Adaptive threshold aero-engine sensor fault diagnosis method based on measurement parameters

Abstract

The invention discloses a self-adaptive threshold aero-engine sensor fault diagnosis method based on measurement parameters. The method comprises the following steps: establishing an analysis redundancy model of an engine input signal and a sensor signal by adopting an OSELM algorithm; a univariate controller based on model-free self-adaptive control is designed, on this basis, a tracker of a baseline model is designed, a fault threshold value of sensor fault diagnosis is calculated through baseline model output, a sensor fault threshold value with self-adaptive capacity within a certain working condition range in an envelope is obtained, and a fault judgment basis is provided. According to the method, the problem that a sensor fault diagnosis threshold is difficult to determine under the conditions of variable working environment parameters and random noise in a traditional sensor fault method is solved. The method employs historical data of a health sensor to train and update an engine baseline model. For a bias fault and a drift fault, the new algorithm realizes a higher fault detection rate, a lower false alarm rate and a smaller minimum detection fault amplitude. The method can adapt to variable working environments and a long service life, and provides a more accurate and reliable diagnosis basis for fault early warning of the engine sensor.

Description

1. Technical Field

[0001] The invention belongs to the technical field of aircraft engine fault diagnosis, and in particular relates to an aircraft engine sensor fault diagnosis method based on an adaptive threshold of measurement parameters. 2. Background technology

[0002] As the main means for the engine to obtain operating information, the accuracy of sensor measurement information is crucial to the normal operation of the engine control system and health management system. Most aircraft engine sensor components work in harsh environments with high temperature, high pressure, high speed, strong vibration and gas corrosion. Therefore, the sensor is a component that is prone to failure. Once the sensor fails, the engine performance will drop significantly, the engine will be damaged, endangering flight safety and causing catastrophic accidents. Therefore, timely and effective fault diagnosis plays an indispensable role in improving the reliability and safety of the control system. The demand for high reliability and low cost of aircraft engines has put forward higher requirements for engine health management technology, so a more intelligent and autonomous diagnosis, prediction and health management system must be developed. For sensor fault diagnosis, a reasonable diagnostic standard is required. The main difficulty in fault diagnosis is to design a threshold setting scheme that can effectively distinguish between faults and false alarms and balance the relationship between fault diagnosis detection rate and false alarm rate. Therefore, it is of great significance to study and design a threshold setting method for sensor fault diagnosis.

[0003] Traditional threshold setting methods require statistical analysis of a large amount of historical data of healthy sensors, and a series of experiments are required to obtain the diagnostic threshold when the sensor is working normally. This method considers a data model-based method to calculate the adaptive threshold of engine sensor failure. In the model-free adaptive control (MFAC) method, a new dynamic linearization method and a new concept called pseudo partial derivative (PPD) are used for discrete-time nonlinear systems. An equivalent dynamic linearization data model is established at each dynamic operating point of the closed-loop system, and then a controller is designed based on the equivalent virtual data model and the control system is theoretically analyzed to achieve adaptive control of the nonlinear system. The MFAC method has the following advantages, which makes it more suitable for the control problems of actual systems. First, MFAC is a data-driven control method that only relies on the real-time measurement data of the controlled system, rather than any mathematical model information of the controlled system. Second, the MFAC method is simple, has a small amount of calculation, is easy to implement, and has strong robustness. For strongly nonlinear systems with model uncertainty such as aircraft engines, MFAC has obvious advantages. Under the control of the MFAC method, an equivalent dynamic linearization data model will be established online at each dynamic operating point, which can represent the working characteristics of the operating point. A health tracker is built based on the MFAC method to track the output of aircraft engines in a healthy state. The adaptive threshold of the sensor is derived based on the data model built based on the MFAC method.

[0004] When a sensor fails, the data model based on the engine measurement parameters cannot represent the normal output of the sensor. Therefore, it is necessary to establish sensor redundancy to characterize the output of normal sensors. OSELM can make full use of the currently acquired data, recursively update the output weights of the neural network according to samples at different times, and update the network parameters online. Based on this feature, a mapping between the estimated variables and other signals can be established through OSELM. During engine maintenance, the data model is updated online according to the sensor output signal. In order to obtain a reliable sensor fault diagnosis threshold, improve the accuracy of sensor fault diagnosis, and reduce the false alarm rate, this paper proposes an adaptive threshold aircraft engine sensor fault diagnosis method based on measurement parameters. This method uses the characteristics of model-free adaptive control to establish a dynamic data model at each operating point to calculate the theoretical threshold of sensor fault diagnosis, provide a basis for sensor fault identification, and improve the reliability of the engine control system. 3. Summary of the invention

[0005] In view of the above technical problems, the present invention provides an adaptive threshold aircraft engine sensor fault diagnosis method based on measurement parameters, which solves the problem that the sensor fault diagnosis threshold is difficult to determine in the case of variable working environment parameters and random noise in traditional sensor fault methods. The controller often uses the speed as the control quantity, so the fault diagnosis of the speed sensor is particularly important. The engine sensor fault diagnosis that adapts to the changeable working environment and long service life has a positive role in improving the accuracy of sensor fault diagnosis, reducing the false alarm rate, and improving the engine control system.

[0006] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is:

[0007] A method for diagnosing faults of aircraft engine sensors based on adaptive thresholds of measurement parameters, characterized in that it comprises the following steps:

[0008] Step A) Based on the historical data of the selected engine sensor and the remaining sensors, an aircraft engine sensor analytical redundancy model is constructed using the OS-ELM algorithm group.

[0009] Step B) Design an MFAC controller based on the tracker of the baseline model in step (A).

[0010] Step C) Calculate sensor fault thresholds using the baseline model tracker and design aero-engine sensor fault judgment criteria.

[0011] Furthermore, in step A), the specific steps of constructing the aircraft engine sensor analytical redundancy model using the OS-ELM algorithm group according to the historical data of the selected engine sensor and the remaining sensors are as follows:

[0012] Step A1), select the sensors y1(t), ..., y1(t) for which the prediction model needs to be established. n (t), y k (t) as the output, y1(t), ...y k-1 (t), y k+1 (t), ..., y n (t) and fuel flow As input; according to the selected input and output, the sensor prediction model constructed by OS-ELM algorithm is:

[0013]

[0014] Among them, f1,…,f n It can be approximated by training the OS-ELM algorithm, where p is the embedding dimension. Since the aircraft engine can be approximated as a second-order object, p = 2.

[0015] Step A2) uses the engine component-level model to generate sensor training samples. The acquisition of sensor signals and the online training of the predictor are carried out simultaneously. In each sampling period, the predictor collects the outputs of each sensor and arranges and groups them according to different prediction units. The sensor measurement value to be estimated is used as the expected output, and the remaining sensor measurement signals and their historical signals are used as input to construct the training samples. The samples are trained online by the OS-ELM model, and recursively updated according to the continuously acquired new samples to obtain an accurate sensor baseline model.

[0016] Furthermore, in the step B), the MFAC controller is designed, and the specific steps of the tracker based on the baseline model in the MFAC controller design step (A) are as follows:

[0017] Step B1), selecting the control variable and the controlled variable of the MFAC controller designed for the engine, using the control input criterion function and the PPD estimation standard function to solve the control law and the PPD estimation algorithm respectively, and obtaining the MFAC-based engine controller.

[0018] Its control law is:

[0019]

[0020] if or

[0021]

[0022] Among them, σ1>0, σ2>0 are weight coefficients used to limit and control the amplitude of the input signal change; y d (k+1) is the expected output signal ε, which is a sufficiently small positive number; η2∈(0,1) is the step size factor; is the estimate of the pseudo partial derivative φ(k); for The initial value of .

[0023] Step B2), designing a nonlinear tracker based on the MFAC controller to track the obtained baseline model.

[0024] The designed trackers are:

[0025]

[0026] Where ρ is the feedback gain, is the estimation error, Δu N (k) is the change in engine control output when the sensor is normal.

[0027] Furthermore, the step C) uses the baseline model tracker to calculate the sensor fault threshold and designs the aircraft engine sensor fault judgment criterion. The specific steps are as follows:

[0028] Step C1), define the system for normal operation of the engine as:

[0029] y N (k+1)=y N (k)+φ N T (k)Δu N (k)

[0030] Substituting this into the tracker gives

[0031]

[0032] in

[0033] Step C2) After scaling, the

[0034]

[0035] in

[0036] Step C3) introduces a sliding window mechanism to further reduce the threshold value and the minimum detection fault amplitude

[0037]

[0038] Where n is the length of the sliding window, which is used to adjust the time span of the historical data used.

[0039]

[0040] in

[0041] Step C4) obtaining the sensor fault threshold and sensor fault judgment criteria:

[0042] The sensor failure thresholds are:

[0043] d f_ad (k) = (1-w)d f_slid (k)

[0044] w=1 / (1+e -q(k) )

[0045]

[0046] Where std is the standard deviation of sensor noise, is the error between the sensor measurement and the baseline model prediction.

[0047] The sensor fault judgment criteria are:

[0048]

[0049] The common logic of sensor fault diagnosis is to analyze whether the prediction residual exceeds a given threshold. At this point, the design of the sensor fault diagnosis method based on adaptive threshold is completed.

[0050] Beneficial effects: The adaptive threshold aircraft engine sensor fault diagnosis method based on measurement parameters provided by the present invention has the following technical effects compared with the prior art by using the above technical solution:

[0051] (1) There is no need to conduct a large number of statistical and simulation experiments on the sensor historical data as in the traditional threshold setting method. Instead, the sensor historical data is used to establish a dynamic linear model online at each working point, and the theoretical adaptive threshold of the diagnosis is calculated based on the data model.

[0052] (2) The OSELM algorithm provides analytical redundancy for the sensors used in turboprop engines, has good prediction accuracy during acceleration, deceleration, and steady-state processes, and improves the reliability of the baseline model of the high-pressure turbine speed sensor.

[0053] (3) The threshold setting method can be applied to a wider engine operating range, and the threshold can be adjusted according to the sensor data under various engine operating conditions. Therefore, it has flexible and convenient working characteristics.

[0054] (4) In view of the slight deviation fault of the engine sensor, a threshold correction coefficient is introduced so that the threshold setting method can still maintain a good false alarm rate and fault detection rate. 4. Description of the drawings

[0055] Figure 1 It is a structural diagram of the adaptive threshold aircraft engine sensor fault diagnosis based on measurement parameters of the present invention.

[0056] Figure 2 It is the simulated fuel flow rate variation diagram.

[0057] Figure 3 It is a comparison chart of the predicted and actual output curves of a turboprop engine.

[0058] Figure 4 This is the control effect diagram of the MFAC controller.

[0059] Figure 5 is the tracking error diagram of the tracker under sensor failure.

[0060] Figure 6 This is the Nh sensor slight offset fault diagnosis diagram.

[0061] Figure 7 It is the Nh sensor drift fault diagnosis diagram. 5. Specific implementation methods

[0062] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings.

[0063] The invention describes a method for diagnosing faults of aircraft engine sensors based on adaptive thresholds of measurement parameters, characterized in that it comprises the following steps:

[0064] Step A) Based on the historical data of the selected engine sensor and the remaining sensors, an aircraft engine sensor analytical redundancy model is constructed using the OS-ELM algorithm group.

[0065] Step B) Design an MFAC controller based on the tracker of the baseline model in step (A).

[0066] Step C) Calculate sensor fault thresholds using the baseline model tracker and design aero-engine sensor fault judgment criteria.

[0067] Step A1), select the sensors y1(t), ..., y1(t) for which the prediction model needs to be established. n (t), y k (t) as the output, y1(t), ...y k-1 (t), y k+1 (t), ..., y n (t) and fuel flow As input; according to the selected input and output, the sensor prediction model constructed by OS-ELM algorithm is:

[0068]

[0069] Among them, f1,…,f n It can be approximated by training the OS-ELM algorithm, where p is the embedding dimension. Since the aircraft engine can be approximated as a second-order object, p = 2.

[0070] Step A2) uses the engine component-level model to generate sensor training samples. The acquisition of sensor signals and the online training of the predictor are carried out simultaneously. In each sampling period, the predictor collects the outputs of each sensor and arranges and groups them according to different prediction units. The sensor measurement value to be estimated is used as the expected output, and the remaining sensor measurement signals and their historical signals are used as input to construct the training samples. The samples are trained online by the OS-ELM model, and recursively updated according to the continuously acquired new samples to obtain an accurate sensor baseline model.

[0071] Step B1), selecting the control variable and the controlled variable of the MFAC controller designed for the engine, using the control input criterion function and the PPD estimation standard function to solve the control law and the PPD estimation algorithm respectively, and obtaining the MFAC-based engine controller.

[0072] Its control law is:

[0073]

[0074] if

[0075]

[0076] Among them, σ1>0, σ2>0 are weight coefficients used to limit and control the amplitude of the input signal change; y d (k+1) is the expected output signal ε, which is a sufficiently small positive number; η2∈(0,1) is the step size factor; is the estimate of the pseudo partial derivative φ(k); for The initial value of .

[0077] Step B2), designing a nonlinear tracker based on the MFAC controller to track the obtained baseline model.

[0078] The designed trackers are:

[0079]

[0080] Where ρ is the feedback gain, is the estimation error, Δu N (k) is the change in engine control output when the sensor is normal.

[0081] Step C1), define the system for normal operation of the engine as:

[0082] y N (k+1)=y N (k)+φ N T(k)Δu N (k)

[0083] Substituting this into the tracker gives

[0084]

[0085] in

[0086] Step C2) After scaling, the

[0087]

[0088] in

[0089] Step C3) introduces a sliding window mechanism to further reduce the threshold value and the minimum detection fault amplitude

[0090]

[0091] Where n is the length of the sliding window, which is used to adjust the time span of the historical data used.

[0092]

[0093] in

[0094] Step C4) obtaining the sensor fault threshold and sensor fault judgment criteria:

[0095] The sensor failure thresholds are:

[0096] d f_ad (k) = (1-w)d f_slid (k)

[0097] w=1 / (1+e -q(k) )

[0098]

[0099] Where std is the standard deviation of sensor noise, is the error between the sensor measurement and the baseline model prediction.

[0100] The sensor fault judgment criteria are:

[0101]

[0102] The common logic of sensor fault diagnosis is to analyze whether the prediction residual exceeds a given threshold. At this point, the design of the sensor fault diagnosis method based on adaptive threshold is completed.

[0103] In order to verify the effectiveness of the adaptive threshold aircraft engine sensor fault diagnosis method based on measurement parameters designed in the present invention, the baseline model output accuracy simulation of the engine takeoff and acceleration and deceleration processes within a certain envelope and the digital simulation of the high-pressure compressor speed sensor fault diagnosis were carried out in the MATLAB environment.

[0104] The present invention uses a nonlinear component model of a certain type of three-rotor turboprop engine as the research object. The model is constructed by object-oriented programming method, including important components of the aircraft engine such as the air inlet, fan, compressor, combustion chamber, gas turbine, power turbine and tail nozzle, and is easy to call in the MATLAB environment.

[0105] According to the historical data of the selected engine sensor and the remaining sensors, the OS-ELM algorithm group is used to build an aircraft engine sensor analytical redundancy model.

[0106] First, the high-pressure compressor outlet pressure, high-pressure compressor outlet temperature, and power turbine outlet pressure are selected to build a predictor, and OSELM is used as the signal predictor. The expression of each sensor signal prediction model is as follows:

[0107] N h Predictor:

[0108]

[0109] P3 Predictor:

[0110]

[0111] T3 Predictor:

[0112]

[0113] P5 Predictor:

[0114]

[0115] The historical sensor data of turboprop engines in service is stored in a database. The OSELM algorithm uses the sensor data in the database to train the prediction model of each sensor to obtain a baseline model of the engine.

[0116] The experiment adopts an open-loop control process, which is conducive to eliminating the influence of other interferences. Figure 2The fuel flow rate shown in the figure is used as the input to obtain the output state and performance parameters. The simulation lasts for 80 seconds, and the fuel flow rate changes from 40% to 100%, and goes through acceleration, steady state and deceleration processes in turn. The engine operating environment is set to H = 0km, Ma = 0. All sensors are normalized. In order to simulate the working environment of the engine in the actual process, the mean and variance are set to 0 and 0.003 respectively. 2 Gaussian noise is added to the sensor output. For each sensor prediction model, the number of hidden neurons is set to 21, and the activation function is the sigmoid function. In order to evaluate the accuracy of the sensor predictor, RMSE is selected as the evaluation indicator of the algorithm prediction performance. Its training and test accuracy are shown in Table 1. Figure 3 The prediction curve of the prediction model is given in .

[0117] Table 1 Training and testing accuracy of turboprop engine dataset

[0118]

[0119] According to the fault characteristics, bias fault and drift fault are taken as typical fault modes. The following model is used for simulation.

[0120] Bias Fault:

[0121]

[0122] Among them, y t is the sensor output, is the true value of the sensor, Δy t is the offset, t is the current time, t0 is the time when the fault occurs, and 1(·) is the unit step function.

[0123] Drift fault:

[0124]

[0125] Among them, ε t is the drift rate.

[0126] The sensor fault diagnosis simulation is simulated using the simulation conditions in the baseline model. The parameters of MFAC are set as follows: The initial value of the adaptive law Design parameters σ1=2, σ2=1, η1=0.3, η2=1, ε=10 -5 , the feedback gain of the tracker ρ=0.3, and the length of the sliding window n=5. The experiment includes bias fault simulation and drift fault simulation. The relevant fault parameters are set as follows: Δy t =0.008,ε t=0.002 / s. In order to evaluate the adaptive threshold setting method proposed in this paper, the performance of the adaptive threshold and two fixed threshold methods are compared. The two fixed thresholds selected are d max and d 3*std Threshold d max It is the maximum value under normal circumstances. 3*std The threshold is 3 times the standard deviation of the noise.

[0127] The control effect of MFAC in closed-loop control of a given desired speed is as follows: Figure 4 As shown, from Figure 4 It can be seen that the MFAC controller can quickly adjust to the reference value within a wide operating range, and the steady-state error during the simulation is close to zero. The adjustment time of the engine acceleration and deceleration dynamic process is also shorter, and satisfactory dynamic performance is obtained, which ensures that the high-pressure gas turbine speed can accurately run to the target speed and achieve the ideal power of the turboprop engine. The change of the control quantity is also within a reasonable range.

[0128] Figure 5 The figure shows how well the MFAC-based tracker tracks the baseline model under both bias sensor fault and drift sensor fault. As can be seen from the figure, the tracking error is within 1% and the tracker has a high estimation accuracy. Therefore, the theoretical threshold calculated by the tracker can represent the range of the sensor under normal conditions.

[0129] Figure 6 The speed change curve of the engine from 88% gas turbine speed to the design point speed is shown. Within 60-70 seconds, there is a bias fault in the speed sensor. For a small bias fault with a fault amplitude of 0.008, the method has a high fault detection rate.

[0130] exist Figure 7 In the figure, when the drift fault occurs at 60 seconds, the adaptive threshold has difficulty detecting the sensor fault because the fault amplitude is small at the beginning. As the fault time increases, the fault amplitude of the drift fault gradually increases. When the fault amplitude reaches 0.00328, the adaptive threshold begins to detect the sensor fault.

[0131] Table 2 lists the specific performance indicators of the three methods.

[0132] Table 2 Sensor fault diagnosis performance indicators

[0133]

[0134] Therefore, the proposed adaptive threshold aero-engine sensor fault diagnosis method based on measurement parameters has good sensor fault diagnosis accuracy.

[0135] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of aircraft engine sensors based on adaptive thresholds of measurement parameters, characterized in that: The following steps are involved: Step A) Based on the historical data of the selected engine sensor and the remaining sensors, an aircraft engine sensor analytical redundancy model is constructed using the OS-ELM algorithm group. Step B) Design an MFAC controller based on the tracker of the baseline model in step (A). Step C) Calculate the sensor fault threshold using the baseline model tracker and design the aero-engine sensor fault judgment criteria.

2. The method for diagnosing aero-engine sensor faults based on adaptive thresholds of measurement parameters according to claim 1, characterized in that: In step A), the specific steps of constructing an aircraft engine sensor analytical redundancy model using the OS-ELM algorithm group according to the historical data of the selected engine sensor and the remaining sensors are as follows: Step A1), select the sensors y1(t), ..., y1(t) for which the prediction model needs to be established. n (t), y k (t) as the output, y1(t), ...y k-1 (t), y k+1 (t), ..., y n (t) and fuel flow As input; according to the selected input and output, the sensor prediction model constructed by OS-ELM algorithm is: Among them, f1,…,f n It can be approximated by training the OS-ELM algorithm, where p is the embedding dimension. Since the aircraft engine can be approximated as a second-order object, p = 2. Step A2) uses the engine component-level model to generate sensor training samples. The acquisition of sensor signals and the online training of the predictor are carried out simultaneously. In each sampling period, the predictor collects the outputs of each sensor and arranges and groups them according to different prediction units. The sensor measurement value to be estimated is used as the expected output, and the remaining sensor measurement signals and their historical signals are used as input to construct the training samples. The samples are trained online by the OS-ELM model, and recursively updated according to the continuously acquired new samples to obtain an accurate sensor baseline model.

3. The method for diagnosing aero-engine sensor faults based on adaptive thresholds of measurement parameters according to claim 1, characterized in that: The specific steps of designing the MFAC controller in step B) and designing the tracker of the baseline model based on the MFAC controller are as follows: Step B1), selecting the control variable and the controlled variable of the MFAC controller designed for the engine, using the control input criterion function and the PPD estimation standard function to solve the control law and the PPD estimation algorithm respectively, and obtaining the MFAC-based engine controller. Its control law is: if or Among them, σ1>0, σ2>0 are weight coefficients used to limit and control the amplitude of the input signal change; y d (k+1) is the expected output signal ε, which is a sufficiently small positive number; η2∈(0,1) is the step size factor; is the estimate of the pseudo partial derivative φ(k); for The initial value of . Step B2), designing a nonlinear tracker based on the MFAC controller to track the obtained baseline model. The designed trackers are: Where ρ is the feedback gain, is the estimation error, Δu N (k) is the change in engine control output when the sensor is normal.

4. The intelligent multivariable control method for an aircraft engine with online correction according to claim 1, characterized in that: The specific steps of step C) using the baseline model tracker to calculate the sensor fault threshold and designing the aeroengine sensor fault judgment criterion are as follows: Step C1), define the system for normal operation of the engine as: and N (k+1)=y N (k)+φ N T (k)Δu N (k) Substituting this into the tracker gives in Step C2) After scaling, the in Step C3) introduces a sliding window mechanism to further reduce the threshold value and the minimum detection fault amplitude Where n is the length of the sliding window, which is used to adjust the time span of the historical data used. The sensor failure thresholds are: d f_ad (k)=(1-w)d f_slid (k) w=1 / (1+e -q(k) ) Where std is the standard deviation of sensor noise, is the error between the sensor measurement value and the baseline model prediction value. The sensor fault judgment criterion is: The common logic of sensor fault diagnosis is to analyze whether the prediction residual exceeds a given threshold. When , the sensor failure of the aircraft engine can be detected at k.

Images