Aero-engine remaining service life prediction method considering confidence Gaussian process regression network

Abstract

The invention relates to the technical field of aero-engine residual service life prediction and artificial intelligence, in particular to an aero-engine residual service life prediction method considering a confidence Gaussian process regression network, which comprises the following steps: acquiring a multi-source monitoring signal and a corresponding residual service life label value; acquiring a hidden state corresponding to the optimal time dependency relationship; obtaining a residual service life prediction value of the aero-engine and a confidence interval thereof; and outputting the residual service life prediction value of the aero-engine in the optimal confidence interval. According to the CGPRN model, the RUL confidence interval is predicted by learning historical data, and the problem of uncertainty modeling is solved in a more organized mode. Meanwhile, the CGPRN is combined with a deep learning process, and inherent complex time sequence modes and dynamic behaviors of the aero-engine are effectively captured. The CGPRN improves the accuracy of the RUL prediction of the aero-engine, and provides transparent and interpretable uncertainty analysis.

Description

Technical Field

[0001] This invention relates to the fields of aircraft engine remaining service life prediction and artificial intelligence technology, specifically to a method for predicting aircraft engine remaining service life considering a confidence Gaussian process regression network. Background Technology

[0002] Prognostics and Health Management (PHM) for aero-engines plays a crucial role in improving engine availability, reducing maintenance costs, and optimizing maintenance event scheduling. By predicting potential failures, maintenance activities can be planned in advance, thereby reducing downtime. Predicting the Remaining Useful Life (RUL) of an aero-engine using monitoring data is central to effective prediction. With the development of sensor technology, modern aero-engines frequently and continuously collect sensor data, forming time-series data with inherent temporal characteristics. Therefore, RUL prediction is characterized as a time-series regression problem capable of capturing operational and performance characteristics.

[0003] Current methods for predicting RUL (Relative Limits) fall into two main categories: model-based and learning-based methods. Model-based methods include physical and statistical methods. Physical methods involve complex physical modeling, which limits their feasibility in engineering applications. In contrast, statistical methods, particularly Gaussian Process Regression (GPR), offer a more flexible approach to RUL prediction. GPR is a nonparametric Bayesian method that effectively models various degradation mechanisms and outputs uncertainty metrics, making it suitable for complex systems where physical models may fail. While GPR is robust in generating prediction intervals, using it alone presents resource consumption challenges when processing large-scale time-series data typical of industrial environments.

[0004] On the other hand, learning-based methods utilize machine learning and deep learning techniques, such as Long Short-Term Memory networks and Transformers, to improve prediction performance using historical data. These methods rely on supervised learning to achieve accurate results. During computation, the focus when handling high-dimensional data is on incorporating temporal continuity and the dynamic characteristics of system behavior. However, many machine learning methods focus only on single-point predictions without quantifying the uncertainty of the predictions, which limits their reliability in practical applications.

[0005] Therefore, there is a need to provide a method for predicting the remaining service life of aero-engines that considers a confidence Gaussian process regression network to solve the above problems. Summary of the Invention

[0006] This invention provides a method for predicting the remaining service life of aero-engines that considers a confidence Gaussian process regression network, in order to solve existing problems.

[0007] The present invention provides a method for predicting the remaining service life of an aero-engine considering a confidence Gaussian process regression network, which adopts the following technical solution, including: S1. Acquire multi-source monitoring signals of the aero-engine throughout its entire lifecycle and the corresponding remaining service life tag values ​​during aircraft operation. The multi-source monitoring signals include: fan inlet total temperature, low-pressure compressor outlet total temperature, high-pressure compressor outlet total temperature, low-pressure turbine outlet total temperature, fan inlet pressure, bypass total pressure, high-pressure compressor outlet total pressure, actual fan speed, actual core engine speed, engine pressure ratio, high-pressure compressor outlet static pressure, fuel flow ratio, fan corrected speed, core engine corrected speed, bypass ratio, combustor air-fuel ratio, bleed air enthalpy, fan speed, core engine converted speed, high-pressure turbine bleed air cooling flow rate, and low-pressure turbine bleed air cooling flow rate. S2. Perform data preprocessing on multi-source monitoring signals to form data packets; construct a time-series feature extraction model based on SLTM, input the data packets into the time-series feature extraction model, and use the sequence learning principle to design a time-series feature extraction method to capture the time dependency between multi-source monitoring signals, and output the hidden state corresponding to the optimal time dependency; S3. The CGPRN model is trained based on multi-source monitoring signals and corresponding remaining service life label values ​​to obtain a trained CGPRN model; the hidden state is input into the CGPRN model to obtain the predicted remaining service life value of the aero-engine and its confidence interval. S4. Obtain performance indicators using multiple performance evaluation methods and the remaining service life prediction value output by the CGPRN model; if each performance indicator meets its corresponding threshold, the remaining service life prediction value output by the CGPRN model is the remaining service life prediction value of the aero-engine with the best confidence interval; if any performance indicator does not meet its corresponding threshold, return to step S2 until each performance indicator meets its corresponding threshold, so that the remaining service life prediction value of the aero-engine with the best confidence interval output by the CGPRN model is obtained.

[0008] A further technical solution of the present invention includes multi-source monitoring signals including: fan inlet total temperature, low-pressure compressor outlet total temperature, high-pressure compressor outlet total temperature, low-pressure turbine outlet total temperature, fan inlet pressure, bypass total pressure, high-pressure compressor outlet total pressure, actual fan speed, actual core engine speed, engine pressure ratio, high-pressure compressor outlet static pressure, fuel flow ratio, fan corrected speed, core engine corrected speed, bypass ratio, combustor air-fuel ratio, bleed air enthalpy, fan speed, core engine converted speed, high-pressure turbine bleed air cooling flow rate, and low-pressure turbine bleed air cooling flow rate data generated by the engine under a single operating condition in a fault mode of high-pressure compressor performance degradation.

[0009] A further technical solution of the present invention includes the following step in setting the tag value corresponding to the multi-source monitoring signal: Set the remaining service life threshold to 125; A piecewise linear degradation model is developed based on the lifetime threshold, and the label value is obtained based on the piecewise linear degradation model:

[0010] in, This represents the actual remaining service life of the aircraft engine. This refers to the remaining service life label value of the aircraft engine.

[0011] A further technical solution of the present invention includes the following step: preprocessing multi-source monitoring signals to form data packets: The target multi-source monitoring signal is obtained by processing the multi-source monitoring signal based on the actual value of the remaining service life; After standardizing the multi-source monitoring signals of the target, the data is smoothed and then sampled using a sliding window to form data packets.

[0012] A further technical solution of the present invention involves processing the multi-source monitoring signals based on the true value of the remaining service life to obtain the target multi-source monitoring signal, as follows: The changing trends of multi-source monitoring signals are obtained, and multi-source monitoring signals with no obvious monotonicity in their changing trends are eliminated to obtain the target multi-source monitoring signal.

[0013] A further technical solution of the present invention involves the following steps: after standardizing the target multi-source monitoring signals, smoothing the data, and forming data packets using sliding window sampling: z-score normalization was used to process the target multi-source monitoring signals, and the processing results were smoothed using an exponential smoothing equation. Within a set number of cycles, the signal is sampled by a sliding window with a set step size. The smoothed multi-source monitoring signal is then standardized into a three-dimensional data packet, where the three dimensions are: batch size, time window size, and number of types of multi-source monitoring signals.

[0014] A further technical solution of the present invention involves inputting data packets into a target temporal feature extraction model, designing a temporal feature extraction method based on sequence learning principles to capture the temporal dependencies between multi-source monitoring signals, and outputting the hidden state corresponding to the optimal temporal dependency. The steps are as follows: The data packet is input into the target temporal feature extraction model and initialized and mapped to the initial hidden state; A target temporal feature extraction model is used to capture the temporal dependencies between multi-source monitoring signals; Set a loss function to optimize the feature extraction process of the target temporal feature extraction model; The hidden state is continuously calculated and adjusted based on the loss function value until the loss function value is minimized. At this point, the time dependency of the target time series feature extraction model output reaches the optimal value, and the hidden state corresponding to the optimal time dependency is the optimal hidden state.

[0015] A further technical solution of the present invention provides that the expression for the loss function is:

[0016] In the formula, This represents the value of the loss function; The tradeoff coefficient represents the relative weight of the two types of errors. Indicate batch dimensions; Indicates the sample index; Indicates the first i True RUL values ​​(labels) for each sample; Indicates the first i RUL predicted values ​​for each sample.

[0017] A further technical solution of the present invention includes performance evaluation methods such as root mean square error, normalized average width, and coverage width standard evaluation methods.

[0018] The beneficial effects of this invention are: This invention collects multi-source monitoring signals from aero-engines; preprocesses these signals to form data packets; employs a time-series feature extraction model to capture the time dependencies between the signals, obtaining the hidden state corresponding to the optimal time dependency; based on the hidden state, it uses the CGPRN model to predict the remaining service life (RUL) of the aero-engine and provides a confidence interval for the RUL prediction; it establishes multiple performance indicators to measure the prediction performance of the CGPRN model; and it adjusts the CGPRN model based on the performance indicator results to achieve the best RUL prediction with the optimal confidence interval. In essence, the CGPRN model learns from historical data to predict the RUL confidence interval, solving the uncertainty modeling problem in a more systematic way. Furthermore, CGPRN incorporates a deep learning process to effectively capture the inherent complex time-series patterns and dynamic behaviors of aero-engines. CGPRN improves the accuracy of aero-engine RUL prediction and provides transparent and interpretable uncertainty analysis. Therefore, this invention can quantify the confidence interval of the predicted value to better achieve the prediction of the remaining service life of aero-engines. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0020] Figure 1 This is a flowchart illustrating a method for predicting the remaining service life of an aero-engine that considers a confidence Gaussian process regression network, according to the present invention. Figure 2 This is a simplified diagram of a simulation structure for acquiring multi-source monitoring signals from an aero-engine, according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the processing steps of an aero-engine remaining service life prediction method considering a confidence Gaussian process regression network according to the present invention. Figure 4 This is a schematic diagram of the confidence interval for predicting the remaining service life of an aero-engine in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some 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 are within the scope of protection of the present invention.

[0022] An embodiment of the present invention, which considers a confidence Gaussian process regression network for predicting the remaining service life of an aero-engine, is as follows: Figure 1 As shown, it includes: S1. Obtain multi-source monitoring signals and corresponding remaining service life tag values.

[0023] Specifically, it acquires multi-source monitoring signals of the aero-engine throughout its entire lifecycle and the corresponding remaining service life tag values ​​during aircraft operation.

[0024] For example, in one specific embodiment, the effectiveness of the present invention is verified using the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset as an example. The specific parameters of the embodiment are: using the C-MAPSS dataset, and the aero-engine schematic diagram used is as follows... Figure 2 As shown, Figure 2 The simulated structure of the medium-pressure turbofan engine includes the engine fan, combustion chamber, fan shaft, core shaft, low-pressure turbine, low-pressure compressor, high-pressure compressor, high-pressure turbine, and nozzle. This embodiment uses the FD001 dataset, which contains multi-source monitoring signals generated under a single operating condition and in a single fault mode of high-pressure compressor performance degradation. In this embodiment, the multi-source monitoring signals include 21 types of monitoring signals, as shown in Table 1.

[0025] Table 1

[0026] For example, in one specific embodiment, the step of setting the tag value corresponding to the multi-source monitoring signal is as follows: Set the remaining service life threshold to 125; A piecewise linear degradation model is developed based on the lifetime threshold, and the label value is obtained based on the piecewise linear degradation model:

[0027] in, This represents the actual remaining service life of the aircraft engine. This refers to the remaining service life label value of the aircraft engine.

[0028] S2, the hidden state corresponding to obtaining the optimal time dependency; Specifically, data preprocessing is performed on multi-source monitoring signals to form data packets; a time-series feature extraction model is constructed based on LSTM, the data packets are input into the time-series feature extraction model, and a time-series feature extraction method is designed using the principle of sequence learning to capture the time dependency between multi-source monitoring signals, and the hidden state corresponding to the optimal time dependency is output.

[0029] For example, in one specific embodiment, the step of preprocessing multi-source monitoring signals to form data packets includes: processing the multi-source monitoring signals according to the true value of remaining lifetime to obtain the target multi-source monitoring signal; performing data standardization on the target multi-source monitoring signal, followed by data smoothing, and using sliding window sampling to form data packets. Specifically, in this embodiment, multi-source monitoring signal screening is performed first: (1): Read the changing trend of multi-source monitoring signals, remove monitoring signals with no obvious monotonicity in the changing trend, and retain 14 parameters: total temperature at low-pressure compressor outlet (T24), total temperature at high-pressure compressor outlet (T30), total temperature at low-pressure turbine outlet (T50), total pressure at high-pressure compressor outlet (P30), actual fan speed (Nf), actual core engine speed (Nc), static pressure at high-pressure compressor outlet (Ps30), fuel flow ratio (Phi), corrected fan speed (NRf), corrected core engine speed (NRc), bypass ratio (BPR), enthalpy of bleed air (htBleed), cooling flow rate of high-pressure turbine bleed air (W31), and cooling flow rate of low-pressure turbine bleed air (W32).

[0030] (2): Data standardization: for the first In the nth sample Data from one sensor Data normalization is performed according to the following formula:

[0031] In the formula, and Representing the first The average and standard deviation of the sensor readings. The normalized data show a clear trend within an acceptable range.

[0032] (3): Data smoothing: Formulate an exponential smoothing equation to preserve the trend of sensor signals and eliminate unnecessary features:

[0033] In the formula, for Time monitoring value The smoothed value. The smoothing parameter represents the weight of the current monitored value relative to the previous smoothed value.

[0034] (4): Data transformation: The monitoring signal is sampled through a sliding window with a time window of 25 cycles and a step size of 1. The monitoring signal is standardized into a three-dimensional data packet with three dimensions: batch size, time window size, and number of types of multi-source monitoring signals.

[0035] For example, in one specific embodiment, the steps of constructing a time-series feature extraction model based on LSTM and training the time-series feature extraction model based on multi-source monitoring signals and corresponding remaining lifetime label values ​​to obtain a trained target time-series feature extraction model are as follows: inputting data packets into the target time-series feature extraction model and randomly initializing them into initial hidden states; using the target time-series feature extraction model to capture the time dependencies between multi-source monitoring signals; setting a loss function to optimize the feature extraction process of the target time-series feature extraction model; continuously resolving and adjusting the hidden states according to the loss function value until the loss function value is minimized, at which point the time dependency output by the target time-series feature extraction model reaches its optimal state, that is, the hidden state corresponding to the optimal time dependency is the optimal hidden state.

[0036] Specifically, in this embodiment, a temporal feature extraction model is used to capture the temporal dependencies between multi-source monitoring signals. This model mimics the gating mechanism in long short-term memory networks. The specific solution process is as follows: (1) Data packet Initialized and mapped to the initial hidden state ; (2) The time dependence between multi-source monitoring signals is captured by the time series feature extraction model. The calculation process is as follows: (a) Update the Forgotten Gate: ; (b) Update the input gate: ; (c) Update cell state: , ; (d) Update the output gate: ; (e) Obtain the time dependencies, represented as hidden states: ; In the formula, For the Sigmoid function, It is the hyperbolic tangent function. For a moment Input, For a moment The hidden state. This represents the current input cell state. This represents the current state of the unit, used for storing memories. For the weight of the forget gate, For input gate weights, For unit weights, These are the output gate weights, and each weight is adaptively updated during the model training process. Correspondingly, For the forgetting gate offset, For input gate bias, For unit bias, This is used for output gate bias. It's important to note that the hidden states encapsulate the most crucial information from the time series, compressing the multi-source time series into a simplified form and serving as input to the CGPRN model.

[0037] (3): Set a loss function to optimize the feature extraction process to obtain the optimal hidden state:

[0038] In the formula, This represents the value of the loss function; The tradeoff coefficient represents the relative weight of the two types of errors. Indicate batch dimensions; Indicates the sample index; Indicates the first i True RUL values ​​(labels) for each sample; Indicates the first i RUL predicted values ​​for each sample.

[0039] (4): Based on the loss function value in step (3), repeat steps (2) and (3) to process the hidden state. Through continuous calculation and adjustment, the time dependency relationship after the time series feature extraction model is optimized, and then the hidden state corresponding to the optimal time dependency relationship can be obtained.

[0040] S3. Obtain the predicted remaining service life of the aero-engine and its confidence interval; Specifically, the CGPRN model is trained based on multi-source monitoring signals and corresponding remaining service life labels to obtain a trained CGPRN model; the hidden state is input into the CGPRN model to obtain the predicted remaining service life of the aero-engine and its confidence interval.

[0041] For example, in one specific embodiment, the step of inputting the hidden state into the CGPRN model to obtain the predicted value of the remaining service life of the aero-engine and its confidence interval is as follows: the CGPRN model receives the optimal hidden state output in step S2. This provides a probabilistic framework for predicting the RUL (Remaining Useful Life) of aero-engines, simultaneously outputting the predicted remaining useful life value and its confidence interval. The specific solution process includes the following steps: (1): For hidden states Its corresponding RUL ( ), For the first i The actual remaining useful life value (label); following the noise-included... The multivariate Gaussian distribution, i.e.:

[0042] In the formula, the mean function It is usually zero mean; It is the identity matrix; The standard deviation of the observed noise; Let be the covariance matrix and the covariance function. The quadratic exponential function is typically defined as follows:

[0043] In the formula, The magnitude of the covariance. The feature length scale; and For the first i and the j A hidden feature.

[0044] (2): For multi-source monitoring signals Gaussian regression was used to calculate the true value of remaining useful life (label). ) and the predicted remaining useful life ( The joint prior Gaussian distribution of )

[0045] In the formula, Based on new monitoring data The captured latent features. The posterior mean function and posterior variance function are calculated as follows:

[0046]

[0047] In the formula, It is the posterior mean function; It is the posterior variance function; It is the cross covariance vector; (3): Calculate the RUL prediction value for a given time, i.e., the point estimate:

[0048] In the formula, Predicted remaining useful life Point estimate.

[0049] (4): Based on Gaussian regression, calculate the confidence interval of the RUL prediction:

[0050] In the formula, , , This represents the standard Gaussian distribution quantile. Typically... If set to 0.05, then the corresponding value is... , .

[0051] Thus, we have obtained the following: Figure 4 The image shows the predicted remaining service life (RUL) and confidence intervals for the aero-engine. It can be seen that around the 85th cycle, the predicted RUL changes from near-level to a continuous decline; simultaneously, the bottom residual changes from predominantly positive to crossing zero and gradually becoming negative, indicating that the engine's health is entering a significant degradation phase. Meanwhile, the confidence band gradually tightens after the 140th cycle, thanks to increased information used for life prediction, making the RUL trajectory more certain and tending towards a conservative estimate. If the lower confidence bound is taken as the conservative decision-making benchmark, by the 150th cycle, the lower bound of the RUL has approached a low level, making it suitable as a candidate range for maintenance windows.

[0052] S4. The predicted remaining service life of the aero-engine with the best confidence interval. Specifically, performance indicators are obtained using multiple performance evaluation methods and the remaining service life prediction value output by the CGPRN model. If each performance indicator meets its corresponding threshold, the remaining service life prediction value output by the CGPRN model is the remaining service life prediction value of the aero-engine with the best confidence interval. If any performance indicator does not meet its corresponding threshold, the process returns to step S2 until each performance indicator meets its corresponding threshold, so that the remaining service life prediction value of the aero-engine with the best confidence interval output by the CGPRN model is obtained.

[0053] For example, in one specific embodiment, the steps for obtaining performance indicators using multiple performance indicator evaluation methods and the remaining useful life prediction value output by the CGPRN model are as follows: The root-mean-square error (RMSE), normalized averaged width (NAW), and coverage width-based criterion (CWC) are used to measure the RUL prediction values ​​obtained in step S3. RMSE represents the difference between the predicted and actual RUL values, while NAW and CWC are used to evaluate the model's interval prediction performance. The performance metric used, RMSE, is as follows:

[0054] In the formula, This is the predicted value for the remaining useful life; This represents the actual value of the remaining useful life. This indicates the number of engine units.

[0055]

[0056] In the formula, This indicates the range of values ​​for the target variable during the prediction period. The NAW value indicates the number of predicted points; a lower NAW value suggests that the prediction interval is more effective. Indicates the first j The upper bound of the interval for each prediction point (set as a 95% confidence upper bound in this invention). Indicates the first j The lower bound of the interval for each prediction point.

[0057]

[0058] In the formula, if but ,otherwise . This represents the penalty intensity coefficient, adjusted when the coverage falls below the nominal confidence level. To increase punishment; The CWC score indicates interval coverage; a lower CWC score suggests that interval prediction is more effective. The core concept of CWC is that if the coverage is below the nominal confidence level, the score should be higher regardless of the interval width; while when the coverage exceeds this level, NAW becomes the dominant factor.

[0059] Specifically, in this embodiment, if each performance indicator meets its corresponding threshold, the remaining service life prediction value output by the CGPRN model is the remaining service life prediction value of the aero-engine within the optimal confidence interval; if any performance indicator does not meet its corresponding threshold, the process returns to step S2 until each performance indicator meets its corresponding threshold, resulting in the remaining service life prediction value of the aero-engine within the optimal confidence interval output by the CGPRN model. In this embodiment, interval coverage... Set at 95%, target variable range during forecast period The validation set is set to the 5th–95th percentile difference. Therefore, the thresholds for each indicator are determined as follows: , , .

[0060] The following compares the prediction results of the prior art with those of the present invention: The method described in this embodiment is compared with shallow learning models and deep learning models. The shallow learning models are Support Vector Regression (SVR) and Sparse Autoencoder (SAE), while the deep learning models are Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM). The results in Table 2 show that the CGPRN model of this invention consistently outperforms existing methods and provides the most reliable remaining lifetime prediction range. This confirms that the CGPRN model provides highly reliable predictions with minimal bias.

[0061] Table 2

[0062] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining service life of an aero-engine considering a confidence Gaussian process regression network, characterized in that, include: S1. Acquire multi-source monitoring signals of the aero-engine throughout its entire lifecycle and the corresponding remaining service life tag values ​​during aircraft operation. The multi-source monitoring signals include: fan inlet total temperature, low-pressure compressor outlet total temperature, high-pressure compressor outlet total temperature, low-pressure turbine outlet total temperature, fan inlet pressure, bypass total pressure, high-pressure compressor outlet total pressure, actual fan speed, actual core engine speed, engine pressure ratio, high-pressure compressor outlet static pressure, fuel flow ratio, fan corrected speed, core engine corrected speed, bypass ratio, combustor air-fuel ratio, bleed air enthalpy, fan speed, core engine converted speed, high-pressure turbine bleed air cooling flow rate, and low-pressure turbine bleed air cooling flow rate. S2. Perform data preprocessing on multi-source monitoring signals to form data packets; construct a time-series feature extraction model based on LSTM, input the data packets into the time-series feature extraction model, and use the sequence learning principle to design a time-series feature extraction method to capture the time dependency between multi-source monitoring signals, and output the hidden state corresponding to the optimal time dependency; S3. The CGPRN model is trained based on multi-source monitoring signals and corresponding remaining service life label values ​​to obtain a trained CGPRN model; the hidden state is input into the CGPRN model to obtain the predicted remaining service life value of the aero-engine and its confidence interval. S4. Obtain performance indicators using multiple performance evaluation methods and the remaining service life prediction value output by the CGPRN model; if each performance indicator meets its corresponding threshold, the remaining service life prediction value output by the CGPRN model is the remaining service life prediction value of the aero-engine with the best confidence interval; if any performance indicator does not meet its corresponding threshold, return to step S2 until each performance indicator meets its corresponding threshold, so that the remaining service life prediction value of the aero-engine with the best confidence interval output by the CGPRN model is obtained.

2. The method for predicting the remaining service life of an aero-engine considering a confidence Gaussian process regression network according to claim 1, characterized in that, The multi-source monitoring signals include: under single operating conditions, in the fault mode of high-pressure compressor performance degradation, the following data are generated: fan inlet total temperature, low-pressure compressor outlet total temperature, high-pressure compressor outlet total temperature, low-pressure turbine outlet total temperature, fan inlet pressure, bypass total pressure, high-pressure compressor outlet total pressure, actual fan speed, actual core engine speed, engine pressure ratio, high-pressure compressor outlet static pressure, fuel flow ratio, fan corrected speed, core engine corrected speed, bypass ratio, combustor air-fuel ratio, bleed air enthalpy, fan speed, core engine converted speed, high-pressure turbine bleed air cooling flow rate, and low-pressure turbine bleed air cooling flow rate.

3. The method for predicting the remaining service life of an aero-engine considering a confidence Gaussian process regression network according to claim 1, characterized in that, The steps to set the tag values ​​corresponding to multi-source monitoring signals are as follows: Set the remaining service life threshold to 125; A piecewise linear degradation model is developed based on the lifetime threshold, and the label value is obtained based on the piecewise linear degradation model: in, This represents the actual remaining service life of the aircraft engine. This refers to the remaining service life label value of the aircraft engine.

4. The method for predicting the remaining service life of an aero-engine considering a confidence Gaussian process regression network according to claim 1, characterized in that, The steps for preprocessing multi-source monitoring signals to form data packets are as follows: The target multi-source monitoring signal is obtained by processing the multi-source monitoring signal based on the actual value of the remaining service life; After standardizing the multi-source monitoring signals of the target, the data is smoothed and then sampled using a sliding window to form data packets.

5. The method for predicting the remaining service life of an aero-engine considering a confidence Gaussian process regression network according to claim 4, characterized in that, The steps for processing multi-source monitoring signals based on the true value of remaining service life to obtain the target multi-source monitoring signal are as follows: The changing trends of multi-source monitoring signals are obtained, and multi-source monitoring signals with no obvious monotonicity in their changing trends are eliminated to obtain the target multi-source monitoring signal.

6. The method for predicting the remaining service life of an aero-engine considering a confidence Gaussian process regression network according to claim 4, characterized in that, The steps for standardizing and smoothing the multi-source monitoring signals of the target, and then forming data packets using sliding window sampling are as follows: z-score normalization was used to process the target multi-source monitoring signals, and the processing results were smoothed using an exponential smoothing equation. Within a set number of cycles, the signal is sampled by a sliding window with a set step size. The smoothed multi-source monitoring signal is then standardized into a three-dimensional data packet, where the three dimensions are: batch size, time window size, and number of types of multi-source monitoring signals.

7. The method for predicting the remaining service life of an aero-engine considering a confidence Gaussian process regression network according to claim 1, characterized in that, The steps for inputting data packets into the target temporal feature extraction model, designing a temporal feature extraction method based on sequence learning principles to capture the temporal dependencies between multi-source monitoring signals, and outputting the hidden state corresponding to the optimal temporal dependency are as follows: The data packet is input into the target temporal feature extraction model and randomly initialized and mapped to the initial hidden state; A target temporal feature extraction model is used to capture the temporal dependencies between multi-source monitoring signals; Set a loss function to optimize the feature extraction process of the target temporal feature extraction model; The hidden state is continuously calculated and adjusted based on the loss function value until the loss function value is minimized. At this point, the time dependency of the target time series feature extraction model output reaches the optimal value, and the hidden state corresponding to the optimal time dependency is the optimal hidden state.

8. The method for predicting the remaining service life of an aero-engine considering a confidence Gaussian process regression network according to claim 7, characterized in that, The expression for the loss function is: In the formula, This represents the value of the loss function; The tradeoff coefficient represents the relative weight of the two types of errors. Indicate batch dimensions; Indicates the sample index; Indicates the first i True RUL values ​​(labels) for each sample; Indicates the first i RUL predicted values ​​for each sample.

9. The method for predicting the remaining service life of an aero-engine considering a confidence Gaussian process regression network according to claim 1, characterized in that, Performance evaluation methods include: root mean square error, normalized average width, and coverage width standard evaluation methods.

Images