Computer-implemented method for optimizing operation of drive train of wind turbine

Abstract

A computer-implemented method for training a classifier for failure diagnosis of a driveline of a wind turbine, comprising: training a physics-based model (100) of the driveline of the wind turbine with operational data (151, 152) of the driveline of the wind turbine under normal conditions, thereby providing a normal hybrid model (200) of the drive train of the wind turbine; modeling at least one abnormal condition in a normal hybrid model (200) and training it with operational data (151, 152 ') of the driveline in a failure condition, thereby providing a fail hybrid model (500) of the driveline of the wind turbine; applying a set of input data (151) of the driveline in the normal condition to the normal hybrid model (200), thereby obtaining a set of combined data (280) in the normal condition for a particular parameter of the driveline; applying a set of combined input data (710) associated with at least one abnormal condition of the driveline and a set of failure data (150 '') to the fault hybrid model (500), thereby generating a set of combined failure data (750) for the at least one abnormal condition; obtaining a set of condition indicators (8105) for the at least one abnormal condition based on a set of output data (152, 152 ') of the driveline, the obtained set of composition data (280) in the normal condition, and the obtained set of composition failure data (750) for the at least one abnormal condition; a classifier (900) is trained with a set of condition indicators (8105) for at least one anomalous condition. A computer-implemented method for optimizing the operation of a drive train of a wind turbine comprises using a classifier (900) as previously trained.

Description

Technical Field

[0001] The present invention relates to the field of drive systems in wind turbines. In particular, the present invention relates to computer modeling and digitalization-based methods and systems for improving efficiency in operation of a power conversion system of a wind turbine and / or improving maintenance of the power conversion system. The present invention is particularly applicable to the field of wind energy. Background Art

[0002] Wind energy conversion is one of the most promising and reliable energy technologies available today. Europe has already installed 220 GW of wind capacity and plans to install an additional 105 GW over the next five years (Wind Energy in Europe: 2020 Statistics and the outlook for 2021-2025. WindEurope, https: / / windeurope.org / intelligence-platform / product / wind-energy-in-europe-in-2020-trends-and-statistics / ). The players involved in this energy are constantly researching this technology with the goal of achieving the best levelized cost of energy (LCOE). According to WindEurope, the costs of operations and maintenance (O&M) account for 25% to 35% of the LCOE of a wind turbine (Wind Energy Digitalisation to 2030. Reduced costs, better performance, safer operation. Published November 2021 (https: / / windeurope.org / intelligence-platform / product / wind-energy-digitalisation-towards-2030 / )), and the costs of corrective maintenance are 30% to 60% of O&M costs (Hansen AD, Wind Energy Engineering: A Handbookfor Onshore and Offshore Wind Turbines, Chapter 8 - Wind Turbine Technologies, 2017, pp. 145-160. doi:10.1016 / B978-0-12-809451-8.00008-4). Today, the possibilities offered by digitalization and artificial intelligence (AI) can significantly contribute to increasing energy production at wind farms, reducing unplanned outages, optimizing O&M work, and extending the service life of components.

[0003] Wind turbine systems can be classified according to the type of generator, gearbox and power converter used. Doubly fed induction generator (DFIG) with multi-stage gearbox and partial ratio converter is a widely used technology (Hansen AD, Iov F, Blaabjerg F, Hansen LH Review of Contemporary Wind Turbine Concepts and Their Market Penetration. Wind Engineering. 2004; Vol. 28(3): pp. 247-263. doi:10.1260 / 0309524041590099). In the DFIG topology (S. Muller, M. Deicke, and R. W. DeDoncker, “Doubly fed induction generator systems for wind turbines,” IEEE Industry Applications Magazine, Vol. 8, No. 3, pp. 26-33, May-June 2002, doi: 10.1109 / 2943.999610), the stator windings are directly connected to the constant frequency grid, while the rotor windings are connected to the grid through a pulse width modulation (PWM) power converter using a set of slip rings. The power converter can control the rotor circuit current, frequency, and phase angle offset (F. Blaabjerg, M. Liserre, and K. Ma, “Power electronics converters for wind turbine systems,” 2011 IEEE Energy Conversion Congress and Exposition, 2011, pp. 281-290, doi: 10.1109 / ECCE.2011.6063781). Such induction generators can operate within a wide slip range, typically ±30% of synchronous speed. As a result, they offer many advantages, such as high energy yield, reduction of mechanical stress and power fluctuations, and controllability of reactive power. The disadvantage of DFIG is the inevitable need for slip rings.

[0004] Wind turbines are also equipped with a control system that will ensure proper operation of the wind turbine in all operating conditions and keep the wind turbine within its normal operating range. Wind turbines contain electrical, mechanical, hydraulic or pneumatic systems and require sensors to sense the variables that will determine the required control actions. The most common variables monitored in the control system are wind speed, rotor speed, active and reactive power, voltage and frequency at the connection point of the wind turbine. In addition, the control system must be able to stop the wind turbine if necessary. One of the control strategies is pitch angle control (MK Dhar, M. Thasfiquzzaman, RK Dhar, M. T. Ahmed, and A. A. Ohsin, "Study on pitchangle control of a variable speed wind turbine using different control strategies," 2017 IEEE International Conference on Power, Control, Signals and Instrumentation Engineering (ICPCSI), 2017, pp. 285-290, doi: 10.1109 / ICPCSI.2017.8392258), which has proven to be an attractive option for variable speed operation and wind turbines larger than 1 MW. Using this control, the blades can be quickly moved away from the wind or into the wind when the power output becomes too high or too low, respectively. The pitch system can be based on a hydraulic system, controlled by a computer system, or an electronically controlled electric motor. The pitch control system must be able to adjust the pitch angle by a fraction of a degree at a time, corresponding to changes in wind speed, to maintain a constant power output.

[0005] The critical failure modes of wind turbine drive train systems, and in particular the generator and power conversion systems, have been analyzed (K. Alewine and W. Chen, "A review of electrical winding failures in windturbine generators," IEEE Electrical Insulation Magazine, Vol. 28, No. 4, pp. 8-13, July-August 2012, doi: 10.1109 / MEI.2012.6232004). Based on the sources of failure of generators (Tavner, PJ & Ran, Li & Penman, Jim & Sedding, Howard (2008) Condition Monitoring of Rotating Electrical Machines. Bibliovault OAI Repository, University of Chicago Press, doi: 10.1049 / PBPO056E), the types of failures are as follows: thermal failures, due to the effects of currents circulating through the windings and overcurrents on the insulation, and taking into account the maximum temperature allowed according to the insulation type and operating conditions; electrical failures, due to the voltage applied to the conductors under normal operating conditions and under abnormal conditions such as surges from the converter; environmental failures, due to the effects of environmental conditions and the degradation of the insulating polymers, corrosion phenomena, etc. that they may cause; mechanical failures, mainly due to vibrations; and thermomechanical failures, due to cyclic operating conditions with sudden or continuous changes in temperature, which have different effects on each material that is part of the cable and its accessories (insulation, shielding, ...). Generators and power converters are the components that have a greater impact on the reliability, failure rate and unavailability of wind turbines. Their failure rate is 15% per year for generators and 6.8% in the case of power converters in offshore wind farms (E.Byon, L.Ntaimo, C.Singh and Y.Ding, "Wind energy facility reliability and maintenance", Handbook of Wind Power Systems, pp. 639-672; Shafiee y Dinmohammadi-An FMEA-Based Risk Assessment Approach for Wind Turbine Systems: A Comparative Study of Onshore and Offshore, 2014). These components are equipped with sensors (temperature, vibration, electrical measurements...) and are connected to wind turbine supervisory control and data acquisition (SCADA) and condition-based monitoring (CBM) systems.Therefore, for each turbine in the wind farm, there is a long historical real-world operational data set. Sometimes, this data set includes recorded anomalies or failures in the operation of the turbine.

[0006] Computer modeling and digitalization are increasingly important in the wind energy sector as they provide tools to improve the design and performance of wind turbines to reduce both capital and operating costs. The large-scale introduction of sensors and improvements in big data processing capabilities over the past decade have allowed for efficient collection and analysis of data, enabling the development and use of digital models of wind turbines, sometimes referred to as digital twins. Data-driven models apply AI techniques to extract knowledge from real-world measurements, which can help identify meaningful patterns in large amounts of data collected from complex systems. In the field of wind energy generation, there are several approaches to this type of model. For example, spectral analysis of measured currents has been used for health monitoring of main bearings and stator and rotor windings of wind turbines (T. Gerber, N. Martin, and C. Mailhes, “Time-Frequency Tracking of Spectral Structures Estimated by a Data-Driven Method,” IEEE Transactions on Industrial Electronics, vol. 62, no. 10, pp. 6616–6626, October 2015, doi: 10.1109 / TIE.2015.2458781). In S. Yin, W. Guang and HR Karimi, “Data-driven design of robust fault detection system for wind turbines”, Mechatronics, Vol. 24, No. 4, pp. 298-306, December 2014, a data-driven model was directly constructed with the aim of detecting and isolating sensor and actuator faults in wind turbines, while E. Alizadeh, N. Meskin and K. Khorasani’s study Anegative selection immune system inspired methodology for fault diagnosis of wind turbines. IEEE Trans Cybern, pp. (99): 1-15, 2016, developed a hierarchical negative selection algorithm (NSA) library to detect and isolate common faults in wind turbines.The study by M. Li, D. Yu, Z. Chen, K. Xiahou, T. Ji, and Q. H. Wu, “A Data-Driven Residual-Based Method for Fault Diagnosis and Isolation in Wind Turbines,” IEEE Transactions on Sustainable Energy, Vol. 10, No. 2, pp. 895–904, April 2019, doi:10.1109 / TSTE.2018.2853990, uses a data-driven fault diagnosis and isolation (FDI) approach for wind turbines that implements a long short-term memory (LSTM) network for the residual generator and applies a random forest algorithm to make decisions. The main drawback of data-driven FDI methods is that they are designed based on experimental and historical data acquired from the system under normal and fault conditions. Such a process requires the development of a good database that includes labeled anomaly / failure data. In addition, the amount of anomaly / failure data is inherently scarce, resulting in an unbalanced dataset. Therefore, the accuracy of data-driven methods is generally poor for cases that are not included in the training dataset. Additionally, black-box models, such as deep learning models, exhibit low interpretability, making it difficult for domain experts to interpret the results and gain the required trust to make decisions based on the model's output.

[0007] Therefore, there is a need to develop new methods and systems for supervising and / or optimizing operation and / or maintenance of wind turbines and / or predicting failures therein that overcome the shortcomings of conventional methods and systems. Summary of the invention

[0008] The present invention overcomes the disadvantages of conventional methods and systems for condition monitoring of wind turbines by providing a computer-implemented method and system for optimizing the operation and / or maintenance of a drive train of a wind turbine and improving its maintenance workflow. In the context of the present invention, a drive train is understood as a set of components involved in wind energy conversion. The optimization is achieved using information obtained from a classifier used by the proposed method.

[0009] The method is based on different models of the transmission system, including its physics-based model, which is trained with data analysis techniques using real operating data. The physics-based model of the transmission system is a mathematical model that represents physical phenomena. The physics-based model involves a number of model parameters. It describes the relationship between physical quantities that characterize the transformation of wind speed energy into mechanical power and the transformation of mechanical power into electrical power. The physics-based model is trained with data analysis techniques using historical and real operating data, generating a so-called normal hybrid model.

[0010] Training a physics-based model with a data-driven algorithm using operational data under normal conditions enables calibration of the values ​​of at least some of the model parameters defining the physics-based model, thereby producing a trained physics-based model referred to as a normal hybrid model. In other words, when a physics-based model is trained with normal data (data representing normal operation), the model is a normal hybrid model. In the context of the present invention, the term "hybrid" refers to the combination of a mathematical model involving physical parameters and a data analysis involving real data (measurements). The selected values ​​of those parameters are obtained by minimizing the difference between the predictions obtained by the physics-based model using a data-driven optimization algorithm trained with real operational data inputs and their corresponding known real operational data outputs.

[0011] Once the normal hybrid model is constructed, it can be extended or adjusted to represent failure conditions, thereby generating a so-called fault model. The resulting new model is also trained using historical and final actual operating data for both normal and failure operations. This is done using real failure operating data inputs, which are fed to the failure model. In other words, when the normal hybrid model is adjusted to represent failures and trained with failure data (data representing failure operations), the normal hybrid model becomes a fault hybrid model. Feeding failure data to the fault model enables calibration of the values ​​of the fault model parameters that define the fault model. The selected values ​​of these fault model parameters are obtained by minimizing the difference between the prediction obtained by the fault model using the failure operating data input and the corresponding known real operating data failure output. Therefore, a so-called fault hybrid model of the power conversion system (transmission system) of a wind turbine is obtained, which takes into account both the data of the transmission system under normal operation and failure operation.

[0012] A digital twin can be defined as a virtual representation of a real system (in this case, the drive train of a wind turbine) that behaves the same as the real system. Therefore, the digital twin of the drive train consists of a physics-based model, a fault model, and its connection to the real system. This connection refers to the information that is transferred from the wind turbine (automatically or manually) to the digital twin and the information that can potentially be transferred from the digital twin to the system and the operator. Digital twins can be used for different applications:

[0013] On the one hand, the fault hybrid model can be used to generate synthetic data of failure scenarios, and even of events that have never occurred before. This allows to acquire knowledge of the behavior of the transmission system under certain conditions, which would not be possible to acquire otherwise. Thus, the digital twin enables the simulation of real scenarios that are difficult or costly to create in the real system. These scenarios can be used for prescriptive analysis of new operating conditions, or for testing extreme conditions and responses to anomalies or failures. The synthetic data is generated using real failure operating data and the fault hybrid model.

[0014] On the other hand, digital twins allow to enhance existing data-driven models (e.g., machine learning classifiers) for estimating the state of the transmission system and assisting the decision-making process during its life cycle (e.g., diagnosing abnormal operation of the transmission system or measuring anomalies in sensors and triggering alarms). In fact, the synthetic data generated by the digital twin (in particular by the fault hybrid model) can be used to enhance the training data set and improve the generalization capabilities of existing data-driven models (e.g., classifiers). Additionally, the digital twin can be used to create new features (also called condition indicators), which can be used as new inputs in classifiers, for example. These features are calculated by comparing real operations (e.g., SCADA data) and / or synthetic data with normal data (predictions) generated by a normal hybrid model. In fact, the normal data generated by the normal hybrid model is used as a baseline to compare the real asset in normal conditions (no faults) with each digital replica under the same operating conditions (e.g., wind speed, pitch angle, torque, cabin temperature).

[0015] In summary, the proposed method combines a physics-based model and a data-driven model for the drive train of a wind turbine to represent specific operating scenarios under both normal and abnormal or failure conditions. By enhancing the physics-based model with data-driven techniques, the physics-based model is optimized for the specific design and operating characteristics of the wind turbine, so that it can be used to generate synthetic data, classify failures and predict anomalies, and optimize the operation and maintenance of the power conversion system (drive train) of the wind turbine.

[0016] The development of data-driven algorithms for diagnosing normal or failed conditions is a complex task that involves: i) defining condition indicators, ii) labeling correct and failed operating data, iii) conceptualization of a classification model (supervised or unsupervised), iv) validation of the model and analysis of its accuracy (e.g., number of false positives or false negatives), and v) analysis of whether it is representative of one machine or a group of machines (e.g., whether the model is valid only for a single wind turbine or for a group of different wind turbines). The proposed methods and models can add value in this work by providing additional synthetic data to enhance the dataset used later by the classification algorithm.

[0017] In a first aspect of the present invention, a computer-implemented method for training a classifier for failure diagnosis of a transmission system of a wind turbine is provided. The method comprises: training a physics-based model of the transmission system of the wind turbine using operation data of the transmission system of the wind turbine under normal conditions, thereby providing a normal hybrid model of the transmission system of the wind turbine; modeling at least one abnormal situation in the normal hybrid model and training it using operation data of the transmission system under failure conditions, thereby providing a fault hybrid model of the transmission system of the wind turbine; applying a set of input data of the transmission system under normal conditions to the normal hybrid model, thereby obtaining a set of synthetic data under normal conditions for specific parameters of the transmission system; applying a set of synthetic input data and a set of failure data associated with at least one abnormal situation of the transmission system to the fault hybrid model, thereby generating a set of synthetic failure data for at least one abnormal situation; obtaining a set of condition indicators for at least one abnormal situation based on a set of output data of the transmission system, the obtained set of synthetic data under normal conditions, and a set of synthetic failure data for at least one abnormal situation; training a classifier using the set of condition indicators for at least one abnormal situation.

[0018] In an embodiment of the invention, a physics-based model of a drive train of a wind turbine comprises a first module representing an aerodynamic conversion stage of the drive train and a second module representing an electromechanical conversion stage of the drive train.

[0019] In an embodiment of the present invention, a normal hybrid model is obtained as follows: providing a physics-based model of a transmission system of a wind turbine, the physics-based model being represented by a set of equations representing energy conversion in the transmission system of the wind turbine and a plurality of model parameters; feeding the physics-based model with operating data of the transmission system under normal conditions, the operating data comprising real operating data input and real operating data output; obtaining a prediction of at least one of the following parameters: generated electric power, phase current, phase voltage, and electromagnetic torque; selecting values ​​of at least some of the model parameters, wherein the values ​​minimize the difference between the prediction and the real operating data output corresponding to the real operating data input, thereby obtaining a physics-based model trained with the operating data.

[0020] In an embodiment of the present invention, a fault hybrid model is obtained as follows: at least one abnormal situation is modeled in a normal hybrid model; operating data of a transmission system under a failure condition is fed to a normal hybrid model including at least one modeled abnormal situation; a prediction of parameters indicating a fault situation is obtained; and values ​​of at least some of the model parameters of the normal hybrid model that is defined by modeling the at least one abnormal situation are selected, wherein the values ​​minimize the difference between the prediction and the actual operating data output corresponding to the actual operating data input, thereby obtaining a calibrated fault hybrid model.

[0021] In an embodiment of the invention, at least one abnormal situation is modeled with the aid of at least one failure model.

[0022] In an embodiment of the invention, at least one failure model is represented by a set of equations and a plurality of failure parameters representing failures in energy conversion of a drive train of a wind turbine.

[0023] In a second aspect of the invention, a computer-implemented method for optimizing the operation of a drive train of a wind turbine comprises using a classifier trained as in the first aspect of the invention.

[0024] In an embodiment of the present invention, the method includes: obtaining a status indicator based on normal synthetic data obtained from a normal mixed model when the normal mixed model is fed with actual operation input data and actual operation output data; providing the status indicator to a trained classifier; and outputting by the trained classifier an output indicating whether the status indicator corresponds to a failure condition or a non-failure condition.

[0025] In an embodiment of the invention, the method further comprises: triggering an alarm when the classifier determines that the transmission system is operating under an abnormal condition.

[0026] In an embodiment of the invention, the method further comprises actuating on a drive train of the wind turbine when the status indicator corresponds to a failure condition.

[0027] In a third aspect of the invention, there is provided an apparatus comprising processing means configured to perform the method according to the first aspect of the invention.

[0028] In a fourth aspect of the invention, there is provided an apparatus comprising processing means configured to perform the method according to the second aspect of the invention.

[0029] In a fifth aspect of the present invention, there is provided a computer program product comprising computer program instructions / codes for executing the method according to the first aspect of the present invention and / or according to the second aspect of the present invention.

[0030] In a sixth aspect of the present invention, a computer readable memory / medium storing program instructions / codes is provided, wherein the program instructions / codes are used to execute the method according to the first aspect of the present invention and / or the method according to the second aspect of the present invention.

[0031] The proposed methods and models address the limitations of current data-driven models described previously. On one hand, they enable the generation of synthetic data for anomaly / failure conditions and for events that have never occurred, which can be used to augment and balance existing datasets and improve the generalization capabilities of data-driven models. On the other hand, the hybrid model preserves physical design relationships, which helps the interpretability of the model and increases the trust of end users.

[0032] The proposed method and system have been compared and validated with operational data from a real wind farm, including flagged failure conditions. In particular, a use case for a drive train based on a doubly-fed induction generator (DFIG) and its corresponding partial ratio power converter is presented. However, the proposed method and system are applicable to other wind power conversion technologies (e.g. permanent magnet generators).

[0033] Other advantages and features of the present invention will become apparent from the following detailed description and will be particularly pointed out in the appended claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to complete the description and to provide a better understanding of the present invention, a set of drawings is provided. These drawings form an integral part of the description and illustrate embodiments of the present invention, which should not be interpreted as limiting the scope of the present invention, but only as examples of how the present invention can be implemented. The drawings include the following figures:

[0035] Figure 1 A physics-based model of a specific system according to an embodiment of the present invention and a process of training the physics-based model using real operational data are schematically illustrated.

[0036] Figure 2A A block diagram of a physics-based model including two modules of a power conversion drive train of a wind turbine according to an embodiment of the present invention is shown. Figure 2B Shown in more detail Figure 2A A physically based model. Figure 2C A block diagram showing a normal hybrid model of a power conversion drive train of a wind turbine according to an embodiment of the present invention is shown.

[0037] Figure 3A A block diagram of a fault hybrid model of a power conversion drive system of a wind turbine according to an embodiment of the invention is shown.The fault hybrid model is based on the normal hybrid model, wherein additional models or conditions representing abnormal or fault situations are appropriately calibrated. Figure 3B An example of a normal hybrid model with additional models, in particular a thermal model, representing abnormal or failure situations according to an embodiment of the invention is schematically shown. Figure 3C The fault model of the power conversion transmission system according to the embodiment of the present invention and the process of training the fault model using real operation data are schematically shown.

[0038] Figure 4A A block diagram of synthetic data generation using a fault mixture model according to an embodiment of the present invention is shown. Figure 4B A block diagram of synthetic data generation using multiple fault mixture models according to an embodiment of the present invention is shown.

[0039] Figure 5A A schematic block diagram of a proposed method for training a failure classifier according to an embodiment of the present invention is shown. Figure 5B A schematic block diagram of a proposed method for optimizing the operation of a drive train of a wind turbine using a trained classifier according to an embodiment of the present invention is shown.

[0040] Figure 6 A representation of a physics-based model of a wind turbine drive train according to an embodiment of the present invention in Matlab-Simulink (a visual simulation tool in Matrix Lab) is shown.

[0041] Figure 7 Shows Figure 6 Generated active power (kW) versus wind speed (m / s) for a wind turbine drive train modeled in Figure 1. Four different power curves are shown: real data (measured in a SCADA system); simulation using the model without calibration; simulation using the model after a first calibration (6 parameters); and simulation using the model after a second calibration (7 additional parameters).

[0042] Figure 8 The representation in Matlab-Simulink of a thermal model (thermal loop of the generator stator winding) added to the normal hybrid model of a wind turbine drive train is shown, which is used to model failures in the thermal loop of the stator winding of the drive train generator.

[0043] 9A to 9D Five marked abnormal situations (wind speed and active power signals) of overheating during real operation are shown.

[0044] FIG. 10A to FIG. 10E Generator stator winding temperature values ​​(real and simulated) during an overtemperature failure flag condition are shown.

[0045] Fig.11A method for synthetic data generation for a generator stator winding overheating scenario formed by a data-driven random failure model and a normal hybrid model according to an embodiment of the present invention is shown.

[0046] Fig.12 Shown are the stator winding temperature calculated by the fault hybrid model (in this case, the thermal model) from the synthetic input variables (non-bold line) and the stator winding temperature as measured by the SCADA system (bold line). DETAILED DESCRIPTION

[0047] The following description should not be taken in a limiting sense but is given merely for the purpose of describing the broad principles of the invention.The following embodiments of the invention will be described by way of example with reference to the above mentioned drawings showing apparatus and results according to the invention.

[0048] Next, different stages of a method for improving / optimizing the operation and / or maintenance of a drive train of a wind turbine are disclosed.Non-limiting examples of components of a drive train involved in wind energy conversion are: main bearings, pitch systems, gearboxes, generators, power converters, bearings and control systems.

[0049] The method is based on computer modeling, simulation and digitization.In a first stage of the proposed method, a physics-based model 100 of a power conversion drive train of a wind turbine is provided. The physics-based model 100 may be a newly created model or an already existing physics-based model. Figure 1 A physics-based model 100 of something (eg, a device or system such as a power drive train of a wind turbine) is generally disclosed. Figure 1 Also shown is an optimization process 10 for training of a physics-based model 100. The physics-based model 100 of the power transmission system of a wind turbine is represented by a set of equations representing the conversion of kinetic energy of wind into electrical energy and a plurality of model parameters. In other words, it is a mathematical model representing the physical phenomena of power conversion occurring in the transmission system, transforming kinetic energy into electrical energy. The physics-based model 100 is constructed taking into account the system design parameters. It describes the relationship between the physical quantities characterizing the conversion of kinetic energy in the wind into mechanical power and the conversion of mechanical power into electrical power.

[0050] Figure 2A A block diagram of a non-limiting example of a physics-based model 100 of a power conversion drive train of a wind turbine is illustrated in FIG. Figure 2BThe physics-based model is depicted in more detail in the example of , where the modeled drive train uses a doubly fed induction generator (DFIG) with back-to-back power converters. However, other wind energy conversion technologies such as a permanent magnet generator (PMG) with a full converter or a squirrel cage induction generator (SCIG), etc., may be used instead. The physics-based model 100 models a set of components involved in wind energy conversion such as main bearings, pitch systems, multipliers (gearboxes), generators, power converters and their control systems. In FIG. 2A to FIG. 2B In the physics-based model 100, the model is divided into two main modules 101, 102. These models can be used together or operated in an independent manner, depending on the available operational data. The first module 101 represents the conversion of kinetic energy in the wind into mechanical power. Figure 2B The wind hits the blades of the wind turbine and causes the rotor to rotate at a low speed in the main shaft. This low speed can be increased using a gearbox. The mechanical torque obtained after the gearbox in 101 feeds the second module 102. The second module 102 represents the electromechanical conversion stage of the transmission system. Figure 2B Schematically represented in FIG. 1 by a doubly fed induction generator (DFIG), a partial ratio power converter and a corresponding control system (not shown), which allows for optimal operation of the drive system. The generator and the corresponding power converter enable the conversion from mechanical torque to electrical power.

[0051] In the transformation performed in the first module, the mechanical power extracted from the kinetic energy in the wind is determined by parameters such as air density, rotor sweep, wind speed and power coefficient. Other parameters related to the main shaft of the transmission system are usually also involved in the transformation. Non-limiting examples of these parameters are wind turbine inertia constant, shaft spring constant, shaft mutual damping and pitch angle.

[0052] In the transformation performed in the second module, where mechanical power is converted into electrical power, the key elements involved are the generator and its corresponding power converter. When the modeled transmission system uses a DFIG, the electrical part of the machine can be modeled, for example, by a fourth-order state-space model, and the mechanical part can be modeled, for example, by a second-order system. It involves solving several equations from both the electrical and mechanical systems. Non-limiting examples of parameters involved in this transformation are: stator resistance and leakage inductance, magnetizing inductance, total stator inductance, q-axis stator voltage and current, d-axis stator voltage and current, stator q and d-axis fluxes, rotor angular velocity, rotor angular position, pole pair number, reference frame angular velocity, mechanical angular velocity, electrical angular velocity, mechanical rotor angular position, electrical rotor angular position, electromagnetic torque, shaft mechanical torque, rotor and load combined inertia coefficient, rotor and load combined inertia constant, rotor and load combined viscous friction coefficient, rotor total inductance, rotor resistance and leakage inductance, q-axis rotor voltage and current, d-axis rotor voltage and current, rotor q and d-axis fluxes, DC bus voltage regulator gain, grid side converter current regulator gain, rotor side converter current regulator gain.

[0053] Depending on the nature of the equipment, it may be difficult to obtain values ​​for the full set of design parameters. In this case, estimates are required, which may affect model performance. The nominal values ​​of DFIG values, such as rated power, voltage, current, frequency, number of pole pairs or angular speed of the rotor are usually known, but the specific values ​​corresponding to the design of the stator windings (stator resistance and leakage inductance, magnetizing inductance, total stator inductance) are rarely known.

[0054] The inputs 151 to the physics-based model 100 (the inputs 121, 122 to the first module 101) are real values ​​(measurements) under normal conditions, such as SCADA values. Figure 2BIn the model shown, these inputs are the real values ​​of the wind speed 121 (e.g., in meters per second, i.e., m / s) striking the turbine blades and the pitch angle 122 (e.g., in degrees, i.e., deg) of the blades, which are acquired at certain time intervals, e.g., periodically. In an embodiment of the invention, the modules 101, 102 may be modeled independently of each other. In this case, when the second module 102 is modeled independently, the real value of the normal condition of the mechanical torque 131 in the main shaft of the wind turbine may also be an input to the model 100. More precisely, although the data 151, 152 used to train the physics-based model 100 are generally considered to be data under normal conditions, in fact, some input data 151 may not be distinguished between failure and normal due to their nature. For example, this is the case with the input data corresponding to the wind speed 121, which therefore cannot be considered to represent a failure condition. Other input data 151, such as the mechanical torque 131, may actually represent a normal condition or a failure condition. Similarly, the output data 152, 141 may generally represent a normal condition or a failure condition. Now, the data 151, 152 used to train the physics-based model 100 is data under normal conditions. The data used to train the physics-based model 100 (in particular, certain input data 131 and output data 152, 141, since other input data 121, 122 are essentially data under normal conditions) have previously been labeled as "data under normal conditions" preferably by a human operator. Therefore, the physics-based model 100 is trained using data under normal conditions. The input values ​​151, 121, 122, 131 are discrete values ​​of historical data under normal conditions collected at multiple moments within a certain time range. Typically, the available data corresponds to SCADA data that is registered, for example, every 10 minutes. However, data with a higher acquisition frequency can also be obtained. The acquisition frequency can range from microseconds (1μs=10 -6 The acquisition frequency may vary from 100 μs or 1 ms to 10 seconds or minutes, such as 1 second, 1 minute or 10 minutes. The acquisition frequency may depend on different situations, such as the type of parameter being acquired.

[0055] Reference again Figure 2BIn an exemplary model, the output 131 of the first module 101 is a prediction of the generated mechanical torque acting on the shaft of the doubly fed induction generator (DFIG) (which belongs to the second module 102). The prediction is obtained by calculating the equations included in the model. The second module 102 takes the output 131 of the first module 101, that is, the mechanical torque acting on the shaft of the DFIG, as input. The output 141 of the second module 102 includes a prediction of the generated electric power and its related signals such as phase current, phase voltage, frequency and / or electromagnetic torque to be delivered to the power grid at certain time intervals. In a preferred embodiment, the output 141 is a prediction of the generated electric power and its corresponding phase current. The prediction is obtained by calculating the equations included in the model. The prediction is in Figure 1 It is called 180.

[0056] Thus, the real operating data under normal conditions 150 includes real inputs (discrete historical values ​​of real operating data, usually measurements) 151 to the physics-based model 100 under normal conditions (e.g., measured values ​​of certain parameters such as wind speed, pitch angle, or mechanical torque, such as Figure 2B The actual operation data 150 are discrete values ​​of historical data collected at multiple moments within a specific time range. Reference numeral 180 represents a prediction of a parameter, such as a prediction of the generated electric power ( Figure 2B predictions in 141).

[0057] Once the physics-based model 100 is established, it is optimized (trained) by applying one or more data-driven optimization algorithms 190 thereon using historical real operating data 150 under normal conditions so as to obtain Figure 1 and Figure 2C A normal mixture model 200 is shown, which is an enriched model with real operating data. Figure 1 The optimization / training process is illustrated in . Figure 1 A training process or algorithm 10 for calibration of a physics-based model 100 is shown. Figure 1, real operating data 150 under normal conditions are fed as real input 151 to the physics-based model 100, or provided as real output 152 to another stage of the calibration process. The calibration process comprises optimizing the design parameters (also referred to as model parameters) by means of an objective function to minimize the residual 160, the exact values ​​of the design parameters possibly being estimated taking into account certain physical constraints (e.g. within a given realistic interval). Preferably, among the design parameters to be optimized, there are parameters associated with the electromechanical conversion stage (e.g. but not limited to the generator, power converter and wind turbine control), parameters associated with the aerodynamic conversion, parameters associated with the control strategy, and parameters associated with the pitch system and the mechanical part of the transmission system. Non-limiting examples of parameters associated with the electromechanical conversion stage that may be selected for optimization are: stator winding resistance, rotor winding resistance, generator inertia constant and generator friction coefficient, power converter grid-side coupling resistance, power converter grid-side coupling inductance, converter line filter capacitor. Non-limiting examples of parameters associated with the aerodynamic conversion stage that may be selected for optimization are: wind turbine inertia constant, shaft mutual damping, and shaft spring constant. Non-limiting examples of parameters associated with the control strategy that may be selected for optimization are: DC bus voltage regulator gain, speed regulator gain, and wind speed at nominal speed and Cp max.

[0058] Therefore, the training process 10 allows estimating the values ​​of certain design variables (parameters) of the physical model 100, which are a priori unknown, from a set of historical operating data (data set) comprising a sufficient representation of the system behavior under normal conditions 151, 152, 121, 122, 131, so as to minimize the difference between the output values ​​(predictions) 141, 180 of the variables of the physical model 100 and the true system values ​​(outputs) 152. Figure 1The training process 10 shown aims to minimize one or more cost functions (or objective functions). In a particular embodiment, there are two cost functions, the first cost function is the difference between the active power estimated (predicted) by the physics-based model 100 and the real active power (generally represented by reference numeral 152) typically provided by SCADA, and the second cost function is the difference between the phase current estimated (predicted) by the physics-based model 100 and the real phase current (generally represented by reference numeral 152) typically provided by SCADA, but there may be a different number of cost functions. In the example, the first cost function aims to minimize the difference in power, and the second cost function aims to minimize the difference in phase current. The objective function of the process is the minimization of the residual 160, which is defined as the difference between the physics-based model output (e.g., prediction of power) 180 and the real operating data 152 (e.g., output power) for a given real input 151 (e.g., wind speed or torque). The optimization algorithm aims to find a combination of parameter values ​​that minimizes the difference between the output (prediction 180) of the physics-based model 100 and the measured data 152.

[0059] In an embodiment of the present invention, the optimization algorithm used by the training process for calibrating (tuning or adjusting) parameters is a global solver that searches for a cost function min with multivariate input variables x x The global minimum of f(x) is obtained for a multivariate input variable x subject to linear and nonlinear constraints and some finite bounds. In a specific example, the objective function is defined as the mean absolute percentage error (MAPE) between the estimated parameter (prediction) and the true measured value of the parameter (e.g., the parameter is active power or phase current).

[0060] In other words, the training process is performed as follows: using real operating data 150 under normal conditions (real input data 151), a first prediction 180 of one or more parameters (e.g., the generated active power and / or at least one related signal or parameter (phase current, phase voltage and / or electromagnetic torque) is performed. A residual 160 is obtained by comparing the real output 152 (operating data under normal conditions, such as output power and / or a measured value of at least one related signal or parameter provided by the drive system of the wind turbine for a given real input 151) with the obtained prediction 180. Based on the residual 160, one or more model parameters defining the physics-based model 100 are modified (stage 170). The training process is repeated n times, such as Figure 1, schematically represented in the lower part of . After n iterations, the physical model 100 is considered to be a normal hybrid model 200. Typically, the training process continues for a given number of iterations, or until a point is found where the objective function is below a threshold. Therefore, the normal hybrid model 200 consists of a physics-based model enriched with operational data (e.g., SCADA data) under normal scenario conditions. In other words, the normal hybrid model is the resulting calibrated physical model 200, wherein the term hybrid refers to the need for both a physics-based model 100 and a data-driven optimization algorithm 190 trained with real data in its construction.

[0061] Once the normal hybrid model 200 is developed, it can be extended or adjusted to include abnormal or failure conditions to study the behavior of the system (the drive train of the wind turbine) under abnormal or failure conditions. Figure 3A . The normal hybrid model 200 together with one or more additional models 301 to 306 representing possible fault conditions form an extended model or fault model 400. The additional models 301 to 306 (also referred to as failure models) may represent a variety of models representing different failures of the wind turbine drive system, including but not limited to aerodynamic and / or mechanical failures in the dynamic conversion stage 101 of the wind turbine drive system and / or magnetic, electrical, mechanical and / or thermal failures in the electromechanical conversion stage 102, such as overheating or short circuits in stator or rotor windings, demagnetization of permanent magnets, mechanical failures in shafts or bearings, etc. The failure models 301 to 306 are represented by a set of equations and a plurality of model parameters representing a fault condition or anomaly. Thus, the failure model enables the creation or modeling of at least one anomaly or failure condition. The failure model may be generated with new separate additional physics-based modules 301 to 306 representing anomalies or failure conditions, or may represent a specific condition (usually a fault condition) integrated within the main normal hybrid model 200. For example, a separate block may be added to the normal hybrid model 200 to represent the failure model, or only certain parameters of the last normal hybrid model 200 may be modified. Alternatively, the failure model may be essentially integrated within the physics-based model 100 or the normal hybrid model 200, adding potential fault conditions to the physics-based model 100 or the normal hybrid model 200. Figure 3AExemplary additional models (failure models) shown in and implemented as separate blocks are: aerodynamic model 301 and mechanical model 302 associated with the dynamic conversion module 101 and electrical model 303, magnetic model 304, mechanical model 305 and thermal model 306 associated with the electromechanical conversion module 102. Non-limiting examples of other additional models that are essentially integrated within the normal hybrid model 200 and can be used to enhance the normal hybrid model 200 with other failure cases are: short circuit models in the stator or rotor (e.g. between coils, between phases or to ground), readings of specific sensors (readings indicative of failures), partial slot discharge models, core failure models, structural failure models, demagnetization of permanent magnets, etc.

[0062] The extended model 400 formed by the normal hybrid model 200 plus one or more additional models 301 to 306 representing possible fault conditions (failure models) may be referred to as a fault model 400 . Figure 3B An example of a fault model 400 formed by a normal hybrid model 200 and a thermal model 306 representing overheating of a DFIG stator winding is shown. The thermal model 306 is required to represent an overheating anomaly. In the example shown, the input to the fault model (thermal model) 306 is a prediction 306-2 of one or more parameters as predicted by the normal hybrid model 200 (e.g., a prediction of the stator phase current) ( Figure 1 The prediction 280 in DT) and the actual value 306-1 of one or more parameters (e.g., measured values ​​collected by the SCADA system), such as the actual value of the cabin temperature. The failure model 306 provides a prediction output 350, which in this exemplary case corresponds to the temperature of the DFIG stator winding. In this case, when the actual value 306-1 of the stator winding temperature for a specific current differs from the value estimated by the DT (prediction 350) by a certain amount, such as a certain amount higher, an abnormality in the cooling system of the DFIG can be identified.

[0063] Figure 3B The thermal model 306 shown in is an example of an additional model (failure model) that is independent of the normal hybrid model 200, since the thermal model 306 is only linked to the normal hybrid model 200 through the prediction 306-2. Depending on the failure scenario to be modeled, it may happen that the failure model is more relevant to the normal hybrid model 200, for example, essentially integrated within the normal hybrid model 200. For example, the resulting fault model 400 can be based on the normal hybrid model 200 modified to represent the failure scenario. In this case, there are no external failure models 301 to 306, but the failure scenario is internally integrated within the normal hybrid model 200, forming the extended model 400 (failure model).

[0064] Once the fault model 400 (normal hybrid model 200 together with a representation of one or more possible fault conditions) is established, the fault model 400 is optimized (trained) by applying one or more data-driven optimization algorithms 450 trained with real operating data in the fault conditions 150', so that a fault hybrid model 500 is obtained, such as Figure 3A and Figure 3C As shown, the fault hybrid model 500 is an enriched (calibrated) model with real fault data. Figure 3C The training process illustrated in FIG. 1 is similar to that applied to the physics-based model 100 to obtain the normal mixture model 200 ( Figure 1 ). The calibrated new model (also called the fault mixture model) is referred to as 500.

[0065] Calibration or training of the fault model 400 is performed as follows using real operating data under fault conditions 150'. The real operating data under fault conditions 150' include real inputs to the fault model 400 under normal conditions (discrete historical values ​​of real operating data, usually measurement results, which cannot be considered to refer to failure conditions due to their nature (e.g. wind speed)) 151 (e.g. measured values ​​of certain parameters such as wind speed, pitch angle or mechanical torque) and real outputs under failure conditions 152' of the transmission system of the wind turbine being modeled (e.g. measured outputs of certain parameters such as generated electrical power and phase current that correspond to or are associated with the real inputs 151 but are affected by the failure model). The actual operating data 150' are discrete values ​​of historical data (e.g. SCADA data) collected at multiple moments within a certain time range. The output data 152' used to train the fault model 400 has previously been preferably labeled as "data under failure conditions" by a human operator. Therefore, the fault model 400 is trained with data under failure conditions. Using real operating data 150' (real input data 151, such as the nacelle temperature or the real room temperature where the generator is installed), a first prediction 180' is made, such as a prediction 180' of the stator winding temperature. By comparing the real output under the failure condition 152' (operating data, such as, in the example, the measured value of the stator winding temperature provided by the transmission system of the wind turbine and corresponding to the given real input 151) with the obtained prediction 180', a residual 160' is obtained. According to this residual 160', at least some model parameters defining the extended model 400 are modified (stage 170'). The parameters to be modified can be parameters of the additional models 301 to 306, parameters of the normal hybrid model 200, or parameters of both (complete fault model 400), depending on the representation of the failure condition. For example, some model parameters defining and representing the failure can be optimized. The exact values ​​of these parameters have been previously estimated by imposing certain physical constraints (e.g., imposing values ​​within a given realistic interval range). Non-limiting examples of parameters to be optimized in failure models (e.g., in particular in thermal models) are material thermal conductivity, area perpendicular to the direction of heat flow, thickness of the layer through which conduction phenomena occur, heat transfer coefficient and surface area in contact with air. Figure 3C , reference numeral 155 denotes a prediction of one or more parameters as predicted by the normal mixture model 200 and provided to the additional models 301 to 306 ( Figure 3B 306-2 in (eg, prediction of stator phase current). The calibration process is repeated n' times, as in Figure 3C After n' iterations, the fault model 400 is considered to be a faulty mixed model 500. Typically, the calibration process continues for a given number of iterations or until a point is found where the objective function (cost function) is below a threshold.

[0066] The fault model 400 is trained by applying one or more data-driven optimization algorithms 450 trained using real operational data 150', such as Figure 3A The optimization algorithm allows estimating the values ​​of certain design variables (parameters) of the fault model 400, which are a priori unknown, from a set of historical operating data 150' comprising a sufficient representation of the system behavior, in order to minimize the difference between the output values ​​(predictions) 180' of the variables of the fault model 400 and the actual system values ​​(outputs) 152'. Figure 3C The optimization algorithm shown in aims to minimize one or more cost functions (or objective functions). The optimization algorithm aims to find a combination of parameter values ​​that minimizes the difference between the output (prediction) of the extended model 400 and the measured operational data 150'. In an embodiment of the present invention, the optimization algorithm 450 for calibrating (tuning or adjusting) parameters is a global solver that searches for a cost function min with multivariate input variables x. x The global minimum of f(x) is obtained for a multivariate input variable x subject to linear and nonlinear constraints and some finite bounds. In a specific example, the objective function is defined as the mean absolute percentage error (MAPE) between the estimated parameter (prediction) and the actual measured value of the parameter (e.g., the temperature of the stator winding).

[0067] The training process of the fault model 400 leading to the fault hybrid model 500 that has been described allows simulation of failure conditions present in the training operational data 150 ′ (that is, present in the set of historical data such as SCADA data that has been used to train the fault model 400 ).

[0068] So far, a fault hybrid model 500 that takes into account abnormal or failure conditions has been obtained. There may be one or more fault hybrid models 500 that include one or several failures. In particular, there may be as many fault hybrid models 500 as different failure conditions to be simulated. Therefore, the thermal model used as an example in this specification should not be considered in a limiting manner. The thermal model should also not be considered as the only possible failure model within the fault hybrid model 500. A single fault hybrid model 500 will represent (model) one failure condition or model. Alternatively, there may be as many fault hybrid models 500 as failure conditions to be modeled.

[0069] The fault hybrid model 500 has been calibrated to simulate failure conditions present in the training operational data (e.g., training SCADA data). However, this does not guarantee that these model executions are representative of the entire abnormal feature space. In fact, in real operational data (e.g., SCADA data 150, 150'), the frequency of abnormal conditions and failures is relatively low, and typically these instances are not annotated (labeled). Therefore, relying solely on deterministic models to generate synthetic failure scenarios will provide narrow data sample constraints for patterns that have been seen before.

[0070] To address this limitation, the proposed method combines a stage of synthetic data generation, e.g. FIG. 4A to FIG. 4B The process uses the previously obtained fault mixture model 500 to generate synthetic data 750. The synthetic data generation includes a random model 705 (also referred to as a random synthetic data generation model), which estimates the occurrence probability distribution of failure data and generates synthetic input data 710 associated with each fault accordingly. Figure 4A , a single fault condition or scenario labeled "Failure Mode 1" is shown. Therefore, two sets of failure data are fed to the fault hybrid model 500: real data 150" (e.g., measured historical data) representing or associated with the fault condition, and generated synthetic data 710 associated with the fault condition. The synthetic input data 710 and the real data 150" for each fault are used as inputs to the fault hybrid model 500 to generate synthetic failure data 750. The actual labeled failure data 150" is the failure data that has been identified and labeled ( Figure 3C The synthetic failure data 750 is generated using the fault mixture model 500 activated by the predetermined probability distribution 705 of the occurrence of the failure data 710. The occurrence probability distribution of the failure data is calculated based on the observation of the marked failure data 150".

[0071] exist Figure 4B In the process of synthetic data generation, Figure 4A It adds as many fault hybrid models 500 as needed to model the set of failure conditions (labeled as "failure mode 1", ..., "failure mode n") and the corresponding set of actual labeled failure data 150" (if any) and / or synthetic data 710 representing the corresponding failure conditions. Figure 4A, the synthetic data 710 is obtained from the stochastic model 705, which estimates the occurrence probability distribution of the failure data associated with the corresponding failure conditions ("failure mode 1", ..., "failure mode n") and accordingly generates synthetic input data associated with each fault 710. Each fault hybrid model 500 can generate an unlimited number of synthetic failure data 750 for a specific failure type ("failure mode 1", ..., "failure mode n") based on the real failure data 150" and using the stochastic model 705.

[0072] Figure 5A and Figure 5B A schematic block diagram of the proposed method for optimizing the operation of a drive train of a wind turbine is shown. Figure 5A The training of a failure classifier is schematically illustrated. Figure 5B Optimization of operation and / or maintenance of a drive train of a wind turbine using a trained classifier is schematically illustrated.

[0073] exist Figure 5A In FIG. 9 , different stages for training the classifier 900 are represented based on different models of the wind turbine drive train. In particular, both the normal mixture model 200 (trained under normal conditions) and the fault mixture model 500 (trained under fault conditions) are used (the fault mixture model 500 is not shown, but it is implicitly included in the stages 700, 700′ for synthetic data generation, which refer to FIG. FIG. 4A to FIG. 4B There may be different fault mixture models 500 to represent different failure conditions ( FIG. 4A to FIG. 4B Depending on the failure to be modeled, different failure scenarios can be included in the single fault hybrid model 500, such as Figure 3A The stages of normalization / feature creation 800 and training of a failure classifier 900 are also schematically represented. Figure 5B 9, the use of a trained failure classifier 900 to perform failure detection from actual real operating data of one or more transmission systems is schematically represented.

[0074] In the first stage, the Figure 1 The normal hybrid model 200 generated by the process shown in is fed with a set of real operating data (usually SCADA data) under normal conditions 151. The normal hybrid model 200 then provides a prediction 280 (a corresponding set of predictions). In other words, the prediction 280 is synthetic data associated with normal conditions. The prediction 280 is, for example, normal synthetic data representing the generated electrical power and at least one related signal or parameter, such as phase current.

[0075] In another phase, which may be concurrent with the first phase, there is a failure synthetic data generation process 700, 700'. One or more fault hybrid models 500 are used to generate failure synthetic data 750, as shown in FIG. FIG. 4A to FIG. 4B For example, but not limited to, if the fault model is the thermal model 306, the synthetic data 750 may be a collection of synthetic failure data representing stator winding temperature. The synthetic failure data 750 is obtained from Figure 5A A set of actual marked failure data 150" is shown, as well as synthetic data 710 (such as FIG. 4A to FIG. 4B Publicly available, but not Figure 5A obtained as shown in the schematic representation of ).

[0076] Thus far, a set of normal synthetic data 280 representing a broad range of normal conditions and a set of failure synthetic data 750 for each failure type being modeled have been generated (see, e.g., FIG. 3A to FIG. 3C ). Then, a new set of features, called condition indicators, is created for each type of failure. In the context of the present invention, a condition indicator is some feature by which an abnormal condition can be detected. Features such as difference, division, gradient, root mean square (RMS), peak, peak-to-peak, kurtosis, and crest factor are widely used for wind turbine fault diagnosis. The condition indicator 8105 is generated based on a labeled data set consisting of the real output data 152, 152' (which can correspond to both normal conditions or failure conditions), the enhanced normal synthetic data 280 generated using the method already described, and the enhanced failure synthetic data 750.

[0077] Both the normal hybrid model 200 and the fault hybrid model 500 are used as a baseline to create a condition indicator 8105 that can improve the accuracy of the failure classifier 900. The failure classifier 900 is a supervised classifier for failure diagnosis. In order to train the classifier 900, a stage of classifying a normal condition or a failure condition of the wind turbine is included (block 900). The classifier 900 is trained using the corresponding condition indicator 8105 generated previously (block 800). In other words, the classification is trained using the condition indicator 8105, and the value of the condition indicator 8105 comes from the relationship between the actual operational output data 152, 152', the prediction (synthetic data under normal conditions) 280 and the synthetic failure data 750.

[0078] At stage 800, condition indicators 8105 are calculated by establishing comparisons between the different data input to this stage (stored output measurements 152, 152' (e.g. SCADA data), normal synthetic data 280 and failure synthetic data 750). For example, depending on the type of failure represented by the corresponding data, the condition indicator 8105 can be, for example, the result of comparing the normal synthetic data 280 with the operational output data 152, 152'; or the result of comparing the synthetic failure data 750 with the operational output data 152, 152'; or the result of comparing the synthetic failure data 750 with the normal synthetic data 280; or the result of comparing the synthetic failure data 750 with the normal synthetic data 280 and with the operational output data 152, 152'. In addition, different types of comparison operations applied to the compared data can be established, such as difference, division, gradient, etc. Therefore, at stage 800, a set of condition indicators 8105 are obtained. They are condition indicators for each type of failure being modeled. For example, under ideal conditions (non-faulty system), the condition indicator 8105 is close to zero (minimal difference) when a comparison is made between real operating data and normal synthetic data, whereas the condition indicator 8105 will have a high value under a fault condition when the real operating data 152' under a fault condition is compared to the normal synthetic data 280. A fault type may have one or more associated condition indicators.

[0079] For example, the condition indicator 8105 can be the difference between the real current 152, 152' measured in the SCADA system and the current 280 simulated by the normal hybrid model 200. In a non-fault condition, the difference will be close to zero. If the difference exceeds a threshold, it indicates an anomaly that will be classified in the next stage of training the failure classifier 900. Another example of a condition indicator 8105 can be the difference between the real stator temperature and the estimated temperature value under the determined failure condition. In this case, a high value of the real stator winding temperature can be compared with its corresponding failure synthetic data. If the result of the comparison is within a specific threshold, it indicates an anomaly that will be classified in the next stage of training the failure classifier 900.

[0080] The aforementioned condition indicators 8105 can also enhance the generalization capabilities of the classifier 900, thereby allowing a single classifier to be trained for multiple turbines (e.g., all turbines) of an energy field. Typically, if only real data is used to train the classifier 900, it will be difficult to obtain a universal threshold for a wide collection of turbines. In fact, due to differences in the manufacturing process, different locations, etc., each turbine typically has different behaviors under the same operating conditions. The fact of having a single classifier 900 for multiple wind turbines provides many advantages. On the one hand, it facilitates governance and scalability in cloud / edge environments because it is easier to manage a single model than to manage a single model for each turbine. On the other hand, having a single classifier 900 allows us to train it with a wider range of normal and failure data under different conditions, making the model more robust.

[0081] The classification process is based on a classifier 900 developed using supervised machine learning techniques. Non-limiting examples of machine learning techniques on which the classifier 900 is based may be k-nearest neighbors (kNN), random forests (RF), support vector machines (SVM), extreme learning machines (ELM), etc. Due to the supervised and machine learning nature of the classifier 900, the classifier 900 needs to be pre-trained before it is further used for classification of normal or failure conditions. Training is performed using labeled data represented by condition indicators 8105, which are associated with different types of failures / anomalies and thus inform about possible anomalies. The classifier 900 is fed with a set of condition indicators 8105, each of which is associated with a label indicating "failure" or "no failure", and in the case of a "failure" label, another label representing the type of failure / anomaly.

[0082] So far, the training of the classifier 900 has been disclosed. Once the classifier 900 is trained, it is used to detect failures in the operation of wind turbines of a wind farm (stage 1000). Figure 5B , which schematically refers to the classifier 900 acting as a failure detector when fed with condition indicators created from actual real data of the wind turbine (as opposed to condition indicators created from historical data used to train the classifier 900). Figure 5BAs shown, the actual status indicator 18105 is obtained by comparing the actual operation output data 1152 (e.g., SCADA data) with the normal synthetic data 1280 obtained therefrom when the normal hybrid model 200 is fed with the actual operation input data (real parameters) 1151. The comparison may involve different mathematical operations applied to the compared data, such as difference, division, gradient, etc. The obtained status indicator 18105 is provided to the trained classifier 900, which outputs a label indicating "failure" or "no failure". In the case of a "failure" label, another label representing the failure / abnormal type is output. In the case of a failure, an early alarm 1050 is generated by the classifier 900.

[0083] The detection of the failure of stage 1000 can be performed in real time, that is, while the operating data (discrete values) are collected, for example, by a SCADA system. The operating data are collected at a certain collection frequency, which can be from microseconds (1 μs = 10 -6 The acquisition frequency may be changed from every 100 μs or every 1 ms to every second or minute, such as every 1 second, every 1 minute or every 10 minutes. The acquisition frequency may be based on different situations, such as the type of parameter being acquired.

[0084] Thus, the state of the drive system of the wind turbine can be diagnosed by means of a classifier 900, which receives as input the condition indicators 18105 at different time intervals within the considered time range. The diagnosis result can indicate valid operation, non-valid operation or a fault or anomaly. Based on the diagnosis result, a message (e.g., an alarm) 1050 can be provided, from which an action can be performed on the drive system of the wind turbine.

[0085] The proposed method optimizes the operation of a power conversion drive train of a wind turbine and improves the efficiency of the operation of the drive train and / or the maintenance of the drive train. The different models on which the method is based or used by the method as well as the training of the models and the execution of the trained models are preferably implemented in an IoT digital architecture, e.g. in the cloud, although it can alternatively be done locally. The different models are fed with field data (real operating data) over a certain time range, including historical data and actual data.

[0086] The method is implemented in one or more processing devices, which may be part of a processing system. One or more processing devices include at least one memory, which may include instructions, for example in the form of computer program code, so that when the instructions are executed, the method according to the invention is performed. The processing device may include a communication module, at least in an embodiment in which it is communicatively coupled to other processing devices.

[0087] The proposed method has been validated with real operating data of a doubly-fed induction generator wind farm with a rated power of 1.5 MW. The synthetic data generation method has been applied to a specific failure mode, namely stator winding overheating. The whole methodology and the obtained results are summarized next.

[0088] The method is applied to the use case: 1.5MW DFIG wind turbine

[0089] The method has been applied and validated with real SCADA data from an operating wind turbine. The drive train of the wind turbine has a 1.5MW DFIG and its corresponding back-to-back power converter. Real operating data collected over three years has been imported, organized and pre-processed. The collected real operating data includes: wind speed, blade pitch angle, mechanical torque, active power, phase current, stator winding temperature, nacelle temperature, etc. A detailed understanding of the available data is required as it has an impact on the specific model topology. For example, if the measurement corresponding to the mechanical torque 131 is available, a single module (second module 102) can be used to model the physics-based model 100. In the case where the mechanical torque is not available, the complete physics-based model (module 101 + module 102) must be used.

[0090] During data exploration and pre-processing of SCADA data, the relationships between parameters represented in the collected real operational data have been analyzed in order to detect possible outliers. Non-limiting examples of such relationships are wind speed vs. generated electrical power and current vs. nacelle temperature, etc. For example, in the power curve, unusual values ​​corresponding to outliers have been observed and removed to avoid confusing the model. It has been observed that these unusual values ​​correspond to, for example, periods when there is no power generation despite being a wind source, or periods with high temperatures in the stator windings when the generator is not generating power.

[0091] Once the initial data analysis was performed, a physical model 100 of the power conversion drive train of the wind turbine was developed in Simulink-Matlab R2020b. Figure 6 A representation of the model consisting of a first module 101 and a second module 102 is shown. Design information for both the generator and the power converter has been used as a starting point for building the model. However, some parameters are calculated or estimated in the absence of their information. Wind speed and pitch angle are the input parameters required for operating the physical model 100. In the case shown, the output of the model 100 is the generated electrical power and associated parameters (generated output current, generated output voltage, generated speed, rotor angle and electromagnetic torque).

[0092] In the first transformation (first module 101), the mechanical power extracted from the kinetic energy in the wind is determined by the air density, rotor sweep, wind speed and power coefficient. The extracted mechanical power is as follows:

[0093]

[0094] Where P is the output power (W), ρ is the air density (kg / m 3 ), A is the rotor swept area (m), V is the wind speed (m / s), and Cp is the power coefficient (dimensionless), which is the ratio of the power extracted by the wind turbine to the total power available in the wind source. The Betz limit of 0.59 is the maximum possible value of Cp, but wind turbines cannot be operated at this value due to design requirements for reliability and durability. In practice (real world), Cp values ​​are usually comprised between 0.25 and 0.45.

[0095] In the second transformation (second module 102), the DFIG block implements a three-phase wound rotor asynchronous motor operating in generator mode. The electrical part of the machine is represented by a fourth-order state space model, and the mechanical part is represented by a second-order system. All electrical variables and parameters are referenced to the stator, indicated by the prime sign in the following machine equations. All stator and rotor quantities are in an arbitrary dual-axis reference frame dq (Table 1).

[0096]

[0097] Table 1. Equivalent circuits and equations involved in DFIG conversion. The parameters involved in solving the DFIG conversion equations are those indicated in Table 2.

[0098]

[0099] Table 2. Parameters involved in DFIG operation

[0100] Other parameters related to the main shaft of the transmission system are usually also involved in the transformation. Non-limiting examples of these parameters are wind turbine inertia constant, shaft spring constant, shaft mutual damping and pitch angle.

[0101] The initial parameters of the physics-based model 100 are generalizations of the true parameters involved in the operation of a given turbine (e.g., stator resistance and leakage inductance, rotor resistance and leakage inductance, magnetizing inductance, total stator inductance, total rotor inductance, friction coefficient, inertia constant). The true values ​​of these parameters can be adjusted using an optimization algorithm ( Figure 1). The algorithm aims to find a combination of parameter values ​​that minimizes the difference between the output of the physics-based model 100 and the measured SCADA data. In this case, the parameters have been tuned (or calibrated) using the surrogate optimization algorithm (surrogateopt) in Matlab (Surrogate Optimization Algorithm - MATLAB & Simulink) (https: / / es.mathworks.com / help / gads / surrogate-optimization-algorithm.html). This optimization algorithm is a global solver specifically for situations where the objective function is computationally expensive. The algorithm searches for a cost function min with multivariate input variables x x The global minimum of f(x) with multivariate input variables x subject to linear and nonlinear constraints and some finite bounds. The resulting objective function can be non-convex and non-smooth. The algorithm starts by learning a surrogate model of the objective function by interpolating radial basis functions through random evaluations of the objective function within given bounds. In the next stage, a value function is minimized as an approximation of the minimization of the objective function. The value function f m Based on a weighted combination of the evaluation of the surrogate model computed in the previous stage and the distances between points sampled from the objective function, as follows:

[0102] f m (x) = wS(x) + (1-w)D(x) (14)

[0103]

[0104] Where S(x) is the scaled proxy output and D(x) is the scaled distance between points evaluated by the objective function. This distance reflects the uncertainty in the proxy model’s estimate. The value function min is performed by random search. X f m (x). The global minimum obtained is then evaluated by the objective function and the result is used to update the proxy model. The minimization of the value function is now calculated on the updated model. This process continues for a given number of iterations, or until a point is found where the objective function is below a threshold.

[0105] In the case of a drive train of a wind turbine, the objective function is defined as the mean absolute percentage error (MAPE) between the active power estimated by the physics-based model 100 and the active power measured by the SCADA system.

[0106]

[0107] During the optimization process in the case shown, 13 parameters are involved: 4 parameters associated with the electromechanical conversion, 3 parameters related to the aerodynamic conversion, 3 parameters of the control strategy and, finally, 3 parameters associated with the mechanical transmission system (Table 3).

[0108]

[0109] Table 3. Parameters involved in the optimization process

[0110] The calibration of the physics-based model was performed in two steps: in the first step, 6 variables were considered, while in the second step, more than 7 variables were added. The table shows the initial values ​​defined for each parameter (design values) and the values ​​adopted after the second calibration step (calibrated values).

[0111]

[0112] Table 4: Design and calibration values ​​of parameters involved in the optimization process

[0113] New values ​​of the calibration parameters are established, always maintaining their physical meaning. In fact, intervals defining the lower and upper allowed values ​​for each parameter have previously been established. The resulting model is a physically based enriched model 200 .

[0114] As a result, the mean absolute percentage error (MAPE) between the real active power measured in the SCADA system and the value obtained in the simulation using the calibration model (model 200) has been improved from 15% to 2.4%. Figure 7 As shown, it represents the relationship between active power (kW) and wind speed (m / s).

[0115] Once the physical model 100 is calibrated, an additional model 300 (in this case, a thermal model) has been added to the normal hybrid model (trained physics-based model) 200 already developed in Simulink to estimate the temperature in each phase of the stator windings, thereby generating a fault model. A schematic diagram of the thermal model is shown in Figure 8 It must be taken into account that the insulation class of the stator winding is F class, which means that it is designed to withstand 155°C. Figure 8 As shown, the thermal circuit takes into account the heat transfer caused by the stator current, taking into account the following phenomena: conduction (between the windings of each of the three stator phases) and convection (between the windings of each of the three stator phases, between each stator winding and the environment, and between each stator winding and the rotor). Radiation has been neglected.

[0116] The conductive heat transfer block models heat transfer in a thermal network by conduction through material layers. The heat transfer rate is governed by Fourier's law (18) and is proportional to the temperature difference, the material thermal conductivity, the area perpendicular to the direction of heat flow, and inversely proportional to the layer thickness:

[0117]

[0118] The Convective Heat Transfer block models heat transfer in a thermal network by convection due to the movement of a fluid (in this case, air). The heat transfer rate (19) is proportional to the temperature difference, the heat transfer coefficient, and the surface area in contact with the fluid:

[0119] Q conv =h c A dT (19)

[0120] The inputs fed to the thermal model are the stator current and the room temperature where the generator is installed (in this case the temperature of the nacelle), while the output is the temperature of each phase of the stator winding.

[0121] Following the same process used in the normal hybrid model, this fault model is trained using operational real SCADA data ( Figure 3C ). Similarly, the training consists in optimizing the values ​​of some independent design parameters representing failure, whose exact values ​​have been estimated between given realistic intervals. In this case, these parameters are the material thermal conductivity, the area perpendicular to the direction of heat flow, the thickness of the layer through which conduction phenomena occur, the heat transfer coefficient and the surface area in contact with the air. After this training, a fault hybrid model is generated.

[0122] In the real data available during the analysis and execution of this use case, five abnormal conditions were flagged as overheating in the stator windings, such as Fig. 9A , Fig. 9B , Fig. 9C and Fig.9D The fault modeling has been verified using the data during these five abnormal situations, and the FIG. 10A to FIG. 10E Results of the estimated stator winding temperatures are shown in Figure 1, where they are compared with the true SCADA winding temperatures. The value of the MAPE between the true stator winding temperature values ​​measured in the SCADA system and the values ​​obtained in the simulation using the fault mixture model is 11%, with a maximum percentage error of 16% in the worst case. This value could still be improved if more accurate design data were available for the thermal model.

[0123] Fig.11The method for generating synthetic failure data for a generator stator winding overheating scenario formed by a data-driven random failure model and a fault hybrid model is schematically shown. Fig.11 The learned statistical distribution shown randomly generates synthetic inputs associated with faults, and the synthetic inputs are fed as inputs to the developed fault mixture model 500 ( FIG. 4A to FIG. 4B ). The fault hybrid model 500 generates the rest of the synthetic failure data (eg, stator winding temperature).

[0124] Fig.12 Both the stator winding temperature values ​​calculated by the fault hybrid model (non-bold line), and the true values ​​of the stator winding measured by the SCADA system (bold line) are shown, the former taking those values ​​(ambient temperature, wind speed, cabin temperature) generated synthetically as input variables.

[0125] A condition indicator has been created as the difference between the stator winding temperature calculated using the fault mixture model and the actual stator winding temperature measured in the SCADA and allows classification to detect actual cooling problems. This will lead to the triggering of an alarm. This is a simple rule-based classifier. Additionally, a benchmark analysis has been performed using multiple features (such as wind speed, torque, difference in stator winding temperature, difference in nacelle temperature, etc.) and different algorithms such as KNN and random forests with other more complex classifiers.

[0126] In summary, a hybrid model-based digital twin that combines the advantages of physics-based models and advanced data analysis techniques has been implemented. The proposed digital twin performs very effectively to control / supervise the operation of the power conversion drive system of a wind turbine, improve the efficiency (optimization) in the operation of the drive system, and / or improve the maintenance of the drive system.

[0127] Among the different advantages of the proposed digital twin-based approach: on the one hand, a procedure for generating synthetic failure data based on real data using different statistical techniques. On the other hand, a classification model for detecting anomalies in the operation of wind turbine drive trains. These two innovations make it possible to address the main limitations of current digital twin approaches in terms of accuracy, interpretability and lack of sufficient training data.

[0128] In this document, the term “comprise” and its derivatives (such as “comprising”, etc.) should not be understood as excluding the possibility that the described and defined content may include other elements, steps, etc.

[0129] In the context of the present disclosure, the term "approximately" and its cognate terms (such as "approximately" etc.) should be understood to indicate values ​​that are very close to those values ​​accompanying the aforementioned terms. That is, deviations within reasonable limits of the exact value should be accepted, as those skilled in the art will understand that such deviations from the values ​​represented are inevitable due to measurement inaccuracies, etc. The same applies to the terms "about", "approximately" and "substantially".

[0130] The present invention is obviously not limited to the specific embodiments described herein, but also includes any modifications (e.g. regarding the choice of materials, dimensions, components, configurations, etc.) that may be considered by a person skilled in the art within the general scope of the invention defined in the claims.

Claims

1. A computer-implemented method for training a classifier for failure diagnosis of a drive train of a wind turbine, comprising: training a physics-based model (100) of the drive train of a wind turbine using operational data (151, 152) of the drive train of the wind turbine under normal conditions, thereby providing a normal hybrid model (200) of the drive train of the wind turbine; Modeling at least one abnormal situation in the normal hybrid model (200) and training it using operation data (151, 152') of the transmission system under a failure condition, thereby providing a fault hybrid model (500) of the transmission system of the wind turbine; Applying a set of input data (151) of the transmission system under normal conditions to the normal hybrid model (200), thereby obtaining a set of synthetic data (280) under normal conditions for specific parameters of the transmission system; Applying a set of synthetic input data (710) and a set of failure data (150") associated with the at least one abnormal condition of the transmission system to the fault mixture model (500) to generate a set of synthetic failure data (750) for the at least one abnormal condition; Obtaining a set of condition indicators (8105) for the at least one abnormal condition based on a set of output data (152, 152') of the transmission system, the set of synthetic data (280) obtained under normal conditions, and the set of synthetic failure data (750) for the at least one abnormal condition; A classifier (900) is trained using the set of condition indicators (8105) for the at least one abnormal condition.

2. The method according to claim 1, wherein: The physics-based model (100) of the drive train of the wind turbine comprises a first module (101) representing an aerodynamic conversion stage of the drive train and a second module (102) representing an electromechanical conversion stage of the drive train.

3. A method according to any one of the preceding claims, wherein: The normal mixture model (200) is obtained as follows: providing a physics-based model (100) of the drive train of the wind turbine, the physics-based model (100) being represented by a set of equations and a plurality of model parameters representing energy conversion in the drive train of the wind turbine; feeding the physics-based model (100) with operating data (151, 152) of the transmission system under normal conditions, the operating data comprising real operating data input (151) and real operating data output (152); obtaining a prediction (180) of at least one of the following parameters: generated electrical power, phase current, phase voltage, and electromagnetic torque; Values ​​of at least some of the model parameters are selected, wherein the values ​​minimize the difference between the prediction (180) and the real operational data output (152) corresponding to the real operational data input (151), thereby obtaining the physics-based model (200) trained with the operational data.

4. A method according to any one of the preceding claims, wherein: The fault hybrid model (500) is obtained as follows: Modeling at least one abnormal situation in the normal mixture model (200); feeding the normal hybrid model (200) including at least one modeled abnormal situation with operating data (151, 152') of the transmission system under a failure condition; obtaining a prediction of a parameter indicative of a fault condition (180'); The values ​​of at least some of the model parameters defining the normal hybrid model (200) modeled with at least one abnormal situation are selected, wherein the values ​​minimize the difference between the prediction (180') and the real operating data output (152') corresponding to the real operating data input (151), thereby obtaining the calibrated fault hybrid model (500).

5. The method according to claim 4, wherein: The at least one abnormal situation is modeled with the aid of at least one failure model (301 to 306).

6. The method according to claim 5, wherein: The at least one failure model (301 to 306) is represented by a set of equations and a plurality of failure parameters representing failures in energy conversion of the drive train of the wind turbine.

7. A computer-implemented method for optimizing the operation of a drive train of a wind turbine, comprising using the classifier (900) trained in any one of the preceding claims.

8. The computer-implemented method of claim 7, comprising: Obtaining a status indicator (18105) based on normal synthetic data (1280) obtained from the normal mixture model (200) when the normal mixture model (200) is fed with actual operational input data (1151) and actual operational output data (1152); providing the condition indicator (18105) to the trained classifier (900); The condition indicator (18105) output by the trained classifier (900) is whether it corresponds to a failure condition or a non-failure condition.

9. The computer-implemented method of claim 7, further comprising: When the classifier (900) determines that the transmission system is operating under abnormal conditions, an alarm is triggered.

10. The computer-implemented method of any one of claims 8 to 9, further comprising: When the status indicator (18105) corresponds to a failure condition, actuating on the drive train of the wind turbine.

11. An apparatus comprising processing means configured to perform the method according to any one of claims 1 to 6.

12. An apparatus comprising processing means configured to perform the method according to any one of claims 7 to 10.

13. A computer program product comprising computer program instructions / codes for performing the method according to any one of claims 1 to 6 and / or the method according to any one of claims 7 to 10.

14. A computer readable memory / medium storing program instructions / codes for executing the method according to any one of claims 1 to 6 and / or the method according to any one of claims 7 to 10.

Images