JP5845705B2 - Gas turbine performance estimation device - Google Patents

Description

  The present invention relates to a gas turbine performance estimation apparatus.

  Conventionally, when estimating the performance of each module of a gas turbine engine fan, compressor, turbine, etc., a measurement sensor such as a pressure sensor or a temperature sensor is added to the inlet / outlet of each element, and the sensor measurement value is acquired to analyze the performance. As a result, module performance such as efficiency and air flow rate was confirmed, and element deterioration and abnormality were judged.

  On the other hand, adding a large number of measurement sensors to each part of the engine for the purpose of ensuring highly accurate performance estimation increases the weight of the gas turbine engine. Therefore, conventionally, a large number of measurement sensors are added to each part of the engine only during an engine development experiment, but the mass production engine is not equipped. Therefore, the measurement sensors installed in the mass production engine are installed only for the purpose of collecting information necessary for control, and the number and types thereof are limited. For example, only sensors for obtaining the fan inlet temperature T2, the fan outlet temperature T28, the compressor inlet temperature T3, the compressor outlet pressure CDP, the turbine exhaust gas temperature T42, the low pressure shaft rotation speed N1, and the high pressure shaft rotation speed N2 are installed. It was.

  However, since the performance of each module cannot be confirmed with a limited measurement sensor, an expert has to judge based on past experience and the like by trend analysis or the like.

  As a technique for making this expert's judgment by a computer, for example, Japanese Patent Laid-Open No. 2000-213395 (Patent Document 1) uses a neural network to perform self-learning to optimize engine model parameters, and A system for controlling the air-fuel ratio using the neural network is disclosed. Japanese Patent Application Laid-Open No. 2001-174366 (Patent Document 2) calculates parameters by operating an engine model using operating conditions and sensor values of an actual engine as a system for monitoring the performance of a turbine engine. A system for diagnosing the operating state of an engine using a neural network is disclosed. Furthermore, JP-A-11-343916 (Patent Document 3) relates to a normal engine, but in engine control using data relating to the engine state as a control parameter, data relating to the engine state as a control parameter is disclosed. At least, a technique for estimating using a fuzzy neural network having a plurality of different data as input information is disclosed.

JP 2000-213395 A JP 2001-174366 A JP-A-11-343916

  The present invention improves a conventional gas turbine performance estimation device, learns from measurement data of a limited measurement sensor by a neural network, and estimates the performance characteristics of each module. An object of the present invention is to provide a gas turbine performance estimation device capable of accurately estimating the performance characteristics of a module and capable of detecting an abnormality or the like.

The present invention provides a limited number of a plurality of types and a plurality of control sensors installed at respective predetermined locations of a gas turbine engine, a temperature sensor, a pressure sensor and a rotational speed installed at the inlet of the gas turbine engine. A sensor, a fuel flow sensor installed at a fuel inlet of the gas turbine engine, and a measured value of the temperature sensor, a measured value of the pressure sensor, a measured value of the rotation speed sensor, and a measured value of the fuel flow sensor, The first engine dynamic model software that uses the nominal model is operated to estimate the measured values of the limited number of the plurality of types and the plurality of control sensors installed at the respective predetermined locations of the gas turbine. a first engine simulator for outputting an estimated value, the limited number of more, before and respective measurement values of a plurality control sensor of A difference calculator for obtaining a difference from each estimated value of the first engine simulator, the temperature measurement value, the pressure measurement value, the fuel flow rate measurement value and the rotation speed measurement value, and a plurality of types and a plurality of differences by the difference calculator A neural network that estimates a performance characteristic coefficient that is an internal state of the dynamic model of the gas turbine engine, and the temperature measurement value, the pressure measurement value, the fuel flow rate measurement value, and the rotation speed measurement value. It is difficult to sense the gas turbine engine by inputting a performance characteristic coefficient estimation value output from the neural network as a state variable group to construct a deterioration model and operating second engine dynamic model software using the deterioration model. second engine simulator that output the estimate of the predetermined portion each physical quantity comprising a portion Wherein the gas turbine performance estimation apparatus having a.

  According to the present invention, the first engine simulator that operates the first engine dynamic model software that uses the nominal model and the second engine simulator that operates the second engine dynamic model software that uses the deterioration model are arranged in cascade. The difference between the measured value of the control sensor of each part and the estimated value of the corresponding control sensor of the first engine simulator is input to the neural network, and the performance characteristic coefficient estimated and output by the neural network is the deterioration of the second engine simulator. By using the state variable group of the model, the feedback system, which tends to be unstable in estimation calculation, is separated, learning from a limited number of sensor measurement data with a neural network, and estimating the performance characteristic coefficient of each module Due to the small number of controls The performance characteristics of each module can be stably and accurately estimated by the sensor.

The functional block diagram of the gas turbine performance estimation apparatus of the 1st Embodiment of this invention. The functional block diagram of the gas turbine engine electronic control apparatus containing the gas turbine performance estimation apparatus of the said 1st Embodiment. Explanatory drawing which shows arrangement | positioning of the sensor for control which measures the measured value of each part of the gas turbine engine acquired with the gas turbine performance estimation apparatus of the said 1st Embodiment. Explanatory drawing of the neural network in the gas turbine performance estimation apparatus of the said embodiment. The graph group 1 which showed the pressure of each part of the gas turbine engine by the gas turbine performance estimation apparatus of the said 1st Embodiment, the measured value of temperature, an estimated value, and the value of the model at the time of a nominal. The graph group 2 which showed both the measured value of the pressure of each part of the gas turbine engine by the gas turbine performance estimation apparatus of the said 1st Embodiment, temperature, the estimated value, and the value of the model at the time of nominal. The graph group 1 which showed the difference of the measured value and estimated value of the pressure of each part of the gas turbine engine by the gas turbine performance estimation apparatus of the said 1st Embodiment, and an estimated value. The graph group 2 which showed the difference of the measured value and estimated value of the pressure of each part of the gas turbine engine by the gas turbine performance estimation apparatus of said 1st Embodiment, and an estimated value. The functional block diagram of the gas turbine performance estimation apparatus of the 2nd Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[First Embodiment]
FIG. 1 shows a configuration of a gas turbine performance estimation device 1 according to a first embodiment of the present invention, and FIG. 2 shows a configuration of a gas turbine engine electronic control device 2 including the gas turbine performance estimation device 1. The gas turbine performance estimation apparatus 1 includes a first engine simulator 11, a difference calculator 12, filters 13 and 14, a neural network 15, and a second engine simulator 16. The first engine simulator 11 and the second engine simulator 16 are equipped with engine dynamic model software, and the operation of the actual gas turbine engine 3 is simulated by the operation, and in real time with respect to the input of the fuel flow rate and the like. In response dynamically, the estimated values y0 (i) and y_hat (i) of the pressure and temperature of each sensor installation part are calculated.

  As shown in FIG. 3, the gas turbine engine 3 which is a real machine includes an intake port 31, a fan 32, a high compressor (HPC) 33, a combustor 34, a high pressure turbine (HPT) 35, and a low pressure turbine (LPT) 36 as modules. , A mixer 37 and an exhaust port 38 are provided. The high compressor 33 and the high pressure turbine 35 are coupled by a high pressure shaft 39, and the high compressor 33 is rotated by the high pressure turbine 35 to highly compress the air. The fan 32 and the low-pressure turbine 36 are coupled by a low-pressure shaft 310, and the fan 32 is rotated by the rotation of the low-pressure turbine 36, and the air is compressed low and sent to the high compressor 33. A part of the air 30 taken in from the intake port 31 becomes a core flow 311, is compressed through the high compressor 33, is mixed with fuel in the combustor 34, is burned, and is sequentially sent to the high pressure turbine 35 and the low pressure turbine 36. These are rotated and fed into the mixer 37. The remainder of the air 30 taken in from the air supply port 31 becomes a bypass flow 312 and is sent to the mixer 37 through the bypass duct 40. In the mixer 37, the hot core stream 311 and the bypass stream 312 are mixed and then discharged from the exhaust port 38 as the exhaust stream 313.

  Various sensors are installed at predetermined locations of the gas turbine engine 3. Inlet inlet (0), fan inlet (2), bypass duct inlet (11), bypass duct outlet (15), high compressor inlet (28), combustor inlet (3), high pressure turbine inlet (4) The temperature sensors T0, T2, T11, T28, T3, T42 and the pressure sensors P2, P11, P28, P3, P41s, P42, P5s are installed at the low pressure turbine inlet (41) and the low pressure turbine outlet (42), respectively. ing. The low pressure shaft rotational speed sensor N1 senses the low pressure shaft rotational speed, and the high pressure shaft rotational speed sensor N2 senses the high pressure shaft rotational speed. Here, the subscript x of Tx and Px corresponds to the numbers of the respective parts in the gas turbine engine 3. These Tx and Px will be described assuming that they also represent the measured values of the sensors at the corresponding locations. The same applies to the rotational speeds N1 and N2.

In the gas turbine performance estimation device 1 of the present embodiment, the first engine simulator 11, the filters 13, 14 and the second engine simulator 16 are each supplied with the fuel flow rate WF and the high compression shown in Table 1 as the operating point signal u (i). The machine inlet variable vane angle VSV and the engine inlet temperature T0 are input, and P0 is input as the engine inlet pressure. Table 1 is a table showing the types of data groups acquired as measurement data of the operating point of the gas turbine engine 3.

In the first engine simulator 11, all 0 is set as the performance characteristic coefficient c. Then, the first engine simulator 11 performs a simulation calculation of the operation of the gas turbine engine 3 in a nominal state (a state in which there is no deterioration or abnormality at the beginning of use), and estimated values of temperature, pressure, and rotation speed at each sensor measurement point. y0 (i) (i = 1 to 16) is output. Here, the performance characteristic coefficient is a multiple value with respect to the nominal state of the heat insulation efficiency and the flow rate indicating the performance of each module, and 0 is set in the first engine simulator 11.

The difference calculator 12 inputs the measured values Tx and Px of the temperature sensor and the pressure sensor installed in various places of the actual machine 3 shown in Table 2, and also inputs the low pressure shaft rotation speed N1 and the high pressure shaft rotation speed N2. Table 2 is a table showing the correspondence between the control sensor group of the gas turbine engine (actual machine) 3 and the measured values thereof. Specifically, the sensor measured values Tx, Px, N1, N2 and the sensor identification number y (i ).

  The difference calculator 12 subtracts the estimated value y0 (i) of the sensor measurement value calculated by the first engine simulator 11 from these sensor measurement values y (i) to obtain the difference Δy_hat (i), and applies it to the filter 13. Output.

Each of the filters 13 and 14 outputs an input value to the neural network 15 after a time delay of a constant time constant with respect to the input. Since the estimated value c_hat (i) of the performance characteristic coefficient is affected by noise included in the sensor measurement value y (i), it is slower than the engine time constant (about 1 second) so as not to interfere with the engine dynamics. It is preferable to set a time constant that is faster than the desired band according to the application (instrument, alarm, etc.), approximately 5 seconds.

  FIG. 4 shows the inside of the neural network 15. To the input layer of the neural network 15, outputs ΔN1, ΔN2,..., ΔP5s of the difference calculator 12 and WF, VSV, and T0 that are operating point inputs u (i) are input. In the neural network 15, the weights of the units in the intermediate layer with respect to the inputs of ΔN1, ΔN2,..., ΔP5s, WF, VSV, and T0 are set to be optimal by initial learning. That is, a sample input and output data set (teacher data) is given, and the combined load of the intermediate layer and the output layer is obtained by iterative calculation so as to reproduce the input / output relationship as much as possible.

The neural network 15 outputs a parameter C (i) (= c_hat (i)) to be estimated from the output layer. Table 3 is a correspondence table between the performance characteristic coefficient c calculated by the neural network 15 and the physical quantity of each part of the gas turbine engine.

As shown in Table 3, the parameter C (i) to be estimated by the neural network 15 is a performance characteristic coefficient, C1 is sw_fh: FAN hub side flow rate, C2 is se_fh: FAN hub side efficiency, and C3 is sw_ft: FAN chip. Side flow rate, and so on. The performance characteristic coefficient C (i) (i = 1 to 12) obtained by the neural network 15 is input to the second engine simulator 16.

The second engine simulator 16 has the same calculation method as the first engine simulator, and calculates and outputs an estimated value y_hat (i) of each sensor measurement value using the performance characteristic coefficient C (i) at the time of deterioration. In addition to the types shown in Table 2, the estimated value y_hat (i) includes physical quantities (thrust, air flow rate, etc.) at each location shown in FIG. 3 including the engine thrust estimated by the second engine simulator 16 which is normally difficult to sense. Pressure, temperature, enthalpy, momentum, flow rate, etc.).

  In the system configuration of the gas turbine engine electronic control unit 2 shown in FIG. 2, the rating calculation unit 21 calculates the rotation speed requests N1ref, N2ref and the like in response to the command of the throttle 4, and the difference unit 22 is the rotation speed sensor of the actual machine 3. The difference ΔN1, ΔN2 between the low-pressure shaft rotation speed N1 and the high-pressure shaft rotation speed N2 and the rotation speed requests N1ref, N2ref is obtained and output to the controller 23. Based on the rotational speed difference value, the controller 23 calculates the supply fuel flow rate WF (= u1) to the actual machine 3, and simultaneously calculates the HPC inlet variable stator blade angle VSV (= u2) and outputs it to the actual machine 3, thereby The rotation of the actual gas turbine engine 3 is controlled. In addition, other parameters (operation variables) that are input to the actual machine for controlling the engine include the flow rate of bleed air that discharges air compressed by the compressor from the main flow, and the clearance between the turbine blades and the case. Although there is a turbine case cooling flow rate for adjustment, WF and VSV are shown here.

  At the same time, in the gas turbine performance estimation apparatus 1, in the first engine simulator 11, the fuel flow rate WF, the HPC inlet variable stator blade angle VSV, the engine inlet temperature T0 from the sensor T0, and the engine inlet pressure P0 are operated from the controller 23. The data u1 to u4 are input, the operation of the gas turbine engine 3 in the nominal state is simulated, and the nominal estimated value y_0 (i) of the measured value of the sensor y (i) (i = 1 to 16) is output.

  Subsequent calculation operations in the gas turbine performance estimation device 1 are as described above, and the second engine simulator 16 uses the sensor y (i) (i = 1 to 1) in the deterioration model (deterioration state model of the gas turbine engine 3). The estimated value y_hat (i) of 16) is output.

  As described above, in the gas turbine performance estimation apparatus 1 according to the present embodiment, the measured value y (i) of the limited control sensor and the estimated value by the simulation calculation of the engine dynamic model in the nominal state in the first engine simulator 11. The performance characteristic of each module is estimated by the neural network 15 having the difference Δy_hat (i) from y_0 (i) as an input, and thereby the performance characteristic coefficient (nominal performance) which is the internal state of the engine dynamic model in the second engine simulator 16 And tune c_hat (i)). Then, an estimated value y_hat (i) corresponding to the measured value y (i) of each sensor is output by the simulation calculation in the deterioration model by the second engine simulator 16.

  As a result, according to the present embodiment, the engine dynamic model of the second engine simulator 16 is always subjected to simulation calculation while keeping the state identified online with the actual machine 3, and corresponds to the measured value y (i) of each sensor. The estimated value y_hat (i) to be obtained can be obtained. The estimated value y_hat (i) in the deterioration model of each sensor can be determined by the engineer to determine the state of each module of the gas turbine engine 3.

  In addition, in the present embodiment, the first engine simulator 11 and the second engine simulator 16 that execute the same engine dynamic model software are arranged in a cascade (referred to as a nominal model and a deterioration model), and the measured value y (i) By using the difference from the estimated value y0 (i) of the nominal model as an input to the neural network 15 and using the performance characteristic coefficient c_hat (i) as the output of the neural network 15 as a state variable of the degradation model, the feedback system is separated. Yes.

  Thereby, according to the present embodiment, as shown in the graph groups of FIGS. 5A, 5B, 6A, and 6B, the oscillation of the performance characteristic coefficient is suppressed, and the sensor measurement value y (i) and the estimated value y_hat are suppressed. (I) matches well, and the performance characteristics of each module can be estimated with high accuracy. 5A and 5B show the measured value y (i) (i = 1 to 16) of each sensor and the estimated value y_hat (i) obtained by the present embodiment, which are plotted simultaneously. 6A and 6B show the difference Δy_hat (i) (= y (i) −y_hat (i)) between the measured value y (i) of each sensor and the estimated value y_hat (i) obtained by the present embodiment. ing. In FIG. 5A and FIG. 5B, it is meant that the accuracy of the estimated value is higher when both graphs match. 6A and 6B, the closer the graph value is to 0, the higher the accuracy of the estimated value. And according to the calculation result of this Embodiment, it can be understood that the difference between the two is small and the estimation can be performed with high accuracy. For example, the variable y1 is the low-pressure shaft rotation speed N1, and the y2 is the high-pressure shaft rotation speed N2, but the difference between the measured value and the estimated value is within a range of approximately ± 1%. Also for the high-pressure turbine pressure y13 and the low-pressure turbine pressure y14, the difference between the measured value and the estimated value is within a range of approximately ± 0.5%.

[Second Embodiment]
A gas turbine performance estimation apparatus 1A according to the second embodiment will be described with reference to FIG. As shown in FIG. 3, the second embodiment is characterized by a predetermined number of control sensors Tx and Px installed at predetermined locations of the gas turbine engine 3 and an inlet of the gas turbine engine 3. Temperature sensor T0 and pressure sensor P0, fuel flow sensor WF installed at the fuel inlet or combustor inlet of gas turbine engine 3, measured value of temperature sensor T0, measured value of pressure sensor P0, measured value of fuel flow sensor WF An engine model simulator 51 that inputs each of them, operates the engine dynamic model software, estimates the measured values of each of a plurality of types and a plurality of control sensors installed at predetermined locations of the gas turbine 3, and outputs the estimated values. And the measured values y (i) of a plurality of types and a plurality of control sensors Tx, Px and the engine model simulator 5 The difference calculator 52 for obtaining the difference Δy_hat (i) from the estimated values y_hat (i) of the plurality of types and the plurality of control sensors, the temperature measurement value T0, the pressure measurement value P0, and the fuel flow rate measurement value F0 And a plurality of types and a plurality of difference values Δy_hat (i) by the difference calculator 52, and the performance characteristic coefficient C (i) of each of the plurality of modules of the gas turbine engine 3 is estimated, and the engine model simulator 51 And a neural network 15 that tunes a performance characteristic coefficient that is an internal state by feedback.

The neural network 15 has a function of creating a sensitivity matrix of Δc / Δy from a difference Δy between characteristic data of a nominal model (a model having new characteristics) and a degradation model by a sensitivity matrix calculation unit 151 and performing online learning. I have. That is, the performance characteristic coefficient parameter obtained by multiplying Δy by the sensitivity matrix Δc / Δy is used as teacher data, and a weight for minimizing the square sum of Δy is searched by the steepest descent method with momentum during the learning time. This neural network 15 calculates a performance model coefficient C (i) with respect to a model that has deteriorated due to the use of the gas turbine engine, that is, a degradation model, and feeds it back to the engine model simulator 51. Therefore, the performance characteristic coefficient which is the internal state of the engine model simulator 51 is sequentially updated so as to match the actual machine 3 deteriorated by online learning.

  According to the gas turbine performance estimation apparatus of the present embodiment, the difference Δy_hat (i) between the limited control sensor (measured value: y (i)) and the estimated value y_hat (i) of the engine dynamic model is input. By estimating the performance characteristics of each module using the neural network 15 and tuning the performance characteristic coefficient (difference from nominal performance: c_hat (i)), which is the internal state of the engine dynamic model, the engine dynamic model is always The state identified online with the real engine can be maintained, and accurate estimation can be performed.

1,1A Gas turbine performance estimation device 2 Electronic control device 3 Gas turbine engine (actual machine)
4 Throttle 11 First engine simulator 12 Difference calculator 13 Filter 14 Filter 15 Neural network 16 Second engine simulator 21 Rating calculator 22 Difference calculator 23 Controller 30 Air flow 31 Inlet 32 Fan 33 High compressor 34 Combustor 35 High pressure Turbine 36 Low-pressure turbine 37 Mixer 38 Nozzle 39 High-pressure rotating shaft 40 Bypass duct 310 Low-pressure rotating shaft 311 Core flow 312 Bypass flow 313 Exhaust flow 51 Engine model simulator 52 Difference calculator 151 Sensitivity matrix calculator Px Pressure sensor (pressure measurement value)
Tx temperature sensor (temperature measurement value)
Nx Rotational speed sensor (Rotational speed)

Claims (3)

A limited number of a plurality of types and a plurality of control sensors installed in each predetermined location of the gas turbine engine;
A temperature sensor, a pressure sensor and a rotational speed sensor installed at an inlet of the gas turbine engine;
A fuel flow sensor installed at a fuel inlet of the gas turbine engine;
The gas turbine is operated by inputting the measured value of the temperature sensor, the measured value of the pressure sensor, the measured value of the rotation speed sensor, and the measured value of the fuel flow rate sensor, and operating first engine dynamic model software that uses a nominal model. A first engine simulator installed at each of the predetermined locations, estimating the measured values of each of the limited number of the plurality of types and the plurality of control sensors, and outputting the estimated values;
A difference calculator for calculating a difference between each of the measurement values of the limited number of plural types and the plurality of control sensors and the estimated value of the first engine simulator;
The temperature measurement value, the pressure measurement value, the fuel flow rate measurement value, the rotation speed measurement value, and a plurality of types and a plurality of difference values by the difference calculator are input, and is an internal state of the dynamic model of the gas turbine engine. A neural network for estimating the performance characteristic coefficient;
The temperature measurement value, the pressure measurement value, the fuel flow rate measurement value, and the rotation speed measurement value are input, the performance characteristic coefficient estimation value output from the neural network is input as a state variable group, and a deterioration model is constructed. and characterized in that a second engine simulator sensing of the second of said gas turbine engine to engine dynamic model software is operated to output the estimated values of the predetermined portion each physical quantity including a difficult place to use Gas turbine performance estimation device.     The fuel flow measurement value, temperature measurement value, pressure measurement value and rotation speed measurement value at the inlet of the gas turbine engine and a plurality of types and a plurality of difference values by the difference calculator are accompanied by a time delay of a predetermined time constant. The gas turbine performance estimation apparatus according to claim 1, further comprising a filter that inputs to the neural network. The sensor data after changing the performance parameter of each element of the engine model and the sensor data when the performance parameter of each element is not changed are used, and the difference between these sensor data is used as teacher data for the neural network . The gas turbine performance estimation device according to claim 1 or 2, wherein a coupling load between the intermediate layer and the output layer is obtained .