Source-linked AI summary
Scalable and Data-Driven Decision Support in the Maintenance, Repair, and Overhaul Process
Houkun Zhu, Helena Ebel, Dominik Scheinert, Florian Schmidt, Jens Altenkirch, Odej Kao
TL;DR
Manual, static MRO procedures can be inefficient, and component-specific repair knowledge must be captured for wider use. SES combines digital RSM rules and data-driven repair suggestions in a scalable system; it reports repair-prediction results and reduced latency through scaling.
Problem
Manual, static repair procedures are inefficient and error-prone, while component-specific knowledge accumulated by experts needs to be documented and shared.
Source: §I · p. 1Method
SES provides decision support through digital RSM application, historical-decision models for defects outside RSMs, and repair-cost and feasibility estimates.
Source: §I · p. 1, 1) Digital Repair Specification Manual · p. 3Results
SES achieved 88.5% F1 score and 5.0% under-estimation ratio, and scaling out mitigated and could prevent latency increases.
Source: A. Prediction Performance · p. 4, §VI · p. 5, 2) Different Scale-out Under Same Loads · p. 5Takeaways & Limitations
SES automates parts of the conventional MRO process while assisting experts with defects undefined in the RSM and adapting to varying demand.
Source: §VI · p. 5
Abstract
from arXiv · showhide
Several businesses apply maintenance, repair, and overhaul (MRO) principles to the life-cycle of their existing products. In cases like casted gas turbine component Product Lifecycle Management (PLM), repairing components in frequent intervals can extend the lifetime expectation of the product, provide higher cost efficiency compared to newly produced components, and even improve the part design during the repair cycle. Another aspect of repair concerns sustainability, as products often contain rare materials. The emissions produced by the repair process are usually smaller than mining materials and casting new components. To optimize the repair process further, we propose the Smart Expert System (SES), which assists engineering experts with machine learning-based decision support throughout the repair process. We elaborate on its IT architecture and present machine learning models employed for representative MRO use cases. The SES is evaluated using actual industry data from a leading gas turbine company and demonstrably fulfills formulated requirements concerning the suitability of the overall decision support and the stability of the enclosing IT architecture.
I. INTRODUCTION
MRO can improve gas-turbine component life-cycle performance, but conventional repair relies on manual, static decisions and knowledge that must be captured for wider use. SES offers machine-learning-based decision support for several repair tasks and a scalable architecture.
I. INTRODUCTION
- MRO is used in gas-turbine refurbishment to minimize life-cycle costs and optimize part performance; maintenance is also of interest because components may outlast their own limited life expectations.Source: §I · p. 1
- Manual, static repair procedures can be inefficient, error-prone, and slow to reflect infrequently updated Repair Specification Manuals (RSMs).Component-specific knowledge about novel issues and complex repairs also accumulates with individual experts and needs to be documented for wider use.Source: §I · p. 1
- SES applies data-driven decision support to automate repair tasks and improve the extraction and integration of expertise from unseen repair operations.Its use cases cover digital RSM application, disposition inference for defects not defined in RSMs, and repair-cost and feasibility estimation.Source: §I · p. 1
- The paper presents an ML-based MRO decision-support solution with fine-granular component scaling and reports high repair-operation prediction performance and reliable handling of varying loads through horizontal scaling.Source: §I · p. 1
II. PROBLEM STATEMENT
The problem statement derives digital-assistant needs from the conventional MRO process, where inspection findings and RSM rules guide repair decisions, while rare or complex defects require expert investigation.
II. PROBLEM STATEMENT
- The problem statement uses the conventional process to derive problems and requirements, including potential automation through digital RSM rules or decisions inferred from historical data.QI decisions are recorded in PLM, and RSMs are periodically adapted, commonly every one or two years.Source: §II · p. 2, A. Conventional Engineering Process · p. 2
A. Conventional Engineering Process
- A used component is inspected for damage, then an engineering expert uses RSM rules to assess severity and choose repair, retention, or scrapping.Source: A. Conventional Engineering Process · p. 2, Figure 1 · p. 2
- Rare or complex defects may fall outside standardized RSM patterns, prompting Quality Inspection in which experts determine a repair process from their experience.Source: A. Conventional Engineering Process · p. 2
B. Requirements for an Assistant System
Practitioner interviews and user stories yield four core requirements for an assistant: preserve engineer oversight, use decision-relevant information, adapt models, and make reasoning understandable.
B. Requirements for an Assistant System
- The assistant should suggest decisions for engineers to review, rather than decide automatically, and should incorporate their feedback to improve models.Source: B. Requirements for an Assistant System · p. 2
- The system should access and incorporate decision-relevant data and expert knowledge, linking data sources with intermediate and final results usable in MRO.Source: B. Requirements for an Assistant System · p. 2
- Because unseen conditions and changing inputs can exceed model limits, the system should regularly retrain models to adapt.Source: B. Requirements for an Assistant System · p. 2
- The system should provide engineers understandable explanations of its reasoning, including the decision process and prediction confidence.Source: B. Requirements for an Assistant System · p. 2
III. ANALYSIS PIPELINES
SES uses a distributed architecture to support data-driven suggestions across engineering phases. Its layered design separates request handling and modeling with a transportation layer between them.
III. ANALYSIS PIPELINES
- The analysis pipelines describe the SES design in response to the problem statement and its requirements.Source: §III · p. 2
A. Architecture Design
- SES is a distributed system of loosely coupled components intended to support scalability and maintenance of individual components.Source: A. Architecture Design · pp. 2–3
- The six-layer architecture routes requests through ingestion and transportation to prediction services, while persistence and MLOps support feedback and model operations and distribution returns processed results.Containerized components can be scaled individually to meet changing request demand or accommodate additional ML models.Source: Figure 2 · p. 2, A. Architecture Design · p. 3
B. Use Cases
The SES addresses repair decision support through two use cases: digitizing the RSM and applying machine learning to Quality Inspection decisions.
B. Use Cases
- The use cases are intended to improve the conventional engineering process.Source: B. Use Cases · p. 3
1) Digital Repair Specification Manual:
- The SES converts the Repair Specification Manual into machine-readable form so engineers can receive repair-operation recommendations directly from defect measurements.Source: 1) Digital Repair Specification Manual · p. 3
2) Quality Inspection:
- For Quality Inspection, decision trees learn from experts’ past decisions to generalize expertise and automate decisions for defects not covered by the RSM.Experts investigate these defects case by case, and their expertise is generalized every one to two years for integration into the RSM.Source: 2) Quality Inspection · p. 3
3) Repair Scope:
The SES digitizes repair-scope logic for individual shop floors and customers, while its design addresses requirements for engineer oversight, interpretable data and models, adaptation, and understandable predictions.
3) Repair Scope:
- Repair scope captures defect severity and repair complexity to estimate cost and decide whether a component is worth repairing; the logic varies by shop floor and customer.The SES digitizes this logic for each shop floor and customer.Source: 3) Repair Scope · p. 3
C. Fulfillment of Requirements
- Engineers retain final decision authority: they can correct unsuitable recommendations, and their corrections are returned to the SES for model fine-tuning.Source: C. Fulfillment of Requirements · p. 3
- The SES uses existing RSM documents and develops its models with engineers to incorporate the process and meet engineers’ interpretability requirements.Source: C. Fulfillment of Requirements · p. 3
- Decision trees can adapt to new conditions through periodic retraining from scratch or fine-tuning a warm-started model.Source: C. Fulfillment of Requirements · p. 3
- The system can provide prediction certainty alongside each model prediction to make results more understandable to engineers.Source: C. Fulfillment of Requirements · p. 3
IV. EVALUATION
The evaluation examines SES prediction performance and end-to-end latency. On real industry data, conservative prediction lowered underestimation while incurring a small F1-score drop.
A. Prediction Performance
- The Quality Inspection training data are imbalanced, with half of the examples labeled “use as is.”Source: Figure 3 · p. 4, A. Prediction Performance · p. 4
- Conservative prediction favors more serious dispositions when their probability exceeds a threshold; with a 30% threshold and probabilities of 40% use as is, 22% rework, and 38% scrap, it selects scrap.The approach addresses skewed disposition data and the preference for cautious predictions.Source: A. Prediction Performance · p. 4
- SES evaluates predictions with F1 score and defines underestimation as predicting use as is or rework for scrap, or use as is for rework.Source: A. Prediction Performance · p. 4
- Underestimation ratio fell from 8.0% to 5.0% at a cost of a 0.5% drop in F1 score; conservative prediction achieved 88.5% F1 score with a 5.0% underestimation ratio.The evaluation data came from a leading gas turbine company.Source: A. Prediction Performance · p. 4
B. End-to-End Latency
The end-to-end latency experiments assess SES processing speed and scalability under different request loads and scale-out levels.
B. End-to-End Latency
- One experiment runs all SES components with two instances and sends requests at different pressure levels.Source: B. End-to-End Latency · p. 4
- A second experiment uses 50 request threads and varies the scale-out level, keeping all components at the same instance count except the message broker.Source: B. End-to-End Latency · p. 4
- Each setting is run five times, with request latency measured during the first 60 seconds.Source: B. End-to-End Latency · p. 4
1) Same Scale-out Under Different Loads:
The pressure test holds SES scale-out at two instances while increasing request threads to create greater load. Latency rises sharply as thread count grows.
1) Same Scale-out Under Different Loads:
- Latency rises from around 20 seconds to almost 80 seconds as request threads increase from 10 to 200.Source: 1) Same Scale-out Under Different Loads · p. 4, Figure 4 · p. 4
- The experiment varies the number of request threads to measure SES latency under different loads.Source: 1) Same Scale-out Under Different Loads · p. 4
- Figure 4 fixes scale-out at two instances; increasing thread count generates greater load.Source: Figure 4 · p. 4
2) Different Scale-out Under Same Loads:
Under a fixed 50-thread load, scaling out SES decreases latency, and the experiments indicate that SES adapts to demand surges. The paper contrasts this decision-support focus with related work emphasizing manufacturing interaction and assistance systems.
2) Different Scale-out Under Same Loads:
- Scaling out SES noticeably decreases latency, and both experiments indicate adaptation to demand surges to ensure process efficiency.Source: 2) Different Scale-out Under Same Loads · p. 5
V. RELATED WORK
- Related work explores local and global assistants, cloud-assisted agents for edge environments, and systematic system reconfiguration,,.Source: §V · p. 5
- Many engineering-assistance studies focus on augmented reality, including voice-supported assistance, 3D-model inspection, and digital manuals,,,,.Source: §V · p. 5
- Reviews classify production worker-assistance systems and discuss manufacturing assistance benefits such as accessibility, speed, customization, and interaction,.Source: §V · p. 5
- Against work focused on reinventing manufacturing interaction, SES emphasizes data-driven engineer decision support through interactive suggestions.Source: §V · p. 5
VI. CONCLUSION
The conclusion presents SES as a scalable, data-driven assistant for gas turbine repair, reporting prediction performance and latency adaptation through scale-out. Transferability across repair objects remains future work.
VI. CONCLUSION
- SES assists gas turbine repair with machine-learning decision support, including for defects undefined in the RSM.Source: §VI · p. 5
- The reported prediction results are 88.5% F1 score and 5.0% under-estimation ratio.Source: §VI · p. 5
- SES can mitigate and even prevent latency increases by scaling out; evaluating and strengthening model transferability to different repair objects is future work.Source: §VI · p. 5