Robust distance measures for on-line monitoring

Abstract

An apparatus and associated method are utilized for monitoring an operation of a system characterized by operational parameters. A non-parametric empirical model generates estimates of parameter values in response to receiving a query vector of monitored parameters for a model characterizing the system. A distance estimation engine (a) determines robust distances between the query vector and each of a set of predetermined historical vectors for the non-parametric empirical model based on an implementation of an elemental kernel function; (b) determines weights for the monitored parameters based on the robust distances; and (c) combining the weights with the predetermined historical vectors to make predictions for the system.

Description

BACKGROUND OF THE INVENTION

Traditionally, the calibration of safety critical nuclear instrumentation has been performed at each refueling cycle. However, many nuclear plants have moved toward condition-directed rather than time-directed calibration. This condition-directed calibration is accomplished through the use of on-line monitoring which commonly uses an autoassociative predictive modeling architecture to assess instrument channel performance. An autoassociative architecture predicts a group of correct sensor values when supplied with a group of sensor values that is corrupted with process and instrument noise, and could also contain faults such as sensor drift or complete failure.

In the U.S. nuclear power industry, millions of dollars are spent annually on the calibration of instrument chains that are performing within the required specifications. For the past twenty years, several nuclear utilities have investigated methods to monitor the calibration of safety critical process instruments. In 2000, the U.S. Nuclear Regulatory Commission (NRC) issued a safety evaluation report (SER) on an EPRI submitted Topical Report (TR) 104965, “On-Line Monitoring of Instrument Channel Performance”. This SER concluded that the generic concept of on-line monitoring (OLM) for tracking instrument performance as discussed in the topical report is acceptable. However, additional requirements were identified that must be addressed by plant specific license amendments if the calibration frequency of safety-related instrumentation is to be relaxed. Since the applicability of an OLM system is directly related to the ability of an empirical model to correctly predict sensor values when supplied faulty data, methods must be developed to ensure that robust empirical models can be developed.

The autoassociative architecture for predicting correct sensor values has also been adapted for use in equipment fault detection and health monitoring. Accordingly, it is known to provide a nonparametric empirical model such as a kernel regression model or a similarity-based model that generates estimates of sensor values responsive to input of measurements of those sensor values in real-time. The estimates are subtracted from the measured values to provide residuals, which are used to detect deviations indicative of incipient equipment failure. Such approaches are known from, for example, U.S. Pat. No. 4,937,763 to Mott; and in U.S. Pat. No. 5,764,509 to Gross et al. In these approaches, a kernel function incorporating a distance function is used to compare the measured values of the sensors arranged as an observation vector, to a set of reference observations. The kernel function, also called a similarity operator, returns a scalar value indicative of the similarity of the input observation vector to each of the reference observation vectors, and these scalar values are used in generating an estimate observation of the sensor values as an adaptive linear combination of at least some of the reference observations. Kernel regression and similarity-based modeling differ in the details of how the adaptive linear combination is formed; however the kernel function is used in both instances. The scalar value or similarity value of the kernel function typically is designed to range between zero and one, where a value of one indicates the compared vectors are identical, and values approaching zero indicate increasing dissimilarity or distance between the vectors.

One of the drawbacks of the kernel functions in use is susceptibility to outlier inputs, especially when the kernel function is executed on the elements of the compared vectors. In such a case, the kernel function compares individual like elements of the vectors, and generates a scalar comparison outcome for each element, then combines those to form an observation level scalar value. When a particular sensor reading is very different from the sensor reading in a reference observation, the observation-level kernel result can be dominated by the outlier sensor value, resulting in a reduced similarity scalar value for the comparison of the input vector to the reference observation in question than might otherwise be implied by the other sensor readings.

SUMMARY OF THE INVENTION

The invention provides improved kernel-based model performance with more robust distance metrics, for sensor calibration and equipment health monitoring. Accordingly, robust distance measures for use in nonparametric, similarity based models are disclosed. The alternative robust distance functions have performance advantages for the common task of sensor drift detection. In particular, a robust Euclidean distance function according to the invention produces significant robustness improvements in nonparametric, similarity based models, such as kernel regression and the multivariate state estimation technique (MSET).

The invention can be used in software for monitoring the health of equipment and sensors, especially for nonparametric empirical model based systems. Accordingly, equipment is instrumented with sensors for determining a variety of physical or statistical measurements of equipment performance; the sensor data is provided to the software which generates estimates of the sensor data using the nonparametric empirical model; and the estimates are compared to the measured values to determine if an incipient deviation exists. Residuals can be processed through a variety of alerting, trending and pattern recognition techniques, to provide an autonomous software system for annunciation of probable and/or incipient equipment failures. People responsible for maintaining or operating the equipment can rely on the software to call out exceptional conditions in the equipment requiring intervention.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 comprises charts (a) data from turbine pressure sensors and (b) data from steam pressure sensors in a nuclear power plant, used to train nonparametric estimation models as may be used with embodiments of the present invention;

FIG. 2 is a bar graph showing the computed accuracy for five sensors in a nonparametric estimation model, for three alternatives of the model employing (a) an ordinary Euclidean distance metric, (b) an L¹-norm distance metric, and (c) a robust Euclidean distance metric according to an embodiment of the invention;

FIG. 3 is a bar graph showing the computed robustness for five sensors in a nonparametric estimation model, for three alternatives of the model employing (a) an ordinary Euclidean distance metric, (b) an L¹-norm distance metric, and (c) a robust Euclidean distance metric according to an embodiment of the invention;

FIG. 4 is a bar graph showing the computed spillover for five sensors in a nonparametric estimation model, for three alternatives of the model employing (a) an ordinary Euclidean distance metric, (b) an L¹-norm distance metric, and (c) a robust Euclidean distance metric according to an embodiment of the invention; and

FIG. 5 illustrates a monitoring apparatus 500 for monitoring a monitored system 505 and according to an embodiment of the invention.

DETAILED DESCRIPTION OF THE INVENTION

An empirical model's architecture may be either defined by a set of parameters and functional relationships (parametric) or a set of data and algorithmic estimation procedures (nonparametric). In a parametric model, training data is used to fit the model to the data according to a pre-defined mathematical structure. For example, consider the following polynomial model:y=b₀+b₁x₁+b₂x₂+b₃x₁x₂+b₄x₁²+b₅x₂²  (1)

In order to completely define this model for a given set of training observations, the polynomial coefficients, are optimized to minimize some objective function, usually the sum of the squared error (SSE). Once the optimal polynomial coefficients have been estimated, the model is completely specified by Equation 1 and the estimated coefficients. Therefore, a parametric model may be roughly defined as a model that may be completely specified by a set of parameters and a functional relationship for applying these parameters to new data in order to estimate the response.

A non-parametric model, by contrast, stores historical data exemplars in memory and processes them when a new query is made. For instance, rather than modeling a whole input space with a parametric model such as a neural network or linear regression, local non-parametric techniques may be used to construct a local model in the immediate region of the query. These models are constructed “on the fly” not beforehand. When the query is made; the algorithm locates historical exemplars in its vicinity and performs a weighted regression with the nearby observations. The observations are weighted with respect to their proximity to the query point. In order to construct a robust local model, one must define a distance function to measure what is considered to be local to the query, implement locally weighted regression, and in some cases consider additional regularization techniques.

As an example, the mathematical framework of a modeling technique such as autoassociative kernel regression (AAKR) is composed of three basic steps. First, the distance between a query vector (the observation comprised of the readings of the multiple sensors in the model) and each of the historical exemplar (memory) vectors is computed using the conventional Euclidean distance or L²-norm:

$\begin{array}{cc}{u}_j=\sqrt{\sum _{i=1}^n{\left(x_{q,i}-m_{j,i}\right)}^2}& \left(2\right)\end{array}$ where, u_j is the distance between the query vector (x) and j^th memory vector, n is the number of variables in the data set, x_q,i is the i^th variable of the query vector, and m_j,i is the i^th variable of the j^th memory vector.

Second, these distances are used to determine weights by evaluating the standard, Gaussian kernel, expressed by:

$\begin{array}{cc}w=K\left(u,h\right)=\frac{1}{\sqrt{2\pi ·h^2}}{e}^{-\frac{u^2}{h^2}}& \left(3\right)\end{array}$ where, h is the kernel's bandwidth. Finally, these weights are combined with the memory vectors to make predictions according to:

$\begin{array}{cc}{\hat{x}}_q=\frac{\sum _{i=1}^M{w}_i·m_i}{\sum _{i=1}^M{w}_i}& \left(4\right)\end{array}$

Here, w_i are the weights, m_i are the memory vectors, M is the number of memory vectors, and {circumflex over (x)}_q is the prediction for the query vector. Since the monitoring system's objective is to detect and quantify sensor drift, the model should be made as immune as possible to sensor drift. In order to improve the robustness of the AAKR modeling routine, distance functions other than the standard Euclidean distance may be used. Before discussing the alternative distance functions, the parameters used to measure model performance must be discussed.

The performance of autoassociative OLM systems is measured in terms of its accuracy, robustness, and spillover. Accuracy measures the ability of the model to correctly and accurately predict sensor values and is normally presented as the mean squared error (MSE) between the prediction and the correct sensor value. Robustness measures the ability of the model to make correct sensor predictions when the respective sensor value is incorrect due to some sort of fault. Spillover measures the effect a faulty sensor input has on the other sensor predictions in the model. An ideal system would be accurate and would not have sensor predictions affected by degraded inputs.

The most basic form of the AAKR modeling technique makes use of the Euclidean distance or L²-norm described above in Equation 2. Since this distance function squares the individual differences, the effects of a faulty input may be amplified, resulting in parameter predictions which are more affected by input variations and therefore less robust. In order to improve robustness, it is desirable to have distance measures which are not affected by errant sensor readings.

A first robust distance function is the L¹-norm, which is defined by the following equation.

$\begin{array}{cc}{u}_j=\sum _{i=1}^nx_{q,i}-m_{j,i}& \left(5\right)\end{array}$ Notice that rather than square the individual differences, the L¹-norm uses the absolute value. This alteration provides a modest improvement in robustness, but the distance will still be affected by faulty input. Therefore, an additional step can be taken in the robust distance function to remove faulty input from the distance calculation and improve model robustness. Accordingly, the largest elemental difference contributing to the distance metric is removed, as shown in the following equation:

$\begin{array}{cc}{u}_j=\sqrt{\sum _{i=1}^n{\left(x_{q,i}-m_{j,i}\right)}^2-\underset{i=1}{\overset{n}{\max }}\left[{\left(x_{q,i}-m_{j,i}\right)}^2\right]}& \left(6\right)\end{array}$

Here,

$\underset{i=1}{\overset{n}{\max }}\left[{\left(x_{q,i}-m_{j,i}\right)}^2\right]$ is the maximum squared difference of the query vector from the j^th memory vector. Simply speaking, one “bad performer” is assumed to exist and its influence is removed from the calculation. To more clearly illustrate Equation 6, consider the following example vectors.x_q=[0.9501 0.2311 0.6068 0.4860]m_j=[0.8913 1.7621 0.4565 0.0185]The squared differences are found to be:(x_q,i−m_j,i)²=[0.0035 2.3438 0.0226 0.2185]  (7)Notice that the largest squared difference is 2.3438. Therefore, the robust Euclidean distance is defined to be the square root of the sum of the squared distances minus the largest squared difference.u_j=√{square root over (2.5884−2.3438)}=0.4946  (8)According to the invention, the robust Euclidean distance is the Euclidean distance with the largest distance or worst performer removed.

This improved robust kernel function can be extended to a variety of similarity operators. Thus, the L¹-norm distance function can be improved by subtracting the largest city block distance element from the sum of the city block distances:

$\begin{array}{cc}{u}_j=\sum _{i=1}^nx_{q,i}-m_{j,i}-\underset{i=1}{\overset{n}{\max }}\left[x_{q,i}-m_{j,i}\right]& \left(9\right)\end{array}$

In fact, any elemental kernel function (one in which the kernel function scalar output is determined by averaging or otherwise combining the scalar comparison results for each element of the compared vectors) is amenable to the improvement of the present invention, by leaving out contribution of the maximally different element from the kernel function calculation.

It may not always be desirable to leave out the maximally different element with each calculation of similarity between two observation vectors. Therefore, in a preferred embodiment of the present invention, a threshold difference is assigned for each element, and the maximum elemental difference is left out of the distance function calculation only if that elemental difference is greater than the threshold specified for that element. The threshold can be determined in a number of ways, either for each element (sensor) individually, or across all the variables (sensors) uniformly. By way of example, a percentage of the range seen in the data for the sensor can be used as a threshold for maximal elemental difference. Another particularly effective method according to the invention comprises the steps of (a) first scaling data for each sensor to a zero-mean centered range where +/−1 is set to one standard deviation; (b) setting the threshold for excluding maximal elemental difference equal to a multiplier of the standard deviation, e.g., 0.5 times the standard deviation. Furthermore, more than one elemental difference can be excluded from the calculation of the distance metric if they exceed their thresholds. A maximum limit on the number of elements that can be excluded can be set, such that, for example, in a 9-variable model, if 4 elemental differences are larger than their exclusionary threshold, but the maximum cap on excluded elements is 3, then the 3 elements with the largest elemental differences are excluded, and the fourth is included in the distance metric, even though it exceeds its threshold.

The weights described in Equation 3 can also be derived in several alternative ways. Regardless of the exact manner in which the distance metric is used to determine weights, the important aspect is that the weights are greatest (the absolute value of the kernel function is maximum) when the two vectors being compared are identical, and the weights diminish as the two vectors being compared are increasingly different. For example, the weights can be determined according to:

$\begin{array}{cc}w=K\left(u,R\right)=\frac{1}{1+\frac{u^{\lambda }}{R}}& \left(10\right)\end{array}$ where R and the power λ are tuning factors. Another way to determine the weights according to the present invention is:

$\begin{array}{cc}w=K\left(u,R\right)=1-\frac{u^{\lambda }}{R}& \left(11\right)\end{array}$ where again R and the power λ are tuning factors.

Furthermore, the mathematical framework of Equation 4 (AAKR) is just one framework in which the memory vectors can be combined according to the weights w. In the framework of similarity-based modeling, the memory vectors are also compared to each other using the kernel function to produce a M-square matrix G of scalar values for the comparison of the M memory vectors:G=K(M,h)  (12)where M is the matrix formed by all memory vectors as columns, and h is a vector of n bandwidths h for each sensor. The weights can similarly be written in matrix notation as:w=K(U,h)  (13)where w is the weight vector of n weights w_i, U is a the vector of distance function results of input observation x and the memory vectors of M, and h is the bandwidths vector. Then, the estimate vector {circumflex over (x)} can be determined with similarity-based modeling as:{circumflex over (x)}q=M·G⁻¹·w  (14)

EXAMPLE

Data was collected from an operating nuclear power plant steam system and used to compare and evaluate the robust distance metrics. The model is used to monitor steam system sensor calibration at an operating plant and contains 5 plant sensors, primarily from one loop, which include 2 turbine pressure sensors and 3 steam pressure sensors. The quoted sensor units are as follows: 1) turbine pressure in pounds per square inch atmospheric (PSIA) and 2) steam pressure in pounds per square inch gauge (PSIG). The training data for each of the sensor types is shown in FIG. 1. The data presented in FIG. 1 was selected from data collected every two minutes over a two-month period. Overall, the training and test data spans approximately 2 weeks of data observing every 5^th sample or every 10 seconds.

The training data was chosen to be 1,600 observations from steady state plant operation. The test data were chosen to be a successive set of 400 observations sampled from steady state plant operation. The training data were used to develop the empirical models and the test data were used to evaluate the performance of the empirical models.

For completeness, the AAKR model was developed with 800 memory vectors and a bandwidth of 0.5, using the form of Equation 3. The resulting accuracy, robustness, and spillover performance metrics are listed in Table 1.

TABLE 1
Accuracy, robustness, and spillover performance
for compared distance functions.
	Turbine Pressure	Steam Pressure
	#1	#2	#3	#4	#5	Average
Accuracy	Euclidean	0.23	0.60	0.44	0.21	0.29	0.35
	L1-norm	0.08	0.20	0.28	0.07	0.02	0.17
	Robust	0.59	2.80	0.89	0.42	0.36	1.10
	Euclidean
Robust-	Euclidean	0.56	0.63	0.29	0.33	0.37	0.44
ness	L1-norm	0.64	0.73	0.21	0.25	0.24	0.41
	Robust	0.20	0.23	0.23	0.18	0.13	0.19
	Euclidean
Spillover	Euclidean	0.11	0.11	0.18	0.18	0.16	0.15
	L1-norm	0.11	0.12	0.12	0.15	0.12	0.13
	Robust	0.09	0.12	0.06	0.08	0.09	0.09
	Euclidean

Turning to FIG. 2, the accuracy of the respective distance metrics are compared. FIG. 3 shows the respective robustness performance metrics. FIG. 4 shows the respective spillover. Generally, a lower value indicates better performance for the metric. These figures show a decrease in robustness and spillover for the robust distance functions. In other words, the models that use the robust distance functions are less affected by faulty input and are considered to be more robust. This increased robustness is not without consequence though, as all of the variable accuracy metrics (MSE) for the robust Euclidean distance function are larger than those of the model with the L²-norm. Even though there may be an increase in the accuracy metric (predictive error of the model), using the normal L²-norm, the decreases in robustness and spillover metrics using the L¹-norm and robust Euclidean distance more than validate its effectiveness in detecting sensor drift. Ultimately, what is important in equipment fault detection and sensor drift detection is robustness, and accuracy does not need to be exacting. This is because accuracy can represent merely overfitting, in which case the fault may not be identified in a timely fashion. Robustness on the other hand is critical to fault detection as the input becomes increasingly incorrect. A less accurate model may still outperform a more accurate model on fault detection because the error between the estimates and actual values serves as a baseline against which the more robust, less accurate model evidences drifts and deviations better.

FIG. 5 illustrates a monitoring apparatus 500 for monitoring a monitored system 505 and according to an embodiment of the invention. The monitored system 505 may comprise, for example, a fossil fueled power plant environment. A set of sensors 510 monitor various parts, sections, or areas within the monitored system 505. For example, the sensors may monitor temperatures or flow rates at various locations in the monitored system 505.

The sensors 510 provide a query vector, based on the measurements of the set of sensors 510, to the monitoring apparatus 500. The monitoring apparatus 500 includes several devices or modules, such as a non-parametric empirical model 515, a distance estimation engine 520, and a memory 525. The non-parametric empirical model 515 generates estimates of parameter values in response to receiving the query vector of monitored parameter. The distance estimation engine 515 determines robust distances between the query vector and each of a set of predetermined historical vectors for the non-parametric empirical model based on an implementation of an elemental kernel function. The distance estimation engine 515 also determines weights for the monitored parameters based on the robust distances and combines the weights with the predetermined historical vectors to make predictions for the system.

It should be appreciated that a wide range of changes and modifications may be made to the embodiments of the invention as described herein. Thus, it is intended that the foregoing detailed description be regarded as illustrative rather than limiting and that the following claims, including all equivalents, are intended to define the scope of the invention.

Claims

1. A method for making predictions based on a non-parametric empirical model used in monitoring a system, the method comprising: providing a processor;receiving at the processor a query vector of multiple query sensor values for different monitored parameters for the non-parametric empirical model characterizing the system;determining, with the processor, robust distances between the query vector and each of a set of predetermined historical vectors, each historical vector having multiple historical sensor values, for the non-parametric empirical model based on an implementation of an elemental kernel function including: performing an elemental calculation between each query sensor value of the query vector and a corresponding historical sensor value of each historical vector, wherein each elemental calculation results in a single elemental contributor for each pair of corresponding query and historical sensor values,eliminating at least one, but less than all, of the elemental contributors formed from at least one comparison between the query vector and the historical vector, and depending on the values of the elemental contributors,calculating the robust distance between the query vector and the historical vector using a calculation with the remaining elemental contributors;determining weights for the monitored parameters based on the robust distances calculated using the query vector;combining the weights with the predetermined historical vectors to make predictions for the system.

2. The method of claim 1, wherein the non-parametric empirical model is based on an autoassociative kernel regression model.

3. The method of claim 1, wherein the non-parametric empirical model is based on a similarity based model.

4. The method of claim 1, wherein the elemental kernel function is Euclidean distance.

5. The method of claim 1, wherein the elemental kernel function is a city block distance.

6. The method of claim 1, wherein the determining robust distances comprises removing a largest elemental contributor to the elemental kernel function.

7. The method of claim 1, wherein the determining robust distances comprises removing at least one of a set of largest elemental contributors to the elemental kernel function based on a threshold distance assigned for each element.

8. The method of claim 1 wherein at least one elemental contributor is eliminated from the robust distance calculation for each query-historical vector comparison.

9. An apparatus for monitoring an operation of a system characterized by operational parameters, comprising: a non-parametric empirical model for generating estimates of parameter values in response to receiving a query vector of multiple query sensor values for different monitored parameters for a model characterizing the system;a distance estimation engine for determining robust distances between the query vector and each of a set of predetermined historical vectors of multiple historical sensor values for the non-parametric empirical model based on an implementation of an elemental kernel function including: performing an elemental calculation between each query sensor value of the query vector and a corresponding historical sensor value of the historical vector wherein each elemental calculation results in a single elemental contributor for each pair of corresponding query and historical sensor values,eliminating at least one, but less than all, of the elemental contributors formed from at least one comparison between the query vector and the historical vector, andcalculating the robust distance between the query vector and the historical vector using a calculation with the remaining elemental contributors;determining weights for the monitored parameters based on the robust distances calculated using the query vector; andcombining the weights with the predetermined historical vectors to make predictions for the system.

10. The apparatus of claim 9, wherein the non-parametric empirical model is based on an auto-associative kernel regression model.

11. The apparatus of claim 9, wherein the non-parametric empirical model is based on a similarity based model.

12. The apparatus of claim 9, wherein the elemental kernel function is Euclidean distance.

13. The apparatus of claim 9, wherein the elemental kernel function is a city block distance.

14. The method of claim 9, wherein the distance estimation engine determines the robust distances at least partially by removing a largest elemental contributor to the elemental kernel function.

15. The apparatus of claim 9, wherein the distance estimation engine determines the robust distances at least partially by removing at least one of a set of largest elemental contributors to the elemental kernel function based on a threshold distance assigned for each element.

16. The apparatus of claim 9 wherein at least one elemental contributor is eliminated from the robust distance calculation for each query-historical vector comparison.