About client
Our client is a largest global manufacturer of industrial vacuum pumps used across semiconductor, pharmaceutical, chemical, and advanced manufacturing industries. These pumps are mission-critical assets where performance, uptime, and maintenance efficiency directly impact critical semiconductor production outcomes.
A key factor influencing pump performance is oil condition. Oil degradation affects reliability, efficiency, and service intervals. Traditionally, oil health assessment relied on manual visual inspection through pump sight glasses during maintenance cycles.
However, this manual approach was subjective, inconsistent, and limited in its ability to detect early degradation patterns or support predictive maintenance strategies.
The Challenge: Why manual oil monitoring was no longer sustainable
Oil condition monitoring is critical for ensuring pump reliability and avoiding costly downtime. When the process relied on manual inspection and experience-based judgment, several operational challenges emerged:
- Oil condition assessment was subjective, leading to inconsistent decisions across technicians
- Early signs of degradation were often missed, increasing failure risks
- Maintenance followed fixed schedules instead of actual equipment condition
- No structured way to correlate oil health with operational data
- Limited visibility into Remaining Useful Life (RUL) of oil and components
These limitations prevented the organization from transitioning toward a more intelligent, data-driven maintenance model.
From manual inspection to AI-driven oil health monitoring
| Before |
After |
| Oil condition was visually inspected through sight glass, relying on technician judgment. |
Oil images are analyzed using AI models to classify oil health objectively. |
| Assessment varied across individuals, leading to inconsistent maintenance decisions. |
Standardized classification ensures consistent and reliable oil health evaluation. |
| Maintenance schedules were time-based rather than condition-based. |
Maintenance decisions are driven by actual oil condition and predicted degradation. |
| No visibility into oil degradation trends or failure risks. |
AI models provide insights into degradation patterns and Remaining Useful Life (RUL). |
| No digital system to analyze or store oil condition data. |
A centralized portal enables image upload, analysis, and result visualization. |
Solution: AI-based oil condition monitoring using computer vision
To address these challenges, the client partnered with Saviant to develop a computer vision-based AI solution that digitizes and standardizes oil condition monitoring. The solution is designed as a scalable digital capability aligned with modern industrial product innovation and predictive maintenance goals.
01. AI model for oil health classification
A Convolutional Neural Network (CNN) model was developed using Azure Custom Vision to analyze oil sight glass images.
- Trained on ~1,000 historical oil images
- Images labeled into three categories: Healthy, Degrading, Critical
- Integrated metadata such as pump ID, operating hours, and service history
- Provides classification results with prediction confidence scores
The model enables automated, objective oil condition assessment, eliminating reliance on manual inspection.
02. Remaining Useful Life (RUL) estimation
In addition to classification, the solution introduces predictive insights through Remaining Useful Life estimation.
- Historical operational data analyzed to determine oil degradation trends
- Oil degradation rates calculated over time
- Estimated remaining operating hours before oil replacement or failure risk
This shifts maintenance from reactive or scheduled to predictive and condition-based.
03. Web-based analytics portal
A web application was developed to make the AI capability accessible to end users. Operators and engineers can:
- Upload oil images for analysis
- View classification results with confidence levels
- Monitor oil health trends
- Enable data-driven maintenance decisions
The portal acts as the foundation for future integration with broader equipment monitoring systems.
Technology foundation
The solution was built using a modern, scalable technology stack aligned with industrial digital transformation needs:
- AI/ML: Convolutional Neural Networks using Azure Custom Vision
- Data inputs: Image datasets + operational metadata
- Application layer: Web-based portal for user interaction
- Cloud platform: Azure-based architecture for scalability and deployment readiness
Saviant’s engagement approach: Built to solve the right problem first
The engagement was executed as a Proof of Concept (PoC), emphasizing rapid validation and measurable business value.
- Started with problem and value discovery & value design
Saviant AI consulting & development team worked closely with the client’s engineering and domain experts to understand oil degradation behavior, current inspection limitations, and define where AI could deliver measurable value in improving maintenance decisions.
- Identified the right data and approach
Relevant datasets were curated, cleaned, and labeled by combining historical oil images with operational metadata, ensuring the foundation was aligned to real-world conditions and domain expertise.
- Built and refined what mattered first
Our initial focus was on developing an AI model to accurately classify oil health. The model was iteratively improved through multiple training cycles and continuous validation with subject matter experts.
- Validated through real-world usage
An interactive web portal was developed to allow stakeholders to upload images and test predictions, enabling early validation, feedback, and confidence in the solution.
- Planned for scalable productization
The solution architecture was designed with future scalability in mind, laying the groundwork for transitioning from a Proof of Concept to a production-ready predictive maintenance platform.
The impact so far
- Potentially reducing failure of pumps due to contaminated or degraded oil by 30%
- Reducing expenditure on warranty by 20% in 1 year
- Successful AI model capable of classifying oil health from images
- Standardized and objective oil condition assessment
- Early detection of degradation patterns enabling proactive action
- Data-driven estimation of Remaining Useful Life (RUL)
- Interactive portal enabling real-time analysis and validation
- Scalable architecture ready for integration into industrial IoT ecosystems
By digitizing oil monitoring and introducing AI-driven insights, the client has taken a significant step toward intelligent, condition-based maintenance.
Current adoption and roadmap
Following the successful PoC, the client is progressing toward expanding the solution’s capabilities and industrial adoption. Key focus areas include:
- Enhancing model accuracy with larger and more diverse datasets
- Expanding use across different pump types and environments
- Integrating with existing equipment monitoring platforms
What’s next: Toward full predictive maintenance
The next phase focuses on building a multivariate predictive model that combines multiple data sources:
- Oil image analytics
- Laboratory testing data
- ISO particle contamination metrics
- Pump vibration sensor data
- Temperature and operational runtime data
This will enable:
- More accurate oil health predictions
- Improved Remaining Useful Life estimation
- Early detection of potential pump defects
The roadmap also includes:
- Advanced anomaly detection capabilities
- Deeper integration with connected equipment ecosystems
- Transition from PoC to a production-ready predictive maintenance platform