Aero-engine health diagnosis method based on unbalanced data set
Through adaptive sparse coding and one-dimensional convolutional neural network combined with phase space reconstruction and adaptive synthesis sampling technology, a balanced data set is generated and a sparse support vector data description and fuzzy neural network model is constructed, which solves the problems of excessive parameters and data imbalance in aircraft engine health diagnosis, and improves diagnostic accuracy and model generalization capabilities.
Patent Information
- Application Number
- CN202510897946.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-01
- Publication Date
- 2025-08-01
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
There are problems in the existing aircraft engine health diagnosis methods such as excessive parameters, unbalanced data and poor generalization capabilities of diagnostic models, resulting in low accuracy of classification models.
Adaptive sparse coding technology is used to perform feature extraction and dimensionality reduction processing, and a normal operating state model is built with one-dimensional convolutional neural network. A balanced data set is generated through phase space reconstruction and adaptive synthesis sampling technology to build a single-class fault diagnosis model for sparse support vector data description and a multi-classified fault classification model for fuzzy neural networks.
Improves the accuracy and efficiency of engine health diagnosis, enables rapid detection and classification of faults, optimizes maintenance strategies, extends engine life and reduces maintenance costs.
Smart Images
Figure CN120408419A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of aero - engine health diagnosis, and particularly relates to aero - engine health diagnosis method based on an imbalanced data set. Background Art
[0002] The health diagnosis of aero - engines is of great importance, which is significant for ensuring flight safety, improving combat effectiveness, and optimizing maintenance strategies. As the core component of an aircraft, the health condition of the engine directly determines the performance and reliability of the aircraft. By real - time monitoring and analyzing various parameters of the engine, potential faults can be detected early, flight accidents caused by mechanical failures can be avoided, and flight safety can be significantly improved. At the same time, by long - term accumulating and analyzing the operation data of the engine, the usage and maintenance strategies of the engine can be optimized, the life of the engine can be extended, the maintenance cost can be reduced, and the operation efficiency of the entire aviation equipment can be improved.
[0003] Traditional diagnosis methods often rely on expert experience, and judge the health state of the engine by manually analyzing the engine parameter signals. To overcome the problems of strong subjectivity, low efficiency, and high misdiagnosis rate of this method, researchers have gradually started to explore engine diagnosis methods based on deep learning and data - driven approaches. However, there are still some problems at present: First, most of the current research is carried out based on balanced data, while in actual engineering applications, there are serious data skew problems in engine data, that is, the number of fault samples is much smaller than the number of normal samples, and the characteristics of faults are incomplete. Second, there are many parameters of the engine, which increases the difficulty of data analysis and may lead to over - fitting phenomena. These problems will result in poor generalization ability of the classification model and affect the accuracy of the model. Summary of the Invention
[0004] Aiming at the above - mentioned deficiencies in the prior art, the aero - engine health diagnosis method based on an imbalanced data set provided by the present invention solves the problems existing in the existing engine health diagnosis algorithms, such as too many parameters, data imbalance, and poor generalization ability of the diagnosis model.
[0005] To achieve the above - mentioned invention purpose, the technical solution adopted by the present invention is: An aero - engine health diagnosis method based on an imbalanced data set, comprising the following steps: S1: Obtain the original data of the aero - engine, and adopt the adaptive sparse coding technology to perform feature extraction and dimensionality reduction processing on the original data to obtain a dimensionality - reduced data set containing significant feature parameters; S2: Construct an aero - engine normal operation state model based on a one - dimensional convolutional neural network; S3: Based on the aero - engine normal operation state model and phase - space reconstruction technology, expand the number of features of the dimensionality - reduced data set; S4: Augment the number of groups of minority-class samples in the dataset after expanding the feature number based on the adaptive synthetic sampling technique to generate a balanced dataset; S5: Construct a one-class fault diagnosis model based on sparse support vector data description and a multi-class fault classification model based on a fuzzy neural network; S6: Input the balanced dataset into the one-class fault diagnosis model and the multi-class fault classification model to perform fault diagnosis and fault classification on the aero-engine.
[0006] Furthermore, the S1 includes the following sub-steps: S11: Obtain the original data of the aero-engine containing vibration signals; S12: Construct an initial dictionary according to the original data of the aero-engine; S13: Through constructing an objective function and minimizing the data reconstruction error, perform sparse coding. The formula is:
[0007] where, represents minimization, is the input signal, is the dictionary matrix, is the sparse coefficient vector, is the reconstruction error, is the sparse regularization coefficient, is norm, is norm; S14: Perform sparse adaption adjustment by dynamically adjusting the sparse regularization coefficient; S15: After dimensionality reduction, retain the high-pressure rotor speed, low-pressure rotor speed, normal overload, high-pressure compressor guide vane angle, low-pressure compressor guide vane angle, afterburner nozzle position, longitudinal vibration value, lateral vibration value, and normal vibration value of the engine as the first feature parameters to obtain a dimensionality-reduced dataset containing significant feature parameters.
[0008] Furthermore, the normal operating state model of the aero-engine in the S2 includes an input layer, a convolutional layer, a pooling layer, and a fully connected layer. The activation function of the convolutional layer is the ReLU activation function, and a Dropout layer is added to prevent overfitting.
[0009] Furthermore, the S3 includes the following sub-steps: S31: Select a partial sample set in the dimensionality-reduced dataset as the original data and input it into the normal operating state model of the aero-engine, and calculate the difference between the output of the normal operating state model of the aero-engine and all sample sets in the dimensionality-reduced dataset. Take the difference as the second feature parameter; S32: Reconstruct the phase space of the high-pressure rotor speed of the engine together with the longitudinal vibration value, lateral vibration value, and normal vibration value to obtain a third characteristic parameter. The formula is as follows:
[0010]
[0011]
[0012]
[0013] where, is the characteristic parameter after the phase space reconstruction of the high-pressure rotor speed, is the high-pressure rotor speed, is the characteristic parameter after the phase space reconstruction of the longitudinal vibration value, is the longitudinal vibration value, is the characteristic parameter after the phase space reconstruction of the lateral vibration value, is the lateral vibration value, is the characteristic parameter after the phase space reconstruction of the normal vibration value, is the normal vibration value, is the sample serial number; S33: Introduce the characteristic parameter of the engine maintenance duration to obtain a dataset with expanded characteristic numbers. The engine maintenance duration is:
[0014] where, is the engine maintenance duration, is the time since the last engine replacement, is the time since the last repair.
[0015] Furthermore, the S4 includes the following sub-steps: S41: Calculate the distance between the minority class samples and all majority class samples, and calculate the difficulty coefficient of the minority class samples , and the formula is:
[0016] where, is the number of the nearest neighbors of the majority class samples, is the number of the nearest neighbors of the minority class samples; S42: According to the difficulty coefficient of the minority class samples, assign a sample weight to each minority class sample. The formula is:
[0017] where, is the sample weight, is the difficulty coefficient of the th minority class sample, is the number of minority class samples, is the total number of samples to be generated; S43: For each minority class sample, randomly select its nearest neighbor to generate a synthetic sample, and the formula is:
[0018] where, [[ID=1⑤]]is the synthetic sample, is the minority class sample, is the nearest neighbor of the minority class sample, is a random number; S44: Add the generated synthetic samples to the dataset after feature number augmentation to obtain a balanced dataset.
[0019] Furthermore, the objective function of the one-class fault diagnosis model based on sparse support vector data description in the S5 is:
[0020] where, is the center of the hypersphere, is the Lagrange multiplier, is the input vector, is the th input vector, is the total number of samples; By calculating the distance from the synthetic sample to the center of the sphere , fault diagnosis is realized, and the distance is calculated by the formula: .
[0021] Furthermore, the multi-class fault classification model based on fuzzy neural network in the S5 includes an input layer, a fuzzification layer, a rule layer, a normalization layer and an output layer, and specifically includes the following sub-steps: a1: Use the trapezoidal function as the membership function, set parameters based on the historical data statistical quantile, and fuzzify the input data to generate membership function values; a2: Based on the fuzzy C-means clustering analysis algorithm, set the number of clusters and the fuzzy index according to the number of fault types, and generate corresponding fuzzy rules for each cluster; a3: Based on the membership function values and fuzzy rules, calculate the rule activation strength and output the activation strength vector; a4: Perform normalization processing on the activation strength vector, and input the normalized weight vector into the output layer for weighted summation; a5: After defuzzification decision-making processing, discrete fault types are output.
[0022] Another technical solution adopted by the present invention is: a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the aviation engine health diagnosis method based on an imbalanced data set disclosed above is implemented.
[0023] The beneficial effects of the present invention are: (1) Aiming at the problems of excessive engine parameters and weak characterization ability, the present invention adopts the ASC technology to extract the main features and reduce the data dimension. At the same time, custom feature parameters such as vibration values and high-pressure rotor speeds after phase space reconstruction are added to achieve efficient feature extraction and enhance the characterization ability of complex signals.
[0024] (2) Aiming at the problem that the number of fault samples is much less than that of normal samples, the present invention adopts the ADASYN method to generate minority class sample data with higher representativeness, thereby balancing the data set. And the S-SVDD algorithm is adopted to improve the accuracy and efficiency of anomaly detection. Description of the Drawings
[0025] Figure 1 It is a flowchart of an aviation engine health diagnosis method based on an imbalanced data set.
[0026] Figure 2 It is a graph showing the change in the deviation degree between the predicted value and the true value based on CNN regression.
[0027] Figure 3 It is a comparison graph before and after the expansion of the number of minority class sample groups.
[0028] Figure 4 It is a comparison graph of the positive class rates of fault detection between SVDD and S-SVDD.
[0029] Figure 5 It is a comparison graph of the accuracy rates of the test sets of the FNN classification model. Detailed Embodiments
[0030] The present invention will be further described below with reference to the drawings and specific embodiments.
[0031] To solve the problems of excessive parameters, data imbalance, and poor generalization ability of the existing engine health diagnosis algorithms, the present invention uses the Adaptive Sparse Coding (ASC) technology to extract the significant features of the data, reduce the dimension of the data, and avoid the influence of excessive parameters on the diagnosis. Then, a one-dimensional convolutional neural network (1D-CNN) is used to establish a model of the normal operating state of the engine and new feature parameters are added. Next, the Adaptive Synthetic Sampling (ADASYN) technology is used to process the minority class samples to generate synthetic samples to balance the data set. Finally, a single-class fault diagnosis model Sparse Support Vector Data Description (S-SVDD) and a multi-class fault classification model Fuzzy Neural Network (FNN) are established to achieve the health diagnosis and fault classification of the engine. The specific steps are as follows.
[0032] As Figure 1 shown, aero-engine health diagnosis method based on an imbalanced data set, comprising the following steps: S1: Obtain the original data of the aero-engine and perform feature extraction and dimensionality reduction processing on the original data by using the adaptive sparse coding technology to obtain a dimensionality-reduced data set containing significant feature parameters; S2: Construct a model of the normal operating state of the aero-engine based on the one-dimensional convolutional neural network; S3: Expand the number of features of the dimensionality-reduced data set based on the model of the normal operating state of the aero-engine and the phase space reconstruction technology; S4: Expand the number of groups of minority class samples in the data set after expanding the number of features based on the adaptive synthetic sampling technology to generate a balanced data set; S5: Construct a single-class fault diagnosis model based on sparse support vector data description and a multi-class fault classification model based on fuzzy neural network; S6: Input the balanced data set into the single-class fault diagnosis model and the multi-class fault classification model to perform fault diagnosis and fault classification on the aero-engine.
[0033] The S1 includes the following sub-steps: S11: Obtain the original data of the aero-engine containing vibration signals; In this embodiment, the data used comes from the ground data of a certain type of aero-engine, including vibration signals collected from bearings in different health states under time-varying rotational speed conditions. It includes a total of 6 data sets, namely D1, D2, D3, D4, D5, and D6. Among them, D1 is unbalanced vibration data, including 2 samples; D2 is misalignment vibration data, including 2 samples; D3 is component looseness data, including 4 samples; D4 is surge data, including 3 samples; D5 is normal data, including 80 samples for training; D6 is data mixed with normal and fault states, including 4 normal samples and 1 sample for each of the 4 types of faults. Each sample has approximately 1000 data points, and each point has 126 parameters; S12: Construct an initial dictionary based on the original aero-engine data; Select the initial dictionary and input the data in the data set, with the 126 engine parameters as the initial dictionary , where, is the sparse coefficient vector for each sample, is the sparse coefficient matrix, is the number of basis vectors, is the number of samples.
[0034] S13: Through constructing an objective function to minimize the data reconstruction error for sparse coding, the formula is:
[0035] where, represents minimization, is the input signal, is the dictionary matrix, is the sparse coefficient vector, is the reconstruction error, is the sparse regularization coefficient, is norm, is norm; S14: Through dynamically adjusting the sparse regularization coefficient for sparse adaptability adjustment, it is verified that is set to 0.05; S15: After dimensionality reduction, retain the high-pressure rotor speed, low-pressure rotor speed, normal overload, high-pressure compressor guide vane angle, low-pressure compressor guide vane angle, tail nozzle position, longitudinal vibration value, lateral vibration value, and normal vibration value of the engine as the first characteristic parameters to obtain a dimensionality-reduced data set containing significant characteristic parameters, and the processed data set is named D11 - D16.
[0036] The normal operating state model of the aero-engine in S2 includes an input layer, a convolutional layer, a pooling layer, and a fully connected layer. The activation function of the convolutional layer is the ReLU activation function, and a Dropout layer is added to prevent overfitting. The learning rate is set to 0.004, the batch size is set to 40, and the number of epochs is set to 20. Figure 2 It is the deviation degree between the predicted value based on CNN regression and the true value.
[0037] S3 in the above includes the following sub-steps: S31: Select a part of the sample set (D15) from the dimensionality-reduced dataset as the original data and input it into the normal operating state model of the aero-engine. Then, calculate the difference between the output of the normal operating state model of the aero-engine and all sample sets (D11 - D16) in the dimensionality-reduced dataset, and use the difference as the second characteristic parameter. S32: Perform phase space reconstruction on the high-pressure rotor speed, longitudinal vibration value, lateral vibration value, and normal vibration value of the engine to obtain the third characteristic parameter. The formula is:
[0038]
[0039]
[0040]
[0041] Among them, is the characteristic parameter after phase space reconstruction of the high-pressure rotor speed, is the high-pressure rotor speed, is the characteristic parameter after phase space reconstruction of the longitudinal vibration value, is the longitudinal vibration value, is the characteristic parameter after phase space reconstruction of the lateral vibration value, is the lateral vibration value, is the characteristic parameter after phase space reconstruction of the normal vibration value, is the normal vibration value, is the sample serial number; S33: Introduce the characteristic parameter of the engine maintenance duration to obtain the dataset with expanded characteristic numbers, denoted as D21 - D26. The engine maintenance duration is:
[0042] Among them, is the engine maintenance duration, is the time since the last engine replacement, is the time since the last repair.
[0043] S4 in the above includes the following sub-steps: S41: Calculate the distances between the minority class samples and all majority class samples, and calculate the difficulty coefficient of the minority class samples , the formula is:
[0044] where, is the number of the nearest neighbors of the majority class samples, is the number of the nearest neighbors of the minority class samples; If is larger, it indicates that the sample is near the classification boundary and the classification is difficult; if is smaller, it indicates that the sample is far from the classification boundary and the classification is relatively easy; S42: According to the difficulty coefficient of the minority class samples, assign and generate sample weights for each minority class sample, the formula is:
[0045] where, is the sample weight, is the difficulty coefficient of the th minority class sample, is the number of the minority class samples, is the total number of samples to be generated; S43: For each minority class sample, randomly select its nearest neighbor to generate a synthetic sample, the formula is:
[0046] where, is the synthetic sample, is the minority class sample, is the nearest neighbor of the minority class sample, is a random number; S44: Add the generated synthetic samples to the dataset after feature number expansion to obtain a balanced dataset, and name the new dataset D31 - D36, Figure 3 is the comparison of the number of data groups before and after expansion.
[0047] The objective function of the one - class fault diagnosis model based on sparse support vector data description in the above - mentioned S5 is:
[0048] where, is the center of the hypersphere, is the Lagrange multiplier, is the input vector, is the th input vector, is the total number of samples; By calculating the synthetic sample Distance to the center of the sphere to achieve fault diagnosis. If the distance exceeds the radius of the sphere, the sample is determined to be a faulty sample and an alarm is triggered; otherwise, it is marked as normal. The distance The calculation formula is as follows: .
[0049] The multi-class fault classification model based on the fuzzy neural network in S5 includes an input layer, a fuzzification layer, a rule layer, a normalization layer, and an output layer, and specifically includes the following sub-steps: a1: Use the trapezoidal function as the membership function, set parameters based on the historical data statistical quantiles, and fuzzify the input data to generate membership function values; a2: Based on the fuzzy C-means clustering analysis algorithm, set the number of clusters and the fuzzy exponent according to the number of fault types, and generate corresponding fuzzy rules for each cluster; a3: Based on the membership function values and fuzzy rules, calculate the rule activation strength and output the activation strength vector; a4: Perform normalization processing on the activation strength vector, and input the normalized weight vector into the output layer for weighted summation; a5: After defuzzification decision processing, output discrete fault types.
[0050] Input the dataset D36 into the single-class fault diagnosis model and the multi-class fault classification model to test the fault diagnosis and fault classification effects of the models. Attached Figure 4 And attached Figure 5 Are respectively the positive class rate of fault detection of S-SVDD and the accuracy rate of the test set of the FNN classification model.
[0051] An aero-engine health diagnosis method based on an imbalanced dataset provided by the present invention can eliminate the problems of poor model generalization ability and low diagnosis accuracy rate caused by excessive engine parameters, weak characterization ability, and data imbalance, enabling faults to be detected and classified more quickly, which is conducive to quickly locating the fault location so as to take measures.
[0052] The embodiment of the present invention also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned aero-engine health diagnosis method based on an imbalanced dataset.
[0053] Those of ordinary skill in the art will realize that the embodiments described herein are provided to assist the reader in understanding the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention based on these technical revelations disclosed in the present invention, and these deformations and combinations are still within the scope of protection of the invention.
Claims
1. An aero-engine health diagnosis method based on an imbalanced dataset, characterized in that It includes the following steps: S1: Obtain the original data of the aero-engine, and perform feature extraction and dimensionality reduction processing on the original data by using the adaptive sparse coding technology to obtain a dimensionality reduction data set containing significant feature parameters; S2: Build a normal operation state model of the aero-engine based on the one-dimensional convolutional neural network; S3: Based on the normal operation state model of the aero-engine and the phase space reconstruction technology, expand the number of features of the dimensionality reduction data set; S4: Based on the adaptive synthetic sampling technology, expand the number of groups of minority class samples in the data set after expanding the number of features to generate a balanced data set; S5: Build a one-class fault diagnosis model based on sparse support vector data description and a multi-class fault classification model based on the fuzzy neural network; S6: Input the balanced data set into the one-class fault diagnosis model and the multi-class fault classification model to perform fault diagnosis and fault classification on the aero-engine.
2. The aero-engine health diagnosis method based on an imbalanced data set according to claim 1, wherein, The S1 includes the following sub-steps: S11: Obtain the original data of the aero-engine containing vibration signals; S12: Build an initial dictionary according to the original data of the aero-engine; S13: By building an objective function and minimizing the data reconstruction error, perform sparse coding. The formula is: Among them, represents minimization, is the input signal, is the dictionary matrix, is the sparse coefficient vector, is the reconstruction error, is the sparse regularization coefficient, is the norm, is the norm; S14: Perform sparse adaptability adjustment by dynamically adjusting the sparse regularization coefficient; S15: After dimensionality reduction, retain the high-pressure rotor speed, low-pressure rotor speed, normal overload, high-pressure compressor guide vane angle, low-pressure compressor guide vane angle, tail nozzle position, longitudinal vibration value, lateral vibration value and normal vibration value of the engine as the first feature parameters to obtain a dimensionality reduction data set containing significant feature parameters.
3. The aero-engine health diagnosis method based on an imbalanced dataset according to claim 1, wherein In the S2, the normal operation state model of the aero-engine includes an input layer, a convolutional layer, a pooling layer and a fully connected layer. The activation function of the convolutional layer is the ReLU activation function, and a Dropout layer is added to prevent overfitting.
4. The aero-engine health diagnosis method based on an imbalanced data set according to claim 2, characterized in that The S3 includes the following sub-steps: S31: Select a part of the sample set in the dimensionality reduction data set as the original data and input it into the normal operation state model of the aero-engine, and calculate the difference between the output of the normal operation state model of the aero-engine and all sample sets in the dimensionality reduction data set, and use the difference as the second feature parameter; S32: Perform phase space reconstruction on the high-pressure rotor speed, longitudinal vibration value, lateral vibration value and normal vibration value of the engine to obtain the third feature parameter. The formula is: Among them, is the characteristic parameter after the phase space reconstruction of the high-pressure rotor speed, is the high-pressure rotor speed, is the characteristic parameter after the phase space reconstruction of the longitudinal vibration value, is the longitudinal vibration value, is the characteristic parameter after the phase space reconstruction of the lateral vibration value, is the lateral vibration value, is the characteristic parameter after the phase space reconstruction of the normal vibration value, is the normal vibration value, is the sample serial number; S33: Introduce the engine maintenance duration feature parameter to obtain a data set after expanding the number of features. The engine maintenance duration is: Among them, is the engine maintenance duration, is the time since the last engine replacement, is the time since the last repair.
5. The aero-engine health diagnosis method based on an imbalanced data set according to claim 4, characterized in that The S4 includes the following sub-steps: S41: Calculate the distances between the minority class samples and all the majority class samples, and calculate the difficulty coefficient of the minority class samples , and the formula is: Among them, is the number of the nearest neighbors of the majority class samples, is the number of the nearest neighbors of the minority class samples; S42: According to the difficulty coefficient of the minority class samples , assign and generate sample weights for each minority class sample, and the formula is: Among them, is the sample weight, is the difficulty coefficient of the th minority-class sample, is the number of minority-class samples, is the total number of samples to be generated; S43: For each minority class sample, randomly select its nearest neighbor to generate a synthetic sample. The formula is: Among them, is a synthetic sample, is a minority sample, is the nearest neighbor of the minority sample, is a random number; S44: Add the generated synthetic samples to the data set after expanding the number of features to obtain a balanced data set.
6. The aero-engine health diagnosis method based on an imbalanced data set according to claim 5, characterized in that, The objective function of the one-class fault diagnosis model based on sparse support vector data description in the S5 is: Among them, is the center of the hypersphere, is the Lagrange multiplier, is the input vector, is the th input vector, is the total number of samples; By calculating the synthetic samples to the center of the sphere the distance is obtained for fault diagnosis, and the distance is calculated by the formula: 。 7. The aero-engine health diagnosis method based on an imbalanced data set according to claim 5, wherein The multi-class fault classification model based on the fuzzy neural network in the S5 includes an input layer, a fuzzification layer, a rule layer, a normalization layer and an output layer. Specifically, it includes the following sub-steps: a1: Use the trapezoidal function as the membership function, set parameters based on the statistical quantiles of historical data, and fuzzify the input data to generate membership function values; a2: Based on the fuzzy C-means clustering analysis algorithm, set the number of clusters and the fuzzy index according to the number of fault types, and generate corresponding fuzzy rules for each cluster; a3: Based on the membership function values and fuzzy rules, calculate the rule activation strength and output the activation strength vector; a4: Normalize the activation strength vector, and input the normalized weight vector into the output layer for weighted summation; a5: After defuzzification decision processing, output discrete fault types.
8. A computer-readable storage medium, characterized in that, For storing a computer program; wherein, when the computer program is executed by a processor, it implements the aero-engine health diagnosis method based on an imbalanced dataset according to any one of claims 1-7.