Plant Maintenance Prediction Using Explainable Failure Rules
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current predictive maintenance methods, such as preventive maintenance and simple heuristic-based approaches, are inefficient due to their lack of optimality and inability to detect abnormalities effectively, often leading to unnecessary downtimes and resource wastage, and lack transparency as they rely on 'black box' machine learning algorithms.
Innovation Solution
An explainable predictive maintenance process using inductive logic programming (ILP) that extracts events from a plant ontology, assigns them to failure or non-failure classes, generates interpretable class expressions describing failure conditions, and initiates maintenance based on predictive accuracy thresholds, ensuring maintenance is timed optimally and targeted.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If preventive maintenance is performed at fixed intervals, then equipment reliability is maintained, but productivity decreases due to unnecessary interruptions and resource wastage
Solution Approach 1:
The system performs preliminary analysis of equipment data using multiple indicators (vibration, temperature, pressure, etc.) and machine learning models to predict potential failures before they occur. This allows maintenance to be scheduled based on actual equipment condition rather than fixed intervals, preventing unnecessary maintenance interruptions while ensuring reliability when truly needed.
Solution Approach 2:
The system continuously monitors equipment data from multiple sensors and provides feedback through a comprehensive risk assessment model. The model updates maintenance recommendations based on real-time equipment status changes, allowing dynamic adjustment of maintenance timing to optimize both reliability and productivity.
2Device complexity
If simple heuristic methods are used for predictive maintenance, then implementation is simple, but measurement precision decreases in detecting abnormalities
Solution Approach 1:
The system merges multiple data sources (sensor data, maintenance history, equipment specifications) and multiple analysis methods (statistical analysis, machine learning models, expert systems) into a unified predictive maintenance platform. This combination enhances abnormality detection precision while maintaining manageable complexity through integrated architecture.
Solution Approach 2:
The system employs a universal machine learning framework that can handle various types of equipment and failure modes through multiple indicators and models. The platform is designed to be adaptable to different equipment types while maintaining consistent high-precision detection capabilities across applications.
3Measurement precision
If black box machine learning algorithms are used, then predictive accuracy improves, but ease of operation decreases due to lack of transparency
Solution Approach 1:
The system introduces an intermediary layer between the black box machine learning models and the user interface. This layer includes explainable AI components that translate model predictions into human-understandable explanations, showing which indicators and factors contributed to failure predictions. This maintains high predictive accuracy while improving transparency and ease of operation.
Solution Approach 2:
The system segments the predictive maintenance process into distinct components: data collection, multiple indicator analysis, machine learning prediction, and explainable output generation. Each segment handles specific functions, allowing the black box models to operate at peak accuracy while the segmentation enables transparent presentation of results to users.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates a method for maintaining a plant comprising the following operations performed by at least one processor: a) extracting events from a given ontology (O) representing knowledge (BG) about the plant equipment and plant properties and about the states of the plant equipment at given points of time, b) obtaining defined failure and non-failure classes, c) assigning at least one of the events (E) to each class by identifying failure and non-failure events from all extracted events, d) applying inductive logic programming (ILP) by feeding it with the failure and non-failure events, whereby the inductive logic programming generates a set of class expressions whereby each class expression (CE) describes under which condition a failure or non-failure of its class occurs, e) outputting, by the at least one processor accessing an output device, a predictive failure occurrence and corresponding failure class estimation to a user in order to initiate user interaction determining maintenance measures for the maintenance of the plant when the plant equipment enters into another state than said states and if the other state matches at least one class expression of the generated set of class expressions.