Aero-engine fault prediction method based on long and short term memory network

Abstract

The invention relates to the technical field of aero-engines, and discloses an aero-engine fault prediction method based on a long and short term memory network, which comprises the following steps: extracting a plurality of groups of operation data group samples from a fast access recorder, and screening out characteristic parameters related to surge faults in each group of data samples; dividing each group of operation data group samples into a training set and a test set; an LSTM model is constructed, an input layer is used for receiving 8 feature parameters reflected by a training set under a multi-dimensional time sequence, the number of hidden units of an LSTM layer is 6, an output layer of a full connection layer is 2 nodes, EGT and PS3 are predicted, and model iteration is performed through training set data; acquiring operation data of the aero-engine in real time, and predicting EGT and PS3 in real time by using an LSTM model; when the EGT suddenly rises or the PS3 suddenly drops, an alarm is triggered; according to the method, high-precision prediction of the aero-engine fault is realized, the false alarm rate of the prediction result is low, and the method can adapt to extreme working conditions, so that the robustness of the model under the extreme working conditions is ensured, and the generalization ability of the model is improved.

Description

Technical Field

[0001] This invention relates to the field of aero-engine technology, and in particular to an aero-engine fault prediction method based on long short-term memory networks. Background Technology

[0002] Aircraft engines are crucial components for normal aircraft flight. Due to the complex and varied operating conditions and relatively harsh operating environment, engine malfunctions can pose a significant threat to flight safety and passenger safety.

[0003] Traditional methods for predicting aero-engine failures (such as threshold-based alarms and expert systems) have the following drawbacks: 1. Reliance on human experience: Fixed thresholds need to be pre-set, making it impossible to dynamically adapt to engine performance degradation. 2. Ignoring temporal characteristics: Strong temporal coupling exists between aero-engine parameters (such as the linkage between EGT and N2 speed), making it difficult for traditional methods to capture long-term dependencies. 3. Low data utilization: Multi-source parameters (such as vibration, pressure, and temperature) lack joint modeling, resulting in prominent information silos.

[0004] To address the aforementioned shortcomings, existing technologies attempt to solve them using data-driven methods such as SVM, CNN, and RNN. However, SVM has limited ability to model nonlinear relationships; CNN excels at spatial feature extraction but lacks sensitivity to temporal dependencies; and RNN struggles to handle long sequence data due to gradient vanishing or exploding problems. Consequently, the accuracy of aero-engine fault predictions using these data-driven methods remains poor, and the false alarm rate of the prediction results is high. Summary of the Invention

[0005] This invention proposes an aero-engine fault prediction method based on long short-term memory networks to overcome the shortcomings of the prior art. This prediction method can achieve high-precision prediction of aero-engine faults and has a low false alarm rate.

[0006] The technical solution of this invention is: a method for predicting aircraft engine faults based on long short-term memory networks, comprising the following steps: Several sets of operational data samples were extracted from the fast access recorder of the aircraft engine, and eight characteristic parameters strongly correlated with surge faults were screened out in each set of data samples through time series correlation analysis. The characteristic parameters include EGT and PS3.

[0007] Specifically, the fast access recorder is a flight data recording system used to record the operating parameters of aircraft engines.

[0008] EGT stands for Exhaust Gas Temperature, which is one of the key thermodynamic parameters of an aero-engine.

[0009] The full English name of PS3 is High Pressure Compressor, which means the total outlet pressure of the high-pressure compressor. PS3 is a core parameter that reflects the performance of the engine's compression system.

[0010] The eight feature parameters selected from several sets of operational data samples are divided into training and testing sets, ensuring the integrity of the time series of the operational data samples during the division process.

[0011] An LSTM model is constructed, with the input layer receiving eight feature parameters from the training set in a multi-dimensional time series. The LSTM layer has six hidden units, and the fully connected layer has two output nodes to predict EGT and PS3. An initial learning rate and a learning rate decay factor are set, and the model is iterated using the training set data. Specifically, LSTM stands for Long Short-Term Memory, which is a recurrent neural network with a gating mechanism and is good at processing time series data.

[0012] The test set data is input into the constructed LSTM model to obtain the predicted values ​​of EGT and PS3. The predicted values ​​of EGT and PS3 are evaluated, the model performance is analyzed, and the initial learning rate and learning rate decay factor are optimized.

[0013] During fault prediction, the eight characteristic parameters of the aero-engine operation data are acquired in real time, and the EGT and PS3 are predicted in real time using the LSTM model. An alarm is triggered when the actual values ​​of EGT and PS3 exceed ±3 times the RMSE of the LSTM model prediction value. Specifically, RMSE stands for Root Mean Square Error.

[0014] In at least one embodiment of the present invention, the feature parameters further include: N11, N21, FF1, T25, PT2 and VIB.

[0015] Specifically, the full English name of N11 is Low-Pressure Rotor (or Fan) Speed-stage 1, which means: low-pressure compressor speed.

[0016] The full English name of N21 is High-Pressure Rotor Speed-stage 1, which means "high-pressure compressor speed" in Chinese.

[0017] The full English name of FF1 is Fuel Flow channel 1, which means fuel emission flow rate.

[0018] The full English name of T25 is Total Temperature at the exit of the Low-PressureCompressor (Booster), which means: Low-Pressure Compressor Outlet Temperature.

[0019] The full English name of PT2 is Total Pressure at the engine inlet (Station 2), which means: low-pressure compressor inlet pressure.

[0020] VIB stands for Vibration, which means engine vibration value.

[0021] In at least one embodiment of the present invention, the plurality of sets of operational data samples include samples of low-temperature start-up and high-altitude cruise conditions.

[0022] In at least one embodiment of the present invention, after dividing the training set and test set of several sets of running data samples, Min-Max normalization is used to map the eight feature parameters in each set of running data of the training set and test set to the interval [0, 1]. When constructing the LSTM model, the input layer receives the eight feature parameters of the training set mapped to the interval [0, 1] under multi-dimensional time series.

[0023] In at least one embodiment of the present invention, in the LSTM model, the activation function is the ReLU activation function, and the parameter configuration includes: the Adam optimizer algorithm and the RMSE loss function.

[0024] In at least one embodiment of the present invention, the number of samples in the running data set is 1000; in step S2, the ratio of the training set to the test set is 7:3.

[0025] In at least one embodiment of the present invention, the number of iterations in the LSTM model is less than or equal to 800, the initial learning rate of the LSTM model is 0.01, and the learning rate reduction factor is 0.1.

[0026] In at least one embodiment of the present invention, MAE, RMSE and R are used. 2 The predicted value was evaluated, and the initial learning rate was adjusted to 0.005, and the learning rate reduction factor was adjusted to 0.3.

[0027] Specifically, MAE stands for Mean Absolute Error, which means the average absolute error between the predicted value and the actual value.

[0028] RMSE stands for Root Mean Square Error. RMSE is the square root of the average of the squared errors between the predicted and actual values.

[0029] R 2 The full English name is Coefficient of Determination, which means "determination coefficient" in Chinese; R 2 It is an indicator that measures how well the model's predicted values ​​fit the actual values.

[0030] In at least one embodiment of the present invention, inference is accelerated by TensorRT during fault prediction.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows: 0. The aero-engine fault prediction method proposed in this invention employs multi-parameter temporal coupling modeling technology. Through temporal correlation analysis, eight feature parameters strongly correlated with surge faults are selected, and a multi-source parameter joint prediction model is constructed. This joint prediction model, by using LSTM, a memory network specifically designed for processing nonlinear sequence data, can better capture long-term dependencies in the sequence data. This results in a strong nonlinear modeling capability and high sensitivity to temporal dependencies, enabling it to effectively process long sequence data. Furthermore, the selection of eight feature parameters strongly correlated with surge faults solves the accuracy problem caused by isolated parameter modeling in traditional methods. Moreover, the fault prediction method of this invention features a dynamic threshold adaptive early warning mechanism, dynamically adjusting the alarm threshold based on the LSTM model output. Unlike traditional alarm methods using fixed thresholds, this fault prediction method significantly reduces the false alarm rate.

[0032] 1. The aero-engine fault prediction method proposed in this invention has special case samples such as low temperature start-up and high-altitude cruise in its LSTM model training dataset. By optimizing the LSTM structure through parameter tuning and analyzing the temporal dependencies, the prediction method can still maintain stable prediction under extreme scenarios such as low temperature start-up and high-altitude cruise, so as to ensure the robustness of the model under extreme conditions and improve the model's generalization ability.

[0033] 3. The aero-engine fault prediction method proposed in this invention adopts a lightweight LSTM optimization strategy, proposing an optimized combination of initial learning rate of 0.005 + descent factor of 0.3, combined with TensorRT acceleration, to achieve a balance between model inference speed and accuracy, and meet the millisecond-level response requirements of aero-engines. Attached Figure Description

[0034] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0035] The accompanying drawings in this invention are not strictly drawn to scale, and the specific dimensions and quantity of each structure can be determined according to actual needs. The drawings described in this invention are merely structural schematic diagrams.

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the described embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. As used herein, the words “comprising” or “including” and similar terms mean that an element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects.

[0038] Existing solutions to the technical problem: 1. The SVM model cannot effectively handle multivariate time series data and has weak generalization ability.

[0039] 2. Univariate LSTM models only a single parameter and ignores the coupling relationship between parameters.

[0040] 3. Unoptimized LSTM structures and default parameters (such as a learning rate of 0.01) can easily lead to slow convergence and overfitting.

[0041] 4. Lack of multi-condition verification; the robustness of the model under extreme conditions such as surge and stall has not been verified.

[0042] Specifically, SVM stands for Support Vector Machine.

[0043] The full English name of CNN is Convolutional Neural Network.

[0044] RNN stands for Recurrent Neural Network.

[0045] The technical problem to be solved by this invention: 1. Multi-parameter time-series coupled modeling: realizes joint prediction of multiple parameters such as compressor outlet total pressure (PS3) and exhaust temperature (EGT).

[0046] 2. Long sequence dependency capture: Optimize the LSTM gating mechanism to improve the prediction accuracy of long-cycle events such as engine start-stop and operating condition switching.

[0047] 3. Dynamic threshold early warning: An adaptive alarm system is built based on model output to reduce the false alarm rate (target <5%).

[0048] Combination Figure 1 As shown, a method for predicting aero-engine faults based on long short-term memory networks includes the following steps: Several sets of operational data samples were extracted from the fast access recorder of the aircraft engine, and eight characteristic parameters strongly correlated with surge faults were screened out in each set of data samples through time series correlation analysis. The characteristic parameters include EGT and PS3.

[0049] The operational data samples are divided into training and testing sets, ensuring the integrity of the time series of the operational data samples during the division process. Specifically, the training dataset includes special case samples such as low-temperature start-up and high-altitude cruise, and the model input includes contextual features such as ambient temperature and air pressure.

[0050] An LSTM model is constructed, with an input layer that receives 8 feature parameters from the training set under multidimensional time series. The LSTM layer has 6 hidden units, and the fully connected layer has 2 output nodes to predict EGT and PS3. An initial learning rate and a learning rate decrease factor are set, and the model is iterated using the training set data.

[0051] The test set data is input into the constructed LSTM model to obtain the predicted values. The predicted values ​​are then evaluated to analyze the model performance.

[0052] During fault prediction, real-time aero-engine operating data is acquired, and EGT and PS3 are predicted in real time using an LSTM model. An alarm is triggered when EGT suddenly increases or PS3 suddenly decreases. The threshold for sudden increase / decrease is defined by the prediction error quantification based on the LSTM model. An alarm is triggered when the actual value exceeds the range of ±3 times RMSE (root mean square error) of the model prediction value.

[0053] Specifically, the above method sets only two parameters, EGT and PS3, as the output nodes of the LSTM model. This reverses the process, treating the physical quantity that best reveals surge as the model reconstruction target, and has the following advantages: 1. The most direct physical meaning: When surge occurs, the airflow separates and reverses instantly. The first reaction is a sharp drop in the high-pressure compressor outlet pressure PS3 and a sharp rise in the combustion chamber outlet temperature EGT. Taking EGT and PS3 as the reconstruction targets is the most direct.

[0054] 2. The error amplification effect is most pronounced: Under normal operating conditions, the timing curves of EGT and PS3 are smooth and predictable; however, once rotational stall or slight surge occurs, the signal changes instantaneously. When using an LSTM model for prediction, the prediction error RMSE in the normal segment is very small; when entering the warning segment, the predicted value "doesn't have time to react," and the error is instantly amplified by 5 to 10 times. Setting the alarm threshold at ±3 times RMSE can utilize this natural amplification to turn minute warning signs into observable alarms without the need to design complex threshold logic.

[0055] 3. Dimensional compression, lightweight model, and low latency: Only 2 quantities are reconstructed, and the output layer has 2 nodes, reducing the number of parameters by more than 75% compared to directly reconstructing 8 quantities; during inference, a forward pass only takes tens of microseconds, and real-time looping at the 10ms level can be achieved on onboard FPGAs or edge computing boxes.

[0056] 4. Training data is easier to "clean": Sensor jumps and missing data often occur in the Quick Access Recorder (QAR). The EGT and PS3 have high sampling frequency and many redundant channels, making them the easiest to perform quality inspection and repair. Using these two high-quality curves for supervision can avoid the tedious work of interpolating and aligning all 8 signals, reducing data cleaning costs.

[0057] As an alternative embodiment, the characteristic parameters include: N11 (low-pressure compressor speed), N21 (high-pressure compressor speed), FF1 (fuel discharge flow rate), T25 (low-pressure compressor outlet temperature), PT2 (low-pressure compressor inlet pressure), and VIB (engine vibration value).

[0058] As an alternative embodiment, several sets of operational data samples include samples of low-temperature start-up and high-altitude cruise conditions. By adjusting the parameters and optimizing the LSTM structure to analyze the temporal dependencies, the prediction method can maintain stable predictions under extreme scenarios such as low-temperature start-up and high-altitude cruise, thereby ensuring the robustness of the model under extreme conditions and improving the model's generalization ability.

[0059] As an alternative embodiment, after dividing the training and test sets of the running data groups into several sets, Min-Max normalization is used to map the eight feature parameters in each set of the running data in the training and test sets to the interval [0, 1]. When constructing the LSTM model, the input layer receives the eight feature parameters of the training set mapped to the interval [0, 1] under multi-dimensional time series. Normalization can eliminate the dimensional differences of different features (e.g., N1 rotation speed may be thousands, while vibration value may be a decimal), avoid the influence of a certain feature on the loss function too much, and accelerate model convergence.

[0060] As an alternative implementation, the LSTM model uses the ReLU activation function to enhance nonlinear expressiveness. The parameter configuration includes: the Adam optimizer algorithm and the RMSE loss function. The Adam optimizer addresses the problem of the fixed learning rate in traditional SGD by adaptively adjusting the learning rate. The RMSE loss function is more sensitive to larger errors and is suitable for scenarios that require accurate prediction (such as small changes in EGT that may indicate performance degradation).

[0061] As an alternative embodiment, the number of samples in the running data set is 1000; in step S2, the ratio of the training set to the test set is 7:3; this ratio balances the reliability of model learning and performance evaluation.

[0062] As an alternative implementation, the number of iterations in the LSTM model is less than or equal to 800, with a maximum of 800 iterations, which is a result of balancing training sufficiency and the risk of overfitting; the initial learning rate of the LSTM model is set to 0.01, and the learning rate decay factor is set to 0.1.

[0063] As an alternative implementation, during the model evaluation process, the predicted values ​​are evaluated using MAE, RMSE, and R2. The initial learning rate is adjusted to 0.005, and the learning rate decrease factor is adjusted to 0.3. The initial learning rate of 0.005 balances convergence speed (avoiding excessive oscillation) and convergence accuracy (avoiding excessively small convergence). The learning rate decrease factor of 0.3 avoids the problems of the learning rate decreasing too quickly (getting stuck in local optima) or too slowly (low convergence efficiency), allowing the model to finely adjust parameters even in the later stages of training.

[0064] Experimental verification results: The experimental data above shows that: real-time prediction of EGT and PS3 has an error of ≤5%, which can detect performance degradation in advance; when EGT suddenly increases or PS3 suddenly drops, an alarm is triggered with an accuracy of 99%.

[0065] As an alternative implementation, TensorRT accelerates inference during fault prediction, with a single prediction taking less than 50ms, meeting the requirements for online monitoring of aero-engines.

[0066] The aero-engine fault prediction method proposed in this invention has the following advantages: 1. Multi-parameter time-series coupled modeling technology: By screening key parameters (8 features such as PS3, EGT, N2 speed, VIB, etc.) through time-series correlation analysis, a multi-source parameter joint prediction model is constructed to solve the problem of insufficient accuracy caused by isolated parameter modeling in traditional methods.

[0067] 2. Dynamic threshold adaptive early warning mechanism: The alarm threshold is dynamically adjusted based on the model output (instead of a fixed threshold), achieving a false alarm rate of <5%, which is significantly better than traditional threshold alarm methods.

[0068] 3. Lightweight LSTM optimization strategy: An optimized combination of initial learning rate of 0.005 and descent factor of 0.3 is proposed, which, combined with TensorRT acceleration, achieves a balance between model inference speed and accuracy, meeting the millisecond-level response requirements of aero-engines.

[0069] 4. Robust design under extreme conditions: Verify model performance under extreme conditions such as surge and stall, and improve model generalization ability through data augmentation (such as simulating sensor noise).

[0070] Furthermore, this LSTM model can maintain stable predictions even under extreme scenarios such as low-temperature startup and high-altitude cruise. The specific reasons for this are as follows: 1. Data-driven adaptive characteristics The root cause is that the LSTM model is trained with a large amount of historical QAR data containing extreme operating conditions, and has learned the nonlinear mapping relationship between different environmental parameters and engine state.

[0071] Specifically, the training dataset includes special case samples such as -40℃ low-temperature start-up and 12km high-altitude cruise; the model input includes contextual features such as ambient temperature and air pressure (parameters such as PT2 and T25); and temporal dependencies are analyzed by optimizing the LSTM structure through parameter tuning.

[0072] 2. The anti-interference capability of the gating mechanism Key technology: The synergistic effect of the forget gate and the input gate can dynamically adjust the memory content.

[0073] Extreme operating condition response mechanism: When starting at low temperature, the forget gate automatically reduces the impact of abnormal transient fluctuations caused by low temperature, and the input gate strengthens key parameters such as fuel flow (FF1); when cruising at high altitude, the output gate adjusts the weight ratio of high pressure rotor speed (N2) and vibration value (VIB) to suppress air pressure fluctuation interference.

[0074] 3. Multi-scale feature extraction design Network structure advantages: Deep LSTM networks (6-layer structure) can simultaneously capture: second-level fluctuations (such as the instantaneous characteristics of compressor surge); minute-level trends (such as the preheating process during low-temperature startup).

[0075] The ability to express non-stationary signals is enhanced by using a ReLU activation layer, and the Adam optimizer is used to automatically adjust the parameter update amplitude.

[0076] The above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions implemented in the present invention, and should all be covered within the protection scope of the present invention.

Claims

1. A method for predicting aero-engine faults based on long short-term memory networks, characterized in that, Includes the following steps: Several sets of operational data samples were extracted from the fast access recorder of the aircraft engine, and eight characteristic parameters strongly correlated with surge faults were screened out in each set of data samples through time series correlation analysis. The characteristic parameters include EGT and PS3. The eight selected feature parameters are divided into training and test sets, ensuring the integrity of the time series during the division process; An LSTM model is constructed, with the input layer receiving eight feature parameters from the training set under multidimensional time series, and the output layer of the fully connected layer having two nodes to predict EGT and PS3. The initial learning rate and learning rate reduction factor of the LSTM model are set, and the model is iterated using the training set data. The test set data is input into the constructed LSTM model to obtain the predicted values ​​of EGT and PS3. The predicted values ​​of EGT and PS3 are evaluated, the model performance is analyzed, and the initial learning rate and learning rate decay factor are optimized. During fault prediction, the eight characteristic parameters of the aero-engine operation data are acquired in real time, and the EGT and PS3 are predicted in real time using the LSTM model; when the actual values ​​of EGT and PS3 exceed ±3 times the RMSE of the LSTM model prediction value, an alarm is triggered.

2. The aero-engine fault prediction method based on long short-term memory networks as described in claim 1, characterized in that, The characteristic parameters also include: N11, N21, FF1, T25, PT2, and VIB.

3. The aero-engine fault prediction method based on long short-term memory networks as described in claim 1, characterized in that, The aforementioned sets of operational data samples include samples of low-temperature start-up and high-altitude cruise conditions.

4. The aero-engine fault prediction method based on long short-term memory networks as described in claim 1, characterized in that, After dividing the training and test sets of the running data groups into several sets, Min-Max normalization is used to map the eight feature parameters in each set of the running data to the interval [0, 1]. When constructing the LSTM model, the input layer receives the eight feature parameters of the training set mapped to the interval [0, 1] under multi-dimensional time series.

5. The aero-engine fault prediction method based on long short-term memory networks as described in claim 1, characterized in that, In the LSTM model, the ReLU activation function is used, and the parameter configuration includes the Adam optimizer algorithm and the RMSE loss function.

6. The aero-engine fault prediction method based on long short-term memory networks as described in claim 1, characterized in that, The number of samples in the running data set is 1000; the ratio of the training set to the test set is 7:

3.

7. The aero-engine fault prediction method based on long short-term memory networks as described in claim 4, characterized in that, The number of iterations in the LSTM model is less than or equal to 800. The initial learning rate of the LSTM model is set to 0.01, and the learning rate decay factor is 0.

1.

8. The method for predicting aero-engine faults based on long short-term memory networks as described in claim 1, characterized in that, After the LSTM model is built, MAE, RMSE, and R are used to measure its performance. 2 The predicted value was evaluated, and the initial learning rate was adjusted to 0.005, and the learning rate reduction factor was adjusted to 0.

3.

9. The method for predicting aero-engine faults based on long short-term memory networks as described in claim 1, characterized in that, In fault prediction, inference is accelerated using TensorRT.

Images