Methods and computing systems for performing predictive health analysis on assets

Abstract

To provide a method for executing a predicted soundness analysis for an asset, and a computing system.SOLUTION: A power system 10 includes local controllers 21 to 23, a central system 20, and / or a remote server system 24. For assets 11 to 13, a plurality of independent probabilistic simulations are performed by using the transition probability of a discrete Markov chain model. From the plurality of independent probabilistic simulations, the expansion of the state of a predicted asset soundness over a certain period is calculated. On the basis of the calculated expansion of the state of the predicted asset soundness, an output is generated.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] Field of Invention This invention relates to a technology for evaluating the soundness of assets. In particular, this invention relates to a method and apparatus for predictive evaluation of asset soundness. [Background technology]

[0002] Background of the Invention Power systems, including power generation, transmission, and / or distribution systems, as well as industrial systems, include assets. Transformers, generators, and distributed energy resource (DER) units are examples of such assets. These assets degrade during operation. For planning purposes, and for scheduling maintenance or replacement work, it is desirable to estimate the remaining useful life (RUL) of the assets.

[0003] RUL estimation can be performed based on sensor data from a group of assets of the same or similar type as the asset being estimated. By labeling the sensor data with fault signatures, it can be indicated whether the sensor data corresponds to the asset's normal functioning state, degraded state, or faulty state. Combining such sensor data for RUL estimation can be challenging, as different types of sensor data may be available for different assets within a group of assets, for example, depending on the manufacturer or the auxiliary sensors installed on the asset.

[0004] Y. Yu et al., “Remaining Useful Life Prediction Using Elliptical Basis Function Network and Markov Chain”, World Academy of Science, Engineering and Technology, 47, 2010, describes a method for predicting remaining useful life using an elliptic basis function (EBF) network and a Markov chain. To account for missing covariates, the EBF structure is trained using a modified expectation-maximization (EM) algorithm. The Markov chain is constructed to represent the evolution of the external covariates, but explicit extrapolation to the internal covariates is not required. [Overview of the project] [Means for solving the problem]

[0005] overview There is a need for improved technology to predict the temporal evolution of asset deterioration. In particular, there is a need for technology that can predict the deterioration of asset health without necessarily requiring the capture of sensor data for the specific asset being predicted. Alternatively, or in addition to this, there is a need for technology that can use historical sensor data from a group of assets for prediction, even when different types of sensor data are available for various assets within a group of assets.

[0006] Embodiments of the present invention are provided, as are the methods and computing systems described in the independent claims. Dependent claims clarify preferred embodiments.

[0007] According to embodiments of the present invention, asset performance degradation can be predicted by using a discrete Markov chain model. By employing the Markov Chain Monte Carlo (MCMC) method, numerous simulations can be performed. A Markov chain model may have a set of discrete states corresponding to various asset health states.

[0008] In some cases, a large amount of captured historical sensor data can be condensed into a small number of transition probabilities. Specifically, a Markov chain model can be configured such that only two or three transition probabilities determine the transitions between states, which can represent a healthy asset state, a degraded asset state where the asset is still functioning, and a failed asset state where the asset has failed. The present invention can also be applied when little or no historical data is available. In such cases, the transition probabilities of the Markov chain model may be set by an expert.

[0009] The transition probabilities between states in a Markov chain model can be determined based on sensor data captured for a group of assets, which are labeled with failure signatures.

[0010] Different transition probabilities can be used in simulations run in parallel. These different transition probabilities can be associated with different operating conditions and / or ambient conditions.

[0011] Information regarding variance, confidence intervals, or other reliability indicators can be obtained from the simulation.

[0012] Using MCMC techniques with a small state space and a small number of transition probabilities, predictive asset health analysis can be performed over various periods, including periods of several years or more. While various other simulation methods aimed at more detailed modeling of asset behavior may be suitable for assessing asset degradation on short time scales, they may be affected by instabilities over longer periods relevant to the health analysis of power system assets or industrial assets.

[0013] A method for performing predictive health analysis on an asset, particularly for determining the remaining useful life (RUL) or the probability of failure (PoF), according to an embodiment, includes performing a plurality of independent probabilistic simulations using the transition probabilities of a discrete Markov chain model. The discrete Markov chain model has a state space that includes a set of asset health states. Each of the plurality of independent probabilistic simulations simulates the future evolution in the state space of the discrete Markov chain model over a prediction period. The method may further include calculating the development of the predicted asset health state over the prediction period from the plurality of independent probabilistic simulations. The method further includes generating an output based on the calculated development of the predicted asset health state.

[0014] The asset may be a power system asset. The asset may be an industrial asset.

[0015] Calculating the development of the predicted asset health state may include calculating the RUL. Calculating the development of the predicted asset health state may include calculating the probability of failure of the asset over the prediction period as a function of time.

[0016] The method may further include calculating, from the plurality of independent probabilistic simulations, confidence information about the development of the predicted asset health state as a function of time over the prediction period.

[0017] The above output may further be generated based on reliability information. The above reliability information may include the time evolution of the confidence interval over the prediction period.

[0018] The above method further includes calculating, as a function of time over the prediction period, dispersion information about the development of the predicted asset health state from the above independent multiple probabilistic simulations.

[0019] The above output may further be generated based on the above dispersion information. The above dispersion information may include the time evolution of the dispersion interval over the prediction period.

[0020] The above reliability information or dispersion information may include the time evolution of the lower limit and the time evolution of the upper limit. The above lower limit can be associated with a first set of transition probabilities, and the above upper limit can be associated with a second set of transition probabilities different from the first set of transition probabilities.

[0021] The above output may include an expression of the development of the asset health state, an alarm or warning generated based on the development of the asset health state, and / or a control signal for controlling the operation of the asset based on the development of the asset health state.

[0022] This method may include deriving a first set of transition probabilities from first sensor data of an asset operating under a first condition (e.g., a first operating condition and / or ambient conditions), and deriving a second set of transition probabilities from second sensor data of an asset operating under a second condition different from the first operating condition (e.g., a second operating condition and / or ambient conditions).

[0023] The above state space may be composed of three states, four states, or five or more states.

[0024] The above state space may include at least one state in which the operation of the above asset is not affected by the adverse effects of a failure.

[0025] The above state space may include at least one state in which the operation of the asset is adversely affected by a failure, but the asset continues to operate.

[0026] The above state space may include states in which the above asset is not functioning due to a failure. Calculating the development of the predicted asset health state described above may involve calculating the probability distribution in the state space multiple times over the forecast period and mapping the probability distribution to a scalar.

[0027] The above scalar may be the probability, determined by the above-mentioned independent simulations, that the asset will become non-functional due to a failure.

[0028] The development of the predicted asset health state described above may be obtained as the time evolution of the scalar described above. The above method may further include determining transition probabilities from historical data, which includes sensor data for multiple assets.

[0029] The sensor data described above may be labeled with a fault signature that indicates which state in the state space of the Markov chain model each asset was in at what point in time.

[0030] The above fault signatures may be calculated automatically, assigned based on input from an expert, or assigned by a combination of the above measures.

[0031] Determining the transition probabilities described above may involve calculating a time-dependent scalar function from sensor data for the multiple assets mentioned above.

[0032] The above scalar function may also be calculated using a heuristic that takes sensor measurements as input and outputs the above scalar function.

[0033] Determining the above transition probabilities may involve identifying transitions within the state space of the Markov chain model based on the above time-dependent scalar function.

[0034] Determining the above transition probabilities may involve calculating the transition probabilities based on the transitions within the state space of the Markov chain model described above.

[0035] The above steps may include comparing the above scalar function with one or more thresholds.

[0036] The above scalar function may also represent a health index indicating the degree of deterioration. The calculation of the above transition probabilities may be performed independently for different groups of sensor data representing different operating and / or ambient conditions of the above-mentioned group of assets.

[0037] The above method may include determining the transition probabilities of multiple sets, each associated with different operating conditions and / or ambient conditions.

[0038] Transition probabilities for multiple sets can be determined for different operating and / or ambient conditions, while the state space of the Markov chain model may remain unchanged.

[0039] The above method may include selecting one of several sets of transition probabilities for performing a simulation in accordance with the above operating conditions and / or ambient conditions that are intended to be conditions for the above asset.

[0040] The above method may include performing the above-described independent probabilistic simulations such that one or more first simulations are performed using a first set of transition probabilities associated with a first operating condition and / or a first surrounding condition, and one or more second simulations are performed using a second set of transition probabilities associated with a second operating condition and / or a second surrounding condition, wherein the first transition probabilities are different from the second transition probabilities.

[0041] The above-mentioned different operating conditions may, but are not limited to, different loads, different voltages, different currents, different operating points, and different insulating fluids.

[0042] The above-mentioned different ambient conditions may, but are not limited to, different temperatures and different relative humidity.

[0043] The above-mentioned independent probabilistic simulations may also be Markov chain Monte Carlo (MCMC) simulations.

[0044] The above Markov chain may be homogeneous. The above transition probabilities may be independent of time.

[0045] The above Markov chain may have a degree of 1; that is, transitions depend on the current state of the Markov chain model but may be independent of past transitions to that state.

[0046] The above Markov chain model may have states that can have monotonically arranged qualitative interpretations, that is, it is always possible to compare two states in terms of their degree of degradation.

[0047] Each state in the state space can have a non-zero transition probability to at most one other state in the state space. This explains why the degree of degradation is greater and why the transition probability to itself is non-zero.

[0048] The above Markov chain model may also be such that each state in the state space that does not correspond to an asset failure has a non-zero transition probability to just one other state in the state space.

[0049] The above Markov chain model may also be one in which the states in the state space corresponding to asset failures do not have non-zero transition probabilities to other states.

[0050] The above Markov chain model may also be a finite Markov chain model. The above state space may consist of three states, four states, or five or more states.

[0051] The above state space may consist of n states (where n is equal to 3, 4, or 5 or more), and for the n-1 states of the state space corresponding to a functioning asset, the transition probability to any other state in the state space is a non-zero transition probability, and for the state corresponding to a failed asset, the transition probability to any other state in the state space is not a non-zero transition probability.

[0052] The above method may further include receiving sensor measurement data captured during the operation of the asset.

[0053] The above method may further include updating the development of the predicted asset health status based on the received sensor measurement data.

[0054] The above-mentioned simulations may include simulations for different ambient and / or operating scenarios.

[0055] The above multiple simulations may be executed in parallel. The above multiple simulations may be run simultaneously.

[0056] The above method may also be implemented by a computer. The above method may be carried out by at least one integrated circuit.

[0057] The above method may be performed by at least one integrated circuit of the central controller of a distributed control system.

[0058] The above method may be performed by at least one integrated circuit of a local controller in a distributed control system.

[0059] The above method may include at least one integrated circuit receiving information about transition probabilities through a communication network.

[0060] The above assets may be power transformers, distributed energy resource (DER) units, or generators.

[0061] The above forecast period may be 1 year or more, 2 years or more, 3 years or more, 4 years or more, 5 years or more, 10 years or more, 15 years or more, or 20 years or more.

[0062] The above forecast period may be one week or longer, one month or longer, etc. The above forecast period may be measured over multiple cycles, which may include, for example, a certain number of aircraft operation cycles, shipping route cycles, train route cycles, etc.

[0063] A method for operating and / or maintaining an asset includes performing a predictive asset health analysis on the asset using the method described in one embodiment, and performing an automatic control or output operation based on the predictive asset health analysis.

[0064] The above control or output operations may include at least one of the following: generating an alarm or warning based on the calculated predicted asset health state development; generating a control signal for asset control operations based on the calculated predicted asset health state development; scheduling asset downtime based on the calculated predicted asset health state development; scheduling maintenance work based on the calculated predicted asset health state development; scheduling replacement work based on the calculated predicted asset health state development; or changing the maintenance interval based on the calculated predicted asset health state development.

[0065] The above control or output operation may include outputting information via the interface about the probability of failure as a function of operating time, information about scheduled or rescheduled maintenance intervals, or information about scheduled replacement intervals.

[0066] A computing system according to the present invention, which functions to perform predictive health analysis on an asset, is configured to perform a plurality of independent probabilistic simulations using the transition probabilities of a discrete Markov chain model, the discrete Markov chain model having a state space containing a set of asset health states, and each of the plurality of independent probabilistic simulations simulates the future development of the state space of the discrete Markov chain model over a forecast period. The computing system is configured to compute the development of the predicted asset health state over the forecast period from the plurality of independent probabilistic simulations. The computing system is configured to control the generation of an output based on the computed development of the predicted asset health state.

[0067] The computing system described above may include a local controller having one or more integrated circuits that function to perform the independent probabilistic simulations described above and to calculate the development of the predicted asset health state described above. One or more of the ICs described above may function to perform the operations detailed herein.

[0068] The computing system described above may include a central controller having one or more integrated circuits that function to perform the independent probabilistic simulations described above and to calculate the development of the predicted asset health state described above. One or more of the ICs described above may function to perform the operations detailed herein.

[0069] The above assets may also be power system assets. The above assets may also be industrial assets.

[0070] The computing system described above may function such that calculating the development of the predicted asset health state may include calculating the RUL.

[0071] The computing system described above may function such that calculating the development of the predicted asset health state includes calculating the probability of asset failure as a function of time over the forecast period.

[0072] The computing system described above may function to calculate confidence information about the development of the predicted asset health state as a function of time over a forecast period, based on the above-mentioned independent probabilistic simulations.

[0073] The computing system described above may function such that the output can be further generated based on the confidence information described above.

[0074] The computing system described above may function such that the confidence information includes the time evolution of the confidence interval over the forecast period.

[0075] The computing system described above may function to calculate distributed information about the development of the predicted asset health state as a function of time over the forecast period, based on the above-mentioned independent probabilistic simulations.

[0076] The computing system described above may function in such a way that the output can be further generated based on distributed information.

[0077] The computing system described above may function such that the output may include a representation of the development of the asset health state, alarms or warnings generated based on the development of the asset health state, and / or control signals for controlling the operation of the asset based on the development of the asset health state.

[0078] The above computing system may function such that the distributed information includes the time evolution of the distributed interval over the forecast period.

[0079] The computing system described above may function such that the confidence information or distributed information may include a lower time evolution and an upper time evolution.

[0080] The computing system described above may function such that the lower bound can be associated with a first set of transition probabilities, and the upper bound can be associated with a second set of transition probabilities that is different from the first set of transition probabilities.

[0081] The computing system described above may function to derive a first set of transition probabilities from first sensor data of an asset operating under a first condition (e.g., a first operating condition and / or ambient condition), and to derive a second set of transition probabilities from second sensor data of an asset operating under a second condition different from the first operating condition (e.g., a second operating condition and / or ambient condition).

[0082] The computing system described above may function such that the state space can consist of three states, four states, or five or more states.

[0083] The computing system described above may function such that the state space may include at least one state in which the operation of the asset is not adversely affected by a failure.

[0084] The computing system described above may function such that the state space may include at least one state in which the operation of the asset is adversely affected by a failure, but the asset continues to operate.

[0085] The computing system described above may function such that the state space may include a state in which the asset is not functioning due to a failure.

[0086] The computing system described above may function such that calculating the development of the predicted asset health state may include calculating the probability distribution in the state space multiple times over the forecast period and mapping the probability distribution to a scalar.

[0087] The computing system described above may function such that the scalar can be the probability, determined by the multiple independent simulations, that the asset will become non-functional due to a failure.

[0088] The computing system described above may function such that the development of the predicted asset health state is obtained as the time evolution of the scalar described above.

[0089] The computing system described above may function to determine transition probabilities from historical data, including sensor data for multiple assets.

[0090] The computing system described above may function to label the sensor data with a fault signature indicating which state in the state space of the Markov chain model each asset was in at a given time.

[0091] The computing system described above may function to automatically calculate the fault signature or to receive input from an expert used to determine the fault signature.

[0092] The computing system described above may function such that determining the transition probabilities may involve calculating a time-dependent scalar function from sensor data for the multiple assets described above.

[0093] The computing system described above may function in such a way that it can calculate the scalar function using a heuristic that takes sensor measurements as input and outputs the scalar function.

[0094] The computing system described above may function such that determining the transition probabilities may involve identifying transitions in the state space of the Markov chain model based on the time-dependent scalar function described above.

[0095] The computing system described above may function such that determining the transition probabilities includes calculating the transition probabilities based on the transitions in the state space of the Markov chain model described above.

[0096] The computing system described above may function such that the steps described above include comparing the scalar function with one or more thresholds.

[0097] The above scalar function may also represent a health index indicating the degree of deterioration. The computing system described above may function such that the determination of the transition probabilities can be performed independently for different groups of sensor data representing different operating and / or ambient conditions of the set of assets.

[0098] The computing system described above may function to determine multiple sets of transition probabilities, each associated with different operating and / or ambient conditions.

[0099] Transition probabilities for multiple sets can be determined for different operating and / or ambient conditions, while the state space of the Markov chain model may remain unchanged.

[0100] The computing system described above may function to select one of several sets of transition probabilities for performing a simulation in accordance with the operating conditions and / or ambient conditions that are intended to be conditions for the assets described above.

[0101] The computing system described above may function to perform the above-mentioned independent probabilistic simulations such that one or more first simulations are performed using a first set of transition probabilities associated with a first operating condition and / or a first surrounding condition, and one or more second simulations are performed using a second set of transition probabilities associated with a second operating condition and / or a second surrounding condition, wherein the first transition probabilities are different from the second transition probabilities.

[0102] The computing system described above may function such that the different operating conditions may, but are not limited to, different loads, different voltages, different currents, different operating points, and different insulating fluids.

[0103] The above-mentioned different ambient conditions may, but are not limited to, different temperatures and different relative humidity.

[0104] The computing system described above may function such that the multiple independent probabilistic simulations described above can become Markov chain Monte Carlo (MCMC) simulations.

[0105] The computing system described above may function in such a way that the Markov chain is homogeneous. The transition probabilities may be independent of time.

[0106] The above computing system may function such that the Markov chain can have a degree of 1, that is, transitions depend on the current state of the Markov chain model but are independent of past transitions to that state.

[0107] The computing system described above may function such that each state in the state space may have a non-zero transition probability to at most one other state in the state space.

[0108] The above computing system may function such that the above Markov chain model can be a model in which states in the state space that do not correspond to asset failures have a non-zero transition probability to only one other state in the state space.

[0109] The above computing system may function such that the above Markov chain model can be a model in which the state space state corresponding to an asset failure does not have a non-zero transition probability to any other state.

[0110] The computing system described above may function such that the Markov chain model is a finite Markov chain model.

[0111] The computing system described above may function such that the state space can consist of four or three states.

[0112] The above computing system may have a state space consisting of n states (where n is equal to 3, 4, or 5 or more), and may function such that for the n-1 states in the state space corresponding to a functioning asset, the transition probability to any other state in the state space is a non-zero transition probability, and for the state corresponding to a failed asset, the transition probability to any other state in the state space is not a non-zero transition probability.

[0113] The computing system described above may also function to receive sensor measurement data captured during the operation of the asset described above.

[0114] The computing system described above may function to update the development of the predicted asset health status based on the received sensor measurement data.

[0115] The computing system described above may function such that the multiple simulations described above may include simulations for different environments and / or operating scenarios.

[0116] The computing system described above may function in such a way that the multiple simulations described above are executed in parallel.

[0117] The computing system described above may function in such a way that the multiple simulations described above are executed simultaneously.

[0118] The computing system described above has at least one integrated circuit that performs the operations described. The operations described may be performed in a distributed computing system, for example, a partitioned control system, and / or using a cloud-based computing system.

[0119] The computing system described above may have an interface for receiving information about transition probabilities via a communication network.

[0120] The above assets may be power transformers, distributed energy resource (DER) units, or generators.

[0121] The above forecast period may be 1 year or more, 2 years or more, 3 years or more, 4 years or more, 5 years or more, 10 years or more, 15 years or more, or 20 years or more.

[0122] The above forecast period may be one week or longer, one month or longer, etc. The above forecast period may be measured over multiple cycles, which may include, for example, a certain number of aircraft operation cycles, shipping route cycles, train route cycles, etc.

[0123] The asset control system includes a computing system for performing a predictive asset health analysis on an asset using the method described above according to one embodiment, and an output interface for automatically triggering or performing control or output actions based on the predictive asset health analysis.

[0124] The above control or output operations may include at least one of the following: generating an alarm or warning based on the calculated predicted asset health state development; generating a control signal for asset control operations based on the calculated predicted asset health state development; scheduling asset downtime based on the calculated predicted asset health state development; scheduling maintenance work based on the calculated predicted asset health state development; scheduling replacement work based on the calculated predicted asset health state development; or changing the maintenance interval based on the calculated predicted asset health state development.

[0125] The above control or output operation may include outputting information via the interface about the probability of failure as a function of operating time, information about scheduled or rescheduled maintenance intervals, or information about scheduled replacement intervals.

[0126] An industrial or power system includes assets and a computing system that performs predictive asset health analysis on those assets.

[0127] The computing system described above may be a distributed controller for the industrial or power system used to control assets.

[0128] Various effects and advantages are related to the present invention. When using a Markov chain model, if the model has a discrete state space and a non-zero transition probability to at most one other state of the Markov chain model corresponding to the next degree of degradation, very few parameters are required. This is especially true when the Markov chain model has a state space that can consist of a fairly small number of states (e.g., three or four states, but arbitrarily many states), in which case the number of transition probabilities between states required to perform this method is very small. Even when different types of sensor data are available for different assets in a group of assets, the transition probabilities can be set based on user input or automatically determined from sensor data with fault signatures.

[0129] RUL curves or other predictive asset health information can be obtained from sensor data of any dimension using probabilistic simulation techniques and based on Markov chain models.

[0130] The distribution of RUL curves or other predictive asset health information can be determined based on probabilistic simulations for different ambient and / or operating conditions. This allows for quantitative information not only about the predicted RUL curve but also its confidence interval. The precision and / or variance of the RUL distribution may be quantified and output.

[0131] High-speed computation can be achieved by integrating multiple probabilistic simulations into a single simulation step that allows these simulations to be executed in parallel.

[0132] Additional information may be generated, such as quantitative information about the variance or confidence interval of the RUL curve, which can be used in prescriptive tools or applications.

[0133] The subject matter of the present invention will be described in more detail with reference to the preferred specific embodiments shown in the accompanying drawings. [Brief explanation of the drawing]

[0134] [Figure 1] This is a schematic diagram of a power system having a computing system according to a certain embodiment. [Figure 2] This is a schematic diagram of a power system having a computing system according to a certain embodiment. [Figure 3] This figure shows the Markov chain model used in the embodiment. [Figure 4] This is a flowchart illustrating a method according to a certain embodiment. [Figure 5] This is a graph showing a specific example of output generated by a method and computing system according to a certain embodiment. [Figure 6] This is a bar graph illustrating the time evolution of state occupancy probabilities. [Figure 7] This is a graph showing a specific example of output generated by a method and computing system according to a certain embodiment. [Figure 8A] This graph shows the results of the predicted asset health analysis combined with the observed asset health status. [Figure 8B] This graph shows the results of the predicted asset health analysis combined with the observed asset health status. [Figure 9] This is a flowchart illustrating a method according to a certain embodiment. [Figure 10] This is a block diagram of a computing system according to a certain embodiment. [Modes for carrying out the invention]

[0135] Detailed description of the embodiment Embodiments of the present invention will be described with reference to drawings in which the same or similar reference numerals indicate the same or similar components. Some embodiments will be described in the context of assets of a power system such as a distributed energy resource (DER) unit or a transformer, but the embodiments are not limited thereto. Features of the embodiments can be combined with each other unless otherwise specified.

[0136] Figures 1 and 2 are schematic diagrams of power systems 10 and 15. Power systems 10 and 15 include multiple assets. These assets may include generators, transformers, or other power system assets such as distributed energy resource (DER) units 11-13, 16-18, etc.

[0137] The power systems 10 and 15 include a control system, which includes local controllers 21 to 23, each associated with an asset. The control system may include a central system 20. The central system 20 may be communicatively coupled to the local controllers. The central system 20 may be communicatively coupled to a remote (e.g., cloud-based) server system 24.

[0138] As will be described in more detail below, the local controllers 21-23, the central system 20, and / or the remote server system 24 may function to perform predictive asset health analysis using a Markov chain model. The Markov chain model may have a specific configuration as described below. The local controllers 21-23, the central system 20, and / or the remote server system 24 may function to perform predictive asset health analysis by running multiple independent probabilistic simulations, particularly Markov chain Monte Carlo (MCMC) simulations.

[0139] The results of the predictive asset health analysis can be used by local controllers 21-23, the central system 20, and / or the remote server system 24 to schedule downtime, maintenance work, replacement work, or to automatically perform control operations. The local controllers 21-23, the central system 20, and / or the remote server system 24 may function to generate and output control or output data. The output may be provided via a human-machine interface (HMI) 26. This HMI can be connected to the local controllers 21-23, the central system 20, and / or the remote server system 24 via the internet or another wide area network (WAN).

[0140] As will be explained in more detail with reference to Figures 3 to 10, the predictive asset health analysis may include simulating the time evolution of assets. MCMC simulations and further processing of the MCMC results can be performed on local controllers 21 to 23. Doing so facilitates the incorporation of local sensor measurements to update the predictive asset health analysis.

[0141] The techniques described herein can be used even when sensor data is unavailable for a particular asset for which predictive asset health analysis is being performed. Specifically, a remaining useful life (RUL) curve or other predictive asset health forecast can be calculated using sensor data and a historical data repository that includes failure information captured for similar assets, preferably assets of the same type as the asset for which predictive asset health analysis is being performed, even when sensor data is not yet available.

[0142] Figure 3 is a graph of a Markov chain model that can be used in the method and computing system according to the embodiment. The state space of the Markov chain model consists of a set of n states S1, ..., Sn. In this case, n=4. However, a state space with a different number of states (for example, n=3, n=5, or n>5) can be used instead.

[0143] The states in the state space may be arranged so that the degree of deterioration of asset health increases from S1 to S2, S2 to S3, and so on. In other words, all states except the last state of the Markov chain model may be followed by another state with a greater degree of deterioration. The last state of the Markov chain model can represent the greatest degree of deterioration.

[0144] In a Markov chain model, the 1st, 2nd, ... (n-1)th states 41-43 have a non-zero transition probability p to the only other state in the state space. 12 , p 23 , p 34 It can be set to have the following: The nth state 44 does not have a non-zero transition probability to any state other than this state.

[0145] As a concrete example, in a Markov chain model, the transition from the first state 41 to the second state 42 has a finite transition probability p. 12 However, the model can be such that the transition probability from the second state 42 back to the first state 41 is zero. The probability is 1-p 12 The first state 41 is maintained in the repeated probabilistic simulations.

[0146] In the Markov chain model, the transition from the second state 42 to the third state 43 has a finite transition probability p. 23 However, the model can be such that the transition probability from the third state 43 back to the second state 42 is zero. The probability is 1-p 23 The second state 42 is maintained in the repeated probabilistic simulations.

[0147] The Markov chain model can be a model where the finite transition probability p is from the third state 43 to the fourth state 44, but the transition probability from the fourth state 44 back to the third state 43 is zero. The third state 43 with a probability of 1 - p is maintained in the repetition of the probabilistic simulation. 34 The final state 44 of the Markov chain model can correspond to a state where the asset has failed to such an extent that it no longer functions. 34 The other states 41 - 43 of the state space can correspond to different degrees of deterioration.

[0148] As a specific example, the first state (which may also be called the "unknown" state S1) can correspond to an asset state where there is no known deterioration that affects the operation of the asset. As a specific example, the first state 41 can correspond to a state where no failure is recorded or a failure cannot be recorded.

[0149] The second state 42 (which may also be called the "initial" state S2) can correspond to a slightly detectable failure that does not immediately affect the performance of the asset. Such an initial failure is usually characterized by a short mean time to repair (MTTR), a low repair cost, and a small impact on the overall performance. If proper maintenance management is not carried out, the initial failure may develop into a more deteriorated failure. The third state 43 (which may also be called the "deterioration" state S3) can correspond to a mode that explains a failure that significantly degrades the performance of the system but does not immediately stop the asset. Usually, such a failure is caused by the deterioration of components. This deterioration will eventually become a fatal failure if not addressed.

[0150] The fourth state 44 (which may also be called the "fatal" state S4) can correspond to the most severe failure mode that immediately and completely stops the asset. Usually, this is characterized by a long and high - cost MTTR (due to total production loss).

[0151]

[0152] ​​

[0153] The transition probabilities of a Markov chain model may be obtained from an expert via a user interface, or they may be determined using historical data as described below. Transition probability p 12 , p 23 , p 34 Various sets of can be used. For example, for N>1, and especially for N>2, the transition probability p 12 , p 23 , p 34 By using various sets of these, it is possible to simulate the development of asset health states under different ambient and / or operating conditions.

[0154] Figure 4 is a flowchart of method 50 according to one embodiment. This method 50 may be automatically executed by one or more ICs in local controllers 21-23, a central system 20, and / or a remote server system 24.

[0155] In step 51, an MCMC simulation of the Markov chain model or other probabilistic simulation is performed. Step S51 may include performing more than 100, more than 1000, more than 2000, more than 5000, more than 10000, more than 50000, more than 100000, more than 500000, or more than one million simulations. As computing power increases, there is no upper limit on the number of simulations. Simulations may be performed in parallel.

[0156] Numerous simulations (e.g., more than 100 or more than 1000) of the transition probability p of a Markov chain model 12 , p 23 , p 34 This can be run for any set of values, but the transition probabilities do not need to be the same in all simulations. Transition probability p 12 , p 23 , p 34By using various sets of these, it is possible to quantitatively evaluate the effects of different operating conditions and / or ambient conditions.

[0157] Simulations can be run over a certain period of time. This period may depend on the specific asset. For power system assets such as transformers, the typical lifespan is more than 10 years, more than 20 years, or longer. Therefore, probabilistic simulations may be run over periods of more than 10 years, more than 20 years, or longer. Depending on the asset, this period may be shorter. Specifically, the forecast period may be more than one week, more than one month, etc. The forecast period may be measured over multiple cycles, which may include a certain number of aircraft operation cycles, shipping route cycles, train route cycles, etc.

[0158] The initial state of a simulation can be selected depending on the available information about the asset. If information about the asset is unavailable, all simulations may start from state 41, which has no information about detectable faults. If information about the asset is available, such as sensor data collected after installation, this sensor data can be used to initialize the simulation. The distribution of initial states for various MCMC or other probabilistic simulations can be selected depending on whether the sensor data already collected indicates that there are no recognizable faults affecting the asset's performance, or whether there are any detectable problems affecting the asset's performance.

[0159] Furthermore, initialization can be probabilistic. For example, if the available information is not definitive about whether an asset is in state 41 or 42, but rather that the probability of it being in either state is equal, then the system can be initialized with a Bayesian prior distribution such that the probability of the asset being in state 41 is 50% and the probability of the asset being in state 42 is 50%. Other probabilistic initializations with more states having different probabilities may also be possible.

[0160] In step 52, the results of the probabilistic simulation are processed. This may include calculating the probability, as a function of time over the above period, that the Markov chain model evolves into a fatal state corresponding to a non-functioning asset. The processing in step 53 may include calculating the time evolution of the health indicator at any point in the forecast period, depending on the probability that the Markov chain model is in the 1st, 2nd, ... nth state 41-44 of the Markov chain model. This processing may include calculating the RUL curve, or calculating another output that shows the probability of asset failure as a function of operating time.

[0161] An output may be generated in step 53. This output may include information about the remaining useful life of the asset as a function of time. This output may include information about the probability of failure of the asset as a function of time. This output may include control and / or output data obtained by further processing of the simulation results, such as scheduling of asset inspection, maintenance, or replacement work.

[0162] Figure 5 is a schematic diagram of the automatically generated and outputtable output 60. Output 60 can represent the probability that the Markov chain model will evolve into a fatal state corresponding to a non-functioning asset. Output 60 can be obtained by calculating the proportion of simulations in which the Markov chain model is in a fatal state S4 in the state space, out of multiple simulations, for each of the multiple time periods over the above period.

[0163] Other outputs or alternative outputs may be generated. For example, by processing RUL curves or other information indicating asset deterioration, inspection, maintenance, or replacement work can be automatically scheduled, and the scheduling information can be output to the operator and / or downtime can be automatically scheduled.

[0164] Alternatively, or in addition to this, the RUL curve, or other information indicating asset deterioration, may be processed using comparisons with thresholds or other triggers to determine whether and when an alarm, warning, or other signal should be issued to the operator.

[0165] Figure 6 shows the probabilistic distributions 61-64 of the population of various states S1-S4 in the state space of the Markov chain model. Distribution 61 corresponds to the first iteration, where most of the simulations of the Markov chain model are still in state S1, corresponding to an asset with no detectable degradation. Distributions 62 and 63 correspond to the subsequent second and third iterations, where initial or more advanced degradation increases. Distribution 64 corresponds to the fourth iteration, where the fatal state S4, corresponding to asset failure, becomes the most frequent, reflecting that up to that point, the probability of the asset being in a non-functional state is higher than the probability of it being in a functional state.

[0166] The probability of an asset reaching the fatal state S4, which is the final state of the Markov chain model, allows us to obtain a related predicted asset health forecast. However, the output, which is the result of processing a probabilistic simulation, is the probability p1, p2, ..., p obtained by the probabilistic simulation that the Markov chain model reaches the 1st, 2nd, ..., nth states. n It can depend on all of these.

[0167] As a concrete example, for any time j within the period during which a probabilistic simulation is performed, a scalar function

[0168]

number

[0169] It is possible to calculate p in the formula. i (j) represents the probability, obtained by probabilistic simulation, that the Markov chain model is in the i-th state at time j, and m iThis represents a scalar value that is a monotonic, especially strictly monotonic, function of the state label i. As a concrete example, all m i m1≦m2≦…≦m n , especially m1 <m2<…m n You may select from a certain section to achieve this result.

[0170] By outputting the function d(j) or the information derived from it, degradation that results in a reduction in RUL can be more adequately reflected, even if the asset has not reached the critical state S4.

[0171] If the function d(j) indicates degradation and is constrained to take values ​​between 0 and 1, then it can be associated with the health index h(j) by h(j) = 1 - d(j).

[0172] Figure 7 shows the output of curve 80, which represents the deterioration of assets as a function of time, obtained by probabilistic simulation. Curve 80 represents the probabilities p1, p2, ..., p of the Markov chain model to be in the 1st, 2nd, ..., nth state, obtained by probabilistic simulation. n It can depend on the development of all time.

[0173] Various states in a Markov chain model can be associated with multiple intervals 71-74. For example, if the health index h is within interval 71, the asset can be determined to be in state S1, where there is no known degradation. If the health index h is within interval 72, the asset can be determined to be in state S2, where there is no initial degradation that does not affect performance. If the health index h is within interval 73, the asset can be determined to be in state S3, where there is greater degradation that affects performance but does not lead to immediate asset shutdown. If the health index h is within interval 74, the asset can be determined to be in state S4, a critical state that leads to immediate asset shutdown.

[0174] Threshold TH1,...TH n-1This allows us to define the upper and lower bounds of the interval 71-74. When initializing the probabilistic simulation of assets, thresholds TH1, ... TH n-1 A comparison can be made with the following. For example, sensor data available for an asset may be processed and converted into a scalar representing the asset's health index h or deterioration index d=1-h, and this scalar can be used as a threshold TH1,...TH n-1 By comparing this, it is possible to determine how the simulation should be initialized.

[0175] By performing probabilistic simulations such as MCMC, it is possible to automatically calculate and output not only the development of the asset's health state, but also the reliability associated with that development.

[0176] Reliability information can take various forms. For example, the development of confidence intervals around curves 60 and 80 may be obtained as a function of time over the forecast period. The time development of the confidence interval can indicate the lower and upper bounds of the probability of fatal failure 60 or the health index h for any time j in the forecast period. The upper and lower bounds can be determined such that at least a certain percentage of the probabilistic simulations (e.g., at least 70%, 80%, 90%, or 95%) are within the range of the probability of fatal failure 60 or the health index h between the upper and lower bounds. Upper and lower bounds 81 and 82, as specific examples of the time development of the confidence interval, are shown in Figure 7.

[0177] Alternatively, or in addition to this, the upper and lower bounds 81 and 82 may reflect the variance in the behavior and / or surrounding conditions to which the asset may be subject. Specifically, each of the curves 80, 81, and 82, as explained with reference to Figure 3, can be obtained by performing multiple probabilistic simulations using Markov chain models with different sets of transition probabilities.

[0178] Figures 8A and 8B show the time evolution of the health index 80 obtained from probabilistic simulations, as well as the upper and lower bounds 81 and 82 representing the time evolution of the confidence interval. Curves 80-82 were obtained using transition probabilities derived from historical sensor data associated with failure signatures via MCMC, and the transitions between states S1-S4 can be identified in the historical sensor data.

[0179] Curves 84 and 86 represent the observed actual degradation of assets of the same type, but are not included in the historical data used to determine the transition probabilities of the Markov chain model. This degradation is shown as the evolution of a continuous function, which is the degradation index. Different states of the Markov chain model can be mapped to different ranges of the degradation index function.

[0180] The results of the predicted asset health analysis 80-82 are probabilistic results. The actual state of assets 84 and 86 may show a different development from curve 80 and / or confidence interval development 81 and 82, but the predicted asset health analysis results 80-82 reliably show the development of asset deterioration obtained for a large probabilistic sample of assets.

[0181] Figure 9 is a flowchart of method 90 according to one embodiment. This method 90 may be automatically executed by local controllers 21-23, a central system 20, and / or a remote server system 24. In one implementation, steps 91-92 of this method may be executed by the central system or the remote server system 24, and steps 93-94 may be executed by local controllers 21-23. In other variations, steps 91-94 may be distributed among multiple ICs of the processing system.

[0182] In step 91, sensor data is received. The sensor data may be historical sensor data of the same type of asset as the asset on which the predictive asset health analysis is performed (for example, a solar panel with a specific power rating range, a wind turbine with a specific power rating range, or a transformer with a rating for a specific section).

[0183] Sensor data may be associated with a period exceeding the length of the prediction period. The sensor data may be labeled sensor data that includes information about the state, for example, states S1 to Sn. Specifically, for any set of sensor data, there may be information that associates this sensor data with one of the states S1, ..., Sn of a Markov chain model.

[0184] If the sensor data does not contain a fault signature, step 91 may include receiving information from an expert via a user interface, assigning this sensor data to the states S1, ... Sn of a Markov chain model.

[0185] If the sensor data does not contain fault signatures, step 91 may include calculating a scalar function from the sensor data that represents a degradation or health indicator for each asset in the asset group, and determining the time at which transitions occurred between S1, ..., Sn based on the sensor data by comparing the scalar function with one or more thresholds (such as thresholds TH1 to TH3 in Figure 7).

[0186] The scalar function may also be calculated from sensor measurements using heuristics. A scalar function may take sensor measurements captured at various points in time as input, process these measurements to form a scalar function that represents the observed development of asset health, reflecting a health index h or a deterioration index d.

[0187] Various techniques can be used to compute the scalar function used to identify transitions between discrete states. Specifically, sensor measurements may be compared to a range of operating values. A penalty may be imposed for each sensor measurement outside this range. Weighted sums or other methods may be used, combining the product of a weighting coefficient for the sensor measurement with a value dependent on the deviation of the sensor measurement from the normal operating range. The weighting coefficient is determined by each sensor and indicates the importance of the asset health measurement.

[0188] Tools are known that provide mappings from sensor measurements to continuous health or degradation functions for a wide range of assets, including but not limited to circuit breakers, batteries (such as lithium-ion batteries), or transformers. For example, tools such as Ellipse APM or RelCare process sensor measurements to provide a function indicating asset health with values ​​within a continuous range. Normalization may be used to normalize the health or degradation function to a desired range (e.g., from 0 to 1).

[0189] The techniques disclosed herein allow any health or degradation function to be mapped to discrete states of a state model using arbitrary normalization and threshold comparisons.

[0190] In step 92, the transition probabilities of the Markov chain model may be determined. The transition probabilities may also be automatically determined from the sensor data and the associated state labels.

[0191] In one specific implementation, the transition probability may be calculated based on the condition probability. For example, the transition probability of the transition from the i-th state to the (i+1)-th state (1≦i≦n-1) at time j can be calculated as follows.

[0192]

number

[0193] In equation (2), the numerator represents the number of assets that were in the i-th state at time j and transitioned to the (i+1)-th state at time j+1. The denominator represents the number of assets that were in the i-th state at time j.

[0194] Transition probabilities may be set and / or adjusted based on user input. By performing averaging or other processing, a homogeneous Markov chain model can be obtained.

[0195] If sensor data is available for various groups of assets of the same type (e.g., solar panels with a specific power rating range, wind turbines with a specific power rating range, transformers with ratings for a specific section) but with different operating and / or ambient conditions, the transition probability may be determined independently for each of these groups.

[0196] In step 93, the RUL calculation may be performed using transition probabilities, or the time evolution of the degradation of one or more assets may be determined. This may involve obtaining the RUL curve by performing an MCMC simulation and then processing the results of the MCMC simulation, or by obtaining another measure of time-dependent degradation for one or more assets.

[0197] Step 93 may include receiving sensor data captured for the asset while the asset is in operation, and fitting the RUL calculation based on this sensor data once it becomes available. This may include updating the MCMC simulation based on this sensor data once it becomes available.

[0198] In step 94, control and / or output actions may be automatically performed based on the results of RUL calculations or other predictive asset health analyses.

[0199] As a concrete example, the RUL curve may be output. Information regarding the time evolution of confidence intervals or variance may also be output simultaneously.

[0200] Alternatively, or in addition to this, the operating point of the asset may be automatically adjusted by local controllers 21-23 associated with the asset.

[0201] Alternatively, or in addition to this, inspection, maintenance, and / or replacement work may be automatically scheduled.

[0202] Alternatively, or in addition to this, downtime for inspection, maintenance, and / or replacement work may be automatically scheduled.

[0203] Alternatively, or in addition to this, alarms, warnings, or other outputs may be generated for output via the HMI in response to the development of the RUL curve or other predictive asset health conditions.

[0204] Figure 10 is a schematic diagram of computing system 100. Computing system 100 may include one or more ICs 103. The ICs may be application-specific integrated circuits (ASICs), processors, controllers, field-programmable gate arrays (FPGAs), or a combination of multiple such integrated circuits.

[0205] IC103 may be located in the central system 20, in one of the local controllers 21-23, in the server system 24, or distributed among these entities.

[0206] IC103 may function to simulate the future development of assets using a Markov chain model by running a probabilistic simulation engine 104. The probabilistic simulation engine 104 may be adapted to perform MCMC simulations.

[0207] The transition probabilities of the Markov chain model used by the probabilistic simulation engine 104 can be received via interface 101 (for example, if IC 103 is one of the local controllers 21-23 and the central system 20 calculates the transition probabilities). IC 103 may calculate the transition probabilities based on historical sensor data of an asset group consisting of assets of the same type as the asset subject to the predicted asset health analysis. The historical sensor data can be received via interface 101 or stored locally in the data storage device 102.

[0208] IC103 may function to run a prediction engine 105. The prediction engine 105 may calculate an RUL curve or other prediction information associated with the development of asset health status based on the results of simulations performed by the probabilistic simulation engine 104.

[0209] IC103 may function to run output engine 106. Output engine 106 may generate output data or output signals for controlling the HMI and / or for implementing control operations on an asset or a system in which the asset is used. Specifically, output engine 106 may function to generate and output data to the HMI so that a RUL curve is output. Output engine 106 may function to generate and output data to the HMI so that information on the time evolution of confidence intervals or variance can be output simultaneously.

[0210] Alternatively, or in addition to this, the output engine 106 may function to automatically adjust its operating point in response to the output of the predictive engine 105.

[0211] Alternatively, or in addition to this, the output engine 106 may function to automatically generate and output information relating to inspection, maintenance, and / or replacement work.

[0212] Alternatively, or in addition to this, the output engine 106 may function to automatically generate and output information regarding downtime so that downtime for inspection, maintenance, and / or replacement work can be automatically scheduled.

[0213] Alternatively, or in addition to the above, the output engine 106 may function to automatically generate and output alarms, warnings, or other outputs via the HMI in response to the development of the RUL curve or other predicted asset health conditions.

[0214] The present invention offers various effects and advantages. Using a Markov chain model with a state space that may consist of a relatively small number of states (e.g., 3 to 4 states), the number of state transition probabilities required to perform this method is very small. RUL curves or other predictive asset health information can be efficiently obtained. Quantitative information can be provided not only for the expected RUL curve but also for the confidence space. Accuracy and / or variance information can be quantified and output.

[0215] The methods and systems according to the present invention can be used in connection with assets of power systems, such as assets of power generation, distribution, and / or transmission systems, or assets of industrial systems, but are not limited thereto.

[0216] The present invention has been described in detail in the drawings and the preceding description, but such description should be considered illustrative or specific and not limiting. A person skilled in the art who practices the claimed invention will be able to understand and practice variations of the disclosed embodiments by examining the drawings, disclosure and appended claims. In the claims, the word “comprising” does not exclude other elements or steps, and the indefinite article “a” or “an” does not exclude plurals. The fact that a particular element or step is described in a particular claim does not in itself indicate that a combination of these elements or steps cannot be used advantageously, and specifically, any more important combination of claims should be considered disclosed in addition to the dependency of the actual claims.

Claims

1. A method for performing predictive health analysis on assets (11-13; 16-18) of power system assets or industrial assets, wherein the method comprises: This includes the step of performing multiple independent probabilistic simulations using the transition probabilities of a discrete Markov chain model, The discrete Markov chain model has a state space that includes a set of asset health states (41-44), Each of the aforementioned independent probabilistic simulations simulates the future evolution of the state space of the discrete Markov chain model over the forecast period. A different set of transition probabilities between the asset health states (41-44) is used for one of the independent probabilistic simulations run in parallel, the different set of transition probabilities is associated with different operating and / or ambient conditions, and the method further, The step includes calculating the development of the predicted asset health state (60;80) over the forecast period from the aforementioned independent probabilistic simulations, The step of calculating the development of the predicted asset health state includes calculating the probability distribution (61-64) in the state space multiple times within the prediction period, and mapping the probability distribution to a scalar representing deterioration. method.

2. The method according to claim 1, wherein the step of calculating the development of the predicted asset health state includes the step of calculating RUL(60;80).

3. The method according to claim 1 or 2, further comprising the step of calculating confidence information or variance information (81, 82) for the development of the predicted asset health state as a function of time over the forecast period.

4. The method according to claim 3, wherein the confidence information includes the future development of the confidence interval over the forecast period, or the variance information includes the future development of the variance over the forecast period.

5. The method according to claim 3 or 4, wherein the confidence information or variance information includes a lower time evolution and an upper time evolution, the lower being associated with a first set of transition probabilities, and the upper being associated with a second set of transition probabilities different from the first set of transition probabilities.

6. The aforementioned state space is At least one state (41) in which the operation of the asset is not adversely affected by a malfunction, There is at least one state (42, 43) in which the operation of the asset is adversely affected by a malfunction, but the asset continues to operate. The method according to any one of claims 1 to 5, further comprising (44) a state in which the asset is not functioning due to a malfunction.

7. The method according to any one of claims 1 to 6, wherein the development of the predicted asset health state is obtained as the time evolution of the scalar.

8. The method according to any one of claims 1 to 7, further comprising the step of determining the transition probabilities used for simulation among the independent probabilistic simulations from historical data including sensor data for a plurality of assets, the sensor data being labeled with a fault signature indicating what state in the state space each asset was in at what point in time.

9. The method according to any one of claims 1 to 8, wherein the aforementioned independent multiple probabilistic simulations are Markov chain Monte Carlo MCMC simulations.

10. The method according to any one of claims 1 to 9, wherein the discrete Markov chain model is homogeneous and each state (41 to 44) in the state space has a non-zero transition probability to at most one other state in the state space.

11. The steps include receiving sensor measurement data from different groups of sensors captured during the operation of the asset, The method further includes the step of updating the development of the predicted asset health state by determining the transition probability used in one of the independent probabilistic simulations from each group of received sensor measurement data, The sensor measurement data received from the different groups described above indicates the different operating conditions and / or ambient conditions. The method according to any one of claims 1 to 10, wherein the sensor measurement data is labeled with a fault signature indicating what state in the state space each asset was in at what point in time.

12. The aforementioned assets are power transformers, distributed energy resource (DER) units, or generators, and / or The method according to any one of claims 1 to 11, wherein the forecast period is one year or more, two years or more, three years or more, four years or more, five years or more, ten years or more, fifteen years or more, or twenty years or more.

13. A method for operating and / or maintaining assets (11-13; 16-18) of a power system asset or an industrial asset, the method comprising: A step of performing a predictive asset health analysis for the asset using the method described in any one of claims 1 to 12, A method comprising the step of automatically performing at least one of the following: generating an alarm or warning based on the calculated predicted asset health state development; generating a control signal for the control operation of the asset based on the calculated predicted asset health state development; scheduling downtime for the asset based on the calculated predicted asset health state development; scheduling maintenance work based on the calculated predicted asset health state development; scheduling replacement work based on the calculated predicted asset health state development; and changing the maintenance interval based on the calculated predicted asset health state development.

14. A computing system (20-24; 100) that functions to perform predictive health analysis on power system assets or industrial assets, the computing system comprising at least one integrated circuit The at least one integrated circuit functions to perform multiple independent probabilistic simulations using the transition probabilities of a discrete Markov chain model. The discrete Markov chain model has a state space that includes a set of asset health states (41-44), Each of the aforementioned independent probabilistic simulations simulates the future evolution of the state space of the discrete Markov chain model over the forecast period. A different set of transition probabilities between the asset health states (41-44) is used for one of the independent probabilistic simulations run in parallel, the different set of transition probabilities is associated with different operating and / or ambient conditions, and the integrated circuit further, The system functions to calculate the development of the predicted asset health state over the forecast period from the aforementioned independent probabilistic simulations. A computing system for calculating the development of the predicted asset health state, comprising the steps of calculating the probability distributions (61-64) in the state space multiple times within the prediction period, and mapping the probability distributions to scalars representing deterioration.

15. Industrial or power systems (10; 15), Assets (11-13; 16-18), An industrial or power system comprising a computing system (20-24; 100) according to claim 14, which performs predictive asset health analysis on the aforementioned assets (11-13; 16-18), wherein the computing system is optionally a distributed controller of the industrial or power system for controlling the aforementioned assets.

Images