Use of irrelevant filters to facilitate efficient RUL analysis of utility system assets

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 generally relate to techniques for improving the reliability of electrical utility systems. More specifically, the disclosed embodiments relate to techniques that use an irrelevant filter to facilitate efficient Remaining Useful Life (RUL) analysis and improve the reliability of utility system assets in the field.

Background Art

[0002] Related Art Utility system assets such as power transformers are important for ensuring interruption-free power delivery from power generation facilities to consumers within the power distribution network. Whenever a power plant fails, power can usually be drawn through the power distribution network to meet consumer demand, so the power distribution network is generally "fault tolerant" with respect to the loss of power generation assets. In contrast, a failure of a power transformer generally leads to a "power outage", which can affect consumers in a small area containing a few blocks, and in some cases, consumers across a large service area covering multiple square miles. Also, a failure of a single transformer can potentially propagate a very large voltage spike throughout the power distribution network, which can cause other transformers to fail and lead to a large-scale regional power outage affecting hundreds of square miles. A transformer explosion can also cause a fire, which can result in significant asset damage and loss of life. Therefore, it is desirable to be able to monitor the health of power transformer operation as much as possible to identify degraded transformers before they fail.

[0003] The current state of the art for power transformer monitoring is Dissolved Gas Analysis (DGA). DGA is the dissolved hydrocarbon It works by detecting the presence of elementary gases. When components inside the transformer become hot enough to generate hydrocarbon gases, it indicates a thermal 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 DGA is performed infrequently, such as once a year. Furthermore, DGA is inherently "reactive" rather than "predictive," as it detects downstream symptoms of a problem after a problem has occurred that has created a hot spot large enough to "bake out" hydrocarbon gases. [Overview of the project] [Problems that the invention aims to solve]

[0004] Several researchers are investigating the possibility of determining the remaining useful life (RUL) of a power transformer using predictive monitoring techniques that analyze time-series sensor signals generated by the power transformer. (For example, "Estimating the Remaining Useful Life of a Power Transformer" by inventors Kenny C. Gross et al., filed March 7, 2019, incorporated herein by reference.) Transformer based on Real-Time Sensor Data and Periodic Dissolved GAS Analyses See U.S. Patent Application No. 16 / 295,613, titled "..." However, one challenge that needs to be addressed for these predictive monitoring technologies is that utility system assets tend to fail only infrequently. This means that there may not be enough failure history data to determine whether an unusual pattern in sensor signals indicates an impending failure or is simply a new pattern in sensor signals unrelated to an impending failure. This lack of failure history data is a challenge for predictive monitoring technologies. However, this means there is a high probability of generating frequent false alarms, leading to unnecessary maintenance work and potentially causing utility system assets to be replaced prematurely.

[0005] Therefore, what is needed is a technology for evaluating the operational soundness and remaining useful life of utility system assets without the aforementioned shortcomings of existing technologies. [Means for solving the problem]

[0006] overview The disclosed embodiments provide a system for estimating the remaining useful life (RUL) of a utility system asset. In monitoring mode, the system iteratively performs the following operations: First, the system receives a set of current time-series signals collected from sensors within the utility system asset. Next, the system generates estimates of the current set of time-series signals using an inference model and performs a pairwise difference operation between the actual values ​​and the estimates of the current set of 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, the irrelevance filter removing SPRT alarms for signals that do not correlate 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 needs to be replaced.

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

[0008] In some embodiments, to reduce the computational workload, RUL-based metrics are calculated for utility system assets only when a number of filtered SPRT alarms exceeding a threshold are generated during the preceding time interval.

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

[0010] In some embodiments, during a RUL training mode preceding a monitoring mode, the system receives a RUL training set containing time-series signals collected from sensors within a similar utility system asset while the similar utility system asset is operational until it fails. The system also receives the relevant failure time for the similar utility system asset. The system then uses an inference model to generate estimates of the RUL training set of time-series signals. Next, the system performs a pairwise difference operation between the actual and estimated values ​​of the RUL training set of time-series signals to generate residuals. Next, the system performs SPRT on the residuals to generate an SPRT alarm with a relevant tripping frequency. Finally, the system trains a logistic regression model to predict the RUL of the utility system asset based on the correlation between the tripping frequency of the SPRT alarm and the failure time of the similar utility system asset.

[0011] In some embodiments, during RUL training mode, the system is also irrelevant. The system configures filters. During this process, the system identifies relevant SPRT alarms generated during the time interval prior to a utility system asset failure and configures irrelevant filters to remove irrelevant SPRT alarms.

[0012] In some embodiments, when training a logistic regression model to predict the RUL of utility system assets, the system considers only the tripping frequencies of SPRT alarms associated with the relevant SPRT alarms.

[0013] In some embodiments, time-series signals collected from sensors within a utility system asset include signals specifying one or more of the following: temperature, current, voltage, resistance, capacitance, vibration, dissolved gas metric, 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 includes a power transformer. [Brief explanation of the drawing]

[0016] [Figure 1] This figure shows an exemplary predictive monitoring system for utility system assets according to the disclosed embodiments. [Figure 2] This is a flowchart of the process for estimating the RUL of utility system assets according to the disclosed embodiments. [Figure 3] This flowchart shows the process for training an inference model for utility system assets according to the disclosed embodiments. [Figure 4] This flowchart shows the process for training a logistic regression model to predict the RUL of utility system assets and constructing relevant irrelevant filters according to the disclosed embodiments. [Modes for carrying out the invention]

[0017] Detailed explanation The following description is provided to enable any person skilled in the art to construct and use this embodiment, and is provided in relation to a specific application and its requirements. Various modifications to the disclosed embodiment 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 this embodiment. Thus, this embodiment is not limited to the embodiments shown and should be given the broadest scope consistent with the principles and features disclosed herein.

[0018] The data structures and codes described in this detailed description are typically stored on computer-readable storage media, which can be any device or medium capable of storing code and / or data for use by a computer system. Computer-readable storage media may include, but are not limited to, volatile memory, non-volatile memory, magnetic and optical storage devices (such as disk drives, magnetic tapes, CDs (Compact Discs), DVDs (Digital Multipurpose Discs or Digital Video Discs)), or other media currently known or to be developed that are capable of storing computer-readable media.

[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 code and / or data stored in a computer-readable storage medium, the computer system will read the data structure and code It is embodied as code and executes the methods and processes stored in a computer-readable storage medium. Further, the methods and processes described below may be included in hardware modules. For example, the hardware modules may include, but are not limited to, application specific integrated circuit (ASIC) chips, field programmable gate arrays (FPGA), and other programmable logic devices now known or later developed. When a hardware module is activated, the hardware module executes the methods and processes included within the hardware module. Overview The disclosed embodiments utilize a novel "irrelevant filter" that mimics the functionality of the basal ganglia of the human brain and facilitates improved RUL prediction decisions for large collections of high-cost utility grid assets such as high voltage transformers. Currently, many industries are benefiting from a new science called "biomimicry" that analyzes nature's best ideas and adapts them to engineering use cases. The invention disclosed herein provides an example of biomimicry.

[0020] Researchers in Sweden who performed MRI studies on the human brain discovered that the basal ganglia act as an "irrelevant filter" that plays an important role in human memory and cognition. If the human brain tried to process and remember all the inputs coming in through the senses, the brain would be overwhelmed. The basal ganglia exclude unnecessary information, thereby leaving only the details essential for forming memories that contribute to the survival of the species, such as, for example, food acquisition, danger avoidance, species reproduction, and ensuring that basic needs are met. Humans with the best memories have been shown to have highly active basal ganglia.

[0021] This core paradigm may be useful for facilitating certain types of engineering-related tasks. For example, researchers are beginning to explore the potential of using machine learning (ML) based on monitoring time-series signals acquired from sensors within utility system assets to facilitate scheduling maintenance work. ML pattern recognition techniques can be trained using sensor signals generated when an asset is considered to be operating without fault, and can further be used to detect anomalous signal patterns in that asset, which has been demonstrated to be used to schedule predictive maintenance to correct the underlying causes of the anomalous signal patterns.

[0022] It should be noted that it is extremely beneficial for asset operators to receive alarms, including early warnings of potential problems. This allows for the asset to be taken 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 a RUL estimate, which provides an estimate of how long an asset can operate safely before the probability of a destructive failure reaches a critical threshold (e.g., a 95% probability of failure). For example, two transformers in a utility grid may both trigger early warning alarms. However, if a service organization knows that the first transformer has a 2-month RUL estimate, but the second transformer is likely to fail in the next 72 hours, it is more beneficial to schedule an emergency repair operation for 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 detect anomalous signal patterns that correlate with asset failures. For example, a given asset may generate an anomalous time-series signal pattern that has not been previously observed on the asset, which would consequently generate an alarm. However, Such anomaly patterns may be considered "irrelevant" if the same pattern has been previously observed in other healthy assets that have operated without incident for several years. This means that many predictive monitoring alarms will be false alarms. For example, an anomaly (but harmless) pattern in a time-series signal may be linked to a relatively new asset, an asset operating in an environment with large temperature fluctuations, or an asset operating in an environment with large fluctuations in electricity flow (e.g., from population changes or utility grid reconfiguration). Such alarms may be triggered by new patterns in the time-series data of individual assets, but may not have significant importance to healthy predictions.

[0024] Therefore, what is needed is an "irrelevant filter" that processes time-series signals for utility system assets performed until failure and generates optimal weighting coefficients for the relevant RUL methodology. Note that this is analogous to the functionality of the basal ganglia "filter" in the human brain, which receives large streams of neural "signals" associated with the five primary senses and periodically "alerts" humans for patterns directly related to danger, survival, or species reproduction opportunities.

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

[0026] Our anomaly detection process is called "Sequential Probability Ratio Test" (SPRT). A systematic binary hypothesis technique called the Ratio Test is used for large amounts of time-series signals. It identifies small subsets of time-series signals used as filters to ensure further pattern recognition analysis to facilitate anomaly detection. Thus, our new technology substantially reduces the cost of RUL analysis by systematically and safely filtering anomaly alarms generated for individual utility system assets so that RUL analysis operations are performed only on "relevant" signature patterns that are likely to be linked to the initial failure condition.

[0027] Before further explaining how our new RUL estimation technique works, we first describe an exemplary predictive monitoring system in which the new technique operates. Predictive monitoring system Figure 1 shows an exemplary predictive monitoring system 100 according to a disclosed embodiment. As shown in Figure 1, the predictive monitoring system 100 operates on a set of time-series sensor signals 104 acquired from sensors within a utility system asset 102, such as a power transformer. Note that the time-series signals 104 may originate from any type of sensor that may be located in the components within the utility system asset 102, such as voltage sensors, current sensors, pressure sensors, rotational speed sensors, and vibration sensors. These time-series sensor signals

[0028]

number

[0029] This is expressed as, and here

[0030]

number

[0031] This represents the value of the time-series sensor signal at time t. During the operation of the predictive monitoring system 100, the time-series signal 104 is supplied to a time-series database 106 that stores the time-series signal 104 for subsequent analysis. Next, the time-series signal 104 is supplied to a non-linear, non-parametric (NLNP) regression model 108, either directly from the utility system asset 102 or from the time-series database 106. Upon receiving the time-series sensor signal 104, the NLNP regression model 108 performs a non-linear, non-parametric regression analysis on the sample (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 regression analysis using the Multivariate State Estimation Technique (MSET). Please note that the term MSET used in this document broadly refers to a class of pattern recognition techniques. (See, for example, [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.) Therefore, the term "MSET" as used herein can refer to any outline technique mentioned in [Gribok], which includes Ordinary Least Squares (OLS), Support Vector Machines (SVM), and Artificial Neural Networks. This includes twerks (ANNs: Artificial Neural Networks), MSET, or standardized MSET (RMSET: Regularized MSET). MSET is used for pattern recognition purposes. While this is advantageous, the disclosed embodiments can generally use any one of a general-purpose class of pattern recognition techniques called nonlinear nonparametric (NLNP) regression, which includes neural networks, support vector machines (SVMs), auto-associative kernel regression (AAKR), and even simple linear regression (LR).

[0033] Before MSET is used to monitor the system, a model is built and an estimation of the system's correct operating state is made from the model. The model is empirically derived from observations made during the 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 algorithm's monitoring phase to compute estimates of the system state.

[0034] Observing the system state is done using a vector of length n.

[0035]

number

[0036] This is represented by , 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 degree of correlation (not necessarily linear) between the element values.

[0037] The state vectors collected during the training phase are placed in a state matrix with m column vectors, each column vector being a state vector of length n that stores all signal values ​​at a specific point in time during the training phase. The quality of the estimates generated 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, state vector

[0039]

number

[0040] It is defined as follows:

[0041]

number

[0042] Here,

[0043]

number

[0044] is time

[0045]

number

[0046] Signals in

[0047]

number

[0048] These are the measured values ​​from. Next, the state matrix.

[0049]

number

[0050] teeth,

[0051]

number

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

[0053]

number

[0054] This acts as a fixed model of the system from which the signal values ​​are estimated. For the time being, we assume that the signal measurements represent a linear correlation phenomenon. And,

[0055]

number

[0056] Assuming that it sufficiently extends to the system's operating space, the state matrix

[0057]

number

[0058] The state vector can be represented as a linear combination of the column vectors stored in the system. Estimated state vector

[0059]

number

[0060] and the actual state vector

[0061]

number

[0062] Minimizing the Euclidean norm between and gives the following:

[0063]

number

[0064] This yields a linearly optimal estimate.

[0065]

number

[0066] However, real-world systems are mostly nonlinear. (Most importantly, the concept of expressing the estimated state vector as a function of both the current state vector and the history of the learned state vectors.) It is desirable to extend the linear approach form in Equation 1 to nonlinear systems. Sticking to the general form in Equation 1 is,

[0067]

number

[0068] It is also attractive from the perspective that the model can be extended by adding new state vectors. With these advantages in mind, some linear matrix operators can be replaced with nonlinear matrix operators.

[0069]

number

[0070] By substituting with this, we can maintain the form of a linear estimation equation.

[0071]

number

[0072] The equation is given by Equation 2

[0073]

number

[0074] The value of is observed

[0075]

number

[0076] This is called MSET estimation for nonlinear operators.

[0077]

number

[0078] The following characteristics must be preserved. 1. Matrix

[0079]

number

[0080] It must be nonspecific. 2. Estimated vector

[0081]

number

[0082] teeth,

[0083]

number

[0084] Some of the elements

[0085]

number

[0086] It must represent the best estimate even when it is outside the range of the same element (i.e., when the observed signal value is smaller than the minimum value of the signal observed during the training phase or larger than the maximum value). 3. Observed vectors

[0087]

number

[0088] but

[0089]

number

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

[0091]

number

[0092] teeth

[0093]

number

[0094] It must be identical to [the other one]. 4.

[0095]

number

[0096] and

[0097]

number

[0098] The difference must be kept to a minimum. It has been shown that nonlinear operators satisfying these conditions exist and that they have been successfully used to monitor real-world systems.

[0099] Returning to Figure 1, the NLNP regression model 108 is “trained” to learn correlation patterns between time-series signals 104. This training process involves a single computationally intensive calculation performed offline using anomalous stored data. The pattern recognition system is then placed into “real-time monitoring mode,” and the trained NLNP regression model 108 predicts what each signal should be based on other correlation variables. This is the “estimated signal value” 110 shown in Figure 1. The system then uses the difference module 112 to perform a pairwise difference operation between the actual signal values ​​and the estimated signal values ​​to generate residuals 114, which are passed to the SPRT module 116. In embodiments of the present invention using MSET regression analysis, the residuals can be calculated using the following formula:

[0100]

number

[0101] Next, the SPRT module 116 performs a "detection operation" on the residual 114 to detect anomalies and generate an SPRT alarm 118 as early as possible. The SPRT module uses the Sequential Probability Ratio Test (SPRT) proposed by Wald to detect subtle statistical changes in a stationary, noisy observation sequence as early as possible. (Wald, Abraham, June 1945, "Sequential Tests of Statistical Hypotheses," "Annals of See Mathematical Statistics, 16(2):117-186. (Details of SPRT are clarified.) The process signals are monitored for this purpose.

[0102]

number

[0103] Typically, processes with a mean of zero and a standard deviation of σ (non-zero mean μ) are observed from each observation.

[0104]

number

[0105] Assume that the process signal is distributed as follows (it can be converted to a zero-average process by subtracting from it).

[0106]

number

[0107] teeth,

[0108]

number

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

[0110]

number

[0111] If it appears to be distributed in a certain way, it is said to be degraded, and here,

[0112]

number

[0113] This represents the magnitude of a predetermined system disturbance. SPRT provides a quantitative framework for determining between two hypotheses related to this concept of signal degradation. (1)

[0114]

number

[0115] :

[0116]

number

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

[0118]

number

[0119] :

[0120]

number

[0121] The observation was that, on average

[0122]

number

[0123] And it is subtracted from a normal distribution with a standard deviation σ.

[0124]

number

[0125] or

[0126]

number

[0127] If true, then with probability (1-α) or (1-β),

[0128]

number

[0129] or

[0130]

number

[0131] We want to determine this, assuming that α and β represent error discrimination probabilities (and therefore 0 ≤ α, β ≤ 1). In other words, α is

[0132]

number

[0133] When the condition is true

[0134]

number

[0135] The error alarm probability is the probability of accepting the error, and β is,

[0136]

number

[0137] When the condition is true

[0138]

number

[0139] This is the probability of accepting the missing alarm, which is the probability of the alarm being missed.

[0140]

number

[0141] teeth,

[0142]

number

[0143] Assuming that is true, the sequence

[0144]

number

[0145] Assuming this is the probability of observing, the likelihood ratio

[0146]

number

[0147] teeth,

[0148]

number

[0149] n observations

[0150]

number

[0151] of

[0152]

number

[0153] It can be calculated after doing so. Taking the natural logarithm of the likelihood ratio, independent observations

[0154]

number

[0155] Assume that it will generate. Sampling

[0156]

number

[0157] teeth

[0158]

number

[0159] Wald's theory states that this continues as long as the condition is met, where A and B are the acceptable thresholds related to the error misclassification probabilities α and β.

[0160]

number

[0161] We

[0162]

number

[0163] When sampling occurs...

[0164]

number

[0165] Stop it immediately

[0166]

number

[0167] We decided that

[0168]

number

[0169] When that happens, sampling stops immediately.

[0170]

number

[0171] This is determined. 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 has not deteriorated 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 values

[0174]

number

[0175] Assuming that the distribution is normal, we obtain a particularly compact equation for (Equation 3).

[0176]

number

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

[0178] Note that the above SPRT is a parametric test, which means that the parameters related to the probability density function must be known before applying the SPRT. Equation 4 is for the normal distribution observation of the process signal

[0179]

Number

[0180] is derived for. It is also possible to derive equations for other distributions (e.g., exponential, Poisson, binomial). However, in an actual computing system, it may be difficult to assume the following. 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] There are nonparametric sequential detection tests, but the mathematics behind them is considerably more complex than that presented above for parametric SPRT. Even if the a priori distribution is known, the third assumption is often violated in an actual computing system. A nominally stationary Gaussian random process can enter a new regime of operation (characterized by a different mean value or different higher moments) in response to the influence from a stimulus. When running a computing system, for example, a sudden change in workload can cause the monitored voltage or current signal to have an upward or downward step change at its nominal value. The observation results no longer seem to be drawn from the original

[0182]

Number

[0183] hypothesis-conforming distribution, so the SPRT recognizes such a step change as a degradation signal.

[0184] Referring to FIG. 1, the NLNP regression model 108 and the difference module 112 cooperate to remove (filter) the dynamics in the signal so that the residuals are a stationary random process when the system is in good condition. When the system ages or deteriorates due to a failure mechanism, the statistical characteristics of the residuals change. This change is detected by the SPRT module 116 that generates the corresponding SPRT alarm 118. When the system is in good condition, the residuals

[0185]

Number

[0186] are a stationary random process, so that the

[0187]

Number

[0188] dynamics in the signal are removed (filtered). When the system ages or deteriorates due to a failure mechanism, the statistical characteristics of the residuals change. This change is detected by the SPRT module 116 that generates the 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 some of the residuals are statistically different from the residuals corresponding to the non-deteriorated state of the system.

[0190] The SPRT alarm 118 is then sent through an irrelevant filter 120 that removes the SPRT alarm for signals that are not correlated with previous failures of similar utility system assets in order to generate a filtered SPRT alarm 124. The filtered SPRT alarm 124 is supplied to a logistic regression model 126, which generates a RUL estimate 128, and the RUL estimate 128 can be represented as a "quantitative risk metric" as will be described in more detail below.

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

[0192]

Number

[0193] where

[0194]

Number

[0195] are the values of the prediction parameters at time t. Therefore, at time t,

[0196]

Number

[0197] is true. Next, the logistic regression model 126 calculates the RUL of the utility system asset 102 as follows. Assuming the current conditions determined by the tripping frequency F of the current SPRT alarm, we represent the probability that the system S will fail within the next T hours as p(T,F). The relationship between p and the current conditions F is modeled using a linear logistic regression model rule.

[0198]

Number

[0199] where

[0200]

number

[0201] and

[0202]

number

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

[0204] RUL estimation technology Figure 2 shows a flowchart of the process for estimating the RUL of a utility system asset according to the disclosed embodiment. In monitoring mode, the system iteratively performs the following operations: First, the system receives a set of current time-series signals collected from sensors within the utility system asset (step 202). Next, the system generates estimates of the current set of time-series signals using an inference model (step 204), and also generates residuals by performing a pairwise difference operation between the actual values ​​and the estimates for the current set of time-series signals (step 206). The system then generates SPRT alarms by performing a successive probability ratio test (SPRT) on the residuals (step 208). Next, the system generates filtered SPRT alarms by applying an irrelevance filter to the SPRT alarms, the irrelevance filter removing SPRT alarms for signals that do not correlate with previous failures of similar utility system assets (step 210). The system then calculates a RUL-based risk index for the utility system asset based on the tripping frequencies of the filtered SPRT alarms using a logistic regression model (step 212). If the risk indicator exceeds the risk indicator 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 irrelevance filter based on time-series signals from the additional failed utility system asset (step 216).

[0205] Figure 3 is a flowchart showing the process for training an inference model according to the disclosed embodiment. During the inference training mode preceding the monitoring mode, the system receives an inference training set of time-series signals collected from sensors within the utility system assets 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] Figure 4 is a flowchart showing the process for training a logistic regression model to predict RUL for an asset and configuring relevant irrelevance filters according to the disclosed embodiment. During the RUL training mode preceding the monitoring mode, the system runs until a similar utility system asset fails, while similar utility system assets are running. The system receives a RUL training set containing time-series signals collected from sensors within the utility system asset (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 may include automatically clustering assets to form clusters containing "similar manufacturers / models" from a list of asset manufacturers / models, or empirically based on the number and type of internal sensors. Note that clusters of similar manufacturers / models may have different banks of transducers and different numbers of sensors, but what is relevant to the purpose of RUL estimation is the general pattern in these time-series signals.) Next, the system generates estimates of the RUL training set for the time series signal using an inference model (step 406). Then, the system generates residuals by performing a pairwise difference operation between the actual values ​​and the estimates of the RUL training set for the time series signal (step 408). Next, the system performs SPRT on the residuals to generate SPRT alarms with relevant tripping frequencies (step 410). Then, the system trains a logistic regression model to predict the RUL for utility system assets 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 configure an irrelevant filter, the system identifies relevant SPRT alarms generated during the time interval prior to the failure of the utility system asset (step 414), and then configures an irrelevant filter to remove irrelevant SPRT alarms (step 416).

[0207] 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 uses without departing from the spirit and scope of the invention. Therefore, the invention is not limited to the embodiments shown, but should be given the broadest scope that conforms to the principles and features disclosed herein.

[0208] The foregoing description of embodiments is provided for illustrative and illustrative purposes only. It is not intended to be exhaustive or to limit this description to the disclosed forms. Therefore, many modifications and variations will be apparent to those skilled in the art. Furthermore, the foregoing disclosure is not intended to limit this description. The scope of this description is defined by the appended claims.

Claims

1. A method for estimating the remaining useful life (RUL) of electronic equipment, wherein the method, during monitoring mode, The electronic device receives a set of time-series signals collected from sensors within the electronic device while it is operating. To detect statistical changes in the set of time-series signals that are considered to be abnormal signal patterns, The method includes generating a set of abnormal alarms, each of which is generated for each of the abnormal signal patterns, and the method The method includes applying an irrelevance filter to the set of abnormal alarms to generate filtered abnormal alarms, wherein the irrelevance filter removes abnormal alarms associated with abnormal signal patterns that do not correlate with previous failures of similar electronic devices, and the method is Using a logistic regression model, calculate a RUL-based risk index for the electronic equipment based on the filtered anomaly alarms, A method comprising generating a notification indicating that the remaining useful life of the electronic device is short when the risk indicator exceeds a risk indicator threshold.

2. The method according to claim 1, wherein when the abnormal signal pattern matches a similar signal pattern previously observed in other similar electronic devices that have operated without incident, the irrelevant filter removes the abnormal alarm associated with the abnormal signal pattern.

3. Detecting statistical changes in the aforementioned set of time-series signals is This includes performing a successive probability ratio test (SPRT) on the set of time-series signals or residual signals generated from the set of time-series signals, wherein the SPRT generates an SPRT alarm for the abnormal signal pattern. The method according to claim 1 or 2, wherein the SPRT alarm is the abnormal alarm.

4. Detecting statistical changes in the set of time-series signals is at least partially based on detecting statistical changes in the residual signals generated from the set of time-series signals. The above method, before the detection, Using an inference model to generate estimates of the set of time-series signals, The method according to any one of claims 1 to 3, further comprising generating the residual signal by performing a pairwise difference operation between the actual values ​​of the set of time-series signals and the estimated values ​​of the set of time-series signals.

5. During the RUL training mode preceding the aforementioned monitoring mode, The RUL training set, which includes time-series signals collected from sensors within a similar electronic device, is received while the similar electronic device is operating until it fails. Receiving the relevant failure time for the aforementioned similar electronic device, Using an inference model to generate estimates of the RUL training set for time-series signals, The process involves generating residuals by performing a pairwise difference operation between the actual values ​​of the RUL training set and the estimated values ​​of the time series signal, A successive probability ratio test (SPRT) is performed on the residuals to generate an SPRT alarm with the associated tripping frequency, The method according to any one of claims 1 to 4, further comprising training the logistic regression model to predict the RUL of the electronic device based on the correlation between the tripping frequency of the SPRT alarm and the failure time of the similar electronic device.

6. The above method additionally configures the irrelevant filter during the RUL training mode by the following operations, and the following operations are: Identifying relevant SPRT alarms generated during time intervals close to the relevant failure time of similar electronic devices, The method according to claim 5, further comprising configuring the irrelevant filter to remove irrelevant SPRT alarms.

7. The method according to claim 6, wherein, while training the logistic regression model to predict the RUL of the electronic device, the method takes into account the tripping frequency of the SPRT alarm associated with the relevant SPRT alarm.

8. The method according to any one of claims 1 to 7, wherein the time-series signal collected from the sensor in the electronic device includes a signal specifying one or more combinations of temperature, current, voltage, resistance, capacitance, vibration, dissolved gas measurement criteria, cooling system parameters, and control signals.

9. The method according to claim 4 or 5, wherein the inference model includes a multivariate state estimation technique (MSET) model.

10. The method according to any one of claims 1 to 9, wherein the electronic device is a utility system asset, a vehicle component, or a computing system device.

11. A computer program for causing a computer to perform the method according to any one of claims 1 to 10.

12. A memory storing the computer program described in claim 11, A system comprising a processor for executing the aforementioned computer program.

Images