PRO-AID

Parameter-Free Reasoning Operator for Automated Identification and Diagnosis (PRO-AID) combines physics-based diagnostic tools with artificial intelligence to detect potential issues in nuclear facilities with unprecedented precision and provide clear explanations to operators in real time. The technology represents a significant advancement in nuclear plant operations, where safety and efficiency are paramount concerns. 

PRO-AID is creating a digital twin of reactor systems, merging physics-based modeling with data-driven machine learning to provide component-level fault diagnostics with unprecedented accuracy. 

What sets PRO-AID apart is its self-learning capability that automatically calibrates model parameters during operation, adapts to system changes, and continues functioning even during sensor failures. The system communicates with operators using natural language processing through large language models (LLMs), translating complex technical diagnostics into clear, actionable information. 

Benefits and Advantages

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.

• Plant-specific input is limited to the piping and instrumentation diagram (P&ID). This avoids the need for a subject-matter expert to construct a detailed model of the engineered system.
• False alarm diagnoses are minimized by requiring that the internal representation of the underlying physics in the algorithm be consistent with the observed sensor readings.
• The fault diagnosis is immune to fault-induced process perturbations that can propagate through a system of components. This is achieved by treating the engineered system in its entirety rather than as a collection of standalone components taken in isolation of each other.
• The fault diagnosis is ​“explainable” and understandable by the System Engineer. The inclusion of process physics in the underlying method provides for this.
• A list of candidate faults is not required to be provided.
• The fault diagnosis is ​“explainable” and understandable by the maintenance engineer. The user can query the code about the reasoning that led to a fault diagnosis and the code returns output from a large language model (LLM) that describes how the code arrived at its answer. This mimics the fault resolution process a maintenance engineer performs upon an indication of an anomaly.
• Inputting a system description to PRO-AID by the user is made efficient by reducing the manual effort required. Generative AI is used to automate the conversion of a visual image of the Piping and Instrumentation Diagram (P&ID) to generate the code-specific syntax and structure that describes the application.

In testing at Argonne’s Mechanisms Engineering Test Loop Facility (METL) and on various light-water reactors, PRO-AID successfully diagnosed faulty sensors and system anomalies while providing ranked probabilities of potential faults. This innovation addresses critical industry challenges like reducing maintenance costs, minimizing downtime, and supporting workforce transitions, arriving at a pivotal moment as AI’s energy demands drive renewed interest in nuclear power. 

By combining advanced diagnostics with natural language explanation capabilities, PRO-AID bridges the gap between complex technical data and human decision-making, potentially reducing costs, increasing plant reliability, and supporting the next generation of nuclear energy systems.

Technical Concepts

PRO-AID consists of the following major components:

• System model, taking input from the user supplied Piping and Instrument Diagram (P&ID) for the engineering system of interest, automatically assembles a physics-based model from a library of plant components that include flow branch and junction, valve, single-phase heat exchanger, condensing heat exchanger, centrifugal pump, induction motor and other components.
• Virtual sensor model, taking input from the system model and measurements, derives the values of unmeasured process variables to yield an augmented sensor set that provides more detailed knowledge of the current state of the physical system.
• Automated reasoning module, taking input from the system model and the augmented sensor set, uses a forward chaining procedure to diagnose component faults and sensors that are out of calibration or failed. If a sensor is determined to be failed, it is automatically removed from the list of physical sensors followed by a reinitialization of the list of virtual sensors. The reasoning process incorporate a generalized likelihood ratio test to include the effects of system model uncertainty and measurement uncertainty on the diagnosis.
• Rank-ordered fault probability module, taking output from the automated reasoning module, determines when anomaly thresholds are exceeded and generates a list of candidate faults rank ordered from greatest to least probable.
• Large Language Model (LLM) enhances diagnostics in PRO-AID by translating complex, physics-based fault detection into clear, human-understandable explanations.
• Visual Language Model (VLM) composes the input file that defines P&ID components, sensors, and connections based on the VLM’s understanding of code-specific syntax and structure acquired from the code documentation and references P&IDs.

Applications

PRO-AID has been used to perform anomaly diagnosis in nuclear plant applications that include operator human factors studies conducted with nuclear plant operators using a full scope plant simulator and equipment performance monitoring using plant data sets from two U.S. nuclear utilities.