Aero-engine high-pressure turbine blade crack prediction method

Abstract

This invention relates to a method for predicting cracks in high-pressure turbine blades of aero-engines, comprising: S1: acquiring various engine operating parameters; S2: preprocessing the engine operating parameters to obtain engine data; S3: processing the engine data in conjunction with engine location to obtain engine feature vectors for each flight segment; S4: calculating cumulative features for each flight segment based on the engine feature vectors; S5: constructing a sample sequence based on the engine feature vectors and cumulative features; S6A: calculating the prediction probability using a fusion algorithm based on the engine feature vectors and cumulative features; S6B: calculating an influence factor based on the engine feature vectors; S7: comparing the predicted probability with an optimal classification threshold, and comparing the influence factor with an influence factor threshold to obtain the crack prediction result. The method for predicting cracks in high-pressure turbine blades of aero-engines described in this invention has the advantages of high recall and high accuracy in predicting early-stage aero-engine cracks.

Description

Technical Field

[0001] This invention relates to the field of detection technology, and in particular to a method for predicting cracks in high-pressure turbine blades of aero-engines. Background Technology

[0002] Aero engines are the core components of aircraft. Taking the CFM56-7B engine used in the Boeing 737NG series aircraft as an example, its high-pressure turbine blades are subjected to harsh environments of high temperature, high pressure and high speed for a long time, which makes them prone to fatigue cracks, affecting the operational reliability of the aero engine and directly impacting flight safety.

[0003] Currently, the inspection of high-pressure turbine blades mainly employs periodic inspections based on non-destructive testing (NDT) and trend analysis based on airborne sensor data. Periodic inspections using NDT methods such as ultrasonic or eddy current testing cannot achieve real-time online monitoring, and have long monitoring cycles and high costs. Trend analysis based on airborne sensor data primarily triggers alarms by monitoring thresholds for parameters such as exhaust temperature, vibration, and speed; however, this method also suffers from latency and a high false alarm rate. Furthermore, methods using convolutional neural networks to predict blade remaining life, which have emerged with the development of machine learning, also suffer from overfitting or insufficient generalization, making it difficult to operate stably under complex real-world conditions. Summary of the Invention

[0004] Based on this, the purpose of this invention is to achieve the ability of early warning of cracks with high recall and high accuracy, and to provide a method for predicting cracks in high-pressure turbine blades of aero engines suitable for CFM56 series engines.

[0005] A method for predicting cracks in high-pressure turbine blades of aero-engines includes the following steps: S1: Obtain various engine operating parameters; S2: Preprocess the engine operating parameters to obtain engine data; the preprocessing includes removing files with severe data loss, standardizing the date and time, generating a crack label column based on the crack record, and aligning it with the engine serial number and time label.

[0006] S3: Combine the engine location with the engine data to obtain the engine feature vectors for each flight segment; S4: Calculate the cumulative sum, cumulative average, and non-zero percentage of the characteristic values ​​in each flight segment based on the characteristic vectors of each engine to obtain the cumulative characteristics; S5: Construct a sample sequence based on the engine feature vector and cumulative features; S6A: Based on the engine feature vector and cumulative features, a fusion algorithm is used to calculate the prediction probability. The fusion algorithm includes three models, which respectively perform stable baseline prediction and feature importance screening, accurately fit complex features for verification and learn high-order combinations between features, capture long-term temporal dependencies and understand the evolution patterns between flight phases; the prediction probability is the weighted average of the prediction results of the three models. S6B: Calculate the influence factor based on the engine feature vector; the influence factor is used to conduct a more granular classification assessment of crack risk based on the physical damage mechanism; S7: Compare the predicted probability with the optimal classification threshold, and compare the influence factor with the influence factor threshold to obtain the crack prediction result; when the predicted probability is greater than or equal to the optimal classification threshold and the influence factor is greater than or equal to the influence factor threshold, it is determined that there is a crack; when the predicted probability is greater than or equal to the optimal classification threshold and the influence factor is less than the influence factor threshold, it is determined that there is an adjacent crack; when the predicted probability is less than the optimal classification threshold and the influence factor is greater than or equal to the influence factor threshold, it is determined that attention is needed; when the predicted probability is less than the optimal classification threshold and the influence factor is less than the influence factor threshold, it is determined that there is no crack.

[0007] Compared to existing technologies, this invention employs a fusion algorithm for comprehensive prediction and utilizes dual verification of influence factors, enabling it to simultaneously achieve high recall, high accuracy, generalization ability, and adaptability in crack prediction. Furthermore, because the prediction process fully integrates the engine's damage mechanism, the prediction results possess both physical interpretability and engineering applicability.

[0008] Furthermore, the impact factor is calculated as follows: S6B1: Obtain the initial climb time and cruise fuel flow for each flight segment; S6B2: Calculate the average initial climb time and cruise fuel flow rate for each item to obtain the average initial climb time and average cruise fuel flow rate; S6B3: Normalize the average initial climb time and average cruise fuel flow rate; S6B4: The influence factor is calculated using a weighted harmonic average based on the normalized average initial climb time and average cruise fuel flow rate.

[0009] Furthermore, the expression for the weighted harmonic average is: , in, The initial climb time weighting coefficient is denoted by 'a', where 'a' is the average initial climb time and 'b' is the average cruise fuel flow rate. When 'a' or 'b' is zero, the influencing factor is directly assigned a value of zero.

[0010] Further, in step S6B1, only the initial climb time (INIT_CLIMB_time) and the cruise fuel flow rate (ENG1FF_avg) or the left engine fuel flow rate (ENG2FF_avg) are acquired when the initial climb time (INIT_CLIMB_time) is greater than zero and the right engine fuel flow rate (ENG1FF_avg) is greater than zero.

[0011] Furthermore, linear stretching optimization is performed on the distribution set of influencing factors, and its expression is: , , , in, The influencing factors before stretching, The influencing factor after stretching, It is the smallest impact factor.

[0012] Furthermore, the sample sequence is constructed by tail-aligned windowing, and the feature sequences within each window are aggregated to obtain a sample sequence labeled with the crack label corresponding to the last flight segment of the window.

[0013] Furthermore, the engine feature vector includes flight-related features, N2 rotational speed features, vibration features, EGT features, rainflow count and spectrum features, and thermal fatigue damage features.

[0014] Furthermore, the three models of the fusion algorithm are the Random Forest model, the XGBoost model, and the Transformer model.

[0015] Furthermore, during the training process of the fusion algorithm, after each round of training, the fusion algorithm searches for the optimal threshold on the validation set with the F2 score as the target.

[0016] Based on the same inventive concept, the present invention also includes an electronic device comprising a processor; a memory for storing a computer program executed by the processor; wherein, when the processor executes the computer program, it implements the method for predicting cracks in high-pressure turbine blades of aero-engines as described above.

[0017] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0018] Figure 1 This is a flowchart of the crack prediction method of the present invention. Detailed Implementation

[0019] This invention addresses the problems of long monitoring cycles and high costs associated with aero-engines. It carefully analyzes aero-engine flight data and designs a high-pressure turbine blade crack prediction method suitable for CFM56 series engines, based on the characteristics of blade cracks. This method deeply explores the physical meaning of information from each flight stage of the aero-engine and the cumulative damage characteristics, and obtains prediction results with high recall and high accuracy after processing with multiple models.

[0020] Specifically, please refer to Figure 1 , Figure 1 This is a flowchart of the crack prediction method of the present invention. A crack prediction method for high-pressure turbine blades of an aero-engine includes: S1: Acquire various engine operating parameters. Each flight segment corresponds to one flight and generates a log file; the log file is stored in the Quick Access Recorder (QAR) and includes aircraft history and flight status data. The aircraft history includes engine serial number, installation location, registration number, and crack occurrence date; the flight status data records various engine parameters in corresponding fields, including file name, date, time, air and ground conditions, flight phase, left engine N2 speed, right engine N2 speed, left engine EGT, right engine EGT, left and right engine N2 rotor vibration, left and right engine fuel flow, left engine fuel cut-off status, right engine fuel cut-off status, left engine throttle lever angle, right engine throttle lever angle, left engine compressor outlet pressure, right engine compressor outlet pressure, initial climb time, and cruise fuel flow; where N2 speed is the speed of the high-pressure rotor of the aero engine, and EGT is the exhaust temperature.

[0021] S2: Preprocess the aforementioned engine operating parameters to obtain engine data. This preprocessing includes removing files with severe data loss, standardizing the date and time, and generating a crack label column based on crack records, aligning it with the engine serial number and time label. The crack label column is marked as 1 for the first 1500 flight segments after a crack occurs, and marked as 0 for all flight segments before the number of flight segments is reached or there are no cracks.

[0022] S3: Process the engine data according to the engine location to obtain the engine feature vectors for each flight segment. The engine feature vectors include 80 basic feature vectors, including flight-related features, N2 speed features, vibration features, EGT features, rainflow count and spectrum features, and thermal fatigue damage features.

[0023] The flight-related features include: flight time; flight category, which is 0 if the flight segment is long-range, otherwise 1; engine shutdown cooling time category (ENG_SHUT_DOWN_time_category), which is 0 if the number of shutdown time records is greater than 300, otherwise 1. This shutdown time includes the number of records during the landing phase, landing taxiing phase, throttle lever angle between 0 and 45 degrees, fuel cut-off signal (FCO) not in the cut-off state, and fuel flow (FF) less than 950; and warm-up time category (ENG_start_warmup_time_category), which is 0 if the start-up warm-up time is greater than 180, otherwise 1. The `armup_time` flag represents the number of times a flight occurs during engine start-up, taxiing, climb, or cruise phases, with the throttle lever angle between 0 and 45 degrees, the fuel cutoff signal (FCO) not in the cut-off state, the fuel flow (FF) greater than 400, and the air-to-ground status not in the air (AIR) state. The `DERATE_conf` flag is set to 0 if the `DERATE` flag is present during takeoff and the average N1 is less than or equal to 95 or if thrust is reduced; otherwise, it is set to 1. The `Flight_training_category` flag in the `AIR_GND_1` state sequence represents the interval between three air-to-ground transitions in the "AIR→-→AIR→-→AIR" mode within a specified range; otherwise, it is set to 0.

[0024] The N2 rotational speed characteristics include the cumulative time (Key1_2_c) for N2 greater than or equal to 99.0, the mean (ENG_N2_avg) of the N2 greater than or equal to 99.0 interval, the standard deviation (ENG_N2_std) for N2 greater than or equal to 99.0, and the 95th quantile (ENG_N2_p95) for N2 greater than or equal to 99.0.

[0025] The vibration characteristics include the number of vibration values ​​> 0.8 (ENG_VIBTN2_count_high), the maximum value of the vibration (ENG_VIBTN2_max), and the standard deviation of the vibration (ENG_VIBTN2_std).

[0026] The EGT characteristics include: statistical values ​​of EGT when the fuel cutoff signal (FCO) is not in the cutoff state, including the engine maximum exhaust temperature (ENG_EGT_max), engine average exhaust temperature (ENG_EGT_avg), engine exhaust temperature standard deviation (ENG_EGT_std), and the 95th percentile of engine exhaust temperature (ENG_EGT_p95); the duration of EGT exceeding 920°C (ENG_EGT_overtemp_duration) and the maximum value of EGT exceeding 920°C (ENG_EGT_overtemp_max); the absolute maximum value of the rate of change of EGT (ENG_EGT_rate_max) and the mean value of the rate of change of EGT (ENG_EGT_rate_avg); and the average value of the 11th stage high-pressure compressor pressure (PS3) when EGT is greater than 700°C (ENG_EGT_gt_700_PS3_SEL_avg).

[0027] The rainflow count and spectral features include rainflow cycle count (N2_rainflow_count) based on N2 sequence, rainflow cycle cumulative damage (N2_cumulative_damage) based on N2 sequence, N2 sequence FFT dominant frequency (N2_fft_max_freq), rainflow count based on vibration (VIBTN2_rainflow_count), and rainflow damage based on vibration (VIBTN2_cumulative_damage). The spectral characteristics include the dominant frequency of vibration (VIBTN2_fft_dominant_freq), total power (VIBTN2_fft_total_power), spectral centroid (VIBTN2_fft_centroid), spectral spread (VIBTN2_fft_spread), low-frequency power (VIBTN2_fft_band_0_10_power), and low-frequency peak frequency (VIBTN2_fft_band_0_10_peak_freq); and the frequency domain characteristics of exhaust gas temperature (EGT) include (EGT_fft_dominant_freq), standard deviation of rate of change (EGT_rate_std), exhaust gas temperature stability (EGT_temp_stability), quantified over-temperature damage (EGT_overtemp_ratio), total number of anomalies (EGT_temp_anomaly_count), overall anomaly level (EGT_temp_anomaly_ratio), average temperature gradient (EGT_temp_gradient_mean), and temperature gradient fluctuation level (EGT_temp_gradient_std). The thermal fatigue damage characteristics include thermal fatigue damage (EGT_thermal_fatigue_damage), thermal fatigue cycles (EGT_thermal_fatigue_cycles), and thermal fatigue damage rate (EGT_thermal_fatigue_damage_rate) calculated based on the EGT rainflow cycle and Coffin-Manson model.

[0028] Meanwhile, time features (days_since_first_flight), month (days_since_first_flight), and week (days_since_first_flight) are added to each of the basic feature vectors, expanding the above 80-dimensional basic feature vectors to 240 dimensions. That is, the record file for each flight segment yields a 240-dimensional engine feature vector.

[0029] S4: Calculate the cumulative sum, cumulative average, and non-zero percentage of the feature values ​​in each segment based on the feature vectors of each engine to obtain the cumulative feature. The cumulative sum is the sum of the feature values ​​from the first segment to the current segment, the cumulative average is the cumulative sum divided by the current segment number, and the non-zero percentage is the proportion of the number of segments with non-zero feature values ​​to the number of current segments.

[0030] Furthermore, a data integrity correction coefficient based on historical statistics is introduced, and the cumulative sum is divided by this data integrity correction coefficient for correction; subsequently, the cumulative sum is normalized to obtain a normalized cumulative sum.

[0031] S5: Construct a sample sequence based on the engine feature vector and cumulative features. The sample sequence is constructed using a tail-aligned windowing method, and the feature sequences within each window are aggregated to obtain a sample sequence labeled with the crack label corresponding to the last flight segment of the window; the aggregation selects one or more of the last flight segment value, mean, maximum value, and minimum value. In this embodiment, a window with a length of 20 and a step size of 1 is used for construction; the aggregation uses mean aggregation.

[0032] S6A: Based on the sample sequence, a fusion algorithm is used to calculate the predicted probability. The fusion algorithm simultaneously employs a random forest model, an XGBoost model, and a Transformer model. The random forest model performs stable baseline prediction and initial feature importance screening; the XGBoost model performs accurate fitting of complex features and learns higher-order combinations between features; and the Transformer model captures long-term temporal dependencies and understands evolutionary patterns between flight phases. The three models complementarily predict the sample sequence. After prediction, each model obtains a set of predicted probabilities, and the predicted probabilities are then averaged with equal weights to obtain the predicted probability.

[0033] During the training process of the fusion algorithm, the cracked areas of the engines with cracks are densely sampled as crack samples, while the engines without cracks are uniformly sampled from the most recent 1500 flight segments as normal samples. Subsequently, the crack samples and normal samples are divided into a validation set and a test set. In this embodiment, a maximum of 300 crack samples with a step size of 1 and a maximum of 500 normal samples are collected.

[0034] The random forest model constructs multiple independent decision trees to vote on or average the prediction results for each input feature. When processing the sample sequence, the random forest model first standardizes each input feature, converting it to a standard normal distribution to eliminate the influence of different feature dimensions, allowing the model to objectively evaluate the contribution of each feature. To prevent overfitting to noise in the training data, this random forest model limits the tree depth to a very shallow level to construct a "weak learner," capturing only the simplest interactions between features, such as "when EGT_overtemp_duration exceeds a certain threshold and vibration value exceeds a certain threshold, the crack probability increases." For each input feature in the sample sequence, the expression for the predicted probability of the random forest model is: , in, For the number of decision trees, For the first The prediction results of each decision tree. The growth of each decision tree follows the CART (Classification and Regression Tree) algorithm, using minimizing Gini impurity as the splitting criterion, and its expression is: , in, The sample set for the current node. For the number of categories, Let be the proportion of samples of class k in the node. During splitting, the node selects the feature and threshold that maximizes the decrease in Gini impurity for splitting.

[0035] During the training of the random forest model, since leaf cracking is a cumulative damage process, although early signs are weak, some features remain almost unchanged throughout the entire lifecycle and do not contribute to prediction. Therefore, by calculating the feature variance and setting a variance threshold, these "lazy" features are automatically identified and removed to reduce model dimensionality and prevent the introduction of noise, thereby improving training efficiency and generalization ability. The expression for the feature variance is: , Where X represents the data obtained from each dimension.

[0036] Meanwhile, since crack samples are far fewer than normal samples, to prevent the model from predicting samples as normal to obtain high accuracy, a balance parameter is used to automatically adjust the weights based on the inverse proportion of the number of samples in each category. This makes the random forest model pay more attention to crack samples to improve recall. The formula for calculating this weight is: , in, The total number of samples in the training set. Let be the number of samples of a specific category in node j. Let be the total number of samples in node j.

[0037] In this embodiment, 100 decision trees are used, and the variance threshold is set to 0.001.

[0038] The XGBoost model, based on the boosting concept, iteratively trains new decision trees to fit the residuals of the previous decision tree. Through an additive model, it progressively optimizes the objective function through sequential learning to accurately approximate complex nonlinear relationships. Specifically, the XGBoost model employs the hist tree method to automatically discretize and bin continuous features, accelerating computation and automatically capturing the nonlinear effects of feature intervals. Taking feature k as an example, its value range is divided into b bins, each corresponding to a discretized interval, expressed as follows: .

[0039] The XGBoost model predicts the following model at the t-th iteration: , in, Let t be the newly added decision tree.

[0040] The objective function consists of a loss function and a regularization term, and its expression is: , The loss function used is LogLoss, and its expression is: , The loss function is approximated by a second-order Taylor expansion to obtain the optimization objective: , , , in, For the Sigmoid function, i.e. , and These are the first and second derivatives of the loss function, respectively. The regularization direction is used to control model complexity, and its expression is: , in, The number of leaf nodes. The weight of the leaf node. This is a structural complexity penalty, representing the minimum loss reduction required for splitting; The L2 regularization coefficient is... This is the L1 regularization coefficient. For aero-engine data that is complex and contains significant noise and random fluctuations, the structural complexity penalty and regularization coefficient penalize the model's complexity, forcing the model to learn only the most robust and essential patterns from the data. This ensures that the decision tree only splits when a certain condition is met, suppressing invalid splits based on accidental correlations. This splitting condition is: , in, and These are samples representing the left and right child nodes, respectively. Former parent node For weight regularization terms, This is a splitting regularization term.

[0041] During the training of the XGBoost model, each sample is first randomly sampled to increase model diversity and reduce the risk of overfitting. Simultaneously, to prevent excessive weighting of normal samples when normal samples are extremely scarce, which could lead to an overemphasis on a very small number of normal samples and a surge in misclassification rates, the weights of normal samples are adjusted based on the ratio of cracked samples to normal samples, and the maximum weight is limited to improve recall while maintaining reasonable precision. The expression for this weight is: , in, To determine the maximum weight, it is set to less than or equal to 5 in this embodiment.

[0042] The Transformer model captures long-distance dependencies in sequence data through a self-attention mechanism and a multi-layer architecture. The self-attention mechanism directly calculates the correlation between any two segments in the sample sequence. Specifically, before the Transformer encoder, three convolutional kernels of different sizes are used to perform multi-scale analysis on the features of each segment, extracting feature patterns from different time spans. Before the attention mechanism, sinusoidal position encoding injects temporal information into the samples of each segment, enabling the attention mechanism, which does not contain sequential information, to distinguish the temporal order of different segments. The encoder employs a multi-head attention mechanism and attention pooling to aggregate the most important information from the sample sequence. Simultaneously, the context vector obtained from attention pooling is weighted and fused after predicting cracks from different dimensions using three independent classifiers. Finally, temperature scaling is applied to pull the model back from ambiguous predictions, making more definitive judgments and improving recall while maintaining a certain level of accuracy.

[0043] The convolution operation is as follows: , , in, This represents the convolution operation. Let be the weight matrix with kernel size k, BN be batch normalization, and GELU be the Gaussian error linear unit activation function. When k is 3, it may capture short-term fluctuations such as an anomalous vibration spike between adjacent flight segments, while when k is 7, it can capture long-term trends such as the gradual rise of the EGT baseline across multiple flight segments.

[0044] The expression for the sinusoidal position encoding is: , , in, For location index, For dimensional indexing.

[0045] The weighted expression for each classifier is: , in, The context vector obtained from attention pooling. This is the Sigmoid activation function.

[0046] For the input sequence, the attention mechanism obtains the query, key, and value matrix through a linear transformation, as follows: , The output of self-attention is calculated as follows: , , Where Q is the query matrix, K is the key matrix, and V is the value matrix. T is the sequence length. For model dimensions, divide by This is to prevent the softmax gradient from vanishing due to an excessively large dot product result.

[0047] The expression for the attention pooling mechanism is: , , in, For the first The hidden state at each time step This is a learnable attention weight vector. The normalized attention score. This is the final context vector.

[0048] The expression for temperature scaling is: , Where z is the model output logits, and T is the temperature parameter. When the temperature parameter is greater than 1, the probability distribution becomes smoother.

[0049] During training, the model uses a complete sequence of samples the size of a window as the raw input instead of aggregated single-point features, enabling it to learn the evolutionary trends between a series of flight segments before the crack occurs. An adaptive Focal Loss function is employed, expressed as: , in, This represents the model's predicted probability of the true class. To focus parameters, For class weights. This loss function uses... Factors focus on samples that are difficult to classify, and these difficult-to-classify samples are likely to be segments of the flight path at the critical point of crack initiation. By focusing on these samples, the recall rate of the model can be significantly improved.

[0050] In this embodiment, the temperature parameter is 1.2. Take 2.0.

[0051] After each model in the fusion algorithm completes one round of training, the fusion algorithm searches for the optimal threshold in the range of 0.1 to 0.9 on the validation set, focusing on recall, with the F2 score as the target. The expression for the F2 score is: .

[0052] S6B: Calculate the influence factor based on the engine's characteristic vector. The influence factor is used for a more granular classification assessment of crack risk based on physical damage mechanisms, and its calculation method is as follows: S6B1: Extract the initial climb time and cruise fuel flow for each flight segment from the log file. Only the initial climb time (INIT_CLIMB_time) greater than zero, and the fuel flow for either the left engine (ENG1FF_avg) or the right engine (ENG2FF_avg) greater than zero, are retrieved. The initial climb phase is a crucial transition from takeoff to cruise altitude. During this phase, the engine operates at high speed and high load, and the turbine blades experience significant thermo-mechanical stress. A shorter initial climb time means the engine rapidly climbs to high speed within a short period. This rapid acceleration leads to more severe thermal and centrifugal force impacts on the blades, and the cumulative effect accelerates the initiation and propagation of fatigue cracks. The fuel flow directly reflects the engine's thrust output and thermal load level. When cracks appear in the high-pressure turbine blades, their aerodynamic performance deteriorates, leading to reduced turbine efficiency. To maintain the same thrust output, the engine control system automatically increases fuel supply to compensate for performance loss. Therefore, a higher cruise fuel flow is not only an indicator of high-load operation but also a significant external manifestation of blade performance degradation. Therefore, engines with a high probability of cracking are more likely to actually have cracks if they also have a short initial climb time and a large cruise fuel flow.

[0053] S6B2: Calculate the average of each initial climb time and cruise fuel flow rate. The average initial climb time reflects the engine's climb habits throughout its entire lifespan, and its expression is: ; The average cruise fuel flow rate reflects the recent state of blade performance degradation, and its expression is: , Where N is always the effective flight segment, Let be the initial climb time for the i-th segment.

[0054] S6B3: Normalize the average initial climb time and average cruise fuel flow rate. The normalization process normalizes the average initial climb time and average cruise fuel flow rate to [values ​​to be filled in]. The interval, expressed as: , , in, This is the theoretical maximum value for the initial climb; exceeding this value is considered a normal climb. In this embodiment, it is set to 200 seconds. The theoretical maximum value for cruise fuel flow is 4500 pounds per hour in this embodiment. A larger value of 'a' indicates a shorter initial climb time and a higher risk of cumulative damage; a larger value of 'b' indicates a larger cruise fuel flow and a more significant performance degradation.

[0055] S6B4: The influencing factor is calculated using a weighted harmonic mean based on the normalized average initial climb time and average cruise fuel flow. By using a weighted harmonic mean, sensitivity to smaller values ​​is improved, allowing for earlier detection of abnormal changes in any indicator, and preventing an overall value of zero when a single indicator is zero. Its expression is: , in, The initial climb time weighting coefficient is set to 0.65 in this embodiment. When a or b is zero, the influence factor is directly assigned a value of zero.

[0056] Furthermore, linear stretching optimization is performed on the aforementioned influencing factors. When the distribution of influencing factors across all engines in the entire fleet is relatively concentrated, making it difficult to distinguish engines with different risk levels, the distribution of influencing factors is mapped to... The interval enhances the visual differentiation of risk levels; the expression for this linear stretching is: , , , in, The influencing factors before stretching, The influencing factor after stretching, It is the smallest impact factor.

[0057] S6: The predicted probability is compared with the optimal classification threshold, and the influence factor is compared with the influence factor threshold to obtain the crack prediction result. Specifically, when the predicted probability is greater than or equal to the optimal classification threshold and the influence factor is greater than or equal to the influence factor threshold, a crack is identified; when the predicted probability is greater than or equal to the optimal classification threshold and the influence factor is less than the influence factor threshold, a crack is identified as near; when the predicted probability is less than the optimal classification threshold and the influence factor is greater than or equal to the influence factor threshold, a crack requires attention; when the predicted probability is less than the optimal classification threshold and the influence factor is less than the influence factor threshold, a crack is identified as not present. The optimal classification threshold is obtained by optimizing the validation set with the F2 score as the target. The influence factor threshold is a preset fixed value, set to 0.5 in this embodiment. By introducing the influence factor and the prediction result for dual validation, the false positive rate is reduced by 5%.

[0058] Compared to existing technologies, this invention employs a fusion algorithm for comprehensive prediction and utilizes dual verification of influence factors, enabling it to simultaneously achieve high recall, high accuracy, generalization ability, and adaptability in crack prediction. Furthermore, because the prediction process fully integrates the engine's damage mechanism, the prediction results possess both physical interpretability and engineering applicability.

[0059] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the embodiments of this application. The singular forms “a,” “the,” and “the” used in the embodiments and claims of this application are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that, unless otherwise stated, “a plurality” means two or more; the terms “first,” “second,” “third,” etc., are used only to distinguish and not to describe a particular order or sequence, nor should they be construed as indicating or implying relative importance. The term “and / or” as used herein refers to and includes any or all possible combinations of one or more associated listed items. When the above description relates to drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. In the description of this application, those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0060] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and the present invention also intends to include these modifications and variations.

Claims

1. A method for predicting cracks in high-pressure turbine blades of aero-engines, characterized in that, Includes the following steps: S1: Obtain various engine operating parameters; S2: Preprocess the engine operating parameters to obtain engine data; the preprocessing includes removing files with severe data loss, standardizing the date and time, generating a crack label column based on the crack record, and aligning it with the engine serial number and time label; S3: Combine the engine location with the engine data to obtain the engine feature vectors for each flight segment; S4: Calculate the cumulative sum, cumulative average, and non-zero percentage of the characteristic values ​​in each flight segment based on the characteristic vectors of each engine to obtain the cumulative characteristics; S5: Construct a sample sequence based on the engine feature vector and cumulative features; S6A: Based on the engine feature vector and cumulative features, a fusion algorithm is used to calculate the prediction probability; the fusion algorithm includes three models, which respectively perform stable baseline prediction and feature importance screening, accurately fit complex feature verification and learn high-order combinations between features, as well as capture long-term temporal dependencies and understand the evolution patterns between flight phases; the prediction probability is the weighted average of the prediction results of the three models; S6B: Calculate the influence factor based on the engine feature vector; the influence factor is used to conduct a more granular classification assessment of crack risk based on the physical damage mechanism; S7: Compare the predicted probability with the optimal classification threshold, and compare the influence factor with the influence factor threshold to obtain the crack prediction result; when the predicted probability is greater than or equal to the optimal classification threshold and the influence factor is greater than or equal to the influence factor threshold, it is determined that there is a crack; when the predicted probability is greater than or equal to the optimal classification threshold and the influence factor is less than the influence factor threshold, it is determined that there is an adjacent crack; when the predicted probability is less than the optimal classification threshold and the influence factor is greater than or equal to the influence factor threshold, it is determined that attention is needed; when the predicted probability is less than the optimal classification threshold and the influence factor is less than the influence factor threshold, it is determined that there is no crack.

2. The method for predicting cracks in high-pressure turbine blades of aero-engines according to claim 1, characterized in that, The impact factor is calculated as follows: S6B1: Obtain the initial climb time and cruise fuel flow for each flight segment; S6B2: Calculate the average initial climb time and cruise fuel flow rate for each item to obtain the average initial climb time and average cruise fuel flow rate; S6B3: Normalize the average initial climb time and average cruise fuel flow rate; S6B4: The influence factor is calculated using a weighted harmonic average based on the normalized average initial climb time and average cruise fuel flow rate.

3. The method for predicting cracks in high-pressure turbine blades of aero-engines according to claim 2, characterized in that, The expression for the weighted harmonic average is: ， in, The initial climb time weighting coefficient is denoted by 'a', where 'a' is the average initial climb time and 'b' is the average cruise fuel flow rate. When 'a' or 'b' is zero, the influencing factor is directly assigned a value of zero.

4. The method for predicting cracks in high-pressure turbine blades of aero-engines according to claim 2, characterized in that, In step S6B1, only the initial climb time (INIT_CLIMB_time) and the cruise fuel flow rate (ENG1FF_avg) or the right engine fuel flow rate (ENG2FF_avg) are acquired when the initial climb time (INIT_CLIMB_time) is greater than zero.

5. The method for predicting cracks in high-pressure turbine blades of aero-engines according to claim 2, characterized in that, The linear stretching optimization is performed on the aforementioned distribution set of influencing factors, and its expression is as follows: ， ， ， in, The influencing factors before stretching. The influencing factor after stretching, It is the smallest impact factor.

6. The method for predicting cracks in high-pressure turbine blades of aero-engines according to claim 1, characterized in that, The sample sequence is constructed by tail-aligned windowing, and the feature sequences within each window are aggregated to obtain a sample sequence labeled with the crack label corresponding to the last flight segment of the window.

7. The method for predicting cracks in high-pressure turbine blades of aero-engines according to claim 1, characterized in that, The engine feature vector includes flight-related features, N2 speed features, vibration features, EGT features, rainflow count and spectrum features, and thermal fatigue damage features.

8. The method for predicting cracks in high-pressure turbine blades of aero-engines according to claim 1, characterized in that, The three models of the fusion algorithm are the Random Forest model, the XGBoost model, and the Transformer model.

9. The method for predicting cracks in high-pressure turbine blades of aero-engines according to claim 8, characterized in that, During the training process of the fusion algorithm, after each round of training, the fusion algorithm searches for the optimal threshold on the validation set with the F2 score as the target.

10. An electronic device, characterized in that, processor; Memory for storing computer programs executed by the processor; The processor executes the computer program to implement the method for predicting cracks in high-pressure turbine blades of aero-engines as described in any one of claims 1-9.

Images