CN116070159A - Method and device for realizing fault diagnosis of main power supply system of airplane under unbalanced sample - Google Patents
Method and device for realizing fault diagnosis of main power supply system of airplane under unbalanced sample Download PDFInfo
- Publication number
- CN116070159A CN116070159A CN202310002066.3A CN202310002066A CN116070159A CN 116070159 A CN116070159 A CN 116070159A CN 202310002066 A CN202310002066 A CN 202310002066A CN 116070159 A CN116070159 A CN 116070159A
- Authority
- CN
- China
- Prior art keywords
- fault
- sample
- main power
- countermeasure network
- optimal transmission
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000003745 diagnosis Methods 0.000 title claims abstract description 63
- 238000000034 method Methods 0.000 title claims abstract description 57
- 230000005540 biological transmission Effects 0.000 claims abstract description 79
- 238000012549 training Methods 0.000 claims abstract description 66
- 238000003066 decision tree Methods 0.000 claims abstract description 57
- 238000012360 testing method Methods 0.000 claims abstract description 30
- 238000009826 distribution Methods 0.000 claims description 32
- 238000004422 calculation algorithm Methods 0.000 claims description 13
- 238000010276 construction Methods 0.000 claims description 10
- 238000012545 processing Methods 0.000 claims description 9
- 230000006870 function Effects 0.000 description 20
- 230000008569 process Effects 0.000 description 18
- 238000004804 winding Methods 0.000 description 14
- 230000000694 effects Effects 0.000 description 11
- 238000013145 classification model Methods 0.000 description 10
- 238000004088 simulation Methods 0.000 description 10
- 238000005457 optimization Methods 0.000 description 5
- 230000008859 change Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 238000012952 Resampling Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 238000013138 pruning Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 230000008034 disappearance Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000016507 interphase Effects 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 239000002689 soil Substances 0.000 description 2
- 239000000243 solution Substances 0.000 description 2
- 230000001131 transforming effect Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000009472 formulation Methods 0.000 description 1
- 238000002347 injection Methods 0.000 description 1
- 239000007924 injection Substances 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000000737 periodic effect Effects 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000013179 statistical model Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Supply And Distribution Of Alternating Current (AREA)
Abstract
The invention discloses a method and a device for realizing fault diagnosis of an aircraft main power supply system under sample unbalance, wherein the method comprises the following steps: constructing an optimal transmission generation countermeasure network model, and training the optimal transmission generation countermeasure network model by utilizing data of each parameter in each fault state in an unbalanced data set of an aircraft main power system to obtain an optimal transmission generation countermeasure network model set; generating an countermeasure network model set by utilizing the optimal transmission, and expanding the data of each parameter under each fault state to obtain a balance data set; constructing a fault sample classification decision tree model, and training the fault sample classification decision tree model by utilizing the balance data set to obtain a trained fault sample classification decision tree model; and obtaining balance sample data of a main power supply system of the target aircraft to construct a test set, and inputting the test set data of the main power supply system of the target aircraft into the trained fault sample classification decision tree model to obtain a fault diagnosis result of the main power supply system of the target aircraft.
Description
Technical Field
The invention relates to the technical field of aerospace, in particular to a method and a device for realizing fault diagnosis of an aircraft main power supply system under sample unbalance.
Background
The power supply system of the airplane is a general name of the parts of generating, adjusting, transforming, controlling, protecting and the like of the electric energy on the airplane, and mainly comprises a main power supply, an auxiliary power supply, an emergency power supply, a secondary power supply, an external power supply and the like. As a key component of a large-scale aircraft power supply system, if faults of a constant-frequency constant-speed alternating current power supply cannot be diagnosed and processed in time, the faults may be further expanded to cause cascading reactions, so that on-board electric equipment cannot work normally, even the whole flying process cannot continue, therefore, the fault diagnosis technology research of the constant-frequency constant-speed alternating current power supply is necessary to be carried out, the fault location is carried out according to the fault diagnosis result, and further, the timely treatment and recovery of the faults are realized by means of the fault-tolerant management strategy of the aircraft power supply system. However, there are often a number of difficulties in obtaining primary power failure data during various stages of aircraft design, development, pilot flight, and normal use. Therefore, for the training and testing process of the fault diagnosis model of the main power supply of the airplane, the unbalanced sample data condition that a large number of normal state samples exist and the fault samples are insufficient is often faced, so that the research of the fault diagnosis technology of the main power supply of the airplane under the unbalanced sample condition is very important for improving the generalization capability of the fault diagnosis model.
In order to solve the unbalanced sample problem, the unbalanced sample processing thought of the data layer and the model layer can be introduced into a deep network by utilizing the powerful complex function approximation capability of deep learning. As one of the research hotspots in the deep learning method in recent years, the generation of an countermeasure network (Generative Adversarial Network, GAN) and its improved model are widely used to solve the problem of fault diagnosis under the sample imbalance condition. The GAN can fully learn and fit the distribution of the original signals of the rotary machine, generate new samples very close to the minority samples based on the approximate distribution, increase the number and diversity of the minority samples, and further improve the fault diagnosis effect under the unbalanced sample condition. However, the GAN adopts KL divergence and JS divergence to measure the distance between the real data distribution and the generated data distribution in the training process, and when there is no or negligible overlap between the two distributions, the problems of gradient disappearance, unstable training, mode collapse and the like can occur.
Disclosure of Invention
The invention provides a method and a device for realizing fault diagnosis of an aircraft main power supply system under sample unbalance, so as to solve the problems of insufficient sample generation capacity, poor generation effect and change of original data distribution of the existing unbalanced sample processing method.
The embodiment of the invention provides a method for realizing fault diagnosis of an aircraft main power supply system under sample unbalance, which comprises the following steps:
constructing an optimal transmission generation countermeasure network model, and training the optimal transmission generation countermeasure network model by utilizing data of each parameter in each fault state in an unbalanced data set of an aircraft main power system to obtain an optimal transmission generation countermeasure network model set;
generating an countermeasure network model set by utilizing the optimal transmission, and expanding the data of each parameter under each fault state to obtain a balance data set;
constructing a fault sample classification decision tree model, and training the fault sample classification decision tree model by utilizing the balance data set to obtain a trained fault sample classification decision tree model;
and obtaining balance sample data of a main power supply system of the target aircraft to construct a test set, and inputting the test set data of the main power supply system of the target aircraft into the trained fault sample classification decision tree model to obtain a fault diagnosis result of the main power supply system of the target aircraft.
Preferably, said constructing an optimal transmission generation countermeasure network model comprises:
constructing a generated countermeasure network model, and obtaining a generated sample by utilizing the generated countermeasure network model;
and calculating the optimal transmission distance between the generated sample and the real sample distribution, and replacing the KL divergence and JS divergence in the generated countermeasure network model with the optimal transmission distance to obtain the optimal transmission generated countermeasure network model.
Preferably, the method further comprises:
n fault states of an aircraft main power supply system are determined, and M parameters under each fault state are determined;
wherein, N and M are positive integers.
Preferably, the training the optimal transmission generation countermeasure network model by using the data of each parameter in each fault state in the unbalanced data set of the main power system of the aircraft, and obtaining the optimal transmission generation countermeasure network model set includes:
training the optimal transmission generation countermeasure network model by using M-N parameter data in the unbalanced data set of the aircraft main power system to obtain an optimal transmission generation countermeasure network sub-model corresponding to the M-N parameters, and forming an optimal transmission generation countermeasure network model set by the optimal transmission generation countermeasure network sub-model corresponding to the M-N parameters.
Preferably, the generating the countermeasure network model set by using the optimal transmission performs expansion processing on the data of each parameter in each fault state, and obtaining a balanced data set includes:
generating minority fault state sample data of each parameter in each fault state in the main power supply system of the airplane by using the optimal transmission generation countermeasure network model set, and obtaining a balance data set.
Preferably, a classification and regression tree CART algorithm is adopted to construct a fault sample classification decision tree model.
The embodiment of the invention provides a device for realizing fault diagnosis of an aircraft main power supply system under sample unbalance, which comprises the following components:
the first construction and training module is used for constructing an optimal transmission generation countermeasure network model, and training the optimal transmission generation countermeasure network model by utilizing data of each parameter in each fault state in the unbalanced data set of the aircraft main power supply system to obtain an optimal transmission generation countermeasure network model set;
the data sample expansion processing module is used for generating an countermeasure network model set by utilizing the optimal transmission to perform expansion processing on the data of each parameter under each fault state to obtain a balance data set;
the second construction and training module is used for constructing a fault sample classification decision tree model, and training the fault sample classification decision tree model by utilizing the balance data set to obtain a trained fault sample classification decision tree model;
the fault diagnosis module is used for acquiring balance sample data of the main power supply system of the target aircraft to construct a test set, and inputting the test set data of the main power supply system of the target aircraft into the trained fault sample classification decision tree model to obtain a fault diagnosis result of the main power supply system of the target aircraft.
The beneficial effects of the invention include as follows:
1. the intelligent data driving method WGAN (Wasserstein Generative Adversarial Network, optimal transmission generation countermeasure network) can complete training based on a small number of fault state samples, generate a large number of new samples approximately distributed with the real fault state samples, effectively solve the problem of sample unbalance in the fault diagnosis field, and save the cost brought by collecting the fault state samples and developing a large number of fault injection tests.
2. The adopted WGAN model indicates the countermeasure game process of the generator and the discriminator by means of the optimal transmission Wasserstein distance, so that the distribution of the generated samples of the generator can be sufficiently close to the distribution of the real fault samples of the main power supply of the aircraft, the number and diversity of the fault state samples of the main power supply are further increased, and the problems of insufficient generation capacity, poor generation effect, change of the original data distribution and the like of a classical resampling method are effectively avoided.
3. The method can construct a WGAN model set for generating high-dimensional parameters based on a plurality of WGAN models with the same generator and discriminator model architecture aiming at the high-dimensional nonlinear monitoring parameters of the aircraft main power supply, and can effectively reduce the space complexity of the model. By introducing optimal transmission Wasserstein distance and gradient penalty terms, training stability of the model and high-quality generation capacity of high-dimensional parameters of a main power supply are enhanced.
Drawings
FIG. 1 is a flow chart of a method for implementing fault diagnosis of an aircraft main power system under sample imbalance provided by the invention;
FIG. 2 is a detailed flow chart of a method for performing fault diagnosis of an aircraft main power system under sample imbalance provided by the present invention;
FIG. 3 is a schematic diagram of a fault diagnosis flow based on DT algorithm provided by the present invention;
FIG. 4 is a schematic diagram of the operation principle of the constant-speed constant-frequency alternating current power supply provided by the invention;
FIG. 5 is a schematic diagram of a simulation model of a main power supply of an aircraft according to the present invention
FIG. 6 is a graphical illustration of a true sample versus generated sample of AC exciter A phase armature winding current provided by the present invention;
FIG. 7 is a graphical illustration of a true sample versus a generated sample of a main generator phase-to-phase short circuit fault provided by the present invention;
FIG. 8 is a comparison of fault diagnosis results for different imbalance ratio enhanced data sets provided by the present invention;
FIG. 9 is a schematic diagram showing comparison of confusion matrices of fault diagnosis results of different unbalance ratio enhancement data sets provided by the invention.
Detailed Description
It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention. In the following description, suffixes such as "module", "part" or "unit" for representing elements are used only for facilitating the description of the present invention, and have no particular meaning in themselves. Thus, "module," "component," or "unit" may be used in combination.
According to the invention, the Wasserstein distance representing the optimal transmission path is used for measuring the distance between the real data distribution and the generated data distribution, even if the two distributions are not overlapped, the Wasserstein distance can still reflect the distance between the two distributions, and a gradient is provided for model optimization, so that a stable WGAN model and high-quality generated data are obtained. Therefore, the invention provides a fault diagnosis method WGAN-DT for generating samples based on a WGAN model set and combining DT to identify fault modes.
The invention mainly comprises two phases, wherein the first phase is to generate few fault state samples based on the WGAN model set and construct an enhanced data set. The second stage is to train DT fault mode classification model based on the enhanced data set to realize fault diagnosis. The invention tests the effectiveness and the universality of a fault diagnosis model method for generating an countermeasure network based on optimal transmission based on simulation sample set data of an aircraft main power supply. Experimental results show that the method provided by the invention can effectively solve the problems of insufficient sample generation capacity, poor generation effect, change of original data distribution and the like of the existing unbalanced sample processing method. The invention adopts the optimal transmission distance GAN to generate a few fault state samples of the main power supply system of the airplane, and solves the problems of poor fault diagnosis effect and the like caused by unbalanced samples.
Fig. 1 is a flowchart of a method for implementing fault diagnosis of an aircraft main power system under sample imbalance, where the method may include:
step S101: constructing an optimal transmission generation countermeasure network model, and training the optimal transmission generation countermeasure network model by utilizing data of each parameter in each fault state in an unbalanced data set of an aircraft main power system to obtain an optimal transmission generation countermeasure network model set;
step S102: generating an countermeasure network model set by utilizing the optimal transmission, and expanding the data of each parameter under each fault state to obtain a balance data set;
step S103: constructing a fault sample classification decision tree model, and training the fault sample classification decision tree model by utilizing the balance data set to obtain a trained fault sample classification decision tree model;
step S104: and obtaining balance sample data of a main power supply system of the target aircraft to construct a test set, and inputting the test set data of the main power supply system of the target aircraft into the trained fault sample classification decision tree model to obtain a fault diagnosis result of the main power supply system of the target aircraft.
Preferably, said constructing an optimal transmission generation countermeasure network model comprises: constructing a generated countermeasure network model, and obtaining a generated sample by utilizing the generated countermeasure network model; and calculating the optimal transmission distance between the generated sample and the real sample distribution, and replacing the KL divergence and JS divergence in the generated countermeasure network model with the optimal transmission distance to obtain the optimal transmission generated countermeasure network model.
The embodiment of the invention also comprises the following steps: n fault states of an aircraft main power supply system are determined, and M parameters under each fault state are determined; wherein, N and M are positive integers.
Further, the training the optimal transmission generation countermeasure network model by using the data of each parameter in each fault state in the unbalanced data set of the main power supply system of the aircraft, and obtaining the optimal transmission generation countermeasure network model set includes: training the optimal transmission generation countermeasure network model by using M-N parameter data in the unbalanced data set of the aircraft main power system to obtain an optimal transmission generation countermeasure network sub-model corresponding to the M-N parameters, and forming an optimal transmission generation countermeasure network model set by the optimal transmission generation countermeasure network sub-model corresponding to the M-N parameters.
Specifically, the generating the countermeasure network model set by using the optimal transmission performs expansion processing on the data of each parameter in each fault state, and obtaining a balanced data set includes: generating minority fault state sample data of each parameter in each fault state in the main power supply system of the airplane by using the optimal transmission generation countermeasure network model set, and obtaining a balance data set.
The embodiment of the invention adopts a classification and regression tree CART algorithm to construct a fault sample classification decision tree model.
The embodiment of the invention also provides a device for realizing the fault diagnosis of the main power supply system of the airplane under the unbalance of the sample, which comprises the following components: the first construction and training module is used for constructing an optimal transmission generation countermeasure network model, and training the optimal transmission generation countermeasure network model by utilizing data of each parameter in each fault state in the unbalanced data set of the aircraft main power supply system to obtain an optimal transmission generation countermeasure network model set; the data sample expansion processing module is used for generating an countermeasure network model set by utilizing the optimal transmission to perform expansion processing on the data of each parameter under each fault state to obtain a balance data set; the second construction and training module is used for constructing a fault sample classification decision tree model, and training the fault sample classification decision tree model by utilizing the balance data set to obtain a trained fault sample classification decision tree model; the fault diagnosis module is used for acquiring balance sample data of the main power supply system of the target aircraft to construct a test set, and inputting the test set data of the main power supply system of the target aircraft into the trained fault sample classification decision tree model to obtain a fault diagnosis result of the main power supply system of the target aircraft.
Fig. 2 is a detailed flowchart of a method for implementing fault diagnosis of a main power supply of an aircraft under sample imbalance, which is provided in the invention, and as shown in fig. 2, includes:
step one: optimal transmission generation opposing network WGAN construction
The design of the generation countermeasure network framework is used for referencing the idea of zero and game, and two parts of the generator and the arbiter are similar to two parties in the game process, and when the income of one party is good, the income of the other party is poor. The core structure of GAN consists of two parts, namely a Generator (G) and a discriminator (D). The training process of GAN is to continuously update the network parameters of both parties according to the feedback based on the generator D and the arbiter G, so that the benefits of both parties are balanced in nash, i.e. the arbiter cannot distinguish whether the sample comes from the generator or the real data.
1) A generator
The generator takes random noise as input and outputs a generated sample. In the training process, the training of the generator is that the generated synthesized sample can 'cheat' the discriminator, so that the discriminator outputs a probability close to 1, and the generated sample is true. When the training process of the generator and the arbiter reaches Nash equilibrium, the generator can accurately reconstruct enough "spurious" generated samples from the noise. The loss function of the generator is shown in equation (1).
2) Distinguishing device
The discriminator is randomly inputted with a true sample or generates a sample and outputs a discrimination probability value of the sample source. In the training process, the aim of the discriminator is to accurately distinguish the real sample from the generated sample, i.e. the probability that the output of the real sample is close to 1 indicates that the sample is true; and for the probability of generating a sample output approaching 0, the sample is indicated as "false". When the training process of the generator and the arbiter reaches the Nash equilibrium, the arbiter will not be able to determine whether the sample is from the generator or the real sample set. The loss function of the arbiter is shown in equation (2).
3) Loss function
The loss function of the overall GAN challenge training can be formulated uniformly with the following formulation.
The formula is a dynamic game reactance process of the generator G and the arbiter D. The solution of this equation requires the calculation of one parameter while the other parameter is fixed. Due to the first term in the formula
And give birth toThe constructors are independent, so minimizing equation (3) is equivalent to minimizing the generator loss function represented by equation (2).
Let the derivative of the arbiter loss function with respect to D (x) be 0, the optimal arbiter is:
in the method, in the process of the invention, P g representing the distribution of the generated samples.
Substituting the formula (4) into the formula (3) and transforming to obtain:
it is possible to introduce KL and JS divergences into equation (4):
(6)
L＝2JS(P r ||P g )-2log2
the KL divergence and JS divergence are respectively:
minimizing the generator loss function is equivalent to minimizing P with optimal arbiter parameters r And P g JS divergence between. In a high-dimensional space, however, P r And P g There is little overlap between the two distributions, so regardless of how far apart the two distributions are, the JS divergence is a constant log2, which easily results in a generator loss function gradient of 0, i.e., results in GAN gradient cancellation and training instability. The invention replaces KL divergence and JS divergence in GAN with Wasserstein distance, and is used for measuring the distance between real data distribution and generated data distribution. Even if the two distributions do not overlap or overlap is negligibleThe Wasserstein distance can also measure the distance between the two distributions, provide gradient for loss function optimization, effectively avoid the problem of gradient disappearance, make the training process more stable, and then improve the quality of the generated data.
True sample P r And generate sample P g The Wasserstein distance between them is defined as follows:
Π(P r ,P g ) Is P r And P g All possible joint distribution sets of the combination are formed. x is the true sample, y is the generated sample, and (x, y) is the sample from the joint distribution γ. I x-y i is the distance between the x and y samples, then
The expected value of the distance between two samples at the joint distribution gamma is characterized. The minimum solution of equation (9), which is the expected value infinitesimal of all joint distributions, is defined as the Wasserstein distance, also known as the bulldozer (Earth-Mover, EM) distance. Distance expectation +>
Comparing with the path planning gamma, the soil pile P r To the soil pile P g The EM distance is the "minimum consumption" corresponding to the "optimal transmission path plan".
Due to
Cannot be directly solved, so Lipschitz is introduced to continuously transform the formula (9) to obtain:
wherein,, I f I L Lipschitz constant, which is a function f (x), has an upper limit of K.
Further by introducing a function f containing a parameter omega ω (x) Distance of EM W (P r ,P g ) The approximation is expressed as:
the generator approximately minimizes the EM distance, i.e., minimizes L in equation (11), resulting in a generator and arbiter loss function for WGAN:
the training effect can be represented by inverting the expression (13), and the smaller the value thereof is, the smaller the EM distance between the generated sample distribution and the real sample distribution is, namely, the closer the generated sample output by the generator is to the real sample, and the better the generating effect is.
The loss function of the constructed model generator is shown in a formula (12), the loss function of the discriminator is shown in a formula (13), wherein the loss function of the discriminator is calculated based on Wasserstein distance, and the discriminator is trained and updated by adopting an RMSProp optimization algorithm, and in the training process, the generator and the discriminator update parameters in turn until a Nash equilibrium state is reached. The training objective function for the entire generated countermeasure network is as follows:
in the model built by the invention, the generator and the discriminator are both built based on the multi-layer perceptron, wherein the generator is mainly composed of two fully connected layers, and the discriminator model is mainly composed of three fully connected layers.
Step two: optimal transmission generation countermeasure network training
The WGAN sub-model is trained using the imbalance samples using the WGAN algorithm model constructed above. For each iteration (epoch), the training process is as follows.
The pseudocode of the WGAN algorithm is as follows:
and carrying out training of a plurality of iteration loops on the constructed optimal transmission generation countermeasure network according to the preset total number of epochs until the loss function is no longer reduced, and completing the training process of the optimal transmission generation countermeasure network.
Step three: sample generation based on WGAN model set
Based on the WGAN model which is completed by the training, as the main power supply system of the airplane comprises a plurality of key sensitive parameters and each parameter is a time sequence type periodic signal, each parameter of each fault mode is adopted to train one WGAN model, a plurality of WGAN models are trained together, a WGAN model set is built, and then an enhanced data set is constructed.
1) Construction of WGAN model set
A WGAN submodel is trained based on each parameter of each fault condition of the aircraft main power supply. The initial parameters and the architecture of the WGAN submodel corresponding to each parameter are identical, and only the model superparameter after training is different. Further, a set of WGAN models is constructed based on the complete training of all WGAN sub-models for generating various parameters of various fault conditions of the aircraft main power supply.
2) Constructing an enhanced dataset
Generating various parameters of the aircraft main power system under various fault states by utilizing the WGAN model set, and constructing a minority class of fault state samples, generating and expanding the fault state samples in the original unbalanced sample set, increasing the number and diversity of the fault state samples, further obtaining an enhanced data set with different balance ratios, and providing a data basis for further training a DT fault mode classification model and verifying sample generation effects.
Step four: failure mode classification based on DT model
Based on the enhanced data set constructed by utilizing the WGAN model set, a Decision Tree (DT) algorithm is adopted to classify the fault modes.
Decision trees are very common modeling methods in statistics, data mining and machine learning, and mainly consist of root nodes, internal nodes and leaf nodes. The root node and the internal node correspond to decision rules of sample characteristics, and the leaf node corresponds to a category to which the sample belongs. Among the three more classical classes of decision trees are CART (Classification And Regression Trees, classification and regression tree), ID3 and C4.5, respectively. In the classification problem, CART adopts a kunning coefficient to select the optimal feature and the splitting point, ID3 is sample feature selection according to the information gain, and C4.5 algorithm performs sample feature selection according to the information gain rate. In various decision tree algorithms, the CART algorithm adopts a simplified binary tree to construct a decision tree model, can process continuous features and select features by using a coefficient of a radix, and has stronger generalization capability through a cross-validated pruning strategy optimization model. Therefore, aiming at the problem of fault diagnosis of the main power supply of the airplane, which is to be solved by the invention, a CART algorithm is adopted to construct a fault sample classification decision tree model.
And taking the DT model as a fault mode classification model of the main power supply system of the aircraft, and generating an enhanced data set of a sample structure as a training set thereof. The DT model first discretizes the continuous parameters in the training set, then, the coefficient of the kunning of each value of each parameter value is calculated, selecting the value of the parameter with the minimum coefficient of the radix as a dividing point, generating a binary tree, and continuously dividing downwards until the number of samples of a certain subtree is lower than a threshold value or the coefficient of the base is lower than the threshold value, so as to complete model construction. And then pruning optimization is carried out on the model by adopting a post pruning strategy, so as to obtain a final DT model. And further testing the trained DT fault mode classification model by adopting a test set, and outputting fault diagnosis results of the DT model on the training set and the test set, wherein the flow is shown in figure 3.
The technical contents of the present invention will be described in detail with specific examples of "diagnosis of main power supply failure of aircraft under sample imbalance based on WGAN-DT
Description of test data: the test adopts fault simulation data of the main power supply of the airplane to construct an unbalanced sample set, and the provided fault diagnosis method is subjected to experimental verification. The invention mainly researches a main power supply of an airplane with a constant-speed constant-frequency alternating-current power supply type, and the main composition structure of the main power supply comprises a permanent magnet auxiliary exciter, an alternating-current main exciter, a rotary rectifier, a main generator and the like. The principle of operation of this type of main power supply is shown in fig. 4.
And constructing four main parts of the main power supply of the airplane such as a permanent magnet auxiliary exciter, an alternating current main exciter, a rotary rectifier, a main generator and the like in a MATLAB/Simulink environment, connecting interfaces of the four models according to the operation principle of the main power supply of the airplane, and finally forming the simulation model of the main power supply of the airplane as shown in fig. 5.
And determining key monitoring parameters according to fault diagnosis requirements of 9 typical fault modes of the main power supply of the airplane, such as a normal state, a single-phase open-circuit fault of the armature winding of the main generator, an interphase short-circuit fault of the armature winding of the AC exciter, a single-diode open-circuit fault of the rotary rectifier, an open-circuit fault of the exciting winding of the AC exciter, a short-circuit fault of the exciting winding of the AC exciter and the armature winding of the AC exciter on the motor shell, and the like, and setting corresponding measuring points in a simulation model of the main power supply of the airplane. The measurement point list is shown in table 1:
table 1 typical failure modes and sample numbers
Injecting and simulating a typical fault mode based on an airplane main power supply simulation model, and constructing an unbalance ratio of 1000:10 as a training set of fault diagnosis models. In the acquired 10s simulation data, the first 6000 sampling points after the simulation model runs for 9s are selected for constructing a test set, namely, each fault mode comprises 50 test samples, and the test set comprises 450 test samples of 9 typical fault modes.
Because the data researched by the invention are simulation signals of the main power supply of the aircraft, the diversity is low, noise is added to the initial simulation data according to a certain signal-to-noise ratio SNR, the diversity of samples is increased, and the data generated by the WAGN-GP method is ensured to have a certain diversity.
SNR＝10lg(P s /P N )
Wherein P is s For signal energy, P N Is the energy of noise
Step 1: optimal transmission generation countermeasure network construction
Because the aircraft main power supply comprises a plurality of key sensitive parameters, and each parameter in different fault states has physical meaning and data characteristics, and constructing a WGAN model for each fault state model is too complex, a WGAN model set for generating high-dimensional parameters of the main power supply is constructed by adopting the WGAN model with the same generator and discriminator architecture. Each parameter of each fault state corresponds to one sub-model in the WGAN model set, the few fault states studied by the invention are 8, each fault state comprises 12 monitoring parameters, and therefore 96 WGAN sub-models are required to be trained together to form the WGAN model set.
Considering that the multi-layer perceptron can achieve fitting of nonlinear functions, a generator and a discriminator of the WGAN are built with the multi-layer perceptron to achieve nonlinear mapping from random noise to generated samples, and from samples to true and false decision values.
The setup parameters for the WGAN-based sample generation model are shown in the following table:
table 2WGAN model parameter set table
As shown in the table, in the WGAN-based sample generation model, both the generator and the discriminant are constructed based on a multi-layer perceptron, wherein the generator is mainly composed of two fully connected layers, and the discriminant model is mainly composed of three fully connected layers. The WAGN model adopts RMSprop as an optimizer, the learning rate is 0.001, the number of batch samples is set to 40, the iteration number is set to 10000, the sample normalization range is 0 to 1, and the random noise range input by the generator is also 0 to 1.
Step 2: optimal transmission generation countermeasure network training
Aiming at the sample unbalance problem existing in the main power supply of the aircraft, namely the situation that the number of samples in each fault state is 10 and is far less than the number (1000) of samples in a normal state under 8 typical fault states such as single-phase open-circuit faults of an armature winding of the main generator, phase short-circuit faults of an armature winding of the main generator, single-phase open-circuit faults of an armature winding of an alternating-current exciter, phase short-circuit faults of an armature winding of the alternating-current exciter, single-diode open-circuit faults of a rotating rectifier, open-circuit faults of an exciting winding of the alternating-current exciter, short-circuit faults of an exciting winding of the alternating-current exciter and the like, the number of samples in each fault state is far less than the number (1000) of samples in the normal state, a WGAN model set is adopted to generate few fault state samples, and the original unbalance sample set is expanded. Constructing one WGAN model for each fault state model is overly complex, so a set of WGAN models for generating high-dimensional parameters of the main power supply is constructed using WGAN models with the same generator and arbiter architecture.
The WGAN sub-model is trained using the imbalance samples using the WGAN algorithm model constructed above. Each pair of network parameters is updated once, and the number of used samples in a batch is 40. The number of challenge training epochs was 10000 times, using RMSprop as an optimizer.
And training the constructed optimal transmission generation countermeasure network in a plurality of iteration loops according to 10000 times of the preset epoch total number until the loss function is no longer reduced, and completing the training process of the optimal transmission generation countermeasure network.
Step 3: sample generation based on WGAN model set
A minority class fault state sample of the aircraft main power supply is generated based on the trained WGAN model set. Fig. 6 shows the parameter waveforms of the real samples in 8 fault states and the parameter waveforms generated based on the WGAN model set, taking the parameter of the ac exciter a-phase armature winding current as an example (the upper graph is a sample in each fault waveform graph, and the lower graph is a real sample).
Further, taking the main generator phase-to-phase short circuit fault as an example, comparing each parameter waveform of the real sample with the generated sample parameter waveform obtained based on the WGAN model, as shown in fig. 7, wherein the upper graph in each parameter waveform graph is the generated sample, and the lower graph is the real sample.
By comparing the real samples and the generated samples of the alternating current exciter A-phase armature winding current under various fault states, the generated samples under various fault states have higher similarity with the real samples, and the WGAN model set well learns the characteristic differences of original parameters among different fault modes. Meanwhile, by comparing the real sample and the generated sample of the interphase short-circuit fault of the main generator, the WGAN model set can accurately learn the distribution characteristics of each parameter in a single fault mode, and the generating effect is good. Through transverse and longitudinal comparison, the training based on a few fault state samples is illustrated, the WGAN model set can learn the characteristics of various types of parameters under various fault states of the main power supply of the aircraft well, and a good data basis is provided for expanding an unbalanced sample set and realizing fault diagnosis in the follow-up implementation.
Step 4: failure mode classification based on DT model
In order to fully analyze the quality of generated samples obtained based on the WGAN and the effect of the generated samples on the performance improvement of the fault mode classifier, a few fault state samples are generated by using a trained WGAN model set, and the initial unbalanced sample set of the aircraft main power supply is expanded. In addition, the performance of the sample set under different sample conditions is tested by sequentially adjusting the imbalance ratio of the sample set from the initial 100:1 to 50:1, 10:1,5:1 and 1:1, and providing enhanced data sets with different balance ratios for the DT fault mode classification model. The DT fault mode classification model is trained 5 times on the enhancement data sets with different unbalance proportions, and the fault diagnosis accuracy average values of the statistical model on the training set and the test set are shown in the table 3:
table 1 fault diagnosis accuracy of different balanced ratio enhancement datasets
In order to more intuitively show the enhanced data set of the main power supply of the airplane based on the data structure generated by the WGAN model set, the improvement condition of the diagnosis accuracy of the DT fault mode classification model is shown in a histogram form, and the change condition of the accuracy of the DT fault mode classification model on the training set and the testing set is shown in fig. 8 after the enhanced data set is constructed by adopting different numbers of generated samples.
The training set of the DT fault pattern classification model is used as the original unbalanced sample set with the unbalance ratio of 100:1 and the enhancement data sets with the unbalance ratios of 50:1, 10:1,5:1 and 1:1 based on the subsequent extended sample structure, and the confusion matrix of the fault diagnosis results (the experimental result with the highest test set accuracy in 5 experiments is selected) of the model on the training set and the test set is shown in FIG. 9.
In summary, the invention constructs the WGAN model set based on the WGAN model with the same architecture of the generator and the arbiter, and is used for generating various parameters of the fault state sample so as to expand the original unbalanced sample set to obtain the enhanced data set, and is used for training the DT fault mode classification model. Along with the increasing of the number of generated samples in the enhanced data set, the fault diagnosis accuracy of the DT model is also improved, and when the number of fault state samples and normal state samples in the enhanced data set reach balance, the fault diagnosis accuracy is as high as 95.73%. The WGAN-DT-based fault diagnosis method effectively solves the problem that a classical fault diagnosis model is inclined to a plurality of types of samples under the unbalanced sample condition and the fault diagnosis accuracy is not high by constructing a balanced enhanced data set. In addition, the fault diagnosis accuracy of the WGAN-DT method provided by the invention on the unbalanced sample set of the main power supply of the aircraft is obviously higher than that of a fault diagnosis method combining a classical fault diagnosis model and a resampling method with the DT model, and is 7.06 percent higher than that of the RUS-DT method with the fault diagnosis accuracy of 2 nd in a comparison method. Compared with a classical fault diagnosis model, the WGAN-DT method adopts a training set for generating a sample structure fault diagnosis model, and can remarkably improve the problems of overfitting, weak generalization capability and the like of the fault diagnosis model caused by sample unbalance. Compared with a classical resampling method, the WGAN model set can generate few fault state samples close to the probability distribution of a real sample, so that the number and diversity of original unbalanced sample sets are effectively increased, and the fault diagnosis effect is further improved.
The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, and thus do not limit the scope of the claims of the present invention. Any modifications, equivalent substitutions and improvements made by those skilled in the art without departing from the scope and spirit of the present invention shall fall within the scope of the appended claims.
Claims (7)
1. A method for performing fault diagnosis of an aircraft main power system under sample imbalance, the method comprising:
constructing an optimal transmission generation countermeasure network model, and training the optimal transmission generation countermeasure network model by utilizing data of each parameter in each fault state in an unbalanced data set of an aircraft main power system to obtain an optimal transmission generation countermeasure network model set;
generating an countermeasure network model set by utilizing the optimal transmission, and expanding the data of each parameter under each fault state to obtain a balance data set;
constructing a fault sample classification decision tree model, and training the fault sample classification decision tree model by utilizing the balance data set to obtain a trained fault sample classification decision tree model;
and obtaining balance sample data of a main power supply system of the target aircraft to construct a test set, and inputting the test set data of the main power supply system of the target aircraft into the trained fault sample classification decision tree model to obtain a fault diagnosis result of the main power supply system of the target aircraft.
2. The method of claim 1, wherein said constructing an optimal transmission generation countermeasure network model comprises:
constructing a generated countermeasure network model, and obtaining a generated sample by utilizing the generated countermeasure network model;
and calculating the optimal transmission distance between the generated sample and the real sample distribution, and replacing the KL divergence and JS divergence in the generated countermeasure network model with the optimal transmission distance to obtain the optimal transmission generated countermeasure network model.
3. The method according to claim 1 or 2, further comprising:
n fault states of an aircraft main power supply system are determined, and M parameters under each fault state are determined;
wherein, N and M are positive integers.
4. A method according to claim 3, wherein training the optimal transmission generation countermeasure network model using data for each parameter in each fault condition in the aircraft main power system imbalance data set to obtain the optimal transmission generation countermeasure network model set comprises:
training the optimal transmission generation countermeasure network model by using M-N parameter data in the unbalanced data set of the aircraft main power system to obtain an optimal transmission generation countermeasure network sub-model corresponding to the M-N parameters, and forming an optimal transmission generation countermeasure network model set by the optimal transmission generation countermeasure network sub-model corresponding to the M-N parameters.
5. The method of claim 4, wherein generating the set of countermeasure network models using the optimal transmission expands the data for each parameter in each fault condition to obtain a balanced data set, comprising:
generating minority fault state sample data of each parameter in each fault state in the main power supply system of the airplane by using the optimal transmission generation countermeasure network model set, and obtaining a balance data set.
6. The method of claim 5, wherein the fault sample classification decision tree model is constructed using a classification and regression tree CART algorithm.
7. An apparatus for implementing fault diagnosis of an aircraft main power system under sample imbalance, comprising:
the first construction and training module is used for constructing an optimal transmission generation countermeasure network model, and training the optimal transmission generation countermeasure network model by utilizing data of each parameter in each fault state in the unbalanced data set of the aircraft main power supply system to obtain an optimal transmission generation countermeasure network model set;
the data sample expansion processing module is used for generating an countermeasure network model set by utilizing the optimal transmission to perform expansion processing on the data of each parameter under each fault state to obtain a balance data set;
the second construction and training module is used for constructing a fault sample classification decision tree model, and training the fault sample classification decision tree model by utilizing the balance data set to obtain a trained fault sample classification decision tree model;
the fault diagnosis module is used for acquiring balance sample data of the main power supply system of the target aircraft to construct a test set, and inputting the test set data of the main power supply system of the target aircraft into the trained fault sample classification decision tree model to obtain a fault diagnosis result of the main power supply system of the target aircraft.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310002066.3A CN116070159A (en) | 2023-01-03 | 2023-01-03 | Method and device for realizing fault diagnosis of main power supply system of airplane under unbalanced sample |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310002066.3A CN116070159A (en) | 2023-01-03 | 2023-01-03 | Method and device for realizing fault diagnosis of main power supply system of airplane under unbalanced sample |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116070159A true CN116070159A (en) | 2023-05-05 |
Family
ID=86174332
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310002066.3A Pending CN116070159A (en) | 2023-01-03 | 2023-01-03 | Method and device for realizing fault diagnosis of main power supply system of airplane under unbalanced sample |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116070159A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117973235A (en) * | 2024-04-01 | 2024-05-03 | 安徽农业大学 | Centrifugal pump fault mode diagnosis method for simulating kinematic vibration data |
-
2023
- 2023-01-03 CN CN202310002066.3A patent/CN116070159A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117973235A (en) * | 2024-04-01 | 2024-05-03 | 安徽农业大学 | Centrifugal pump fault mode diagnosis method for simulating kinematic vibration data |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | Partial transfer learning of multidiscriminator deep weighted adversarial network in cross-machine fault diagnosis | |
| Bai et al. | A novel parameter identification approach via hybrid learning for aggregate load modeling | |
| CN109033702A (en) | A kind of Transient Voltage Stability in Electric Power System appraisal procedure based on convolutional neural networks CNN | |
| CN108241873B (en) | A kind of intelligent failure diagnosis method towards pumping plant main equipment | |
| Baltas et al. | A comparative analysis of decision trees, support vector machines and artificial neural networks for on-line transient stability assessment | |
| CN108551167A (en) | A kind of electric power system transient stability method of discrimination based on XGBoost algorithms | |
| CN106897821A (en) | A kind of transient state assesses feature selection approach and device | |
| CN108828436A (en) | Analog-circuit fault diagnosis method based on the adaptive glowworm swarm algorithm of chaos cloud | |
| CN111523778A (en) | Power grid operation safety assessment method based on particle swarm algorithm and gradient lifting tree | |
| CN114006370B (en) | Power system transient stability analysis and evaluation method and system | |
| Wang | Application of extension theory to vibration fault diagnosis of generator sets | |
| Li et al. | Deep learning-based transient stability assessment framework for large-scale modern power system | |
| CN116070159A (en) | Method and device for realizing fault diagnosis of main power supply system of airplane under unbalanced sample | |
| Li et al. | Fault identification in power network based on deep reinforcement learning | |
| Foggo et al. | pmuBAGE: The benchmarking assortment of generated PMU data for power system events | |
| CN117370568A (en) | Power grid main equipment knowledge graph completion method based on pre-training language model | |
| CN106055761B (en) | A kind of load model parameters method of calibration based on decision tree classification | |
| Wang | Research on the fault diagnosis of mechanical equipment vibration system based on expert system | |
| Zhang et al. | Prediction for the maximum frequency deviation of post-disturbance based on the deep belief network | |
| CN116566061A (en) | Grid-connected inverter system stability on-line monitoring method and system | |
| CN115713032A (en) | Power grid prevention control method, device, equipment and medium | |
| CN115936926A (en) | SMOTE-GBDT-based unbalanced electricity stealing data classification method and device, computer equipment and storage medium | |
| CN113255213B (en) | Power converter composite fault identification method | |
| CN109815108A (en) | A kind of combined test set of uses case priorization sort method and system based on weight | |
| CN116204771A (en) | Power system transient stability key feature selection method, device and product |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |