Use of irrelevant filter to facilitate efficient RUL analysis of utility system asset

Abstract

To provide a method and system for improving the reliability of a utility system asset.SOLUTION: The method includes receiving a time series signal collected from a sensor in a utility system asset, using an inference model to generate an estimated value of the time series signal, performing a pairwise difference operation between the actual value and the estimated value of the time series signal to generate a residual, and performing a sequential probability ratio test (SPRT) on the residuals to generate an SPRT alarm, applying an irrelevance filter that removes SPRT alarm for a signal that does not correlate with previous failures of similar utility system assets to the SPRT alarm to generate a filtered SPRT alarm, using a logistic regression model to calculate an RUL-based risk indicator for the utility system asset on the basis of the filtered SPRT alarm, and generating a notification indicating that the utility system asset needs to be replaced when the risk indicator exceeds a threshold value.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] Field The disclosed embodiments relate generally to techniques for improving the reliability of electrical utility systems. More specifically, the disclosed embodiments use irrelevance filters to facilitate efficient Remaining Useful Life (RUL) analysis and in-field This paper relates to a technology for improving the reliability of utility system assets in a power grid. [Background technology]

[0002] Related Technology Utility system assets such as power transformers are critical to ensure uninterrupted delivery of power from generating facilities to consumers in the distribution grid. The distribution grid is generally "fault tolerant" to the loss of generating assets because whenever a generating station fails, replacement power can usually be drawn through the distribution grid to meet consumer demand. In contrast, the failure of a power transformer typically leads to a "blackout," which may affect consumers in a small area comprising a few blocks, and in some cases, across a large service territory comprising multiple square miles. Also, the failure of a single transformer can potentially propagate very large voltage spikes throughout the distribution grid, which may cause other transformers to fail, leading to a large regional blackout affecting hundreds of square miles. Transformer explosions can also cause fires, which can result in considerable asset damage, and loss of life. It is therefore desirable to be able to monitor the operational health of power transformers whenever possible in order to identify degraded transformers before they fail.

[0003] The current state of the art for power transformer monitoring is Dissolved Gas Analysis (DGA). DGA measures dissolved hydrocarbons in the oil used to bathe the transformer windings. DGA works by detecting the presence of hydrocarbon gases. When components within the transformer get hot enough to produce hydrocarbon gases, it represents a heat-related problem within the transformer. The problem with DGA is that it requires periodic extraction of oil samples from the transformer and chemical analysis to detect the presence of hydrocarbon gases. This process is time-consuming and expensive, which means that DGAs are performed infrequently, for example once a year. Furthermore, DGAs are "reactive" and not "predictive" in nature, as they detect downstream symptoms of a problem after a problem has occurred that creates enough hot spots to "bake out" the hydrocarbon gases. Summary of the Invention [Problem to be solved by the invention]

[0004] Several researchers have investigated the possibility of determining the remaining useful life (RUL) of a power transformer using predictive monitoring techniques that analyze time series of sensor signals generated by the power transformer. (See, for example, “Estimating THE Remaining Useful Life of a Power Transformer,” by inventors Kenny C. Gross et al., filed March 7, 2019, which is incorporated herein by reference.) Transformer based on Real-Time Sensor Data and Periodic Dissolved GAS Analyzes (See U.S. Patent Application Serial No. 16 / 295,613, entitled "Predictive Monitoring Techniques for Utility Systems," 2009). However, one challenge that needs to be addressed for these predictive monitoring techniques is that utility system assets tend to fail only infrequently. This means that there may not be sufficient historical failure data to determine whether an unusual pattern in a sensor signal is indicative of an impending failure or is simply a new pattern in the sensor signal that is not associated with an impending failure. This lack of historical failure data can make predictive monitoring techniques difficult to implement. However, this means that there is a high probability of generating high frequency false alarms, which may result in unnecessary maintenance work and premature replacement of utility system assets.

[0005] Therefore, what is needed is a technique for assessing the operational health and remaining useful life of utility system assets without the above-mentioned shortcomings of existing techniques. [Means for solving the problem]

[0006] overview The disclosed embodiment provides a system for estimating remaining useful life (RUL) of a utility system asset. During a monitoring mode, the system iteratively performs the following operations: First, the system receives a set of current time series signals collected from sensors in the utility system asset. Next, the system uses an inference model to generate estimates of the set of current time series signals and performs a pairwise difference operation between the actual values ​​and the estimates of the set of current time series signals to generate residuals. Next, the system performs a Sequential Probability Ratio Test (SPRT) on the residuals to generate SPRT alarms. Next, the system applies an irrelevance filter to the SPRT alarms to generate filtered SPRT alarms, where the irrelevance filter removes SPRT alarms for signals that are not correlated with previous failures of similar utility system assets. The system calculates a RUL-based risk index for the utility system asset based on the filtered SPRT alarms using a logistic regression model. Finally, when the risk index exceeds a risk index threshold, the system generates a notification indicating that the utility system asset should be replaced.

[0007] In some embodiments, the system periodically updates the logistic regression model and the irrelevance filter based on time series signals from the additional failed utility system assets.

[0008] In some embodiments, to reduce the computational workload, the RUL-based index is calculated for a utility system asset only when more than a threshold number of filtered SPRT alarms have been generated during the preceding time interval.

[0009] In some embodiments, during an inference training mode that precedes the monitoring mode, the system receives an inference training set of time series signals collected from sensors in the utility system assets during normal fault-free operation, and then trains an inference model to predict values ​​of the time series signals based on the inference training set.

[0010] In some embodiments, during a RUL training mode preceding the monitoring mode, the system receives an RUL training set including time series signals collected from sensors in similar utility system assets while the similar utility system assets are operated until failure. The system also receives associated failure times for the similar utility system assets. The system then uses an inference model to generate estimates of the RUL training set of time series signals. The system then performs pairwise differencing operations between the actual and estimated values ​​of the RUL training set of time series signals to generate residuals. The system then performs an SPRT on the residuals to generate SPRT alarms having associated tripping frequencies. Finally, the system trains a logistic regression model to predict the RUL of the utility system asset based on the correlation between the tripping frequencies of the SPRT alarms and the failure times of the similar utility system assets.

[0011] In some embodiments, during the RUL training mode, the system also Configure Filters: During this process, the system identifies relevant SPRT alarms generated during the time interval before the utility system asset failed, and configures irrelevance filters to remove unrelated SPRT alarms.

[0012] In some embodiments, while training a logistic regression model to predict the RUL of a utility system asset, the system considers only the tripping frequencies of SPRT alarms that are linked to the associated SPRT alarm.

[0013] In some embodiments, the time series signals collected from sensors in the utility system assets include signals specifying one or more of temperature, current, voltage, resistance, capacitance, vibration, dissolved gas metrics, cooling system parameters, and control signals.

[0014] In some embodiments, the inference model includes a Multivariate State Estimation Technique (MSET) model.

[0015] In some embodiments, the utility system asset comprises a power transformer. [Brief description of the drawings]

[0016] [Figure 1] FIG. 1 illustrates an example predictive monitoring system for utility system assets in accordance with disclosed embodiments. [Diagram 2] 1 is a flowchart of a process for estimating the RUL of a utility system asset in accordance with disclosed embodiments. [Diagram 3] 1 is a flowchart illustrating a process for training an inference model for a utility system asset according to disclosed embodiments. [Figure 4] 1 is a flowchart illustrating a process for training a logistic regression model to predict the RUL of a utility system asset and configuring an associated irrelevance filter in accordance with disclosed embodiments. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0017] Detailed Description The following description is presented to enable any person skilled in the art to make and use the present embodiments, and is provided in the context of a particular application and its requirements. Various modifications to the disclosed embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other embodiments and applications without departing from the spirit and scope of the present embodiments. Thus, the present embodiments are not limited to the embodiments shown, but are to be accorded the widest scope consistent with the principles and features disclosed herein.

[0018] The data structures and code described in this detailed description are typically stored on a computer-readable storage medium, which may be any device or medium capable of storing code and / or data for use by a computer system. The computer-readable storage medium may include, but is not limited to, volatile memory, non-volatile memory, magnetic and optical storage devices (such as disk drives, magnetic tapes, CDs (compact disks), DVDs (digital versatile disks or digital video disks)), or other media capable of storing computer-readable media now known or later developed.

[0019] The methods and processes described in the detailed description section may be embodied as code and / or data, which may be stored in a computer-readable storage medium as described above. When a computer system reads and executes the code and / or data stored in a computer-readable storage medium, the computer system executes the data structures and code. The hardware module may be embodied as a programmable logic device (PLC) and executes the methods and processes stored in a computer-readable storage medium. Additionally, the methods and processes described below may be included in a hardware module. For example, the hardware module may include, but is not limited to, an application specific integrated circuit (ASIC) chip, a field programmable gate array (FPGA), and other programmable logic devices now known or later developed. When the hardware module is activated, the hardware module executes the methods and processes included in the hardware module. overview The disclosed embodiments utilize a novel "irrelevance filter" that mimics the functionality of the basal ganglia of the human brain and facilitates improved RUL prediction decisions, especially for large collections of high-cost utility grid assets such as high-voltage transformers. Many industries are currently benefiting from an emerging science called "biomimetics," which analyzes nature's best ideas and adapts them to engineering use cases. The inventions disclosed herein provide examples of biomimicry.

[0020] Swedish researchers performing MRI studies on the human brain have found that the basal ganglia act as an "irrelevance filter" that plays a key role in human memory and cognition. If the human brain were to attempt to process and remember all input coming in through the senses, it would be overwhelmed. The basal ganglia filter out unnecessary information, thereby leaving only the details essential to form memories that contribute to the survival of the species, for example, obtaining food, avoiding danger, reproducing the species, and ensuring that basic needs are met. Humans with the best memories have been shown to have highly active basal ganglia.

[0021] This basal ganglia paradigm can be useful to facilitate certain types of engineering-related tasks. For example, researchers have begun to explore the possibility of using machine learning (ML) based on monitoring of time series signals obtained from sensors in utility system assets to facilitate scheduling of maintenance activities. It has been demonstrated that ML pattern recognition techniques can be trained using sensor signals generated when an asset is considered to be operating without faults and can be further used to detect anomalous signal patterns for that asset, which can be used to schedule predictive maintenance that will correct the underlying causes of the anomalous signal patterns.

[0022] Note that it is extremely beneficial for asset operators to receive alarms with early warnings about potential problems. This allows them to take the asset out of service as soon as possible to quickly diagnose the root cause of the abnormal signal. However, when there is a large set of similar assets, what is more beneficial is the RUL estimate, which provides an estimate of how long the asset would be able to safely operate before the probability of catastrophic failure reaches a critical threshold (e.g., 95% probability of failure). For example, two transformers in a utility grid can both issue early warning alarms. However, if a service organization knows that the first transformer has an RUL estimate of two months, but the second transformer is likely to fail in the next 72 hours, it is more beneficial to schedule an emergency repair action on the second transformer and wait for a "convenient maintenance window" to repair the first transformer.

[0023] However, for high-cost, high-reliability utility assets that do not fail frequently, existing RUL estimation techniques may not work well because only a small number of assets actually fail during operation. This means that there may not be enough training data for ML techniques to be able to detect anomalous signal patterns that correlate with asset failures. For example, a given asset may generate anomalous time series signal patterns not previously seen on the asset, which would result in an alarm. However, Such anomalous patterns may be considered "irrelevant" if the same patterns have been seen previously in other healthy assets that have operated without incident for several years. This means that many predictive monitoring alarms are false alarms. For example, anomalous (but harmless) patterns in time series signals may be associated with relatively new assets, assets operating in environments with large temperature fluctuations, or assets operating in environments with large fluctuations in electricity flows (e.g., from population changes or utility grid reconfigurations). Such alarms may be triggered by new patterns in the time series data of individual assets, but may not have health predictive significance.

[0024] What is needed, therefore, is an "irrelevance filter" that processes time series signals for utility system assets that have run to failure and generates optimal weighting coefficients for the associated RUL methodology. Note that this is similar in functionality to the basal ganglia "filter" for the human brain, which receives a large stream of neural "signals" linked to the five primary senses and periodically "alerts" humans to patterns that directly relate to danger, survival, or reproductive opportunities of the species.

[0025] Our novel ML-based technique works by processing data historian files. More specifically, when a collection of utility system assets, such as high voltage transformers, are monitored, time-series telemetry signals are continuously stored in data historian files, with one (logical) data historian file per monitored asset. These data historian files can be "collected" continuously (e.g., on a 1-15 minute basis) and added to a large database, where they are processed to discover trends, anomalies, environmental issues, and other incipient problems.

[0026] Our anomaly detection process is based on the Sequential Probability Ratio Test (SPRT). We apply a systematic binary hypothesis technique called the Unrelated Ratio Test (IRRT) to large sets of time series signals. Used as a filter, it identifies a small subset of time series signals that warrant further pattern recognition analysis to facilitate anomaly detection. Thus, our novel technique substantially reduces RUL analysis costs by systematically and safely filtering anomaly alarms generated for individual utility system assets such that RUL analysis operations are performed only on "relevant" signature patterns that are likely to be linked to an incipient fault condition.

[0027] Before further describing the operation of our new RUL estimation technique, we first describe an exemplary predictive monitoring system in which the new technique operates. Predictive Monitoring System FIGURE 1 illustrates an exemplary predictive monitoring system 100 in accordance with the disclosed embodiments. As shown in FIGURE 1, the predictive monitoring system 100 operates on a set of time series sensor signals 104 obtained from sensors in a utility system asset 102, such as a power transformer. It should be noted that the time series signals 104 may come from any type of sensor that may be located on a component in the utility system asset 102, for example, a voltage sensor, a current sensor, a pressure sensor, a rotational speed sensor, and a vibration sensor. These time series sensor signals may be

[0028]

number

[0029] where

[0030]

number

[0031] is the value of the time series sensor signal at time t. During operation of the predictive monitoring system 100, the time series signals 104 are provided to a time series database 106, which stores the time series signals 104 for subsequent analysis. The time series signals 104 are then provided to a Non-Linear, Non-Parametric (NLNP) regression model 108, either directly from the utility system assets 102 or from the time series database 106. Upon receiving the time series sensor signals 104, the NLNP regression model 108 performs a non-linear, non-parametric regression analysis on the samples (including the "current sample"). Once the analysis is complete, the NLNP regression model 108 outputs an estimated signal value 110.

[0032] In one embodiment of the present invention, the NLNP regression model 108 performs the regression analysis using a Multivariate State Estimation Technique (MSET). Please note that the term MSET, as used herein, refers to a technique that broadly describes a class of pattern recognition techniques. (See, e.g., [Gribok] "Use of Kernel Based Techniques For Sensor Validation in Nuclear Power Plants," by Andrei V. Gribok, J. Wesley Hines, and Robert E. Uhrig, THE Third American Nuclear Society International Topical Meeting on Nuclear PLANT Instrumentation and Control and Human-Machine Interface Technologies, Washington DC, November 13-17, 2000.) Thus, the term "MSET" as used herein can refer to any of the general techniques mentioned in [Gribok], including Ordinary Least Squares (OLS), Support Vector Machines (SVM), Artificial Neural Networks (AINN), and others. These include Artificial Neural Networks (ANNs), MSETs, or Regularized MSETs (RMSETs). Using MSETs for pattern recognition purposes is Advantageously, the disclosed embodiments may use any one of a general class of pattern recognition techniques generally referred to as nonlinear nonparametric (NLNP) regression, which includes neural networks, support vector machines (SVM), auto-associative kernel regression (AAKR), and even simple linear regression (LR).

[0033] Before MSET is used to monitor a system, a model is constructed and an estimate of the system's correct operating state is made from the model. The model is empirically derived from observations made during a training phase on the actual system under expected normal operating conditions. Relationships between signals are learned during the training phase and these relationships are used in the monitoring phase of the algorithm to compute an estimate of the system state.

[0034] The observations of the system state are expressed as a vector of length n

[0035]

number

[0036] where n is the number of signals in the system. The state vector elements do not need to be linearly independent, but there must be some correlation (not necessarily linear) between the element values.

[0037] The state vectors collected during the training phase are arranged into a state matrix with m column vectors, each of which is a state vector of length n that stores the values ​​of all signals at a particular time during the training phase. The quality of the estimates produced during the MSET monitoring phase depends on how well a subset of the m state vectors represents the expected behavior of the system observed during the training phase.

[0038] More formally, the state vector

[0039]

number

[0040] is defined as follows:

[0041]

number

[0042] Where:

[0043]

number

[0044] is time

[0045]

number

[0046] Signals in

[0047]

number

[0048] Then, the state matrix

[0049]

number

[0050] teeth,

[0051]

number

[0052] It can be defined as: When MSET monitors the system for degradation during its monitoring phase, the state matrix

[0053]

number

[0054] acts as a fixed model of the system whose signal values ​​are to be estimated. For the moment, we assume that the signal measurements represent a linear correlation phenomenon. Then,

[0055]

number

[0056] Assuming that fully spans the system's operating space, the state matrix

[0057]

number

[0058] The state vector can be expressed as a linear combination of the column vectors stored in the estimated state vector

[0059]

number

[0060] and the actual state vector

[0061]

number

[0062] By minimizing the Euclidean norm between

[0063]

number

[0064] A linear optimal estimate of is obtained.

[0065]

number

[0066] However, real systems are rarely linear. (The most important is the notion of expressing the estimated state vector as a function of both the current state vector and the history of the trained state vector.) It is desirable to extend the form of the linear approach in Eq. 1 to nonlinear systems. Adhering to the general form in Eq. 1 is

[0067]

number

[0068] It is also attractive from the viewpoint that the model can be extended by adding new state vectors to . With these benefits in mind, we consider replacing some of the linear matrix operators with nonlinear matrix operators.

[0069]

number

[0070] while still maintaining the form of the linear estimation equation.

[0071]

number

[0072] Given by Eq.

[0073]

number

[0074] The value of is the observation

[0075]

number

[0076] This is called the MSET estimate for the nonlinear operator

[0077]

number

[0078] must preserve the following properties: 1. Queue

[0079]

number

[0080] must be non-specific. 2. Estimation Vector

[0081]

number

[0082] teeth,

[0083]

number

[0084] Some elements of

[0085]

number

[0086] (i.e., when the observed signal value is less than the minimum or greater than the maximum signal value observed during the training phase) should represent an optimal estimate even outside the range of the same element. 3. Observation Vector

[0087]

number

[0088] but

[0089]

number

[0090] If it is identical to one of the column vectors of

[0091]

number

[0092] teeth

[0093]

number

[0094] must be identical to 4.

[0095]

number

[0096] and

[0097]

number

[0098] The difference must be kept to a minimum. It has been shown that nonlinear operators that satisfy these conditions exist and have been successful in monitoring real systems.

[0099] Returning to FIG. 1, the NLNP regression model 108 is "trained" to learn correlation patterns between the time series signals 104. This training process involves a single computationally intensive calculation performed offline using accumulated data that does not contain anomalies. The pattern recognition system is then placed in "real-time monitoring mode" and the trained NLNP regression model 108 predicts what each signal should be based on other correlated variables. This is the "estimated signal value" 110 shown in FIG. 1. The system then uses a differencing module 112 to perform pairwise differencing operations between the actual and estimated signal values ​​to generate residuals 114 that are passed to the SPRT module 116. In an embodiment of the invention using MSET regression analysis, the residuals can be calculated using the following formula:

[0100]

number

[0101] The SPRT module 116 then performs "detection operations" on the residuals 114 to detect anomalies and possibly generate SPRT alarms 118. The SPRT module uses sequential probability ratio tests (SPRTs) proposed by Wald to detect subtle statistical changes in stationary noisy observation sequences at the earliest possible time. (Wald, Abraham, June 1945, "Sequential Tests of Statistical Hypotheses," Annals of (See Mathematical Statistics, 16(2):117-186.) For details of SPRT, The process signal to be monitored for the purpose of

[0102]

number

[0103] is usually calculated from each observation for a process with mean zero and standard deviation σ (non-zero mean μ

[0104]

number

[0105] (which can be converted to a zero-mean process by subtracting

[0106]

number

[0107] teeth,

[0108]

number

[0109] Observations made on have a normal (Gaussian) distribution rather than a zero mean.

[0110]

number

[0111] is said to be degraded if it appears to be distributed with respect to

[0112]

number

[0113] is the magnitude of the desired system disturbance. SPRT provides a quantitative framework for deciding between two hypotheses related to this concept of signal degradation. (1)

[0114]

number

[0115] :

[0116]

number

[0117] The observations are drawn from a normal distribution with mean zero and standard deviation σ. (2)

[0118]

number

[0119] :

[0120]

number

[0121] The observations are average

[0122]

number

[0123] and is drawn from a normal distribution with standard deviation σ.

[0124]

number

[0125] or

[0126]

number

[0127] is true, with probability (1-α) or (1-β), respectively.

[0128]

number

[0129] or

[0130]

number

[0131] Suppose we wish to determine the probability of misclassification (so that 0≦α, β≦1). In other words, α is

[0132]

number

[0133] is true

[0134]

number

[0135] is the false alarm probability, which is the probability of accepting

[0136]

number

[0137] is true

[0138]

number

[0139] is the missed alarm probability, which is the probability of accepting

[0140]

number

[0141] teeth,

[0142]

number

[0143] Assuming that is true, the sequence

[0144]

number

[0145] is the probability of observing

[0146]

number

[0147] teeth,

[0148]

number

[0149] n observations

[0150]

number

[0151] of

[0152]

number

[0153] It can be calculated after taking the natural logarithm of the likelihood ratio and assuming that the independent observations

[0154]

number

[0155] Suppose we generate the following:

[0156]

number

[0157] teeth

[0158]

number

[0159] Wald's theory continues as long as,where A and B are tolerance thresholds related to the error misidentification probabilities,α,and,β,.

[0160]

number

[0161] We

[0162]

number

[0163] When sampling

[0164]

number

[0165] Stop immediately

[0166]

number

[0167] We decided

[0168]

number

[0169] Sampling stops immediately when

[0170]

number

[0171] Note that in order to continuously monitor the system, the SPRT algorithm resets itself as soon as a decision is made. For example, if the likelihood ratio exceeds threshold A, indicating that the process signal is not degraded at that point, the next sample is treated as the first sample (n=1 according to Equation 3) in a new sequence of observations.

[0172]

number

[0173] Observed value of

[0174]

number

[0175] Assuming that is normally distributed, we obtain a particularly compact formula for (Eq. 3).

[0176]

number

[0177] Following Wald's sequential analysis, it has been shown that decision tests based on SPRTs have optimal properties: for given probabilities α and β, there is no other procedure that has an error probability or expected risk at least as low as the SPRT and has a shorter average sampling time. It is because of this property, and the inherent simplicity of equation (2), that SPRTs have become popular for monitoring stationary Gaussian random processes.

[0178] It should be noted that the above SPRT is a parametric test, meaning that the probability density function and associated parameters must be known before applying the SPRT. Equation 4 expresses the process signal

[0179]

number

[0180] It is derived for normally distributed observations of . Formulas for other distributions (e.g., exponential, Poisson, binomial) can also be derived. However, in a practical computing system, it may be difficult to assume that 1. The distribution of the process signal is known in advance. 2. The distribution of the process signal does not change over time. 3. The parameters of the distribution do not change over time.

[0181] Non-parametric sequential detection tests exist, but the mathematics behind them are considerably more complex than those presented above for the parametric SPRT. Even if the a priori distribution is known, the third assumption is often violated in real computing systems. A nominally stationary Gaussian random process may enter a new operating regime (characterized by a different mean value or different higher moments) in response to influences from stimuli. When running a computing system, for example, a sudden workload change may cause the monitored voltage or current signal to have an upward or downward step change in its nominal value. The observations no longer correspond to the original

[0182]

number

[0183] The SPRT recognizes such step changes as degradation signals because they do not appear to be drawn from a distribution that fits the hypotheses.

[0184] Referring to FIG. 1, the NLNP regression model 108 and the differencing module 112 are When the system is in good condition, the residual

[0185]

number

[0186] is a stationary random process, so that the signal

[0187]

number

[0188] As the system ages or deteriorates due to failure mechanisms, the statistical properties of the residuals change. This change is detected by the SPRT module 116 which generates a corresponding SPRT alarm 118.

[0189] During operation, the SPRT module 116 applies a sequential probability ratio test to the residuals and generates an alarm when one or more residuals are statistically different from the residuals corresponding to the undegraded state of the system.

[0190] The SPRT alarms 118 are then sent through an irrelevance filter 120, which removes SPRT alarms for signals that are not correlated with prior failures of similar utility system assets, to produce filtered SPRT alarms 124. The filtered SPRT alarms 124 are fed into a logistic regression model 126, which produces a RUL estimate 128, which may be expressed as a "quantitative risk index," as described in more detail below.

[0191] While computing the RUL estimate 128, the logistic regression model 126 records each instance of alarm in the filtered SPRT alarms 124 and uses these instances to determine the tripping frequency of the current alarm. As degradation progresses, the tripping frequency of the filtered alarms increases. We define the tripping frequency of these alarms as

[0192]

number

[0193] where:

[0194]

number

[0195] is the value of the predicted parameter at time t. Therefore, at time t,

[0196]

number

[0197] It is. The logistic regression model 126 then calculates the RUL for the utility system asset 102 as follows: We denote the probability that a system S will fail within the next T hours given the current conditions determined by the current SPRT alarm tripping frequency F as p(T,F). The relationship between p and the current conditions F is expressed by the linear logistic regression model: The model is based on the

[0198]

number

[0199] and

[0200]

number

[0201] and

[0202]

number

[0203] is estimated from the historical or experimental failure data of the system. Note that the tripping frequency is normalized to have a value between 0 and 1 to simplify this calculation.

[0204] RUL estimation technology FIG. 2 shows a flow chart of a process for estimating the RUL of a utility system asset according to a disclosed embodiment. During a monitoring mode, the system iteratively performs the following operations: First, the system receives a set of current time series signals collected from sensors in the utility system asset (step 202). Next, the system uses an inference model to generate estimates of the set of current time series signals (step 204), and also performs pairwise differencing between the actual values ​​and the estimates for the set of current time series signals to generate residuals (step 206). The system then performs a sequential probability ratio test (SPRT) on the residuals to generate SPRT alarms (step 208). Next, the system applies an irrelevance filter to the SPRT alarms to generate filtered SPRT alarms, which removes SPRT alarms for signals that are not correlated with previous failures of similar utility system assets (step 210). The system then calculates an RUL-based risk index for the utility system asset based on the tripping frequency of the filtered SPRT alarms using a logistic regression model (step 212). If the risk index exceeds the risk index threshold, the system generates a notification indicating that the utility system asset needs to be replaced (step 214). Finally, the system periodically updates the logistic regression model and the irrelevance filter based on the time series signals from the failed additional utility system assets (step 216).

[0205] 3 is a flow chart illustrating a process for training an inference model according to a disclosed embodiment. During an inference training mode, which precedes a monitoring mode, the system receives an inference training set of time series signals collected from sensors in a utility system asset during normal fault-free operation (step 302). The system then trains an inference model to predict values ​​of the time series signals based on the inference training set (step 304).

[0206] 4 is a flow chart illustrating a process for training a logistic regression model to predict the RUL for an asset and to configure an associated irrelevance filter, according to a disclosed embodiment. During the RUL training mode, which precedes the monitoring mode, the system performs RUL training on similar utility system assets while they are running until failure. The system receives an RUL training set that includes time series signals collected from sensors in utility system assets (step 402). The system also receives failure times for similar utility system assets (step 404). (Note that the process for determining which utility system assets are similar can include automatically clustering the assets to form clusters that include "similar makes / models" from a list of asset makes / models or empirically based on the number and type of internal sensors. Note that clusters of similar makes / models may have different banks of transducers and different numbers of sensors, but it is the general patterns in these time series signals that are relevant for purposes of RUL estimation.) Next, the system uses the inference model to generate estimates of the RUL training set of the time series signal (step 406). The system then performs pairwise differencing between the actual values ​​and the estimates of the RUL training set of the time series signal to generate residuals (step 408). The system then performs SPRT on the residuals to generate SPRT alarms with associated tripping frequencies (step 410). The system then trains a logistic regression model to predict the RUL for the utility system asset based on the correlation between the tripping frequencies of the SPRT alarms and the failure times of similar utility system assets (step 412). Next, to construct an irrelevance filter, the system identifies relevant SPRT alarms generated during the time interval before the utility system asset failed (step 414), and then constructs an irrelevance filter to remove unrelated SPRT alarms (step 416).

[0207] Various modifications to the disclosed embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments and applications without departing from the spirit and scope of the invention. Thus, the present invention is not limited to the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

[0208] The foregoing description of the embodiments has been presented for purposes of illustration and description only. They are not intended to be exhaustive or to limit the description to the disclosed forms. Thus, many modifications and variations will be apparent to those skilled in the art. In addition, the above disclosure is not intended to limit the description. The scope of the description is defined by the appended claims.

Claims

1. 1. A method for estimating remaining useful life (RUL) of a utility system asset, the method comprising: iteratively performing the following operations during a monitoring mode, the operations comprising: receiving a set of current time series signals collected from sensors within the utility system asset; generating estimates of the current set of time series signals using an inference model; performing pairwise differencing between the actual values ​​of the current set of time series signals and the estimated values ​​to generate residuals; performing a sequential probability ratio test (SPRT) on the residuals to generate a SPRT alarm; and applying an irrelevance filter to the SPRT alarms to generate filtered SPRT alarms, the irrelevance filter removing SPRT alarms for signals that are not correlated with prior failures of similar utility system assets, the operations of: calculating a RUL-based risk index for the utility system assets based on the filtered SPRT alarms using a logistic regression model; and generating a notification indicating a need to replace the utility system asset when the risk index exceeds a risk index threshold.

2. The method of claim 1 , further comprising periodically updating the logistic regression model and the irrelevance filter based on time series signals from additional utility system assets that have failed.

3. 2. The method of claim 1, wherein to reduce computational workload, the RUL-based index is calculated for a utility system asset only when more than a threshold number of filtered SPRT alarms are generated during a preceding time interval.

4. During an inference training mode preceding said monitoring mode, receiving an inference training set of time series signals collected from sensors within the utility system asset during normal fault-free operation; The method of claim 1 , further comprising: training the inference model to predict values ​​of the time-series signal based on the inference training set.

5. During a RUL training mode preceding the monitoring mode, receiving a RUL training set including time series signals collected from sensors in similar utility system assets while the similar utility system assets are operated until failure; receiving associated downtimes for the similar utility system assets; generating estimates of a RUL training set of time series signals using the inference model; performing pairwise differencing between the actual values ​​of the RUL training set of time series signals and the estimated values ​​to generate residuals; performing SPRT on the residual to generate a SPRT alarm having an associated tripping frequency; and training the logistic regression model to predict a RUL for the utility system asset based on a correlation between a tripping frequency of the SPRT alarm and the failure times of the similar utility system assets.

6. The method further comprises, during the RUL training mode, additionally configuring the irrelevant filter by the following operations: Identifying associated SPRT alarms generated during a time interval prior to the utility system asset failing; and configuring the irrelevance filter to remove unrelated SPRT alarms.

7. During training the logistic regression model to predict the RUL of the utility system assets, the method further comprises: The method of claim 6 , further comprising considering only the tripping frequencies of SPRT alarms associated with the associated SPRT alarm.

8. 2. The method of claim 1 , wherein the time series signals collected from sensors in the utility system asset include signals specifying one or more of temperature, current, voltage, resistance, capacitance, vibration, dissolved gas metrics, cooling system parameters, and control signals.

9. The method of claim 1 , wherein the inference model comprises a multivariate state estimation technique (MSET) model.

10. The method of claim 1 , wherein the utility system asset comprises a power transformer.

11. 1. A non-transitory computer readable storage medium storing instructions that, when executed by a computer, cause the computer to perform a method for estimating remaining useful life (RUL) of a utility system asset, the method comprising: repetitively performing the following operations during a monitoring mode, the operations comprising: receiving a set of current time series signals collected from sensors within the utility system asset; generating estimates of the current set of time series signals using an inference model; performing pairwise differencing between the actual values ​​of the current set of time series signals and the estimated values ​​to generate residuals; performing a sequential probability ratio test (SPRT) on the residuals to generate a SPRT alarm; and applying an irrelevance filter to the SPRT alarms to generate filtered SPRT alarms, the irrelevance filter removing SPRT alarms for signals that are not correlated with prior failures of similar utility system assets, the operations of: calculating a RUL-based risk index for the utility system assets based on the filtered SPRT alarms using a logistic regression model; and generating a notification indicating a need to replace the utility system asset when the risk index exceeds a risk index threshold.

12. 12. The non-transitory computer-readable storage medium of claim 11, wherein the method further comprises periodically updating the logistic regression model and the irrelevance filter based on time series signals from additional utility system assets that have failed.

13. 12. The non-transitory computer-readable storage medium of claim 11, wherein to reduce computational workload, the RUL-based index is calculated for a utility system asset only when more than a threshold number of filtered SPRT alarms are generated during a preceding time interval. storage medium.

14. The method comprises: During an inference training mode preceding said monitoring mode, receiving an inference training set of time series signals collected from sensors within the utility system asset during normal fault-free operation; and training an inference model to predict values ​​of the time-series signal based on the inference training set.

15. The method comprises: During a RUL training mode preceding the monitoring mode, receiving a RUL training set including time series signals collected from sensors in similar utility system assets while the similar utility system assets are operated until failure; receiving associated downtimes for the similar utility system assets; generating estimates of the RUL training set of time series signals using the inference model; performing pairwise differencing between the actual values ​​of the RUL training set of time series signals and the estimated values ​​to generate residuals; performing SPRT on the residual to generate a SPRT alarm having an associated tripping frequency; and training the logistic regression model to predict a RUL of the utility system asset based on a correlation between a tripping frequency of the SPRT alarm and the failure time of the similar utility system assets.

16. The method comprises: During the RUL training mode, the irrelevant filter is additionally configured by the following operations: Identifying associated SPRT alarms generated during a time interval prior to the utility system asset failing; and configuring the irrelevance filter to remove unrelated SPRT alarms.

17. The method comprises: During training the logistic regression model to predict the RUL of the utility system assets, 20. The non-transitory computer-readable storage medium of claim 16, further comprising: considering only tripping frequencies of SPRT alarms associated with an associated SPRT alarm.

18. 1. A system for estimating remaining useful life (RUL) of a utility system asset, comprising: At least one processor and at least one associated memory; a notification mechanism executing on said at least one processor; The notification mechanism, during a monitoring mode, receiving a set of current time series signals collected from sensors within the utility system asset; generating estimates of the current set of time series signals using an inference model; performing pairwise differencing between the actual values ​​of the current set of time series signals and the estimated values ​​to generate residuals; performing a sequential probability ratio test (SPRT) on the residuals to generate a SPRT alarm; applying an irrelevance filter to the SPRT alarms to generate filtered SPRT alarms, the irrelevance filter removing SPRT alarms for signals that are not correlated with prior failures of similar utility system assets; Calculating a RUL-based risk index for the utility system assets based on the filtered SPRT alarms using a logistic regression model; The system generates a notification indicating that the utility system asset needs to be replaced when the risk index exceeds a risk index threshold.

19. During a RUL training mode preceding said monitoring mode, The notification mechanism includes: receiving a RUL training set including time series signals collected from sensors in similar utility system assets while the similar utility system assets are operated until failure; receiving associated downtimes for the similar utility system assets; generating estimates of the RUL training set of time series signals using the inference model; performing pairwise differencing between the actual values ​​and the estimated values ​​of the RUL training set of time series signals to generate residuals; performing SPRT on the residual to generate a SPRT alarm having an associated tripping frequency; 20. The system of claim 18, wherein the logistic regression model is trained to predict a RUL for the utility system asset based on a correlation between a tripping frequency of the SPRT alarm and the failure time of the similar utility system assets.

20. During the RUL training mode, The notification mechanism further configures the irrelevance filter, and during the configuration process, the notification mechanism: Identifying associated SPRT alarms generated during a time interval prior to the utility system asset failing; The system of claim 19 , wherein the irrelevant filter is configured to remove unrelated SPRT alarms.

Images