A method for fault diagnosis of aircraft engines
Optimizing the parameters of SVM multi-classifiers through sparse autoencoder and particle swarm algorithms, the problem of poor generalization of fault diagnosis methods that rely on physical models and expert knowledge in the prior art is solved, and a more accurate and stable aircraft engine fault prediction is achieved.
Patent Information
- Application Number
- CN202211479507.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-24
- Publication Date
- 2025-05-06
- Estimated Expiration
- 2042-11-24
AI Technical Summary
Existing aircraft engine fault diagnosis methods rely on physical models and expert prior knowledge, have poor generalization and are unable to effectively process nonlinear data, resulting in unstable fault prediction.
The data-driven fault diagnosis method is adopted to reduce the dimension and feature extraction of aircraft engine fault data through sparse autoencoder (SAE), and optimize the parameters of SVM multi-classifiers in combination with particle swarm algorithm to achieve fault type prediction.
Without the need for expert prior knowledge, it can effectively explore key parameters in the degradation process of equipment, improve the accuracy and stability of fault prediction, and reduce the economic burden caused by finding faults during the use of aircraft engines.
Smart Images
Figure CN115859177B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of aircraft engine fault diagnosis, and in particular relates to an aircraft engine fault diagnosis method. Background Art
[0002] An aircraft engine is a highly sophisticated system, most of whose components operate in extremely harsh environments such as high temperature, high pressure, vibration, and rotation. In addition, due to the large number of parts in each component, there are many failure modes, and even multiple fault types and combined failures may occur, which increases the difficulty of fault prediction.
[0003] Currently, the mainstream fault diagnosis solution is to use the state detection data of aircraft engine components collected by sensors to establish a data-driven fault diagnosis model to infer the health status of aircraft engines. The fault diagnosis problem of equipment is essentially a classification problem. This solution first pre-processes the collected raw sensor state monitoring data, and then determines the fault type through feature extraction and classification algorithm classification, which improves the work efficiency of equipment maintenance personnel and saves equipment maintenance costs.
[0004] The diagnostic method based on physical models relies heavily on a large amount of expert prior knowledge, with a single judgment basis and poor generalization. The diagnostic method based on statistical models cannot mine features and process nonlinear data. The prediction method based on data-driven models usually uses SVM (Support Vector Machine) as the classification module, but the parameters in SVM are uncertain. Manually adjusting the parameters will make the model classification results unstable. Summary of the invention
[0005] In order to solve the problems in the background, the present invention provides a method for diagnosing a fault of an aircraft engine, comprising:
[0006] S1: obtaining original aircraft engine fault data with label information and dividing the original aircraft engine fault data into a training set and a test set; the label information includes: aircraft engine fault category;
[0007] S2: Input the training set into the sparse autoencoder SAE and train the sparse autoencoder SAE through the back-propagation mechanism;
[0008] S3: Input the samples in the test set into the trained sparse autoencoder SAE to obtain a sparse sample vector through hidden layer dimensionality reduction;
[0009] S4: input the sparse sample vector into the SVM multi-classifier to output the prediction result of the sparse sample vector, and use the particle swarm algorithm to calculate the optimal parameters of the SVM multi-classifier according to the prediction result of the sparse sample vector and the label information of the sparse sample vector;
[0010] S5 obtains target aircraft engine fault data and calculates a sparse sample vector of the target aircraft engine fault data, and inputs the sparse sample vector of the target aircraft engine fault data into an SVM multi-classifier to output a fault category of the target aircraft engine fault data.
[0011] The present invention has at least the following beneficial effects
[0012] The present invention predicts aircraft engine failures based on a data-driven model. Its greatest advantage is that it does not require expert prior knowledge of physics, nor does it require understanding of the mechanism of the gradual failure process of the equipment and its complex internal structure. It only uses the particle swarm algorithm to mine key parameters that can effectively represent the equipment degradation process from the collected monitoring data, and uses the model to predict possible failure types, greatly reducing the huge economic burden caused by finding faults during the use of aircraft engines.
[0013] Instruction Manual
[0014] Figure 1 is a flowchart of the method of the present invention;
[0015] Figure 2 This is a diagram of the sparse autoencoder structure of the present invention. Specific implementation methods
[0016] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.
[0017] See also Figure 1 The present invention provides a method for diagnosing a fault of an aircraft engine, comprising:
[0018] S1: obtaining original aircraft engine fault data with label information and dividing the original aircraft engine fault data into a training set and a test set; the label information includes: aircraft engine fault category;
[0019] The data source can be obtained by using the fault data of a certain type of aircraft engine simulated by the simulation platform TE developed by Eastman Company of the United States. The data that needs to be obtained here include the parameters of each sensor under the operation state of the aircraft engine, and the type of engine fault caused, such as normal state, fan fault, compressor fault, high-pressure turbine fault, and low-pressure turbine fault.
[0020] The acquired aircraft engine fault source data cannot be directly input into the model, and the data needs to be standardized. For example, the fault type cannot be directly processed by the computer and needs to be standardized.
[0021] The normalized data needs to be stored in the local database, and the data is uniformly stored and named through the table structure. Storing the data in the database can also improve the efficiency of data retrieval and reuse and the mapping of relationships between tables.
[0022] To facilitate model training, different fault types are labeled accordingly. For example, the operating status is 0, the fan fault is 1, the compressor fault is 2, the high-pressure turbine fault is 3, and the low-pressure turbine fault is 4.
[0023] Common methods for data normalization include Min-Max normalization and Z-score normalization. Although different normalization methods have different calculation methods, the actual effects are almost the same. In order to simplify the calculation amount of the model and improve the prediction accuracy, this article uses the Min-Max normalization method to scale the data to the interval [0,1] through linear transformation. The Min-Max normalization formula is as follows:
[0024]
[0025] in, Represents the state detection data of the kth column at the i-th cycle in the data set, yes The value obtained after Min-Max normalization, min(x k ) represents the minimum value in the kth column data, max(x k ) represents the maximum value in the kth column.
[0026] Both model training and validation require a certain amount of data. This paper adopts the conventional 5-fold cross-validation method, and the ratio of training set to test set for each fault type is 4:1.
[0027] S2: Input the training set into the Sparse Auto Encoder (SAE) and train the Sparse Auto Encoder (SAE) through the back-propagation mechanism;
[0028] In the basic AE (Auto Encoder) structure, the number of neurons in the hidden layer is usually less than that in the input layer. This is to retain the valid information in the original data. However, an unreasonable setting of the difference in the number of nodes between the two may lead to the loss of valid information. On the contrary, once the number of neurons in the hidden layer is greater than or equal to the number of neurons in the input layer, the purpose of data dimensionality reduction cannot be achieved, and the output information may contain redundant information.
[0029] In order to enable the autoencoder to adaptively adjust the number of neurons required according to the characteristics of the data, the extraction of effective information can be constrained by adding sparsity constraints to the hidden layer. The sparsity constraint is to add sparse regularization constraints to the neurons of AE, so that the hidden layer can use the least number of neurons to extract effective information from the original data. SAE does not need to reduce the number of neurons in the hidden layer. By adding sparsity constraints on the basis of the neural network loss function, the weight parameters of the neurons can be adjusted, that is, neurons with output close to 1 in the hidden layer are activated, and neurons with output close to 0 are inhibited. Figure 2 shown.
[0030] The information of each neuron in the hidden layer is interrelated, so the activation degree of each neuron is completely determined by the characteristics of the input data. The activation degree of the neuron is represented by ρ, and the average activation degree is The calculation formula is:
[0031]
[0032] Where N is the number of samples, is the activity of the jth neuron in the i-th group. Then, in order to achieve the sparsity constraint, it is necessary to add the average activation function and a penalty term different from the sparsity regularization to the objective function of SAE. Relative entropy is a method to measure the difference between two distributions, so the relative entropy function is used as the penalty term. The formula is as follows:
[0033]
[0034] At present, in order to better adapt to different types of data, the variants of autoencoders are constantly being expanded. Commonly used autoencoder variants include sparse autoencoders, variational autoencoders, and denoising autoencoders. According to the high-dimensional and nonlinear characteristics of aircraft engine condition monitoring data, this paper uses sparse autoencoders.
[0035] S3: Input the samples in the test set into the trained sparse autoencoder SAE to obtain a sparse sample vector through hidden layer dimensionality reduction;
[0036] S4: Input the sparse sample vector into the SVM (Support Vector Machine) multi-classifier to output the prediction result of the sparse sample vector, and use the particle swarm algorithm to calculate the optimal parameters of the SVM multi-classifier according to the prediction result of the sparse sample vector and the label information of the sparse sample vector;
[0037] S41: Set the number of iterations, spatial dimension and number of particles of the particle swarm, and randomly initialize the speed and position of each particle;
[0038] In the present invention, the spatial dimension is two-dimensional, and the component of each dimension of the particle in the two-dimensional space corresponds to the kernel parameter C and the penalty parameter g of the SVM multi-classifier;
[0039] S42: Using the component of the particle's position in each dimension of space as a parameter of the SVM multi-classifier, and using the Euclidean distance between the prediction result of the SVM multi-classifier for the sparse sample vector and the label information of the sparse sample vector under the parameter as the particle's fitness value, updating the particle's optimal fitness value and the group's optimal fitness value; wherein, the smaller the Euclidean distance between the prediction result of the SVM multi-classifier for the sparse sample vector and the label information of the sparse sample vector, the higher the particle's fitness value, and vice versa.
[0040] If the fitness value of the current particle is greater than the best historical fitness value of the particle, the fitness value of the current particle is used to replace the best historical fitness value of the particle;
[0041] If the historical best fitness value of the current particle is greater than the historical best fitness value of the group particles, the historical best fitness value of the current particle is used to replace the historical best fitness value of the group particles;
[0042] S43: Update the velocity and position of the particle;
[0043]
[0044]
[0045]
[0046] Among them, w t represents the nonlinear inertia weight, η represents the curvature adjustment parameter, and α represents the nonlinear inertia weight w t The amplitude parameter of the present invention is 0.6, b represents the nonlinear inertia weight w t The horizontal displacement parameter of the present invention is 0.3, t represents the number of iterations, V id represents the velocity of the ith particle in the dth dimension, T max represents the maximum number of iterations, represents the solution corresponding to the historical best fitness value of the particle at the tth iteration, represents the solution corresponding to the historical best fitness value of the group particles at the tth iteration, X id represents the position of the ith particle in the dth dimension, γ i is a random number between [0,1], c1 and c2 represent learning factors, and their values are non-negative constants.
[0047] S44: Repeat steps S43-S44 until a preset number of iterations is reached, and output the parameters of the SVM multi-classifier corresponding to the historical best fitness value of the current population particles to obtain the optimal parameters of the SVM multi-classifier.
[0048] S5 obtains target aircraft engine fault data and calculates a sparse sample vector of the target aircraft engine fault data, and inputs the sparse sample vector of the target aircraft engine fault data into an SVM multi-classifier to output a fault category of the target aircraft engine fault data.
[0049] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.
Claims
1. A method for diagnosing faults of an aircraft engine, characterized in that: include: S1: Obtain original aircraft engine fault data with label information and divide the original aircraft engine fault data into a training set and a test set; The label information includes: the fault category of the aircraft engine; S2: Input the training set into the sparse autoencoder SAE and train the sparse autoencoder SAE through the back-propagation mechanism; S3: Input the samples in the test set into the trained sparse autoencoder SAE to obtain a sparse sample vector through hidden layer dimensionality reduction; S4: input the sparse sample vector into the SVM multi-classifier to output the prediction result of the sparse sample vector, and use the particle swarm algorithm to calculate the optimal parameters of the SVM multi-classifier according to the prediction result of the sparse sample vector and the label information of the sparse sample vector; The method of calculating the optimal parameters of the SVM multi-classifier using the particle swarm algorithm according to the prediction results of the sparse sample vector and the label information of the sparse sample vector includes: S41: Set the number of iterations, spatial dimension and number of particles of the particle swarm, and randomly initialize the speed and position of each particle; S42: taking the component of the particle position in each dimension of the space as the parameter of the SVM multi-classifier, and taking the prediction result of the SVM multi-classifier on the sparse sample vector under the parameter and the Euclidean distance of the label information of the sparse sample vector as the fitness value of the particle, and updating the optimal fitness value of the particle and the optimal fitness value of the group of particles; S43: Update the velocity and position of the particle; S44: Repeat steps S43-S44 until a preset number of iterations is reached, and output the parameters of the SVM multi-classifier corresponding to the historical best fitness value of the current population particles to obtain the optimal parameters of the SVM multi-classifier; S5 obtains target aircraft engine fault data and calculates a sparse sample vector of the target aircraft engine fault data, and inputs the sparse sample vector of the target aircraft engine fault data into an SVM multi-classifier to output a fault category of the target aircraft engine fault data.
2. The method for diagnosing a fault of an aircraft engine according to claim 1, characterized in that: If the fitness value of the current particle is greater than the best historical fitness value of the particle, the fitness value of the current particle is used to replace the best historical fitness value of the particle; If the historical best fitness value of the current particle is greater than the historical best fitness value of the group particles, the historical best fitness value of the current particle is used to replace the historical best fitness value of the group particles.
3. The method for diagnosing a fault of an aircraft engine according to claim 1, characterized in that: The updating of particle speed and position includes: In the formula, w t represents the nonlinear inertia weight, η represents the curvature adjustment parameter, and α represents the nonlinear inertia weight w t The amplitude parameter, b represents the nonlinear inertia weight w t The horizontal displacement parameter, t represents the number of iterations, V id represents the velocity of the ith particle in the dth dimension, T max represents the maximum number of iterations, represents the solution corresponding to the historical best fitness value of the particle at the tth iteration, represents the solution corresponding to the historical best fitness value of the group particles at the tth iteration, X id represents the position of the ith particle in the dth dimension, γ i is a random number between [0,1], c1 and c2 represent learning factors, and their values are non-negative constants.