A proactive approach to anomaly detection using automated tools could detect subtle signs of anomalies in nuclear power plant (NPP) components before they escalate into unexpected equipment failures. Once an anomalous event has been detected and if its cause is not immediately obvious, the next step is to identify the cause of the anomaly so the condition can be corrected. This is currently a labor-intensive (i.e., costly) process made more difficult by the size and complexity of NPP systems. As an alternative, automated root cause analysis (RCA) methods could be used to aid subject matter experts (SMEs) in diagnosing anomalies, which would reduce labor costs, enable more efficient responses to detected events, and prevent unexpected outages. The objective of the present effort was to assess RCA methods in terms of their feasibility for application in NPP time-series data. RCA can be performed via two main approaches. The first is model based, relying on expert- or physics-based models encompassing every type of cause and effect produced at the component, equipment, and process scales. This approach was deemed impractical due to the size and complexity of NPPs. The second approach is data-driven, relying on analysis of patterns within measured data in order to make root cause assessments. Building on the data-driven anomaly detection methods in past research, a data-driven approach to RCA was researched herein. This effort assessed two types of data-driven RCA methods. The first type is contribution-based methods, which assume that the sensor or process causing the anomaly is usually the largest anomaly contributor. These methods use a data-driven model of normal (i.e., anomaly-free) behavior to identify each variable’s contributions to a detected event. The specific methods assessed were an existing method based on dynamic principal component analysis (DPCA), and two novel methods based on leave-one-variable-out (LOVO) models, one each for linear and nonlinear dynamic systems. The second type of data-driven RCA method is time-based. Such methods assume that the sensor or process causing the anomaly shows anomalous signs before other sensors and processes. The time-based method used for the present research is a Granger causality (GC)-based method that uses models of both normal and anomalous behavior to detect problems in upstream processes that cascade down to their dependent processes. An existing method was extended from univariate cases to multivariate cases. All these methods utilize unsupervised machine learning—meaning there are no historical anomalies or cause-effect models to compare against. As such, they rely purely on SME interpretations of the data-driven results in order to identify root causes. To evaluate the methods of interest in a controlled environment, two datasets were used. First, a synthetic dataset based on spring-mass-damper (SMD) systems (commonly found in mechanical engineering references) was used with known anomalies introduced into the system, enabling quantitative comparisons of the methods on both linear and nonlinear datasets. Afterward, a test case using real data from a NPP feedwater system was utilized. The results for the SMD datasets were assessed based on interpretability and repeatability. The interpretability question is a subjective but necessary consequence of using unsupervised and data-driven approaches. Because root causes are highly dependent on the system of interest and there is a lack of previously labeled data, SMEs will always be required to interpret the results. The repeatability question is more quantifiable, and is necessary because the results are stochastic, meaning they are a function of noise and uncertainty. As such, the results should be examined over multiple instances of each possible root cause. The interpretability metric pertains to whether the information provided by the respective algorithm aids a SME in identifying (or narrowing down) the root cause of the anomaly. Per the results for the SMD datasets, both contribution-based and time-based methods qualitatively demonstrated the ability to help SMEs with the RCA task. Regarding the contribution-based methods, this assessment was conducted by considering anomaly contributions as vectors and then determining that the directions did indeed align with some intuitive knowledge of the system. However, for the DPCA method, the assessment was sensitive to a user-selected parameter that proved difficult to select. For the time-based method, the assessment was conducted by ranking calculated features for the anomaly types and noting that features that were intuitively associated with those respective types often appeared either first or second. By contrast, repeatability was quantified via cosine similarity, resulting in an F1 score that served as a measure of both uniqueness and repeatability (i.e., how distinct the average directions for each anomaly type were, and how much variation e
