A method for predicting the remaining lifespan of aero-engines based on dual-particle filter hybrid prediction

Abstract

This invention discloses a method for predicting the remaining life of aero-engines based on a dual-particle filter hybrid model, belonging to the field of data prediction technology. It includes the following steps: data preprocessing, model building, performing MO-RVM modeling on each existing measurement data point, and establishing an initial... n y A MO-RVM prediction model is used; based on the prediction results, it is determined whether to dynamically update the model; health parameters are predicted: the system state and hidden health parameters are predicted using a state / health parameter two-particle filter, realizing the prediction of state and health parameters. k Forward prediction; Remaining lifetime prediction: Plot the probability density function of the health parameter prediction particles in the DPF to obtain the moment with the highest probability. First, a data-driven method based on dynamic updates is used to achieve long-term measurement data prediction. Then, a two-particle filter is used to estimate the predicted values ​​of the health parameters, thereby predicting the remaining lifetime from the current moment to the system failure moment, improving the accuracy and reliability of lifetime prediction.

Description

Technical Field

[0001] This invention relates to the field of data prediction technology, and in particular to a method for predicting the remaining life of aero-engines based on a hybrid two-particle filter. Background Technology

[0002] As the heart of an aircraft, the performance stability and reliability of an aero-engine are directly related to aircraft safety. With the rapid development of the aviation industry, the demand for real-time monitoring and life prediction of aero-engine performance is increasing. Traditional life prediction methods mainly rely on empirical formulas or simple statistical analysis, which often ignore the complexity and variability of engine operation.

[0003] Existing technical solutions mainly include analytical methods based on physical models and statistical methods based on data-driven approaches. Physical model-based methods predict performance trends by constructing physical equations for engine performance. However, these methods require detailed physical parameters and accurate initial conditions, and in practical applications, the difficulty in directly measuring the internal state of the engine limits their accuracy and practicality.

[0004] Data-driven methods, such as artificial neural networks and support vector machines, build predictive models by analyzing historical data. While these methods can reflect the statistical regularities of engine performance to some extent, they typically lack a deep understanding of the internal physical processes of the engine and have high requirements for data quality and quantity. Their predictive effectiveness is limited when data is insufficient or of poor quality.

[0005] Considering both safety and economic factors, the rapid development of the aviation industry has led to an increasingly urgent need for fault diagnosis, prediction, and health management in the aviation field. However, research in this area in my country is still in its early stages, exhibiting a mismatch between demand and reality. Therefore, this invention focuses on the aviation field, specifically aero-engines, and studies the application of a hybrid prediction method based on a two-particle filter in aero-engines, thereby improving the remaining service life prediction method. Summary of the Invention

[0006] This invention addresses the shortcomings of existing data-driven and physics-based prediction methods, as well as the problem that traditional observers only estimate system states and cannot directly obtain health parameters. It provides a method for predicting the remaining lifespan of aero-engines based on a hybrid two-particle filter approach. This method combines the advantages of data-driven and physics-based approaches, estimating the state and health parameters of the aero-engine using a two-particle filter and performing long-term predictions based on the measurement data using a dynamically updated multi-output correlation vector machine, effectively improving the accuracy and reliability of lifespan prediction.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A method for predicting the remaining life of an aero-engine based on a hybrid two-particle filter includes the following steps:

[0009] Step 1: Data preprocessing and model building: Select the state x, observed sensor measurements y, and health parameter θ for calculation and estimation. Perform MO-RVM modeling on each existing measurement data y to establish the initial n. y One MO-RVM prediction model;

[0010] Step 2: Based on the prediction results of the prediction model, decide whether to perform dynamic model updates: Determine whether the MO-RVM model needs to be updated according to the state update rules. If the result indicates that an update is needed, then the DMO-RVM model will be further established, and then the sensor observations will be predicted. If dynamic updates are not needed, then prediction will be performed directly.

[0011] Step 3: Predict health parameters: Use the state / health parameter dual-particle filter to predict the system state and hidden health parameters based on the sensor observations predicted in Step 2, and achieve k-step forward prediction of state and health parameters.

[0012] Step 4, Remaining Lifetime Prediction: Using the predicted health parameter values ​​obtained in Step 3, plot the probability density distribution of the predicted health parameters. By plotting the probability density function of the predicted particles of the health parameters in the DPF, the moment with the highest probability is obtained, which is the estimated remaining lifetime.

[0013] Furthermore, in step 1, MO-RVM modeling is performed on each existing initial measurement data to obtain n. y The initial prediction models are as follows: According to the formula:

[0014] ......

[0016] The initial k-step forward prediction result is obtained through iteration: P represents the MO-RVM prediction model.

[0017] Furthermore, the state update rule in step 2 is determined by the difference between the predicted and actual measured values. This difference is quantified using Mahalanobis distance, which is defined as follows:

[0018]

[0019] In the formula, Σ is the covariance matrix of the multidimensional random variables, μ is the mean of the reference sample, and () TLet x be the transpose of the matrix, and x be the state value.

[0020] Furthermore, the judgment criterion in step 2 is as follows: the average Mahalanobis distance of the i-th measurement value within the monitoring window m is D. M (m,i) If any of the n average distances is greater than the threshold σ1 and is σ2 times greater than the average Mahalanobis distance of the previous window, it is determined that an update is needed. The determination rule is as follows:

[0021]

[0022] In the formula: σ1 is threshold 1, used to determine the magnitude of the distance, and σ2 is threshold 2, used to determine the rate of change of the distance;

[0023] Since it is not necessary to update the prediction model for all n measurements, a sub-rule is established under the above decision rule for secondary decision-making. The sub-rule is as follows:

[0024] D M (m,j) <σ3

[0025] In the formula: σ3 is the second-order decision threshold, and σ3 < σ1. When the above sub-rule is satisfied, the decision does not need to update the i-th measurement value.

[0026] Furthermore, in step 3, the prediction process is as follows:

[0027] For time steps r = 1, ..., k:

[0028] (1) Perform r-step forward prior state prediction:

[0029] according to Predict the prior state of the particle, where The particle noise added to the state particle at time t+r has a distribution that conforms to... and These are the estimated values ​​of the state x and the health parameter θ at time t+r-1, respectively.

[0030] according to Estimate the prior state covariance matrix;

[0031] (2) Perform r-step forward and backward state prediction:

[0032] Through resampling, calculation

[0033] Obtain the weights of the filter for estimating the state x at time t+r. in, for The probability density function;

[0034] Calculate the approximate posterior state distribution:

[0035] Where Nreg is the number of regularizations, A t Depend on Calculated;

[0036] Based on filter weights Resampling state to estimate particles, this is using The resulting posterior state estimate distribution after approximate resampling is: δ(·) is the Dirac function;

[0037] according to Approximate posterior state estimation;

[0038] (3) Perform the first r-step forward and backward parameter prediction:

[0039] Calculate the corrected output prediction error h t (·) represents the nonlinear equation between state, health parameters, and measured values;

[0040] Calculate correction parameters in, These are the adaptive coefficients during the calibration process. Let be the Jacobian determinant of the j-th particle;

[0041] The kernel shrinkage algorithm is used to calculate the first step of posterior parameter prediction:

[0042] A is the contraction matrix.

[0043]

[0044] The noise distribution added to each particle;

[0045] (4) Perform the second r-step forward and backward parameter prediction:

[0046] Calculate parameters to estimate filter weights

[0047] Perform resampling.

[0048] The first posterior parameter distribution is approximated using the following method.

[0049]

[0050] Particle prediction based on resampling parameters Resampling is performed, and the approximate posterior parameter prediction distribution is as follows:

[0051] Posterior parameter prediction

[0052] Calculate the posterior parameter prediction covariance as follows:

[0053] Based on the above prediction process, the predicted measurement data is used, and resampling is maintained for future moments to generate the health parameter particle probability density function predicted k steps forward.

[0054] Furthermore, in step 4, for D-dimensional data x, the formula for calculating the health parameter particle probability density function P(x|θ) is:

[0055]

[0056] Where μ is the mean of x, Σ is the covariance of x, and x is the state parameter of the engine.

[0057] Furthermore, in step 4, the formula for calculating the remaining lifetime RUL(t) at time t is:

[0058] RUL(t) = t j -t

[0059]

[0060] In the formula: t j The time it takes for any relevant parameter to reach a critical value. The parameter threshold at which failure occurs. Life expectancy is the healthy lifespan.

[0061] The present invention has the following advantages:

[0062] 1. By collecting and integrating available historical measurement data, a dynamic multi-output correlation vector machine prediction model (DMO-RVM) is established to predict sensor observations. Based on these observations, the system state and hidden health parameters are predicted using a state / parameter two-particle filter (DPF). Furthermore, K-step forward prediction of the state and health parameters is achieved, predicting the remaining lifetime from the current moment to the system failure moment.

[0063] 2. An initial dynamic prediction model, DMO-RVM, is established based on historical data, and k-step forward predictions are performed to obtain measured prediction values, thereby improving the adaptability of the MO-RVM model to changes in system state and thus improving prediction accuracy. Based on this, the state and health parameters are predicted using a dynamic prediction function (DPF). The design goal of the DPF is to address the convergence of the estimator, enabling it to reach the true optimal estimate.

[0064] 3. Compared to traditional health monitoring schemes that estimate state parameters using particle filters before making judgments, dual-particle filters can achieve joint estimation of state and health parameters. When directly monitoring health parameters, the failure threshold is easier to obtain and more intuitive. Furthermore, the proposed hybrid prediction method, compared to prediction schemes based on physical models and data-driven algorithms, has stronger practicality and stability, and can more accurately predict the remaining lifespan of the system. Attached Figure Description

[0065] Figure 1 : A schematic diagram of the process of this invention.

[0066] Figure 2 Schematic diagram of a two-particle filter.

[0067] Figure 3 : A diagram illustrating the dynamically updated sample monitoring window.

[0068] Figure 4 : Schematic diagram of the probability density distribution of predicted remaining lifetime. The actual remaining lifetime is 220.

[0069] Figure 5 : Schematic diagram of the probability density distribution of predicted remaining lifetime. The actual remaining lifetime is 460.

[0070] Figure 6 : Schematic diagram of the probability density distribution of predicted remaining lifetime. The actual remaining lifetime is 740.

[0071] Figure 7 Schematic diagram of remaining life prediction under fouling conditions in high-pressure compressors.

[0072] Figure 8 : Schematic diagram of the process for predicting measurement data and health parameters based on dynamic models.

[0073] Figure 9 : Schematic diagram of the update process of some measurement values ​​under the damage caused by scale buildup.

[0074] Figure 10 Schematic diagram of dynamic estimation measurements under scale damage.

[0075] Figure 11 Comparison chart of long-term predictions of measurements based on DMO-RVM and ARMA. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0077] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0078] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0079] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0080] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. In addition, the terms "first," "second," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0081] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0082] refer to Figure 1-10 As shown, one embodiment of the present invention is as follows:

[0083] The method for predicting the remaining life of aero-engines based on dual-particle filter hybrid prediction includes the following steps:

[0084] Step 1: Data preprocessing and model building: Select the state x, observed sensor measurements y, and health parameter θ for calculation and estimation. Perform MO-RVM modeling on each existing measurement data y to establish the initial n. y One MO-RVM prediction model.

[0085] Where, state x = [T CC ,N1,N2,P LT ,P CC ,P LC ,P HT ], where T CC The combustion chamber temperature is given by N1 and N2, which are the shaft speeds connecting the high-pressure turbine and the high-pressure compressor, and the low-pressure turbine and the low-pressure compressor, respectively. P is the combustion chamber temperature. LT P CC P LC and P HT These represent the pressures of the low-pressure turbine, combustion chamber, low-pressure compressor, and high-pressure turbine, respectively.

[0086] Observe the sensor measurement value y=[N1,N2,P HC ,P LC ,T HC ,T LC ,T HT ,T LT ], where T HC T LC T HT and T LT P represents the outlet temperatures of the high-pressure compressor, low-pressure compressor, high-pressure turbine, and low-pressure turbine, respectively. HC This indicates the pressure of the high-pressure air compressor.

[0087] Health parameters

[0088] The efficiency θ of high-pressure compressor (HC), low-pressure compressor (LC), high-pressure turbine (HT), and low-pressure turbine (LT) η and traffic

[0089] For each existing initial measurement data point, MO-RVM modeling is performed to obtain n. y The initial prediction models are as follows: According to the formula:

[0090] ......

[0092] The initial k-step forward prediction result is obtained through iteration: P represents the MO-RVM prediction model.

[0093] Step 2: Based on the prediction results of the prediction model, decide whether to perform dynamic model updates: Determine whether the MO-RVM model needs to be updated according to the state update rules. If the result indicates that an update is needed, then the DMO-RVM model will be further established, and then the sensor observations will be predicted. If dynamic updates are not needed, then prediction will be performed directly.

[0094] The need to update the MO-RVM model in the current data window m is determined based on the state update rules. These rules are determined by the difference between the predicted and actual measurements, which is quantified using Mahalanobis distance. The Mahalanobis distance for a single-point comparison is defined as follows:

[0095]

[0096] In the formula, Σ is the covariance matrix of the multidimensional random variables, μ is the mean of the reference sample, and () T Let x be the transpose of the matrix, and x be the state value. The larger the Mahalanobis distance, the farther the distance between the observation point and the sample. Therefore, when the Mahalanobis distance in the above formula is greater than the threshold, the MO-RVM prediction model is updated.

[0097] First, select the current data window to be updated. Second, calculate the Mahalanobis distance of all measurements within the monitoring window. Finally, update the prediction model using the following criteria.

[0098] The average Mahalanobis distance of the i-th measurement within the monitoring window m is

[0099] D M (m,i) When any one of the n average distances is greater than the threshold σ1, and is also greater than the average Mahalanobis distance of the previous window...

[0100] When σ is multiplied by 2, it is determined that an update is needed, and the determination rules are as follows:

[0101]

[0102] In the formula: σ1 is threshold 1, used to determine the magnitude of the distance, and σ2 is threshold 2, used to determine the rate of change of the distance;

[0103] Since it is not necessary to update the prediction model for all n measurements, a sub-rule is established under the above decision rule for secondary decision-making. The sub-rule is as follows:

[0104] D M (m,i) <σ3

[0105] In the formula: σ3 is the second-order decision threshold, and σ3 < σ1. When the above sub-rule is satisfied, the decision does not need to update the i-th measurement value.

[0106] Step 3: Predict Health Parameters: Using a state / health parameter dual-particle filter, the system state and hidden health parameters are predicted based on the sensor observations predicted in Step 2, achieving k-step forward prediction of the state and health parameters. The k-step forward prediction of state and health parameters is implemented iteratively, and the prediction process is as follows:

[0107] For time steps r = 1, ..., k:

[0108] (1) Perform r-step forward prior state prediction:

[0109] according to

[0110] Predict the prior state of the particle, where The particle noise added to the state particle at time t+r has a distribution that conforms to... and These are the estimated values ​​of the state x and the health parameter θ at time t+r-1, respectively.

[0111] according to Estimate the prior state covariance matrix;

[0112] (2) Perform r-step forward and backward state prediction:

[0113] Through resampling, calculation

[0114] Obtain the weights of the filter for estimating the state x at time t+r. in, for The probability density function;

[0115] Approximate posterior state distribution calculation

[0116] Where Nreg is the number of regularizations, A t Depend on Calculated;

[0117] Based on filter weights Resampling state to estimate particles, this is using The resulting posterior state estimate distribution after approximate resampling is: δ(·) is the Dirac function;

[0118] according to Approximate posterior state estimation;

[0119] (3) Perform the first r-step forward and backward parameter prediction:

[0120] Calculate the corrected output prediction error h t (·) represents the nonlinear equation between state, health parameters, and measured values;

[0121] Calculate correction parameters in, These are the adaptive coefficients during the calibration process. Let be the Jacobian determinant of the j-th particle;

[0122] The kernel shrinkage algorithm is used to calculate the first step of posterior parameter prediction:

[0123] A is the contraction matrix.

[0124]

[0125] The noise distribution added to each particle;

[0126] (4) Perform the second r-step forward and backward parameter prediction:

[0127] Calculate parameters to estimate filter weights

[0128]

[0129] Perform resampling.

[0130] The first posterior parameter distribution is approximated using the following method.

[0131]

[0132] Particle prediction based on resampling parameters Resampling is performed, and the approximate posterior parameter prediction distribution is as follows:

[0133] Posterior parameter prediction

[0134] Calculate the posterior parameter prediction covariance as follows:

[0135] Based on the above prediction process, the predicted measurement data is used, and resampling is maintained for future moments to generate the health parameter particle probability density function predicted k steps forward.

[0136] Step 4, Remaining Lifetime Prediction: Using the predicted health parameter values ​​obtained in Step 3, plot the probability density distribution of the predicted health parameters. By plotting the probability density function of the predicted particles of the health parameters in the DPF, the moment with the highest probability is obtained, which is the estimated remaining lifetime.

[0137] Using the predicted health parameters obtained in step 3, plot the probability density distribution of the predicted health parameters. For D-dimensional data x, the formula for calculating the particle probability density function P(x|θ) of the health parameters is:

[0138]

[0139] Where μ is the mean of x, Σ is the covariance of x, and x is the state parameter of the engine. It is assumed that the data follows a Gaussian distribution. By plotting the probability density function of the particles predicted by the health parameters in the DPF, the moment with the highest probability is obtained, and this moment is considered the estimated remaining lifetime.

[0140] Step 4, when calculating the remaining lifespan of the system using the obtained health parameters, also requires determining the time of failure. The formula for calculating the remaining lifespan RUL(t) at time t is:

[0141] RUL(t) = t j -t

[0142]

[0143] In the formula: t j The time it takes for any relevant parameter to reach a critical value. The parameter threshold at which failure occurs. This refers to the remaining lifespan of the healthy service life. Once one of these parameters reaches a critical value, maintenance must be performed. It's important to note that the critical value is crucial for failure assessment. In most applications, an exact value is not chosen as the estimate of remaining lifespan; instead, a confidence interval or probability distribution for the predicted remaining lifespan is provided.

[0144] The effects of this invention are further illustrated by the following experiments:

[0145] 1. Experimental conditions:

[0146] This invention applies a prediction algorithm to a twin-shaft turbofan engine developed by Meskin et al. This model establishes a Simulink model by considering the nonlinear dynamic characteristics of both rotor dynamics and volumetric dynamics. By assuming an adiabatic process, the model calculates the pressure and temperature of the intake system in the engine's inlet duct. For the low-pressure compressor, the pressure ratio (π...) LCThe volumetric dynamics calculations between the high-pressure and low-pressure compressors are used. The rotor speed (N1) is obtained by performing energy balance calculations on the low-pressure shaft connecting the low-pressure compressor and the low-pressure turbine. Once the pressure ratio and speed are calculated, the corrected mass flow rate and efficiency can be obtained from the performance diagram, thus yielding the temperature rise. Similar to the low-pressure compressor, for the high-pressure compressor, the pressure is obtained through volumetric dynamics, and the speed (N2) is obtained by solving the energy balance problem on the high-pressure shaft connecting the high-pressure compressor and the high-pressure turbine. Finally, the pressure ratio of the high-pressure turbine is obtained through volumetric dynamics calculations between the high- and low-pressure turbines, and the pressure ratio of the low-pressure turbine is obtained through volumetric dynamics calculations.

[0147] 2. Experiment Content:

[0148] (1) Introduction to the Experimental Dataset: First, it is necessary to select the state, health parameters, and measurement parameters used for calculation and estimation. Select x = [T] CC ,N1,N2,P LT ,P CC ,P LC ,P HT [T] represents the state. CC The combustion chamber temperature is given by N1 and N2, which are the shaft speeds connecting the high-pressure turbine and the high-pressure compressor, and the low-pressure turbine and the low-pressure compressor, respectively. P is the combustion chamber temperature. LT P CC P LC and P HT These represent the pressures of the low-pressure turbine, combustion chamber, low-pressure compressor, and high-pressure turbine, respectively.

[0149] Choose y = [N1, N2, P] HC ,P LC ,T HC ,T LC ,T HT ,T LT As the measured value of the observation sensor, T HC T LC T HT and T LT These represent the outlet temperatures of the high-pressure compressor, low-pressure compressor, high-pressure turbine, and low-pressure turbine, respectively. Additionally, health parameters include the efficiency θ of the high-pressure compressor (HC), low-pressure compressor (LC), high-pressure turbine (HT), and low-pressure turbine (LT). η and traffic Health parameters

[0150] (2) Data processing:

[0151] In the high-pressure compressor fouling damage experiment, it is assumed that 100 data points (1 second each) of measurements are already known input to establish the initial model of DMO-RVM and obtain the initial estimated values ​​of the measurements. When new measurement data is collected, the Mahalanobis distance between the estimated value of the DMO-RVM prediction model and the measured value is used to determine whether the DMO-RVM model should be updated. Figure 9 As shown, the dynamic model DMO-RVM for the measured value N1 was updated three times, at 8.6 seconds, 11.4 seconds, and 13.4 seconds; the measured value T HT After two updates, at seconds 9.2 and 12. The DMO-RVM update times for other measurements are shown in the table below. The total number of updates is four. This total number of updates refers to the uniform update count for all measurements. In other words, DMO-RVM updates for different measurements at the same time are counted as one.

[0152] T in the table below LC and T HC No update is needed; the reasons can be found here. Figure 10 This can be seen from the measured value T. LC and T HC The initial predicted values ​​are highly accurate and closely approximate the actual measurements, therefore no updates are necessary. The final estimated values ​​obtained from the 8 sets of measurements through a dynamic update mechanism are as follows: Figure 10 As shown, the estimated values ​​all fall within the 95% confidence interval of the measured values. When the estimated value exceeds this interval, the DMO-RVM model is updated through a dynamic update mechanism to calculate a new estimated value.

[0153] Schedule of Measurement Updates for Scale Damage

[0154]

[0155]

[0156] (3) Analysis of experimental results:

[0157] First, the DMO-RVM and ARMA algorithms were used to model and predict the initial 100 data points (1 second), respectively. Figure 11 As can be seen from the results, the DMO-RVM algorithm demonstrates superior long-term predictive ability compared to the ARMA algorithm.

[0158] Secondly, to verify the impact of the dynamic measurement prediction method based on the DMO-RVM algorithm on remaining lifetime estimation, Figure 7This paper compares the remaining lifetime estimates obtained by the DMO-RVM algorithm and the ARMA algorithm under dynamic updates of measured values ​​under degradation caused by high-pressure compressor fouling. Overall, the remaining lifetime estimated using the hybrid prediction method is more accurate than the data-driven prediction method. Furthermore, the DMO-RVM algorithm demonstrates higher accuracy in long-term remaining lifetime estimation when dynamically estimating measured values. In other words, when process data is limited, the DMO-RVM-based hybrid prediction method is more accurate than the ARMA-based hybrid prediction method.

[0159] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining life of an aero-engine based on a dual-particle filter hybrid method, characterized in that: Includes the following steps: Step 1: Data preprocessing and model building: Select the state x, observed sensor measurements y, and health parameter θ for calculation and estimation. Perform MO-RVM modeling on each existing measurement data y to establish the initial n. y One MO-RVM prediction model; Step 2: Based on the prediction results of the prediction model, decide whether to perform dynamic model updates: Determine whether the MO-RVM model needs to be updated according to the state update rules. If the result indicates that an update is needed, then the DMO-RVM model will be further established, and then the sensor observations will be predicted. If dynamic updates are not needed, then prediction will be performed directly. Step 3: Predict health parameters: Use the state / health parameter dual-particle filter to predict the system state and hidden health parameters based on the sensor observations predicted in Step 2, and achieve k-step forward prediction of state and health parameters. Step 4, Remaining Lifetime Prediction: Using the predicted health parameter values ​​obtained in Step 3, plot the probability density distribution of the predicted health parameters. By plotting the probability density function of the predicted particles of the health parameters in the DPF, the moment with the highest probability is obtained, which is the estimated remaining lifetime.

2. The method for predicting the remaining life of an aero-engine based on a dual-particle filter hybrid method according to claim 1, characterized in that: In step 1, MO-RVM modeling is performed on each existing initial measurement data to obtain n. y The initial prediction models are as follows: According to the formula: The initial k-step forward prediction result is obtained through iteration: j = 1,...,n y P is the MO-RVM prediction model.

3. The method for predicting the remaining life of an aero-engine based on a dual-particle filter hybrid method according to claim 1, characterized in that: The state update rule in step 2 is determined by the difference between the predicted and actual measured values. This difference is quantified using Mahalanobis distance, which is defined as follows: In the formula, Σ is the covariance matrix of the multidimensional random variables, μ is the mean of the reference sample, and () T Let x be the transpose of the matrix, and x be the state value.

4. The method for predicting the remaining life of an aero-engine based on a dual-particle filter hybrid method according to claim 1, characterized in that: The judgment criterion in step 2 is as follows: the average Mahalanobis distance of the i-th measurement value within the monitoring window m is D. M (m,i) If any of the n average distances is greater than the threshold σ1 and is σ2 times greater than the average Mahalanobis distance of the previous window, it is determined that an update is needed. The determination rule is as follows: In the formula: σ1 is threshold 1, σ2 is threshold 2; Since it is not necessary to update the prediction model for all n measurements, a sub-rule is established under the above decision rule for secondary decision-making. The sub-rule is as follows: D M (m,i) <σ3 In the formula: σ3 is the second-order decision threshold, and σ3 < σ1. When the above sub-rule is satisfied, the decision does not need to update the i-th measurement value.

5. The method for predicting the remaining life of an aero-engine based on a dual-particle filter hybrid method according to claim 1, characterized in that: In step 3, the prediction process is as follows: For time steps r＝1,...,k： (1) Perform r-step forward prior state prediction: according to Predict the prior state of the particle, where The particle noise added to the state particle at time t+r has a distribution that conforms to... and These are the estimated values ​​of the state x and the health parameter θ at time t+r-1, respectively. according to Estimate the prior state covariance matrix; (2) Perform r-step forward and backward state prediction: Through resampling, calculation Obtain the weights of the filter for estimating the state x at time t+r. in, for The probability density function; Calculate the approximate posterior state distribution: Where Nreg is the number of regularizations, A t Depend on Calculated; Based on filter weights Resampling state to estimate particles, this is using The resulting posterior state estimate distribution after approximate resampling is: δ(·) is the Dirac function; according to Approximate posterior state estimation; (3) Perform the first r-step forward and backward parameter prediction: Calculate the corrected output prediction error h t (·) represents the nonlinear equation between state, health parameters, and measured values; Calculate correction parameters in, These are the adaptive coefficients during the calibration process. Let be the Jacobian determinant of the j-th particle; The kernel shrinkage algorithm is used to calculate the first step of posterior parameter prediction: A is the contraction matrix. The noise distribution added to each particle; (4) Perform the second r-step forward and backward parameter prediction: Calculate parameters to estimate filter weights Perform resampling. The first posterior parameter distribution is approximated using the following method. Particle prediction based on resampling parameters Resampling is performed, and the approximate posterior parameter prediction distribution is as follows: Posterior parameter prediction Calculate the posterior parameter prediction covariance as follows: Based on the above prediction process, the predicted measurement data is used, and resampling is maintained for future moments to generate the health parameter particle probability density function predicted k steps forward.

6. The method for predicting the remaining life of an aero-engine based on a dual-particle filter hybrid method according to claim 5, characterized in that: In step 4, for D-dimensional data, the formula for calculating the health parameter particle probability density function P(x|θ) is: Where μ is the mean of x, Σ is the covariance of x, and x is the state parameter of the engine.

7. The method for predicting the remaining life of an aero-engine based on a dual-particle filter hybrid method according to claim 1, characterized in that: In step 4, the remaining lifetime RUL(t) at time t is calculated using the following formula: RUL(t)=t j -t In the formula: t j The time it takes for any relevant parameter to reach a critical value. The parameter threshold at which failure occurs. Life expectancy is the healthy lifespan.

Images