CN-121997125-A - Large-bypass-ratio turbofan engine fleet all-condition anomaly monitoring method based on self-attention reconstruction network

CN121997125ACN 121997125 ACN121997125 ACN 121997125ACN-121997125-A

Abstract

The invention discloses a large bypass ratio turbofan engine fleet all-condition anomaly monitoring method based on a self-attention reconstruction network. The method comprises the steps of carrying out sliding window sampling on multivariable flight parameter time sequence data, adopting mean standard deviation normalization, utilizing a reconstruction network comprising a self-attention encoder and a decoder to learn a health baseline, outputting a target parameter reconstruction value and taking a residual error between an actual measurement value and the reconstruction value as an abnormal measurement. Aiming at the problem that the distribution difference of the residual error is obvious between different flight phases and different state intervals, the method estimates a bit-dividing threshold value based on the three-dimensional conditions of an engine individual, the flight phases and the state boxes of the healthy residual error, sets a level rollback mechanism to cope with strange individuals of different fleets or small sample condition intervals, combines a continuous duration consistency rule during online monitoring, and triggers an alarm only when the overrun lasts for not less than K sampling periods. The invention can realize steady abnormal monitoring under the conditions of full working condition and individual difference, reduce false alarm and promote the engineering deployment.

Inventors

LU FENG
JIANG WENLONG
ZHOU XIN
Zou Zelong
LIU YI

Assignees

南京航空航天大学

Dates

Publication Date: 20260508
Application Date: 20251231

Claims (6)

1. The large bypass ratio turbofan engine fleet all-condition anomaly monitoring method based on the self-attention reconstruction network is characterized by comprising the following steps of: Step A) obtaining multivariable time sequence operation data of an aircraft engine fleet, at least comprising exhaust temperature EGT and an individual identifier e/flight phase identifier p related parameter thereof, and preprocessing the multivariable time sequence operation data. And B), training Informer-AE reconstruction network on a training set only comprising healthy/normal fleet operation data to establish an all-condition baseline model, and calculating residual errors according to actual measurement values of target parameters and the reconstruction values of the baseline model to establish a healthy residual error library. Step C) selecting at least one continuous state feature s on the healthy residual library and performing equal frequency binning on the continuous state feature s to obtain a state bucket b, and estimating a conditional quantile threshold value at the quantile level for each three-dimensional condition group g= (e, p, b). And D) selecting a bit division threshold corresponding to the working condition information from a threshold library based on the working condition information of the sample to be monitored as a judgment threshold, and executing level rollback to obtain an available judgment threshold when the threshold library is missing or does not meet a preset stability condition. And E) comparing the reconstructed residual error of the sample to be monitored with the judging threshold value, and outputting an alarm result in combination with a continuous duration consistency rule, wherein the alarm output is triggered only when the residual error meets the threshold exceeding condition and lasts for not less than K sampling periods.
2. The method for monitoring the abnormal conditions of the large bypass ratio turbofan engine fleet based on the self-attention reconstruction network according to claim 1, wherein the preprocessing of the multivariate time series data in the step A) is specifically as follows: step A1) the flight phase identifier p is determined by adopting a phase code recorded by QAR, and the corresponding relation between the phase code and the flight phase satisfies the following table 1: TABLE 1 flight phase definition Step A2) selecting engine related measurement parameters based on engineering experience and a common sensor combination, and carrying out Pearson correlation analysis on the engine related measurement parameters, wherein the specific extracted flight parameters and meanings thereof are shown in a table 2: Table 2 QAR parameter extraction control Table Step A3) calculating the mean value and standard deviation of each variable based on the health training data, and adopting the mean value and standard deviation to carry out consistent standardized processing on the training, verification and online monitoring stage data, wherein the specific formula is as follows: wherein mu j 、σ j is the mean value and standard deviation of the characteristic channel j on the training set respectively.
3. The method for monitoring the all-condition anomalies of the large-bypass-ratio turbofan engine fleet based on the self-attention reconstruction network according to claim 1, wherein the specific steps of establishing the baseline model and the healthy residual library in the step B) are as follows: step B1) firstly, carrying out linear mapping on each time step to obtain three groups of vectors of query, key and value, then obtaining the correlation weight between positions through scaling dot product and normalization, and under the combined action of multi-head parallel and sparse selection mechanism, focusing information contributes to larger dependency relationship so as to improve efficiency and robustness, wherein the basic calculation formula is as follows: Q=E 0 W Q ,K＝E 0 W K ,V＝E 0 W V Where W Q 、W K 、W V is a trainable projection matrix, then for any head attention, calculate weights in the form of scaled dot products and weight sum the value vectors: To enhance the diversity and directionality of feature expressions, a multi-head attention architecture comprising h attention heads can be employed, each using independent parameter sets Splicing the outputs of each head according to channels and linearly mapping the outputs back to the dimension of the model: M(E 0 )＝Concat(Attn(Q (1) ,K (1) ,V (1) ),...,Attn(Q (h) ,K (h) ,V (h) ))W O In the formula, W O is output projection, so that the multi-head structure enables the model to model the dependency relationship at the same time in different angles and different time scales, and the representation capability of the complex sequence is improved. Based on the standard multi-head attention, a sparse idea is introduced for engineering constraint of adapting to the time sequence data of a long-sequence engine, namely, scoring vector of an ith query q l is introduced The information contribution degree is defined as follows: Selecting the top n query index sets with the largest contribution degree Performing complete attention only on queries within S and forming a sparse weight matrix The method comprises the following steps: Z S ＝A S V, Where ContextFill (·) represents a sparse reconstruction operation that fills/interpolates the locations of the unselected queries with neighboring or global statistics. The full connection attention of O (L 2 ) can be replaced by the expected complexity of similar O (nL) through sparsity selection, and simultaneously the sensitivity to remote dependence and the cross-working condition alignment capability are reserved, so that the method is very suitable for overlength and heteroscedastic flight parameter data. Step B2) through the above steps, the cross-position timing information has been extracted in the time dimension, but the outputs are still arranged in time steps, and each channel needs to be subjected to average pooling in the dimension of length L, so that the sequence level representation is compressed into a latent variable of a fixed dimension As input to the decoder, the basic calculation formula is: in the formula, Representing that each channel is independently averaged over the time dimension t, the result remains the channel dimension unchanged. The decoder is then responsible for mapping the latent variable z back to the target in the observation space, and reconstructing the target variable, wherein the basic calculation formula is as follows: in the formula, the function is activated Select GELU, h 1 ,h 2 are respectively represented by the first hidden layer and the second hidden layer, and finally output Representing an estimate of the target variable at the end of the corresponding window. The parameters of the hidden layer are jointly optimized, so that the encoding end refines the representation of the health domain in the latent space, and the decoding end non-linearly maps the representation of the health domain back to the target parameters, thereby improving the fitting degree of the reconstruction value and the true value on the normal sample, and simultaneously, establishing a health residual library by utilizing the reconstruction residual on the normal sample, thereby being convenient for subsequent threshold learning and online monitoring.
4. The method for monitoring the abnormal conditions of the large bypass ratio turbofan engine fleet according to claim 1, wherein the specific steps of estimating the conditional quantile threshold in the step C) are as follows: Step C1) to characterize the difference of normal fluctuation scale under the same condition, the equal frequency bin is executed on the health library for the single continuous state feature. And dividing the value taking domain into k intervals according to sample dividing by setting an equal-frequency dividing box to ensure that the sample quantity in each barrel is approximately equal, thereby ensuring that the estimation of each barrel dividing has similar variance level. Score box boundary { A 0 ＜A 1 ＜…＜A k }, then the state-to-bucket mapping is: if A J-1 <s t ≤A J then b t =j,j=1,...,k After the completion of the equal frequency grouping of the status dimension, the effects of individual differences and flight phases need to be further processed. Thus, in the threshold learning phase, the health library residuals are aggregated according to the three-dimensional conditions of the engine individual e, the flight phase p, and the state bucket b, and the high ranking threshold is estimated on each condition set. Let group tag g= (e, p, b) with health residual set of Given a quantile level α∈ (0, 1), the conditional threshold is defined as:
5. The method for monitoring the abnormal conditions of the large bypass ratio turbofan engine fleet based on the self-attention reconstruction network according to claim 1 is characterized in that the step D) comprises the following steps of: In step D1), in order to ensure that the anomaly monitoring system still has good accuracy and stability when facing to samples of a strange condition group, a level rollback strategy is introduced, and the system can automatically rollback according to other conditions such as a flight stage and a working condition state, so as to obtain a health threshold under a similar model or state, and a threshold selection rule is given in a piecewise function form: After the hierarchical rollback mechanism determines the condition threshold value which should be adopted at each moment, single-point detection is carried out on the current absolute residual error by using the group threshold value in discrimination: exceed t ＝1(R t ＞T * (e,p t ,b i )) Where 1 (·) is the indicator function.
6. The method for monitoring the abnormal conditions of the large bypass ratio turbofan engine fleet based on the self-attention reconstruction network according to claim 1, wherein the continuous duration consistency rule in the step E) comprises the following specific steps: Step E1) on a threshold selection system with a three-dimensional condition matching and rollback mechanism, an alarm is triggered only when the overrun lasts for not less than K sampling periods, K can be calibrated according to the sampling rate and the acceptable response delay, and the balance of robustness and sensitivity is achieved. Let the sampling period be Δt The alert criteria are:

Description

Large-bypass-ratio turbofan engine fleet all-condition anomaly monitoring method based on self-attention reconstruction network Technical Field The invention relates to the technical field of aeroengine health management and state monitoring, in particular to a large bypass ratio turbofan engine fleet all-condition anomaly monitoring method based on a self-attention reconstruction network. Background The engineering goal of the anomaly monitoring of the aero-engine is to identify the anomaly state as early as possible under the conditions of wide-envelope running and complex tasks so as to facilitate the subsequent processing. Along with long-term accumulation of flight parameter recorded data, the multi-parameter long sequence provides conditions for establishing a data-driven healthy baseline, and related research also gradually develops from experience threshold overrun to feature space detection based on machine learning, and then directly learns a normal mode from a multivariable time sequence by utilizing a deep network. Particularly, under the actual condition of scarce labels, an unsupervised idea of learning health characterization by taking normal data as a main part and forming an alarm basis by the deviation degree is widely adopted, wherein the reconstruction method can deviate by residual quantification and is convenient to link with an engineering threshold system, so that the method has practicability. However, the anomaly monitoring should not be limited to a single stable condition, but should work stably under all conditions and have cross-station applicability. If a base line is established on a single or narrow working condition and a unified threshold is adopted, the alarm is easily mismatched along with the change of the working condition, and the false alarm is increased or the missing alarm is increased. Meanwhile, the supervision strategy relying on fault labels is limited by sample scarcity and incomplete category in reality, and long-term stable deployment is difficult to support. In addition, if the threshold lacks conditional adaptation capability, the model tends to significantly degrade during unfamiliar or low sample operating conditions. Disclosure of Invention Aiming at the problems, the invention provides an unsupervised anomaly monitoring method for a full-working condition scene, which takes the sparse self-attention of Informer as a core to construct a reconstruction type self-coding network, learns healthy baselines of multi-parameter long sequences on healthy frame historical data, takes reconstruction residual errors as unified anomaly measurement, does not carry out hard regulation on residual error distribution at a detection end, but learns conditional bit-dividing thresholds based on healthy domain statistics, so that the thresholds can be adaptively changed along with engine individuals, flight phases and state boxes, and simultaneously introduces a level rollback mechanism to automatically return to threshold levels of similar conditions when strange individuals or small sample condition intervals in different fleets are met, thereby avoiding incapacity of discrimination or improper threshold selection due to data sparsity. The method is suitable for anomaly monitoring under different working conditions, and improves engineering usability and cross-condition generalization capability. The technical scheme adopted by the invention is as follows: The large bypass ratio turbofan engine fleet all-condition anomaly monitoring method based on the self-attention reconstruction network is characterized by comprising the following steps of: Step A) obtaining multivariable time sequence operation data of the aeroengine, at least comprising exhaust temperature EGT and an individual identifier e/flight phase identifier p related parameter thereof, and preprocessing the multivariable time sequence operation data. And B), training Informer-AE reconstruction network on a training set only comprising healthy/normal flight data to establish an all-condition baseline model, and calculating residual errors according to actual measured values of target parameters and the reconstruction values of the baseline model to establish a healthy residual error library. Step C) selecting at least one continuous state feature s on the healthy residual library and performing equal frequency binning on the continuous state feature s to obtain a state bucket b, and estimating a conditional quantile threshold value at the quantile level for each three-dimensional condition group g= (e, p, b). And D) selecting a bit division threshold corresponding to the working condition information from a threshold library based on the working condition information of the sample to be monitored as a judgment threshold, and executing level rollback to obtain an available judgment threshold when the threshold library is missing or does not meet a preset stability condition. And E) comparing the reconstructed re