CN-121997036-A - Turbofan engine residual life prediction method based on two-stage degradation labels

CN121997036ACN 121997036 ACN121997036 ACN 121997036ACN-121997036-A

Abstract

The invention relates to a turbofan engine residual life prediction method based on a two-stage degradation label, which comprises the steps of obtaining sensor data of the turbofan engine and preprocessing, constructing a differential feature based on physical attributes in the sensor data, adopting piecewise linear fitting and elbow method self-adapting to the sensor data to determine the optimal segmentation number of the sensor data of each physical attribute and obtain degradation turning points, constructing a two-stage degradation label based on the degradation turning points, and constructing a TPRUL-Net prediction model fusing BiLSTM network and an attention mechanism, wherein the TPRUL-Net prediction model is provided with a time sequence feature branch used for inputting the differential feature and the two-stage degradation label, a manual feature branch used for inputting the statistical feature obtained by the sensor data, and the TPRUL-Net prediction model outputs a residual life prediction value in a double-branch fusion feature.

Inventors

SUO BIN
Gu ao
FANG YANHONG

Assignees

西南科技大学

Dates

Publication Date: 20260508
Application Date: 20260120

Claims (10)

1. The turbofan engine residual life prediction method based on the two-stage degradation label is characterized by comprising the following steps of: S1, acquiring sensor data of a turbofan engine and preprocessing; s2, constructing differential features based on physical attributes in sensor data; S3, adaptively determining the optimal segmentation number of the sensor data of each physical attribute by adopting segmentation linear fitting and elbow method combination aiming at the sensor data, and obtaining the integral degradation turning point of the turbofan engine through a multi-sensor breakpoint fusion algorithm; s4, constructing a two-stage degradation label based on the degradation turning point, wherein a stable degradation label is constructed before the degradation turning point, and a nonlinear acceleration degradation label is constructed after the degradation turning point; S5, constructing a TPRUL-Net prediction model fusing BiLSTM network and an attention mechanism, wherein the TPRUL-Net prediction model is provided with a time sequence feature branch for inputting a differential feature and a two-stage degradation label, a manual feature branch for inputting a statistical feature obtained by sensor data, and the TPRUL-Net prediction model outputs a residual life prediction value in a double-branch fusion feature; S6, acquiring sensor data of the turbofan engine to be tested, and predicting the residual service life of the turbofan engine to be tested by using the TPRUL-Net prediction model.
2. The method for predicting remaining life of turbofan engine based on two-stage degradation labels of claim 1, wherein in step S1, sensor data of the turbofan engine is obtained and preprocessed, the sensor data is time-series data, and the preprocessing comprises outlier rejection and normalization processing.
3. The two-stage degradation tag-based turbofan engine residual life prediction method according to claim 1 or 2, wherein in the step S2, in the step of constructing the differential feature based on the physical properties in the sensor data, the consistency change feature evolving over time is constructed by proportional operation for the sensor data whose physical properties are the pressure type and the flow type, the anomaly feature accumulating over time is constructed by difference operation for the sensor data whose physical properties are the temperature type and the vibration type, and the sensor data is enhanced based on the consistency change feature and the anomaly feature to construct the differential feature.
4. The method for predicting remaining life of turbofan engine based on two-stage degradation label according to claim 3, wherein in step S2, in the step of constructing the differential feature based on the physical attribute in the sensor data, the sensor data of the first working cycle of the turbofan engine is used as reference sensor data, and the ratio operation or the difference operation is performed on the sensor data of each subsequent working cycle and the reference sensor data to construct the corresponding feature.
5. The method for predicting the remaining life of a turbofan engine based on a two-stage degradation label according to claim 3, wherein in step S3, the step of adaptively determining the optimal number of segments of the sensor data of each physical attribute by combining piecewise linear fitting with an elbow method for the sensor data, and obtaining the degradation turning point of the whole turbofan engine by a multi-sensor breakpoint fusion algorithm comprises the following steps: performing principal component analysis on the preprocessed sensor data, and reserving a preset number of principal components as PCA dimension reduction data; performing low-pass filtering processing on the PCA dimensionality reduction data to eliminate noise interference; Performing piecewise linear fitting on the PCA dimensionality reduction data of the sensor subjected to the low-pass filtering treatment, calculating a second-order residual error square sum after the completion of piecewise linear fitting on different orders, finally finding out the position corresponding to the minimum value of the second-order difference as an elbow turning point, and determining the optimal piecewise number of the sensor data of each physical attribute; And screening the piecewise linear fitting result under the optimal piecewise number by a multi-sensor breakpoint fusion algorithm based on slope mutation calibration to obtain the integral degradation turning point of the turbofan engine.
6. The method for predicting the remaining life of a turbofan engine based on a two-stage degradation label of claim 5, wherein the step of screening the piecewise linear fitting result under the optimal piecewise number by a multi-sensor breakpoint fusion algorithm based on slope mutation calibration to obtain the degradation turning point of the turbofan engine as a whole comprises the steps of: Obtaining fitting break points of all segments based on a piecewise linear fitting result, screening effective break points in the fitting break points, and constructing a clustering feature vector based on a correlation coefficient, wherein the correlation coefficient is obtained by calculating the numerical value of the sensor data at the first moment and the last moment and the absolute value of the pearson correlation coefficient of the corresponding operation period; Classifying similar effective break points into a plurality of groups through K-means clustering, removing extreme values of each group of effective break points by adopting a 3 sigma criterion, and carrying out intra-group weighted fusion by taking a correlation coefficient as a weight; And if the number of the effective break points is less than the clustering number, directly adopting the weighted average of the correlation to obtain a fusion break point as a degradation turning point.
7. The turbofan engine remaining life prediction method based on the two-stage degradation label according to claim 6, wherein in the step of constructing the two-stage degradation label based on the degradation turning point in step S4, a stationary degradation label is constructed based on a linear attenuation model construction, and the nonlinear acceleration degradation label is constructed by introducing an acceleration attenuation ratio, and the constructed two-stage degradation label can be expressed as: Wherein, the Representing a degraded tag that is to be displayed, Indicating the total number of cycles of the turbofan engine during the observation period, The number of the current circulation of the turbofan engine is represented, Represents the degradation turning point of the turbofan engine as a whole, Indicating the acceleration damping rate.
8. The two-stage degradation tag-based turbofan engine remaining life prediction method according to claim 7, wherein the acceleration attenuation ratio introduced by the nonlinear acceleration degradation tag is obtained based on the steps of: Acquiring a first average degradation slope of stable degradation of the turbofan engine based on sensor data before the degradation turning point; Acquiring sensor data after the degradation turning point based on the degradation turning point to acquire a second average degradation slope of nonlinear acceleration degradation of the turbofan engine; Acquiring degradation slope multiplying power of each sensor in the turbofan engine based on the second average degradation slope and the first average degradation slope; Acquiring an average slope multiplying power of the turbofan engine based on the degradation slope multiplying power of each sensor, and expanding the average slope multiplying power to acquire an effective multiplying power range of the turbofan engine; and carrying out interval summarization based on the effective multiplying power ranges of all turbofan engines to obtain an acceleration damping multiplying power reference interval of the whole engine, and taking the value of the acceleration damping multiplying power reference interval to obtain the value of the acceleration damping multiplying power to be introduced.
9. The method for predicting the remaining life of a turbofan engine based on the two-stage degradation label according to claim 8, wherein in the step S5, in the step of constructing a TPRUL-Net prediction model integrating BiLSTM network and attention mechanism, the TPRUL-Net prediction model comprises a time sequence feature branch, a manual feature branch, a feature splicing layer and an output layer; The time sequence characteristic branch comprises a first input layer, a BiLSTM network, an attention mechanism module and a first linear layer; the first input layer is used for receiving the differential feature and the two-stage degradation label and processing the differential feature and the two-stage degradation label in a sliding time window to output a first feature tensor sequence with fixed step length; the BiLSTM network captures degradation information based on forward and backward dependency based on a first characteristic tensor sequence; the attention mechanism module generates attention weight in a time dimension based on the degradation information, highlights characteristic contribution of the degradation information at a key moment, and fuses the attention weight and the degradation information by adopting residual connection to generate time sequence characteristics; The first linear layer receives the time sequence characteristics and outputs the time sequence characteristics after flattening; The manual feature branch comprises a second input layer, a statistics module, a Z-Score normalization layer and a second linear layer; the second input layer is used for receiving sensor data and processing the sensor data in a sliding time window to output a second characteristic tensor sequence with fixed step length; the statistical module extracts statistical features based on the second feature tensor sequence, wherein the statistical features comprise an average value, a standard deviation and a degradation trend coefficient; the Z-Score normalization layer is used for receiving the statistical characteristics and performing Z-Score normalization; The second linear layer receives the statistical characteristics after Z-Score standardization, and outputs the statistical characteristics after flattening; The characteristic splicing layer receives the time sequence characteristics output by the time sequence characteristic branch and the statistical characteristics output by the manual characteristic branch and splices the time sequence characteristics and the statistical characteristics in characteristic dimension to obtain a double-branch fusion characteristic; and the output layer maps out a single non-negative real number based on the fusion characteristic and is used as a predicted value of the residual life of the turbofan engine.
10. The two-stage degradation label-based turbofan engine residual life prediction method according to claim 1 or 5 or 6, wherein the optimal number of segments is 5.

Description

Turbofan engine residual life prediction method based on two-stage degradation labels Technical Field The invention relates to the field of equipment fault prediction and health management, in particular to a turbofan engine residual life prediction method based on a two-stage degradation label. Background In recent years, prediction and health management (Prognostics AND HEALTH MANAGEMENT, PHM) has received a great deal of attention in both academia and industry as a key research area for ensuring the safety and reliability of complex engineering systems. The PHM has the core mission that the prediction of the residual service life (REMAINING USEFUL LIFE, RUL) of the equipment is accurately realized according to the real-time monitoring data of the running state of the equipment by means of intelligent analysis, so that scientific and reasonable basis is provided for preventive maintenance and state maintenance, and the loss and risk caused by the sudden failure of the equipment are effectively avoided. In the field of aviation, turbofan engines are the most critical core components of aircraft, and the performance degradation process has a direct and significant impact on flight safety and operation and maintenance costs. In the event of a severe failure of the engine during flight, it is highly likely to have catastrophic consequences. Thus, achieving accurate predictions of its RUL during engine operation has become a central issue in improving aviation safety levels, reducing maintenance costs, and optimizing operation and maintenance strategies. In the related research literature, the RUL prediction method for turbofan engines mainly covers a mechanism modeling-based method, a data driving-based method and a hybrid method. Based on the method of mechanism modeling, the life prediction is achieved by constructing a degradation model of thermodynamic and aerodynamic coupling processes mainly by deeply describing the physical mechanism of the engine. However, because the turbofan engine has extremely complex structure, the working mechanism of the turbofan engine presents high nonlinearity, and comprehensive and accurate modeling is difficult to perform, so that the method faces a plurality of limitations in practical engineering application, and the real degradation condition of the engine cannot be accurately reflected. With the continuous progress of sensor technology and the great improvement of computing power, the data driving method gradually becomes the main stream mode of the RUL prediction of the turbofan engine. The method uses statistical learning or deep learning models to develop and predict the future service life of the engine by deeply excavating degradation modes contained in the historical operation data of the engine. In recent years, recurrent Neural Networks (RNNs), long-short-term memory networks (LSTM), bi-directional LSTM (BiLSTM), convolutional Neural Networks (CNNs), and fusion structures thereof have exhibited outstanding performance in this field. For example, LSTM can effectively alleviate gradient vanishing problem and capture long-term dependency relationship in data, biLSTM further enhances prediction capability by modeling forward and backward sequences, CNN has strong feature extraction capability, and can realize joint modeling of space-time degradation features after combining with time sequence model. Meanwhile, the attention mechanism and the multi-scale modeling method are widely applied, so that the key degradation stage can be highlighted, the sensitivity of the model to different scale features is enhanced, and the prediction accuracy is improved. In order to better fit the actual laws of multistage, nonlinear degradation of turbofan engines, researchers have begun focusing their eyes on staged predictive models, which have also gradually evolved into important research trends in recent years. The model is usually based on degradation point detection, the life cycle of the engine is divided into a plurality of health stages, and a special prediction model is designed for different stages, so that the limitation that a single model is difficult to adapt to the full-cycle degradation rule of the engine is overcome. For example, part of researches adopt a two-stage division strategy, a healthy period and a degenerate period are identified by using variable point detection, then CNN-LSTM and LSTM models are used for modeling respectively, the research is carried out through a double-task LSTM network, the degenerate stage classification and RUL regression are carried out simultaneously, and fine-granularity degenerate modeling is realized based on seven-stage division, in addition, a general framework of classification-regression is constructed through the research, and continuous sensing data is divided into three states of new, medium and heavy based on membership functions so as to adapt to the requirements under different industrial scenes. In general, the stage