Aero-engine wing washing time prediction method, device, medium and product

Abstract

The application discloses an aero-engine in-wing water washing time prediction method, equipment, medium and product. The method comprises the following steps: acquiring a flight data time sequence set corresponding to an aero-engine; after filling missing flight parameters, abnormal parameter value clipping and standardization correction are performed on each initial flight parameter set in the flight data time sequence set, at least one stable flight parameter value set is extracted; a training sample set is constructed according to a plurality of stable flight parameter value sets corresponding to a plurality of aero-engines respectively, and a pre-constructed three-level fusion network is trained using the training sample set to obtain a margin prediction model; a prediction input sequence of an in-wing aero-engine is input into the margin prediction model to obtain an exhaust temperature margin prediction value of the in-wing aero-engine, and the in-wing water washing time of the in-wing aero-engine is determined according to the difference value between the exhaust temperature margin prediction value and a margin standard value. The in-wing water washing time of the aero-engine can be predicted with high precision.

Description

Technical Field

[0001] This invention relates to the field of aero-engine health management and deep learning prediction technology, and in particular to a method, device, medium and product for predicting the timing of wing washing in aero-engines. Background Technology

[0002] Aero engines are the core power units of modern civil and military aircraft, and their performance directly affects flight safety, operational efficiency, and economic benefits. With the continued growth of the global air transport industry, the need for engine health management technology is becoming increasingly urgent. Unplanned engine downtime not only directly threatens flight safety but also leads to economic losses such as flight delays, cancellations, and high maintenance costs.

[0003] Condition-Based Maintenance (CBM) effectively reduces maintenance costs and improves fleet availability by continuously monitoring key engine performance parameters and developing maintenance plans based on their actual degradation status. Gas path parameter trend prediction, a core component of the CBM technology system, aims to build performance degradation evolution models based on historical monitoring data to predict the future trajectory of key parameters. Among these, key gas path parameters are the most sensitive comprehensive indicators reflecting the performance degradation status of aero-engines.

[0004] However, predicting the timing of wing washing for aero-engines faces the following technical challenges: First, the engine houses hundreds of sensors with strong nonlinear and coupled relationships between their parameters, significantly increasing the complexity of feature extraction and modeling. Second, the engine's operating environment is complex, with time-varying external environmental factors such as altitude, temperature, and humidity introducing significant measurement noise and environmental interference. Third, engine performance degradation typically exhibits a slow, gradual, long-term time-series process, with weak trend signals easily masked by short-term fluctuations. Fourth, traditional prediction methods such as multiple linear regression and support vector regression struggle to simultaneously model time-series dependencies, multivariate interactions, and complex nonlinear mapping relationships. Summary of the Invention

[0005] This invention provides a method, device, medium, and product for predicting the timing of wing washing of aero engines, so as to achieve high-precision prediction of the timing of wing washing of aero engines.

[0006] According to one aspect of the present invention, a method for predicting the timing of wing washing in an aero-engine is provided, comprising: Acquire the flight data time series set corresponding to the aero-engine, wherein the flight data time series set includes multiple initial flight parameter value sets sorted by acquisition time; The missing flight parameter values ​​in each initial flight parameter set are filled to obtain each complete flight parameter value set. The complete flight parameter value set includes the exhaust temperature value that matches the set acquisition time and multiple gas path parameter values ​​that correspond to the exhaust temperature value. Abnormal parameter values ​​are pruned from each complete flight parameter value set after the filling process is completed, and the parameter values ​​of different types of flight parameters in each complete flight parameter value set after pruning are smoothed respectively. Based on the temperature correction factor and pressure correction factor calculated from the first type of flight parameter values ​​in each complete flight parameter value set, after standardizing and correcting the second type of flight parameter values ​​in each complete flight parameter value set, at least one stable flight parameter value set is extracted from each complete flight parameter value set. Based on multiple stable flight parameter value sets corresponding to multiple aero engines, a training sample set is constructed, and the pre-constructed three-level fusion network is trained using the training sample set to obtain a margin prediction model. The three-level fusion network includes a bidirectional long short-term memory temporal network for extracting multi-parameter bidirectional local temporal features, a multi-head self-attention global context encoding network for extracting global context enhancement features, and a sparrow search optimized gradient boosting tree prediction network for obtaining exhaust temperature margin prediction values. The predicted input sequence of the wing-mounted aero-engine is input into the margin prediction model to obtain the predicted value of the exhaust temperature margin of the wing-mounted aero-engine. The predicted input sequence includes multiple target gas path parameter value groups sorted according to the acquisition time, and each target gas path parameter value group includes multiple gas path parameter values. When the difference between the predicted exhaust temperature margin and the standard margin value exceeds the difference threshold, it is determined that the on-wing aero-engine needs to be washed with water.

[0007] According to another aspect of the present invention, a device for predicting the timing of wing washing in an aircraft engine is provided, the device comprising: The flight data time series set acquisition module is used to acquire the flight data time series set corresponding to the aero-engine. The flight data time series set includes multiple initial flight parameter value sets sorted by acquisition time. The filling processing module is used to fill the missing flight parameter values ​​in each initial flight parameter set to obtain each complete flight parameter value set. The complete flight parameter value set includes an exhaust temperature value that matches the set acquisition time and multiple gas path parameter values ​​that correspond to the exhaust temperature value. The trimming and smoothing module is used to trim abnormal parameter values ​​from each complete flight parameter value set after the filling process, and to smooth the parameter values ​​of different types of flight parameters in each complete flight parameter value set after trimming. The stable flight parameter value set extraction module is used to standardize and correct the second type of flight parameter values ​​in each complete flight parameter value set based on the temperature correction factor and pressure correction factor calculated from the first type of flight parameter values ​​in each complete flight parameter value set, and then extract at least one stable flight parameter value set from each complete flight parameter value set. The margin prediction model training module is used to construct a training sample set based on multiple stable flight parameter value sets corresponding to multiple aero engines, and to train a pre-constructed three-level fusion network using the training sample set to obtain the margin prediction model. The three-level fusion network includes a bidirectional long short-term memory temporal network for extracting multi-parameter bidirectional local temporal features, a multi-head self-attention global context encoding network for extracting global context enhancement features, and a sparrow search optimized gradient boosting tree prediction network for obtaining exhaust temperature margin prediction values. The exhaust temperature margin prediction value determination module is used to input the prediction input sequence of the wing-mounted aero-engine into the margin prediction model to obtain the exhaust temperature margin prediction value of the wing-mounted aero-engine. The prediction input sequence includes multiple target gas path parameter value groups sorted according to the acquisition time, and each target gas path parameter value group includes multiple gas path parameter values. The on-wing washing timing determination module is used to determine that on-wing washing is required for the on-wing aero-engine when the difference between the predicted value of the exhaust temperature margin and the standard value of the margin is greater than the difference threshold.

[0008] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: At least one processor; and a memory communicatively connected to said at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the method for predicting the timing of wing washing of an aircraft engine according to any embodiment of the present invention.

[0009] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the method for predicting the timing of wing washing of an aero-engine as described in any embodiment of the present invention.

[0010] According to another aspect of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of the method for predicting the timing of wing washing of an aero-engine as described in any embodiment of the present invention.

[0011] The technical solution of this invention improves the data quality of model training samples by performing operations such as filling missing flight parameters, pruning and smoothing abnormal parameter values, standardizing correction, and extracting stable flight parameter value sets on the time series set of aero-engine flight data. The margin prediction model trained using a three-level fusion network composed of a bidirectional long short-term memory time series network, a multi-head self-attention global context encoding network, and a sparrow search optimized gradient boosting tree prediction network can accurately predict exhaust temperature margin. Finally, by performing a Boolean comparison between the predicted exhaust temperature margin value and the standard value, the on-wing water washing timing of the aero-engine is directly determined, facilitating engineering applications.

[0012] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart of a method for predicting the timing of wing washing of an aero-engine according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the feature processing branch in a bidirectional long short-term memory temporal network applicable to Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the structure of a multi-head self-attention global context coding network provided in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the structure of the Sparrow Search Optimization Gradient Boosting Tree Prediction Network provided in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram illustrating the principle of the method for predicting the timing of wing washing of an aero-engine according to an embodiment of the present invention.

[0015] Figure 6 This is a schematic diagram of the structure of a predictive device for the timing of wing washing of an aircraft engine according to Embodiment 2 of the present invention; Figure 7 This is a schematic diagram of the structure of an electronic device that implements the method for predicting the timing of wing washing of an aero-engine according to embodiments of the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] Example 1 Figure 1 This is a flowchart of a method for predicting the timing of wing-mounted airflow washing of an aero-engine according to Embodiment 1 of the present invention. This embodiment is applicable to situations where the timing of wing-mounted airflow washing of an aero-engine is accurately predicted based on multiple target airflow parameter values ​​continuously collected over a recent period. This method can be executed by a device for predicting the timing of wing-mounted airflow washing of an aero-engine. This prediction device can be implemented in hardware and / or software and is generally configured in an electronic device with data processing capabilities. Figure 1 As shown, the method includes: S110. Obtain the flight data time series set corresponding to the aero-engine.

[0019] The flight data time series set includes multiple initial flight parameter value sets sorted by acquisition time.

[0020] In this embodiment, the flight data time series set corresponding to the aero-engine can be formed by collecting the full flight data (referred to as QAR flight data) recorded in the airborne QAR (Quick Access Recorder).

[0021] At each data acquisition time, the aircraft records a set of initial flight parameter values, forming an initial flight parameter value set. This initial flight parameter value set specifically includes: timestamp information, EGT (Exhaust Gas Temperature), and a set of gas path parameter values ​​associated with the EGT. The gas path parameter value set includes the low-pressure rotor speed (N1), high-pressure rotor speed (N2), fuel flow rate (FF), lubricating oil inlet temperature (OIT), lubricating oil pressure (OIP), total atmospheric temperature (TAT), high-pressure compressor outlet temperature (T3), low-pressure turbine outlet temperature (T5), fan outlet total temperature (T25), and full-temperature probe. The following values ​​are listed: Head Total Pressure (PT2), Fan Outlet Total Pressure (P25), High Pressure Compressor Outlet Total Pressure (P3), Bleed Air Value (Bleeds), Variable Bleed Valve Value (VBV), Variable Stator Blade Value (VSV), High Pressure Turbine Clearance Control Value (HPT), Low Pressure Turbine Clearance Control Value (LPT), Mach Number (Ma), Static Atmosphere Temperature (SAT), Corrected Airspeed (CAS), and Flight Altitude (ALT).

[0022] In this embodiment, the initial flight parameter value set acquired at a single acquisition time should theoretically include the aforementioned EGT and the corresponding gas path parameter value set. By sorting multiple initial flight parameter sets according to their acquisition time, the flight data time series set can be obtained.

[0023] S120. Fill in the missing flight parameter values ​​in each initial flight parameter set to obtain each complete flight parameter value set.

[0024] The complete flight parameter value set includes the exhaust temperature value that matches the set acquisition time, as well as multiple gas path parameter values ​​that correspond to the exhaust temperature value.

[0025] The inventors discovered through research on QAR flight data of a specific type of turbofan engine that some flight parameter values ​​were missing in the initial flight parameter sets at certain acquisition times. Therefore, before constructing training samples to train the margin prediction model, it is necessary to first fill in the missing flight parameter values ​​in each initial flight parameter set.

[0026] In an optional implementation of this embodiment, filling in the missing flight parameter values ​​in each initial flight parameter set to obtain each complete flight parameter value set may include: S1201. Perform cluster analysis on the parameter values ​​of each flight parameter in each initial flight parameter value set, calculate the missing rate of each flight parameter in the entire initial flight parameter value set, and identify missing flight parameters whose missing rate is greater than or equal to the preset missing rate threshold.

[0027] In this embodiment, processing is performed based on the time series set of QAR flight data of a specific type of turbofan engine used in the aforementioned study. For example, the K-means clustering algorithm can be used to group the parameter values ​​of each flight parameter (e.g., low-pressure rotor speed, high-pressure rotor speed, or total atmospheric temperature) in the entire initial flight parameter value set, count the number of parameter values ​​included in each flight parameter, divide the calculated number of parameter values ​​by the total number of initial flight parameter value sets, and obtain the effectiveness rate of each flight parameter. Then, based on the effectiveness rates of each aircraft parameter, the missing rate (1 - effectiveness rate) of each flight parameter in the entire initial flight parameter value set can be calculated.

[0028] In one optional implementation of this embodiment, a missing rate threshold of 5% can be preset. Statistically, the missing rates for the two flight parameters, total fan outlet pressure and low-pressure turbine outlet temperature, reached 10.4% and 6.9%, respectively, both exceeding the 5% threshold. Therefore, both total fan outlet pressure and low-pressure turbine outlet temperature can be considered as missing flight parameters, while the missing rates for other flight parameters are all below this threshold and do not require filling.

[0029] S1202. Based on the thermodynamic coupling relationship between various flight parameters, select multiple key coupling parameters that match the missing flight parameters from among the various flight parameters.

[0030] In this embodiment, the flight parameters inside the aero-engine follow the thermodynamic cycle law of the gas turbine, and there are fixed aerodynamic and thermodynamic coupling relationships between different flight parameters. This relationship serves as the basis for selecting key coupling parameters. That is, multiple flight parameters with high thermodynamic coupling relationships with the missing flight parameter can be obtained as multiple key coupling parameters that match the missing flight parameter.

[0031] In a specific example, for the missing flight parameter of total fan outlet pressure, the key coupling parameters can be selected as follows: low-pressure rotor speed, high-pressure rotor speed, exhaust temperature, fuel flow rate, total atmospheric temperature, static atmospheric temperature, flight altitude, and Mach number. For the missing flight parameter of low-pressure turbine outlet temperature, the key coupling parameters can be selected as follows: high-pressure rotor speed, exhaust temperature, total high-pressure compressor outlet pressure, high-pressure compressor outlet temperature, fuel flow rate, variable bleed valve, and variable stator blades. The key coupling parameters selected above are strongly correlated with the corresponding missing flight parameters, and their mapping relationship can be fitted by a regression model, providing reliable input features for subsequent missing value filling.

[0032] S1203. After performing Z-score standardization on the parameter values ​​of each flight parameter in the entire initial flight parameter value set, extract the complete target initial flight parameter value set.

[0033] In this embodiment, considering that the dimensions and numerical ranges of different flight parameters vary greatly, directly using them for regression modeling would cause an imbalance in model weights. Therefore, Z-score standardization can be used first to eliminate dimensional interference.

[0034] Specifically, the standardized formula for Z-scores is as follows: ,in, Let j be the mean of the j-th flight parameter calculated over the entire set of initial flight parameter values. Let be the standard deviation of the j-th flight parameter calculated over the entire set of initial flight parameter values. Let j be the original parameter value of the flight parameter in the i-th initial flight parameter value set. The standard parameter value is obtained by performing Z-score standardization on the j-th flight parameter in the i-th initial flight parameter value set.

[0035] After completing the standardization of all flight parameters, each initial flight parameter value set is traversed to perform a flight parameter integrity check. Only those initial flight parameter value sets with no missing flight parameter values ​​are retained as the target initial flight parameter value set.

[0036] S1204. In the initial flight parameter value set of each target, extract the parameter values ​​of each key coupling parameter as the input features of the preset regression model, and extract the parameter values ​​of each missing flight parameter as training labels, and solve for the optimal coefficients used to calculate the missing flight parameters based on each key coupling parameter.

[0037] In this embodiment, the preset regression model can be the Ridge Regression model, which introduces an L2 regularization term to effectively solve the fitting distortion problem caused by multicollinearity among multiple thermocoupled parameters.

[0038] Specifically, when solving for the optimal coefficient of a missing flight parameter A, a method of the form: The closed-form solution.

[0039] in, The input feature matrix of the regression model consists of the parameter values ​​of each key coupling parameter in the initial flight parameter value set of each target. I is a d*d identity matrix, and d is the total number of key coupling parameters that match the missing flight parameter A. These are the preset regularization parameters. , The training label vector for the model consists of the parameter values ​​of each missing flight parameter from the initial flight parameter value set for each target. By substituting the specific values ​​for each of these terms, the optimal coefficient matching the missing flight parameter A can be obtained. .

[0040] In other words, this embodiment can construct matching closed-form solutions based on the number of missing flight parameters, and finally obtain the optimal coefficients corresponding to each missing flight parameter.

[0041] S1205. Based on the solved optimal coefficients, fill in the missing flight parameter values ​​in each initial flight parameter set that has undergone Z-score standardization to obtain the filled parameter values ​​that match the missing flight parameters.

[0042] In this embodiment, all initial flight parameter value sets with missing parameter values ​​are traversed, and all key coupling parameters after standardization are extracted from these initial flight parameter value sets. These parameters are then substituted into the corresponding ridge regression prediction formula to obtain the filling parameter values ​​that match the missing flight parameters.

[0043] In a specific example, when filling in the missing flight parameter A, the initial flight parameter value set 1, which contains missing values ​​for flight parameter A, is first obtained. Then, all key coupling parameters corresponding to the missing flight parameter A are retrieved from the initial flight parameter value set 1, forming... Then, based on the optimal coefficient corresponding to the missing flight parameter A According to the formula: The filler parameter values ​​corresponding to the missing flight parameter A in the initial flight parameter value set 1 are calculated. .

[0044] Based on the above embodiments, after calculating the predicted values ​​of each missing flight parameter under the standardized scale, the mean and standard deviation of each flight parameter stored in the standardization stage are used for inverse transformation to restore the standardized predicted values ​​to the original values ​​that conform to the physical dimensions, thus obtaining the filled parameter values ​​without boundary constraints.

[0045] S1206. Based on the maximum and minimum parameter values ​​defined by the physical characteristics of the missing flight parameters, each filler parameter value is pruned to obtain a complete set of flight parameter values.

[0046] In this embodiment, by combining the aerodynamic and temperature operating boundaries specified at the factory for a specific type of aero-engine, upper and lower limits of physically reasonable values ​​can be set for each missing flight parameter, serving as the maximum and minimum parameter values. Then, a preset trimming strategy can be used to constrain and correct all filled parameter values. For example, if a filled parameter value is lower than the physical lower limit, it is replaced with a preset minimum value; if a filled parameter value is higher than the physical upper limit, it is replaced with a preset maximum value.

[0047] The above settings avoid the problem of the filling result violating the physical laws of aero-engine operation. By backfilling the filling parameter values ​​after boundary clipping to the corresponding initial flight parameter value set, all flight parameters are free of gaps and invalid values, ultimately generating a complete set of flight parameter values ​​for subsequent outlier handling and smoothing / denoising operations.

[0048] S130. Perform abnormal parameter value pruning on each complete flight parameter value set after the filling process is completed, and perform smoothing on the parameter values ​​of different types of flight parameters in each complete flight parameter value set after pruning.

[0049] In an optional implementation of this embodiment, performing abnormal parameter value pruning on each complete flight parameter value set after the filling process can include: S1301. Based on the complete set of flight parameter values, calculate the mean value of each flight parameter. and standard deviation Based on a preset standard deviation criterion, the upper critical value of each flight parameter is calculated. and lower critical value .

[0050] S1302. Compare the parameter values ​​of each flight parameter in the complete flight parameter value set with the matching upper and lower critical values.

[0051] S1303. If the first target parameter value in the complete set of first target flight parameter values ​​is greater than the target upper critical value of the first target flight parameter, then the first target parameter value is corrected to the target upper critical value.

[0052] S1304. If the second target parameter value in the complete flight parameter value set of the second target is less than the target lower critical value of the second target flight parameter, then the second target parameter value is corrected to the target lower critical value.

[0053] In existing technologies, outlier detection is required before using training samples for model training, and the detected outliers are directly filtered out. However, since sample data for aero-engines is relatively scarce, this embodiment proposes a method for pruning outlier parameter values ​​by using correction instead of direct filtering.

[0054] Specifically, the complete flight parameter value sets can be clustered according to each flight parameter to obtain the parameter value set corresponding to each flight parameter. Then, the mean value corresponding to each flight parameter can be calculated. and standard deviation Then, based on the preset The criteria determine the upper critical value for each flight parameter. and lower critical value Then, outliers of each flight parameter that exceed the upper or lower threshold are cropped to the matching upper or lower threshold.

[0055] In another optional implementation of this embodiment, the parameter values ​​of different types of flight parameters in each complete flight parameter value set after clipping are smoothed, which may specifically include: For temperature-related flight parameters, a Savitzky-Golay filter is used for smoothing; for pressure-related parameters, a Gaussian smoothing algorithm is used; for fuel flow and system control-related parameters, a locally weighted scatter smoothing algorithm is used; and for engine speed-related parameters, a moving average algorithm is used.

[0056] Specifically, for temperature-related flight parameters, such as T3, T5, T25, SAT, and EGT, a Savitzky-Golay filter can be used for smoothing.

[0057] In a specific example, after obtaining the parameter values ​​corresponding to the exhaust temperature flight parameter based on the complete flight parameter value set, the above parameter values ​​can be sorted according to the acquisition time to obtain the time series of parameter values ​​corresponding to the exhaust temperature.

[0058] Therefore, according to the formula: This yields smoothed results for each exhaust temperature value. The Savitzky-Golay filter has a time window length of 2m+1, where m is the window radius. The exhaust temperature value is the value collected at time t. The convolution coefficients at the i-th point within the time window are determined by the time window length of the Savitzky-Golay filter and the preset polynomial order (e.g., 3). To The result is a smooth one.

[0059] In this optional embodiment, the inventors considered that during the takeoff and climb phases of an aircraft, the values ​​of temperature-related flight parameters will have significant peaks. Ordinary smoothing (such as moving average) will "smooth out" these peaks, resulting in the loss of key operational information. The polynomial fitting characteristics of the Savitzky-Golay filter can effectively preserve the peak values, slopes, and other features of the signal while smoothing out noise, making it very suitable for temperature-related flight parameters that change rapidly and have characteristic peaks.

[0060] Specifically, for pressure parameters such as PT2, P25, and P3, Gaussian smoothing algorithm can be used for smoothing.

[0061] In a specific example, after obtaining the parameter values ​​corresponding to the total pressure at the fan outlet based on the complete set of flight parameter values, the parameter values ​​can be sorted according to the acquisition time to obtain the time series of parameter values ​​corresponding to the total pressure at the fan outlet.

[0062] Therefore, according to the formula: This yields the smoothed result corresponding to the total pressure value at each fan outlet. The preset Gaussian smoothing time window length is 2m+1, where m is the window radius. This represents the total pressure at the fan outlet at time t. This represents the weight value of the k-th point within the time window. ,in, The preset Gaussian kernel standard deviation parameter, To The result is a smooth one.

[0063] In this optional embodiment, the inventors considered that the parameter values ​​of pressure-related flight parameters change gradually, and the noise is mostly high-frequency random noise. Gaussian smoothing is a gentle low-pass filter that can effectively suppress noise without introducing "lag" like moving average, making it suitable for smoothly changing pressure-related flight parameters.

[0064] Specifically, for the parameter values ​​of fuel flow and system control parameters, such as FF, HPT, and LPT, a locally weighted scatter smoothing algorithm (referred to as LOWESS) can be used for smoothing.

[0065] In a specific example, after obtaining the parameter values ​​corresponding to the flight parameter of fuel flow based on the complete set of flight parameter values, the above parameter values ​​can be sorted according to the acquisition time to obtain the time series of parameter values ​​corresponding to fuel flow.

[0066] According to the formula: The smoothing results for each fuel flow rate value can be calculated. The time window length for locally weighted scatter smoothing is 2m+1, where m is the window radius. The data is the fuel flow rate at time t. and Preset the intercept and slope of the linear fitting model within the time window. This represents the local weight value at the i-th point within the time window. , , To The result is a smooth one.

[0067] In this optional embodiment, the inventors considered that fuel flow and system control parameters are affected by throttle commands and compressor surge boundary control, and their changes are non-linear. Linear smoothing processes such as moving average or Gaussian smoothing will treat the real changes brought about by control commands as noise and smooth them out, while the LOWESS algorithm can better preserve the non-linear trend and smooth the noise.

[0068] Specifically, for parameters such as speed-related parameters, such as N1 and N2, a moving average algorithm can be used for smoothing.

[0069] In a specific example, after obtaining the parameter values ​​corresponding to the low-pressure rotor speed based on the complete set of flight parameter values, the parameter values ​​can be sorted according to the acquisition time to obtain the time series of parameter values ​​corresponding to the low-pressure rotor speed.

[0070] Therefore, according to the formula: The smoothed results of each low-pressure rotor speed value were calculated. The moving average time window size is 2k+1, where k is the window radius. The low-pressure rotor speed value is collected at time t. To The result is a smooth one.

[0071] In this optional embodiment, the inventors considered that N1 and N2 are among the most stable parameters of an aero-engine, changing slowly and with low noise. The moving average algorithm is simple to calculate and highly efficient, sufficient to meet the noise reduction requirements of such stable signals, without introducing additional computational complexity.

[0072] S140. Based on the temperature correction factor and pressure correction factor calculated from the first type of flight parameter values ​​in each complete flight parameter value set, after standardizing and correcting the second type of flight parameter values ​​in each complete flight parameter value set, at least one stable flight parameter value set is extracted from each complete flight parameter value set.

[0073] In an optional implementation of this embodiment, the first type of flight parameter values ​​can be the total atmospheric temperature, Mach number, and flight altitude, while the second type of flight parameter values ​​can be the low-pressure rotor speed, high-pressure rotor speed, exhaust temperature, and fuel flow rate.

[0074] Accordingly, based on the temperature correction factor and pressure correction factor calculated from the first type of flight parameter values ​​in each complete flight parameter value set, the second type of flight parameter values ​​in each complete flight parameter value set are standardized and corrected, which may include: S1401. In each complete flight parameter value set, obtain the current complete flight parameter value set in sequence.

[0075] S1402. From the current complete set of flight parameter values, extract the total atmospheric temperature (TAT), Mach number (Ma), and flight altitude (ALT) as the first type of flight parameter values.

[0076] S1403, According to the formula: The temperature correction factor was calculated. And according to the formula: The pressure correction factor was calculated. .

[0077] S1404. Extract the original low-pressure rotor speed value from the current complete flight parameter value set. High-voltage rotor speed value Exhaust temperature value and fuel flow rate As a second type of flight parameter value.

[0078] S1405, According to the formula: Calculation yields the pair Standardization correction results .

[0079] S1406, According to the formula: Calculation yields the pair Standardization correction results .

[0080] S1407, According to the formula: Calculation yields the pair Standardization correction results .

[0081] S1408, According to the formula: Calculation yields the pair Standardization correction results .

[0082] S1409. Check whether the processing of all complete flight parameter value sets is complete. If yes, end the standardization correction process; otherwise, return to execute S1401.

[0083] In this optional embodiment, the inventors creatively consider using total atmospheric temperature, Mach number, and flight altitude as standardized parameters to construct correction factors. This eliminates environmental factors from various flight parameters, achieving the technical effect of converting all measured parameters to ISA (International Standard Atmosphere) standard conditions. Compared with existing standardized corrections, the technical solution of this invention creatively introduces Mach number and flight altitude to construct correction factors, which can effectively eliminate the additional influence of flight altitude and flight speed on various flight parameters.

[0084] Specifically, Table 1 shows the numerical comparison between existing standardization correction methods and the improved standardization correction method proposed in the embodiments of the present invention.

[0085] Table 1 As shown in Table 1, the improved standardization correction method, by introducing Mach number and flight altitude values ​​as effect correction terms, effectively controls data dispersion while correcting for environmental influences, making it more suitable for performance degradation trend analysis in practical engineering scenarios. The technical solution of this invention solves the problem that traditional correction methods cannot handle high Mach number ramjet effects, which is the fundamental reason why the variance of the improved model shown in the table is significantly reduced after correction.

[0086] In another optional implementation of this embodiment, at least one stable flight parameter value set is extracted from each complete flight parameter value set, which may specifically include: S1401' According to the preset sliding step size, the preset total number of parameter sets W, and the acquisition time of each complete flight parameter value set, the complete flight parameter value set is divided into multiple sliding time windows, wherein each sliding time window includes W complete flight parameter value sets.

[0087] Here, the total number of parameters W can be understood as the size of the sliding time window. For example, W can be 40, and the sliding step size can be the step size of moving from one sliding time window to another, for example, the sliding step size can be 3.

[0088] In this embodiment, after sorting each complete flight parameter value set on the entire time axis according to the acquisition time, starting from the complete flight parameter value set at the farthest acquisition time, firstly, take W consecutive complete flight parameter value sets and divide them into a sliding time window. After a sliding step of W complete flight parameter value sets, obtain a new complete flight parameter value set as the starting point, and then take W consecutive complete flight parameter value sets and divide them into a new sliding time window. This process is repeated until each complete flight parameter value set is divided into multiple sliding time windows.

[0089] S1402', According to the formula: The range score value of the j-th flight parameter in the i-th sliding time window is calculated. ; in, Let be the range of values ​​for the j-th flight parameter within the i-th sliding time window. , Let j be the parameter value of the j-th flight parameter in the k-th complete flight parameter value set in the i-th sliding time window, where max() is the function to take the maximum value and min() is the function to take the minimum value. Let $\frac{j}{i}$ be the standard deviation of the $j$-th flight parameter within the $i$-th sliding time window. , Let be the mean of the j-th flight parameter within the i-th sliding time window. , The preset stability tolerance threshold is set for the j-th flight parameter.

[0090] In this optional embodiment, The smaller the value, the less the flight parameter fluctuates within its sliding time window, and therefore, the higher the range score of the flight parameter.

[0091] S1403', According to the formula: The standard deviation score of the j-th flight parameter in the i-th sliding time window is calculated. .

[0092] In this optional embodiment, The smaller the value, the more concentrated the values ​​of the flight parameter are within the corresponding sliding time window; therefore, the higher the standard deviation score of the flight parameter.

[0093] S1404', According to the formula: The timing score value of the j-th flight parameter within the i-th sliding time window is calculated. ; in, Let j be the mean absolute difference of the flight parameter in the i-th sliding time window. , The parameter value is the j-th flight parameter in the (k+1)-th complete flight parameter value set within the i-th sliding time window.

[0094] In this optional embodiment, The smaller the value, the slower the flight parameter changes within its sliding time window, and therefore, the higher the timing score of the flight parameter.

[0095] S1405', According to the formula: The overall score for the i-th sliding time window is calculated. ; in, , as well as Here, is the preset weighting coefficient, and m is the total number of flight parameters included in the complete flight parameter value set. The preset weighting coefficient for the j-th flight parameter, The preset validity index value for the j-th flight parameter.

[0096] In an optional implementation of this embodiment, considering that range and standard deviation are more important, therefore, , , .

[0097] S1406' In each sliding time window, select the target sliding time window whose comprehensive score exceeds the preset score threshold, and take the complete flight parameter value set located at the center of each target sliding time window as the stable flight parameter set.

[0098] The preset score threshold can be preset according to the actual situation; for example, it can be selected as 0.83. Based on the above embodiments, after obtaining each stable flight parameter set, Pearson correlation analysis can be used to complete the input feature screening. That is, among the multiple gas path parameters (e.g., 22) associated with exhaust temperature, the gas path parameters most related to exhaust temperature (e.g., 16) are retained for subsequent more accurate model training.

[0099] S150. Based on the multiple stable flight parameter value sets corresponding to multiple aero engines, a training sample set is constructed, and the pre-constructed three-level fusion network is trained using the training sample set to obtain the margin prediction model.

[0100] In an optional implementation of this embodiment, a training sample set is constructed based on multiple stable flight parameter value sets corresponding to multiple aero engines, which may specifically include: S1501. Sort the stable flight parameter value sets according to the acquisition time to obtain a time series data sample set, and continuously extract multiple data sample subsequences of target data length from the time series data sample set.

[0101] S1502. In the set of stable flight parameter values ​​minus 1 of the previous target data length of each data sample subsequence, retain the values ​​of each gas path parameter, and calculate the exhaust temperature margin value that matches the exhaust temperature value of the last stable flight parameter value set as the data label to obtain the training sample set.

[0102] In a specific example, assuming the target data length is 6, multiple data sample subsequences can be constructed, each containing a set of 6 stable flight parameter values. These subsequences may have some overlap in their stable flight parameter value sets.

[0103] For a data sample subsequence A, the gas path parameter values ​​from the first 5 stable flight parameter value sets can be retained to form feature value set 1, feature value set 2, feature value set 3, feature value set 4 and feature value set 5 respectively, and the exhaust temperature margin value from the 6th stable flight parameter value set can be used as data label A. Combine them to obtain a training sample in the form of {feature value set 1, feature value set 2, feature value set 3, feature value set 4, feature value set 5, data label A}, so as to finally obtain the above training sample set.

[0104] Specifically, before training the model using the training sample set, in order to eliminate the impact of differences in the dimensions and numerical ranges of different features on the model training, it is necessary to first normalize each feature value within each feature value set in the training sample set. In a specific example, the min-maximum normalization method can be used to achieve the above normalization process.

[0105] Specifically, it can be done according to the formula: The eigenvalues ​​are then normalized. This represents the feature value of feature x in a specific training sample before normalization. This represents the eigenvalues ​​of feature x after normalization. This represents the minimum value of feature x in the training sample set. This represents the maximum value of feature x in the training sample set. Feature x can be understood as specific flight parameters such as N1, N2, or FF mentioned above.

[0106] The normalized training sample set is divided into a training set and a test set according to a preset ratio. In this embodiment, the preset ratio is preferably 7:3, that is, 70% of the samples are used as the training set and 30% of the samples are used as the test set. The training set is used to train the model, and the test set is used to verify the model's ability to determine the timing of wing washing for unknown samples.

[0107] In this embodiment, a specific implementation of a three-level fusion network is proposed. This fusion network is used to output a one-dimensional exhaust temperature margin prediction value based on the feature value set from multiple consecutive historical acquisition times. The three-level fusion network includes, in sequence, a bidirectional long short-term memory (BiLSTM) network for extracting multi-parameter bidirectional local temporal features, a multi-head self-attention global context encoding network (referred to as the enhanced Transformer module) for extracting global context enhancement features, and a Sparrow Search Algorithm optimized Extreme Gradient Boosting Prediction (SSA-XGBoost) network for obtaining the exhaust temperature margin prediction value.

[0108] Specifically, the BiLSTM network includes multiple parallel feature processing branches, as well as a branch aggregation and fusion layer that is connected to the output of each feature processing branch. Each feature processing branch is used to independently receive the feature subset corresponding to the set type of flight parameters (one or more types of flight parameters) from the original stable flight parameter time-series features of the input time-series sample (i.e., a normalized training sample in the training set). Each feature processing branch includes a forward Long Short-Term Memory (LSTM) layer, a reverse Long Short-Term Memory layer, and an intra-branch fusion layer that are connected to the outputs of the forward Long Short-Term Memory layer and the reverse Long Short-Term Memory layer, respectively, which are independent and run in parallel. The forward long short-term memory layer is used to extract the forward temporal features of the corresponding feature subset along the forward time sequence, the reverse long short-term memory layer is used to extract the reverse temporal features of the corresponding feature subset along the reverse time sequence, and the branch fusion layer is used to fuse the forward and reverse temporal features to obtain and output the bidirectional local temporal features of a single branch. The branch aggregation and fusion layer is used to fuse the bidirectional local temporal features of each individual branch to obtain multi-parameter bidirectional local temporal features.

[0109] Among them, Figure 2 The diagram shows a schematic representation of the feature processing branch in a bidirectional long short-term memory temporal network applicable to Embodiment 1 of the present invention. Figure 2 As shown, multiple forward LSTM units ( Figure 2 The forward LSTM layer is formed by multiple backward LSTM units (referred to as forward LSTM in Chinese). Figure 2The backward LSTM layer is formed by combining the forward LSTM layer and the backward LSTM layer. The outputs of the forward LSTM layer and the backward LSTM layer are fused together through a branch fusion layer (LSTM). Figure 2 (Simplified to a splicing module) is connected to fuse the forward temporal features extracted by the forward LSTM layer with the reverse temporal features extracted by the reverse long short-term memory layer, resulting in and outputting single-branch bidirectional local temporal features. Figure 2 x0, x1, ..., x t ... x n It can be understood as the feature vectors corresponding to different time steps within a time series. The above feature vectors together constitute a feature subset corresponding to a feature processing branch. A single feature vector contains one or more flight parameter feature values.

[0110] It is understandable that bidirectional long short-term memory temporal networks consist of multiple... Figure 2 The feature processing branches shown are arranged in parallel. The outputs of each feature processing branch are connected through a branch aggregation and fusion layer to output the final multi-parameter bidirectional local temporal features.

[0111] Specifically, the bidirectional long short-term memory temporal network divides the multivariate time-series data in the currently output training samples into at least two feature subsets according to feature categories. These subsets are then input into the feature processing branch corresponding to each feature subset. The forward and backward LSTM layers in each feature processing branch process the data separately, and the resulting features are fused using a concatenation method to obtain bidirectional temporal features corresponding to each feature subset. These bidirectional temporal features are then concatenated and fused again by the branch aggregation and fusion layer to output multi-parameter bidirectional local temporal features, which are then transmitted to the enhanced Transformer module. This approach allows for the simultaneous utilization of both forward and backward temporal information from the input data, more effectively capturing local bidirectional temporal correlation features in the multi-source operating parameters of aero-engines.

[0112] In an optional implementation of this embodiment, the multi-head self-attention global context coding network specifically includes: a position coding unit, a multi-head self-attention unit, a feedforward network unit, a decoder unit, and a transpose splicing unit connected in sequence; The location encoding unit is used to embed temporal location information into the input multi-parameter bidirectional local temporal features to obtain a temporal location time series vector, and to fuse the temporal location time series vector with the multi-parameter bidirectional local temporal features to obtain a fused temporal feature with location identifier; The multi-head self-attention unit is used to generate a query matrix, key matrix, and value matrix corresponding to each self-attention channel based on the fused temporal features with location identifiers and the channel weights of each attention channel. Then, it calculates the information matrix of each attention channel based on the query matrix, key matrix, and value matrix corresponding to each self-attention channel. Based on the information matrix of each attention channel and the output weight matrix of the multi-head attention, the initial global context features are obtained. The decoder unit, including a weight decay algorithm layer and a random deactivation layer, is used to obtain the global context feature encoding result corresponding to the initial global context features while effectively reducing computational complexity. The transpose splicing unit is used to transpose the global context feature encoding result and the multi-parameter bidirectional local temporal feature respectively, and then splice each transpose result horizontally along the feature dimension to obtain and output the global context enhancement feature.

[0113] In this optional embodiment, the position encoding unit can be specifically based on the formula: Position encoding is performed on the multi-parameter bidirectional local temporal features output by the bidirectional long short-term memory temporal network.

[0114] in, This refers to the positional encoding value in the even-numbered dimension (2u) within the feature vector corresponding to the pos-th temporal position in the positional encoding matrix. This refers to the encoded value in the (2u+1)th odd-numbered dimension within the feature vector corresponding to the pos-th temporal position in the positional coding matrix. , Let be the length of the time series sequence for the multi-parameter bidirectional local temporal feature (i.e., the maximum number of time steps), and u be the integer index for the dimension grouping. , and These represent the even-numbered and odd-numbered dimensions within the feature vector, respectively. This represents the feature dimension at a single time step in a multi-parameter bidirectional local temporal feature.

[0115] After obtaining the location encoding matrix carrying location information Subsequently, by fusing PE with the multi-parameter bidirectional local temporal feature X”, a fused temporal feature X2 with location identifier can be obtained and input into the multi-head self-attention unit.

[0116] A multi-head self-attention unit contains multiple self-attention channels, and each self-attention channel corresponds to a self-attention head. Each self-attention head has three weight matrices (or channel weights) pre-trained.

[0117] Suppose the weight matrix (or channel weights) corresponding to the e-th self-attention head is: Then, a query matrix corresponding to the e-th self-attention head can be generated based on the fused temporal feature X2. Key matrix AND-value matrix .

[0118] Specifically, , , . h represents the total number of self-attention heads in a multi-head self-attention unit. , , , , and These represent the feature dimensions of the query matrix (Q matrix), key matrix (K matrix), and value matrix (V matrix), respectively, where R is the set of real numbers. In this embodiment, When the number of self-attention channels, or self-attention heads, is 4, h=4.

[0119] Then, the dot product attention calculation formula can be scaled according to the query matrix, key matrix, and value matrix of each attention channel: We solve for the information matrix head corresponding to each self-attention channel (or self-attention head). Here, softmax(.) is the normalized exponential function.

[0120] Correspondingly, the information matrix head corresponds to the e-th self-attention head. e It can be represented as: .

[0121] After obtaining the information matrix for each attention channel, we can use the formula: The output matrix of the multi-head self-attention unit is calculated. , as the initial global context features. The output weight matrix of the multi-head attention system is obtained through pre-training; the Concat() function is the feature concatenation function. .

[0122] The above initial global context features Inputting the data into the decoder unit of a multi-head self-attention global context encoder network yields the same result as... The corresponding global context feature encoding result.

[0123] In an optional implementation of this embodiment, before inputting the initial global context features into the decoder unit, the initial global context features can first be input into the feedforward processing subunit, which includes a first residual connection and normalization layer (referred to as Add&Norm), a feedforward network, and a second residual connection and normalization layer connected in sequence.

[0124] The first residual connection and normalization layer performs residual connection (Add) and layer normalization (Norm) operations: the input features of the multi-head self-attention unit are added element-wise with the output features of the multi-head self-attention unit, and then layer normalization is performed. The normalized features are input into the feedforward network, which independently performs nonlinear feature mapping on the features at each temporal position; the output of the feedforward network is fed into the second residual connection and normalization layer, where residual addition and layer normalization are performed again to obtain the final output features of the encoder, which are then transmitted to the decoder unit.

[0125] Unlike existing decoder structures, the decoder unit in this invention introduces a weight decay algorithm layer and a random deactivation layer (Dropout layer) before the linear layer in existing decoders. Through these improvements, while preserving the model's ability to model global context information, the computational burden is effectively reduced, and overfitting is suppressed. Finally, the encoding result of the multi-head self-attention global context encoding network, combined with the context information of the entire training set, extracts enhanced features of a preset dimension (e.g., 16 dimensions). Compared to the original input data, these enhanced features have a more compact and effective feature representation, which helps reduce the processing complexity in subsequent prediction processes.

[0126] Furthermore, multi-parameter bidirectional local temporal features are... The global context feature encoding result output by the decoder unit The input is fed into the transpose concatenation unit, where it is transposed and then horizontally concatenated along the feature dimension to output global context-enhanced features. As a new sample set, it is input into the Sparrow Search Optimization Gradient Boosting Tree Prediction Network.

[0127] Specifically, in Figure 3 The diagram shows a schematic representation of a multi-head self-attention global context coding network according to Embodiment 1 of the present invention. Figure 3 As shown, the multi-parameter bidirectional local temporal feature X” is input as a regular feature into a multi-head self-attention global context encoding network, and then processed by the position encoding unit ( Figure 3 The embedded layer part can obtain a time series vector with the time position (in the embedded layer part). Figure 3The fused temporal feature X2 (the positional encoding tensor in the code) is obtained. This fused temporal feature X2 is then processed sequentially by the multi-head self-attention unit, the feedforward processing subunit, the decoder unit, and the transpose concatenation unit to obtain the final global context enhancement feature. The output is fed into the Sparrow Search Optimization Gradient Boosting Tree Prediction Network.

[0128] In an optional implementation of this embodiment, the Sparrow Search Optimization Gradient Boosting Tree Prediction Network includes a parameter optimization unit and a gradient boosting tree regression unit.

[0129] The parameter optimization unit is used to perform global optimization of the hyperparameter combination of the gradient boosting tree using a sparrow population iterative search mechanism to obtain the globally optimal hyperparameter combination. The hyperparameter combination includes at least one of the following: model learning step size, maximum depth of decision tree, number of weak learners, random sampling ratio of samples, and random sampling ratio of features. The gradient boosting tree regression unit is used to build an integrated decision tree model based on the globally optimal hyperparameter combination, iteratively train multiple decision trees based on the received global context enhancement features, construct a loss function based on second-order Taylor expansion to optimize the integrated decision tree model, and accumulate the leaf node scores of each decision tree through the optimized integrated decision tree model to obtain and output the exhaust temperature margin prediction value.

[0130] In this embodiment, after the multi-head self-attention global context encoding network fuses the enhanced features and the input multi-parameter bidirectional local temporal features to obtain global context enhancement features, these global context enhancement features are used as prediction input features and input into the Sparrow Search Optimized Gradient Boosting Tree Prediction Network. In this network, the SSA algorithm is used to optimize the hyperparameters of the XGBoost model to obtain an optimized XGBoost prediction model. The optimized hyperparameters include at least one of the following: learning rate, number of weak learners, sample sampling ratio, maximum tree depth, and feature sampling ratio. This embodiment of the invention uses the SSA algorithm to optimize the selection of hyperparameters for the XGBoost model. Through the SSA algorithm, the efficiency and stability of searching the core hyperparameters of the XGBoost model can be improved, thereby enhancing the overall prediction accuracy.

[0131] Specifically, in Figure 4 The diagram shows a schematic of the structure of a sparrow search optimization gradient boosting tree prediction network provided in Embodiment 1 of the present invention.

[0132] Specifically, such as Figure 4 As shown: The parameter optimization unit first initializes the parameters of the gradient boosting tree model.

[0133] Among them, the hyperparameters to be initialized include the learning step size. Maximum depth of decision tree Number of weak learners Random sampling ratio and the proportion of random sampling features The hyperparameters to be initialized are represented as a hyperparameter input matrix (also known as a population parameter matrix). Its expression is as follows: Each data element in the matrix is ​​denoted as . Column index v represents the v-th hyperparameter to be optimized, and row index u represents the u-th candidate solution. Matrix elements This represents the value of the v-th hyperparameter to be optimized in the u-th candidate solution group. Each line represents a complete hyperparameter combination. Suppose there are V hyperparameters to be optimized and U candidate solutions exist.

[0134] According to the formula: This generates initial candidate solutions for the v-th hyperparameter to be optimized in the u-th group. and Let be the lower and upper boundaries of the v-th hyperparameter to be optimized, respectively. A random number between 0 and 1.

[0135] Based on the initial candidate solutions of each hyperparameter to be optimized in each group, an initial candidate solution set is constructed as the sparrow population of the sparrow search algorithm. The fitness value corresponding to each group of candidate solutions is calculated, and the sparrow population is divided according to the fitness: the individuals with the best fitness are selected as discoverers, and the individuals other than the discoverers are selected as followers.

[0136] According to the formula: The location of the discoverer is iteratively updated.

[0137] in, This represents the v-th hyperparameter in the u-th candidate solution group during the t-th iteration, where t is the iteration number. The maximum number of iterations, It is a random number. ; To set the preset warning value, ; To preset a safety value, ; These are random numbers that follow a normal distribution. for A matrix of all 1s.

[0138] After updating the discoverer's location, iterative updates of the follower's location within the sparrow population are needed. The update formula is as follows: .

[0139] in, exp(.) is an exponential function with base e. Let v be the v-th hyperparameter in the u-th candidate solution group in the (t+1)-th iteration after the update. This is the current optimal candidate solution obtained after the discoverer completes the update. This is the worst candidate solution globally within the current iteration period; for The sign matrix, whose elements are randomly assigned the value 1 or -1.

[0140] After updating the positions of the followers, randomly select some individuals within the sparrow population as watchdogs, and follow the formula: The location of the vigilant is updated iteratively.

[0141] in, This represents the current globally optimal solution in the t-th iteration; This represents the step size control parameter, which is a random number that follows a standard normal distribution; Represents a random number. ; This represents the u-th candidate solution in the t-th iteration. fitness value, This represents the fitness value of the current globally optimal solution; This is the worst candidate solution in the t-th iteration. This represents a preset minimum constant, used to avoid cases where the denominator is 0.

[0142] The algorithm iteratively updates the positions of the discoverers, followers, and watchers in the sparrow population. When the maximum number of iterations is reached, or a preset convergence condition is met, the iteration stops, and the current globally optimal solution is output. , which serves as the globally optimal combination of hyperparameters for the gradient boosting tree model.

[0143] The gradient boosting tree regression unit builds an ensemble decision tree model based on the globally optimal hyperparameter combination output by the parameter optimization unit, and receives global context enhancement features from the decoder unit of the multi-head self-attention global context encoding network. The feature is used as a training sample to train the ensemble decision tree model to obtain the final margin prediction model.

[0144] Specifically, suppose the ensemble decision tree model includes The first decision tree, the The objective function of a decision tree can be expressed by the formula: and To express.

[0145] in, To integrate the final predicted value of the decision tree model for the s-th sample, , where n represents the total number of samples; Or simply Let represent the decision function of the c-th decision tree, where c is the traversal index of the decision tree. This provides global context enhancement features for the s-th sample. This represents the preset loss function; Indicates and The corresponding real tags; This represents the cumulative prediction value of the first t-1 decision trees for sample s; Or simply , represents the decision function of the t-th decision tree to be trained in the current iteration process. This represents a predefined regularization function used to characterize the model complexity of the c-th decision tree.

[0146] The loss function is approximated using a second-order Taylor expansion, and the constant term is removed to obtain the... The simplified objective function for a decision tree is shown below: in, For loss function The first derivative at sample s For loss function The second derivative at sample s.

[0147] because It is a constant term and will not affect The solution, and The training is complete, and the values ​​are all fixed. Therefore, the above objective function formula can be further simplified to: Function to obtain the regularization term of the t-th decision tree: ,in, Let be the total number of leaf nodes in the t-th decision tree. Let be the score of the j-th leaf node in the t-th decision tree. and These are the regularization coefficients, used to control the leaf node score and the number of leaf nodes, respectively.

[0148] Based on this regularization function, by combining the coefficients of the first and second terms of the aforementioned objective function formula and removing the constant term, we can obtain the following updated objective function: ; in, , representing the sum of the first derivatives of all samples within the j-th leaf. , representing the sum of the second derivatives of all samples within the j-th leaf. Let represent the set of samples that fall into the j-th leaf.

[0149] Leaf weights By taking the derivative and setting it to zero, we can find the optimal solution for the j-th leaf node. Substituting the optimal solution above into the latest objective function, we can simplify to obtain the optimal objective function for the t-th tree: .

[0150] Based on the above optimization objective function, each decision tree is optimally partitioned. The partitioned decision trees are then used as the basis for the next round of optimization and training. After training, the optimized ensemble decision tree model is then used. Figure 4 (In short, the optimal decision tree) The scores of the leaf nodes of each decision tree are summed to obtain and output the model's exhaust temperature margin prediction. .

[0151] In this embodiment, by using a training sample set to train the fusion network using the aforementioned computational logic, a margin prediction model can be finally trained.

[0152] It is important to reiterate that this embodiment of the invention utilizes a bidirectional long short-term memory temporal network to independently model each operating parameter bidirectionally, and then achieves deep fusion of multi-source features through an aggregation and fusion layer. This avoids the feature crosstalk problem encountered when mixing multiple parameters from engine sensors, while also fully capturing the cooperative degradation coupling relationship between parameters. Simultaneously, the bidirectional mechanism enables the network to utilize contextual information from both historical and future moments, effectively improving the model's sensitivity to identifying engine performance degradation trends, and effectively capturing local bidirectional temporal correlation features among the multi-source operating parameters of aero-engines.

[0153] Furthermore, by using a multi-head self-attention global context encoding network to model the global context of the output features of the bidirectional long short-term memory temporal network, the model can further perceive the global context information and extract global context-enhanced features, effectively improving the prediction accuracy of the model. Simultaneously, introducing a weight decay algorithm and a random deactivation layer into the decoder unit can effectively reduce the model's computational burden and processing complexity, suppress overfitting, and improve the model's prediction accuracy and computational efficiency while preserving the model's ability to model global context features.

[0154] Meanwhile, the gradient boosting tree prediction network is optimized by sparrow search to perform global hyperparameter optimization of the gradient boosting tree. By perturbing and correcting candidate solutions in local optimal regions, boundary regions, and low fitness regions, the population maintains good search diversity during the iteration process, avoids premature convergence to local optima, improves parameter optimization accuracy and model prediction stability, and thus ensures the accuracy of the overall prediction results.

[0155] S160. Input the predicted input sequence of the wing-mounted aero-engine into the margin prediction model to obtain the predicted value of the exhaust temperature margin of the wing-mounted aero-engine.

[0156] The predicted input sequence includes multiple target gas path parameter value groups sorted by acquisition time, and each target gas path parameter value group includes multiple gas path parameter values.

[0157] S170. When the difference between the predicted exhaust temperature margin and the standard margin value is greater than the difference threshold, it is determined that the on-wing aero-engine needs to be washed with water.

[0158] In this embodiment, when it is necessary to predict the timing of wing-mounted air washing for wing-mounted aero-engine A under the current system time, it is necessary to first collect multiple target gas path parameter values ​​of wing-mounted aero-engine A over a continuous period of time before the current system time. For example, within the first 20 minutes of the current system time, multiple gas path parameter values ​​collected every 5 minutes are combined into a prediction input sequence. By comparing the predicted exhaust temperature margin value obtained from the margin prediction model with the margin standard value (e.g., 10°C), if the absolute difference between the two is greater than 10°C, it is determined that wing washing should be performed; if the absolute difference between the two is less than or equal to 10°C, it is determined that flight should continue.

[0159] The technical solution of this invention improves the data quality of model training samples by performing operations such as filling missing flight parameters, pruning and smoothing abnormal parameter values, standardizing correction, and extracting stable flight parameter value sets on the time series set of aero-engine flight data. The margin prediction model trained using a three-level fusion network, comprising a bidirectional long short-term memory time series network, a multi-head self-attention global context encoding network, and a sparrow search optimization gradient boosting tree prediction network, can accurately predict exhaust temperature margin. Finally, by performing a Boolean comparison between the predicted exhaust temperature margin value and the standard value, the on-wing water washing timing of the aero-engine is directly determined, facilitating engineering applications.

[0160] Furthermore, to verify the effectiveness of the model proposed in this embodiment, the inventors conducted comparative experiments under four configurations: CNN, LSTM, XGBoost, and CNN+LSTM. The inventors converted the exhaust temperature margin prediction value output by the model into a binary classification result, where 0 indicates no on-wing washing is required, and 1 indicates on-wing washing is required. According to the formula: The accuracy of the model is calculated. Here, Acu (Accuracy) represents the accuracy rate. This indicates the number of samples that actually require wing washing, and the model prediction also indicates that wing washing is necessary. This indicates the number of samples for which wing washing is actually unnecessary and which the model predicts will also be unnecessary. This represents the total number of samples in the test set. The experimental results are shown in Table 2.

[0161] Table 2 Experiments show that the model of this invention achieves an accuracy of 0.994 in the task of determining the timing of wing washing for aero-engines, which is higher than that of four comparative models: CNN, LSTM, XGBoost, and CNN-LSTM. Compared with the comparative models, the accuracy of the model of this invention is improved by 0.7 percentage points for the optimal result of 0.987. This verifies that extracting multi-parameter bidirectional local temporal features through a bidirectional long short-term memory temporal network, extracting global context enhancement features through a multi-head self-attention global context encoding network, and optimizing parameters through sparrow search to optimize the gradient boosting tree prediction network can effectively improve the accuracy of determining the timing of wing washing. From the comprehensive requirements of aero-engine health management for model robustness, operational condition adaptability, and engineering interpretability, the model of this invention still has more complete engineering application value.

[0162] Figure 5 This is a schematic diagram illustrating the principle of the method for predicting the timing of wing washing in an aero-engine, as described in this embodiment of the invention. Figure 5 As shown, raw QAR flight data of the aero-engine is acquired, and data preprocessing, parameter standardization and correction, stable feature point extraction, and feature selection are performed sequentially to construct a multivariate time series dataset. The data preprocessing steps include missing value imputation, outlier detection and correction, adaptive smoothing and noise reduction, and EGTM calculation. The multivariate time series dataset is sequentially input into a bidirectional long short-term memory network for temporal feature extraction, followed by global context feature enhancement through a multi-head self-attention global context encoding network. Then, it is input into a sparrow search optimized gradient boosting tree prediction network for prediction, outputting a predicted exhaust temperature margin value. This value is compared with a preset on-wing washing threshold; if it exceeds the on-wing washing threshold, on-wing washing is performed; otherwise, flight continues.

[0163] Example 2 Figure 6 This is a schematic diagram of a device for predicting the timing of wing washing in an aircraft engine, provided in Embodiment 2 of the present invention. Figure 6 As shown, the device includes: a flight data time series set acquisition module 610, a filling processing module 620, a clipping and smoothing processing module 630, a stable flight parameter value set extraction module 640, a margin prediction model training module 650, an exhaust temperature margin prediction value determination module 660, and an on-wing water washing timing determination module 670, wherein: The flight data time series set acquisition module 610 is used to acquire the flight data time series set corresponding to the aero-engine, wherein the flight data time series set includes multiple initial flight parameter value sets sorted according to the acquisition time; The filling processing module 620 is used to fill the parameter values ​​of missing flight parameters in each initial flight parameter set to obtain each complete flight parameter value set. The complete flight parameter value set includes an exhaust temperature value that matches the set acquisition time and multiple gas path parameter values ​​that correspond to the exhaust temperature value. The trimming and smoothing module 630 is used to trim abnormal parameter values ​​from each complete flight parameter value set after the filling process, and to smooth the parameter values ​​of different types of flight parameters in each complete flight parameter value set after the trimming process. The stable flight parameter value set extraction module 640 is used to extract at least one stable flight parameter value set from each complete flight parameter value set after standardizing and correcting the second type of flight parameter values ​​in each complete flight parameter value set by calculating the temperature correction factor and pressure correction factor respectively from the first type of flight parameter values ​​in each complete flight parameter value set. The margin prediction model training module 650 is used to construct a training sample set based on multiple stable flight parameter value sets corresponding to multiple aero engines, and to train a pre-constructed three-level fusion network using the training sample set to obtain the margin prediction model. The three-level fusion network includes a bidirectional long short-term memory temporal network for extracting multi-parameter bidirectional local temporal features, a multi-head self-attention global context encoding network for extracting global context enhancement features, and a sparrow search optimized gradient boosting tree prediction network for obtaining exhaust temperature margin prediction values. The exhaust temperature margin prediction value determination module 660 is used to input the prediction input sequence of the wing-mounted aero-engine into the margin prediction model to obtain the exhaust temperature margin prediction value of the wing-mounted aero-engine. The prediction input sequence includes multiple target gas path parameter value groups sorted according to the acquisition time, and each target gas path parameter value group includes multiple gas path parameter values. The on-wing washing timing determination module 670 is used to determine that on-wing washing of the aero-engine is required when the difference between the predicted value of the exhaust temperature margin and the standard value of the margin is greater than the difference threshold.

[0164] The technical solution of this invention improves the data quality of model training samples by performing operations such as filling missing flight parameters, pruning and smoothing abnormal parameter values, standardizing correction, and extracting stable flight parameter value sets on the time series set of aero-engine flight data. The margin prediction model trained using a three-level fusion network composed of a bidirectional long short-term memory time series network, a multi-head self-attention global context encoding network, and a sparrow search optimized gradient boosting tree prediction network can accurately predict exhaust temperature margin. Finally, by performing a Boolean comparison between the predicted exhaust temperature margin value and the standard value, the on-wing water washing timing of the aero-engine is directly determined, facilitating engineering applications.

[0165] Based on the above embodiments, the bidirectional long short-term memory temporal network includes multiple parallel feature processing branches and a branch aggregation and fusion layer that is connected to the output of each feature processing branch. Each feature processing branch of the bidirectional long short-term memory temporal network is used to independently receive the feature subset corresponding to the set type of flight parameters in the original stable flight parameter temporal features of the input temporal sample. Each feature processing branch includes a forward long short-term memory layer, a reverse long short-term memory layer, and an intra-branch fusion layer that are independently and in parallel. The forward long short-term memory layer is used to extract the forward temporal features of the corresponding feature subset along the forward time sequence, the reverse long short-term memory layer is used to extract the reverse temporal features of the corresponding feature subset along the reverse time sequence, and the branch fusion layer is used to fuse the forward and reverse temporal features to obtain and output the bidirectional local temporal features of a single branch. The branch aggregation and fusion layer is used to fuse the bidirectional local temporal features of each individual branch to obtain multi-parameter bidirectional local temporal features.

[0166] Based on the above embodiments, the multi-head self-attention global context coding network includes a position coding unit, a multi-head self-attention unit, a feedforward network unit, a decoder unit, and a transpose splicing unit connected in sequence. The location encoding unit is used to embed temporal location information into the input multi-parameter bidirectional local temporal features to obtain a temporal location time series vector, and to fuse the temporal location time series vector with the multi-parameter bidirectional local temporal features to obtain a fused temporal feature with location identifier; The multi-head self-attention unit is used to generate a query matrix, key matrix, and value matrix corresponding to each self-attention channel based on the fused temporal features with location identifiers and the channel weights of each attention channel. Then, it calculates the information matrix of each attention channel based on the query matrix, key matrix, and value matrix corresponding to each self-attention channel. Based on the information matrix of each attention channel and the output weight matrix of the multi-head attention, the initial global context features are obtained. The decoder unit, including a weight decay algorithm layer and a random deactivation layer, is used to obtain the global context feature encoding result corresponding to the initial global context features while effectively reducing computational complexity. The transpose splicing unit is used to transpose the global context feature encoding result and the multi-parameter bidirectional local temporal feature respectively, and then splice each transpose result horizontally along the feature dimension to obtain and output the global context enhancement feature.

[0167] Based on the above embodiments, the sparrow search optimization gradient boosting tree prediction network includes a parameter optimization unit and a gradient boosting tree regression unit; The parameter optimization unit of the sparrow search optimization gradient boosting tree prediction network is used to perform global optimization of the hyperparameter combination of the gradient boosting tree using a sparrow population iterative search mechanism to obtain the globally optimal hyperparameter combination. The hyperparameter combination includes at least one of the following: model learning step size, maximum depth of decision tree, number of weak learners, random sampling ratio of samples, and random sampling ratio of features. The gradient boosting tree regression unit is used to build an integrated decision tree model based on the globally optimal hyperparameter combination, iteratively train multiple decision trees based on the received global context enhancement features, construct a loss function based on second-order Taylor expansion to optimize the integrated decision tree model, and accumulate the leaf node scores of each decision tree through the optimized integrated decision tree model to obtain and output the exhaust temperature margin prediction value.

[0168] Based on the above embodiments, the filling processing module 620 can be specifically used for: Cluster analysis is performed on the parameter values ​​of each flight parameter in each initial flight parameter value set to calculate the missing rate of each flight parameter in the entire initial flight parameter value set, and missing flight parameters with missing rates greater than or equal to a preset missing rate threshold are identified. Based on the thermodynamic coupling relationship between various flight parameters, several key coupling parameters that match the missing flight parameters are selected from among the various flight parameters; After performing Z-score standardization on the parameter values ​​of each flight parameter in the entire initial flight parameter value set, a complete set of initial flight parameter values ​​for multiple targets is extracted. In each target initial flight parameter value set, the parameter values ​​of each key coupling parameter are extracted as input features of the preset regression model, and the parameter values ​​of each missing flight parameter are extracted as training labels. The optimal coefficients for calculating the missing flight parameters based on each key coupling parameter are then solved. Based on the optimal coefficients obtained, the parameter values ​​of the missing flight parameters in each initial flight parameter set that has undergone Z-score standardization are filled to obtain the filled parameter values ​​that match the missing flight parameters. Based on the maximum and minimum parameter values ​​defined by the physical characteristics of the missing flight parameters, each filler parameter value is pruned to obtain a complete set of flight parameter values.

[0169] Based on the above embodiments, the cropping smoothing processing module 630 can be specifically used for: Based on the complete set of flight parameter values, the mean μ and standard deviation σ of each flight parameter are calculated. Based on the preset standard deviation criterion, the upper critical value μ+3σ and the lower critical value μ-3σ of each flight parameter are calculated. The parameter values ​​of each flight parameter in the complete flight parameter value set are numerically compared with the matching upper and lower critical values; If the first target parameter value in the complete set of first target flight parameter values ​​is greater than the target upper critical value of the first target flight parameter, then the first target parameter value is corrected to the target upper critical value; If the second target parameter value in the complete set of flight parameter values ​​of the second target is less than the target lower critical value of the second target flight parameter, then the second target parameter value is corrected to the target lower critical value.

[0170] Based on the above embodiments, the cropping smoothing module 630 can be further used for: For temperature-related flight parameters, a Savitzky-Golay filter is used for smoothing. For pressure-related parameters, a Gaussian smoothing algorithm is used for smoothing. For the parameter values ​​of fuel flow and system control parameters, a locally weighted scatter smoothing algorithm is used for smoothing; and For the parameter values ​​of speed-related parameters, a moving average algorithm is used for smoothing.

[0171] Based on the above embodiments, the stable flight parameter value set extraction module 640 can be specifically used for: In each complete flight parameter value set, the current complete flight parameter value set is obtained sequentially; In the current complete set of flight parameter values, the total atmospheric temperature (TAT), Mach number (Ma), and flight altitude (ALT) are extracted as the first type of flight parameter values. According to the formula: The temperature correction factor was calculated. And according to the formula: The pressure correction factor was calculated. ; Extract the original low-pressure rotor speed value from the current complete flight parameter value set. High-voltage rotor speed value Exhaust temperature value and fuel flow rate As a second type of flight parameter value; According to the formula: Calculation yields the pair Standardization correction results ; According to the formula: Calculation yields the pair Standardization correction results ; According to the formula: Calculation yields the pair Standardization correction results ; According to the formula: Calculation yields the pair Standardization correction results ; Return to the previous operation and retrieve the current complete flight parameter value set in each complete flight parameter value set sequentially until the processing of all complete flight parameter value sets is completed.

[0172] Based on the above embodiments, the stable flight parameter value set extraction module 340 can be further used for: According to the preset sliding step size, the preset total number of parameter sets W, and the acquisition time of each complete flight parameter value set, each complete flight parameter value set is divided into multiple sliding time windows, wherein each sliding time window includes W complete flight parameter value sets; According to the formula: The range score value of the j-th flight parameter in the i-th sliding time window is calculated; Where represents the numerical range of the j-th flight parameter within the i-th sliding time window. , Let be the parameter value of the j-th flight parameter in the k-th complete flight parameter value set within the i-th sliding time window, where max() is the maximum value function and min() is the minimum value function; and represent the standard deviation of the j-th flight parameter within the i-th sliding time window. , Let be the mean of the j-th flight parameter within the i-th sliding time window. , The preset stability tolerance threshold for the j-th flight parameter; According to the formula: The standard deviation score of the j-th flight parameter in the i-th sliding time window is calculated. ; According to the formula: The timing score value of the j-th flight parameter within the i-th sliding time window is calculated. ; in, Let j be the mean absolute difference of the flight parameter in the i-th sliding time window. , The parameter value of the j-th flight parameter in the (k+1)-th complete flight parameter value set within the i-th sliding time window; According to the formula: The overall score for the i-th sliding time window is calculated. ; Where α, β, and γ are preset weighting coefficients, and m is the total number of flight parameters included in the complete flight parameter value set. The preset weighting coefficient for the j-th flight parameter, The preset validity index value for the j-th flight parameter; Within each sliding time window, target sliding time windows with comprehensive scores exceeding a preset score threshold are selected, and the complete flight parameter value set located at the center of each target sliding time window is used as the stable flight parameter set.

[0173] Based on the above embodiments, the margin prediction model training module 650 can be specifically used for: The stable flight parameter value sets are sorted according to the acquisition time to obtain a time series data sample set, and multiple data sample subsequences of target data length are continuously extracted from the time series data sample set. In each data sample subsequence, within the set of stable flight parameter values ​​minus one of the preceding target data lengths, retain the values ​​of each gas path parameter and calculate the exhaust temperature margin value that matches the exhaust temperature value of the last stable flight parameter value set as the data label to obtain the training sample set.

[0174] The prediction device for the timing of wing washing of an aero-engine provided in the embodiments of the present invention can execute the prediction method for the timing of wing washing of an aero-engine provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method.

[0175] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0176] Figure 7 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0177] like Figure 7 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0178] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0179] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, for example, performing a method for predicting the timing of wing washing of an aircraft engine as described in any embodiment of the present invention.

[0180] In some embodiments, the method for predicting the timing of wing washing of an aircraft engine as described in any embodiment of the present invention can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the method for predicting the timing of wing washing of an aircraft engine as described above in any embodiment of the present invention can be performed. Alternatively, in other embodiments, processor 11 can be configured by any other suitable means (e.g., by means of firmware) to perform the method for predicting the timing of wing washing of an aircraft engine as described in any embodiment of the present invention. That is: Acquire the flight data time series set corresponding to the aero-engine, wherein the flight data time series set includes multiple initial flight parameter value sets sorted by acquisition time; The missing flight parameter values ​​in each initial flight parameter set are filled to obtain each complete flight parameter value set. The complete flight parameter value set includes the exhaust temperature value that matches the set acquisition time and multiple gas path parameter values ​​that correspond to the exhaust temperature value. Abnormal parameter values ​​are pruned from each complete flight parameter value set after the filling process is completed, and the parameter values ​​of different types of flight parameters in each complete flight parameter value set after pruning are smoothed respectively. Based on the temperature correction factor and pressure correction factor calculated from the first type of flight parameter values ​​in each complete flight parameter value set, after standardizing and correcting the second type of flight parameter values ​​in each complete flight parameter value set, at least one stable flight parameter value set is extracted from each complete flight parameter value set. Based on multiple stable flight parameter value sets corresponding to multiple aero engines, a training sample set is constructed, and the pre-constructed three-level fusion network is trained using the training sample set to obtain a margin prediction model. The three-level fusion network includes a bidirectional long short-term memory temporal network for extracting multi-parameter bidirectional local temporal features, a multi-head self-attention global context encoding network for extracting global context enhancement features, and a sparrow search optimized gradient boosting tree prediction network for obtaining exhaust temperature margin prediction values. The predicted input sequence of the wing-mounted aero-engine is input into the margin prediction model to obtain the predicted value of the exhaust temperature margin of the wing-mounted aero-engine. The predicted input sequence includes multiple target gas path parameter value groups sorted according to the acquisition time, and each target gas path parameter value group includes multiple gas path parameter values. When the difference between the predicted exhaust temperature margin and the standard margin value exceeds the difference threshold, it is determined that the on-wing aero-engine needs to be washed with water.

[0181] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in a computer program product comprising one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0182] The computer programs included in the computer program product for implementing the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, partially on a remote machine as a standalone software package, or entirely on a remote machine or server.

[0183] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0184] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0185] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0186] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0187] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0188] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for predicting the timing of wing washing in an aero-engine, characterized in that, The method includes: Acquire the flight data time series set corresponding to the aero-engine, wherein the flight data time series set includes multiple initial flight parameter value sets sorted by acquisition time; The missing flight parameter values ​​in each initial flight parameter set are filled to obtain each complete flight parameter value set. The complete flight parameter value set includes the exhaust temperature value that matches the set acquisition time and multiple gas path parameter values ​​that correspond to the exhaust temperature value. Abnormal parameter values ​​are pruned from each complete flight parameter value set after the filling process is completed, and the parameter values ​​of different types of flight parameters in each complete flight parameter value set after pruning are smoothed respectively. Based on the temperature correction factor and pressure correction factor calculated from the first type of flight parameter values ​​in each complete flight parameter value set, after standardizing and correcting the second type of flight parameter values ​​in each complete flight parameter value set, at least one stable flight parameter value set is extracted from each complete flight parameter value set. Based on multiple stable flight parameter value sets corresponding to multiple aero engines, a training sample set is constructed, and the pre-constructed three-level fusion network is trained using the training sample set to obtain the margin prediction model. The three-level fusion network includes a bidirectional long short-term memory temporal network for extracting multi-parameter bidirectional local temporal features, a multi-head self-attention global context encoding network for extracting global context enhancement features, and a sparrow search optimization gradient boosting tree prediction network for obtaining exhaust temperature margin prediction values. The predicted input sequence of the wing-mounted aero-engine is input into the margin prediction model to obtain the predicted value of the exhaust temperature margin of the wing-mounted aero-engine. The predicted input sequence includes multiple target gas path parameter value groups sorted according to the acquisition time, and each target gas path parameter value group includes multiple gas path parameter values. When the difference between the predicted exhaust temperature margin and the standard margin value exceeds the difference threshold, it is determined that the on-wing aero-engine needs to be washed with water.

2. The method according to claim 1, characterized in that, The bidirectional long short-term memory temporal network includes multiple parallel feature processing branches and a branch aggregation and fusion layer that is connected to the output of each feature processing branch. Each feature processing branch is used to independently receive the feature subset corresponding to the set type of flight parameters in the original stable flight parameter time series features of the input time series sample. Each feature processing branch includes a forward long short-term memory layer, a reverse long short-term memory layer, and an intra-branch fusion layer that are independently and in parallel. The forward long short-term memory layer is used to extract the forward temporal features of the corresponding feature subset along the forward time sequence, the reverse long short-term memory layer is used to extract the reverse temporal features of the corresponding feature subset along the reverse time sequence, and the branch fusion layer is used to fuse the forward and reverse temporal features to obtain and output the bidirectional local temporal features of a single branch. The branch aggregation and fusion layer is used to fuse the bidirectional local temporal features of each individual branch to obtain multi-parameter bidirectional local temporal features.

3. The method according to claim 2, characterized in that, The multi-head self-attention global context coding network includes a position coding unit, a multi-head self-attention unit, a feedforward network unit, a decoder unit, and a transpose splicing unit connected in sequence. The location encoding unit is used to embed temporal location information into the input multi-parameter bidirectional local temporal features to obtain a temporal location time series vector, and to fuse the temporal location time series vector with the multi-parameter bidirectional local temporal features to obtain a fused temporal feature with location identifier; The multi-head self-attention unit is used to generate a query matrix, key matrix, and value matrix corresponding to each self-attention channel based on the fused temporal features with location identifiers and the channel weights of each attention channel. Then, it calculates the information matrix of each attention channel based on the query matrix, key matrix, and value matrix corresponding to each self-attention channel. Based on the information matrix of each attention channel and the output weight matrix of the multi-head attention, the initial global context features are obtained. The decoder unit, including a weight decay algorithm layer and a random deactivation layer, is used to obtain the global context feature encoding result corresponding to the initial global context features while effectively reducing computational complexity. The transpose splicing unit is used to transpose the global context feature encoding result and the multi-parameter bidirectional local temporal feature respectively, and then splice each transpose result horizontally along the feature dimension to obtain and output the global context enhancement feature.

4. The method according to claim 2, characterized in that, The sparrow search optimization gradient boosting tree prediction network includes a parameter optimization unit and a gradient boosting tree regression unit. The parameter optimization unit is used to perform global optimization of the hyperparameter combination of the gradient boosting tree using a sparrow population iterative search mechanism to obtain the globally optimal hyperparameter combination. The hyperparameter combination includes at least one of the following: model learning step size, maximum depth of decision tree, number of weak learners, random sampling ratio of samples, and random sampling ratio of features. The gradient boosting tree regression unit is used to build an integrated decision tree model based on the globally optimal hyperparameter combination, iteratively train multiple decision trees based on the received global context enhancement features, construct a loss function based on second-order Taylor expansion to optimize the integrated decision tree model, and accumulate the leaf node scores of each decision tree through the optimized integrated decision tree model to obtain and output the exhaust temperature margin prediction value.

5. The method according to any one of claims 2-4, characterized in that, The missing flight parameter values ​​in each initial flight parameter set are filled in to obtain each complete flight parameter value set, including: Cluster analysis is performed on the parameter values ​​of each flight parameter in each initial flight parameter value set to calculate the missing rate of each flight parameter in the entire initial flight parameter value set, and missing flight parameters with missing rates greater than or equal to a preset missing rate threshold are identified. Based on the thermodynamic coupling relationship between various flight parameters, several key coupling parameters that match the missing flight parameters are selected from among the various flight parameters; After performing Z-score standardization on the parameter values ​​of each flight parameter in the entire initial flight parameter value set, a complete set of initial flight parameter values ​​for multiple targets is extracted. In each target initial flight parameter value set, the parameter values ​​of each key coupling parameter are extracted as input features of the preset regression model, and the parameter values ​​of each missing flight parameter are extracted as training labels. The optimal coefficients for calculating the missing flight parameters based on each key coupling parameter are then solved. Based on the optimal coefficients obtained, the parameter values ​​of the missing flight parameters in each initial flight parameter set that has undergone Z-score standardization are filled to obtain the filled parameter values ​​that match the missing flight parameters. Based on the maximum and minimum parameter values ​​defined by the physical characteristics of the missing flight parameters, each filler parameter value is pruned to obtain a complete set of flight parameter values.

6. The method according to any one of claims 2-4, characterized in that, Based on the temperature correction factor and pressure correction factor calculated from the first type of flight parameter values ​​in each complete flight parameter value set, the second type of flight parameter values ​​in each complete flight parameter value set are standardized and corrected, including: In each complete flight parameter value set, the current complete flight parameter value set is obtained sequentially; In the current complete set of flight parameter values, the total atmospheric temperature (TAT), Mach number (Ma), and flight altitude (ALT) are extracted as the first type of flight parameter values. According to the formula: The temperature correction factor was calculated. And according to the formula: The pressure correction factor was calculated. ; Extract the original low-pressure rotor speed value from the current complete set of flight parameter values. High-voltage rotor speed value Exhaust temperature value and fuel flow rate As a second type of flight parameter value; According to the formula: Calculation yields the pair Standardization correction results ; According to the formula: Calculation yields the pair Standardization correction results ; According to the formula: Calculation yields the pair Standardization correction results ; According to the formula: Calculation yields the pair Standardization correction results ; Return to the previous operation and retrieve the current complete flight parameter value set in each complete flight parameter value set sequentially until the processing of all complete flight parameter value sets is completed.

7. The method according to any one of claims 2-4, characterized in that, Extract at least one stable flight parameter set from each complete flight parameter value set, including: According to the preset sliding step size, the preset total number of parameter sets W, and the acquisition time of each complete flight parameter value set, each complete flight parameter value set is divided into multiple sliding time windows, wherein each sliding time window includes W complete flight parameter value sets; According to the formula: The range score value of the j-th flight parameter in the i-th sliding time window is calculated. ; in, Let be the range of values ​​for the j-th flight parameter within the i-th sliding time window. , Let j be the parameter value of the j-th flight parameter in the k-th complete flight parameter value set in the i-th sliding time window, where max() is the function to take the maximum value and min() is the function to take the minimum value. Let $\frac{j}{i}$ be the standard deviation of the $j$-th flight parameter within the $i$-th sliding time window. , Let be the mean of the j-th flight parameter within the i-th sliding time window. , The preset stability tolerance threshold for the j-th flight parameter; According to the formula: The standard deviation score of the j-th flight parameter in the i-th sliding time window is calculated. ; According to the formula: The timing score value of the j-th flight parameter within the i-th sliding time window is calculated. ; in, Let j be the mean absolute difference of the flight parameter in the i-th sliding time window. , The parameter value of the j-th flight parameter in the (k+1)-th complete flight parameter value set within the i-th sliding time window; According to the formula: The overall score for the i-th sliding time window is calculated. ; in, , as well as Here, is the preset weighting coefficient, and m is the total number of flight parameters included in the complete flight parameter value set. The preset weighting coefficient for the j-th flight parameter. The preset validity index value for the j-th flight parameter; Within each sliding time window, target sliding time windows with comprehensive scores exceeding a preset score threshold are selected, and the complete flight parameter value set located at the center of each target sliding time window is used as the stable flight parameter set.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and a memory communicatively connected to said at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the aircraft engine wing washing timing prediction method according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the method for predicting the timing of wing washing of an aero-engine as described in any one of claims 1-7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method for predicting the timing of wing washing of an aero-engine according to any one of claims 1-7.

Images