Method and system for predicting the risk of observable damage in a wind turbine gearbox component

Abstract

Embodiments disclosed herein may include methods, systems, and tangible, non-transitory, computer-readable media with instructions for determining a damage state of a wind turbine gearbox. A method executed on, performed by, or implemented as instructions by a processor of a system may include generating a damage state diagnosis, generating a damage progression model for a wind turbine gearbox using operational data and a state transition function, and combining the damage state diagnosis with the damage progression model using a hybrid prognostic model to generate a probability distribution of current damage state estimates. Generating the damage state diagnosis may include obtaining a measurement data set of the wind turbine gearbox, integrating intelligence provided by a field operator into the measurement data set, normalizing the measurement data set with respect to known failure behaviors of the wind turbine gearbox, and combining the normalized measurement data sets with known failure behaviors to generate the damage state diagnosis.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to U.S. Provisional Application No. 62 / 786,817, filed on December 31, 2018, the entire disclosure of which is incorporated herein by reference. Technical Field

[0003] The present disclosure generally relates to wind turbines. More specifically, the present disclosure generally relates to a method of assessing the risk of observable damage to a wind turbine gearbox component. Background Art

[0004] Predicting and mitigating damage accumulation in wind turbine components is a multi-billion dollar opportunity globally, given the exponential growth in the installed base of wind turbines. The rapid increase in the number of wind turbines in operation in the world over the past few years has led to the field-verified realization that damage is accumulating faster than expected in the wind turbine gearbox (WTG) components of these wind turbines, resulting in increased gearbox failures leading to millions of dollars in additional operations and maintenance (O&M) expenditures. As a result, there is a growing demand for solutions that can provide early warning of damage initiation and tracking of damage development with reduced uncertainty, smart asset operations that are closely linked to specific damage / failure modes, and real data-driven reliability analytics to optimize long-term maintenance and component procurement strategies.

[0005] A variety of prediction algorithms have been developed and applied to predict damage accumulation in WTG components. These approaches are either purely physics-based or purely data / statistical and often rely on the integration of additional expensive retrofit sensor suites. Some hybrid approaches have been proposed that attempt to combine physics-based approaches with statistical, data-driven approaches that rely on retrofitting of sensors. These approaches suffer from lack of scalability, high uncertainty (i.e., low accuracy of predictions), and economic impracticality due to the need for expensive retrofit sensors (e.g., CMS). The main reasons for these shortcomings include the lack of accurate representation of sub-component-level physics and multi-body drivetrain interactions; over-reliance on purely data-driven predictions using time-series data (“big data”) from wind turbine SCADA systems; minimal integration of digital “small data,” i.e., site- and customer-specific information (e.g., alarm logs, site maintenance history logs, bill of materials information, geospatial information, and configuration information); poor data governance resulting in large amounts of undetected noise and inconsistencies in information across the various data inputs; and minimal integration of quantitative field intelligence from the experience of actual field operators.

[0006] Therefore, there is a need for improved methods and systems to accurately assess the risk of observable damage in WTG components. Summary of the invention

[0007] An embodiment may be a method for determining a damage state of a wind turbine gearbox of a wind turbine. The method may include generating (i.e., producing) a damage state diagnosis, generating a damage progression model of the wind turbine gearbox using operational data and a state transition function, and combining the damage state diagnosis with the damage progression model using a hybrid prognostics model to generate (i.e., produce) a probability distribution of a current damage state estimate.

[0008] An embodiment may be a system for determining a damage state of a wind turbine gearbox of a wind turbine. The system may include a processor. The processor may be configured to perform a method including generating a damage state diagnosis, generating a damage progression model of the wind turbine gearbox using operational data and a state transition function, and combining the damage state diagnosis with the damage progression model using a hybrid prognostic model to generate a probability distribution of a current damage state estimate.

[0009] An embodiment may be a tangible, non-transitory, computer-readable medium having instructions thereon. When executed by a processor, the instructions may cause the processor to perform a method for determining a damage state of a wind turbine gearbox of a wind turbine, including generating a damage state diagnosis, generating a damage progression model for the wind turbine gearbox using operational data and a state transition function, and combining the damage state diagnosis with the damage progression model using a hybrid prognostic model to generate a probability distribution of a current damage state estimate.

[0010] Generating a damage state diagnosis may include acquiring a measurement data set from a wind turbine, integrating intelligence provided by a field operator into the measurement data set, normalizing the measurement data set relative to known failure behaviors of a wind turbine gearbox to generate a standardized measurement data set, and combining the standardized measurement data set with the known failure behaviors to generate a damage state diagnosis.

[0011] The hybrid prognostic model may include a sample-based predictor. The method, whether executed on a processor of the system, performed, or implemented by the processor as instructions from a tangible, non-transitory computer-readable medium, may further include using the hybrid prognostic model to generate a future state prediction. The future state prediction may include a probability of the wind turbine gearbox reaching a critical damage threshold for a prediction horizon, a time to reach the prediction horizon, and a system state of the wind turbine gearbox at the time step to reach the prediction horizon.

[0012] The damage progression model may include a physics-based model. The physics-based model may include one or more of finite element analysis, computational fluid dynamics, multibody dynamics, or linear elastic fracture mechanics.

[0013] The damage progression model may include a machine learning based model.

[0014] Field intelligence provided by a field operator may include operator findings of previously seen faults or damage. The operator's previously seen findings of faults or damage may include anecdotal evidence not present in the borescope analysis. The method, whether executed on a processor of a system, performed, or implemented by a processor as instructions from a tangible, non-transitory computer-readable medium, may further include constructing a model or constructing a data representation using the operator's previously seen findings or damage.

[0015] The measurement data set may include one or more of direct measurement data, indirect measurement data, global behavior or operational data.

[0016] The measurement data set may include direct measurement data. Obtaining the measurement data set may include performing direct measurement using a measurement unit to obtain the direct measurement data. The measurement unit may include a borescope.

[0017] The measurement data set may include indirect measurement data and obtaining the measurement data set may include performing indirect measurements using a temperature sensor or a vibration sensor.

[0018] The global behavior data may include a comparison of operational data of one or more similar wind turbines similarly configured as the wind turbine.

[0019] The operational data may include supervisory control and data acquisition (SCADA) data. The SCADA data may include one or more of a time series of system input data, system output data, or a system log. The SCADA data may include system input data and the system input data may include one or more of wind speed or wind direction. The SCADA data may include system output data and the system output data may include generated power. The SCADA data may include a time series of system logs, and the time series of system logs may include a time series of wind turbine states or faults.

[0020] Known failure behaviors can be derived from historical data. Historical data can include maintenance records, field failure information, root cause analysis, or bill of materials data. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] For a more complete understanding of the nature and objects of the present invention, reference should be made to the following detailed description in conjunction with the accompanying drawings, in which:

[0022] Figure 1 is a flow chart illustrating an embodiment of this method;

[0023] Figure 2 An embodiment hybrid remaining useful life framework for combined data fusion is shown; and

[0024] Figure 3 An example implementation of a hybrid prognostic model is shown. DETAILED DESCRIPTION

[0025] Although the claimed subject matter will be described in terms of specific embodiments, other embodiments, including embodiments that do not provide all of the benefits and features set forth herein, are also within the scope of this disclosure. Various structural, logical, process step, and electronic changes may be made without departing from the scope of the present disclosure. Therefore, the scope of the present disclosure is defined solely by reference to the appended claims.

[0026] The present disclosure provides methods, systems and apparatus for predicting the risk of observable damage in WTG components such as bearings and gears. Observable damage may refer to the accumulation of damage over time resulting in visually identifiable physical damage in a single or multiple gearbox components. Damage accumulation may refer to the process by which components in a WTG that transmit torque from a rotor to a generator degrade over time due to the formation of microscopic cracks under repeated loading cycles.

[0027] Embodiments of the present disclosure implement a novel combination of material science, data science, and field experience insights to provide accurate predictions of the cumulative risk of damage to WTG components, thereby driving effective optimization of operational maintenance and life extension of wind turbines. These embodiments may include processes that utilize heterogeneous information sets to predict wind turbine reliability. In addition, such embodiments may provide designers, site managers, asset managers, and risk managers with a tool to evaluate the performance of WTG components under a variety of conditions: different materials, manufacturing processes, and operating conditions, without having to rely on expensive and time-consuming testing. The inputs to the embodiment methods and systems using the data fusion process according to the present disclosure may include material physics models, data science models, and field experience insights.

[0028] Material physics driven models can use the laws of physics to express a causal understanding of the physical processes responsible for the initiation and propagation of damage through computational models.

[0029] Material physics models may have advantages in their ability to represent problems and failure modes through a purely physics-based understanding - without the need for any continuous sensor measurements. They may provide a long-term probabilistic view of damage accumulation and time to failure, as well as measures of uncertainty that allow for risk-adjusted decision making. They may be highly interpretable and explainable - that is, because the output is a result of the rules of physics, designers and creators may clearly understand why the output looks and behaves in a particular way.

[0030] Material physics models may be limited to the representation of failure modes that are explicitly represented by a single model. As a result, such approaches are unable to represent or predict failure modes that are unknown or lack a physics-based causal path that follows the WTG design (e.g., problems arising from manufacturing, design, or field-specific operational issues, etc.). Material physics models may also have high computational costs, often requiring supercomputers or large numbers of memory-intensive cloud nodes, and therefore may not be inherently suitable for solutions at scale.

[0031] Data science models can refer to a range of statistical learning techniques focused on anomaly detection and decision tree type techniques. Such models may aim to predict the accumulation and progression of damage using features derived entirely from sensors and the correlations / associations between them. Note that unlike physics driven models, data driven models do not require causal paths to form the basis of a predictive algorithm, as data driven models rely on correlations rather than causation.

[0032] Data science models may have the ability to identify multiple predictive patterns in data without identifying relevant features. They may be well suited to solving short-term damage prediction problems through techniques such as anomaly detection and classification tree type approaches. They may be relatively low computational cost and amenable to solutions at scale. They can integrate a variety of data into a single predictive model (e.g., sensor datasets, log-type datasets, or field intelligence through expert rules, etc.).

[0033] Data science models can be limited to being a direct function of the data and can be prone to issues such as overfitting, which can lead to false positives. They can require large amounts of data for training, testing, and implementation. They can be highly sensitive to noise in the data - if not filtered through domain knowledge, the model can learn noise and signal equally. They may require previously known failures to predict similar types of failures. They may not predict things that were not present in the data used to train the model.

[0034] Field experience insights can refer to the distillation of field personnel / asset managers’ experience in dealing with various faults and issues while operating wind farms in various conditions and locations. Such field experience insights can be combined with predictive material physics and data science models to complete the overall data fusion model.

[0035] The approach presented here may rely on data fusion (i.e., combination) of material physics-driven modeling and data science guided by intelligence derived from field experience. Such data fusion approaches may be based on optimizing the strengths and limitations of each input to the data fusion.

[0036] Data fusion can estimate the current state of damage, provide a prediction of the remaining useful life (RUL), and can be described as decision-level fusion, where a combination of various predictive techniques and field intelligence are used to provide the determination of quasi-independent time windows associated with the predicted visible risk of damage. The core advantage of data fusion may be attributed to the use of a material physics modeling framework, where the actual physical processes involved in component degradation can be modeled rather than relying solely on statistical analysis of the collected data.

[0037] The material physics modeling approach according to the present disclosure can combine and simulate multiple input parameters without assuming parameter correlation or linearity of damage accumulation. This can be achieved by embedding high-fidelity modeling representations of stresses and strains that drive damage accumulation processes to explain common WTG component failure modes, such as rolling contact fatigue and wear.

[0038] A material physics-driven modeling approach can consider the following techniques applied to the various contributing stages of the gearbox component damage accumulation process. This unique combination allows for more accurate and flexible predictions of damage progression. In particular, these techniques can include: wind loading calculated using aeroelastic simulation software, including the effects of wind speed, wind shear, turbulence intensity, and air density; multibody dynamics analysis of load transmission from the rotor through the gearbox to the generator, including shaft and gearbox housing bending; and high-fidelity component loading models that can accurately describe the distribution of macro-level loading to the sub-component level, including simulation of the conditions that govern the tribological behavior of contacting bodies.

[0039] The data science component of the data fusion approach can leverage heterogeneous “big” and “small” data that are transformed in predictive decision data streams through multiple predictive frameworks. Examples of such data streams may include, among others: time series data from various sensors in wind turbines – often referred to as Supervisory Control and Data Acquisition (SCADA) data; various operational logs (e.g., logs of faults / alarms, logs of turbine operating status, or field notes manually entered by field maintenance technicians); digital records of previously known damage / fault diagnosis and reporting (e.g., borescope images, oil analysis or root cause analysis reports, etc.); a history of maintenance and replacement events for each wind turbine (e.g., bearing replacement, gearbox replacement, or routine maintenance); geospatial information, including the geographic coordinates of each turbine or other geographic datasets (land use, topography, etc.) encompassing the entire wind farm; time series data of meteorological data from point locations within the wind farm (and their coordinates), or, if unavailable, files describing measured climatic conditions used to assess expected weather-related risks in wind farm design; and / or a bill of materials (BOM) for each WTG.

[0040] Each of these input data sets can be exploited by a variety of different predictive data science models based on partitioning of the input data set based on percentage representation of specific information and fidelity regarding specific aspects of wind turbine operational behavior. The data science component of the data fusion approach may involve running a collection of time series type classification and regression models and statistical models, which can convert qualitative risk factors into quantifiable indicators. For example, point measurements residing in a SCADA system cannot represent the entire three-dimensional envelope of loads and environmental conditions encountered by a rotating wind turbine. Therefore, a statistical model can combine terrain information, wind turbine coordinate empirical rules, and site-level wind direction time series into a turbine-level risk indicator representing the risk of high turbulent loads from wind farm density and steep terrain. Similarly, known turbine-level alarm logs can be converted into time-based aggregate scores that can be used as damage risk features. All of these features, as well as standard features derived from the time series data set, can be analyzed in training mode against known past failure records of various turbines. This step can inform the operator whether the features used have predictive power for known damage / failure events when represented in a classification type model or a regression type anomaly detector model. Each data science model within the ensemble can produce a risk rating, including the damage level and time window of damage for all modeled wind turbines.

[0041] By executing material physics models and supporting data science models, various modeling outputs can be generated and combined through a hybrid RUL estimation framework.

[0042] The RUL framework may recognize that, at any given point in time, not all known ways a turbine gearbox may damage or fail are represented by the various damage progression and measurement models embedded in the framework. This incomplete representation of the resulting damage problem may manifest itself through incorrect predictions - deemed incorrect by comparison to gearbox borescope analysis (direct measurement) or other known failures of wind turbine gearboxes. Therefore, the present disclosure may utilize a third input: field intelligence obtained from turbine operators and managers.

[0043] Figure 1 This is presented in a flow chart illustrating an embodiment method. For example, each time an inaccurate prediction occurs, a forensic analysis can be performed and compared directly to experience-based field intelligence. This may reveal specific causes of damage or failure that are not represented in the RUL framework. Any such determination of new causes can be further reviewed through detailed analysis and modeling studies to form hypotheses and express them as new damage measurement / progression models.

[0044] It is important to note that the estimate or prediction of the current damage state may be directly related to the on-site inspection of the damage event. Based on the on-site inspection results triggered by the damage estimation method, this framework can update the future predicted RUL or, in the case of incorrect predictions, enhance the embedded model through forensic analysis and on-site intelligence integrated into the model.

[0045] The hybrid RUL framework that integrates data fusion is Figure 2 To maximize the accuracy of the prediction, many different data streams may be used simultaneously. Data from previous events 1, also referred to as “historical data,” may include maintenance records, field failure information, root cause analysis, BOM data, etc. Historical data may be used to extract comparative information for damage that has previously occurred on similarly installed and used systems and / or components (i.e., assets), such as similar wind turbines having a similar configuration to the wind turbine for which the damage status is to be determined. These data from previous events may be separate from information streams derived from direct or indirect measurements of the current asset, a few examples of which are Figure 2 -As shown in Figure 4.

[0046] The current state of damage of an asset can be assessed through direct physical inspection, during which measurement instruments can be used to quantify the state of damage relative to the installed condition. The accuracy and precision of these measurements depends on the instrumentation used. For example, wind turbine technicians often use borescope imaging2 to inspect components in gearboxes. Images of damaged areas can be used to generally classify the state of damage (e.g., damage categories 1-4) based on the size of the damaged area observed. Direct measurements can be affected by instrument errors.

[0047] Indirect measures of damage state can be obtained through sensors that monitor the asset during operation3. Sensor measurements can provide time series data of system parameters that are functionally dependent on the changing damage state. For example, a temperature sensor installed within a WTG can measure a time series of temperatures, and a concomitant study of past known failures may reveal a large number of such events associated with higher temperatures. This has previously been referred to as an anomaly detector data science model. For example, clustering techniques can define the “normal behavior” of a specific relevant temperature variable (based on data analysis of time series from previous known failures), and the adjustment of damage thresholds created by all temperature anomalies through a regression-type temperature predictor model can produce a cumulative measure of such anomalies that is used to provide an indirect measure of damage state. Indirect measurements can be affected by two main sources of error, namely measurement error and mapping error. Measurement error may refer to issues such as calibration, drift, EMI, etc. Mapping error may refer to the uncertainty introduced by the process of relating measured variables to the latent variable of interest (i.e., damage state).

[0048] Another indirect measurement can be obtained by using a data science model that uses measurements of global system behavior. Since the global behavior of the system may change over time as the condition of the asset degrades, this data stream can be a useful independent data stream that can be used to enhance the understanding of the current health state of the asset. An example of such global system behavior data 4 for a wind turbine application may be an alarm / fault code. Fault codes may be generated by a controller based on rules that define normal and abnormal behavior. For example, when the temperature measured by one or more sensors exceeds a predetermined value, a controller may generate a temperature exceedance fault. The method according to the present disclosure may utilize a trained data science model based on a verified relationship between specific sequences of various fault codes and past known damage / fault events. Therefore, the output of such a model run on a time-aggregated feature set of various alarm codes can represent another indirect measurement of the current damage state.

[0049] An understanding of the physics of failure modes associated with an asset of interest can be used to develop models that describe how the asset degrades over time and use. Such physics-based models6 can take various forms depending on the physics involved. Some examples of common physics-based modeling techniques include finite element analysis (FEA), computational fluid dynamics (CFD), multibody dynamics (MBD), and linear elastic fracture mechanics (LEFM). The term “damage” in relation to physical components is often related to changes in the state of the constituent materials, so a good understanding of the material’s response to operating conditions (materials science) is required to develop effective models. Because the models approximate the real physical system, it is also necessary to understand the sources of model uncertainty.

[0050] Measurement systems 3-4 and material physics-based models use operational data 5, which provides information describing the conditions under which the asset is used. Such data can include sensor time series measurements. In wind turbines, a supervisory control and data acquisition (SCADA) system can perform this function, acquiring system input data (e.g., wind speed, wind direction), system output (e.g., power generation), and any time series of system logs (e.g., status / faults, etc.). Operational data may include local measurements (e.g., visual inspections), remote measurements (sensor data). Operational data can also be broken down to provide local and global behavior data for a single asset, which can allow the asset to be compared to a Met Mast (i.e., a tower containing weather measurement equipment) or multiple turbines.

[0051] The definition of a “severe damage level” for an asset can vary widely depending on a number of factors. In general, a severe damage level defines a state of damage at which the asset can no longer safely perform its intended function. Historical data can be used in part to define the state of damage at which similar assets were removed from service. Analysis of a database containing, for example, asset bills of materials, known operating conditions / usage, maintenance records, etc., can extract a set of known failures that are similar in nature to the asset and failure mode of interest7.

[0052] As mentioned above, in addition to direct measurements of field-verified damage (e.g., borescopes), various data science models can be used directly as measurement objects to provide measurements of the current state of damage.

[0053] The output of the measurement data fusion process 8 can be normalized (0≤D≤1) relative to the known fault behavior 7 of the WTG (e.g., based on the past performance of similar WTGs) to form a damage state diagnosis 11. A damage state "diagnosis" can be defined as a single measurement (i.e., observation) of an asset in operation. The diagnostic data must be reported along with a measure of the associated uncertainty. For example, diagnostic data exhibiting Gaussian uncertainty can be reported based on the mean and variance of the distribution.

[0054] Damage progression models can be defined in terms of state transition functions9, where system inputs are functionally related to the rate of damage progression. These functions are likely (with a few exceptions) to be monotonic in nature - a gear or bearing does not "not wear out" with increasing time\usage. Damage progression models may be physics-based6 or may be derived from data analysis (machine learning).

[0055] In order to predict future damage states, future usage needs to be understood. The Future Input Predictor Model 10 can be used to analyze previous operating conditions that an asset has experienced and infer future usage patterns. In the wind turbine example, previous SCADA data can be mined to determine a histogram of wind loads representative of this particular asset each month, allowing seasonal variations to be accounted for.

[0056] Information from measurements (diagnoses) and damage progression models may be combined in a hybrid prognostic model 12. A hybrid prognostic model may generally be defined as combining potential damage progression functions using evidence from multiple independent data streams to produce probabilistic estimates of current and future states.

[0057] A variety of different state estimation algorithms can be used to combine the diagnostic observation data with a damage progression model (Kalman filter, unscented Kalman filter, particle filter, etc.), and the output is a probability distribution of the current damage state estimate 13. A sample-based prediction algorithm can then be used to produce (i.e., generate) future state predictions 15, which may include the probability of the WTG reaching a critical damage threshold for a given prediction horizon, the time to reach this range, and the system state of the WTG at each future time step (e.g., until this range is reached).

[0058] In application, a current state estimate 13 indicating a level of damage of concern to an operator can trigger a manual inspection process or other maintenance action by an expert in the field. Empirical knowledge of nominal behavior gained through experience is a resource that can be used to "tune" prognostic models. Field intelligence 16 can be used to identify out-of-nominal conditions for a specific asset that may lead to unexpected damage progression. Such conditions are typically not observable in online data streams (e.g., lubricant contamination, assembly issues, damage to secondary components, etc.), but may result in significant deviations from the nominal damage progression behavior. Integrating field intelligence as a source of feedback into the prognostic prediction process can provide a formal mechanism for adjusting the algorithm based on the real-world condition of the asset.

[0059] Figure 3 An example implementation of a hybrid prognostic model 12 is shown. The purpose of the prognosticator object is to encapsulate functionality for joint state parameter estimation and uncertainty management. The prognosticator object may be initialized 22 with an initial state estimate. In the example particle filter implementation given, the initial particle state variables may be sampled from an initial diagnosis 21. At this point, the prognosticator may also be provided hooks to a diagnostic data stream 25, an asset operating condition database 26, and assigned one or more further damage progression models 27. The prognosticator object may then be stored 23 for later use.

[0060] When a new diagnostic data point is detected, the predictor can first use the time from the last update (T i -1) to the current time (T i )'s operational data 26, and a damage progression model 27 is used to build an estimate. The particle filter algorithm employed here may be based on sample importance resampling (SIR) and can be easily found in the literature. The particle weights may then be updated 29 based on the likelihood of a given diagnostic observation 28. Particle degeneration occurs when a few particles have high weights and the rest have very low weights. To avoid the degeneracy problem, the particles may be resampled 30 after the weight calculation. The updated and resampled weighted particles thus represent the current damage state estimate 31 as a discrete probability distribution.

[0061] If a forward prediction of end of life (EOL) or remaining useful life (RUL) is required 32, a sample-based predictor 33 can be used to propagate each particle state forward in time, using estimated future operating conditions 34 and a damage progression model 27 until the damage state exceeds a critical level. Due to the uncertainty of the model and future operating conditions, the future state prediction 35 can also be described by a discrete probability density function. The RUL 36 can be calculated as the difference between the current time and some measure (e.g., mean) of the future state distribution.

[0062] In an example embodiment, a method for determining a damage state of a wind turbine gearbox may include generating a damage state diagnosis. The damage state diagnosis may be generated by obtaining a measurement data set of the wind turbine, integrating field intelligence provided by a field operator into the measurement data set, normalizing the measurement data set with respect to known fault behaviors to generate a standardized measurement data set, and combining the standardized measurement data set with the known fault behaviors to generate the damage state diagnosis. Examples of field intelligence provided by the field operator may include faults or damage previously discovered by the operator (e.g., anecdotal evidence that does not exist in any current borescope) that may be used to build a model or represented in the data.

[0063] The measurement data set may include one or more of direct measurement data, indirect measurement data, global behavior, and operational data. Direct measurement may include use of measurement, where the measurement unit is, for example, a borescope. Examples of indirect measurement may include use of a temperature sensor or a vibration sensor. Examples of global behavior may include comparisons of operating data of wind turbines of similar configurations. Operational data may include, for example, supervisory control and data acquisition (SCADA) data, and SCADA data may include, for example, a time series of system input data, system output data, or a system log. System input data may include, for example, wind speed or wind direction. System output data may include, for example, the power generated by the subject wind turbine. The time series of the system log may include, for example, a time series of wind turbine states and / or faults.

[0064] Known failure behaviors may be derived, for example, from historical data, which may include, for example, maintenance records, field failure information, root cause analysis, and / or bill of materials data.

[0065] The method may further include generating a damage progression model for the wind turbine gearbox using the operational data and the state transition function. The hybrid prognostic model may be used to combine the damage state diagnosis with the damage progression model to generate a probability distribution for the current damage state estimate. In some embodiments, the hybrid prognostic model may further use a sample-based predictor to generate a future state prediction. The future state prediction may include one or more of a probability of reaching a critical damage threshold for a prediction horizon, a time to reach the prediction horizon, and a system state for a time step.

[0066] In some embodiments, the damage progression model can be a physics-based model that combines, for example, finite element analysis, computational fluid dynamics, multibody dynamics, and / or linear elastic fracture mechanics. In some embodiments, the damage progression model can be a machine learning-based model. In some embodiments, the damage progression model can include components of a physics-based model and a machine learning model.

[0067] The present disclosure may be embodied as a system for determining a damage state of a wind turbine gearbox, wherein the system may implement any of the methods described herein. Such a system may be, for example, a processor programmed to perform any of the disclosed methods. Such a system may further include a communication link for receiving one or more of SCADA data, direct measurement data, indirect measurement data, operational data, and the like.

[0068] The processor may be in communication with and / or include a memory. The memory may be, for example, a random access memory (RAM) (e.g., dynamic RAM, static RAM), flash memory, removable memory, etc. In some instances, instructions associated with performing the operations described herein (e.g., operating an image sensor, extracting multiple patches from a holographic image) may be stored in the memory and / or storage medium (including, in some embodiments, a database storing instructions) and the instructions executed at the processor.

[0069] In some instances, the processor may include one or more modules and / or components. Each module / component executed by the processor may be any combination of hardware-based modules / components (e.g., field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), digital signal processors (DSPs)), software-based modules (e.g., modules of computer code stored in memory and / or databases and / or executed at processors) and / or a combination of hardware- and software-based modules. Each module / component executed by the processor can perform one or more specific functions / operations as described herein. In some instances, the modules / components included and executed in the processor may be, for example, processing, applications, virtual machines, and / or some other hardware or software modules / components. The processor may be any suitable processor configured to run and / or execute those modules / components. The processor may be any suitable processing device configured to run and / or execute a set of instructions or codes. For example, the processor may be a general-purpose processor, a central processing unit (CPU), an accelerated processing unit (APU), a field programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a digital signal processor (DSP), etc.

[0070] Some examples described herein may involve computer storage products with non-transitory computer-readable media (also referred to as non-transitory processor-readable media) having instructions or computer codes for performing various computer-implemented operations. Computer-readable media (or processor-readable media) are non-transitory in the sense that they do not themselves include transient propagating signals (e.g., propagating electromagnetic waves that carry information on a transmission medium such as space or cables). The media and computer codes (also referred to as codes) may be those designed and constructed for specific purposes. Examples of non-transitory computer-readable media include, but are not limited to: magnetic storage media such as hard disks, floppy disks, and magnetic tapes; optical storage media such as compact disks / digital video disks (CD / DVD), compact disk read-only memories (CD-ROMs), and holographic devices; magneto-optical storage media such as optical magnetic disks; carrier signal processing modules; and hardware devices specifically configured to store and execute program codes, such as application-specific integrated circuits (ASICs), programmable logic devices (PLDs), read-only memories (ROMs), and random access memory (RAM) devices. Other examples described herein may involve computer program products, which may include, for example, instructions and / or computer codes discussed herein.

[0071] Examples of computer code may include, but are not limited to, microcode or microinstructions, machine instructions, such as code generated by a compiler for generating a Web service, and files containing high-level instructions executed by a computer using a translator. For example, Java, C++, .NET, or other programming languages ​​(e.g., object-oriented programming languages) and development tools may be used to implement examples. Additional examples of computer code may include, but are not limited to, control signals, encrypted code, and compressed code.

[0072] In some embodiments, the present disclosure may be embodied as a tangible, non-transitory, computer-readable medium having instructions thereon that, when executed by a processor, execute thereon or cause the processor to perform any of the methods disclosed herein.

[0073] The steps of the methods described in the various embodiments and examples disclosed herein are sufficient to implement the methods of the present invention. Therefore, in an embodiment, the method is essentially composed of a combination of the steps of the methods disclosed herein. In another embodiment, the method consists of these steps.

Claims

1. A method for determining a damage state of a wind turbine gearbox of a wind turbine, comprising: Generates damage status diagnostics including: acquiring a measurement data set from said wind turbine; integrating field intelligence provided by a field operator into the measurement data set; normalizing the measurement data set relative to a known fault behavior of the wind turbine gearbox to produce a normalized measurement data set; combining a standardized measurement data set with the known fault behavior to generate the damage state diagnosis; generating a damage progression model for the wind turbine gearbox using operational data and a state transition function; and combining the damage state diagnosis with the damage progression model using a hybrid prognostic model to produce a probability distribution of current damage state estimates, wherein the hybrid prognostic model combines potential damage progression functions using evidence from multiple independent data streams to produce probability estimates of current and future states.

2. The method of claim 1 , wherein the hybrid prognostic model comprises a sample-based predictor, and wherein the method further comprises using the hybrid prognostic model to generate a future state prediction, the future state prediction comprising: a probability of said wind turbine gearbox reaching a critical damage threshold for the prediction horizon, The time to reach the prediction horizon, and System state of a wind turbine gearbox at one time step.

3. The method according to claim 1, wherein: The damage progression model comprises a physics-based model. 4 . The method of claim 3 , wherein the physics-based model comprises one or more of finite element analysis, computational fluid dynamics, multibody dynamics, or linear elastic fracture mechanics.

5. The method according to claim 1, wherein: The damage progression model comprises a machine learning based model.

6. The method of claim 1, wherein the field intelligence provided by the field operator includes operator findings of previously seen faults or damages.

7. The method according to claim 6, wherein: The previously seen operator findings of failure or damage include anecdotal evidence that was not present in the borescope analysis.

8. The method of claim 6, further comprising building a model using operator findings of previously seen failures or damages.

9. The method of claim 6, further comprising constructing the data representation using operator findings of previously seen failures or corruptions.

10. The method according to claim 1, wherein: The measurement data set includes one or more of direct measurement data, indirect measurement data, global behavior data, or operational data.

11. The method according to claim 10, wherein: The measurement data set includes direct measurement data, and obtaining the measurement data set includes performing direct measurement using a measurement unit to obtain the direct measurement data.

12. The method according to claim 11, wherein: The measuring unit comprises a borescope.

13. The method according to claim 10, wherein: The measurement data set includes indirect measurement data, and obtaining the measurement data set includes performing indirect measurements using a temperature sensor or a vibration sensor.

14. The method of claim 10, wherein the global behavior data comprises a comparison of operational data of one or more similar wind turbines similarly configured as the wind turbine.

15. The method of claim 10, wherein the operational data comprises supervisory control and data acquisition (SCADA) data.

16. The method of claim 15, wherein the SCADA data comprises one or more of a time series of system input data, system output data, or system logs.

17. The method according to claim 16, wherein: The SCADA data includes system input data, and the system input data includes one or more of wind speed or wind direction.

18. The method according to claim 16, wherein: The SCADA data includes system output data, and the system output data includes generated power.

19. The method of claim 16, wherein the SCADA data comprises a time series of system logs, and the time series of system logs comprises a time series of wind turbine status or faults.

20. The method of claim 1, wherein the operational data comprises supervisory control and data acquisition (SCADA) data.

21. The method of claim 20, wherein the SCADA data comprises one or more of a time series of system input data, system output data, or system logs.

22. The method according to claim 21, wherein: The SCADA data includes system input data, and the system input data includes one or more of wind speed or wind direction.

23. The method according to claim 21, wherein: The SCADA data includes system output data, and the system output data includes generated power.

24. The method of claim 21, wherein the SCADA data comprises a time series of system logs, and the time series of system logs comprises a time series of wind turbine status or faults.

25. The method of claim 1, wherein: The known fault behaviors are derived from historical data.

26. The method according to claim 25, wherein: The historical data may include maintenance records, field failure information, root cause analysis, or bill of materials data.

27. A system for determining a damage state of a wind turbine gearbox, comprising a processor configured to perform the method according to any one of claims 1-26.

28. A tangible, non-transitory computer-readable medium having instructions thereon which, when implemented by a processor, cause the processor to perform the method of any one of claims 1-26.

Images