The field of the present disclosure relates, in general, to system monitoring and diagnostics. More particularly, the present disclosure relates to systems, devices and methods to detect and determine faults based on patterns in vibration spectra data.
The working condition of different components of an asset may be monitored by a Conditions Monitoring System (CMS), wherein alerts may be generated to indicate faults and other warnings regarding the operation of the asset. For some assets such as a wind turbine for example, a wind turbine Vibration analyst might monitor the health of the turbine's gearbox and other bearings using a CMS that uses bearing vibration data to detect anomalous spectrum spikes and, accordingly creates alerts based on the detected spectrum spikes. This conventional approach relies heavily on the availability and knowledge of the exact gearbox kinematic information for the wind turbine being monitored and is rule based, and uses limits or changes in trends to alert of possible component wear or damage.
In general, some conventional vibration monitoring and analysis systems usually consist of two steps where the first step is an automated alarming system that stores or has access to physical information such as natural fault frequency and rotational frequencies derived from kinematics information of the wind turbine, such as the count of gear teeth/rolling elements, etc. In a second step, a vibration analyst may review vibration spectrum plots associated with the wind turbine to understand whether the diagnosis from the alarming system is correct. This second step is necessitated since the alarm system usually only looks at absolute amplitude values or changes in trends at specific locations (i.e., frequencies) in the spectrum. By using precise kinematics information, a traditional CMS system may miss a known fault if a signature varies even slightly from a textbook (i.e., theoretical) signature. This is due to the rule based nature of the system that relies on spectrum spikes at specific frequencies to indicate known types of faults. While this approach might work for problems where the fault signature is well defined and well known, it tends to break down when a new (i.e., novel) fault emerges.
Accordingly, conventional systems may potentially miss novel faults and/or may not correctly detect known faults that have a signature that varies, even slightly, from standard fault frequencies or a signature that shifts away from the standard fault frequencies as the wear or damage progresses.
Therefore, there exists a need for methods and systems that support and facilitate a holistic view of a vibration spectrum for an asset, as opposed to specific amplitude values at specific frequencies in signal spectra.
In one aspect, an embodiment of the present disclosure relates to a method including receiving vibration spectrum data from a plurality of different assets; determining, based on a shape of the vibration spectrum data for each of the plurality of assets, clusters for the plurality of assets, assets being grouped in a same cluster having vibration spectrum data of a similar spectral shape; determining for each of the clusters, based on an application of domain derived pattern recognition rules for the vibration spectrum data, one of a plurality of fault classifications; generating an output including an association of each of the plurality of assets with the fault classification of the cluster in which the particular asset is grouped; and saving a record of the output.
In other embodiments, a system including spectrum pattern matching (or similar functionality) module and a root cause identification (or similar functionality) module may be implemented to perform at least some of the features of the methods and processes disclosed herein. In yet another example embodiment, a tangible medium may embody executable instructions that can be executed by a processor-enabled device or system to implement at least some aspects of the processes of the present disclosure.
While the present disclosure describes the use of various methods and systems for the detection of wind turbine bearing, shaft and gearing issues, the present disclosure has applicability to all rotating or reciprocating machinery, including but not limited to and across a number different types of power generation, industrial and manufacturing industries. Illustrative examples of some equipment where methods and systems disclosed herein apply include, but are not limited to, power generation equipment, motors, pumps, gearboxes, bearings, shafts, gears, compressors, metal extrusion equipment, mining and metal production equipment.
These and other features, aspects, and advantages of the present disclosure will become better understood when the following detailed description is read with reference to the accompanying drawings in which like characters represent like parts throughout the drawings, wherein:
Unless otherwise indicated, the drawings provided herein are meant to illustrate features of embodiments of this disclosure. These features are believed to be applicable in a wide variety of systems comprising one or more embodiments of this disclosure. As such, the drawings are not meant to include all conventional features known by those of ordinary skill in the art to be required for the practice of the embodiments disclosed herein.
In the following specification and the claims, a number of terms are referenced that have the following meanings.
The singular forms “a”, “an”, and “the” include plural references unless the context clearly dictates otherwise.
“Optional” or “optionally” means that the subsequently described event or circumstance may or may not occur, and that the description includes instances where the event occurs and instances where it does not.
The present disclosure relates to processes and systems to detect and diagnose faults and other conditions based on a pattern or spectral shape of vibration spectrum from an asset.
Regarding
Data captured by sensors in the drivetrain may primarily be waveform data of the vibrations, displacements or acoustic signatures generated by the drivetrain components (e.g., high speed shaft and low speed shaft bearings) and derivatives of those waveforms. In some aspects, the primary types of data that might be captured can include asynchronous waveforms and synchronous waveforms. An asynchronous waveform is a type of vibration signal that is recorded for a fixed amount of time. Since the assets of the example of
As referred to herein, a synchronous waveform is a waveform that is recorded over a certain number of revolutions of a bearing and may not be stored directly. Instead, derived spectrograms related to the waveforms may be stored for a certain period of time (e.g., every 4 hours). The direct FFT (Fast Fourier Transform) of the waveform is referred to as a high resolution spectrum, whereas a finer and more detailed spectrum derived post enveloping via a digital signal process (DSP) is referred to as a synchronous (or sync)enveloped spectrum. Plots of both a high resolution spectrum and a sync enveloped spectrum may capture and reveal distinctly visible faults including sidebands (if any), while the high-resolution spectrum may display the fault signatures in a plot with a higher noise floor or a plot containing signatures not directly related to component wear or damage.
In some applications, the signatures for high speed stage faults may be more pronounced in the sync enveloped spectrum data. In some aspects, some faults might exhibit their signature in sync enveloped spectrum data before exhibiting any anomaly in the high resolution spectrum data, over a typical lifecycle of a gearbox bearing. In some embodiments herein, sync enveloped spectrum data may be used for HSS fault detection algorithms (discussed in greater detail below).
Regarding frequency spectrum plots, a brief discussion is presented for a better understanding of the frequency ranges they represent. Referring to the example of
Thus, if the gear ratio of HSIS to HSS is 3, the fundamental frequency values will be HSS 1×=24 Hz and HSIS 1×=8 Hz.
It is noted that different faults may have their own 1× values. In some aspects, the 1× values of faults may be viewed as the frequency at which a particular fault creates noise within the system. For example, a crack in the bearing outer race of a bearing may cause a mild disturbance/noise or vibration signature every time a ball traverses the crack. As such, the outer race fault 1× fundamental frequency (also called the Bearing Ball Pass Frequency Outer-race (BPFO) 1×) will be the frequency of that generated noise.
In some aspects, the spectrum of each stage of the drivetrain of the wind turbine of the
For the example of
In some contexts, an analysis of the amplitudes of the fault frequency information can be used to determine whether a particular type of fault is present in a bearing of an asset if the fault frequency information for a particular asset is available and known. However, there are several limitations to this approach that is based on knowledge of the fault frequency information. First, the fault frequency information may not be readily available, if at all. Fault frequencies might be known through mathematical formulae that relate the design parameters such as, for example, the number of rolling elements, pitch angle, roller diameter, pitch diameter, etc. to the particular defect frequencies (e.g., BPFI, BPFO and BSF). However, the design parameters vary greatly from one bearing to another and may not be known or even supplied to a vibration engineer or other entity.
Secondly, the fault frequency might not be present in the spectra (i.e., the “faulty” spectra) of a particular asset. In theory, the natural fault frequencies (e.g., BPFO, BPFI, etc.) should be present where the corresponding fault is present in a bearing. However, in practice that may not always be the case. For example, it is possible that a frequency close to a BPFO is getting excited in a bearing that actually has ORBP issues. As this example demonstrates, an over-reliance on theoretical fault frequencies can be detrimental in a practical (i.e., real-world) application since faults characterized by amplitude spikes might actually occur at a frequency different than the expected/theoretical frequency and might not be seen when looking for faults at the expected/theoretical frequency. The deviation in an actual fault signature is often caused by secondary damage to the bearing in and around the location of the initial fault, causing sideband energy or frequency shifts in the signature produced when the defect is encountered by a rolling element.
These limitations of relying on known fault frequencies often necessitates that a vibration engineer review spectrum data for even common fault patterns (e.g., faults in high speed stages). This manual review by expert(s) can be costly and/or leads to productivity losses.
Some previous techniques, processes, and systems, include a vibration engineer looking at a report with some of the subject turbines flagged as being a concern. For each flagged turbine, the vibration engineer might look at the different spectral plots of that turbine to ascertain whether the suspected fault is indeed present. This type of inspection might include a review of the spectral DEI (Dynamic Energy Index) numbers apart from visual inspection of the spikes and aligning them with the natural fault frequencies, where this amount of effort is necessitated by the fact that these previous types of systems rely on a thresholding algorithm that looks for a rise in amplitude at a specific location (i.e., frequency).
Some embodiments herein may use asset (e.g., gearbox) agnostic algorithms to determine a defect based on the basic defect signatures (i.e., “patterns”) in vibration spectrum data. An overview of some embodiments will be explained in the context of bearing signatures for ORBP, IRBP, BS, and other defects.
For an outer race defect in a bearing, a shock impulse will be recorded every time a rolling element of the bearing traverses the defective location. For these types of generated impulses, the ORBP frequency (BPFO) and its harmonics are present in the resulting spectra.
For an inner race defect in a bearing (i.e., an IRBP signature), the signature of an inner race fault is more complex than an ORBP signature. In some embodiments and configurations, the inner race of a bearing usually rotates with a shaft and thus the inner race defect location moves in and out of the bearing loading zone. As such, the shock impulse created when the bearing is in the loading zone is more pronounced than an impulse created when the bearing moves out of the loading zone. In this manner, the BPFI impulse signal gets amplitude modulated by the shaft's rotational frequency (referred to as the shaft 1×). These factors result in the presence of a spike in amplitude at the BPFI frequency (and its harmonics), as well as the presence of one or two sidebands separated by the shaft 1× frequency. An illustrative IRBP fault signature is depicted in a plot 300 of
For a ball spin defect in a bearing (i.e., a BS signature), the signature pattern of a rolling element defect may be similar to an IRBP defect. As illustrated in
In some instances, there may be other fault signatures for which there are no existing detection algorithms. Some such examples may include some signatures pertaining to low speed stages of a turbine drivetrain that are visible only in data derived from stage 1 of the drivetrain.
For example, a Planetary Bearing fault may be characterized by a distinctive ‘haystack’ pattern in a region between gear mesh harmonics. Unlike the sync enveloped spectrum in the examples of
Another low speed stage (LSS) fault is the Ring Gear fault that exhibits a distinctive “sawtooth” pattern in the stage 1 high res data.
In some embodiments, the present disclosure relates to a system (including software and hardware components) that receives wind turbine gearbox accelerometer data and uses pattern matching techniques to detect assets that display similar spectral shapes of the accelerometer vibration spectrum.
In some embodiments, the assets may each be a wind turbine as discussed in example of
At operation 805, vibration spectrum data is received from a plurality of assets. The vibration spectrum data may be received from, for example, a plurality of wind turbines deployed in a “wind farm”. The data received at operation 805 may include one or more files, wherein vibration spectrum data for each of the subject assets is identifiable in the data received at operation 805.
At operation 810, a pattern matching module or other system or device having similar functionality ingests the vibration spectrum data from each wind turbine asset. In some embodiments, operation 810 may also include other processing aspects such as, but not limited to, averaging the received data for each asset over a certain time frame (i.e., period of time). In some embodiments, operation 810 may use a clustering algorithm or a combination of clustering algorithms to finalize pairing of assets. In some embodiments, one or more features may be extracted from the received vibration spectrum data by operation 810, in an effort to reduce a dimensionality of the data, on a per asset basis.
At operation 815, a Root Cause detection process may use clustering algorithms on the vibration spectrum (e.g., lower dimensional) data to identify the assets (e.g., turbines) that share a similar spectral shape. Assets determined to have a similar spectral shape are group together. The groupings of assets having similar spectral shapes (i.e., patterns) are referred to herein as “clusters”. In some aspects, assets that are in the same cluster are similar by spectral shape and are therefore of similar fault or health condition/status.
In some embodiments, a granularity of the clustering results of operation 810 might be too high for an intended or practical use. For example, in some instances only turbines facing upwind or downwind HSIS IRBP fault might be grouped together, while an vibration engineer may be interested in faults only at a summary fault level of HSIS IRBP (i.e., excluding differentiations in upwind or downwind).
Proceeding to operation 815, a Root Cause Identification module or other system(s) and device(s) having a similar functionality may operate to identify an exact cause of the problem or fault characterized by the spectral shape (i.e., signature) in the clusters of different shape vibration signals.
In some aspects, a nonparametric clustering algorithm of operation 810 might excel at grouping assets based on the spectral shape of the individual asset's vibration spectrum data, and will not be able to identify the exact nature of the faults represented in the clusters (i.e., groups). In some embodiments, algorithm(s) to determine the clusters based on identifying vibration spectrum data having a similar spectral shape may include one of a hierarchical clustering algorithm, a k-means algorithm, a nearest neighbor algorithm, and at least one algorithm based on a combination of clustering methods. As used herein, a hierarchical clustering algorithm assigns each asset to its own cluster and then successively merge pairs of clusters that are closest to each other. The process can be repeated until a desired number of clusters are found. As used herein, a k-means algorithm consists of selecting a few of the spectra randomly as cluster centroids. Each of the other spectra are then assigned to these clusters based on their distance from the centroids. The centroids are updated and the process is repeated until the algorithm converges (i.e., no changes in the cluster assignments or centroids).
In some embodiments, root cause identification algorithm(s) at operation 815 may take a data based, gearbox agnostic approach to detect the particular fault in the clusters. In some instances, a process for detecting or determining a particular fault at operation 815 may be determined in a gearbox agnostic way, to the extent possible. In some embodiments, multiple different algorithms may be used to detect the anomalous spikes, sidebands, and other characteristic features in a specific spectrum and based on the different patterns (e.g., a spike at a certain location along with sidebands, a specific haystack-like pattern, a presence of asynchronous harmonics, a sawtooth pattern, etc.) or signature generate a fault diagnosis such as, for example, Inner Race/Outer Race/Ball Spin issues in higher speed stages and Planet Bearing and other faults in lower speed stages.
Operation 820 may include a final output that combines the outputs of a pattern matching module and a root cause identification module to generate a unified view including a fault determination for the plurality of assets that may be better than a strictly clustering or a strictly rule based approach.
In some embodiments, an output of operation 820 may be presented via a user interface frontend that might include textual, graphical, and other visualization representations that provide a mechanism for a user to view and analyze an asset's spectrum, in isolation and/or in comparison with other assets that may exhibit similar spectral patterns.
At operation 915, an initial clustering of the assets based on a spectral shape of their associated vibration signals is performed. Operation 915 may include executing a hierarchical clustering function. Operation 920 may further merge clusters based on, for example, their centroids being less than some threshold value (e.g., <1σ) away from a mean centroid distance for the clusters.
Having grouped the different assets into clusters based on their spectral shapes at operations 915 and 920, operation 925 operates to identify, based on the centroids of merged clusters, whether the signal in each of the merged clusters is indicative of a fault in the asset.
Operation 930 determines, for each cluster, a distance of each asset (e.g., turbine) in the cluster from the centroid of the cluster. If the asset is relatively far away from the mean based on some threshold value (e.g., >0.75σ) as determined at operation 935, then process 900 advances to operation 940 where the asset may have its condition classified by, for example, a rule engine since the particular asset under consideration is sufficiently different from the centroid asset in its cluster to be classified the same.
If, at operation 935, the asset under consideration is not relatively far away from the mean based on some threshold value (e.g., <0.75σ), then process 900 advances to operation 945 where the asset may be classified the same as the centroid asset in its cluster.
At operation 1020, a linear regression model is executed to compute the slope of the signal. The values of the linear regression output are further subtracted from the normalization values computed at operation 1015. At operation 1025, the RMS energy of the normalized signal is computed. All values less than the RMS value (calculated at operation 1010) are set to 0 (e.g., finalSpectra).
Process 1000 is repeated for all signals of each asset being evaluated, as indicated by feedback loop 1030.
At operation 1125, a cluster (n) is chosen as a reference cluster and operation 1130 operates to determine whether the clusters under consideration are sufficiently close (i.e., similar) to be merged together. Operation 1130 determines whether the centroid of cluster n<1σ away from the centroid of cluster n+1. If the determination of operation 1130 is “yes”, then process 1100 advances to operation 1135 where cluster n and n+1 are merged, the cluster to be checked next is updated, and the centroid of the newly merged cluster is computed.
If the determination of operation 1130 is “no”, then process 1100 advances to operation 1140 where cluster n and n+1 are not merged and the cluster to be checked next is updated. Thereafter, operations 1125-1140 are repeated until no further merging of clusters is possible.
At operation 1230, if >50% of the harmonic peaks exhibit a peak at equal distances from the harmonic peak, then the signal is classified as being indicative of an IRBP fault at operation 1235. If <50% of the harmonic peaks do not exhibit a peak at equal distances from the harmonic peak at 1230, then process 1200 proceeds to operation 1240 where the most common sideband width is identified and for the first 3 harmonic peaks a check is executed to determine whether there are sidebands at the identified sideband width.
If >2 harmonic peaks show sidebands at the identified sideband width, as determined at operation 1245, then the signal is classified as being indicative of an IRBP fault at operation 1255. Otherwise, the signal is classified as being indicative of a “healthy” asset at operation 1250.
At operation 1325, the first 5 harmonics of the tallest peak are calculated to form a checklist. If at operation 1330 the number of harmonics having no sidebands is not ≥2 (i.e., <2), then the asset associated with the signal is deemed to be “healthy” at operation 1335. If operation 1330 determines the number of harmonics having no sidebands is ≥2, then a determination is made at operation 1340 regarding whether there are more than 3 continuous peaks in the checklist (from operation 1325) that are >0. If there are more than 3 continuous peaks in the checklist that are >0, then the signal is classified as being indicative of an ORBP fault at operation 1345. Otherwise, the asset associated with the signal is deemed to be “healthy” at operation 1350.
At operation 1420, the “subList” is examined to identify locations where the slope of the signal changes (i.e., the inflection point). Moreover, only those inflection points where the change in slope is <0.2 (i.e., identifying the triangles that characterize BS faults) are retained for further processing. At operation 1425, the number of triangles present in the signal is computed (“nTriangles”) and process 1400 advances to operation 1430.
Referring to operation 1435, input vibrations signals are received and at operation 1440 if the location of the tallest peak in the signal is >10 and <100 and the first three multiples of the tallest peak are non-zero, then a “BSFlag” is set to 1, otherwise the “BSFlag” is set to 0. From operation 1440, process 1400 advances to operation 1430.
At operation 1430, a determination is made whether the “BSFlag”=1 and “nTriangles”≥5. If the “BSFlag”=1 and “nTriangles”≥5, then the signal is identified as being indicative of asset having a BS fault. Otherwise, the asset associated with the signal is deemed to be “healthy” at operation 1450.
At operation 1520, the grouped high amplitude peaks are filtered based on the number of peaks in the group, where any group having less than 4 peaks is filtered out. A “Group Ratio” is calculated as specified in
Referring to flow 1502 and starting at operation 1540, the assets of the present example are grouped based on common gear mesh frequencies and harmonic signatures. At operation 1545 and for each gearbox group, the following values are calculated: number of haystack patterns in each turbine spectrum in the group (“nGroup”), average number of haystack patterns across the group (“AvgGroup”), and the number of turbines in each group (“size”).
At operation 1550, if the “size”>10, then the standard deviation of “nGroups” in the gearbox is calculated at operation 1560 and saved (e.g., “sigma”). At operation 1565, the turbines in the group are classified based on the value of “nGroups”. For example, for turbines having “nGroups”<1 “sigma”, the turbine is classified as being “healthy”; for turbines having “nGroups”>1 “sigma” and <2 “sigma”, the turbine is classified as being “potentially faulty”; and for turbines having “nGroups”>2 “sigma”, the turbine is classified as being “faulty”.
At operation 1550, if the “size” is <10, then the mean(“AvgGroup”) and mean(“sigma”) across the larger groups are calculated at operation 1552. At operation 1555, the turbines in the larger groups are classified based on the value of “nGroups”. For example, for turbines having “nGroups”<2[mean(“AvgGroup”)] are classified as being “healthy” and for turbines having “nGroups”>2[mean(“sigma”)] are classified as being “faulty”.
In some embodiments, processes and systems herein detect faults and a fault type directly based on high resolution spectra and are capable of significant savings in time and manual effort. In some instances, the technologies disclosed herein may, for example, double the productivity of a CMS expert (e.g., more turbines per analyst).
In some embodiments, processes and systems herein provide a mechanism to analyze and determine faults in vibration spectrum data from a wide variety of assets, including assets for which fault frequency information is not known and/or available to an analyst. In some embodiments, the processes disclosed herein may be executed, at least in part, automatically in response to one or more events or actions.
Some embodiments herein provide signature-based, asset (e.g., gearbox) independent/agnostic detection processes and systems that enable fault detection even when a specific natural fault frequency is absent and/or unknown.
Apparatus 1600 includes processor 1605 operatively coupled to communication device 1615 to communicate with other systems, data storage device 1630, one or more input devices 1610 to receive inputs from other systems and entities, one or more output devices 1620 and memory 1625. Communication device 1615 may facilitate communication with other systems and components, such as other external computational assets and data. Input device(s) 1610 may comprise, for example, a keyboard, a keypad, a mouse or other pointing device, a microphone, knob or a switch, an infra-red (IR) port, a docking station, and/or a touch screen. Input device(s) 1610 may be used, for example, to enter information into apparatus 1600. Output device(s) 1620 may comprise, for example, a display (e.g., a display screen) a speaker, and/or a printer.
Data storage device 1630 may comprise any appropriate persistent storage device, including combinations of magnetic storage devices (e.g., magnetic tape, hard disk drives and flash memory), solid state storages device, optical storage devices, Read Only Memory (ROM) devices, Random Access Memory (RAM), Storage Class Memory (SCM) or any other fast-access memory.
Fault rule engine 1635 may comprise program instructions executed by processor 1605 to cause apparatus 1600 to perform any one or more of the processes described herein, including but not limited to aspects disclosed in
Data 1640 (either cached or a full database) may be stored in volatile memory such as memory 1625. Data storage device 1630 may also store data and other program code for providing additional functionality and/or which are necessary for operation of apparatus 1600, such as device drivers, operating system files, etc. Data 1650 may include data related an asset that may be used in the identification of faults herein.
Although specific features of various embodiments of the disclosure may be shown in some drawings and not in others, this is for convenience only. In accordance with the principles of the disclosure, any feature of a drawing may be referenced and/or claimed in combination with any feature of any other drawing.
This written description uses examples to disclose the embodiments, including the best mode, and also to enable any person skilled in the art to practice the embodiments, including making and using any devices or systems and performing any incorporated methods. The patentable scope of the disclosure is defined by the claims, and may include other examples that occur to those skilled in the art. Such other examples are intended to be within the scope of the claims if they have structural elements that do not differ from the literal language of the claims, or if they include equivalent structural elements with insubstantial differences from the literal language of the claims.