Operation and maintenance (O&M) costs at commercial nuclear plants are more than three times that of gas turbine plants with these costs accounting for more than 70% of operating expenses. Advanced predictive maintenance practices enabled by AI/ML in combination with a digital twin can provide unique capabilities to automate many of the labor-intensive tasks, the objective being to reduce staffing levels to make existing and advanced reactors more cost competitive.
A basic requirement of an optimum maintenance strategy is the capability to monitor the system in real-time to determine the condition of components. That information can then be used to update the likelihood of component failure that serves as input to a maintenance schedule that is continually adapted given evolving plant conditions.
The current industry-installed capability for on-line diagnosis of component anomalies is limited to anomaly detection using a data-driven digital twin. Nuclear utility experience with these methods has been less than satisfactory. The main issue has to do with reliability with a poor trade-off between sensitivity and false-positive alarms.
Many of the shortcomings can be remedied by including physics in the diagnostic process. This takes the form of a physics-based digital twin which allows specific faults to be identified given an adequate sensor set. The advantages compared to the data-driven approach are listed in the table.
Component Fault Diagnosis
The fault detection and diagnosis problem can be solved using physics-based models constructed for each component in a T-H system to represent component behavior during normal working condition. If a sufficient set of sensor readings is available, the outputs of the component models can be compared against measurement data to detect anomalies in the components using quantitative reasoning.
Process variables not measured can be inferred from available measurements by solving the conservation equations represented by the digital twin. The digital twin embodies the performance of the component as a set of analytic relations representing the physical mechanisms of degradation and performance. These so-called virtual sensors provide information for assessing component performance status. In the figure, a digital twin for the system essentially doubles the number of process variables for which values are known.
The calculation of performance indexes may be straight forward, for example centrifugal pump efficiency calculated from measurements of hydraulic variables. But in cases where there is no direct measurement of the degradation process, such as stress corrosion cracking and mechanical fatigue, the digital twin needs to be a multi-physics calculation. Measurements of stressor variables such as temperature and chemistry are input variables to the calculation while the degree of degradation is output as a virtual sensor.
System Fault Diagnosis
Argonne has developed the software package Parameter-Free Reasoning Operator for Automated Identification and Diagnosis (PRO-AID) that performs real-time monitoring and diagnostics for an engineering system using a form of automated reasoning. The code automatically constructs a digital twin from the fluid system Piping and Instrumentation Diagram (P&ID) and electrical One-Line Drawing as provided by the user. The PRO-AID solution to the monitoring and diagnosis problem is unique while its capabilities and ease of use make for an attractive business case.
- An equipment diagnosis is provided that is consistent with sensor readings and the underlying physics of the faulted plant. A list of candidate faults is not required to be provided a priori.
- Only the plant-specific input contained in the P&ID is required for input, avoiding the need for a subject-matter expert to construct a detailed model to represent a fluid system.
- PRO-AID operates at the system level rather than at the component level. It can distinguish among process fault-induced perturbations throughout the system to identify the actual faulted component.
- The fault diagnosis is “explainable” and understandable by the System Engineer.
The code has been used to analyze the adequacy of sensor set coverage for resolving faults in a system. One such study was for the high-pressure feedwater system shown in the figure of a U.S. utility’s PWR.
An optimum maintenance schedule or time for component replacement schedule aims to reduce the cost of maintenance activities. Existing approaches to predictive scheduling of maintenance are based on a generic likelihood of failure for a component without regard for the current component condition or the mechanistic degradation process as it evolves going forward from current time. LPI, Inc. is partnering with Argonne to develop a software package for maintenance optimization. The estimated future reliability of a component is combined with the cost of maintenance or replacement, the savings by deferring scheduled maintenance, and the cost of an in-service failure. The future reliability of a component is derived from PRO-AID diagnostic information combined with an understanding of the mechanisms of component degradation.
Sensor Set Design
The system sensor set is an important factor that needs to be considered when developing improved maintenance procedures for cost reduction. The adjacent figure suggests how a sensor set might be designed as part of the process of maintenance optimization. When applied to the current fleet this process can identify additional sensors to augment the existing set. For the advanced reactor there may be more opportunity. There the design of the sensor set can be integrated with the design of plant hardware to achieve an optimal maintenance strategy.