Large-bypass-ratio turbofan engine fleet full-working-condition abnormity monitoring method based on self-attention reconstruction network

Abstract

The invention discloses a large-bypass-ratio turbofan engine fleet full-working-condition abnormity monitoring method based on a self-attention reconstruction network. According to the method, sliding window sampling is carried out on multivariable flight parameter time sequence data, and normalization is carried out by adopting a mean value standard deviation; and learning the health baseline by using a reconstruction network comprising a self-attention encoder and a decoder, outputting a target parameter reconstruction value, and taking a residual error between a measured value and the reconstruction value as an abnormal measure. Aiming at the problem that the distribution difference of residual errors in different flight stages and different state intervals is obvious, a quantile threshold value is estimated under the three-dimensional conditions of an engine individual, the flight stages and a state sub-box based on healthy residual errors, and a hierarchical backspacing mechanism is set to cope with strange individual or small sample condition intervals of different fleets; and during online monitoring, in combination with a continuous duration consistency rule, an alarm is triggered only when the overrun lasts no less than K sampling periods. According to the invention, robust anomaly monitoring can be realized under all working conditions and individual difference conditions, false alarms are reduced, and the engineering deployability is improved.

Description

Technical Field

[0001] This invention relates to the field of aero-engine health management and condition monitoring technology, and in particular to a method for full-condition anomaly monitoring of a high-bypass turbofan engine fleet based on a self-attention reconstruction network. Background Technology

[0002] The engineering goal of aero-engine anomaly monitoring is to identify abnormal states as early as possible under wide envelope operation and complex mission conditions to facilitate subsequent processing. With the long-term accumulation of flight parameter recording data, long multi-parameter sequences have provided the conditions for establishing data-driven health baselines. Related research has gradually evolved from empirical threshold exceedances to feature space detection based on machine learning, and then to using deep networks to directly learn normal patterns from multivariate time series. Especially under practical conditions of scarce labels, the unsupervised approach of learning health representations primarily from normal data and then forming alarm criteria based on the degree of deviation has been widely adopted. Among these, the reconstruction method can quantify deviations with residuals and is more practical due to its ease of integration with engineering threshold systems.

[0003] However, anomaly monitoring should not be limited to a single stable operating condition, but should operate stably under all operating conditions and be applicable across different engine configurations. Engine data changes drastically; flight phase transitions, changes in control inputs, and fluctuations in ambient temperature and air pressure can all alter the scale and form of normal fluctuations. If a baseline is established on a single or narrow operating condition and a uniform threshold is used, alarms can easily become mismatched with changes in operating conditions, leading to an increase in false alarms or false negatives. At the same time, supervised strategies relying on fault labels are limited in reality by the scarcity of samples and incomplete categories, making it difficult to support long-term stable deployment. In addition, if the threshold lacks conditional adaptation capabilities, the model's performance often degrades significantly in unfamiliar engine configurations or operating condition ranges with few samples. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes an unsupervised anomaly monitoring method for all operating conditions. It constructs a reconstructive autoencoder network based on Informer's sparse self-attention mechanism, learning a multi-parameter long-sequence health baseline from historical health flight data, and using the reconstructed residuals as a unified anomaly metric. At the detection end, instead of rigidly defining the residual distribution, it uses conditional quantile thresholds learned statistically from the health domain, allowing the thresholds to adaptively change with individual engines, flight phases, and state bins. Simultaneously, a hierarchical backoff mechanism is introduced, automatically reverting to a threshold level with similar conditions when encountering unfamiliar individuals or small sample condition intervals within different fleets, avoiding inability to identify anomalies due to data sparsity or inappropriate threshold selection. This method is applicable to anomaly monitoring under various operating conditions, improving engineering usability and cross-condition generalization ability.

[0005] Technical solution: To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0006] A method for full-condition anomaly monitoring of a high-bypass turbofan engine fleet based on a self-attention reconstruction network, characterized by the following steps:

[0007] Step A) Obtain multivariable time-series operating data of the aero-engine, including at least the exhaust temperature EGT and its individual identifier e / flight phase identifier p related parameters, and preprocess the multivariable time-series operating data.

[0008] Step B) Train the Informer-AE reconstruction network on a training set containing only healthy / normal flight data to establish a full-condition baseline model, and calculate the residuals based on the measured values ​​of the target parameters and the reconstructed values ​​of the baseline model to establish a healthy residual library.

[0009] Step C) Select at least one continuous state feature s in the health residual library and perform equal-frequency binning on it to obtain state bucket b. Estimate the conditional quantile threshold at the quantile level for each three-dimensional condition set g = (e, p, b).

[0010] Step D) Based on the operating condition information of the sample to be monitored, select the quantile threshold corresponding to the operating condition information from the threshold library as the judgment threshold, and perform hierarchical backtracking to obtain an available judgment threshold when the threshold library is missing or does not meet the preset stability conditions.

[0011] Step E) Compare the reconstructed residual of the sample to be monitored with the judgment threshold, and output an alarm result in combination with the continuous duration consistency rule, wherein the alarm output is triggered only when the residual meets the threshold condition and lasts for no less than K sampling periods.

[0012] Furthermore, the specific steps for preprocessing the multivariate time series data in step A) are as follows:

[0013] Step A1) The flight phase identifier p is determined using the phase code recorded in QAR, and the correspondence between the phase code and the flight phase satisfies the conditions shown in Table 1:

[0014] Table 1 Definition of Flight Phases

[0015]

[0016]

[0017] Step A2) Based on engineering experience and commonly used sensor combinations, select relevant engine measurement parameters and perform Pearson correlation analysis on them. The specific extracted flight parameters and their meanings are shown in Table 2:

[0018] Table 2 QAR Parameter Extraction Comparison Table

[0019]

[0020] Step A3) Calculate the mean and standard deviation of each variable based on the health training data, and use the mean and standard deviation to perform consistent standardization processing on the data from the training, validation, and online monitoring phases. The specific formula is as follows:

[0021]

[0022] In the formula, μ j σ j denoted as the mean and standard deviation of feature channel j on the training set, respectively.

[0023] Furthermore, the specific steps for establishing the baseline model and health residual library in step B) are as follows:

[0024] Step B1) First, perform linear mapping at each time step to obtain three sets of vectors: query, key, and value. Then, obtain the relevance weights between positions through scaling dot product and normalization. Under the combined effect of multi-head parallelism and sparse selection mechanism, focus on the dependencies that contribute more information to improve efficiency and robustness. The basic calculation formula is as follows:

[0025] Q = E0W Q K = E0W K V = E0W V

[0026] In the formula, W Q W K W V Given a trainable projection matrix, the weights for any attention endpoint are then calculated using a scaled dot product and the value vectors are summed with weights:

[0027]

[0028] To enhance the diversity and directionality of feature representation, a multi-head attention architecture with h attention heads can be adopted, each using an independent set of parameters. The outputs from each head are concatenated by channel and linearly mapped back to the model dimension:

[0029] M(E0) = Concat(Attn(Q) (1) K (1) v (1) ), ..., Attn(Q (h) K (h) V (h) ))w O

[0030] In the formula, W O The output projection enables the multi-head structure to model dependencies simultaneously from different angles and at different time scales, thus improving the ability to represent complex sequences.

[0031] Based on the aforementioned standard multi-head attention, and to adapt to the engineering constraints of long-sequence engine time-series data, a sparsity approach is introduced: for the i-th query q i Scoring vector Its information contribution is defined as:

[0032]

[0033] Based on this, select the top n query index sets with the greatest contribution. Full attention is performed only on queries within S, and a sparse weight matrix is ​​formed. get:

[0034] Z S =A S V,

[0035] In the formula, ContextFill(·) represents a sparse reconstruction operation that fills / interpolates the positions of unselected queries using neighbor or global statistics. Sparsity selection can replace the expected complexity of O(L) with approximately O(nL). 2 Fully connected attention, while retaining sensitivity to remote dependencies and cross-condition alignment capabilities, is ideal for ultra-long, heteroscedastic flight parameter data.

[0036] Step B2) After the above steps, cross-position temporal information has been extracted in the time dimension, but the output is still arranged according to time steps. It is necessary to perform average pooling on each channel in the length L dimension to compress the sequence-level representation into a latent variable of a fixed dimension. As the input to the decoder, its basic calculation formula is:

[0037]

[0038] In the formula, This means that the mean is calculated independently for each channel along the time dimension t, with the result maintaining the channel dimension unchanged. The decoder then maps this latent variable z back to the target quantity in the observation space, reconstructing the target variable. The basic calculation formula is:

[0039]

[0040] In the formula, the activation function Choosing GELU, h1 and h2 are the first and second hidden layer representations respectively, and the final output is... This represents the target variable estimate at the end of the corresponding window. By jointly optimizing the parameters of the hidden layer, the encoder extracts the health domain representation in the latent space, and the decoder nonlinearly maps it back to the target parameters, thereby improving the fit between the reconstructed values ​​and the true values ​​on normal samples. At the same time, a health residual library is established using the reconstruction residuals on normal samples, which facilitates subsequent threshold learning and online monitoring.

[0041] The specific steps for estimating the conditional quantile threshold in step C) are as follows:

[0042] Step C1) To characterize the differences in the scale of normal fluctuations under the same conditions, equal-frequency binning is performed on a healthy database for a single continuous state feature. Assume that equal-frequency binning divides the value range into k intervals according to the sample quantiles, making the sample size in each bin approximately equal, thus ensuring that the quantile estimates of each bin have similar variance levels. The bin boundaries are {A0 < A1 < ... < A...}. k}, then the mapping from state to bucket is:

[0043]

[0044] After completing the equal-frequency grouping of the state dimension, further processing of the influence of individual differences and flight phases is required. Therefore, in the threshold learning phase, the health library residuals are aggregated according to the three-dimensional conditions of engine individual e, flight phase p, and state bucket b, and the high quantile threshold is estimated on each condition group. Let the group label g = (e, p, b), and its health residual set be... Given a quantile level α∈(0,1), the conditional threshold is defined as:

[0045]

[0046] The specific steps of the hierarchical rollback mechanism in step D) are as follows:

[0047] Step D1) To ensure the anomaly monitoring system maintains good accuracy and stability when facing unfamiliar sample groups, a hierarchical backoff strategy is introduced. This strategy automatically backoffs based on other conditions such as flight phase and operating status to obtain health thresholds for similar aircraft models or states. The threshold selection rules are given below in piecewise function form:

[0048]

[0049] After the hierarchical backoff mechanism determines the conditional thresholds to be used at each time step, a single-point detection is performed on the current absolute residual using the group thresholds during the discrimination process:

[0050]

[0051] In the formula, 1(·) is an indicator function.

[0052] The specific steps for the continuous duration consistency rule in step E) are as follows:

[0053] Step E1) In a threshold selection system with existing three-dimensional condition matching and backoff mechanisms, an alarm is triggered only when the exceedance persists for at least K sampling periods. K can be calibrated based on the sampling rate and acceptable response delay to achieve a balance between robustness and sensitivity. Let the sampling period be Δt and... The alarm criteria are as follows:

[0054]

[0055] Beneficial Effects: The present invention provides a method for monitoring anomalies in a high-bypass turbofan engine fleet under all operating conditions based on a self-attention reconstruction network. Compared with existing technologies, the above technical solution has the following technical advantages:

[0056] (1) The applicable operating conditions for abnormal monitoring have been increased. By using a self-attention reconstruction network to learn a full-condition health baseline mapping on health data, it can cover multiple flight phases such as takeoff, climb, cruise, and descent, reducing the dependence on manufacturer benchmark models and single-condition baselines.

[0057] (2) Reduced data labeling requirements. By using conditional quantile thresholds, the health residual statistics are transformed into comparable anomaly discrimination rules, which are more robust to skewed and heteroscedastic residual distributions.

[0058] (3) Improved the generalization ability of the monitoring method across individuals. By using a hierarchical back-off threshold chain, the available threshold can still be output even when there are insufficient samples in the subdivided groups, thus improving the generalization ability across individuals and working conditions. Attached Figure Description

[0059] Figure 1 This is a structural diagram of the high bypass ratio turbofan engine fleet full-condition anomaly monitoring method based on self-attention reconstruction network according to the present invention.

[0060] Figure 2 It is a heatmap showing the correlation between sensor parameters.

[0061] Figure 3 This is a schematic diagram of the EGT reconstruction network based on Informer-AE.

[0062] Figure 4 This is a schematic diagram of a sparse self-attention mechanism.

[0063] Figure 5 These are loss reduction curves for different methods.

[0064] Figure 6 This is a schematic diagram illustrating the effect of the healthy flight reconstruction.

[0065] Figure 7 This is a diagram illustrating the effect of abnormal flight monitoring. Detailed Implementation

[0066] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0067] The invention describes a method for full-condition anomaly monitoring of a high-bypass turbofan engine fleet based on a self-attention reconstruction network, characterized by comprising the following steps:

[0068] Step A) Obtain multivariable time-series operating data of the aero-engine, including at least the exhaust temperature EGT and its individual identifier e / flight phase identifier p related parameters, and preprocess the multivariable time-series operating data.

[0069] Step B) Train the Informer-AE reconstruction network on a training set containing only healthy / normal flight data to establish a full-condition baseline model, and calculate the residuals based on the measured values ​​of the target parameters and the reconstructed values ​​of the baseline model to establish a healthy residual library.

[0070] Step C) Select at least one continuous state feature s in the health residual library and perform equal-frequency binning on it to obtain state bucket b. Estimate the conditional quantile threshold at the quantile level for each three-dimensional condition set g = (e, p, b).

[0071] Step D) Based on the operating condition information of the sample to be monitored, select the quantile threshold corresponding to the operating condition information from the threshold library as the judgment threshold, and perform hierarchical backtracking to obtain an available judgment threshold when the threshold library is missing or does not meet the preset stability conditions.

[0072] Step E) Compare the reconstructed residual of the sample to be monitored with the judgment threshold, and output an alarm result in combination with the continuous duration consistency rule, wherein the alarm output is triggered only when the residual meets the threshold condition and lasts for no less than K sampling periods.

[0073] Furthermore, the specific steps for preprocessing the multivariate time series data in step A) are as follows:

[0074] Step A1) The flight phase identifier p is determined using the phase code recorded in QAR, and the correspondence between the phase code and the flight phase satisfies the conditions shown in Table 1:

[0075] Table 1 Definition of Flight Phases

[0076]

[0077] Step A2) Based on engineering experience and commonly used sensor combinations, select relevant engine measurement parameters and perform Pearson correlation analysis on them. The correlation analysis results are as follows: Figure 2 As shown in Table 2, the specific extracted flight parameters and their meanings are as follows:

[0078] Table 2 QAR Parameter Extraction Comparison Table

[0079]

[0080]

[0081] Step A3) Calculate the mean and standard deviation of each variable based on the health training data, and use the mean and standard deviation to perform consistent standardization processing on the data from the training, validation, and online monitoring phases. The specific formula is as follows:

[0082]

[0083] In the formula, μ j σ j denoted as the mean and standard deviation of feature channel j on the training set, respectively.

[0084] Furthermore, in step B), the network structure is reconstructed as follows: Figure 3 As shown, the specific steps for establishing the baseline model and health residual library are as follows:

[0085] Step B1) First, perform a linear mapping at each time step to obtain three sets of vectors: query, key, and value. Then, obtain the relevance weights between positions through scaling dot products and normalization. Under the combined effect of multi-head parallelism and sparse selection mechanisms, focus on dependencies that contribute more information to improve efficiency and robustness. A schematic diagram of the sparse self-attention mechanism is shown below. Figure 4 As shown, its basic calculation formula is:

[0086] Q = E0W Q K = E0W K V = E0W V

[0087] In the formula, W Q W K W V Given a trainable projection matrix, the weights for any attention endpoint are then calculated using a scaled dot product and the value vectors are summed with weights:

[0088]

[0089] To enhance the diversity and directionality of feature representation, a multi-head attention architecture with h attention heads can be adopted, each using an independent set of parameters. The outputs from each head are concatenated by channel and linearly mapped back to the model dimension:

[0090] M(E0) = Concat(Attn(Q) (1) K (1) V (1) ), ..., Attn(Q (h)K (h) V (h) ))W O

[0091] In the formula, W O The output projection enables the multi-head structure to model dependencies simultaneously from different angles and at different time scales, thus improving the ability to represent complex sequences.

[0092] Based on the aforementioned standard multi-head attention, and to adapt to the engineering constraints of long-sequence engine time-series data, a sparsity approach is introduced: for the i-th query q i Scoring vector Its information contribution is defined as:

[0093]

[0094] Based on this, select the top n query index sets with the greatest contribution. Full attention is performed only on queries within S, and a sparse weight matrix is ​​formed. get:

[0095] Z S =A S V,

[0096] In the formula, ContextFill(·) represents a sparse reconstruction operation that fills / interpolates the positions of unselected queries using neighbor or global statistics. Sparsity selection can replace the expected complexity of O(L) with approximately O(nL). 2 Fully connected attention, while retaining sensitivity to remote dependencies and cross-condition alignment capabilities, is ideal for ultra-long, heteroscedastic flight parameter data.

[0097] Step B2) After the above steps, cross-position temporal information has been extracted in the time dimension, but the output is still arranged according to time steps. It is necessary to perform average pooling on each channel in the length L dimension to compress the sequence-level representation into a latent variable of a fixed dimension. As the input to the decoder, its basic calculation formula is:

[0098]

[0099] In the formula, This means that the mean is calculated independently for each channel along the time dimension t, with the result maintaining the channel dimension unchanged. The decoder then maps this latent variable z back to the target quantity in the observation space, reconstructing the target variable. The basic calculation formula is:

[0100]

[0101] In the formula, the activation function Choosing GELU, h1 and h2 are the first and second hidden layer representations respectively, and the final output is... This represents the target variable estimate at the end of the corresponding window. By jointly optimizing the parameters of the hidden layer, the encoder extracts the health domain representation in the latent space, and the decoder nonlinearly maps it back to the target parameters, thereby improving the fit between the reconstructed values ​​and the true values ​​on normal samples. At the same time, a health residual library is established using the reconstruction residuals on normal samples, which facilitates subsequent threshold learning and online monitoring.

[0102] The specific steps for estimating the conditional quantile threshold in step C) are as follows:

[0103] Step C1) To characterize the differences in the scale of normal fluctuations under the same conditions, equal-frequency binning is performed on a healthy database for a single continuous state feature. Assume that equal-frequency binning divides the value range into k intervals according to the sample quantiles, making the sample size in each bin approximately equal, thus ensuring that the quantile estimates of each bin have similar variance levels. The bin boundaries are {A0 < A1 < ... < A...}. k}, then the mapping from state to bucket is:

[0104]

[0105] After completing the equal-frequency grouping of the state dimension, further processing of the influence of individual differences and flight phases is required. Therefore, in the threshold learning phase, the health library residuals are aggregated according to the three-dimensional conditions of engine individual e, flight phase p, and state bucket b, and the high quantile threshold is estimated on each condition group. Let the group label g = (e, p, b), and its health residual set be... Given a quantile level α∈(0,1), the conditional threshold is defined as:

[0106]

[0107] The specific steps of the hierarchical rollback mechanism in step D) are as follows:

[0108] Step D1) To ensure the anomaly monitoring system maintains good accuracy and stability when facing unfamiliar sample groups, a hierarchical backoff strategy is introduced. This strategy automatically backoffs based on other conditions such as flight phase and operating status to obtain health thresholds for similar aircraft models or states. The threshold selection rules are given below in piecewise function form:

[0109]

[0110] After the hierarchical backoff mechanism determines the conditional thresholds to be used at each time step, a single-point detection is performed on the current absolute residual using the group thresholds during the discrimination process:

[0111]

[0112] In the formula, 1(·) is an indicator function.

[0113] The specific steps for the continuous duration consistency rule in step E) are as follows:

[0114] Step E1) In a threshold selection system with existing three-dimensional condition matching and backoff mechanisms, an alarm is triggered only when the exceedance persists for at least K sampling periods. K can be calibrated based on the sampling rate and acceptable response delay to achieve a balance between robustness and sensitivity. Let the sampling period be Δt and... The alarm criteria are as follows:

[0115]

[0116] Beneficial Effects: The present invention provides a method for monitoring anomalies in a high-bypass turbofan engine fleet under all operating conditions based on a self-attention reconstruction network. Compared with existing technologies, the above technical solution has the following technical advantages:

[0117] (1) The applicable operating conditions for abnormal monitoring have been increased. By using a self-attention reconstruction network to learn a full-condition health baseline mapping on health data, it can cover multiple flight phases such as takeoff, climb, cruise, and descent, reducing the dependence on manufacturer benchmark models and single-condition baselines.

[0118] (2) Reduced data labeling requirements. By using conditional quantile thresholds, the health residual statistics are transformed into comparable anomaly discrimination rules, which are more robust to skewed and heteroscedastic residual distributions.

[0119] (3) Improved the generalization ability of the monitoring method across individuals. By using a hierarchical back-off threshold chain, usable thresholds can still be output even when there are insufficient samples in the subdivided groups or in the case of new engine individuals, thus improving the generalization ability across individuals and operating conditions.

[0120] To verify the effectiveness of the full-condition anomaly monitoring method for high-bypass turbofan engine fleets based on self-attention reconstruction networks designed in this invention, a portion of flight records from 2022 were selected from 3200 flight data of 25 B787 aircraft with GEnx-1B engines from 4 airlines. The training set included 200 normal flight data from each season of the year to characterize the normal operating status of the engines; the test set consisted of 40 normal flights and 10 malfunctioning flights.

[0121] To compare the trainability and convergence stability of different reconstruction methods, a curve showing the training loss changing with epochs is plotted. The training loss curve is shown below. Figure 5 As shown.

[0122] like Figure 5As shown, Informer-AE has the fastest loss reduction speed and the smallest final convergence value, and can stably enter the optimal range within about 50 rounds. This indicates that it can more fully explore global dependencies within a long time window and achieve higher gradient utilization efficiency through parallel computing.

[0123] To further verify the actual reconstruction capabilities of each reconstruction algorithm during the testing phase, two indices, RMSE and mean relative error, were calculated for the same batch of flight data. The specific results are shown in Table 3.

[0124] Table 3 Comparison of Reconstruction Errors Using Different Methods

[0125]

[0126] As shown in Table 3, in terms of overall trend, Informer-AE performs best in both top-error indices, with average values ​​of 4.81% and 0.57%, respectively, and exhibits smaller and more stable reconstructed residuals.

[0127] To verify the discriminative ability of different reconstruction methods in anomaly detection tasks, precision, recall, and F1 score were used as comprehensive evaluation indicators to conduct a horizontal comparison of the four methods. The results are shown in Table 4.

[0128] Table 4 Comparison of monitoring results using different methods

[0129]

[0130] As shown in Table 4, Informer-AE outperformed in all three metrics (P = 0.9281, Recall = 0.9833, F1 = 0.9549), indicating that it can maintain high recall of real anomalies while reducing false alarm rate, and has the most stable overall discrimination.

[0131] To verify the robustness and noise resistance of the reconstruction-based online monitoring process on healthy flights, multiple normal flights were selected for visualization analysis. Comparison charts of measured and reconstructed EGT curves and residual-threshold evolution diagrams were plotted. The results are as follows: Figure 6 As shown.

[0132] like Figure 6 As shown, Informer-AE's reconstruction of EGT in different phase intervals such as takeoff, climb, cruise, and descent is basically synchronized with the actual measurement, and the absolute residuals are within the dynamic threshold band for a long time, indicating that the model can achieve high-fidelity fitting of normal behavior.

[0133] Experimental testing showed that all 10 abnormal flights in the test set were correctly detected. To further verify the method's time-by-time discrimination capability under real-world operational disturbances and multi-stage flight conditions, time-series plots comparing the measured-reconstructed EGT trajectory and absolute residuals with dynamic thresholds were generated. Alarm intervals were highlighted in red. The results are as follows: Figure 7 As shown.

[0134] like Figure 7 As shown, the model maintains stable sensitivity across different flight phases: once an anomaly occurs, the residual quickly surpasses the dynamic threshold learned under the corresponding phase and state conditions, forming a continuous over-limit band consistent with the alarm interval; when the anomaly is resolved or the operating conditions return to normal, the residual quickly falls back to within the threshold band, and the alarm subsides, demonstrating good temporal consistency and low hysteresis characteristics. This indicates that the anomaly measurement and conditional threshold strategy based on reconstructed residuals can effectively adapt to changes in multiple operating conditions and maintain clear and stable discrimination boundaries during the occurrence and subsidence of anomalies.

[0135] It is worth noting that, to evaluate the reconstructed network's ability to monitor unfamiliar engine units, engine AC5 was excluded from the training set as an independent test sample. Experimental results are as follows... Figure 7 As shown in (i) and (j), although the overall reconstruction error of the engine is slightly increased due to individual differences, the dynamic threshold and hierarchical backoff strategy based on hierarchical conditionalization can still achieve accurate alarm and stable discrimination.

[0136] This phenomenon indicates that the reconstructed network has a certain generalization ability when faced with unseen data, and the dynamic threshold strategy can adaptively compensate for cross-device differences, thereby ensuring the reliability and transferability of monitoring results.

[0137] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for full-condition anomaly monitoring of a high-bypass turbofan engine fleet based on a self-attention reconstruction network, characterized in that, Includes the following steps: Step A) Obtain multivariate time-series operational data of the aircraft engine fleet, including at least exhaust gas temperature (EGT) and its individual identifier e / flight phase identifier p related parameters, and preprocess the multivariate time-series operational data. Step B) Train the Informer-AE reconstruction network on a training set containing only healthy / normal fleet operation data to establish a full-condition baseline model, and calculate the residuals based on the measured values ​​of the target parameters and the reconstructed values ​​of the baseline model to establish a healthy residual library. Step C) Select at least one continuous state feature s in the health residual library and perform equal-frequency binning on it to obtain state bucket b. Estimate the conditional quantile threshold at the quantile level for each three-dimensional condition set g = (e, p, b). Step D) Based on the operating condition information of the sample to be monitored, select the quantile threshold corresponding to the operating condition information from the threshold library as the judgment threshold, and perform hierarchical backtracking to obtain an available judgment threshold when the threshold library is missing or does not meet the preset stability conditions. Step E) Compare the reconstructed residual of the sample to be monitored with the judgment threshold, and output an alarm result in combination with the continuous duration consistency rule, wherein the alarm output is triggered only when the residual meets the threshold condition and lasts for no less than K sampling periods.

2. The method for full-condition anomaly monitoring of a high-bypass turbofan engine fleet based on a self-attention reconstruction network according to claim 1 is characterized in that: The specific steps for preprocessing multivariate time series data in step A) are as follows: Step A1) The flight phase identifier p is determined using the phase code recorded in QAR, and the correspondence between the phase code and the flight phase satisfies the conditions shown in Table 1: Table 1 Definition of Flight Phases Step A2) Based on engineering experience and commonly used sensor combinations, select relevant engine measurement parameters and perform Pearson correlation analysis on them. The specific extracted flight parameters and their meanings are shown in Table 2: Table 2 QAR Parameter Extraction Comparison Table Step A3) Calculate the mean and standard deviation of each variable based on the health training data, and use the mean and standard deviation to perform consistent standardization processing on the data from the training, validation, and online monitoring phases. The specific formula is as follows: In the formula, μ j σ j denoted as the mean and standard deviation of feature channel j on the training set, respectively.

3. The method for full-condition anomaly monitoring of a high-bypass turbofan engine fleet based on a self-attention reconstruction network according to claim 1 is characterized in that: The specific steps for establishing the baseline model and health residual library in step B) are as follows: Step B1) First, perform linear mapping at each time step to obtain three sets of vectors: query, key, and value. Then, obtain the relevance weights between positions through scaling dot product and normalization. Under the combined effect of multi-head parallelism and sparse selection mechanism, focus on the dependencies that contribute more information to improve efficiency and robustness. The basic calculation formula is as follows: Q=E0W Q ,K=E0W K ,V=E0W V In the formula, W Q W K W V Given a trainable projection matrix, the weights for any attention endpoint are then calculated using a scaled dot product and the value vectors are summed with weights: To enhance the diversity and directionality of feature representation, a multi-head attention architecture with h attention heads can be adopted, each using an independent set of parameters. The outputs from each head are concatenated by channel and linearly mapped back to the model dimension: M(E0)=Concat(Attn(Q (1) ,K (1) ,V (1) ),...,Attn(Q (h) ,K (h) ,V (h) ))W O In the formula, W O The output projection enables the multi-head structure to model dependencies simultaneously from different angles and at different time scales, thus improving the ability to represent complex sequences. Based on the aforementioned standard multi-head attention, and to adapt to the engineering constraints of long-sequence engine time-series data, a sparsity approach is introduced: for the i-th query q l Scoring vector Its information contribution is defined as: Based on this, select the top n query index sets with the greatest contribution. Full attention is performed only on queries within S, and a sparse weight matrix is ​​formed. get: From S =A S In, In the formula, ContextFill(·) represents a sparse reconstruction operation that fills / interpolates the positions of unselected queries using neighbor or global statistics. Sparsity selection can replace the expected complexity of O(L) with approximately O(nL). 2 Fully connected attention, while retaining sensitivity to remote dependencies and cross-condition alignment capabilities, is ideal for ultra-long, heteroscedastic flight parameter data. Step B2) After the above steps, cross-position temporal information has been extracted in the time dimension, but the output is still arranged according to time steps. It is necessary to perform average pooling on each channel in the length L dimension to compress the sequence-level representation into a latent variable of a fixed dimension. As the input to the decoder, its basic calculation formula is: In the formula, This means that the mean is calculated independently for each channel along the time dimension t, with the result maintaining the channel dimension unchanged. The decoder then maps this latent variable z back to the target quantity in the observation space, reconstructing the target variable. The basic calculation formula is: In the formula, the activation function Choosing GELU, h1 and h2 are the first and second hidden layer representations respectively, and the final output is... This represents the target variable estimate at the end of the corresponding window. By jointly optimizing the parameters of the hidden layer, the encoder extracts the health domain representation in the latent space, and the decoder nonlinearly maps it back to the target parameters, thereby improving the fit between the reconstructed values ​​and the true values ​​on normal samples. At the same time, a health residual library is established using the reconstruction residuals on normal samples, which facilitates subsequent threshold learning and online monitoring.

4. The method for full-condition anomaly monitoring of a high-bypass turbofan engine fleet based on a self-attention reconstruction network according to claim 1 is characterized in that: The specific steps for estimating the conditional quantile threshold in step C) are as follows: Step C1) To characterize the differences in the scale of normal fluctuations under the same conditions, equal-frequency binning is performed on a healthy database for a single continuous state feature. Assume that equal-frequency binning divides the value range into k intervals according to the sample quantiles, making the sample size in each bin approximately equal, thus ensuring that the quantile estimates of each bin have similar variance levels. The bin boundaries are {A0 < A1 < ... < A...}. k }, then the mapping from state to bucket is: if A J-1 <s t ≤A J then b t =j,j=1,...,k After completing the equal-frequency grouping of the state dimension, further processing of the influence of individual differences and flight phases is required. Therefore, in the threshold learning phase, the health library residuals are aggregated according to the three-dimensional conditions of engine individual e, flight phase p, and state bucket b, and the high quantile threshold is estimated on each condition group. Let the group label g = (e, p, b), and its health residual set be... Given a quantile level α∈(0,1), the conditional threshold is defined as:

5. The method for full-condition anomaly monitoring of a high-bypass turbofan engine fleet based on a self-attention reconstruction network according to claim 1 is characterized in that: The specific steps of the hierarchical rollback mechanism in step D) are as follows: Step D1) To ensure the anomaly monitoring system maintains good accuracy and stability when facing unfamiliar sample groups, a hierarchical backoff strategy is introduced. This strategy automatically backoffs based on other conditions such as flight phase and operating status to obtain health thresholds for similar aircraft models or states. The threshold selection rules are given below in piecewise function form: After the hierarchical backoff mechanism determines the conditional thresholds to be used at each time step, a single-point detection is performed on the current absolute residual using the group thresholds during the discrimination process: exceed t =1(R t >T * (e,p t ,b i )) In the formula, 1(·) is an indicator function.

6. The method for full-condition anomaly monitoring of a high-bypass turbofan engine fleet based on a self-attention reconstruction network according to claim 1 is characterized in that: The specific steps of the continuous duration consistency rule in step E) are as follows: Step E1) In a threshold selection system with existing three-dimensional condition matching and backoff mechanisms, an alarm is triggered only when the exceedance persists for at least K sampling periods. K can be calibrated based on the sampling rate and acceptable response delay to achieve a balance between robustness and sensitivity. Let the sampling period be Δt and... The alarm criteria are as follows:

Images