Maintenance system and method for improving the reliability and / or availability of a power converter station

Abstract

A maintenance system (1) for improving the reliability and / or availability of a power converter station (100) by risk-based maintenance, being configured to analyze live data (5) from monitoring devices (2a, 2b, 2c), which monitor subcomponents (7a1-7f) of components (3a, 3b, 3c, 61, 62) of the power converter station (100), wherein the live data (5) is analyzed for reactive risk management (11), which comprises alerting a user when an error event (E) is detected, and for proactive risk management (10), which comprises calculating one or more reliability and / or availability parameters (RF, MTTR) for the subcomponents (7a1-7f) and / or components (3a, 3b, 3c, 61, 62) and one or more system reliability and / or availability parameters (SR, DR) for the entire power converter station (100) based on the reliability and / or availability parameters (RF, MTTR) for the subcomponents (7a1-7f) and / or components (3a, 3b, 3c, 61, 62), being configured for alerting a user and / or providing mitigation suggestions (MS) on the level of the subcomponents (7a1-7f).

Description

[0001] P2023,0923 WO E / P220328WO01 January30,2024 -1 - DescriptionMaintenance system and method for improving the rel iabilityand / oravailabilityofa powerconverterstationThe present disclosure relates to a maintenance sys tem forimproving the reliabilityand / oravailabilityofa power converterstation.With Internet of Things (IoT) technology becoming w idelyapplied, the asset risk and lifecycle management isexperiencing a digitalization transition.However, in thecurrent digital twin markets, operators of power co nverterstations tend to get overwhelmed by various tools f ortargeting failure modes. Without an expert understa nding ofthe system,operatorsmayfailto respond in time.As machine-learning technologies advance rapidly, c ondition-based maintenance is possible with applications lik e anomalydetection and classification, including failure pro gnosis.However, due to insufficient understanding of the i ntrinsiccomponentcharacteristicsand theirfailure modes, thealgorithms are often applied in predictions of rand omfailures,whose patternsare notpredictable.This mayresult in misinterpretation ofthe generated outcomes.Embodiments of the disclosure relate to an improvedmaintenance system forpowerconverterstations.According to a first aspect, a maintenance system f orimproving the reliabilityand / oravailabilityofa powerconverter station by risk-based maintenance is conf igured touse live data from monitoring devicesmonitoring P2023,0923 WO E / P220328WO01 January30,2024 -2 -subcomponents of components of a power converter st ation. Themaintenance system is configured to analyze the liv e databoth forreactive riskmanagement,which comprises alerting auser when an error event is detected, and for proac tive riskmanagement, which comprises calculating one or morereliabilityand / oravailabilityparametersforthesubcomponents and / or components and one or more sys temreliabilityand / oravailabilityparametersforthe entire powerconverterstation.The maintenance system is configuredfor alerting a user and / or providing mitigation sug gestionson the levelofthe subcomponents.Accordingly, failures of subcomponents are not only detectedand predicted,butthese failuresare connected to system reliabilityand / oravailabilityparametersforthe entirepower converter station. Such system reliability pa rameterscan include a shutdown risk (reliability) and a dow ntime risk(availability). The system reliability and / or avail abilityparametersmaybe calculated from the subcomponentreliability and / or availability parameters. By info rming theuser on the risk for the entire system, guidance is given forthe severity and urgency of a maintenance action. T hemaintenance system may be configured for generating an alarmwhen a shutdown risk for the entire power converter stationexceedsa specified threshold level.The maintenance system may comprise a front-end pla tform foralerting the userand / orproviding the mitigationsuggestions. As an example, an alarm may be trigger ed in caseofan error. The live data maycomprise a time seriesofsensor measurements,forexample.Asan example,the live data may P2023,0923 WO E / P220328WO01 January30,2024 -3 -comprise a time series of acoustic measurements. Ac ousticmeasurements may be made for monitoring a bearing o f acooling pump, for example. The live data may compri se loggedevents. In reactive risk management, an alarm may b egenerated when a logged event occurs. It is also po ssiblethatthe live data isanalyzed to detectan error.The maintenance system may be configured to use a m achine-learning algorithm for proactive risk management an d / orreactive risk management. The machine learning algo rithm maycomprise pattern recognition models, for example. B y themachine-learning algorithms for proactive risk mana gement, atrend estimation and prediction can be made of when criticalthresholds are exceeded. From this, failure rates f or thesubcomponentscan be updated.The machine-learning algorithmmay be configured to use training data from a fleet of powerconverterstations.As an example, the machine-learning algorithm may b econfigured to calculate an anomaly score for a time series oflive data. Thereby, a deviation from a normal behav ior of thesubcomponent is determined. Based on the anomaly sc ore, anerroristriggered and / orreliability / availability parameters are calculated. The maintenance system maybe configured to inform a useronmitigation actions when an error has occurred in re activerisk management. Alternatively or additionally, themaintenance system may be configured to inform a us er onmitigationssuggestionsforpreventing a predicted errorin proactive riskmaintenance. P2023,0923 WO E / P220328WO01 January30,2024 -4 -According to a further aspect, a power converter st ationcomprises the maintenance system as disclosed in th eforegoing. The power converter station further comp rises thesubcomponentsand monitoring deviceswhich provide the livedate to the maintenance system. The maintenance sys tem may bea local part of the power converter station. The ma intenancesystem maybe also a remote system.According to a further aspect, a method for improvi ng thereliability and / or availability of a power converte r stationbyrisk-based maintenance comprisesanalyzing live data from monitoring devices,which monitorsubcomponentsof the powerconverter station. The algorithm may have a clear l ink tospecific failure modes of that subcomponent. In the method,the live data is analyzed for reactive risk managem ent, whichcomprises alerting a user when an error event is de tected,and forproactive riskmanagement,which comprisescalculating one or more reliability and / or availabi lityparameters for the subcomponents and one or more sy stemreliabilityand / oravailabilityparametersforthe entirepower converter station. The method may further com prisealerting a user and / or providing mitigation suggest ions onthe levelofthe subcomponents.The present disclosure comprises several aspects an dembodiments. Every feature described with respect t o one ofthe aspects and embodiments is also disclosed herei n withrespectto the otheraspectsand embodiments,even iftherespective feature is not explicitly mentioned in t hiscontext.Further features, refinements and expediencies beco meapparent from the following description of the exem plary P2023,0923 WO E / P220328WO01 January30,2024 -5 - embodimentsin connection with the figures.In the figures,elements of the same structure and / or functionality may bereferenced by the same reference signs. It is to beunderstood that the embodiments shown in the figure s areillustrative representations and are not necessaril y drawn toscale.Figure 1 shows an embodiment of a maintenance syste m in aschematicdiagram,Figure 2 shows an embodiment of a cooling pump syst em in aschematicdiagram,Figure 3 shows an embodiment of a front-end platfor m in aschematicview,Figure 4 shows an embodiment of failure detection b y amachine-learning method in a schematic flow cha rt,Figures5A to 5C show stepsin an analysisoflive data for failure detection bya machine-learning method in schematicdiagram.Figure 1 shows a maintenance system 1 for improving thereliability and / or availability of a power converte r station100 by risk-based maintenance. Figure 2 shows a coo lingsystem 4 of the power converter station 100 as a sp ecificexample forcomponents3a,3b,3cmonitored bythe maintenance system 1. A powerconverterstation comprisesa pluralityof monitoringdevices 2a, 2b, 2c for monitoring the operation sta tus ofcomponents 3a, 3b, 3c (see Fig. 2) of the power con verter P2023,0923 WO E / P220328WO01 January30,2024 -6 -station. The monitoring devices 2a, 2b, 2c may be c ontroland / or protection devices. The monitoring devices 2 a, 2b, 2cprovide live data 5 to a risk analyzing algorithm 9 .As an example, one of the components 3a, 3b, 3c may be acooling pump arrangement 3a. Further components may be heatexchangers3b orfurtherparts3c.The cooling pump arrangement 3a comprises a plurali ty ofcomponents 6 1, 62 and subcomponents 7a 1-7f. In particular, thearrangement 3a comprises two cooling pumps 6 1, 6 2, which canbe connected to the same power supply. The cooling pumps 61,62 are shown in a parallel connection, illustrating th at thecooling pumps 6 1, 6 2 can work independently from each other,wherein the functionality of the cooling pump arran gement 3ais also provided with only one of the cooling pumps 61, 6 2 inoperation. Accordingly, a redundant system is provi ded.The first cooling pump 6 1 comprises a first bearing 7a 1 andthe second cooling pump 6 2 comprises a second bearing 7a 2. Thebearings 7a 1, 7a2 are subcomponents which are monitored andcan be replaced in case ofan error. Furthersubcomponentsofthe arrangement3a maybe a PLC(programmable logic controller) 7b of the cooling s ystem 4, aUPS (Uninterruptible PowerSupply,)7c,a pressure sensor7d,a flow sensor 7e and an AC power supply 7f, for exa mple.One or more of the subcomponents 7a 1-7f are monitored by oneor more monitoring devices 2a, 2b, 2c. It is also p ossiblethat one or more of the subcomponents 7a 1-7f act themselvesas monitoring devices 2a, 2b, 2c and can be additio nallymonitored byothermonitoring devices. P2023,0923 WO E / P220328WO01 January30,2024 -7 -In the specific example of Fig. 2, a monitoring dev ice 2amonitors the first bearing 7a 1. The second bearing 7a 2 may bemonitored by the same monitoring device 2a or a dif ferentmonitoring device.The monitoring device 2a maybe a noise sensororvibration sensor.The live data 5 (see Fig. 1) comprises a time serie s 5a of anoutput of one or more of the monitoring devices 2a, 2b, 2c.As an example, the monitoring device 2a of the spec ificembodimentofFig.2 in the form ofa noise sensor orvibration sensor provides a time series 5a of acous tic sensormeasurements.Furthermore, the live data 5 comprises transient fa ultrecorder files 5b. Transient faults are temporary f aultswhich are caused byan intermittenterror.Furthermore, the live data 5 comprises logged event s 5c. Thelogged events5care criticaleventsgenerating an errormessage. As an example, once a signal exceeds a pre -definedthreshold level,an eventistriggered and logged. The live data 5 maycomprise also furthertypesofoperational data to implement risk-based maintenanc eanalysis.The maintenance system 1 further uses system inform ation 8which can be stored in a memoryofthe maintenance system 1.The system information 8 comprises domain expertise on thespecific station design, its performance and operat ionalexperience. As an example, this includes informatio n of agingof pumps of the same or a similar design and also m aintenance P2023,0923 WO E / P220328WO01 January30,2024 -8 -and mitigation suggestions for preventing or elimin atingfaults.The system information 8 can be regularly updated a nd mayinclude the behaviorofrespective componentsorsubcomponents in a monitored fleet of converter sta tions. Inaddition to that, the system information 8 may also comprisestatic reliability and / or availability parameters f romoriginal equipment manufacturers.Every subcomponent 7a 1-7f has attributed reliability and / oravailabilityparametersindicating the severityof a failureand guiding the maintenance plans. Such reliability andavailability parameters include failure rate (FR) ( frequencyofwhich a device failswithin a specified period) and meantime to restore (MTTR), i.e., the time required for restoringa component.From the reliabilityand availability parametersof the subcomponents 7a 1-7f, also reliability andavailability parameters of the components 3a, 3b, 3 c can becalculated. Therefrom, a shutdown risk on system le vel, i.e.forthe entire powerstation can be calculated.The system information 8 may further include a list offailure modes and a corresponding maintenance task. As anexample, failure modes and maintenance tasks for th e givenexample maybe: •Failure mode 1: “Remaining Useful Life (RUL)- Beari ngfailure due to vibrations” / Reliability-Centered Maintenance 1: “1 year conditi on-based:Measure vibration” P2023,0923 WO E / P220328WO01 January30,2024 -9 - • Failure mode 2: “RUL -Insufficientlubrication of bearing due to defect gasket” / Reliability-Centere dMaintenance 2: “1 yearcondition-based:Measure vibration” • Failure mode 3: “RUL -Bearing failure in pump due to contaminated lubrication” / Reliability-Centered Maintenance 3: “3 month predetermined:Lubrication of bearings”. Ascan be seen from the failure modes,the control device 2ain the form of a vibration sensor addresses specifi c failuremodes 1 and 2.The live data 5 and system information 8 provides t he inputto a risk-analyzing algorithm 9. The risk-analyzing algorithm 9 comprisesproactive riskmanagement 10 and reactive risk management 11. Gene rally,proactive risk management 10 refers to risk managem ent beforean event for a specific component or subcomponent i striggered and reactive risk management 11 refers to riskmanagementafteran eventfora specificcomponent or subcomponentistriggered.Reactive risk management 11 can be based on live da ta 5 inthe form of logged events 5c and system information 8, forexample.Asan example,a logged event5cmaybe a failure ofone of the pumps 6 1, 6 2 or an event where a sensor signal,e.g. provided in a time series 5a, exceeds a thresh old levelorisclassified asabnormal. P2023,0923 WO E / P220328WO01 January30,2024 -10 -Proactive risk management 10 can be based on live d ata 5 inthe form oftime series5a,forexample.Proactive riskmanagement 10 can be used to predict a failure befo re aneventistriggered.Both proactive risk management 10 and reactive riskmanagement 11 can result in information and / or alar msprovided to the user on a front-end platform 12. As anexample, the front-end platform 12 is a user-interf ace suchas a monitor on which risk factors are displayed, m aintenancerecommendations are given and / or an alarm is trigge red.As an example, the system information 8 may include areliability diagram of a failure rate FR and mean-t ime-to-restore MTTR for each subcomponent 7a 1-7f and each component3a,3b,3c.For the given example, FR and MTTR may be as follow s:First pump 6 1: FR= 3.84E-5 / h; MTTR = 170 hSecond pump 6 2: FR= 3.84E-5 / h; MTTR = 170 hPLC 7c:FR= 1.00E-6 / h;MTTR = 24-48 hCooling pump arrangement 3a: FR = 8.87E-3 / y; MTTR = 170 h.From the reliability and availability parameters FR and MTTRof the individual components 3a, 3b, 3c, 6 1, 62 and / orsubcomponents 7a 1-7f, system reliability and availabilityparameters FR including a shutdown risk SR and down time riskDR for the entire system can be calculated and disp layed onthe front-end platform 12. As an example, when an e rror islogged, associated risk factors can be displayed on thefront-end platform 12. P2023,0923 WO E / P220328WO01 January30,2024 -11 -The shutdown risk SR for the entire system is calcu lated asthe product of FR and MTTR in respect of a non-redu ndantcomponent or subcomponent. The downtime risk DR for theentire system is the same as the MTTR of the respec tivecomponent.The reliability and availability parameters FR of a component3a, 3b, 3c and / or subcomponent 7a 1-7f can be initially basedon static design information. As the operation stat us of thesubcomponents 7a 1-7f may shift from original designs over theoperation time, especially after events are trigger ed, pre-defined static reliability and availability paramet ers asprovided in the initial reliability and availabilit y diagrammay notapplyanymore.With the aid of domain expertise and machine learni ng, theparameters are re-calculated dynamically when a tri ggeredevent occurs and / or proactively during continuous m onitoring.As an example, when one of the pumps 6 1, 6 2 has failed and anerror is triggered, the shutdown risk for the entir e systemchanges, as a failure of the other one of the pumps 61, 6 2will lead to a shutdown of the entire system. In ca se of anerror,the relevantupdated riskparameterscan be displayed,and mitigation recommendations can be given. The ri skmitigation recommendations are pre-programmed and a re basedon domain expertknowledge.Figure 3 shows an example of a front-end platform 1 2displaying an error message E, risk levels SR, DR a ndsuggested mitigation actionsMS.The front-end platform 12 may comprise both compone nts ofreactive risk management 11, i.e., an error message E and P2023,0923 WO E / P220328WO01 January30,2024 -12 -proactive risk management 10, i.e., updated risk le vels SR,DR.Also the recommended mitigation actionsMS may comprisecomponents of reactive risk management 11 such as r ecommendedactionsformitigating the errorand componentsof proactive riskmanagement10 such asrecommended actionsfor avoiding future errors.The front-end platform 12 may be also used in proac tive riskmanagement 10 without that an event is triggered. G enerally,output for reactive risk management 11 takes preced ence overoutputforproactive riskmanagement10.In the case of reactive risk management 11, the fro nt-endplatform 12 displays the occurrence of a logged eve nt 5c. Asan example, an error message E may be provided with theinformation that one of the pumps 6 1, 6 2 is faulty and analarm maybe triggered.The front-end platform 12 may further display the u pdatedshutdown risk SR for the entire system and the mean time torestore MTTR for different time intervals. As an ex ample, theshutdown riskSR fora time intervalofthe next2 daysmaybe calculated and displayed as 1%, for the next wee k as 4 %and for the next month as 16 %. The downtime risk D R may becalculated and displayed as 1.8 h for the next 2 da ys, as 6.5hours for the next week and as 26.7 hours for the n extmonths.In addition to that, a more detailed description of thetriggered event and a mitigation suggestion MS is d isplayed.As an example, the following information can be dis played:“Pump faulty. Check pump indications. A pump failur e has aconsiderable availability impact as there is an inc reased P2023,0923 WO E / P220328WO01 January30,2024 -13 -risk that the 2 nd pump will also fail. Contact maintenanceforfurthersteps.”Accordingly, operators are guided for instant actio ns tomitigate the occurred error,informed on potential futurerisks and guided for maintenance actions for avoidi ng futureerrors. Based on the updated likelihood and mitigat ionactions the operator can then decide whether to sen dmaintenance staffto the site based on the updated likelihoodand can take rapid actions following mitigation sug gestionsasthe firstresponse to reduce the shutdown risk. Accordingly,the maintenance system 1 calculatesa shutdownrisk SR for the entire system, based on triggered e ventsand / or analyzed live data 5 for subcomponents 7 a1, 7 f .Thereby,the overallriskforthe entire system is predictedand a maintenance action can be prioritized, accord ingly. Themaintenance system 1 provides a mitigation suggesti on MS atsubcomponentlevel.Thus,the maintenance system 1 usesa down-top and top-down approach.Figure 4 shows an embodiment of failure detection b y amachine-learning method in a schematicflow chart. Theprocess can be used both for reactive risk manageme nt 11 fordetecting an error and for proactive risk managemen t 10 forpredicting future errors. This may include updating failurerates FR of components 6 1, 6 2 and subcomponents 7 a1- 7 f .In the shown embodiment, a cooling pump arrangement 3a ismonitored by one or more monitoring devices 2a in t he form ofacoustic sensors for measuring vibrations of a firs t bearing7a1 of a first cooling pump 6 1 and a second bearing 7 a2 of thesecond cooling pump 6 2. P2023,0923 WO E / P220328WO01 January30,2024 -14 -The time series 5a comprises spectrograms of a meas uredacoustic signal, i.e. the levels of different frequ enciesover time. The spectrograms can be determined by us ing FastFourier Transform of acoustic signals. As an exampl e, theupper spectrograms can be associated to the first b earing 7 a1for different time intervals and the lower spectrog rams canbe associated to the second bearing 7 a2 for different timeintervals.The time series 5a are provided as an input to a ma chine-learning algorithm 14. The machine-learning algorit hm 14comprisesan encoding step,analyzing the inputin latentspace and a decoding step. The result of the algori thm may bean anomaly score AS for the input data 5, leading t o that anerror E is triggered, to an update of the failure r ates FR ora mitigation suggestion MS,forexample.Figures 5A-5C show steps in an analysis of live dat a 5 forfailure detection by a machine-learning method 14 i nschematicdiagrams.As shown in Figure 5A, the time series 5a of Fig. 4 can befurther processed into spectra for different operat ing speedsof the cooling pumps 6 1, 6 2. The generation of the spectra isone of the encoding steps carried out by the auto-e ncoder 13.The spectra are determined for operating speeds, e. g. of 1500RPM, 3000 RPM, 2500 RPM (upper curves) and a pump i n an OFF-status (lowest curve), wherein RPM denotes the numb er ofrevolutionsperminute. P2023,0923 WO E / P220328WO01 January30,2024 -15 -As illustrated in Figure 5B, the high dimension fea tures ofthe spectrogram are compressed byauto-encodersto extractlatent attributes. In this example, the input dimen sion is512 and latent representation is 4-dimensional. Pri nciplecomponentanalysisisapplied to reduce the latent features visualized in the main componentsPC1,PC2.With the compressed data, an anomaly score is calcu lated,specifying the difference between the output G(x i ) of amachine learning model and the ideal output y i with the giveninput x i . A threshold v is calculated as the 99.8 %percentile of the normally distributed reconstructi on errorsofthe normalpatterns:v=P99.8 ((y i - G(x i )) 2).The threshold v is then used to detect anomalous op erationpatterns.Figure 5C shows a diagram of an anomaly score AS de terminedforthe differentpackagesn ofthe time series5. Thethreshold v discriminates normal data from abnormal data.When a high anomaly score AS has been calculated du ring aspecified time interval, an alarm can be triggered, an errormessage E can be displayed, mitigation actions MS c an berecommended and or the failure rates FR can be upda ted.Training data setsformachine-learning modelscan begenerated by monitoring respective components 6 1, 62 andsubcomponents 7a 1-7f and their operational status over time.As an example, in case of bearings 7a 1, 7a2, acoustic data iscollected and an error occurring or not occurring w ithinpredetermined time intervals is attributed to the r espective P2023,0923 WO E / P220328WO01 January30,2024 -16 - data sets.The training data setsmaybe collected also from fleetsofpowerstations. Generally,the maintenance system 1 maycomprise a processorconfigured for carrying out the analysis in particu lar therisk analyzing algorithm 9 comprising the machine-l earningalgorithm 14. The maintenance system 1 may comprise a memorystoring the information 8 and the machine-learning model.Themaintenance system 1 may further comprise a communi cationdevice for exchanging data, in particular live data 5, with afleetofpowerconverterstations.The maintenance system 1 may be a central system fo r risk-based maintenance of several power converter statio ns or alocal system for risk-based maintenance of a single powerconverter station. Also for a local system, data ex changewith other power converter stations may be provided . Themaintenance system 1 can be also a cloud-based syst em.

[0002] P2023,0923 WO E / P220328WO01 January30,2024 -17 -Reference Signs1 maintenance system 2a,2b,2cmonitoring device 3a cooling pump arrangement 3b heatexchangers 3c furthercomponents 4 cooling system 5 live data 5a time series 5b transientfaultrecorderfiles 5c logged events 61firstcooling pump 62second cooling pump 7a1 firstbearing 7a2second bearing 7b PLC (programmable logiccontroller) 7c UPS (Uninterruptible PowerSupply) 7d pressure sensor 7e flow sensor 7f AC powersupply 8 system information 9 riskanalyzing algorithm 10 proactive riskmanagement 11 reactive riskmanagement 12 front-end platform 13 auto-encoder 14 machine-learning algorithm FR failure rate MTTR mean time to restore SR shutdown risk DR downtime risk E errormessage P2023,0923 WO E / P220328WO01 January30,2024 -18 - MS mitigation suggestion AS anomalyscore

Claims

P2023,0923 WO E / P220328WO01 January30,2024 -19 - Claims1. A maintenance system (1) for improving the relia bilityand / or availability of a power converter station (1 00) byrisk-based maintenance,being configured to analyze live data (5) from moni toringdevices (2a, 2b, 2c), which monitor subcomponents ( 7a1-7f) ofcomponents (3a, 3b, 3c, 6 1, 6 2) of the power converterstation (100), wherein the live data (5)isanalyzed forreactive riskmanagement (11), which comprises alerting a user wh en anerror event (E) is detected, and for proactive riskmanagement (10), which comprises calculating one or morereliability and / or availability parameters (RF, MTT R) for thesubcomponents (7a 1-7f) and / or components (3a, 3b, 3c, 6 1, 6 2)and comprises calculating one or more system reliab ilityand / or availability parameters (SR, DR) for the ent ire powerconverter station (100) based on the reliability an d / oravailability parameters (RF, MTTR) for the subcompo nents(7a 1-7f) and / or components (3a, 3b, 3c, 6 1, 6 2) ,being configured for alerting a user and / or providi ngmitigation suggestions (MS) on the level of the sub components(7a 1-7f).2.The maintenance system (1)ofclaim 1,wherein reactive risk management (11) and / or proact ive riskmanagement (10) is based on a machine-learning algo rithm(14).

3. The maintenance system (1) of any of the precedi ng claims,comprising a front-end platform (12) for alerting a userand / orproviding the mitigation suggestions(MS).

4. The maintenance system (1) of any of the precedi ng claims,P2023,0923 WO E / P220328WO01 January30,2024 -20 -configured to inform a user on mitigations suggesti ons (MS)for mitigating an occurred error (E) in reactive ri skmaintenance (11).

5. The maintenance system (1) of any of the precedi ng claims,configured to inform a user on mitigations suggesti ons (MS)forpreventing a predicted errorin proactive risk maintenance (10).

6. The maintenance system (1) of any of the precedi ng claims,wherein the machine-learning algorithm (14) is conf igured touse training data from a fleet of power converter s tations(100).

7. The maintenance system (1) of any of the precedi ng claims,being configured for generating an alarm when a shu tdown risk(SR) for the entire power converter station (100) e xceeds aspecified threshold level. 8.The maintenance system (1)ofclaim 7, wherein live data (5)comprisesa time series(5a) ofmeasurements and wherein the machine-learning algor ithm (14)is configured to determine an anomaly scores (AS) f or thetime series(5a). 9.The maintenance system (1)ofclaim 8,wherein the anomaly score (AS) determines when an e rror (E)istriggered and / orisused forthe calculation ofreliability and / or availability parameters (RF, MTT R).

10. The maintenance system (1) of any of the preced ingdevice, configured for monitoring a bearing (7a 1, 7a 2) of acooling pump (6 1, 6 2), wherein a monitoring device (2a)P2023,0923 WO E / P220328WO01 January30,2024 -21 -provides live data (5a) in the form of acoustic mea surementsof the bearing (7a 1, 7a 2).11.The maintenance system ofclaim 10,configured to detect abnormal patterns in the acous ticmeasurementsin a machine-learning algorithm (14). 12.A powerconverterstation (100)comprising themaintenance system (1) of any of the preceding clai ms.13.A method forimproving the reliabilityand / or availabilityofa powerconverterstation (100)by risk-based maintenance,comprises analyzing live data (5) from monitoring d evices(2a, 2b, 2c), which monitor subcomponents (7a 1-7f) ofcomponents (3a, 3b, 3c, 6 1, 6 2) of the power converterstation (100), wherein the live data (5)isanalyzed forreactive riskmanagement (11), which comprises alerting a user wh en anerror event (E) is detected, and for proactive riskmanagement (10), which comprises calculating one or morereliability and / or availability parameters (RF, MTT R) for thesubcomponents (7a 1-7f) and / or components (3a, 3b, 3c, 6 1, 6 2)and one or more system reliability and / or availabil ityparameters(SR,DR)forthe entire powerconverter station (100)based on the reliabilityand / oravailability parameters(RF, MTTR) for the subcomponents (7a 1-7f) and / or components(3a, 3b, 3c, 6 1, 6 2),and comprising alerting a user and / or providing mit igationsuggestions (MS) on the level of the subcomponents (7a 1-7f).

Images