detection framework, realizing the causal decoupling of operating condition effects and degradation signals, so that health indicators are not affected by thrust operating condition switching; (2) The dual-channel health indicator proposed in this invention integrates physical information residuals and environmental anomaly information, and can capture distributed degradation and local degradation features at the same time; (3) This invention only requires health status data to complete training, the model is lightweight (total number of parameters is about 18,000), and supports cross-engine migration applications. Attached Figure Description
[0018] Figure 1 This is an overall flowchart of an embodiment of the present invention.
[0019] Figure 2 This is a schematic diagram of the thermodynamic mapping network structure according to an embodiment of the present invention.
[0020] Figure 3 This is a comparison chart of the degradation detection performance of different methods in embodiments of the present invention.
[0021] Figure 4 This is a graph showing the evolution of health indicators over the life cycle according to an embodiment of the present invention. Detailed Implementation
[0022] The present invention will now be described in detail with reference to the embodiments and accompanying drawings.
[0023] Reference Figure 1 A method for detecting the condition-invariant degradation of aero-engines based on physical constraint causal decoupling includes the following steps: Step 1: Data Acquisition. Acquire data from the aircraft engine electronic control unit and record the timing. The observation vector is This vector contains three types of signals: (1) Subset of physical coupling parameters Dimension This includes turbine exhaust temperature, engine speed, fuel pump displacement, fuel pump speed, actuator current, engine volumetric flow rate, and combustion chamber pressure; (2) Subset of environmental parameters Dimension This includes the temperature of the thermal control unit, the housing temperature of the electronic control unit, and the fuel temperature; (3) Working condition vector Dimension It includes two thrust command signals; Step Two: Health Status Data Extraction. From the entire lifecycle data of the aero-engine, based on maintenance records and operation logs, determine the time interval for the health status stage, and extract the data within that time interval as the training set. The total number of health samples is denoted as [missing value]. ; Step 3: Training the thermodynamic mapping network. (Refer to...) Figure 2 Construct a four-layer multilayer perceptron network As a thermodynamic mapping network, in This is a set of network parameters.
[0024] The thermodynamic mapping network input is the operating condition vector. The output consists of predicted values for seven physical coupling parameters; the hidden layer dimensions of the thermodynamic mapping network are 64, 128, and 64 respectively, with each hidden layer followed by a modified linear unit activation function. ,Right now ; During the forward propagation of the thermodynamic mapping network, the first The layer is calculated as follows: the input is processed by the weight matrix. and bias vector The linear transformation is followed by activation function processing.
[0025] A degradation detection framework is established based on a structural causal model, in which the physical coupling parameters are represented as follows: in, The output of the thermodynamic mapping network characterizes the deterministic causal relationship between operating conditions and parameters; Degradation deviation characterizes parameter drift caused by equipment aging; For measuring noise; Train the thermodynamic mapping network only on health state data, minimizing the thermodynamic mapping training loss function: in, For the number of healthy samples, For the first The actual observed values of the physical coupling parameters for each sample. This is the predicted value input by the network for this sample operating condition.
[0026] Training employs the Adam optimizer, combined with a cosine annealing learning rate scheduling strategy, with an initial learning rate set to... The training was conducted for 500 rounds; after training, the coefficient of determination of the thermodynamic mapping network on health data was calculated. The result of 0.97 indicates that the thermodynamic mapping network can accurately fit the mapping relationship between the operating conditions and the physical coupling parameters.
[0027] Step 4: Physical Information Residual Calculation. For each time step in the full lifecycle data obtained in Step 2... The observed values are used to calculate the predicted values using the thermodynamic mapping network trained in step three, and the residual vector is obtained. Defined as: in, This indicates taking the absolute value of each element; For the residual vector of the th Calculate the z-score for each feature dimension: in, and The first The mean and standard deviation of the residuals of each feature under healthy conditions; Physical residual score Defined as: In the above calculation, the positive value of the z-score is taken. The physical residual score is obtained by multiplying the mean and maximum value of the positive value by 0.6 and 0.4 respectively and then adding them together. The mean term reflects the overall degree of deviation, and the maximum value term captures the significant anomalies of a single parameter.
[0028] Step 5: Calculate the environmental anomaly score. This involves calculating the subset of environmental parameters collected in Step 1. Calculate the global z-score anomaly score: in, and These are the mean vector and standard deviation vector of environmental parameters under healthy conditions, respectively; the environmental anomaly score reflects the overall degree to which environmental parameters deviate from the healthy baseline and is used to capture signs of degradation related to the thermal environment.
[0029] Step Six: Constructing Dual-Channel Health Indicators. This involves converting the physical residual scores obtained in Step Four into... Compared with the environmental anomaly score obtained in step five The original health indicators are then integrated according to preset weights. in, For physical channel weights, The environmental channel weights are set higher; the larger physical channel weights reflect the prior knowledge that the physical parameter residuals are the main characteristic signal of degradation.
[0030] Step 7: Exponential Moving Average Time Series Smoothing. The original health indicators obtained in Step 6 are then smoothed using an exponential moving average time series method. in, For smoothing coefficients, The smoothed health indicators This is the smoothed health index from the previous time step. A small smoothing coefficient allows the health index to filter out high-frequency noise fluctuations, retain degradation trend information, and avoid over-responding to sudden transient events. The smoothed health index is the final output condition-invariant health index.
[0031] The experimental data in this embodiment comes from the electronic control unit records of a certain type of turbojet aero-engine throughout its entire life cycle, containing a total of 123,705 data records. The data covers the complete service life of the engine from its factory condition to its scrapping, with a recording frequency of once per second. The health phase data is used to train the thermodynamic mapping network, while the degradation phase data is used to verify the detection performance.
[0032] The thermodynamic mapping network was implemented using the PyTorch framework on a workstation equipped with an NVIDIA GPU. The network structure is a four-layer multilayer perceptron with a 2D input layer, hidden layers of 64D, 128D, and 64D respectively, and a 7D output layer. Training was performed using the Adam optimizer with an initial learning rate of 0.001, cosine annealing for learning rate scheduling, and 500 training epochs. After training, the thermodynamic mapping network achieved a determination coefficient of 0.97 on healthy data.
[0033] Reference Figure 3 , Figure 3 This chart compares the performance of the method of this invention with several baseline methods on a degradation detection task. The comparison methods include isolated forest, long short-term memory network autoencoder, statistical methods, global Z-score, and moving average. Evaluation metrics include AUC-ROC, average accuracy, trend, and monotonicity. Experimental results show that the method of this invention achieves an AUC-ROC of 0.973, significantly outperforming isolated forest (0.929), long short-term memory network autoencoder (0.910), and statistical methods (0.870). Regarding the trend metric, the method of this invention achieves 0.768, far exceeding all other comparison methods, indicating that the health index constructed by this invention has a good monotonic upward trend, which is beneficial for predicting degradation trends.
[0034] Reference Figure 4 , Figure 4 The graph shows the evolution curve of the health index output by the method of this invention over the engine's life cycle. As can be observed from the graph, during the healthy phase, the smoothed health index remains at a low level and fluctuates smoothly, indicating that the thermodynamic mapping network can accurately fit the mapping relationship from operating conditions to parameters, and that operating condition switching does not cause spurious fluctuations in the health index. After entering the degradation phase, the health index shows a clear upward trend, and the trend of the curve after exponential moving average smoothing is even clearer, which is beneficial for setting reasonable alarm thresholds.
[0035] Further transfer verification experiments show that the thermodynamic mapping network trained on the health data of one engine can be directly transferred to another engine of the same model for degradation detection. The health indicators after the transfer still maintain the characteristics of being stable in the healthy stage and rising in the degradation stage, which verifies the cross-engine transfer capability of the method of the present invention.
[0036] Ablation experiments further validated the contributions of each component: removing the physical prior constraints of the thermodynamic mapping network reduced the AUC-ROC by 14 percentage points; training with only 20% of the healthy data was sufficient to achieve a performance level of 97.4% of the complete data, demonstrating the good data efficiency of the proposed method. The model has approximately 18,000 parameters, and the inference speed meets the requirements for real-time monitoring.
Claims
1. An aero-engine working condition invariant degradation detection method based on physical constraint causal decoupling, characterized in that, Includes the following steps: Step 1: Collect data from the aircraft engine electronic control unit, including physical coupling parameters, environmental parameters, and operating condition vectors; Step 2: Extract health status data from the entire lifecycle data of the aero-engine as a training set; Step 3: Construct a multilayer perceptron network as a thermodynamic mapping network, with the operating condition vector as input and the physical coupling parameters as output. Train this thermodynamic mapping network only on health state data to learn the causal mapping relationship from operating conditions to physical coupling parameters. Step 4: Use the trained thermodynamic mapping network to predict the full life cycle data, calculate the residual between the predicted value and the actual observed value, and convert it into a physical residual score; Step 5: Calculate the anomaly score of environmental parameters relative to the baseline of health status; Step Six: Merge the physical residual score and the environmental anomaly score according to preset weights to form a dual-channel health indicator; Step 7: Perform exponential moving average smoothing on the merged health indicators to output the final condition-invariant health indicators.
2. The method of claim 1, wherein, In step one, the physical coupling parameters include seven parameters: turbine exhaust temperature, engine speed, fuel pump displacement, fuel pump speed, actuator current, engine volumetric flow rate, and combustion chamber pressure; the environmental parameters include three parameters: thermal control unit temperature, electronic control unit housing temperature, and fuel temperature; and the operating condition vector includes two thrust command signals.
3. The method of claim 1, wherein, In step three, the thermodynamic mapping network is a four-layer multilayer perceptron network with hidden layer dimensions of 64, 128, and 64 respectively, and the activation function is the modified linear unit function; the training loss function of the thermodynamic mapping network is the mean squared error loss function.
4. The method of claim 3, wherein, In step three, the Adam optimizer and cosine annealing learning rate scheduling strategy are used for training. The initial learning rate is set to 0.001, and the training rounds are 500. After training, the determination coefficient of the thermodynamic mapping network reaches 0.
97.
5. The method for detecting degradation of aero-engines under unchanged operating conditions according to claim 1, characterized in that, The step four specifically comprises: for each observation value at each time in the full life cycle data, calculating a predicted value by using a thermodynamic mapping network, taking the absolute value of the difference between the predicted value and the observation value as a residual vector; calculating a z-score for each feature dimension of the residual vector, the z-score being based on the residual mean and standard deviation under the healthy state; and obtaining a physical residual score by weighting the mean and maximum of the positive part of the z-score and weighted fusion.
6. The method for detecting degradation of aero-engines under unchanged operating conditions according to claim 1, characterized in that, The calculation method for the environmental anomaly score in step five is as follows: calculate the absolute value of the difference between each environmental parameter and the mean of the healthy state, divide it by the standard deviation of the healthy state to obtain the z-score, and take the mean of the z-scores of all environmental parameters as the environmental anomaly score.
7. The method for detecting degradation of aero-engines under unchanged operating conditions according to claim 1, characterized in that, The fusion formula for the dual-channel health indicators in step six is the sum of the physical residual score multiplied by the physical channel weight and the environmental anomaly score multiplied by the environmental channel weight, where the physical channel weight... Environmental channel weight .
8. The method for detecting degradation of aero-engines under unchanged operating conditions according to claim 1, characterized in that, The smoothing coefficient of the exponential moving average in step seven The smoothed health index is equal to the smoothing coefficient multiplied by the original health index at the current moment, plus one minus the smoothing coefficient multiplied by the smoothed health index at the previous moment.
9. The method for detecting degradation of aero-engines under unchanged operating conditions according to claim 1, characterized in that, The aforementioned method for detecting invariant degradation of aero-engine operating conditions based on physical constraint causal decoupling is established based on a structural causal model. The structural causal model represents the physical coupling parameters as the sum of the thermodynamic mapping network output, degradation deviation, and measurement noise. The thermodynamic mapping network captures the deterministic causal relationship between operating conditions and parameters, the degradation deviation characterizes parameter drift caused by equipment aging, and the measurement noise characterizes random interference.
10. The method for detecting degradation of an aero-engine under unchanged operating conditions according to claim 1, characterized in that, The thermodynamic mapping network has a total of no more than 18,000 parameters and supports cross-engine migration, meaning that after being trained on the health data of one engine, it can be directly applied to the degradation detection of other engines of the same model without retraining.