A method for self-diagnosis of aircraft engine sensor faults based on neural network

Through the neural network model, the probability distribution function and time series fault classification model of sensor parameters are constructed, which solves the problem of self-diagnosis of faults of aircraft engine sensors when there is insufficient information, and realizes accurate judgment of mutations and long-term time series faults, ensuring the engine's Run safely.

CN114330517BActive Publication Date: 2024-11-12NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
View PDF 9 Cites 0 Cited by

Patent Information

Application Number
CN202111536040.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-15
Publication Date
2024-11-12
Estimated Expiration
2041-12-15

AI Technical Summary

Technical Problem

It is difficult to achieve accurate prediction of fault self-diagnosis in case that aero engine sensors can only access local historical data, especially when there is insufficient information, which leads to an increased risk of failure of the control system.

Method used

The prediction model of sensor parameter probability distribution function and time series fault classification model are adopted based on neural network. By constructing the probability density distribution function of sensor parameters, mutation faults are judged, and faults (hard failure, noise, drift) under long time series are carried out. Classification.

Benefits of technology

It realizes self-diagnosis of sensor failures in case of incomplete information, improves the safe and reliable operation of the engine, and ensures the self-diagnosis capability of the sensor under the distributed control architecture.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114330517B_ABST
    Figure CN114330517B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for self-diagnosis of aero-engine sensor faults based on a neural network, which belongs to the field of aero-engine sensor fault diagnosis and includes the following steps: establishing an aero-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; combining the probability distribution mutation prediction model and the sensor fault classification model to construct a sensor fault self-diagnosis framework, and using the constructed fault self-diagnosis framework to realize sensor fault self-diagnosis of a turbofan engine. The present invention aims to solve the fault diagnosis problem of a sensor when it can only access its own local historical data, realize probability estimation through a neural network to solve the uncertainty and randomness problems under the condition of imperfect information, and perform feature extraction on the sensor time series data to realize fault classification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for self-diagnosis of aeroengine sensor faults based on a neural network, and belongs to the field of aeroengine sensor fault diagnosis. Background Art

[0002] With the rapid development of aviation propulsion technology, modern flight missions have put forward higher requirements for the accuracy and stability of engine control systems. The full authority digital electronic control system (FADEC) has gradually replaced mechanical hydraulic control with its powerful computing and processing capabilities, extremely high control accuracy and relatively low maintenance costs, becoming the mainstream trend of current aviation engine control systems. Digital electronic control systems are composed of a large number of electronic components, sensors and actuators, and aircraft engines are often in high temperature and high pressure, complex and changeable working environments, which makes these components prone to failure during engine operation. Among them, sensor failure is the main cause of control system failure. Therefore, fault diagnosis of engine sensors and real-time grasp of their health status are important measures to ensure the safe operation of the engine.

[0003] The main methods for fault diagnosis of aircraft engine sensors are based on mathematical models (such as system identification and filter estimation) and artificial intelligence (such as support vector basis, wavelet analysis, fuzzy reasoning, neural network). The main principle is based on residual test. For example, the sensor parameter filter residual is established based on the real-time estimation of the engine real-time model and the Kalman filter in the mathematical model, and compared with the corresponding threshold, so as to achieve the purpose of fault detection. In the artificial intelligence-based method, the prior knowledge and training data are used for reasoning to realize the real-time prediction of sensor parameters and perform fault detection based on the residual with the actual parameters. With the trend of distributed control system as the future control architecture of aircraft engines, intelligent sensors that can independently realize self-diagnosis have become a hot research topic, requiring sensors to realize fault diagnosis when they can only access local data. However, the current fault diagnosis methods rely on high-precision real-time models or a large number of training samples containing other sensor information. It is difficult to accurately predict residuals when there is insufficient information.

[0004] Aiming at the problem of self-diagnosis of aircraft engine sensors when only local historical data can be accessed, the present invention proposes a prediction model of sensor parameter probability distribution function based on neural network. By considering the uncertainty caused by incomplete information, the probability density distribution function of the sensor parameters at the next moment is constructed according to historical data, and the real-time judgment of sudden faults is realized according to the probability size. On this basis, the time series fault classification model is introduced to realize the judgment and classification of faults (hard failure, noise, drift) under long time series. Summary of the invention

[0005] The technical problem to be solved by the present invention is to solve the fault self-diagnosis problem of aircraft engine sensors when only local historical data can be accessed. By considering the uncertainty caused by incomplete information, the probability density distribution function of the sensor parameters at the next moment is constructed according to the historical data, and the real-time judgment of sudden faults is realized according to the probability size. A time series fault classification model is introduced to realize the judgment and classification of faults (hard failures, noise, drift) under long time series, so as to ensure the safe and reliable operation of the engine.

[0006] The present invention adopts the following technical solution: a method for self-diagnosis of aircraft engine sensor faults based on neural network, characterized in that the method comprises the following steps:

[0007] Step A: Establish an aircraft engine component-level model and collect a sensor historical parameter database;

[0008] Step B: design a sensor parameter probability distribution function prediction model based on a neural network. The model predicts the probability distribution function of the sensor parameter at the next moment based only on the local historical parameter data in the sensor's own historical parameter database, and determines whether a parameter mutation occurs based on the probability of the actual sensor parameter.

[0009] Step C: designing a sensor time series fault classification model based on a neural network, the model extracts feature vectors of the sensor's historical time series, intermittently determines whether the sensor has a fault and classifies the fault;

[0010] Step D: Combine the probability distribution mutation prediction model and the sensor fault classification model to build a sensor fault self-diagnosis framework, save and output the diagnosis results and reports.

[0011] The step A comprises the following steps:

[0012] Based on the principles of aviation engines, a component-level model of a certain type of engine is established. During the operation of the engine, sensor parameters are collected and stored in the database in time series for neural network training.

[0013] In step B, the specific steps of establishing a sensor parameter probability distribution function prediction model based on a neural network are as follows:

[0014] (1) Designing the neural network structure

[0015] Design neural network input and use the sensor parameter data within the historical access depth p as network input [X t-1 , X t-2 , X t-3 , ..., X t-p ] T , where X iRepresents the sensor parameter data at time i; designs the neural network output, takes the prediction vector as the network output, and the prediction vector dimension corresponds to the sample label with the mean μ and standard deviation σ of the normal distribution function as the network output, and constructs the prediction value probability distribution function N(μ, σ); selects and optimizes the number of hidden layers and the number of nodes in each layer of the neural network based on experience and experimental results; selects the activation function based on the overall structure and function of the neural network, selects the ReLU function max(0, x) as the activation function of the intermediate layer, and selects the tanh function as the output layer

[0016] (2) Training the Neural Network

[0017] Data batch sampling randomly selects part of the data from the database as the training set. Each training randomly samples N training samples from the training set, where the i-th sample feature is represented by [X i-1 , X i-2 , X i-3 ,...,X i-p ] T , the sample label is X i ; Design the loss function to increase the probability value P(X t |X t-1 , X t-2 , ..., X t-p ), the loss function should be positively correlated with this value. In order to reduce the training speed and improve the stability of the training process, the logarithmic function of P is selected to construct the loss function. Since the batch training method uses multiple samples for training at a time, the loss function is the expectation of the sampling probability of multiple samples L = -∑lnP; neural network parameter vector optimization, according to the designed loss function, the gradient descent method is used to optimize the neural network parameter vector θ with the learning rate α. The k-th training process can be expressed as

[0018] (3) Testing the neural network

[0019] Construct a test set and randomly select some data from the database as the test set; calculate the probability distribution of sample labels, input sample features into the neural network, and use the neural network to estimate the mean of the probability normal distribution With standard deviation Constructing estimated probability distributions Determine the mutation threshold and determine the mutation threshold range according to the 3σ criterion. The probability of the normal distribution value within (μ-3σ, μ-3σ] is 99.7%, and the probability of the sample label outside (μ-3σ, μ-3σ] is less than 0.3%, which is called a low-probability event. At this time, it can be determined that the sensor has mutated.

[0020] In step C, the specific steps of establishing a sensor long time series fault classification model based on a neural network are as follows:

[0021] (1) Analyze sensor failure modes

[0022] Sensor failure modes include: normal, the sensor operates normally; hard failure, the sensor does not work, and the parameters remain unchanged without external influences, which can be expressed as t f is the time when the fault occurs; drift, the sensor parameters have a fixed drift offset from the normal value, which can be expressed as X t =f(t)+bias; Noise: If the noise level is too large in the time series, it can be expressed as X t =f(t)+Noise;

[0023] (2) Designing the neural network structure

[0024] Design the neural network input, access the sensor parameter history time series within the sensor fault classification interval T, consider the double margin sensor (the parameters are represented by subscripts 1 and 2), and calculate the difference between the two sensor parameter sequences - the error sequence (the parameters are represented by subscripts e Represented), extract the mean and standard deviation of the eigenvalues ​​of each sequence and error sequence to form a eigenvector

[0025] g=[μ1,μ2,μ e ,σ1,σ2,σ e ] T (1.2)

[0026] Design the output of the neural network. According to the fault classification, the sample label can be designed as a 4-dimensional Boolean vector to represent the fault mode. The output of the neural network is also a 4-dimensional vector.

[0027] The number of hidden layers and nodes in each layer of the neural network are selected and optimized based on experience and experimental results; the activation function is selected based on the overall structure and function of the neural network. The activation function of the intermediate layer is the ReLU function max(0,x), and the output layer is the softmax function to regularize the network output. where z j is the j-th output of the neural network without softmax function activation, is the i-th output after activation by the softmax function. The softmax activation function can convert the elements of the neural network output vector into elements less than 1 and whose sum is 1;

[0028] (3) Training the Neural Network

[0029] Randomly select some data from the sensor historical parameter database, extract the historical time series features as sample features, introduce fault signals into the historical data and give corresponding Boolean vector labels to form a training set;

[0030] Design loss function. In order to improve the accuracy of fault classification, combined with the characteristics of softmax activation function, the loss function is defined as the cross entropy between the network output vector and the sample label vector.

[0031] Neural network parameter vector optimization, according to the designed loss function, the gradient descent method is used to optimize the neural network parameter vector θ with a learning rate α. The k-th training process can be expressed as

[0032] The step D comprises the following steps:

[0033] (1) Initialization

[0034] Initialization time t=0, historical access depth p, fault classification interval T;

[0035] (2) Probability distribution prediction and mutation judgment

[0036] The historical parameters within the historical access depth p before access are used, and the probability distribution function of the sensor parameters at time t is predicted using the probability prediction model, and combined with the actual sensor parameters to determine whether a mutation occurs;

[0037] (3) Fault classification

[0038] Access the historical parameters within the previous time p, use the probability prediction model to predict the probability distribution function of the sensor parameters at time t, and combine the actual sensor parameters to determine whether a mutation occurs; determine whether the fault classification interval time t%T=0 is reached. If so, access the historical parameters within the previous T time, extract the feature vector, and use the fault classification model to determine whether a fault occurs and perform fault classification.

[0039] The present invention has the following beneficial effects:

[0040] (1) The present invention aims at the problem of self-diagnosis of aircraft engine sensor faults when only local historical data can be accessed. A parameter probability distribution function prediction model based on a neural network is proposed. By constructing a probability distribution function of sensor parameters, the problem caused by parameter uncertainty is solved, and single-point mutation fault self-diagnosis of the sensor is realized according to the probability size of the parameters.

[0041] (2) The present invention introduces a sensor time series fault classification model to realize the judgment and classification of faults (hard failure, noise, drift) under long time series, thereby ensuring safe and reliable operation of the engine.

[0042] (3) The sensor parameter probability distribution function prediction model and time series fault classification model of the present invention can perform fault self-diagnosis when only the local historical data of the sensor is accessed. In the context of the distributed control architecture of aircraft engines as the future trend, it can provide a theoretical and experimental basis for the development and application of self-diagnostic intelligent sensors. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a schematic diagram of the fault diagnosis principle of the single-point mutation of the turbofan engine rotor speed sensor parameters based on the probability distribution function prediction model;

[0044] Figure 2 It is a neural network structure diagram of the probability distribution function prediction model;

[0045] Figure 3 It is the neural network structure diagram of the time series fault classification model;

[0046] Figure 4 It is the time series fault classification principle diagram of the turbofan engine rotor speed sensor parameters;

[0047] Figure 5 It is a schematic cross-sectional view of a turbofan engine component-level model;

[0048] Figure 6 It is the overall framework diagram of turbofan engine sensor fault self-diagnosis;

[0049] Figure 7 This is a simulation diagram of fault diagnosis of single-point mutation of turbofan engine rotor speed sensor parameters;

[0050] Figure 8 It is a simulation diagram of the time series fault classification of the turbofan engine rotor speed sensor parameters under normal conditions;

[0051] Fig. 9 It is a simulation diagram of hard failure fault of time series fault classification of turbofan engine rotor speed sensor parameters;

[0052] Fig.10 It is a simulation diagram of drift fault of time series fault classification of turbofan engine rotor speed sensor parameters;

[0053] Fig.11 It is a simulation diagram of the time series fault classification of the turbofan engine rotor speed sensor parameters under excessive noise fault. DETAILED DESCRIPTION

[0054] The technical solution of the present invention is described in detail below with reference to the accompanying drawings:

[0055] The specific implementation scheme of the present invention takes the fault diagnosis of a low-pressure rotor speed sensor of a certain type of turbofan engine as an example. The fault diagnosis principle of the turbofan engine rotor speed sensor parameter single-point mutation based on the probability distribution function prediction model is as follows: Figure 1 At time t, the turbofan engine receives the control quantity u t , the model outputs the low-pressure rotor speed X t The sensor senses the data and stores it in the historical database as the training sample of the probability distribution function prediction model and the input value of the real-time prediction. The training process of the prediction model and the real-time prediction process adopt a dual-thread mode. In the first thread, training samples are randomly sampled from the historical database as the training set. The loss function L = -∑lnP is calculated in each training round, and the neural network is trained to update the neural network parameter vector Where k is the number of training rounds, and the neural network structure of the probability distribution function prediction model is as follows Figure 2 As shown; in the second thread, the historical data of the previous p moments [X t-1 , ..., X t-p ] T As the input of the prediction model, the probability distribution function N(μ,σ) of the sensor parameters at the next moment is predicted in real time, and the probability density size P(X t |X t-1 ,X t-2 ,...,X t-p ), and based on this, we can judge whether the sensor has a single-point mutation fault. The probability distribution function in this experiment is a normal distribution, and the 3σ criterion can be used, that is, when X t ∈(μ-3σ,μ+3σ] is judged as the sensor is normal, otherwise it is judged that the sensor has a single-point mutation fault. The parameter vector of the prediction model of the second thread is directly assigned by the neural network of the first thread.

[0056] On the basis of single-point mutation fault diagnosis, the time series fault classification of turbofan engine rotor speed sensor parameters (including hard failure, excessive noise, drift) and the neural network structure of the sensor time series fault classification model are as follows: Figure 3 As shown in the figure, unlike the dual-thread training process of the sensor parameter probability distribution function prediction model, long time series fault diagnosis adopts the scheme of offline training and offline use of neural network. First, the training data containing label y is used to calculate the cross entropy loss function. The neural network model for time series fault classification is trained and can be used offline after the model training is completed. The principle of time series fault classification of turbofan engine rotor speed sensor parameters is as follows: Figure 4 As shown, the turbofan engine receives the control quantity u t , the model outputs the low-pressure rotor speed X tThe sensor senses the data and stores it in the historical database. When the time t reaches the classification interval t%T=0, the time series is sampled from the historical database. and (considering double margins), and calculate the error sequence The mean and standard deviation of each sequence are extracted as the feature vector g = [μ1, μ2, μ e ,σ1,σ2,σ e ] T , the feature vector is used as the input of the time series fault classification model to obtain the classification result The index of the maximum value in corresponds to the fault model. For example, when , it is judged as failure mode 3 - drift.

[0057] In order to verify the effectiveness of the designed neural network-based aeroengine sensor fault self-diagnosis method, the engine component-level model dynamic link library is called in the Python environment for digital simulation. The component-level model structure is as follows: Figure 5 The overall framework of turbofan engine sensor fault self-diagnosis is shown in Figure 6 As shown, the initialization time is t=0, the historical access depth is p, and the fault classification interval is T. At each moment, the sensor measures the rotor speed and stores it in the historical database. At each moment, the sensor parameters of the previous p moments are accessed as the input of the probability distribution function prediction model to obtain the sensor parameter probability distribution function at this moment and determine whether a single-point mutation fault occurs. If a fault occurs, the probability distribution function mean is used to replace the mutation parameter to achieve parameter reconstruction. When the moment t reaches the fault classification interval t%T=0, the sensor parameters of the previous T moments are accessed for feature extraction and used as a time series fault classification model to achieve normal, hard failure, drift, and excessive noise fault classification. When the engine stops running, the entire turbofan engine sensor fault self-diagnosis process ends at the same time. Combined with the real engine data, the sensor measurement noise is set to 0.2% level and the sensor sampling step is set to 0.02s. The sensor parameter probability distribution function prediction model is used to diagnose the single-point mutation fault of the turbofan engine rotor speed sensor parameter, and the probability mean is used to realize signal reconstruction. When the sensor is fault-free, the accuracy rate of parameter single-point mutation fault diagnosis is 100%; when the sensor has a single-point mutation fault, the accuracy rate of parameter single-point mutation fault diagnosis is 99.8%. Figure 7The simulation diagram of the sensor parameter probability distribution function prediction model during an engine acceleration process is shown. In the 10-second simulation, the rotor speed sensor had a total of 3 single-point mutation faults. All faults were successfully diagnosed by the sensor parameter probability distribution function prediction model, and the sensor signal was reconstructed using the mean of the prediction distribution function. The sensor time series fault classification model is used to classify the turbofan engine rotor speed sensor parameters, and the total fault classification accuracy is 96.7%; when the training sample label is normal y = [1,0,0,0] T When the training sample label is hard failure y=[0,1,0,0], the fault classification accuracy is 99.6%; when the training sample label is hard failure y=[0,1,0,0] T When the training sample label is drifted y = [0,0,1,0], the fault classification accuracy is 99.6%; T When the training sample label is too noisy, the fault classification accuracy is 75.5%; when the training sample label is too noisy, y = [0,0,0,1] T When , the fault classification accuracy is 90.5%. Figure 8 The time series fault classification results of the rotor speed sensor parameters within 40 seconds of engine operation are shown when the sensor is operating normally (mode 0). Fig. 9 The time series fault classification results of the rotor speed sensor parameters when the sensor fails hard (mode 1) are shown. Fig.10 The time series fault classification results of the rotor speed sensor parameters in the case of sensor drift fault (mode 2) are shown. Fig.11 The time series fault classification results of the rotor speed sensor parameters when the sensor noise is too large (mode 3) are shown.

[0058] A method for self-diagnosis of aircraft engine sensor faults based on a neural network of the present invention comprises the following steps:

[0059] Step A: Establish an aircraft engine component-level model and collect a sensor historical parameter database;

[0060] Step B: design a sensor parameter probability distribution function prediction model based on a neural network. The model predicts the probability distribution function of the sensor parameter at the next moment based only on the local historical parameter data in the sensor's own historical parameter database, and determines whether a parameter mutation occurs based on the probability of the actual sensor parameter.

[0061] Step C: designing a sensor time series fault classification model based on a neural network, the model extracts feature vectors of the sensor's historical time series, intermittently determines whether the sensor has a fault and classifies the fault;

[0062] Step D: Combine the probability distribution mutation prediction model and the sensor fault classification model to build a sensor fault self-diagnosis framework, save and output the diagnosis results and reports.

[0063] The step A comprises the following steps:

[0064] Based on the principles of aviation engines, a component-level model of a certain type of engine is established. During the operation of the engine, sensor parameters are collected and stored in the database in time series for neural network training.

[0065] In step B, the specific steps of establishing a sensor parameter probability distribution function prediction model based on a neural network are as follows:

[0066] (1) Designing the neural network structure

[0067] Design neural network input and use the sensor parameter data within the historical access depth p as network input [X t-1 ,X t-2 ,X t-3 ,...,X t-p ] T , where X i Represents the sensor parameter data at time i; Design the neural network output, use the prediction vector as the network output, and the prediction vector dimension corresponds to the sample label with the mean μ and standard deviation σ of the normal distribution function as the network output, and construct the prediction value probability distribution function N(μ,σ); Select and optimize the number of hidden layers and the number of nodes in each layer of the neural network based on experience and experimental results; Select the activation function based on the overall structure and function of the neural network, select the ReLU function max(0,x) for the intermediate layer activation function, and select the tanh function for the output layer

[0068] (2) Training the Neural Network

[0069] Data batch sampling randomly selects part of the data from the database as the training set. Each training randomly samples N training samples from the training set, where the i-th sample feature is represented by [X i-1 ,X i-2 ,X i-3 ,...,X i-p ] T , the sample label is X i ; Design the loss function to increase the probability value P(X t |X t-1 ,X t-2 ,...,X t-p), the loss function should be positively correlated with this value. In order to reduce the training speed and improve the stability of the training process, the logarithmic function of P is selected to construct the loss function. Since the batch training method uses multiple samples for training at a time, the loss function is the expectation of the sampling probability of multiple samples L = -∑lnP; neural network parameter vector optimization, according to the designed loss function, the gradient descent method is used to optimize the neural network parameter vector θ with the learning rate α. The k-th training process can be expressed as

[0070] (3) Testing the neural network

[0071] Construct a test set and randomly select some data from the database as the test set; calculate the probability distribution of sample labels, input sample features into the neural network, and use the neural network to estimate the mean of the probability normal distribution With standard deviation Constructing estimated probability distributions Determine the mutation threshold and determine the mutation threshold range according to the 3σ criterion. The probability of the normal distribution value within (μ-3σ,μ-3σ] is 99.7%, and the probability of the sample label outside (μ-3σ,μ-3σ] is less than 0.3%, which is called a low-probability event. At this time, it can be determined that the sensor has mutated.

[0072] In step C, the specific steps of establishing a sensor long time series fault classification model based on a neural network are as follows:

[0073] (1) Analyze sensor failure modes

[0074] Sensor failure modes include: normal, the sensor operates normally; hard failure, the sensor does not work, and the parameters remain unchanged without external influences, which can be expressed as t f is the time when the fault occurs; drift, the sensor parameters have a fixed drift offset from the normal value, which can be expressed as X t =f(t)+bias; Noise: If the noise level is too large in the time series, it can be expressed as X t =f(t)+Noise;

[0075] (2) Designing the neural network structure

[0076] Design the neural network input, access the sensor parameter history time series within the sensor fault classification interval T, consider the double margin sensor (the parameters are represented by subscripts 1 and 2), and calculate the difference between the two sensor parameter sequences - the error sequence (the parameters are represented by subscripts e Represented), extract the mean and standard deviation of the eigenvalues ​​of each sequence and error sequence to form a eigenvector

[0077] g=[μ1,μ2,μ e ,σ1,σ2,σe ] T (1.3)

[0078] Design the output of the neural network. According to the fault classification, the sample label can be designed as a 4-dimensional Boolean vector to represent the fault mode. The output of the neural network is also a 4-dimensional vector.

[0079] The number of hidden layers and nodes in each layer of the neural network are selected and optimized based on experience and experimental results; the activation function is selected based on the overall structure and function of the neural network. The activation function of the intermediate layer is the ReLU function max(0,x), and the output layer is the softmax function to regularize the network output. where z j is the j-th output of the neural network without softmax function activation, is the i-th output after activation by the softmax function. The softmax activation function can convert the elements of the neural network output vector into elements less than 1 and whose sum is 1;

[0080] (3) Training the Neural Network

[0081] Randomly select some data from the sensor historical parameter database, extract the historical time series features as sample features, introduce fault signals into the historical data and give corresponding Boolean vector labels to form a training set;

[0082] Design loss function. In order to improve the accuracy of fault classification, combined with the characteristics of softmax activation function, the loss function is defined as the cross entropy between the network output vector and the sample label vector.

[0083] Neural network parameter vector optimization, according to the designed loss function, the gradient descent method is used to optimize the neural network parameter vector θ with a learning rate α. The k-th training process can be expressed as

[0084] The step D comprises the following steps:

[0085] (1) Initialization

[0086] Initialization time t=0, historical access depth p, fault classification interval T;

[0087] (2) Probability distribution prediction and mutation judgment

[0088] The historical parameters within the historical access depth p before access are used, and the probability distribution function of the sensor parameters at time t is predicted using the probability prediction model, and combined with the actual sensor parameters to determine whether a mutation occurs;

[0089] (3) Fault classification

[0090] Access the historical parameters within the previous time p, use the probability prediction model to predict the probability distribution function of the sensor parameters at time t, and combine the actual sensor parameters to determine whether a mutation occurs; determine whether the fault classification interval time t%T=0 is reached. If so, access the historical parameters within the previous T time, extract the feature vector, and use the fault classification model to determine whether a fault occurs and perform fault classification.

[0091] It should be pointed out that the above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes and substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for self-diagnosis of aircraft engine sensor faults based on neural network, characterized in that: The method comprises the following steps: Step A: Establish an aircraft engine component-level model and collect a sensor historical parameter database; Step B: design a sensor parameter probability distribution function prediction model based on a neural network. The model predicts the probability distribution function of the sensor parameter at the next moment based only on the local historical parameter data in the sensor's own historical parameter database, and determines whether a parameter mutation occurs based on the probability of the actual sensor parameter. Step C: designing a sensor time series fault classification model based on a neural network, the model extracts feature vectors of the sensor's historical time series, intermittently determines whether the sensor has a fault and classifies the fault; Step D: Combining the probability distribution mutation prediction model and the sensor fault classification model, constructing a sensor fault self-diagnosis framework, saving and outputting the diagnosis results and reports; Step B includes the following steps: (1) Designing the neural network structure Design neural network input and use the sensor parameter data within the historical access depth p as network input [X t-1 ,X t-2 ,X t-3 ,...,X t-p ] T , where X i Represents the sensor parameter data at time i; Design the neural network output, use the prediction vector as the network output, and the prediction vector dimension corresponds to the sample label with the mean μ and standard deviation σ of the normal distribution function as the network output, and construct the prediction value probability distribution function N(μ,σ); Select and optimize the number of hidden layers and the number of nodes in each layer of the neural network based on experience and experimental results; Select the activation function based on the overall structure and function of the neural network, select the ReLU function max(0,x) for the intermediate layer activation function, and select the tanh function for the output layer (2) Training the Neural Network Data batch sampling randomly selects part of the data from the database as the training set. Each training randomly samples N training samples from the training set, where the i-th sample feature is represented by [X i-1 ,X i-2 ,X i-3 ,...,X i-p ] T , the sample label is X i ; Design the loss function to increase the probability value P(X t |X t-1 ,X t-2 ,...,X t-p ), the loss function should be positively correlated with this value. In order to reduce the training speed and improve the stability of the training process, the logarithmic function of P is selected to construct the loss function. Since the batch training method uses multiple samples for training at a time, the loss function is the expectation of the sampling probability of multiple samples L = -∑lnP; neural network parameter vector optimization, according to the designed loss function, the gradient descent method is used to optimize the neural network parameter vector θ with the learning rate α. The k-th training process is expressed as (3) Testing the neural network Construct a test set and randomly select some data from the database as the test set; calculate the probability distribution of sample labels, input sample features into the neural network, and use the neural network to estimate the mean of the probability normal distribution With standard deviation Constructing estimated probability distributions Determine the mutation threshold. According to the 3σ criterion, the mutation threshold range is determined. The probability of the normal distribution value within (μ-3σ,μ-3σ] is 99.7%, and the probability of the sample label outside (μ-3σ,μ-3σ] is less than 0.3%, which is called a low-probability event. At this time, it is judged that the sensor has mutated; In step B, only the local historical information of the sensor itself is used for self-diagnosis; in order to deal with the uncertainty and randomness caused by the change of the working condition of the aircraft engine, the estimated probability distribution function is constructed. Cover all uncertainties; use estimated probability size Instead of the traditional estimated residual To determine whether a mutation occurs, in the normal distribution, the 3σ criterion is used as the basis for mutation judgment.

2. The method for self-diagnosis of aircraft engine sensor faults based on a neural network as claimed in claim 1, characterized in that: Step A specifically includes: establishing a component-level model of a certain type of engine based on the principle of aircraft engines, collecting sensor parameters during the operation of the engine, and storing them in a database in time series order for neural network training.

3. The method for self-diagnosis of aircraft engine sensor faults based on neural network according to claim 1, characterized in that: Step C includes the following steps: (1) Analyze sensor failure modes Sensor failure modes include: normal, the sensor operates normally; hard failure, the sensor does not work, and the parameters remain unchanged without external influences, which is expressed as t f is the time when the fault occurs; drift, the sensor parameters have a fixed drift offset from the normal value, expressed as X t =f(t)+bias; Noise: The noise level is too large in the time series, expressed as X t =f(t)+Noise; (2) Designing the neural network structure Design the neural network input, access the sensor parameter history time series within the sensor fault classification interval T, consider the dual margin sensor, the parameters are represented by subscripts 1 and 2, and calculate the difference between the two sensor parameter sequences - the error sequence, the parameters are represented by subscripts e Indicates that the mean and standard deviation of the eigenvalues ​​of each sequence and error sequence are extracted to form a eigenvector g=[μ1,μ2,μ e ,σ1,σ2,σ e ] T (1.1) Design the neural network output. According to the fault classification, the sample label is designed to be a 4-dimensional Boolean vector, which represents the fault mode. The neural network output is also a 4-dimensional vector. The number of hidden layers and nodes in each layer of the neural network are selected and optimized based on experience and experimental results; the activation function is selected based on the overall structure and function of the neural network. The activation function of the intermediate layer is the ReLU function max(0,x), and the output layer is the softmax function to regularize the network output. where z j is the j-th output of the neural network without softmax function activation, is the i-th output after activation by the softmax function. The softmax activation function converts the elements of the neural network output vector into elements that are less than 1 and sum to 1; (3) Training the Neural Network Randomly select some data from the sensor historical parameter database, extract the historical time series features as sample features, introduce fault signals into the historical data and give corresponding Boolean vector labels to form a training set; Design loss function. In order to improve the accuracy of fault classification, combined with the characteristics of softmax activation function, the loss function is defined as the cross entropy between the network output vector and the sample label vector. Neural network parameter vector optimization, according to the designed loss function, the gradient descent method is used to optimize the neural network parameter vector θ with a learning rate α. The k-th training process is expressed as 4. The method for self-diagnosis of aircraft engine sensor faults based on neural network according to claim 1, characterized in that: Step D comprises the following steps: (1) Initialization Initialization time t=0, historical access depth p, fault classification interval T; (2) Probability distribution prediction and mutation judgment The historical parameters within the historical access depth p before access are used, and the probability distribution function of the sensor parameters at time t is predicted using the probability prediction model, and combined with the actual sensor parameters to determine whether a mutation occurs; (3) Fault classification Access the historical parameters within the previous time p, use the probability prediction model to predict the probability distribution function of the sensor parameters at time t, and combine the actual sensor parameters to determine whether a mutation occurs; determine whether the fault classification interval time t%T=0 is reached. If so, access the historical parameters within the previous T time, extract the feature vector, and use the fault classification model to determine whether a fault occurs and perform fault classification.

Citation Information

Patent Citations

  • Fault detection method applied in intermittent process

    CN108255656A

  • Depth calculation model for aero-engine gas circuit fault diagnosis

    CN110321603A

  • Time sequence prediction method based on GRU neural network

    CN110647980A

  • Aero-engine gas circuit fault diagnosis method based on deep learning and information fusion

    CN111368885A

  • Aero-engine cross-domain fault diagnosis method based on dynamic adversarial residual network

    CN115479773A