Method and system for model calibration

Abstract

Methods and systems for calibrating a model are provided herein. In some embodiments, the method includes receiving test data for operating parameters of a real-world system, such as an engine, from an operating test via a control circuit. The control circuit also receives model data of the operating parameter from a simulation via a model of the engine. The control circuit then compresses the test data and the model data to generate compressed test data and compressed model data, and fuses the compressed test data with the compressed model data to generate fused data. The control circuit performs a parallel Bayesian reasoning simulation using the fused data to identify at least one value of a tuning parameter of the model. The control circuit may identify and select tuning parameters to match model data to test data, which may be one or more model outputs (i.e., output parameters).

Description

Technical Field

[0001] These teachings relate generally to models, and more particularly, to models such as models of engine or cycle performance. Background Art

[0002] Mathematical models (such as models of engine or cycle performance) use mathematics or physics to describe real-world systems. The model can be used as a mapping function from input parameters (x) to output parameters (y) and can be modulated by tuning parameters (θ). The tuning parameters (θ) can be real-valued scalars that are not directly predicted or estimated by the model but define the model architecture. In order to improve the prediction of the output parameters (y) from the model, the model can be calibrated to align the output parameters (y) with observations or test data from the real-world system. Such calibration may involve adjusting or tuning the tuning parameters (θ) in the model, which can take multiple values ​​to ensure that the model provides an accurate description of the behavior of the real-world system. Calibration can be achieved by comparing the predictions from the model with actual observations or measurements made on the real-world system. In order to align the predictions with the test data from the real-world system, the values ​​of one or more tuning parameters (θ) can be adjusted. BRIEF DESCRIPTION OF THE DRAWINGS

[0003] Various needs are met, at least in part, by providing methods and systems for model calibration as described in the following detailed description, particularly when studied in conjunction with the accompanying drawings. A full and enabling disclosure of the various aspects of the present description, including the best mode thereof, is set forth in the specification with reference to the accompanying drawings, wherein:

[0004] Figure 1 A system for calibrating a model including various embodiments according to these teachings;

[0005] Figure 2 A schematic diagram of a tuning process including a model according to various embodiments of these teachings;

[0006] Figure 3A and 3B A flow chart including a method of calibrating a model according to various embodiments of these teachings;

[0007] Figure 4 including schematic diagrams of data compression processes according to various embodiments of these teachings;

[0008] Figure 5 A graph including an exemplary parallelizable Markov Chain Monte Carlo (MCMC) process according to various embodiments of these teachings;

[0009] Figure 6Another graph including an exemplary parallelizable Markov Chain Monte Carlo (MCMC) process according to various embodiments of these teachings;

[0010] Figure 7 including exemplary probability distributions of tuning parameters for derived models according to various embodiments of these teachings;

[0011] Fig. 8A , Figure 8B and Figure 8C is a flow chart of a method for prognostic health monitoring and health management / engine health monitoring (PHM / EHM) of an engine, wherein the PHM / EHM model is tuned using the method described herein;

[0012] Fig. 9 is Figures 8A-8C A graph of exemplary flight data used in the method of FIG. 1 ; and

[0013] Fig.10 is a flow chart of an exemplary method for certifying an engine by performing an analysis using a calibration model that has been calibrated by the methods described herein.

[0014] The elements in the drawings are shown for simplicity and clarity and are not necessarily drawn to scale. For example, the size and / or relative position of some elements in the drawings may be exaggerated relative to other elements to help improve the understanding of the various embodiments of the present teaching. In addition, common but well-understood elements that are useful or necessary in commercially feasible embodiments are generally not described to facilitate less obstructed observation of these different embodiments of the present teaching. Certain actions and / or steps may be described or depicted in a specific order of occurrence, but those skilled in the art will understand that such specificity regarding sequence is not actually required. DETAILED DESCRIPTION

[0015] Model calibration is usually performed by adjusting the tuning of tuning parameters (θ), for example, to reduce errors in model predictions (e.g., with reference to field data and / or test data). In some methods, tuning is accomplished via trial and error. Trial-and-error-based tuning involves manually adjusting tuning parameters (θ) via a human-in-loop method to drive one or more output parameters (y) from the model to test data from a real-world system. In some methods, an experimental design is run on a real-world system to obtain observations or test data. The tuning parameters (θ) are then manually set to specific test values, and a simulation or analysis is run using the model to generate test output parameters (y) from the model. The test output parameters (y) are then compared with the test data from the real-world system to check the alignment and error between the test output parameters (y) and the test data. Experiments are conducted, and one or more tuning parameters (θ) are adjusted each time until the model is fully calibrated. That is, until alignment is achieved between the model output parameters (y) and the test data from the real-world system, and / or the error between the model output parameters (y) and the test data is reduced to an acceptable level.

[0016] Tuning based on trial and error is very time consuming, and in some cases, may take several months. For example, the model output parameters (y) from the experiment and the test data from the real-world system can be in the form of a large time data set with high dimensions. In addition to the size, the test data set may include noise, bias, or measurement errors, which makes it difficult to tune the model data. In addition, the model data may include physical simplifications for capturing the real-world system, so not all real-world effects can be modeled. In addition, a large number of experiments may be required to achieve a fully tuned model. Multiple output parameters (y) of the model may also have to be calibrated together, and the tuning parameters (θ) may have multiple potential solutions, which further complicates the analysis. In addition, the automation of tuning based on trial and error involves a lot of computational work, which may be daunting. The ability of humans to evaluate model data and test data is limited. Automatic calibration via manual intervention methods may be too computationally consuming and may not be able to handle defects in test data and model data. These are major challenges in application settings.

[0017] In general, various aspects of the present disclosure can be used to reduce the time and computational effort involved in model calibration. The methods and systems described herein for calibrating a model use data compression to reduce the size of a large temporal data set (i.e., a data set from a model and a real-world system), and use Bayesian data fusion techniques to fuse the reduced data sets together. In addition, the methods and systems use a parallel Markov chain Monte Carlo (MCMC) algorithm to identify a series of potential values ​​for a tuning parameter (θ). For example, the series of potential values ​​can be in the form of a probability distribution of the tuning parameter (θ).

[0018] The methods and systems for model calibration described herein can be used to select or generate values ​​of tuning parameters (θ) for a model, which optimize or otherwise improve one or more output parameters of the model using parallel Bayesian reasoning techniques. Optimization can involve minimizing or reducing the error between test data reflecting observed values ​​in a real-world system and model data reflecting predicted values ​​generated via the model. For example, for multiple output parameters (y) of a model, the tuning parameters can reduce the error between the test data and the model data. The use of Bayesian reasoning techniques in combination with data compression algorithms reduces the computational work involved in calibrating model data and test data during calibration. In addition, computational efficiency enables the methods and systems described herein to generate values ​​of one or more tuning parameters (θ) together, thereby generating a series of tuning parameters for the model. Therefore, multiple tuning parameters can be selected together to reduce the error between model data and test data, rather than selecting the value of each tuning parameter (θ) individually.

[0019] It is contemplated that the models referred to herein may be any type of model, such as a physics-based model, a numerical model, or a chemistry-based model. Other non-limiting examples of model types include component-level models (CLMs), piecewise linear models (PLMs), computational fluid dynamics (CFD) models, heat transfer models, financial models, or finite element analysis models. Furthermore, the model may be a simple (i.e., simplified) model or a complex model.

[0020] In some approaches, the methods and systems for calibrating models described herein may be used to calibrate models of engines, such as gas turbine engines that may be used in an aviation environment. Because engine models involve dense time series data, calibrating engine models may be computationally expensive, and the methods described herein may be used to calibrate engine models in a computationally efficient manner. However, it is contemplated that in other approaches, these methods and systems may be used to calibrate models of other real-world systems.

[0021] It is also contemplated that the methods and systems for model calibration described herein may be applied to models involving any type of transient data (e.g., in addition to time series data). For example, the model data and test data may be discretized by any other suitable variable (such as flight points, engine cycles), or otherwise instead of being discretized by time. Thus, among other approaches, the methods and systems described herein may be used to calibrate models for predictive health management (PHM) and / or engine health management (EHM). Such PHM and / or EHM models may include data discretized by flight points that, when arranged in chronological order, constitute a date / time axis. It should also be understood that the model data and test data referred to herein may be discretized in equal (or approximately equal) increments, or may be discretized in unequal increments.

[0022] Although the systems and methods described herein refer to calibrating model data to test data from a real-world system, it is also contemplated that the systems and methods may calibrate model data to any other type of data for any purpose. For example, the test data or target data to which the model data is calibrated need not be from a real-world system, but may be any target. In one non-limiting example, the target data may be output from another model, software program, or system.

[0023] The terms and expressions used herein have the ordinary technical meanings given to these terms and expressions by those skilled in the art in the above-mentioned technical fields, unless different specific meanings are set forth herein. Unless otherwise specifically indicated, the word "or" used herein should be interpreted as having a disjunctive structure, not a conjunction structure. Unless otherwise specified herein, the terms "connect", "fix", "attach to", etc. refer to direct connection, fixation or attachment, as well as indirect connection, fixation or attachment through one or more intermediate components or features.

[0024] The singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise.

[0025] As used herein throughout the specification and claims, approximating language is applied to modify any quantitative representation that can be permissibly varied without resulting in a change in the basic function to which it is related. Thus, a value modified by one or more terms such as "approximately," "approximately," and "substantially" is not limited to the precise value specified. In at least some cases, approximate language may correspond to the precision of an instrument used to measure a value, or to the precision of a method or machine used to construct or manufacture a component and / or system. For example, approximate language may mean within a 10% margin.

[0026] The above and other benefits may become more apparent upon a thorough review and study of the following detailed description. Referring now to the drawings, and particularly to the Figure 1, a system 100 compatible with many of these teachings will now be presented. The system 100 can be used to calibrate a model. In particular, the system 100 can be used to generate or determine one or more possible values ​​of a tuning parameter (θ) of a model. Examples of tuning parameters for an engine or cycle performance model include, but are not limited to, compressor efficiency, turbine flow, and heat transfer coefficient.

[0027] The system 100 includes a test device 106 , a display device 108 , a user input device 110 , a model 112 , a tuning module 114 , one or more databases 116 , and a control unit 118 .

[0028] The test rig 106 is configured to perform operational tests on the real-world system 102. In some embodiments, the real-world system is an engine, such as a gas turbine engine or a portion thereof. It is contemplated that the test rig 106 may perform on-wing and / or off-wing engine testing. The test rig 106 may include one or more cameras, probes, and / or sensors that measure, detect, and / or otherwise determine an operational parameter (y′) of the real-world system 102. The sensors employed by the test rig 106 may be temperature sensors, pressure sensors, flow sensors, speed sensors, airflow sensors, mass flow sensors, vibration sensors, liquid level sensors, and the like. Examples of operational parameters (y′) that may be measured by the test rig 106 include, but are not limited to, temperature, pressure, vibration level, rotor speed, oil level, fuel flow rate, fuel level, oil filter differential pressure, fuel filter differential pressure, oil filter differential pressure, and the like. When the real-world system 102 is an engine of an aircraft, the operational parameter (y′) may also include flight data and / or aircraft data, such as altitude and Mach number. The test device 106 collects, generates, or otherwise acquires test data (ie, observation data, site data) that is indicative of actual operation of the real-world system 102 .

[0029] The test data may include data regarding one or more of the above-mentioned operating parameters (y'). However, it should be understood that the test data may include any parameter of the real-world system 102 that can be directly or indirectly measured, sensed, or observed, or any parameter that can be calculated based on the measured, sensed, or observed parameter. In addition, when the real-world system 102 is an engine (such as an engine of an aircraft), the operating parameter may be measured on the wing or on the ground, or may be calculated based on the parameters measured on the wing or on the ground. Although reference is made to Figure 2 The test data is from a real-world system, but the test data does not need to be observation data from a real-world system, but can be any type of target data, such as data from software, output from other models, or data from any other type of system.

[0030] In some examples, when the real-world system 102 is an engine, the test rig 106 may acquire data during startup, acceleration from idle to takeoff power, deceleration from takeoff power to idle, shutdown, etc. The test rig 106 may acquire fuel flow, engine temperature, engine pressure, speed, or similar data.

[0031] The display device 108 is configured to present data, images, or other information related to the calibration of the model 112. The display device 108 can display or present test data collected by the test device 106. In addition, the display device can display or present output parameters (y), input parameters (x), and / or tuning parameters (θ), for example, in a graphical representation or otherwise. In some aspects, the display device 108 is associated with the test device 106. The display device 108 can be of any suitable type, including, but not limited to, a cathode ray tube (CRT) device, a liquid crystal display (LCD) device, a light emitting diode (LED) device, etc. The display device 108 can also present additional information or attributes associated with the real-world system 102, and / or an identifier associated with the system, such as a name, model, type, date of manufacture, etc.

[0032] The user input device 110 is configured to receive input from a user. The user's input may include, but is not limited to, device settings, user preferences, component information, etc. The user input device 110 may take any suitable form. For example, the user input device 110 may include, but is not limited to, a mouse, a touch screen, a keyboard, a track pad, a digital pad, etc.

[0033] Model 112 includes one or more mathematical or physical based functions that represent the behavior of real world system 102. Model 112 simulates, analyzes, or otherwise describes the operation of real world system 102. Model 112 receives one or more input parameters (x) (e.g., x 1 …x p ), and generates one or more output parameters (y) (e.g., y 1 …y m The output parameters (y) may include any combination of the above operating parameters. The model 112 also includes one or more tuning parameters (θ) (e.g., θ 1 …θ n ). The model 112 may model the operation of the real-world system 102 and predict or estimate operating parameters of the real-world system 102, such as temperature, pressure, vibration level, rotor speed, oil consumption, fuel consumption, or similar operating parameters as the output parameter (y). In this manner, the model 112 generates or otherwise determines model data indicative of simulated or predicted operation of the real-world system 102. The model data may include one or more of the above-described output parameters (y) of the model 112.

[0034] In some approaches, model 112 is a physics-based model, and in some aspects, model 112 is a physics-based cycle model. Model 112 can model the thermal response of the engine and predict characteristics, such as fuel flow and temperature. Model 112 can use such characteristics to predict performance characteristics, such as fuel burned to generate thrust. When model 112 is a model of an engine, examples of input parameters (x) include, but are not limited to, altitude, Mach number, ambient temperature, throttle lever angle, etc. Examples of output parameters (y) include, but are not limited to, engine rotor speed, fuel flow, exhaust temperature, etc. Examples of tuning parameters (θ) include, but are not limited to, compressor efficiency, turbine flow, heat transfer coefficient, etc.

[0035] In other approaches, the model 112 may be a physics-based model of any system that can be described by physics. For example, the model may be a physics-based model of combustion dynamics, heat transfer, transient materials, or control and electrical applications (e.g., a proportional-integral-derivative (PID) controller).

[0036] The tuning module 114 includes one or more algorithms for determining one or more possible values ​​of the tuning parameters (θ) of the model 112. The algorithms may include, but are not limited to, data compression algorithms, data fusion algorithms, and parallel MCMC algorithms. The tuning module 114 may be a Bayesian machine. It is contemplated that by employing these three types of algorithms (e.g., data compression, data fusion, and parallel MCMC), the tuning module 114 can make the Bayesian approach faster. Due to the reliance on MCMC that does not occur in parallel, a fully Bayesian approach is typically computationally expensive. In order to circumvent the bottleneck of a typical MCMC, the tuning module 114 employed herein utilizes a parallel MCMC approach to reduce computational overhead by utilizing parallel model processing and by connecting with data compression and data fusion algorithms. In Figure 5 An exemplary parallel MCMC method is shown in .

[0037] A data compression algorithm reduces the size or dimensionality of data (such as test data or model data). The data compression algorithm may involve one or more of principal component analysis (PCA), locally linear embedding (LLE), compressed variational autoencoder (CVAE), and manifold learning. Any suitable data compression algorithm that can reduce the size or dimensionality of a data set may be employed. The data compression algorithm reduces the number of bytes required to represent the data set and the amount of memory required to store the data. In some aspects, the data compression algorithm may reduce the size of a data set from 1000s of data points or more to 10s of data points or less. In Figure 4 An exemplary data compression method is shown in FIG.

[0038] The data fusion algorithm mathematically combines or provides joint analysis data from multiple sources. In some embodiments, the data fusion algorithm can fuse model data and test data in compressed or uncompressed form. Any suitable data fusion algorithm can be used, such as high-fidelity modeling.

[0039] The parallel MCMC algorithm involves systematic random sampling from a distribution to construct the most likely distribution of a parameter. The MCMC algorithm generates samples of several values ​​as tuning parameters of a function (e.g., model 112) to create an approximation of the distribution. Figure 4 and Figure 5 An exemplary MCMC method is described in more detail. In addition to parallel MCMC, it is contemplated that any parallelizable Bayesian inference technique may be used in the tuning module 140. The Markov Chain Monte Carlo (MCMC) algorithm discovers a set of potential solutions that also explain the ground truth (i.e., test data or field data) while using rules that reduce the computational load imposed by a large number of parameters and large time series data sets.

[0040] The one or more databases 116 may include or store one or more types of data useful for calibrating the model 112. The one or more databases 116 may include test data obtained via the test device 106 and / or model data obtained via the model. The one or more databases 116 may also include any combination of data used by the tuning module 114 to determine the value of one or more tuning parameters (θ) of the model 112. In some embodiments, the one or more databases 116 may also include tuning parameters (θ) or possible values ​​of tuning parameters (θ).

[0041] The control unit 118 (also referred to as a controller or control circuit) may be used as a computing device to perform the functions and methods described herein. It is contemplated that the control unit 118 may include one or more of the model 112 and / or the tuning module 114, or may otherwise communicate with these components of the system 100. The control unit 118 may include one or more processors 120, an input / output (I / O) device 122, a transceiver 124, and a memory device 126. The processor 120 may include any suitable processing device, such as a microprocessor, a microcontroller, an integrated circuit, a logic device, or other suitable processing device. The processor 120 may be used to perform or assist in performing the steps of the processes, methods, functions, and techniques described herein, and to control various communications, decisions, programs, content, lists, services, interfaces, logging, reports, and the like. In addition, the one or more processors 120 may access the memory device 126, which may store instructions 130, codes, and the like that are implemented by the processor 120 to implement the intended functions.

[0042] The memory device 126 typically includes one or more processor-readable and / or computer-readable media accessed by at least the processor 120, and may include volatile and / or non-volatile media, such as RAM, ROM, EEPROM, flash memory, and / or other memory technologies. In addition, the memory device 126 is shown as being internal to the control unit 118; however, the memory device 126 may be internal memory, external memory, or a combination of internal and external memory. Similarly, some or all of the memory devices 126 may be internal memory, external memory, or a combination of internal and external memory of the processor 120. The memory device 126 may be substantially any relevant memory, such as, but not limited to, one or more of a solid-state storage device or drive, a hard drive, a universal serial bus (USB) stick or drive, a flash secure digital (SD) card, other memory cards, and other such memories or a combination of two or more such memories, and some or all of the memories may be distributed at multiple locations on a computer network. The memory device 126 may store data 128 such as code, software, executable programs, scripts, data, content, lists, programming, programs, log or historical data, engine information, component information, etc. Although Figure 1 The various components are shown coupled together via a bus, but it should be understood that the various components may actually be coupled directly to the control unit 118 and / or one or more other components.

[0043] Typically, the control unit 118 also includes one or more communication interfaces, ports, or transceivers 124, etc., to allow the control unit 118 to communicate with other devices, and / or other such communications or a combination of two or more such communication methods, via a communication bus, a distributed computer, and / or a communication network (e.g., a local area network (LAN), the Internet, a wide area network (WAN), etc.). In addition, the transceiver 124 can be configured for wired, wireless, optical, fiber optic cable, satellite, or other such communication configurations, or a combination of two or more such communications.

[0044] I / O device 122 may be any relevant port or combination of ports, such as, but not limited to, a USB, Ethernet, or other such port. I / O device 122 may be configured to allow wired and / or wireless communication to connect to external components. For example, I / O device 122 may provide wired and / or wireless communication (e.g., Wi-Fi, Bluetooth, cellular, RF, and / or other such wireless communications), and in some cases, may include any suitable wired and / or wireless interface devices, circuits, and / or connection devices, such as, but not limited to, one or more transmitters, receivers, transceivers, or a combination of two or more such devices.

[0045] The control unit 118 may receive data from various components of the system 100. In addition, the control circuitry may transmit data to various components of the system 100. For example, the control unit 118 may receive data from the test device 106, which may include test data indicating measured or observed operating parameters or performance characteristics of the real-world system 102. The control unit 118 may also receive data from the model 112. For example, the control circuitry may receive data regarding input parameters (x), output parameters (y), and / or tuning parameters (θ) of the model 112. The control unit 118 may also receive input data from the user input device 110 and provide instructions to the test device 106 based on such input data. It is contemplated that the tuning module 114 may be integrated into the control unit 118 or separate from the control unit 118.

[0046] like Figure 1 As shown, the various components of system 100 may communicate directly, or indirectly, such as through one or more distributed communication networks, such as network 132. For example, network 132 may include a LAN, WAN, Internet, cellular, Wi-Fi, and other such communication networks, or a combination of two or more such networks.

[0047] In operation, system 100 calibrates model 112 using test data acquired via test device 106, model data acquired via model 112, and tuning module 114. In some aspects, the test data and model data may be stored in one or more databases 116 for use by tuning module 114.

[0048] The test device 106 performs tests on the real world system 102 to generate test data. The test data includes one or more operating parameters (y') monitored or otherwise observed by the test device 106. Model 112 is used to perform simulations or analyses to generate model data. The model data includes one or more output parameters (y) predicted or estimated using the model 112. The tuning module 114 then reduces the size of the test data and the model data via a data compression algorithm. The compressed model data is fused with the compressed test data. The fused data is then analyzed via a parallel MCMC algorithm to determine possible values ​​of one or more tuning parameters (θ) of the model 112. The determined values ​​of the tuning parameters (θ) can be used in the model 112 to generate a calibration model or a tuning model.

[0049] In some embodiments, the system 100 may also optionally include a manufacturing device 134. The manufacturing device 134 may be any suitable manufacturing device capable of manufacturing components of the real-world system 102. A suitable manufacturing device 134 may be, for example, an additive manufacturing device. In some aspects, the system 100 may cause the manufacturing device 134 to manufacture components of the real-world system 102 based on results from a simulation run using the calibration model.

[0050] In other embodiments, the system 100 may also optionally include a control system 136. The control system 136 may be operable to control the operation of the real-world system 102. When the real-world system 102 is an engine, the control system 136 may be an engine control system. In some aspects, the system 100 may cause the control system 136 to adjust at least one setting of the control system 136 based on the results of a simulation run using the calibration model.

[0051] Go to Figure 2 A tuning module 140 is provided for performing data compression driven tuning. The tuning module 140 may be a Bayesian machine configured to employ Bayes' theorem to determine the conditional probability that one or more tuning parameters (θ) of the model will minimize or otherwise reduce the error between the test data and the model data.

[0052] The tuning module 140 receives test data directly or indirectly from a test device that monitors the real-world system 142. The test device performs operational tests on the real-world system to determine the operating parameters y′ of the real-world system. 1 , y′ 2 , y′ 3 The test data is provided as a function of time operating parameters y ′ 1 , y′ 2 , y′ 3 time series data.

[0053] The tuning module 140 receives model data directly or indirectly from a model 144 of the real world system. The model 144 includes tuning parameters θ 1 …θ n The model 144 predicts or estimates the operating parameters of the real-world system 142. The predicted operating parameters are generated as output parameters y of the model 144. 1 ,y 2 ,y 3 The model data is provided as a function of time output parameter y 1 ,y 2 ,y 3 time series data.

[0054] The tuning module 140 then analyzes the model data and the test data using a data compression algorithm, a Bayesian data fusion algorithm, and a parallel MCMC algorithm to derive a distribution 148 of tuning parameters θ for the model 144. The distribution 148 provides the probability that a particular value of the tuning parameter θ is a suitable candidate for the tuning parameter (i.e., will minimize or otherwise reduce the error between the test data and the model data). Compressing the time series data and applying parallel MCMC on the compressed data speeds up the model calibration process. In addition, tuning via the tuning module improves the accuracy of the calibration process, such that the predicted operating parameters (i.e., the output parameters y) from the model are more accurate than the predicted operating parameters (i.e., the output parameters y) from the model. 1 ,y 2 ,y 3 ) and the observed operating parameters (i.e., operating parameters y′) observed by testing the real-world system 142 1 , y′ 2 , y′ 3 ) are closely aligned. The one or more tuning parameters obtained via tuning module 140 may be used in model 144 to generate a tuning model.

[0055] In some approaches, the tuning module 140 may also generate a difference model 146. The difference model 146 provides a prediction of the operating parameters (i.e., the output parameters y) generated using the tuning model. 1 ,y 2 ,y 3 ) and the observed operating parameters (i.e., operating parameters y′ 1 , y′ 2 , y′ 3 ) between the two (i.e., also called the difference or variance). Figure 2 In , the error between the observed operating parameter and the predicted operating parameter is represented by δy. For example, the difference model 146 may provide time series data that predicts the error δy as a function of time.

[0056] If the difference model 146 can be related to physically based parameters, the difference model 146 can be used to further tune the model. In addition, the difference model 146 can also identify the source of the difference, in particular, as the error δy and one or more model input parameters (such as Figure 1 Once identified, the tuning module 140 can exploit the correlation between the error δy and the model input parameters (x) in the model to further improve the model prediction.

[0057] refer to Figure 3A and Figure 3B , provides an exemplary method for calibrating a model of an engine. In a typical application setting, Figure 3A and Figure 3BMany (but not necessarily all) of the steps shown in are performed using a model calibration system, which in some approaches may include one or more components of system 100. Although the method is described with reference to an engine, it is contemplated that the method may also be applied to other real-world systems.

[0058] At box 150, the model calibration system 100 receives test data from an operational test of the engine. The operational test may be performed via one or more test devices that monitor one or more performance characteristics of the engine. Such operating parameters may include, for example, engine temperature, speed, engine pressure, flow rate, fuel consumption, etc. The test data includes observed operating parameters or actual operating parameters of the engine. It is also contemplated that the test data may include performance characteristics of the engine, which the model calibration system 100 calculates based on the observed operating parameters of the engine.

[0059] At block 152, the model calibration system 100 receives model data from a model of the engine. For example, the model may perform a simulation of the engine to model operating parameters of the engine to generate model data. As described above, the model may be used to predict operating parameters of the engine (such as fuel flow, speed, temperature, pressure, etc.) to generate model data. It is contemplated that in some aspects, the model data may also include performance characteristics, such as thrust-specific fuel consumption (i.e., fuel burned to generate thrust) of the engine calculated by the model based on the operating parameters.

[0060] At block 154, the model calibration system 100 compresses the test data to generate compressed test data. The model calibration system 100 also compresses the model data to generate compressed model data. Compression reduces the dimensionality of the test data and the model data.

[0061] At block 156, the model calibration system 100 fuses the compressed test data with the compressed model data to generate fused data. The model calibration system 100 may fuse the test data and the model data using one or more data fusion algorithms. The fused data may have a segmentation of input parameters, tuning parameters, and output parameters that existed prior to data compression. However, for several performance outputs simultaneously or in conjunction with one another, the data size of the input data (e.g., data prior to fusion) is reduced, for example, from a time scale of 1000s to 10s or less.

[0062] At block 158, the model calibration system 100 performs a parallel MCMC simulation using the fused data to identify at least one possible value for a tuning parameter of the model. It is contemplated that the model calibration system 100 may employ any parallelizable Bayesian inference technique to determine the value of the tuning parameter. In particular, the model calibration system 100 may generate a distribution of the tuning parameter, wherein the distribution provides a probability that a particular value is a suitable candidate for the tuning parameter.

[0063] At box 160, the model calibration system 100 selects a value of a tuning parameter from at least one possible value of the tuning parameter to calibrate the model. At box 162, the model calibration system 100 may then generate a calibration model using the selected value of the tuning parameter. The model calibration system 100 may then use the calibration model to generate calibration model data. The calibration model data may, for example, provide calibrated predictions of model output parameters (such as operating parameters of the engine). For example, the calibration model may predict engine response under conditions not run in the test (i.e., extrapolate the results to an area where the engine is not running), such as at different altitudes, inlet temperatures, control plans, etc.

[0064] The determination of the composition of the calibration model is typically set by the intended use of the model and / or design practice, and is based on the error or difference between the model data and the test data. It is contemplated that any target error can be achieved based on the use case. In some aspects, the calibration model can exhibit an error of about 50% or less, an error of 20% or less, an error of about 10% or less, or in some aspects, an error of 5% or less, where the error reflects the difference between the model data (i.e., the output parameter predicted by the model using the selected tuning parameters) and the test data.

[0065] It is contemplated that this method may be used to generate possible values ​​for a plurality of different tuning parameters of a model.In this manner, the model calibration system 100 may generate a series of tuning parameter values ​​that may be used to generate a calibration model or a tuned model.

[0066] In some approaches, the model calibration system 100 may optionally compare the calibration model data with the test data to generate a difference model at block 163. The difference model provides the error between the calibration model data and the test data.

[0067] In some methods, at box 164, the calibration model can be used to generate the predicted operating parameters of the engine. The calibration model can also be used to predict or predict the performance characteristics of the engine. This performance characteristic can include the performance characteristics of the operation of the engine, including the performance characteristics under new input parameters and / or untested input parameters. The performance characteristics may include, but are not limited to, the thrust, fuel combustion, speed, airflow, pressure and / or temperature of the entire engine system. For example, the calibration model can be used to evaluate the performance parameters of the engine within the flight envelope. In addition, not all parameters of the engine can be feasibly sensed during engine operation (e.g., airflow at the burner inlet, high-pressure turbine blade temperature, etc.). The calibration model can be used to predict such parameters. In addition, the calibration model can be used to simulate overspeed events, for example, to predict the conditions that cause overspeed and the degree of overspeed under various engine conditions. In this way, the calibration model can be used to certify that the engine meets airworthiness standards.

[0068] In some methods, at box 192A, the model calibration system 100 can also be selectively used to determine at least one design parameter of an engine component using the calibration model. For example, the calibration model can be used to determine a design parameter of an actuator (e.g., as an actuator length or rod area), a fuel pump (e.g., a total flow tolerance), or a booster (e.g., based on the stall position of the engine in the flight envelope). The predicted performance characteristics can be used to determine one or more design features and / or parameters of the engine component. That is, the calibration model can be used to predict the performance characteristics and select design features and / or parameters of the engine to achieve a specific performance characteristic.

[0069] At block 192B, the model calibration system 100 may cause the device to manufacture the engine component according to the design parameters. The device may be any suitable manufacturing device, such as an additive manufacturing device. Thus, the engine component may then be manufactured based on the design features and / or parameters.

[0070] For example, a calibrated model of an engine can be used to evaluate preliminary designs of engine components. The calibration model can provide a direct representation of the engine in digital form. To determine the performance of the preliminary design, simulations can be run using the calibration model. For example, the calibration model can be used to evaluate how well the preliminary design performs relative to performance goals or other objectives. In this way, the calibration model can be used to evaluate whether specific design parameters are appropriate. Transient simulations using the calibration model can be run to study the preliminary design under time-varying conditions of power levels, sudden load applications, etc. Figure 3A and Figure 3B The accuracy of the predictions of the calibration model generated by the method may allow the design of engine components to be evaluated without running physical tests on the components or with reduced physical testing of the components.

[0071] In another example, the calibration model may be used to certify the engine. Fig.10 An exemplary certification process using a model calibrated using the methods described herein is detailed. For example, during the engineering development phase, one or more engine design parameters may be adjusted to improve the production design. To proceed with certification, these design modifications may involve additional engine testing or a highly accurate digital model that can be used to confirm the impact of the design changes. Fig.10 The toolset and process detailed in can be used to recalibrate the digital model to ensure it is correctly tuned to match the engine. Following the toolset and modeling approach described in this article improves the efficiency and speed of calibration. Applying traditional engine model calibration methods can take months to calibrate and certify an engine. The model calibration method described in this article can allow for rapid calibration and recalibration for certification.

[0072] In other methods, at box 194A, the model calibration system 100 may selectively use the calibration model to determine the engine control settings. At box 194B, the model calibration system 100 may cause or otherwise affect an adjustment of at least one control parameter of the engine based on the output of the calibration model. In some aspects, the model calibration system 100 may cause the engine control system to adjust at least one control parameter of the engine.

[0073] For example, a calibration model can be used to determine the initial control settings for an engine. The control settings may include, but are not limited to, a power management plan, a fuel management plan, positioning stator blades, and the like. Based on these control settings, the calibration model can predict various parameters about the engine, such as pressure, temperature, and / or flow rate. A calibration model can be used to run a simulation in order to optimize or otherwise improve engine performance by selecting appropriate control settings. The calibration model can help select control settings from several different perspectives to optimize or improve engine performance, including, for example, maximizing range, reducing fuel consumption, increasing thrust, improving operability, reducing fuel burn during step climbs, improving climb performance, and the like. In some methods, initial control settings can be set for testing and also for optimization. For example, control settings can be adjusted from cautious or conservative settings to more aggressive settings to drive the performance of the engine. Using a digital model for this control system optimization can greatly reduce the amount of time required for full engine testing for design, certification, and on-site delivery of initial or upgraded products.

[0074] In another example, the calibration model can be used to evaluate alternative control methods, such as the pressure ratio, speed and flight condition method. The calibration model can be used to run tests to compare one control method to another control method. This method can be used to determine whether controlling based on one type of data may be more relevant or more advantageous than controlling based on another type of data.

[0075] In some approaches, the calibration model can also be used to develop a linear model that can be used in a multi-loop control system with two control loops (an inner loop and an outer loop). For example, the inner loop can be used to determine how to get the variable stator vanes to a certain position, and the outer loop can determine the target position of the variable stator vanes of the engine.

[0076] In other methods, at box 196A, the model calibration system 100 can use the calibration model to determine or assess compliance with engine certification standards. The model calibration system 100 can compare performance parameters predicted using the calibration model with one or more defined performance standards to determine whether the engine meets certification and / or airworthiness standards.

[0077] For example, as part of the certification effort for engine operation, a calibration model can be used to demonstrate stall-free performance around the flight envelope. The calibration model can be developed from a limited number of unique engines tested on a small subset of the flight envelope. Using the calibration model, the characteristics or performance of one engine can also be applied to the entire engine family to further improve the efficiency of the certification effort. In some approaches, the calibration model can be used to certify by analysis without the need for engine testing or with reduced engine testing. Figure 3A and Figure 3B The method can generate a calibrated model with relatively low error (e.g., error of the model output compared to actual engine test values). In some aspects, the error of the predictive model can be between 5% and approximately 10% error before calibration is applied. Once the model is calibrated, the error can be reduced to less than 5% error, or in some aspects, for certification critical aspects, to less than 2% error. The low error in the model output can enable certification of an engine or its components without engine testing or with reduced engine testing. The use of such a low-error model can facilitate certification reports, such as engine operations that would otherwise require full flight envelope testing of the engine, or failure analysis testing that may put engine development hardware at risk.

[0078] By eliminating or reducing the need for engine testing, compared to methods using actual engine testing Figure 3A and Figure 3B The calibrated model generated by the method can also enable certification of an engine or evaluation of engine design parameters or control settings to occur in a relatively short period of time, e.g., in hours or days rather than years. For example, tens of thousands to millions of simulations can be run to evaluate engine operating performance, and with a well-calibrated model, the model can be used instead of running these simulations on the actual engine.

[0079] In other approaches, the method may include a feedback loop whereby test data is iteratively fed back into the model. In this way, the model may be re-tuned using new test data to further improve model accuracy and fine-tune the values ​​of tuning parameters. In some aspects, there may be real-time feedback into the model such that the model calibration system 100 may receive new test data in real-time and use the new test data to re-tune the model. Figure 3A and Figure 3B The method or part thereof is retuned in real time to newly received test data.

[0080] although Figure 3A and Figure 3B The method is applied to engines, but it is envisioned that the method can be used with models of other real-world systems.

[0081] Figure 4-7 An exemplary analysis that may be performed via a tuning module to determine values ​​of tuning parameters of a model is provided.

[0082] Figure 4 An exemplary method for data compression is shown, which reflects a PCA-based mapping to reduce the dimensionality of transient data (i.e., time-varying data). Data compression takes time-varying parameters (y) and projects them into spatially varying components (u, w). There are two input data sets including time-varying parameters (y): an input model data set 166 and an input test data set 166. There are two output data sets including spatially varying components (u, w): an output model data set 168 and an output test data set 169.

[0083] The input model data set 166 (e.g., the first input data set) includes time-varying model data. The input test data set 167 (e.g., the second input data set) includes time-varying test data. The input model data set 166 and the input test data set 167 include multiple (number "n") records of time-varying parameters (y). In some examples, the number "n" is within 1000s.

[0084] The PCA-based algorithm reduces the dimension of the input model dataset 166 to generate an output model dataset 168 (e.g., a first output dataset). As shown in the figure, the input model dataset 166 is divided into parameters (y 1 ) is expressed as a function of the input parameter (x), tuning parameter (θ), and time (t) (i.e., y 1 =f(x, θ, t)). In the input model data set 166, the parameters (y 1 ) is the output parameter predicted by the model of the real world system. The input parameter (x) is the input parameter of the model. The input parameter (x) represents the condition of the real world system. The output model data set 168 includes spatially varying model data and is not time-varying. That is, the output model data set 168 converts the parameter (y) into1 ) is represented by parameter w 1 , which is only a function of the input parameters (x) and the tuning parameters (θ) (i.e., w 1 =f(x,θ)). The PCA-based algorithm transforms the input model dataset 166 into the output model dataset 168.

[0085] In addition, the PCA-based algorithm also reduces the dimension of the input test data set 167 and generates an output test data set 169 (e.g., a second output data set). As shown in the figure, the input test data set 167 reduces the parameters (y 1 ) is expressed as a function of the input parameters (x) and time (t) of the real-world system (i.e., y 1 =f(x, t)). The input parameter (x) is the input parameter of the model, which represents the conditions of the real world system, such as measured or observed conditions. In the input test data set 167, the parameter y 1 is a parameter measured or observed in a real-world system. The output test data set 169 includes model data that varies in space and is not time-varying. That is, the output test data set 169 converts the parameters (y 1 ) is expressed as parameter (u 1 ), which is only a function of the input parameter (x) (i.e., w 1 =f(x,θ)). The PCA-based algorithm transforms the input test data set 167 into the output test data set 169.

[0086] The output data sets (i.e., output model data set 168 and output test data set 169) include fewer records (number "m") than the input data sets (i.e., input model data set 166 and input test data set 167). That is, the PCA-based algorithm compresses "n" records in the input data set into "m" records in the output data set. That is, the number "m" is less than the number "n". In some aspects, the number "m" is within 10s, in some aspects, it is less than 10, and in other aspects, it is between about 5 and about 10.

[0087] like Figure 4 As depicted in FIG. 1 , data compression eliminates the time-varying components of the input model dataset 166 and the input test dataset 167. In this form, the output model dataset 168 can be fused with the output test dataset 169. For example, the output model dataset 168 has the parameters (w 1 ) can be compared with the data of the output test data set 169 about the parameter u 1) to generate a fused data set based on a common value of an input parameter (x). Thus, the data fusion algorithm fuses two spatially varying data sets (e.g., an output model data set 168 and an output test data set 169) to generate a fused data set. A tuning module (such as tuning module 114) can use the fused data set to optimize the tuning parameters (θ) of the model. In some approaches, a mathematical formula or equation (such as a parallel MCMC algorithm) can determine which tuning parameters (θ) best align the output model data set 168 with the output test data set 169.

[0088] Figure 5 A simulation using parallel MCMC processing is shown. The simulation generates a distribution 170 of the tuning parameter (θ) of the model. Distribution 170 is a posterior distribution. Distribution 170 is generated based on distribution 171 used as a prior distribution. The values ​​along vertical line 172 show different values ​​of the tuning parameter (θ). Distribution 170 reflects the probability that a specific value is a suitable candidate for the tuning parameter (θ). Multiple particles are analyzed in parallel - particle 174A, particle 174B, particle 174C, particle 174D, particle 174E - to simplify and increase the speed of analysis. Each particle has an initial sample value to start the stochastic optimization of the tuning parameter (θ). Through analysis, these lines show the trajectory of each particle. Line 175A shows the trajectory of particle 174A. Line 175B shows the trajectory of particle 174B. Line 175C shows the trajectory of particle 174C. Line 175D shows the trajectory of particle 174D. Line 175E shows the trajectory of particle 174E. The dashed lines and dots represent values ​​for particles that were rejected as tuning parameter values. The solid lines and dots represent values ​​that were accepted as tuning parameter values. Figure 5 As shown, the analysis of multiple particles in parallel in parallel MCMC results in faster analysis and improves the computational efficiency of the model calibration process compared to naive (ie, non-parallel) MCMC with computationally heavy single particle operations.

[0089] The simulation may run an MCMC algorithm, may run a set number of iterations, and select the highest value of the tuning parameter based on a probability weight obtained by fusing the likelihood of explaining the test data with a prior value of the value that is the optimal solution.

[0090] Imagine parallel MCMC algorithms such as Figure 5) can be used to provide possible values ​​of tuning parameters (θ) that generate accurate predictions for multiple model output parameters (y). That is, the MCMC algorithm can optimize or reduce the error between the model data and the test data for multiple parameters (y), rather than selecting tuning parameter values ​​for each output parameter separately. In addition, using this parallel MCMC algorithm, the distribution of multiple tuning parameters (θ) in the model can be obtained simultaneously. In this way, the methods described herein can provide a set of solutions (i.e., a series of tuning parameters) that will be applicable to multiple output parameters of the model, or in some aspects, to all output parameters of the model. For example, where the model includes 5 tuning parameters, all 5 parameters can be tuned together using the methods described herein. In addition, all 5 parameters can be tuned to optimize multiple output parameters of the model (e.g., flow, temperature, and pressure), rather than a single output parameter (e.g., temperature).

[0091] Figure 6 A joint graph of possible values ​​for multiple tuning parameters (θ) of a model is shown. The vertical lines - lines 180A, 180B, 180C - each represent a range of possible values ​​for different tuning parameters. Line 180A represents a range of values ​​for tuning parameter q. Line 180B represents a range of values ​​for tuning parameter T0. Line 180C represents a range of values ​​for tuning parameter h. Line 182 extending across the joint graph shows different combinations of values ​​for three different tuning parameters (tuning parameter q, tuning parameter T0, and tuning parameter h). Line 184 represents the "weight" of each different combination. The highest weight on line 184 provides the optimal or best combination of values ​​for the different tuning parameters (i.e., the combination that most closely aligns the test data and the model data).

[0092] The weights are the logarithm of the posterior probability for a given value of the tuning parameter. Fusion of the likelihood with the prior probability allows to derive an estimate of the posterior probability. The fusion is performed using Bayes' theorem. Figure 5 Distribution 170 in φ depicts an exemplary posterior probability distribution, while distribution 171 depicts an exemplary prior distribution. For a given value of the tuning parameter, the weights are approximated by fusing the likelihood that the test data is explained with the prior probability that that value is the optimal solution (i.e., the solution that minimizes the error between the test data and the model data).

[0093] Figure 7 An exemplary distribution generated by a parallel MCMC analysis using test data and model data is shown. A distribution 186 of the tuning parameter h and a distribution 188 of the tuning parameter T0 are provided. A distribution 190 of the tuning parameter q is also provided. Distribution 186 provides Figure 6The range of possible values ​​of the tuning parameter h used in the graphs in . Distributions 186, 188, and 190 are obtained by performing parallel MCMC analysis on the compressed and fused test data and model data via the methods described herein.

[0094] Distribution 186 shows Figure 6 The values ​​on the x-axis of distribution 186 represent possible values ​​of tuning parameter h that will optimize or otherwise improve model tuning. The y-axis of distribution 186 represents the probability that a particular value of h on the x-axis will optimize or otherwise improve model tuning.

[0095] Distribution 188 provides Figure 6 The range of possible values ​​of the tuning parameter T0 used in the graph is shown in Figure 188. Figure 6 The values ​​on the x-axis of distribution 188 represent possible values ​​of tuning parameter T0 that will optimize or otherwise improve model tuning. The y-axis of distribution 188 represents the probability that a particular value of T0 on the x-axis will optimize or otherwise improve model tuning.

[0096] Distribution 190 provides Figure 6 The range of possible values ​​of the tuning parameter q used in the graph is shown in Figure 190. Figure 6 The values ​​on the x-axis of distribution 190 represent possible values ​​of tuning parameter q that will optimize or otherwise improve model tuning. The y-axis of distribution 190 represents the probability that a particular value of q on the x-axis will optimize or otherwise improve model tuning.

[0097] Fig. 8A , Figure 8B and Figure 8C Another exemplary application of the method and system for model calibration described herein is shown. In particular, Figures 8A-8C A method for predictive health monitoring and health management / engine health monitoring (PHM / EHM) of an engine is shown, wherein a PHM / EHM model is tuned using the methods described herein. The PHM / EHM model may be a physics-based model for engine prognostics and / or engine health monitoring. The input data to the PHM / EHM model may be flight data and / or test cell data. Fig. 9Exemplary flight data is shown in FIG. The PHM / EHM model can predict early symptoms of developing engine failures and can therefore prevent engine failures, reducing expensive maintenance. In addition, the PHM / EHM model provides engine health forecasts that may help improve fleet management. Calibrating the PHM / EHM model using the model calibration method described herein can result in a higher confidence level in the prediction of when engine components may fail, and which engine components may fail.

[0098] First go to Fig. 8A At box 200, the PHM / EHM system receives flight data or continuous engine operating data (CEOD). It should be understood that CEOD need not be completely continuous and may have one or more interruptions. CEOD may include any information related to engine or aircraft flight and performance parameters, such as generated by a digital flight recorder or similar device. In some methods, the input data may include one or more of altitude, Mach number, and power parameters such as power level angle (PLA) or rotor speed.

[0099] At block 202, the PHM / EHM system may determine or otherwise receive an indication of a number of flight phases and sampling points in the CEOD to be run by the physics-based model. In some approaches, the physics-based model may run a simulation using CEOD data from all phases and all data points of the CEOD data (see block 202A). In other approaches, the physics-based model may run a simulation using CEOD data from one or more selected flight phases 204 (see block 202B). The flight phases 204 may include, but are not limited to, launch, taxi out, takeoff, climb, cruise, descent, approach, landing, taxi in, and shutdown.

[0100] In some approaches, the PHM / EHM system may determine or otherwise receive an indication of a sub-selection of data points in one or more flight phases. The sub-selection may involve one or more of a minimum value, a maximum value, data points around a mean with a certain confidence interval, data points around a median with a certain confidence interval, and data points around a particular distribution percentile with a certain confidence interval. The sub-selection may also occur at fixed time intervals or using any other logic.

[0101] At box 206, the PHM / EHM system can perform a single flight through a physics-based model. In some methods, the physics-based model can be a numerical propulsion system simulation (NPSS) cycle model. The NPSS model can predict various parameters about the internal conditions of the engine, such as pressure, temperature, and flow rate. The NPSS model can model a gas turbine engine or a hybrid electric engine. There may be several outputs generated by the NPSS model. An example of an NPSS model output is a predicted core rotor speed under operating conditions.

[0102] Physics-based models can use tunable coefficients (e.g., tuning parameters) that represent the health of the engine. Figure 3A and Figure 3B The method described herein is used to tune the physics-based model (see blocks 150-162). In particular, reference Figure 3A and Figure 3B The described method can be used for flight matching to match the output of a physics-based model with CEOD data.

[0103] In some aspects, at block 208, the PHM / EHM system may repeat the flight matching described with reference to block 206 to fine-tune the tuning parameters of the physics-based model. For example, the PHM / EHM system may repeat the flight matching for several flights, and in some aspects, repeat the flight matching for several consecutive flights.

[0104] At box 210, the PHM / EHM system may determine one or more changes in tuning parameters of the physics-based model. Changes in tuning parameters may indicate changes in engine health. Therefore, at box 212, the PHM / EHM system may determine one or more trends in engine health based on changes in tuning parameters. The PHM / EHM system may use statistical methods to determine engine health trends based on changes in tuning parameters that have confirmed statistical significance. In some methods, the PHM / EHM system may determine trends in tunable parameters depending on the flight phase. In other methods, the PHM / EHM system may determine how the change occurred (e.g., gradual shift, sudden shift). Trends in engine health and / or trends in tuning parameters may be used to identify one or more recommended maintenance actions for the engine.

[0105] At box 214, the PHM / EHM system may transmit or cause a signal to be transmitted to provide an electronic notification or report identifying a recommended maintenance action for the engine. In some methods. For example, the PHM / EHM system may send a recommendation to an airline or individual associated with the engine (such as a maintenance person, owner, etc.) to take one or more specific maintenance actions. The maintenance action may be intended to increase on-wing time, or convert an unplanned engine removal into a planned engine removal to improve fleet management and reduce major incidents. In another example, the PHM / EHM system may prioritize the rank order of engines with the greatest risk, or design modifications identified during this period are being implemented into field engines. In yet another example, the maintenance action may support a period of evaluating an engine design or assessing an engine redesign.

[0106] In some embodiments, a single flight data (e.g., single flight matching) may be used to calibrate the physics-based model. In other embodiments, multiple flight data may be used to calibrate the physics-based model, in which case the X-axis of the time series data may be converted to engine hours indicating engine duration.

[0107] Go to Figure 8B At block 250 , the PHM / EHM system may detect symptoms of an abnormality in the engine associated with a possible root cause based on trends in the engine health determined via the calibrated model.

[0108] At box 251, the PHM / EHM system may transmit or otherwise cause a customer notification report or alert to be transmitted to another system (such as a system associated with one or more airlines). The customer notification report may describe observed engine symptoms, possible root causes, and / or recommended maintenance actions. In one example, a symptom may be a sudden increase in oil filter increments in flight compared to its previous personalized history. Upon receiving such a symptom and confirming such a symptom using prior knowledge, a possible root cause may be identified as a sudden blockage of the filter due to various reasons (including but not limited to oil contamination or mechanical failure). For example, depending on the oil level and how far away from the upcoming bypass threshold, the recommended maintenance action may be to replace the oil filter within a specified time frame. In another example, the system may detect a sudden increase in exhaust gas temperature (EGT) and fuel flow, as well as a decrease in core speed. Upon receiving and confirming such a symptom, a possible root cause may be, but is not limited to, a loss of high pressure turbine efficiency due to an external active clearance control system leak or an internal high pressure turbine blade, shroud, or nozzle guide vane interruption. The recommended maintenance action may be to check the operation and piping of the high pressure turbine active clearance control system, and if not found, perform a borescope inspection on the high pressure turbine.

[0109] At block 252, in response to the customer notification report, one or more maintenance actions may be performed on the engine. The maintenance actions may be performed by a customer, such as an airline.

[0110] Through the maintenance action, the possible root cause identified in the customer notification report can be found and confirmed. Therefore, the engine is determined to have a root cause anomaly, and the calibration model correctly identified the root cause anomaly (e.g., a true positive). In this case, maintenance costs may be reduced and time on wing increased. The true positive can be used to further train the PHM / EHM model and improve model accuracy.

[0111] Alternatively, through maintenance actions, a possible root cause may not be found. Therefore, the engine is determined to not have a root cause anomaly, and the calibration model does not correctly identify the root cause anomaly (e.g., a false positive). In this case, maintenance costs may increase and time on wing may be reduced. False positives can be used to train the PHM / EHM model and improve model accuracy. In some approaches, it may also be necessary to rebuild or recalibrate the model.

[0112] At box 254, maintenance action is not performed in response to the customer notification report. As a result, the engine may experience a possible root cause during the estimated time frame. Therefore, it can be determined that the engine has a root cause anomaly (e.g., a true positive). Not resolving the root cause anomaly may increase maintenance costs and reduce on-wing time. Again, the true positive can be used to further train the PHM / EHM model. Alternatively, the engine may not experience a possible root cause during the estimated time frame. Therefore, it can be determined that the engine does not have a root cause anomaly (e.g., a false positive) identified by a calibrated model in the customer notification report. In this case, the false positive results in the same maintenance cost and the same on-wing time. The model can be trained using false positives, or in some methods, it may need to be rebuilt or recalibrated.

[0113] Now go to Figure 8C , at box 256, the trend of the engine health identified using the calibration model may not detect any abnormal symptoms. The engine may experience an abnormal operating mode during the estimated time horizon. Therefore, it can be determined that the engine has an abnormality that is missed by the calibration model (e.g., a false negative). Missing such anomalies may increase maintenance costs and reduce time on wing. As described above, false positives can be used to further train the model, or can indicate that the model needs to be rebuilt or recalibrated.

[0114] The engine may also not experience any anomalies as expected based on the trend of the engine health. Therefore, the engine is healthy as predicted (e.g., a true negative). The resulting maintenance costs and time on wing may be the same. Such true positives can be used to further train the model to improve the accuracy of the calibrated model.

[0115] Fig. 9 Exemplary flight data or continuous engine operating data (CEOD data) are shown. The flight data provides operating data of the engine over time during different defined flight phases.

[0116] Fig.10 An exemplary method for certifying an engine by performing an analysis using a calibration model that has been calibrated by the methods described herein is shown. In a typical application setting, Fig.10 Many (but not necessarily all) of the steps shown in are performed by a model calibration system, which in some approaches may include one or more components of system 100. Although the method is described with reference to an engine, it is contemplated that the method may also be applied to other real-world systems.

[0117] The model calibration system 100 may include or be in communication with an initial transient model of the engine, as shown at block 220. The transient model includes one or more tuning parameters.

[0118] At box 222, an engine test program may be defined for the engine. An engine test program may be any program designed to collect data about the engine, for example, to certify or verify that the engine is safe to operate or fit for use. In one approach, the test program may be designed to certify no-stall performance around a flight envelope.

[0119] At block 224, ground test data for the engine is acquired. Ground testing may be conducted off-wing (e.g., in an engine test cell or other device or facility capable of simulating flight conditions on the ground). Ground testing may include any testing that determines the functionality of an engine system in a manner that reproduces flight or operation of the engine to evaluate engine performance. For example, ground testing may include flight load simulation, material statics and fatigue, structural dynamics, etc.

[0120] At block 226, using the methods described herein (e.g., reference Figure 3A and Figure 3BThe transient model is calibrated using the method described in the previous section. During calibration, the model calibration system 100 may determine values ​​of one or more tuning parameters of the transient model. Ground test data is used to calibrate the transient model; in particular, the output of the transient model is matched to the ground test data. At box 228, a calibrated transient model that has been matched (or calibrated) to the ground test data is obtained. At box 230, the engine control software is updated based on simulations performed using the ground test calibrated transient model.

[0121] At box 232, flight test data for the engine is acquired. Flight test data may be acquired on-wing during flight. Flight testing may include any test that determines the functionality of an engine system in a manner that replicates flight or runs the engine to evaluate engine performance.

[0122] At block 234, using the methods described herein (e.g., reference Figure 3A and Figure 3B During calibration, the model calibration system 100 may determine updated values ​​of one or more tuning parameters of the transient model.

[0123] At box 236, a calibrated transient model is obtained that has been matched (or calibrated) to the flight test data. At box 238, the engine control software is again updated based on simulations performed using the flight test calibrated transient model.

[0124] At box 240, the updated control software may be certified based on simulations run using a transient model calibrated by the flight test. In this way, certification may be performed by analysis rather than via ground testing or actual flight testing of the engine.

[0125] At box 242, there may be an update to the engine hardware, for example, as part of a performance improvement program for the engine. At box 244, the calibrated transient model is then updated to reflect the update to the engine hardware. In this way, at box 246, the simulation may then be run again using the calibrated transient model to perform a second certification of the engine by performing analysis using the calibrated transient model.

[0126] Further aspects of the disclosure are provided by the subject matter of the following clauses:

[0127] A method comprising: receiving test data of operating parameters of an engine from a test device; receiving model data of the operating parameters from a simulation performed via a model of the engine; compressing the test data and the model data to generate compressed test data and compressed model data; fusing the compressed test data with the compressed model data to generate fused data; performing parallel Bayesian inference simulation using the fused data to identify tuning parameter values ​​for the model; using the tuning parameter values ​​in the model to obtain a tuned model; predicting performance parameters of the engine using the tuned model; and modifying the physical design or operation of the engine based on the performance parameters.

[0128] A method as described in any preceding clause, wherein the modification is an adjustment to a design parameter of the engine.

[0129] A method as in any preceding clause, further comprising determining design parameters for an engine component based on the performance parameters; and causing a manufacturing device to form the engine component in accordance with the design parameters.

[0130] A method as in any preceding clause, wherein the modification is an adjustment to a control setting of the engine.

[0131] A method as claimed in any preceding clause, further comprising determining updated engine control settings for the engine based on the performance parameters; and causing an engine control system to adjust operation of the engine based on the updated engine control settings.

[0132] A method as in any preceding clause, further comprising determining compliance with an engine certification standard based on the performance parameter by comparing the performance parameter to a defined performance standard.

[0133] A method as in any preceding clause, wherein the parallel Bayesian inference simulation comprises a parallel Markov Chain Monte Carlo (MCMC) simulation.

[0134] A method according to any preceding clause, wherein the test data includes data regarding multiple operating parameters of the engine, wherein the model data includes predicted values ​​of the multiple operating parameters, the method comprising: using the fused data to perform parallel Bayesian inference simulations to identify at least one value of the tuning parameter that aligns the model data and the test data.

[0135] A method as in any preceding clause, using the tuning parameters in the model to generate calibrated predictions of the operating parameters.

[0136] A method as in any preceding clause, wherein the method further comprises generating a difference model providing a difference between the calibrated predicted and observed values ​​of the operating parameter.

[0137] A method as in any preceding clause, wherein identifying tuning parameters for the model comprises generating a probability distribution for the tuning parameters, the probability distribution providing a probability that a particular value is a suitable candidate for the tuning parameter.

[0138] A method as described in any preceding clause, wherein compressing the test data includes at least one of principal component analysis (PCA), locally linear embedding (LLE), compressed variational autoencoder (CVAE), or manifold learning.

[0139] A method as claimed in any preceding clause, wherein said test data and said model data are transient time series data of said operating parameters of said engine.

[0140] A method as described in any preceding clause, wherein the test device comprises a sensor which determines an observed value of the operating parameter of the engine.

[0141] A method as in any preceding clause, further comprising comparing the performance parameter of the engine to a defined performance standard, thereby certifying the engine.

[0142] A method according to any preceding clause, further comprising at least one of: causing an engine control system to adjust at least one control parameter of the engine based on the engine control setting; or causing an apparatus to manufacture the engine component according to the design parameter.

[0143] A method as described in any preceding clause, wherein the test data comprises observed values ​​of the operating parameters from operational testing, and wherein the model data comprises predicted values ​​of the operating parameters from the simulation.

[0144] A system for calibrating a model, the system comprising: at least one processor; and a memory device storing instructions, the instructions, when executed by the at least one processor, causing the at least one processor to operate, the at least one processor being configured to: receive test data of operating parameters of an engine via a test device in communication with the at least one processor; receive model data of the operating parameters from a simulation performed via a model of the engine; compress the test data and the model data to generate compressed test data and compressed model data; fuse the compressed test data with the compressed model data to generate fused data; perform parallel Bayesian inference simulation using the fused data to identify tuning parameter values ​​of the model; use the tuning parameter values ​​in the model to obtain a tuned model; use the tuned model to predict performance parameters of the engine; and modify the physical design or operation of the engine based on the performance parameters.

[0145] A system according to any preceding clause, wherein the at least one processor is further configured to cause an engine control system to adjust at least one control parameter of the engine based on the engine control setting, or to cause a device to manufacture the engine component according to the design parameter based on the performance parameter.

[0146] The system of any preceding clause, wherein the parallel Bayesian inference simulation comprises a parallel Markov Chain Monte Carlo (MCMC) simulation.

[0147] The system of any preceding clause, wherein the at least one processor is further configured to use the tuning parameters in the model to generate a calibrated prediction of the operating parameter.

[0148] A system as described in any preceding clause, wherein the at least one processor is further configured to generate a difference model that provides a difference between the calibrated predicted value and the observed value of the operating parameter.

[0149] The system of any preceding clause, wherein identifying tuning parameters of the model comprises generating a probability distribution over the tuning parameters, the probability distribution providing a probability that a particular value is a suitable candidate for the tuning parameter.

[0150] A system as claimed in any preceding clause, wherein the test data and the model data are transient time series data of the operating parameters of the engine.

[0151] A system as in any preceding clause, further comprising a testing device in communication with the at least one processor, wherein the test data is acquired by the testing device.

[0152] A system according to any preceding clause, wherein the at least one processor is further configured to perform at least one of: causing an engine control system to adjust at least one control parameter of the engine based on the engine control setting; or causing a device to manufacture the engine component according to the design parameter.

[0153] A system as described in any preceding clause, wherein the at least one processor is further configured to compare the performance parameters of the engine to defined performance criteria to thereby certify the engine.

[0154] A method comprising: receiving test data for an operating parameter of an engine from an operational test of the engine from a test rig, the test data comprising observed values ​​of the operating parameter from the operational test; receiving model data for the operating parameter from a simulation performed via a model of the engine, the model data comprising predicted values ​​of the operating parameter from the simulation; compressing the test data and the model data to generate compressed test data and compressed model data; fusing the compressed test data with the compressed model data to generate fused data; generating a first distribution for a first tuning parameter of the model by performing parallel Bayesian inference simulations using the fused data, the first distribution providing a probability that a particular value is a suitable candidate for the first tuning parameter; generating a second distribution for a second tuning parameter of the model by performing parallel Bayesian inference simulations using the fused data, the second distribution providing a probability that a particular value is a suitable candidate for the second tuning parameter; generating a tuning model by using the value of the first tuning parameter and the value of the second tuning parameter in the model; and modifying a physical design or operation of the engine based on the performance parameter.

[0155] A method as in any preceding clause, further comprising using the tuning model to determine at least one of design parameters of engine components, engine control settings, or compliance with engine certification standards.

[0156] A method according to any preceding clause, further comprising at least one of: causing an engine control system to adjust at least one control parameter of the engine based on the engine control setting; or causing an apparatus to manufacture the engine component according to the design parameter.

[0157] A method as in any preceding clause, further comprising: selecting a value for the first tuning parameter based on the first distribution; and selecting a value for the second tuning parameter based on the second distribution.

[0158] The method according to any of the preceding clauses further includes: generating multiple possible values ​​of the first tuning parameter based on the first distribution; generating multiple possible values ​​of the second tuning parameter based on the second distribution; generating multiple combinations of values ​​of the first tuning parameter and the second tuning parameter based on the possible values; determining the likelihood that each of the multiple combinations will align the model data with the test data; and selecting a specific combination of values ​​that results in the highest likelihood of aligning the model data with the test data.

[0159] According to a method according to any preceding clause, the test data includes data about multiple operating parameters of the engine, and wherein the model data includes predicted values ​​of the multiple operating parameters, the method includes: selecting the specific combination of values ​​to align the multiple operating parameters of the model data and the test data.

[0160] A method comprising: receiving test data associated with a system; receiving model data of the operating parameters from a simulation performed via a model; compressing the test data and the model data to generate compressed test data and compressed model data; fusing the compressed test data with the compressed model data to generate fused data; and performing a parallel Bayesian inference simulation using the fused data to identify at least one value of a tuning parameter of the model.

[0161] A method as in any preceding clause, wherein the system is a real-world system, and wherein the test data is acquired via a test device associated with the real-world system.

[0162] A system for calibrating a model, the system comprising: at least one processor; and a memory device storing instructions which, when executed by the at least one processor, cause the at least one processor to perform operations, wherein the at least one processor is configured to: receive test data of operating parameters of a system via a test device in communication with the at least one processor; receive model data from a simulation performed via a model of the system; compress the test data and the model data to generate compressed test data and compressed model data; fuse the compressed test data with the compressed model data to generate fused data; and perform a parallel Bayesian inference simulation using the fused data to identify at least one value of a tuning parameter of the model.

[0163] A system as described in any preceding clause, wherein the system is a real-world system, and wherein the test data is acquired via a test device associated with the real-world system.

[0164] A method comprising: receiving test data of a system from a test device; receiving model data from a simulation performed via a model of the system, the model data including predicted values ​​of the operating parameters from the simulation; compressing the test data and the model data to generate compressed test data and compressed model data; fusing the compressed test data with the compressed model data to generate fused data; generating a first distribution of a first tuning parameter of the model by performing parallel Bayesian inference simulation using the fused data, the first distribution providing a probability that a specific value is a suitable candidate for the first tuning parameter; and generating a second distribution of a second tuning parameter of the model by performing parallel Bayesian inference simulation using the fused data, the second distribution providing a probability that a specific value is a suitable candidate for the second tuning parameter.

[0165] A method as in any preceding clause, wherein the system is a real-world system, and wherein the test data is acquired via a test device associated with the real-world system.

[0166] A method comprising: receiving test data of operating parameters of a real-world system from a test of the real-world system from a test device, the test data including observed values ​​of the operating parameters from the operational test; receiving model data of the operating parameters from a simulation performed via a model of the real-world system, the model data including predicted values ​​of the operating parameters from the simulation; compressing the test data and the model data to generate compressed test data and compressed model data; fusing the compressed test data with the compressed model data to generate fused data; performing parallel Bayesian inference simulation using the fused data to identify tuning parameter values ​​of the model; using the tuning parameter values ​​in the model to obtain a tuned model; and using the tuned model to determine at least one of a performance parameter of the real-world system, a design parameter of a component of the real-world system, or a control setting of the real-world system.

[0167] A method as in any preceding clause, further comprising adjusting at least one control parameter of the real-world system based on the control setting.

[0168] A method according to any preceding clause, further comprising causing an apparatus to manufacture the engine component in accordance with the design parameters.

[0169] A method comprising: receiving test data of an operating parameter of a real-world system from an operational test of the real-world system from a test device, the test data comprising observed values ​​of the operating parameter from the operational test; receiving model data of the operating parameter from a simulation performed via a model of the real-world system, the model data comprising predicted values ​​of the operating parameter from the simulation; compressing the test data and the model data to generate compressed test data and compressed model data; fusing the compressed test data with the compressed model data to generate fused data; generating a first distribution of a first tuning parameter of the model by performing a parallel Bayesian inference simulation using the fused data, the first distribution providing a probability that a particular value is a suitable candidate for the first tuning parameter, and selecting a value of the first tuning parameter based on the first distribution; generating a second distribution of a second tuning parameter of the model by performing parallel Bayesian inference simulation using the fused data, the second distribution providing a probability that a particular value is a suitable candidate for the second tuning parameter, and selecting a value of the second tuning parameter based on the second distribution; generating a tuning model by using the value of the first tuning parameter and the value of the second tuning parameter in the model; and using the tuning model to determine at least one of a performance parameter of the real-world system, a design parameter of a component of the real-world system, or a control setting of the real-world system.

[0170] A method as in any preceding clause, further comprising adjusting at least one control parameter of the real-world system based on the control setting.

[0171] A method as in any preceding clause, further comprising causing an apparatus to manufacture the component in accordance with the design parameters.

[0172] A system for calibrating a model of a real-world system, the system comprising: at least one processor; and a memory device storing instructions, the instructions, when executed by the at least one processor, causing the at least one processor to operate, the at least one processor being configured to: receive test data of operating parameters of the real-world system from an operational test of the real-world system via a test device in communication with the at least one processor, the test data comprising observed values ​​of the operating parameters from the operational test; receive model data of the operating parameters from a simulation performed via the model, the model data comprising predicted values ​​of the operating parameters from the simulation; compress the test data and the model data to generate compressed test data and compressed model data; fuse the compressed test data with the compressed model data to generate fused data; perform parallel Bayesian inference simulation using the fused data to identify tuning parameter values ​​of the model; use the tuning parameter values ​​in the model to obtain a tuned model; and use the tuned model to predict performance parameters of the real-world system.

[0173] The system of any preceding clause, wherein the at least one processor is further configured to adjust at least one control parameter of the real-world system based on the performance parameter.

[0174] A system as recited in any preceding clause, wherein the at least one processor is further configured to determine at least one design parameter of a component based on the performance parameter and cause an apparatus to manufacture the engine component in accordance with the design parameter.

Claims

1. A method, characterized in that include: receiving test data of operating parameters of the engine from a test device; receiving model data for the operating parameters from a simulation performed via a model of the engine; compressing the test data and the model data to generate compressed test data and compressed model data; fusing the compressed test data with the compressed model data to generate fused data; performing parallel Bayesian inference simulations using the fused data to identify tuning parameter values ​​for the model; using the tuning parameter values ​​in the model to obtain a tuned model; predicting a performance parameter of the engine using the tuning model; as well as Modifications are made to the physical design or operation of the engine based on the performance parameters.

2. The method according to claim 1, characterized in that in, The modification is an adjustment to the design parameters of the engine.

3. The method according to claim 2, characterized in that Further comprising determining design parameters of the engine component based on the performance parameters; and causing a manufacturing device to form the engine component according to the design parameters.

4. The method according to claim 1, characterized in that: in, The modification is an adjustment to the control settings of the engine.

5. The method according to claim 4, characterized in that Further comprising determining updated engine control settings for the engine based on the performance parameters; and causing an engine control system to adjust operation of the engine based on the updated engine control settings.

6. The method according to claim 1, characterized in that Further included determining compliance with engine certification standards based on the performance parameters by comparing the performance parameters to defined performance standards.

7. The method according to claim 1, characterized in that in, The test data includes data on a plurality of operating parameters of the engine, wherein the model data includes predicted values ​​of the plurality of operating parameters, the method comprising: A parallel Bayesian inference simulation is performed using the fused data to identify the tuning parameter values ​​that align the model data and the test data.

8. The method according to claim 1, characterized in that in, The method further comprises: generating a calibrated prediction of the operating parameter using the tuning model; and A difference model is generated that provides differences between the calibrated predicted values ​​and observed values ​​of the operating parameters.

9. The method according to claim 1, characterized in that: in, Identifying tuning parameters for the model includes generating a probability distribution for the tuning parameters, the probability distribution providing a probability that a particular value is a suitable candidate for the tuning parameter.

10. The method according to claim 1, characterized in that in, The test device includes a sensor that determines an observed value of the operating parameter of the engine.

Images