AI-Powered Leak Detection
A proposal for developing in-house leak detection capabilities using artificial intelligence — replacing vendor dependency with internal expertise and data ownership.
Background
This proposal comes from a unique intersection of experience:
- AI & Machine Learning: Background in Computer Programming with active experience building Convolutional Neural Networks and machine learning systems for object detection.
- Leak Detection: 20 years at a pipeline control center, including 8 years as Shift Supervisor where the primary focus was leak detection and related processes.
Opportunity
There is a real opportunity to enhance both safety and cost-efficiency by implementing an AI-powered leak detection system.
Save Money
- Avoid recurring costs by developing and maintaining your own leak detection algorithm
- An advanced leak detection system could lead to lower insurance premiums due to reduced risk
- A more precise system would significantly reduce false alarms — fewer after-hours calls for leak detection engineers, improving work-life balance and reducing overtime costs
Improve Safety
Current leak detection systems are threshold-based. While reliable to an extent, alarms are only triggered when a specific amount of product loss is detected over a defined period. Transitioning to a pattern-recognition-based system — either independently or alongside the existing setup — offers significant advantages:
- Increased Detection: Recognizing subtle patterns can identify leaks that go undetected by threshold-based systems
- Faster Response: Early detection translates into quicker response times, minimizing environmental impact and operational downtime
- Improved Accuracy: AI excels at filtering out noise in the data, leading to more accurate leak identification compared to human observation alone
The Data Advantage
When we engage third-party vendors, we essentially hand over our valuable data, enabling them to build algorithms that they can then market to our competitors. The true value lies in the data itself, not necessarily in the complexity of model development.
The foundation of any successful neural network is data. Building in-house means the data — and the competitive advantage it creates — stays in-house.
Implementation Plan
1. Data Collection
Extensive leak detection data spanning multiple pipeline systems and years is a tremendous asset for training a powerful machine learning model.
Simulators can generate diverse failure scenarios, providing an invaluable resource for training. The ability to simulate a wide range of leak situations is crucial for developing a robust and reliable system.
2. Data Labelling
Using a system like CVAT, in-house experts label the data by drawing boundary boxes around leak signatures.
As the model learns from labeled data, it can begin to automatically label new data, requiring only expert review for quality control. This iterative process allows for rapid model improvement and scalability.
3. Model Training
Determining the most effective approach:
- Image-based Analysis (CNN): Training the model to recognize visual patterns of leaks in images
- Raw Data Analysis (RNN): Training the model to identify leak signatures within raw sensor data
Azure provides a readily available and cost-effective platform for training.
4. SCADA Integration
For a CNN model, integration could be achieved through a dedicated terminal monitoring leak detection patterns generated by the existing SCADA system. The CNN analyzes patterns in real-time, leveraging its training to identify anomalies.
5. Maintenance & Iteration
The initial model provides a strong foundation. Continuous improvement through:
- Incorporating New Data: Regularly feeding the model with new data enhances accuracy over time
- Refining Labels: Ongoing review and refinement leads to better pattern recognition
- Iterative Updates: Regular model updates ensure optimal performance as conditions evolve
Next Steps
- Cost-Benefit Analysis: Thorough comparison of in-house development costs versus third-party solutions — factoring in licensing fees, implementation costs, and ongoing maintenance
- Liability Assessment: Identify and address potential legal or regulatory implications related to data security, privacy, and the use of AI in leak detection
- Resource Evaluation: Determine the optimal team structure, allocate necessary resources, and secure leadership support