A method for diagnosing a dynamic process sensor fault of an apu and control system

Abstract

The application discloses an APU and a method for diagnosing a dynamic process sensor fault of a control system, and the method comprises the following steps: using a small disturbance method to establish a wide-range linear dynamic model of the APU; on the basis, using a Kalman filter cluster to design dynamic process sensor fault diagnosis isolation and reconstruction logic; judging a dynamic process, and then using a residual weighted least square method to obtain a sensor fault period; in the dynamic process, disconnecting the Kalman filter to isolate the fault, bringing the saved flow coefficient which has not entered the dynamic into a measurement equation, and using the same to reconstruct the sensor fault, and calculating to obtain a correct APU fault sensor value. The application provides a solution for the fault diagnosis isolation and reconstruction problem of the APU system in the dynamic process. Compared with the existing research, the application can also effectively eliminate the estimation error and ensure the accuracy of the flow coefficient estimation in view of the deviation caused by the dynamic process and the fault to the flow coefficient estimation.

Description

Technical Field

[0001] This invention relates to the field of APU fault diagnosis, and more particularly to sensor fault diagnosis and reconstruction, as well as accurate estimation of flow coefficients. Background Technology

[0002] The Auxiliary Power Unit (APU), a small gas turbine engine mounted on an aircraft and capable of operating independently without external power, is also a member of the aero-engine system. Due to its long-term operation in harsh environments of high temperature and pressure, APU components and the sensors monitoring them are highly susceptible to failure. For reasons of flight safety and maintenance economy, sensor fault diagnosis is increasingly attracting attention.

[0003] Sensors are crucial components of the APU, and the accuracy of sensor measurements is essential for both flight control and diagnostic systems. There are generally two ways to improve sensor accuracy and reliability: hardware redundancy and analytical redundancy. Hardware redundancy uses additional redundant mechanisms to provide sensor redundancy, while analytical redundancy, in the event of a hardware redundancy failure, provides an estimated parameter based on the mathematical or network model of the object under study. This estimated parameter is then used as a redundancy value to replace the faulty sensor measurement.

[0004] While hardware redundancy offers high reliability and provides relatively accurate redundancy values, it doesn't offer significant advantages in reducing aircraft weight. Furthermore, redundancy mechanisms are susceptible to damage under high temperature and pressure environments. Therefore, in addition to hardware redundancy, analytical redundancy techniques must also be developed. Analytical redundancy methods can be categorized into model-based and data-based methods.

[0005] Data-driven methods typically employ artificial neural networks and extreme learning machines to train a model of sensor fault types using existing normal and fault data. Model-based methods primarily use Kalman filters for estimation, with common methods including linear Kalman filters, extended Kalman filters, and unscented Kalman filters. These methods track sensor measurements through filters and then use residual weighted squares to diagnose sensor faults. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the deficiencies of the prior art by providing a method for diagnosing sensor faults in an APU during dynamic processes, and to avoid the impact of sensor faults and dynamic processes on flow coefficient estimation.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] Step A): Establish an APU state variable model using the small perturbation method, augmenting the flow coefficient into the state variable model. Based on this, construct augmented state variable models at multiple operating points. Use engine speed scheduling to establish a large-range linear dynamic model of the APU with outputs from five sensors, where the two control variables of the model are fuel flow W. f The system outputs five sensors: engine speed N, compressor outlet pressure P3, compressor outlet temperature T3, turbine outlet pressure P5, and turbine outlet temperature T5. The dynamic process continuously alters the system's load power input through changes in fuel flow and load power. A wide-range linear dynamic model can handle both steady-state and dynamic operation.

[0009] Step B) Using the established large-scale linear dynamic model, design a dynamic process sensor fault diagnosis and isolation logic based on a linear Kalman filter cluster. For the five sensors of the APU, the filter cluster contains five filters corresponding to each sensor. Each filter monitors a specific sensor and simultaneously estimates the flow coefficient in real time. When no fault is detected, the average flow coefficient of the last 20 steps is always stored in the buffer. The start time of the dynamic process is determined by the change in the input quantity. During the simulation, the residual weighted square method is used to determine the fault start time. The fault period and fault sensor information of the APU are diagnosed by the relationship between the fault indication signal output by the filter and the detection threshold. When the dynamic process or when a sensor fault is detected, the Kalman filter's tracking result of the measurement value may differ from the actual measurement value due to the influence of constantly changing input quantities or interference from the sensor fault. Therefore, the fault filter needs to be disconnected for fault isolation during the dynamic process and the fault duration. The multi-step flow coefficients residing in the buffer before the fault are introduced into the state variables of the measurement equation, and the sensor fault value is reconstructed using the measurement equation with the correct state variables.

[0010] Step C) evaluates the effectiveness of the dynamic process sensor fault diagnosis method for the APU and control system. The simulation involves controlling the engine speed at 100% using a PID controller, continuously varying the load power input, and then using the PID controller to change the fuel flow input, thus constructing a dynamic process. The designed simulation evaluation experiments include the following scenarios: sensor faults are categorized into bias faults and drift faults for verification; the flow coefficient is categorized into non-degradation and degradation cases for verification. Through these four scenarios, the effectiveness of the fault diagnosis and reconstruction system is verified during the period when the fault indication signal exceeds the threshold. The fault occurrence periods include both steady-state and dynamic processes, thus verifying the effectiveness of the dynamic process sensor fault diagnosis.

[0011] As part of the research on fault diagnosis of dynamic process sensors in APU and control systems, a specific step (step B) for determining the fault period of a dynamic process is proposed as follows:

[0012] Step B1) involves a linearization step to obtain a linear dynamic model consisting of a coefficient matrix and a steady-state base point interpolation table. Interpolation using load power and rotational speed yields the coefficient matrix and steady-state base point for the current state. Since the APU has only one axis, it's crucial to verify the rotational speed for fault detection; otherwise, incorrect speeds may be used for interpolation. If there was no fault in the previous step, the speed sensor measurement is used for interpolation; if there was a fault, the reconstructed speed from the previous step must be used.

[0013] Step B2), in the constructed linear dynamic model, the control variable of the linear dynamic model is the fuel flow rate W. f The model input is changed by varying the load power (PW). Different load powers correspond to different flow coefficients. A PID algorithm is used to control the engine speed at 100%. As the load power changes, the fuel flow is continuously adjusted by the PID controller, creating a dynamic process to simulate the APU's operation. A linear Kalman filter is used to estimate the flow coefficient when the APU is in a steady-state, fault-free condition.

[0014] The state determination signal for determining the occurrence of a dynamic process can be expressed by the following formula:

[0015] |PW1-PW last |>0.1

[0016] |PW2-PW last |>0.1

[0017] |PW3-PW last |>0.1

[0018] |W f1 -W flast |>0.01

[0019] |W f2 -W flast |>0.01

[0020] |W f3 -W flast |>0.01

[0021] In the formula, PW lastThe load power is measured in kW, and the absolute value of the difference between the previous and next time steps is greater than 0.1. This represents the load power at the time the load power remained unchanged. PW1 is the value at the first time the load power changed, PW2 is the value at the second time the load power changed, and PW3 is the value at the third time the load power changed. This difference relationship must be satisfied for three sampling periods after the absolute value of the difference between the previous and next time steps is greater than 0.1 for the system to be considered to have entered a dynamic process based on the load power change. This is to avoid the influence of occasional interference on the judgment. The same principle applies when the fuel flow rate is measured in kg / s. When both the load power and fuel flow rate satisfy the above formula, the system is considered to have entered a dynamic process, and the state judgment signal is true. Through extensive simulation experiments, when a change in load power causes the system to enter a dynamic process, the dynamic process will last no more than 20 seconds, after which the system returns to a steady state.

[0022] To adapt to changes in flow coefficient, the flow coefficient is stored separately for different load powers. When the APU is in a steady state and no fault is detected, the flow coefficient is calculated in real time using a linear Kalman filter algorithm and updated accordingly. Changes in the filter input caused by dynamic processes, and deviations in the filter's tracking of measured values ​​due to sensor malfunctions, can both affect the flow coefficient estimation, leading to incorrect estimates. Based on the state determination signal, when the APU is in a dynamic process or a sensor malfunction is detected, the Kalman filter is disconnected, and the flow coefficient is no longer updated. Instead, the average of the flow coefficient values ​​from the last 20 steps is automatically calculated and used as the correct flow coefficient in the extended measurement equation, as shown in the following formula:

[0023] Δy=CΔx real +DΔu+υ

[0024] Δx real =(Δn,ΔCW) real ,ΔTW real )

[0025] Δx in the formula real It is the state vector ΔCW after substituting the average flow coefficient under fault-free conditions. real It is the compressor flow coefficient when there is no malfunction, ΔTW real This is the turbine flow coefficient when there is no fault. By using the state variables when there is no fault to participate in the reconstruction, the sensor faults in the dynamic process can be reconstructed from the calculation results of a large-scale linear dynamic model to obtain the correct measurement value, while also avoiding the impact of the dynamic process or sensor faults on the flow coefficient estimation.

[0026] Once the fault diagnosis, isolation, and reconfiguration system deems the fault resolved, it reconnects the linear Kalman filter and continues updating the flow coefficients. It maps the previously saved average flow coefficient to the load power. The next time the system operates at the same load power, it directly calls this average value. This method ensures that even when a fault occurs during a dynamic process, the correct flow coefficient can be retrieved, preventing errors in the reconfigured value caused by the inability to obtain an accurate average flow coefficient during a dynamic process.

[0027] Step B3), the fault isolation mechanism relies on the fact that the sensor has non-zero measurement noise, so the sensor's measured value and the Kalman filter's estimated value will also be inconsistent. This is defined through the state variable equation:

[0028]

[0029]

[0030] WSSR = e iT (∑ i ) -1 e i

[0031] Where the superscript i represents the i-th filter, Δy i k The measured value Δy k Remove the subset after the i-th row. For Δy i k The estimated value of C; i D i Remove the remaining portion after the i-th row from matrices C and D. Then e i Defined as the filtering residual of the i-th filter.

[0032] These filter residuals contain information about sensor malfunctions, so the filter residuals will change when a sensor malfunction occurs. The filter residuals are then weighted by a squared power e. iT (∑ i ) -1 e i The filtered residual is processed, and this formula is named the Fault Indication Signal WSSR. Where, ∑ i =diag[σ i ] 2 Vector σ i Let be the standard deviation of the i-th sensor subset.

[0033] When no fault occurs, the sensor measurement and the filter estimate are very close, so the WSSR signal is also small. When a sensor malfunctions, it takes some time for the filter estimate to catch up with the fault measurement, during which time the WSSR signal increases rapidly. By setting an appropriate threshold, the occurrence of a fault can be detected.

[0034] When the filter residual contains fault information from all five sensors, the residual will increase if any one sensor fails, rendering it meaningless for detecting the location of the faulty sensor. To address this, five filters are used to monitor the fault status of each sensor. The input signal to each filter is a subset of the measurements from the remaining four sensors after removing the sensor it is monitoring. For example, the input signal to the i-th filter is the subset of measurements from the remaining four sensors after removing the i-th sensor. If the i-th sensor fails, all four input measurements used by the i-th filter are fault-free, resulting in a smaller WSSR signal. Conversely, if other sensors fail, the four input measurements used by the i-th filter will include fault information, leading to a larger WSSR signal.

[0035] Taking the failure of the first sensor as an example, the WSSR signal at this time should satisfy the following formula:

[0036] WSSR1 < a

[0037] WSSR2>a

[0038] WSSR3>a

[0039] WSSR4>a

[0040] WSSR5>a

[0041] In the formula, 'a' represents the set detection threshold, and WSSR1 to WSSR5 are the fault indication signals output by the filters in the filter cluster that have removed the input values ​​of the first to fifth sensors. Taking the first sensor as an example, only the first filter does not use the faulty input value. When only the WSSR1 signal is below the threshold while the other WSSR signals exceed the threshold, the first sensor is determined to be faulty. The same logic applies to the faults of other sensors. The determination of the start time of sensor fault also adopts a three-step determination method, that is, the fault indication signal must meet the above formula within three sampling steps starting from the beginning of the formula before a sensor fault is considered to have occurred.

[0042] The accuracy of large-scale linear dynamic models decreases during dynamic processes. In the initial period of the dynamic process, the calculation results of the large-scale linear dynamic model have a certain error compared to the actual measured values, causing the fault indication signal to briefly and significantly exceed the normal value at the beginning of the dynamic process. This fault indication signal exceeding the threshold is not caused by sensor failure, but by the inherent accuracy problem of the linear dynamic model itself during the dynamic process. The solution is to use multi-level thresholds. The application of multi-level thresholds is as a fault judgment standard during the fault duration; that is, after the fault is detected, multi-level thresholds are used to determine the duration of the fault. Taking the first sensor failure as an example, the formula is as follows:

[0043] WSSR1<b

[0044] WSSR2>a

[0045] WSSR3>a

[0046] WSSR4>a

[0047] WSSR5>a

[0048] In the formula, 'b' represents the secondary threshold. Using the primary threshold 'a' would cause WSSR1 to exceed threshold 'a' at the beginning of the dynamic process, leading to missed diagnoses. Therefore, when designing the detection threshold, 'b' should generally be set larger than 'a', exceeding the impact of the linear dynamic model's accuracy error on the fault indication signal. Using a multi-level threshold method provides some tolerance for the accuracy of the linear dynamic model in dynamic process fault diagnosis. Even if the linear dynamic model has some accuracy error during the dynamic process, it will not affect the fault diagnosis effect. This method can accurately detect faulty sensors and reconstruct the correct measurement values.

[0049] As part of the simulation analysis of fault diagnosis methods for dynamic process sensors in APU and control systems, the specific steps of step C) are as follows:

[0050] In step C1), the simulation assumes that five sensors experience bias faults sequentially. Each sensor fault occurs during both steady-state and dynamic processes, without degradation of the flow coefficient. The simulation output shows the reconstructed image of sensor P3 and the estimated flow coefficient.

[0051] Step C2) In the simulation, five sensors are set to experience drift faults sequentially, with the drift fault amount increasing at a fixed rate. Each sensor fault occurs during both steady-state and dynamic processes, and the flow coefficient does not degrade. The simulation provides a reconstruction diagram of sensor P3 and an estimated flow coefficient diagram.

[0052] In step C3), the simulation assumes that five sensors will sequentially experience bias faults. Each sensor fault occurs during both steady-state and dynamic processes, and the turbine flow coefficient in the flow coefficient degrades by 0.01. The simulation results show the reconstructed image of sensor P3 and the estimated flow coefficient.

[0053] Step C4): During the simulation, five sensors are set to experience drift failures sequentially, with the drift failure amount increasing at a fixed rate. Each sensor failure occurs during both steady-state and dynamic processes. The turbine flow coefficient in the flow coefficient is degraded by 0.01. The simulation results show the reconstructed image of sensor P3 and the estimated flow coefficient.

[0054] Furthermore, in step B2), when the sensor malfunctions or the entire system is in a dynamic process, the filter is disconnected to avoid the impact of sensor malfunction and dynamic process on the flow coefficient estimation. When the sensor is normal, the flow coefficient is saved. After the filter is disconnected, the normal flow coefficient is brought into the measurement equation to participate in the reconstruction of the value of the faulty sensor.

[0055] Furthermore, the Kalman filter cluster in step B3) is a distributed structure, with five filters computing in parallel. When a sensor fault occurs during a dynamic process, because the filters are disconnected at this time, the erroneous measurement value will not be tracked. Instead, the previously estimated flow coefficient will be substituted into the measurement equation of the state variable model, and the measurement equation will be used for reconstruction. In this way, even during a dynamic process, faulty sensors can be diagnosed and reconstructed.

[0056] In step C), simulation analysis was performed on whether the sensor experienced bias or drift faults, whether the fault occurred during a steady-state or dynamic period, and whether the flow coefficient degraded. This demonstrates that the dynamic process fault diagnosis method is universal and can diagnose different fault types and APU system performance degradation.

[0057] Compared with existing solutions, the present invention, employing the above technical solution, has the following technical advantages:

[0058] (1) This invention proposes a solution to the problem of fault diagnosis of APU and control system in dynamic process, and fills the gap in the use of Kalman filter method for fault diagnosis of APU system in dynamic process.

[0059] (2) By disconnecting the Kalman filter and introducing the flow coefficient estimated at normal time, the present invention can keep the flow coefficient estimation normal throughout the simulation process and will not be affected by sensor failure and dynamic process periods. Based on this, the measurement equation can be used to reconstruct the faulty sensor value in the dynamic process.

[0060] (3) This invention tested the diagnostic effect under two fault types, namely bias fault and drift fault, and under two states, namely whether the flow coefficient degraded. It can diagnose and reconstruct different fault types and different APU performance states. Attached Figure Description

[0061] Figure 1 This is a structural diagram of the APU and control system dynamic process sensor fault diagnosis system;

[0062] Figure 2 This is a block diagram illustrating the fault diagnosis principle of dynamic process sensors in the APU and control system.

[0063] Figure 3 This describes how the load power changes over time.

[0064] Figure 4 The diagram shows the reconstructed effect of the bias fault sensor without degradation. (a)-(e) represent N, P3, T3, P5, and T5, respectively.

[0065] Figure 5 This is a graph showing the variation of the bias fault flow coefficient without degradation;

[0066] Figure 6 This is a diagram showing the reconstruction effect of a drift fault sensor without degradation.

[0067] Figure 7 This is a graph showing the change in the flow coefficient during drift faults without degradation.

[0068] Figure 8 This is a diagram showing the effect of reconstructing a bias fault sensor when TW degrades;

[0069] Figure 9 This is a graph showing the change in the bias fault flow coefficient during TW degradation;

[0070] Figure 10 This is a diagram showing the effect of reconstructing a drift fault sensor during TW degradation;

[0071] Figure 11 This is a graph showing the change in the flow coefficient during TW degradation due to drift faults. Detailed Implementation

[0072] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0073] The present invention uses a small perturbation method to establish a state variable model, and then constructs an augmented state variable model including flow coefficients. After constructing the augmented state variable model at multiple operating points, a large-scale linear dynamic model is obtained. Based on the large-scale linear dynamic model, a linear Kalman filter method is used. Each of the five sensors of the APU has a filter to monitor its fault status. The residual between the filter output value and the actual measurement value is processed by weighted square of the residual to obtain a fault indication signal. The effect is verified by repeatedly trying different thresholds. A fault can be determined when the fault indication signal of the monitoring sensor exceeds the threshold. At this time, the filter is disconnected, and the measurement equation is used for reconstruction.

[0074] The specific implementation of this invention takes a certain type of APU model as the research object. This invention describes a fault diagnosis method for dynamic process sensors in an APU and control system, which specifically includes the following steps:

[0075] Step A): Establish an APU state variable model using the small perturbation method, augmenting the flow coefficient into the state variable model. Based on this, construct augmented state variable models at multiple operating points. Use engine speed scheduling to establish a large-range linear dynamic model of the APU with outputs from five sensors, where the two control variables of the model are fuel flow W. f The system outputs five sensors: engine speed N, compressor outlet pressure P3, compressor outlet temperature T3, turbine outlet pressure P5, and turbine outlet temperature T5. The dynamic process continuously alters the system's load power input through changes in fuel flow and load power. A wide-range linear dynamic model can handle both steady-state and dynamic operation.

[0076] Step B) Using the established large-scale linear dynamic model, design a dynamic process sensor fault diagnosis and isolation logic based on a linear Kalman filter cluster. For the five sensors of the APU, the filter cluster contains five filters corresponding to each sensor. Each filter monitors a specific sensor and simultaneously estimates the flow coefficient in real time. When no fault is detected, the average flow coefficient of the last 20 steps is always stored in the buffer. The start time of the dynamic process is determined by the change in the input quantity. During the simulation, the residual weighted square method is used to determine the fault start time. The fault period and fault sensor information of the APU are diagnosed by the relationship between the fault indication signal output by the filter and the detection threshold. When the dynamic process or when a sensor fault is detected, the Kalman filter's tracking result of the measurement value may differ from the actual measurement value due to the influence of constantly changing input quantities or interference from the sensor fault. Therefore, the fault filter needs to be disconnected for fault isolation during the dynamic process and the fault duration. The multi-step flow coefficients residing in the buffer before the fault are introduced into the state variables of the measurement equation, and the sensor fault value is reconstructed using the measurement equation with the correct state variables.

[0077] Step B1) involves a linearization step to obtain a linear dynamic model consisting of a coefficient matrix and a steady-state base point interpolation table. Interpolation using load power and rotational speed yields the coefficient matrix and steady-state base point for the current state. Since the APU has only one axis, it's crucial to verify the rotational speed for fault detection; otherwise, incorrect speeds may be used for interpolation. If there was no fault in the previous step, the speed sensor measurement is used for interpolation; if there was a fault, the reconstructed speed from the previous step must be used.

[0078] Step B2), in the constructed linear dynamic model, the control variable of the linear dynamic model is the fuel flow rate W. f The model input is changed by varying the load power (PW). Different load powers correspond to different flow coefficients. A PID algorithm is used to control the engine speed at 100%. As the load power changes, the fuel flow is continuously adjusted by the PID controller, creating a dynamic process to simulate the APU's operation. A linear Kalman filter is used to estimate the flow coefficient when the APU is in a steady-state, fault-free condition.

[0079] The state determination signal for determining the occurrence of a dynamic process can be expressed by the following formula:

[0080] |PW1-PW last |>0.1

[0081] |PW2-PW last |>0.1

[0082] |PW3-PWlast |>0.1

[0083] |W f1 -W flast |>0.01

[0084] |W f2 -W flast |>0.01

[0085] |W f3 -W flast |>0.01

[0086] In the formula, PW last The load power is measured in kW, and the absolute value of the difference between the previous and next time steps is greater than 0.1. This represents the load power at the time the load power remained unchanged. PW1 is the value at the first time the load power changed, PW2 is the value at the second time the load power changed, and PW3 is the value at the third time the load power changed. This difference relationship must be satisfied for three sampling periods after the absolute value of the difference between the previous and next time steps is greater than 0.1 for the system to be considered to have entered a dynamic process based on the load power change. This is to avoid the influence of occasional interference on the judgment. The same principle applies when the fuel flow rate is measured in kg / s. When both the load power and fuel flow rate satisfy the above formula, the system is considered to have entered a dynamic process, and the state judgment signal is true. Through extensive simulation experiments, when a change in load power causes the system to enter a dynamic process, the dynamic process will last no more than 20 seconds, after which the system returns to a steady state.

[0087] To adapt to changes in flow coefficient, the flow coefficient is stored separately for different load powers. When the APU is in a steady state and no fault is detected, the flow coefficient is calculated in real time using a linear Kalman filter algorithm and updated accordingly. Changes in the filter input caused by dynamic processes, and deviations in the filter's tracking of measured values ​​due to sensor malfunctions, can both affect the flow coefficient estimation, leading to incorrect estimates. Based on the state determination signal, when the APU is in a dynamic process or a sensor malfunction is detected, the Kalman filter is disconnected, and the flow coefficient is no longer updated. Instead, the average of the flow coefficient values ​​from the last 20 steps is automatically calculated and used as the correct flow coefficient in the extended measurement equation, as shown in the following formula:

[0088] Δy=CΔx real +DΔu+υ

[0089] -x real =(Δn,-CW) real ,-TWreal )

[0090] Δx in the formula real It is the state vector ΔCW after substituting the average flow coefficient under fault-free conditions. real It is the compressor flow coefficient when there is no malfunction, ΔTW real This is the turbine flow coefficient when there is no fault. By using the state variables when there is no fault to participate in the reconstruction, the sensor faults in the dynamic process can be reconstructed from the calculation results of a large-scale linear dynamic model to obtain the correct measurement value, while also avoiding the impact of the dynamic process or sensor faults on the flow coefficient estimation.

[0091] Once the fault diagnosis, isolation, and reconfiguration system deems the fault resolved, it reconnects the linear Kalman filter and continues updating the flow coefficients. It maps the previously saved average flow coefficient to the load power. The next time the system operates at the same load power, it directly calls this average value. This method ensures that even when a fault occurs during a dynamic process, the correct flow coefficient can be retrieved, preventing errors in the reconfigured value caused by the inability to obtain an accurate average flow coefficient during a dynamic process.

[0092] In step B2), when the sensor malfunctions or the entire system is in a dynamic process, the filter is disconnected to avoid the impact of sensor malfunction and dynamic process on the flow coefficient estimation. When the sensor is normal, the flow coefficient is saved. After the filter is disconnected, the normal flow coefficient is substituted into the measurement equation to participate in the reconstruction of the value of the faulty sensor.

[0093] Step B3), the fault isolation mechanism relies on the fact that the sensor has non-zero measurement noise, so the sensor's measured value and the Kalman filter's estimated value will also be inconsistent. This is defined through the state variable equation:

[0094]

[0095]

[0096] WSSR = e iT (∑ i ) -1 e i

[0097] Where the superscript i represents the i-th filter, Δy i k The measured value Δy k Remove the subset after the i-th row. For Δy i k The estimated value of C; i D i Remove the remaining portion after the i-th row from matrices C and D. Then ei Defined as the filtering residual of the i-th filter.

[0098] These filter residuals contain information about sensor malfunctions, so the filter residuals will change when a sensor malfunction occurs. The filter residuals are then weighted by a squared power e. iT (∑ i ) -1 e i The filtered residual is processed, and this formula is named the Fault Indication Signal WSSR. Where, ∑ i =diag[σ i ] 2 Vector σ i Let be the standard deviation of the i-th sensor subset.

[0099] When no fault occurs, the sensor measurement and the filter estimate are very close, so the WSSR signal is also small. When a sensor malfunctions, it takes some time for the filter estimate to catch up with the fault measurement, during which time the WSSR signal increases rapidly. By setting an appropriate threshold, the occurrence of a fault can be detected.

[0100] When the filter residual contains fault information from all five sensors, the residual will increase if any one sensor fails, rendering it meaningless for detecting the location of the faulty sensor. To address this, five filters are used to monitor the fault status of each sensor. The input signal to each filter is a subset of the measurements from the remaining four sensors after removing the sensor it is monitoring. For example, the input signal to the i-th filter is the subset of measurements from the remaining four sensors after removing the i-th sensor. If the i-th sensor fails, all four input measurements used by the i-th filter are fault-free, resulting in a smaller WSSR signal. Conversely, if other sensors fail, the four input measurements used by the i-th filter will include fault information, leading to a larger WSSR signal.

[0101] Taking the failure of the first sensor as an example, the WSSR signal at this time should satisfy the following formula:

[0102] WSSR1 < a

[0103] WSSR2>a

[0104] WSSR3>a

[0105] WSSR4>a

[0106] WSSR5>a

[0107] In the formula, 'a' represents the set detection threshold, and WSSR1 to WSSR5 are the fault indication signals output by the filters in the filter cluster that have removed the input values ​​of the first to fifth sensors. Taking the first sensor as an example, only the first filter does not use the faulty input value. When only the WSSR1 signal is below the threshold while the other WSSR signals exceed the threshold, the first sensor is determined to be faulty. The same logic applies to the faults of other sensors. The determination of the start time of sensor fault also adopts a three-step determination method, that is, the fault indication signal must meet the above formula within three sampling steps starting from the beginning of the formula before a sensor fault is considered to have occurred.

[0108] The accuracy of large-scale linear dynamic models decreases during dynamic processes. In the initial period of the dynamic process, the calculation results of the large-scale linear dynamic model have a certain error compared to the actual measured values, causing the fault indication signal to briefly and significantly exceed the normal value at the beginning of the dynamic process. This fault indication signal exceeding the threshold is not caused by sensor failure, but by the inherent accuracy problem of the linear dynamic model itself during the dynamic process. The solution is to use multi-level thresholds. The application of multi-level thresholds is as a fault judgment standard during the fault duration; that is, after the fault is detected, multi-level thresholds are used to determine the duration of the fault. Taking the first sensor failure as an example, the formula is as follows:

[0109] WSSR1<b

[0110] WSSR2>a

[0111] WSSR3>a

[0112] WSSR4>a

[0113] WSSR5>a

[0114] In the formula, 'b' represents the secondary threshold. Using the primary threshold 'a' would cause WSSR1 to exceed threshold 'a' at the beginning of the dynamic process, leading to missed diagnoses. Therefore, when designing the detection threshold, 'b' should generally be set larger than 'a', exceeding the impact of the linear dynamic model's accuracy error on the fault indication signal. Using a multi-level threshold method provides some tolerance for the accuracy of the linear dynamic model in dynamic process fault diagnosis. Even if the linear dynamic model has some accuracy error during the dynamic process, it will not affect the fault diagnosis effect. This method can accurately detect faulty sensors and reconstruct the correct measurement values.

[0115] The Kalman filter cluster in step B3) is a distributed structure, with five filters computing in parallel. When a sensor fault occurs during a dynamic process, because the filters are disconnected at this time, the erroneous measurement value will not be tracked. Instead, the previously estimated flow coefficient will be substituted into the measurement equation of the state variable model, and the measurement equation will be used for reconstruction. In this way, even during a dynamic process, faulty sensors can be diagnosed and reconstructed.

[0116] Step C) evaluates the effectiveness of the dynamic process sensor fault diagnosis method for the APU and control system. The simulation involves controlling the engine speed at 100% using a PID controller, continuously varying the load power input, and then using the PID controller to change the fuel flow input, thus constructing a dynamic process. The designed simulation evaluation experiments include the following scenarios: sensor faults are categorized into bias faults and drift faults for verification; the flow coefficient is categorized into non-degradation and degradation cases for verification. Through these four scenarios, the effectiveness of the fault diagnosis and reconstruction system is verified during the period when the fault indication signal exceeds the threshold. The fault occurrence periods include both steady-state and dynamic processes, thus verifying the effectiveness of the dynamic process sensor fault diagnosis.

[0117] In step C1), the simulation assumes that five sensors experience bias faults sequentially. Each sensor fault occurs during both steady-state and dynamic processes, without degradation of the flow coefficient. The simulation output shows the reconstructed image of sensor P3 and the estimated flow coefficient.

[0118] Step C2) In the simulation, five sensors are set to experience drift faults sequentially, with the drift fault amount increasing at a fixed rate. Each sensor fault occurs during both steady-state and dynamic processes, and the flow coefficient does not degrade. The simulation provides a reconstruction diagram of sensor P3 and an estimated flow coefficient diagram.

[0119] In step C3), the simulation assumes that five sensors will sequentially experience bias faults. Each sensor fault occurs during both steady-state and dynamic processes, and the turbine flow coefficient in the flow coefficient degrades by 0.01. The simulation results show the reconstructed image of sensor P3 and the estimated flow coefficient.

[0120] Step C4): During the simulation, five sensors are set to experience drift failures sequentially, with the drift failure amount increasing at a fixed rate. Each sensor failure occurs during both steady-state and dynamic processes. The turbine flow coefficient in the flow coefficient is degraded by 0.01. The simulation results show the reconstructed image of sensor P3 and the estimated flow coefficient.

[0121] In step C), simulation analysis was performed to determine whether the sensor experienced bias or drift faults, whether the fault occurred during a steady-state or dynamic period, and whether the flow coefficient degraded. This demonstrates that the dynamic process fault diagnosis method is universal and can diagnose different fault types and APU system performance degradation.

[0122] To verify the effectiveness of the fault diagnosis method for dynamic process sensors in the APU and control system designed in this invention, a ground-point fault diagnosis simulation experiment was conducted in the Simulink software environment. The system structure diagram of the fault diagnosis method for dynamic process sensors in the APU and control system is shown below. Figure 1 As shown, the principle block diagram is as follows: Figure 2 As shown.

[0123] The simulation is set at ground point conditions, with a simulation time of 300 seconds. During the simulation time, the dynamic process of the simulation is created by continuously changing the load power PW in the input quantity. The changes in load power are as follows: Figure 3 As shown, this invention verifies the effectiveness of the fault diagnosis method for dynamic process sensors in APU and control systems under four different scenarios.

[0124] (1) Bias fault in the case of no degradation

[0125] Within a 300-second simulation period, to demonstrate the effectiveness of sensor fault diagnosis, five sensors were sequentially subjected to bias faults, each with an upward bias amplitude of 3%. The fault durations were as follows: speed sensor from 45 to 65 seconds, P3 sensor from 95 to 115 seconds, T3 sensor from 145 to 165 seconds, P5 sensor from 195 to 215 seconds, and T5 sensor from 245 to 265 seconds. The flow coefficient did not degrade, and the changes in input load power were consistent with... Figure 3 This consistency ensures that the 20-second period following each sensor failure includes both steady-state and dynamic processes. The reconstructed signal diagrams from the five sensors are shown below. Figure 4 As shown in the figure, the effect of flow coefficient estimation is as follows: Figure 5 As shown.

[0126] (2) Drift fault in the absence of degradation

[0127] Within a 300-second simulation period, to demonstrate the effectiveness of sensor fault diagnosis, five sensors were sequentially subjected to drift faults. The fault time for the speed sensor was from the 30th to the 60th second, for sensor P3 from the 90th to the 110th second, for sensor T3 from the 140th to the 160th second, for sensor P5 from the 190th to the 210th second, and for sensor T5 from the 240th to the 260th second. The flow coefficient did not degrade, and the changes in input load power were consistent with... Figure 3 Consistent. The fault expressions for the five sensors can be represented as:

[0128] y N =y N +0.0015*y N *(t-30)

[0129] y P3 =y P3 +0.0015*y P3 *(t-90)

[0130] y T3 =y T3 +0.0015*y T3 *(t-140)

[0131] y P5 =y P5 +0.0015*y P5 *(t-190)

[0132] y T5 =y T5 +0.0015*y T5 *(t-240)

[0133] The reconstruction effect of the signals from the 5 sensors is shown in the figure below. Figure 6 As shown in the figure, the effect of flow coefficient estimation is as follows: Figure 7 As shown.

[0134] (3) Bias fault under TW degradation

[0135] Within a 300-second simulation period, to demonstrate the effectiveness of sensor fault diagnosis, five sensors were sequentially subjected to bias faults, each with an upward bias amplitude of 3%. The fault time for each sensor was the same as the time set for the bias fault simulation without degradation. TW was degraded and increased by 0.01; the changes in input load power were similar to... Figure 3 Consistent. The reconstruction results of the five sensor signals are shown in the following figure. Figure 8 As shown in the figure, the effect of flow coefficient estimation is as follows: Figure 9 As shown.

[0136] (4) Drift fault under TW degradation

[0137] Within a 300-second simulation period, to demonstrate the effectiveness of sensor fault diagnosis, five sensors were sequentially subjected to drift faults. The timing and amplitude settings for each sensor's fault occurrence were consistent with the simulation settings for drift faults without degradation. When TW degraded, it increased by 0.01, and the changes in input load power were as follows... Figure 3 Consistent. The reconstruction results of the five sensor signals are shown in the following figure. Figure 10 As shown in the figure, the effect of flow coefficient estimation is as follows: Figure 11 As shown.

[0138] Simulation results show that for bias faults, the reconstructed value essentially replaces the measured fault value from the moment the fault occurs, and the reconstructed value matches the true value well in both steady-state and dynamic processes. For drift faults, a period of time is required to detect and reconstruct the fault after its occurrence, as the fault indication signal needs time to grow from its initial value to exceeding the threshold. After reconstruction, the reconstructed value matches the true value well. The linear Kalman filter automatically disconnects during the dynamic process of the simulation and when a sensor fault is detected, and the previously calculated average flow coefficient is used for calculation. This ensures that the flow coefficient is not incorrectly obtained due to the influence of dynamic processes or sensor faults during the simulation, and the correct flow coefficient can be estimated regardless of whether the flow coefficient degrades.

[0139] It should be noted that the above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations and substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for fault diagnosis of dynamic process sensors in an APU and control system, characterized in that, Includes the following steps: Step A) Using the small perturbation method and speed scheduling, a large-range linear dynamic model of the APU with outputs from several sensors is established; where the control variable of the model is the fuel flow rate W. f The load power is PW, the sensor output is the speed N, the compressor outlet pressure is P3, the compressor outlet temperature is T3, the turbine outlet pressure is P5, and the turbine outlet temperature is T5. The dynamic process continuously changes the load power input of the system through changes in fuel flow and load power. The large-range linear dynamic model can complete the operation between steady state and dynamic state. Step B) Design a fault diagnosis isolation logic for dynamic process sensors based on a linear Kalman filter cluster. The input of each filter in the filter cluster removes one sensor measurement value in turn, while retaining the other sensor measurements values. The start time of the dynamic process is determined by the change in the input value. During the simulation, the fault start time is determined by the residual weighted square method. The fault period of the APU and the fault sensor information are diagnosed by the relationship between the fault indication signal output by the filter and the detection threshold. When a dynamic process or when a sensor fault is detected to begin, the fault filter needs to be disconnected for fault isolation. The multi-step flow coefficients residing in the buffer designed before the fault are brought into the state variables of the measurement equation, and the sensor fault value is reconstructed using the measurement equation with the correct state variables. The specific steps of step B) are as follows: Step B1): The linear dynamic model obtains the coefficient matrix and steady-state base point in the current state by interpolating the load power and rotational speed. If there is no fault in the rotational speed in the previous step, the measured value of the rotational speed sensor is used for interpolation. If there is a fault in the rotational speed in the previous step, the reconstructed rotational speed in the previous step is used for interpolation. Step B2), in the constructed linear dynamic model, the control variable of the linear dynamic model is the fuel flow rate W. f The load power PW is used; the linear Kalman filter method is used to estimate the flow coefficient when the APU is in a steady state and there is no fault; when the current state of the APU is steady and no fault is detected, the flow coefficient value at the current moment is calculated in real time according to the linear Kalman filter algorithm, and the flow coefficient is updated. Based on the status judgment signal, when the APU is in a dynamic process or a sensor fault is detected, the Kalman filter is disconnected, and the flow coefficient is no longer updated. Instead, the average value of the flow coefficient values ​​of the most recent steps is automatically calculated and used as the correct flow coefficient in the extended measurement equation for calculation. When the fault diagnosis, isolation and reconstruction system determines that the fault has ended, the linear Kalman filter is reconnected and the flow coefficient is updated again. Step B3) Calculate the fault indication signal corresponding to each filter; during the fault period, if one fault indication signal is below the threshold and the rest of the fault indication signals are above the threshold, it is determined that a sensor has failed. The filter corresponding to the fault indication signal is missing the measurement value of the faulty sensor, and the faulty sensor is determined accordingly. The state determination signal for determining the occurrence of a dynamic process is expressed by the following formula: |PW1-PW last |>0.1 |PW2-PW last |>0.1 |PW3-PW last |>0.1 |In f1 -IN flast |>0.01 |In f2 -IN flast |>0.01 |In f3 -IN flast |>0.01 In the formula PW last When the load power is expressed in kW, the absolute value of the difference between the previous and next time steps is greater than 0.1, representing the load power at the time when it has not changed. PW1 is the value at the first time step when the load power changes, PW2 is the value at the second time step when the load power changes, and PW3 is the value at the third time step when the load power changes. This difference relationship must be satisfied for three sampling periods after the occurrence of the situation where the absolute value of the difference between the previous and next time steps is greater than 0.1 for the system to be considered to have entered a dynamic process based on the load power change. The same principle applies when the fuel flow rate is expressed in kg / s. When both the load power and fuel flow rate satisfy the above formula, the system is considered to have entered a dynamic process, and the state determination signal is true. To adapt to changes in flow coefficient, the flow coefficient for different load powers is stored separately. When the APU is in a steady state and no fault is detected, the flow coefficient value at the current moment is calculated in real time according to the linear Kalman filter algorithm, and the flow coefficient is updated accordingly. Based on the status determination signal, when the APU is in a dynamic process or detects a sensor malfunction, the Kalman filter is disconnected, and the flow coefficient is no longer updated. Instead, the average value of the flow coefficients from the most recent 20 steps is automatically calculated and used as the correct flow coefficient in the extended measurement equation for calculation, as expressed by the following formula: Δy=CΔx real +DΔu+υ Δx real =(Δn,ΔCW real ,ΔTW real ) Δx in the formula real It is the state vector ΔCW after substituting the average flow coefficient under fault-free conditions. real It is the compressor flow coefficient when there is no malfunction, ΔTW real It is the turbine flow coefficient when there is no malfunction; When the fault diagnosis, isolation and reconfiguration system determines that the fault has ended, it reconnects the linear Kalman filter and continues to update the flow coefficient; it then associates the previously saved average flow coefficient with the load power, so that the average value can be directly called when the system runs at the same load power next time. The Kalman filter cluster has a distributed structure, with all filters computed in parallel. When a sensor failure occurs during a dynamic process, the filter is disconnected and no longer tracks the erroneous measurement value. Instead, the previously estimated flow coefficient is substituted into the measurement equation of the state variable model, and the measurement equation is used for reconstruction.

Images