CN114330517A - Neural network-based aircraft engine sensor fault self-diagnosis method - Google Patents
Neural network-based aircraft engine sensor fault self-diagnosis method Download PDFInfo
- Publication number
- CN114330517A CN114330517A CN202111536040.4A CN202111536040A CN114330517A CN 114330517 A CN114330517 A CN 114330517A CN 202111536040 A CN202111536040 A CN 202111536040A CN 114330517 A CN114330517 A CN 114330517A
- Authority
- CN
- China
- Prior art keywords
- sensor
- neural network
- fault
- function
- historical
- 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
Images
Abstract
The invention discloses a neural network-based aircraft engine sensor fault self-diagnosis method, which belongs to the field of aircraft engine sensor fault diagnosis and comprises the following steps: establishing an aircraft engine component level model, and collecting a sensor historical parameter database; designing a probability distribution function prediction model based on a neural network; designing a time series fault classification model based on a neural network; and combining the probability distribution mutation prediction model and the sensor fault classification model to construct a sensor fault self-diagnosis framework, and realizing the sensor fault self-diagnosis of the turbofan engine by adopting the constructed fault self-diagnosis framework. The invention aims to solve the problem of fault diagnosis of a sensor under the condition that the sensor can only access local historical data of the sensor, realize probability estimation through a neural network to solve the problems of uncertainty and randomness under the condition of incomplete information, and extract the characteristics of time sequence data of the sensor to realize fault classification.
Description
Technical Field
The invention relates to a neural network-based aircraft engine sensor fault self-diagnosis method, and belongs to the field of aircraft engine sensor fault diagnosis.
Background
With the rapid development of aviation propulsion technology, modern flight mission puts higher requirements on the accuracy and stability of an engine control system, and a full authority digital electronic control system (FADEC) gradually replaces mechanical hydraulic control with strong calculation processing capability, extremely high control precision and relatively low maintenance cost, and becomes the mainstream trend of the current aviation engine control system. The digital electronic control system consists of a large number of electronic components, sensors and actuating mechanisms, and an aircraft engine is often in a high-temperature high-pressure complex and variable working environment, so that the components are prone to failure in the operation process of the engine, wherein the failure of the sensors is a main reason for failure of the control system. Therefore, the engine sensor fault diagnosis and real-time health state grasping are important measures for ensuring safe operation of the engine.
The method for diagnosing the faults of the sensor of the aircraft engine mainly comprises a mathematical model (such as system identification and filtering estimation) and an artificial intelligence (such as support vector base, wavelet analysis, fuzzy inference and neural network). The main principle is mostly based on residual error detection, for example, based on real-time estimation of an engine real-time model and a Kalman filter in a mathematical model, a sensor parameter filtering residual error is established, and is compared with a corresponding threshold value, so that the purpose of fault detection is achieved. Under the trend that a distributed control system is used as a future control framework of an aircraft engine, an intelligent sensor capable of independently realizing self diagnosis becomes a popular research, and the sensor is required to realize fault diagnosis under the condition that only local data can be accessed. However, current fault diagnosis methods rely on highly accurate real-time models or a large number of training samples containing other sensor information, and it is difficult to accurately predict residuals when the information is insufficient.
Aiming at the fault self-diagnosis problem of an aircraft engine sensor under the condition that only local historical data can be accessed, a sensor parameter probability distribution function prediction model is provided based on a neural network, the probability density distribution function of the sensor parameter at the next moment is constructed according to the historical data by considering the uncertainty caused by incomplete information, and the real-time judgment of the sudden-change fault is realized according to the probability. On the basis, a time series fault classification model is introduced to judge and classify faults (hard failure, noise and drift) under a long time series.
Disclosure of Invention
The invention aims to solve the technical problem of fault self-diagnosis of an aircraft engine sensor under the condition that only local historical data can be accessed, by considering uncertainty caused by incomplete information, a probability density distribution function of sensor parameters at the next moment is constructed according to the historical data, sudden-change fault real-time judgment is realized according to the probability, and a time series fault classification model is introduced, so that faults (hard failure, noise and drift) under a long-time sequence are judged and classified, and the safe and reliable operation of an engine is ensured.
The invention adopts the following technical scheme: a neural network-based aircraft engine sensor fault self-diagnosis method is characterized by comprising the following steps:
a, establishing an aircraft engine component level model, and collecting a sensor historical parameter database;
b, designing a sensor parameter probability distribution function prediction model based on a neural network, predicting the probability distribution function of the sensor parameter at the next moment only according to local historical parameter data in a historical parameter database of the sensor, and judging whether parameter mutation occurs according to the probability of the real sensor parameter;
step C, designing a sensor time sequence fault classification model based on a neural network, extracting a characteristic vector of a sensor historical time sequence by the model, and discontinuously judging whether the sensor has faults or not and classifying the faults;
and D, combining the probability distribution mutation prediction model and the sensor fault classification model, constructing a sensor fault self-diagnosis framework, and storing and outputting a diagnosis result and a report.
The step A comprises the following steps:
a certain type of engine component level model is established according to the principle of an aircraft engine, sensor parameters are collected in the running process of the engine and are stored in a database in a time sequence for neural network training.
In the step B, the concrete steps of establishing the sensor parameter probability distribution function prediction model based on the neural network are as follows:
(1) designing neural network structures
Designing neural network input using sensor parameter data within historical access depth p time as network input [ X [ ]t-1，Xt-2，Xt-3，...，Xt-p]TWherein X isiSensor parameter data representing time i; designing neural network output, taking a prediction vector as network output, taking a sample label corresponding to the dimension of the prediction vector as network output, and taking the mean value mu and the standard deviation sigma of a normal distribution function as the network output to construct a probability distribution function N (mu, sigma) of a predicted value; selecting and optimizing the number of hidden layers of the neural network and the number of nodes of each layer according to experience and experimental results; selecting an activation function according to the overall structure and the function of the neural network, selecting a ReLU function max (0, x) as the activation function of the middle layer, and selecting a tanh function as the function of the output layer
(2) Training neural networks
Sampling data in batches, randomly selecting partial data from a database as a training set, and randomly sampling N training samples from the training set for each training, wherein the ith sample characteristic is represented as [ X ]i-1，Xi-2，Xi-3,...,Xi-p]TSample label is Xi(ii) a Designing a loss function to increase the probability value P (X) of the occurrence of the label of the training samplet|Xt-1，Xt-2，...，Xt-p) Loss ofThe function is positively correlated with the value, in order to reduce the training speed and improve the stability of the training process, a logarithm function of P is selected to construct a loss function, and the batch training method adopts a plurality of samples for training once, so that the loss function is expected L-sigma lnP of the sampling probability of the samples; optimizing the neural network parameter vector theta by adopting a gradient descent method according to the designed loss function and a learning rate alpha, wherein the kth training process can be expressed as
(3) Testing neural networks
Constructing a test set, and randomly selecting partial data from a database as the test set; calculating the probability distribution of the sample label, inputting the sample characteristics into the neural network, estimating the mean value of the probability normal distribution by the neural network
And standard deviation of
Constructing an estimated probability distribution
Determining mutation threshold, determining mutation threshold range according to 3 sigma criterion, and normally distributing in (mu-3 sigma )]The value probability within is 99.7 percent, and the sample label is (mu-3 sigma )]And when the probability of the other sensor is less than 0.3%, the sensor is called a small probability event, and the sensor can be judged to have sudden change.
In the step C, the concrete steps of establishing the sensor long-time sequence fault classification model based on the neural network are as follows:
(1) analyzing sensor failure modes
The sensor failure modes include: the sensor normally operates; hard failure, sensor off, parameter remaining unchanged from outside influences, may be expressed as
tfIs the time of occurrence of the fault; drift, a fixed drift offset of the sensor parameter from normal, which can be expressed as Xt(t) + bias; noise, which is too noisy in time series, can be expressed as Xt＝f(t)+Noise；
(2) Designing neural network structures
Designing neural network input, accessing historical time sequence of sensor parameters in time T of fault classification interval before sensor, and considering double-margin sensor (subscript for affiliated parameter)1And2expressed), and calculates the difference between the two sensor parameter sequences-the error sequence (the associated parameter is indexed with a subscript)eExpressing), extracting the mean value and standard deviation of the characteristic value of each segment of sequence and error sequence to form a characteristic vector
g＝[μ1,μ2,μe,σ1,σ2,σe]T (1.2)
Designing neural network output, wherein according to fault classification, a sample label can be designed to be a 4-dimensional Boolean vector to represent a fault mode, and the neural network output is also a 4-dimensional vector;
selecting and optimizing the number of hidden layers of the neural network and the number of nodes of each layer according to experience and experimental results; selecting an activation function according to the overall structure and the function of the neural network, selecting a ReLU function max (0, x) as the activation function of the middle layer, selecting a softmax function as the activation function of the output layer to regularize the network output
Wherein z isjFor the jth output of the neural network that has not been activated by the softmax function,
for the ith output after being activated by the softmax function, the softmax activation function can convert the neural network output vector elements into elements which are smaller than 1 and the sum of which is 1;
(3) training neural networks
Randomly selecting partial data from a sensor historical parameter database, extracting historical time sequence characteristics as sample characteristics, introducing fault signals into the historical data, and giving corresponding Boolean vector labels to form a training set;
designing a loss function, and defining the loss function as the cross entropy of a network output vector and a sample label vector by combining the characteristics of a softmax activation function in order to improve the accuracy of fault classification
Optimizing the neural network parameter vector theta by adopting a gradient descent method according to the designed loss function and a learning rate alpha, wherein the kth training process can be expressed as
The step D comprises the following steps:
(1) initialization
Initializing time T as 0, historical access depth p and fault classification interval T;
(2) probability distribution prediction and mutation judgment
Historical parameters in a historical visit depth p moment before visit, a probability prediction model is used for predicting a probability distribution function of sensor parameters at the t moment, and whether mutation occurs is judged by combining actual sensor parameters;
(3) fault classification
Accessing historical parameters in the previous moment p, predicting a probability distribution function of sensor parameters at the moment t by using a probability prediction model, and judging whether mutation occurs or not by combining actual sensor parameters; and judging whether the time T% T of the fault classification interval is 0, if so, accessing historical parameters in the previous T time, extracting a feature vector, and judging whether the fault occurs by using a fault classification model and classifying the fault.
The invention has the following beneficial effects:
(1) the invention provides a neural network-based parameter probability distribution function prediction model for the fault self-diagnosis problem of an aircraft engine sensor under the condition that only local historical data can be accessed, solves the problem caused by parameter uncertainty by constructing the probability distribution function of sensor parameters, and realizes the single-point mutation fault self-diagnosis of the sensor according to the probability of the parameters.
(2) The invention introduces a sensor time sequence fault classification model to judge and classify faults (hard failure, noise and drift) under a long time sequence, thereby ensuring the safe and reliable operation of an engine.
(3) The sensor parameter probability distribution function prediction model and the time series fault classification model can perform fault self-diagnosis under the condition of only accessing local historical data of the sensor, and can provide theoretical and experimental basis for the development and application of a self-diagnosis intelligent sensor under the background that an aircraft engine distributed control architecture is taken as a future trend.
Drawings
FIG. 1 is a schematic diagram of a turbofan engine rotor speed sensor parameter single point sudden change fault diagnosis based on a probability distribution function prediction model;
FIG. 2 is a diagram of a neural network architecture for a probability distribution function prediction model;
FIG. 3 is a diagram of a neural network structure of a time series fault classification model;
FIG. 4 is a time series fault classification schematic of turbofan engine rotor speed sensor parameters;
FIG. 5 is a cross-sectional schematic view of a turbofan engine component level model;
FIG. 6 is a general frame diagram of a turbofan engine sensor fault self-diagnosis;
FIG. 7 is a simulation diagram of single point sudden change fault diagnosis of a turbofan engine rotor speed sensor parameter;
FIG. 8 is a simulation diagram of a normal case of time series fault classification of turbofan engine rotor speed sensor parameters;
FIG. 9 is a simulation plot of a hard failure fault for time series fault classification of turbofan engine rotor speed sensor parameters;
FIG. 10 is a simulation plot of a drift fault under time series fault classification of turbofan engine rotor speed sensor parameters;
FIG. 11 is a simulation plot of a time series fault classification of turbofan engine rotor speed sensor parameters under over-noisy fault.
Detailed Description
The technical scheme of the invention is explained in detail in the following with the accompanying drawings:
the specific embodiment of the invention takes the fault diagnosis of a low-pressure rotor speed sensor of a certain turbofan engine as an example, and the principle of the fault diagnosis of single-point mutation of the rotor speed sensor parameter of the turbofan engine based on a probability distribution function prediction model is shown in fig. 1. At time t, the turbofan engine receives a control quantity utModel output low pressure rotor speed XtAnd the data is stored in a historical database as a probability distribution function prediction model training sample and an input value of real-time prediction by the sensing of a sensor. The training process and the real-time prediction process of the prediction model adopt a double-thread mode. In the first thread, training samples are randomly sampled from a historical database to serve as a training set, each training round calculates a loss function L (sigma) lnP, the neural network is trained, and the parameter vector of the neural network is updated
Wherein k is the number of training rounds, and the neural network structure of the probability distribution function prediction model is shown in FIG. 2; in the second thread, the previous p time history data [ X ] is extracted from the history databaset-1，...，Xt-p]TAs the input of the prediction model, the probability distribution function N (mu, sigma) of the sensor parameter at the next moment is predicted in real time, and the probability density size P (X) of the sensor parameter is calculatedt|Xt-1,Xt-2,...,Xt-p) And whether the sensor has single-point mutation fault is judged according to the probability distribution function, the probability distribution function in the experiment is normal distribution, and 3 sigma criterion can be adopted, namely when X ist∈(μ-3σ,μ+3σ]And judging that the sensor is normal, otherwise, judging that the sensor has single-point mutation fault. The parameter vector of the prediction model of the second thread is directly subjected to parameter assignment by the neural network of the first thread.
On the basis of single-point sudden change fault diagnosis, the time sequence of the parameters of the rotating speed sensor of the turbofan engine rotor isThe fault classification (including hard failure, excessive noise and drift) is carried out, the neural network structure of the sensor time sequence fault classification model is shown in fig. 3, the two-line training process of the sensor parameter probability distribution function prediction model is different, and the scheme of offline training and offline use of the neural network is adopted for long-time sequence fault diagnosis. Firstly, training data containing labels y is utilized to cross entropy loss function
Training the time sequence fault classification neural network model, and using the model offline after the model training is completed. FIG. 4 shows the principle of time series fault classification of the parameters of a turbofan engine rotor speed sensor, where the turbofan engine receives a control quantity utModel output low pressure rotor speed XtAnd the data is stored in a historical database by the induction of a sensor, and when the time T reaches the classification interval T% T which is 0, a time sequence is sampled from the historical database
And
(taking into account double margins) and calculating the error sequence
The mean and standard deviation of each sequence were extracted as the feature vector g ═ μ1,μ2,μe,σ1,σ2,σe]TThe feature vector is used as the input of a time series fault classification model, and a classification result can be obtained
The index of the medium maximum corresponds to the fault model, e.g. when
And then, judging as a failure mode 3-drift.
Designed neural network-based aeronautical hair for verificationAccording to the effectiveness of the fault self-diagnosis method of the motive machine sensor, the dynamic link library of the engine component-level model is called to perform digital simulation in a Python environment, and the structure of the component-level model is shown in FIG. 5. The general framework of the self-diagnosis of the sensor fault of the turbofan engine is shown in fig. 6, wherein the initialization time T is 0, the historical access depth p, the fault classification interval T, the sensor measures the rotor speed at each moment and stores the rotor speed in a historical database, the sensor parameter at the previous p moments is accessed at each moment and is used as the input of a probability distribution function prediction model, the probability distribution function of the sensor parameter at the moment is obtained and whether single-point sudden change fault occurs or not is judged, if the single-point sudden change fault occurs, then the probability distribution function mean value is used for replacing the mutation parameter to realize parameter reconstruction, when the time T reaches the fault classification interval T% T is 0, then accessing the sensor parameters at the T moment before accessing for feature extraction and using the sensor parameters as a time series fault classification model to realize normal, hard failure, drift and over-noise fault classification, when the engine finishes running, the whole turbofan engine sensor fault self-diagnosis process is simultaneously finished. In combination with the engine real data, the sensor measurement noise is set to a 0.2% level and the sensor sampling step is set to 0.02 s. The method comprises the steps that a sensor parameter probability distribution function prediction model is used for diagnosing single-point mutation faults of parameters of a rotating speed sensor of a turbofan engine rotor, signal reconstruction is achieved through probability mean values, and when the sensor is free of faults, the criterion rate of the parameter single-point mutation faults is 100%; when the sensor has single-point mutation fault, the failure rate of parameter single-point mutation fault diagnosis is 99.8%. Fig. 7 shows a working simulation diagram of the sensor parameter probability distribution function prediction model in a section of engine acceleration process, in 10s simulation, a rotor speed sensor has a total of 3 single-point sudden-change faults, all the faults are successfully diagnosed by the sensor parameter probability distribution function prediction model, and a sensor signal is reconstructed by using the mean value of the prediction distribution function. Time series fault classification is realized on the parameters of the rotating speed sensor of the turbofan engine rotor by using a sensor time series fault classification model, and the total accuracy of fault classification is 96.7%; when the training sample label is normal y ═ 1,0,0,0]TIn time, the fault classification accuracy is 99.6%; when the training sample label is hard failure y ═ 0,1,0,0]TIn time, the fault classification accuracy is 99.6%; when the training sample label is drift y ═ 0,0,1,0]TIn time, the fault classification accuracy is 75.5%; when the training sample label is too large y ═ 0,0,0,1]TThe fault classification accuracy is 90.5%. Fig. 8 shows the time-series fault classification results of the rotor speed sensor parameters during the engine operation 40s when the sensor is normally operated (mode 0), fig. 9 shows the time-series fault classification results of the rotor speed sensor parameters when the sensor is hard failed (mode 1), fig. 10 shows the time-series fault classification results of the rotor speed sensor parameters when the sensor is in drift fault (mode 2), and fig. 11 shows the time-series fault classification results of the rotor speed sensor parameters when the sensor is excessively noisy (mode 3).
The invention discloses a neural network-based aircraft engine sensor fault self-diagnosis method, which comprises the following steps of:
a, establishing an aircraft engine component level model, and collecting a sensor historical parameter database;
b, designing a sensor parameter probability distribution function prediction model based on a neural network, predicting the probability distribution function of the sensor parameter at the next moment only according to local historical parameter data in a historical parameter database of the sensor, and judging whether parameter mutation occurs according to the probability of the real sensor parameter;
step C, designing a sensor time sequence fault classification model based on a neural network, extracting a characteristic vector of a sensor historical time sequence by the model, and discontinuously judging whether the sensor has faults or not and classifying the faults;
and D, combining the probability distribution mutation prediction model and the sensor fault classification model, constructing a sensor fault self-diagnosis framework, and storing and outputting a diagnosis result and a report.
The step A comprises the following steps:
a certain type of engine component level model is established according to the principle of an aircraft engine, sensor parameters are collected in the running process of the engine and are stored in a database in a time sequence for neural network training.
In the step B, the concrete steps of establishing the sensor parameter probability distribution function prediction model based on the neural network are as follows:
(1) designing neural network structures
Designing neural network input using sensor parameter data within historical access depth p time as network input [ X [ ]t-1,Xt-2,Xt-3,...,Xt-p]TWherein X isiSensor parameter data representing time i; designing neural network output, taking a prediction vector as network output, taking a sample label corresponding to the dimension of the prediction vector as network output, and taking the mean value mu and the standard deviation sigma of a normal distribution function as the network output to construct a probability distribution function N (mu, sigma) of a predicted value; selecting and optimizing the number of hidden layers of the neural network and the number of nodes of each layer according to experience and experimental results; selecting an activation function according to the overall structure and the function of the neural network, selecting a ReLU function max (0, x) as the activation function of the middle layer, and selecting a tanh function as the function of the output layer
(2) Training neural networks
Sampling data in batches, randomly selecting partial data from a database as a training set, and randomly sampling N training samples from the training set for each training, wherein the ith sample characteristic is represented as [ X ]i-1,Xi-2,Xi-3,...,Xi-p]TSample label is Xi(ii) a Designing a loss function to increase the probability value P (X) of the occurrence of the label of the training samplet|Xt-1,Xt-2,...,Xt-p) The loss function is positively correlated with the value, in order to reduce the training speed and improve the stability of the training process, a logarithm function of P is selected to construct the loss function, and the batch training method adopts a plurality of samples for training once, so that the loss function is expected L-sigma lnP of the sampling probability of the samples; optimizing the neural network parameter vector theta by adopting a gradient descent method according to the designed loss function and a learning rate alpha, wherein the kth training process can be expressed as
(3) Testing neural networks
Constructing a test set, and randomly selecting partial data from a database as the test set; calculating the probability distribution of the sample label, inputting the sample characteristics into the neural network, estimating the mean value of the probability normal distribution by the neural network
And standard deviation of
Constructing an estimated probability distribution
Determining mutation threshold, determining mutation threshold range according to 3 sigma criterion, and normally distributing in (mu-3 sigma )]The value probability within is 99.7 percent, and the sample label is (mu-3 sigma )]And when the probability of the other sensor is less than 0.3%, the sensor is called a small probability event, and the sensor can be judged to have sudden change.
In the step C, the concrete steps of establishing the sensor long-time sequence fault classification model based on the neural network are as follows:
(1) analyzing sensor failure modes
The sensor failure modes include: the sensor normally operates; hard failure, sensor off, parameter remaining unchanged from outside influences, may be expressed as
tfIs the time of occurrence of the fault; drift, a fixed drift offset of the sensor parameter from normal, which can be expressed as Xt(t) + bias; noise, which is too noisy in time series, can be expressed as Xt＝f(t)+Noise；
(2) Designing neural network structures
Designing neural network inputs, accessing sensor parameter history within time T of fault classification interval before sensorTime series, taking into account double-margin sensors (subscripts for the parameters to which they pertain)1And2expressed), and calculates the difference between the two sensor parameter sequences-the error sequence (the associated parameter is indexed with a subscript)eExpressing), extracting the mean value and standard deviation of the characteristic value of each segment of sequence and error sequence to form a characteristic vector
g＝[μ1,μ2,μe,σ1,σ2,σe]T (1.3)
Designing neural network output, wherein according to fault classification, a sample label can be designed to be a 4-dimensional Boolean vector to represent a fault mode, and the neural network output is also a 4-dimensional vector;
selecting and optimizing the number of hidden layers of the neural network and the number of nodes of each layer according to experience and experimental results; selecting an activation function according to the overall structure and the function of the neural network, selecting a ReLU function max (0, x) as the activation function of the middle layer, selecting a softmax function as the activation function of the output layer to regularize the network output
Wherein z isjFor the jth output of the neural network that has not been activated by the softmax function,
for the ith output after being activated by the softmax function, the softmax activation function can convert the neural network output vector elements into elements which are smaller than 1 and the sum of which is 1;
(3) training neural networks
Randomly selecting partial data from a sensor historical parameter database, extracting historical time sequence characteristics as sample characteristics, introducing fault signals into the historical data, and giving corresponding Boolean vector labels to form a training set;
designing a loss function, and defining the loss function as the cross entropy of a network output vector and a sample label vector by combining the characteristics of a softmax activation function in order to improve the accuracy of fault classification
Optimizing the neural network parameter vector theta by adopting a gradient descent method according to the designed loss function and a learning rate alpha, wherein the kth training process can be expressed as
The step D comprises the following steps:
(1) initialization
Initializing time T as 0, historical access depth p and fault classification interval T;
(2) probability distribution prediction and mutation judgment
Historical parameters in a historical visit depth p moment before visit, a probability prediction model is used for predicting a probability distribution function of sensor parameters at the t moment, and whether mutation occurs is judged by combining actual sensor parameters;
(3) fault classification
Accessing historical parameters in the previous moment p, predicting a probability distribution function of sensor parameters at the moment t by using a probability prediction model, and judging whether mutation occurs or not by combining actual sensor parameters; and judging whether the time T% T of the fault classification interval is 0, if so, accessing historical parameters in the previous T time, extracting a feature vector, and judging whether the fault occurs by using a fault classification model and classifying the fault.
It should be noted that the above mentioned embodiments are only specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any changes and substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the appended claims.
Claims (6)
1. A neural network-based aircraft engine sensor fault self-diagnosis method is characterized by comprising the following steps:
a, establishing an aircraft engine component level model, and collecting a sensor historical parameter database;
b, designing a sensor parameter probability distribution function prediction model based on a neural network, predicting the probability distribution function of the sensor parameter at the next moment only according to local historical parameter data in a historical parameter database of the sensor, and judging whether parameter mutation occurs according to the probability of the real sensor parameter;
step C, designing a sensor time sequence fault classification model based on a neural network, extracting a characteristic vector of a sensor historical time sequence by the model, and discontinuously judging whether the sensor has faults or not and classifying the faults;
and D, combining the probability distribution mutation prediction model and the sensor fault classification model, constructing a sensor fault self-diagnosis framework, and storing and outputting a diagnosis result and a report.
2. The neural network-based aircraft engine sensor fault self-diagnosis method of claim 1, characterized in that: the step A specifically comprises the following steps: a certain type of engine component level model is established according to the principle of an aircraft engine, sensor parameters are collected in the running process of the engine and are stored in a database in a time sequence for neural network training.
3. The neural network-based aircraft engine sensor fault self-diagnosis method of claim 1, characterized in that: the step B comprises the following steps:
(1) designing neural network structures
Designing neural network input using sensor parameter data within historical access depth p time as network input [ X [ ]t-1,Xt-2,Xt-3,...,Xt-p]TWherein X isiSensor parameter data representing time i; designing neural network output, taking a prediction vector as network output, taking a sample label corresponding to the dimension of the prediction vector as network output, and taking the mean value mu and the standard deviation sigma of a normal distribution function as the network output to construct a probability distribution function N (mu, sigma) of a predicted value; selecting and optimizing the number of hidden layers of the neural network and the number of nodes of each layer according to experience and experimental results;selecting an activation function according to the overall structure and the function of the neural network, selecting a ReLU function max (0, x) as the activation function of the middle layer, and selecting a tanh function as the function of the output layer
(2) Training neural networks
Sampling data in batches, randomly selecting partial data from a database as a training set, and randomly sampling N training samples from the training set for each training, wherein the ith sample characteristic is represented as [ X ]i-1,Xi-2,Xi-3,...,Xi-p]TSample label is Xi(ii) a Designing a loss function to increase the probability value P (X) of the occurrence of the label of the training samplet|Xt-1,Xt-2,...,Xt-p) The loss function is positively correlated with the value, in order to reduce the training speed and improve the stability of the training process, a logarithm function of P is selected to construct the loss function, and the batch training method adopts a plurality of samples for training once, so that the loss function is the expected L-sigma ln P of the sampling probability of the samples; optimizing the neural network parameter vector theta by adopting a gradient descent method according to the designed loss function and a learning rate alpha, wherein the kth training process can be expressed as
(3) Testing neural networks
Constructing a test set, and randomly selecting partial data from a database as the test set; calculating the probability distribution of the sample label, inputting the sample characteristics into the neural network, estimating the mean value of the probability normal distribution by the neural network
And standard deviation of
Constructing an estimated probability distribution
Determining mutation threshold, determining mutation threshold range according to 3 sigma criterion, and normally distributing in (mu-3 sigma )]The value probability within is 99.7 percent, and the sample label is (mu-3 sigma )]And when the probability of the other sensor is less than 0.3%, the sensor is called a small probability event, and the sensor can be judged to have sudden change.
4. The neural network-based aircraft engine sensor fault self-diagnosis method of claim 3, characterized in that: in the step B, self-diagnosis is realized only by local historical information of the sensor; in order to deal with uncertainty and randomness caused by working condition change of the aircraft engine, an estimation probability distribution function is constructed
All uncertainties are covered; by estimating the size of the probability
Replacing the conventional estimated residual
And judging whether mutation occurs or not, wherein in normal distribution, a 3 sigma criterion is used as a mutation judgment basis.
5. The neural network-based aircraft engine sensor fault self-diagnosis method as claimed in claim 1, wherein the step C comprises the steps of:
(1) analyzing sensor failure modes
The sensor failure modes include: the sensor normally operates; hard failure, sensor off, parameter remaining unchanged from outside influences, may be expressed as
tfIs the time of occurrence of the fault; drift, fixed drift offset of sensor parameters from normalCan be represented as Xt(t) + bias; noise, which is too noisy in time series, can be expressed as Xt＝f(t)+Noise；
(2) Designing neural network structures
Designing neural network input, accessing historical time series of sensor parameters in the time T of fault classification interval before sensor, considering double-margin sensor, using subscript 1 and 2 to represent parameters, calculating the difference between two sensor parameter series-error series, using subscript e to represent parameters, extracting the mean value and standard difference of characteristic values of each series and error series, forming characteristic vector
g＝[μ1，μ2，μe，σ1，σ2，σe]T (1.1)
Designing neural network output, wherein according to fault classification, a sample label can be designed to be a 4-dimensional Boolean vector to represent a fault mode, and the neural network output is also a 4-dimensional vector;
selecting and optimizing the number of hidden layers of the neural network and the number of nodes of each layer according to experience and experimental results; selecting an activation function according to the overall structure and the function of the neural network, selecting a ReLU function max (0, x) as the activation function of the middle layer, selecting a softmax function as the activation function of the output layer to regularize the network output
Wherein z isjFor the jth output of the neural network that has not been activated by the softmax function,
for the ith output after being activated by the softmax function, the softmax activation function can convert the neural network output vector elements into elements which are smaller than 1 and the sum of which is 1;
(3) training neural networks
Randomly selecting partial data from a sensor historical parameter database, extracting historical time sequence characteristics as sample characteristics, introducing fault signals into the historical data, and giving corresponding Boolean vector labels to form a training set;
designing a loss function, and defining the loss function as the cross entropy of a network output vector and a sample label vector by combining the characteristics of a softmax activation function in order to improve the accuracy of fault classification
Optimizing the neural network parameter vector theta by adopting a gradient descent method according to the designed loss function and a learning rate alpha, wherein the kth training process can be expressed as
6. The neural network-based aircraft engine sensor fault self-diagnosis method of claim 1, characterized in that: the step D comprises the following steps:
(1) initialization
Initializing time T as 0, historical access depth p and fault classification interval T;
(2) probability distribution prediction and mutation judgment
Historical parameters in a historical visit depth p moment before visit, a probability prediction model is used for predicting a probability distribution function of sensor parameters at the t moment, and whether mutation occurs is judged by combining actual sensor parameters;
(3) fault classification
Accessing historical parameters in the previous moment p, predicting a probability distribution function of sensor parameters at the moment t by using a probability prediction model, and judging whether mutation occurs or not by combining actual sensor parameters; and judging whether the time T% T of the fault classification interval is 0, if so, accessing historical parameters in the previous T time, extracting a feature vector, and judging whether the fault occurs by using a fault classification model and classifying the fault.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111536040.4A CN114330517A (en) | 2021-12-15 | 2021-12-15 | Neural network-based aircraft engine sensor fault self-diagnosis method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111536040.4A CN114330517A (en) | 2021-12-15 | 2021-12-15 | Neural network-based aircraft engine sensor fault self-diagnosis method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114330517A true CN114330517A (en) | 2022-04-12 |
Family
ID=81053449
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111536040.4A Pending CN114330517A (en) | 2021-12-15 | 2021-12-15 | Neural network-based aircraft engine sensor fault self-diagnosis method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114330517A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115131943A (en) * | 2022-07-07 | 2022-09-30 | 杭州申昊科技股份有限公司 | Acousto-optic linkage early warning method based on deep learning |
| CN116702030A (en) * | 2023-05-31 | 2023-09-05 | 浙江大学 | Blast furnace state monitoring method and device based on sensor reliability analysis |
| CN117074628A (en) * | 2023-10-17 | 2023-11-17 | 山东鑫建检测技术有限公司 | Multi-sensor air quality detection equipment fault positioning method |
-
2021
- 2021-12-15 CN CN202111536040.4A patent/CN114330517A/en active Pending
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115131943A (en) * | 2022-07-07 | 2022-09-30 | 杭州申昊科技股份有限公司 | Acousto-optic linkage early warning method based on deep learning |
| CN115131943B (en) * | 2022-07-07 | 2023-10-31 | 杭州申昊科技股份有限公司 | Acousto-optic linkage early warning method based on deep learning |
| CN116702030A (en) * | 2023-05-31 | 2023-09-05 | 浙江大学 | Blast furnace state monitoring method and device based on sensor reliability analysis |
| CN116702030B (en) * | 2023-05-31 | 2024-01-23 | 浙江大学 | Blast furnace state monitoring method and device based on sensor reliability analysis |
| CN117074628A (en) * | 2023-10-17 | 2023-11-17 | 山东鑫建检测技术有限公司 | Multi-sensor air quality detection equipment fault positioning method |
| CN117074628B (en) * | 2023-10-17 | 2024-01-09 | 山东鑫建检测技术有限公司 | Multi-sensor air quality detection equipment fault positioning method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112131760B (en) | CBAM model-based prediction method for residual life of aircraft engine | |
| CN114330517A (en) | Neural network-based aircraft engine sensor fault self-diagnosis method | |
| CN109766583A (en) | Based on no label, unbalanced, initial value uncertain data aero-engine service life prediction technique | |
| Reddy et al. | Anomaly detection and fault disambiguation in large flight data: A multi-modal deep auto-encoder approach | |
| CN107657250B (en) | Bearing fault detection and positioning method and detection and positioning model implementation system and method | |
| CN108256173B (en) | Gas circuit fault diagnosis method and system for dynamic process of aircraft engine | |
| CN110609524B (en) | Industrial equipment residual life prediction model and construction method and application thereof | |
| CN109376401B (en) | Self-adaptive multi-source information optimization and fusion mechanical residual life prediction method | |
| CN111523081A (en) | Aircraft engine fault diagnosis method based on enhanced gated cyclic neural network | |
| CN114266278B (en) | Dual-attention network-based equipment residual service life prediction method | |
| CN112668775A (en) | Air quality prediction method based on time sequence convolution network algorithm | |
| CN113869563A (en) | Method for predicting remaining life of aviation turbofan engine based on fault feature migration | |
| CN113569338A (en) | Gas compressor rotating stall early warning method based on time expansion convolutional network | |
| CN111860839A (en) | Shore bridge fault monitoring method based on multi-signal fusion and Adam optimization algorithm | |
| CN115481658A (en) | Pulse echo state network model for aircraft engine fault prediction | |
| CN115081484A (en) | Aircraft engine sensor fault diagnosis method based on CRJ-OSELM algorithm | |
| CN108459933B (en) | Big data computer system fault detection method based on deep recursion network | |
| CN114169091A (en) | Method for establishing prediction model of residual life of engineering mechanical part and prediction method | |
| CN116662743A (en) | Engine residual life prediction method based on multi-mode deep learning | |
| CN116521406A (en) | Method for detecting anomaly of non-overrun flight parameter data of aero-engine based on residual gate GRU-VAE model | |
| CN115859791A (en) | Method and device for estimating residual life of aero-engine | |
| CN112560252B (en) | Method for predicting residual life of aeroengine | |
| CN114841063A (en) | Aero-engine residual life prediction method based on deep learning | |
| Zhi-hong et al. | Sensor Fault Diagnosis Based on Wavelet Analysis and LSTM Neural Network | |
| Wang et al. | Degradation pattern identification and remaining useful life prediction for mechanical equipment using SKF-EN |
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 |