AI based Optimization for Equipment Maintenance Scheduling
Challenge
Industrial Internet has given rise to availability of sensor data from numerous machines belonging to various domains such as agriculture, energy, manufacturing etc. These sensor readings can indicate health of the machines. This has led to increased business desire to perform maintenance of these machines based on their condition rather than following the current industry practice of time-based maintenance. It has also been shown that condition-based maintenance can lead to signi cant nancial savings. Such goals can be achieved by building models for prediction of remaining useful life (RUL) of the machines, based on their sensor readings.
Below are some of the milestones for creating an on-demand or condition based maintenance strategy:
-
Provide optimal windows for operational duration.
-
Provide novel operational insights for equipment.
-
Proactively identify potential failure modes and time to failure before planned maintenance windows.
-
Improve decision-making by increasing visibility into equipment health, including predictions of performance degradation and/or equipment failure.
Solution
To achieve the desired on-demand approach for optimizing the maintenance schedule, a multi fault simultaneous Prognostics system will be created. This Prognostics framework will consist of the following: Anomaly detection capturing abnormal operation, allowing experts to define and discover more faults. Fault Classification to detect specific faults and their time to failure. Remaining Useful Life Estimators to provide time to failure counters for each type of fault.
Approach
Our approach is to create two main solutions that can be hosted on cloud or on premise.
-
Maintenance Optimization Framework that allows the client to correctly forecast future failure events.
-
Full prognostics framework that allows Client SME identify new failures and detect early degradation.
Maintenance Optimization
Framework for optimal operational duration. To provide optimal maintenance scheduling, we will generate time-to-failure algorithms, also known as Remaining Useful Life (RUL) estimations. The output of RUL estimations will then be overlain with the equipment run hour forecasting algorithms. By overlaying plant forecast run hours with time-to-failure counters, maintenance scheduling can be optimized. This will properly place the future time window when outages are more likely needed. In addition to building and designing prognostics algorithms. The run hours forecast algorithm uses statistical techniques including ARIMA (Auto Regressive Integrated Moving Average)
Prognostics Framework
To provide optimal maintenance scheduling, we will generate time-to-failure algorithms, also known as Remaining Useful Life (RUL) estimations. Our goal is to leverage Remaining Useful Life estimation (RUL) models for each fault type identified within the target list of pumps. RUL estimation will serve as an “odometer” for maintenance scheduling. To estimate RUL with variable operating conditions (e.g. BHA mud motors, Turbines, pumps, Wind Turbine blades, and gearboxes). This approach relies on a Deep Learning technique called Long Short Term Memory (LSTM) Autoencoders.
In addition to estimating RUL, we would build a semi-supervised early degradation detection algorithm using Deep Learning based Autoencoders and One Class Classification algorithms. This approach will make the Prognostics system more complete. Early detection algorithm will help identify early onset of system faults and therefore harden the time to failure estimation. This early detection process will also allow identifying and adding more faults to the system in the future as the usefulness of the platform is expanded.
Demo
Conclusion
Here at BKO services, we use your operations data along with some of the state-of-the-art machine learning techniques to transform your organization from periodic maintenance strategy to on-demand maintenance strategy and help save Millions of dollars annually.
Digitizing Pre-Sales for Small Scale LNG Plants
Challenge
Our Client, an energy company focused on producing and providing LNG, wanted to improve their sales engagement with their clients. This would help them improve their sales process and customer engagement. The client wanted to show prospective clients a life like experience of what it takes to own and operate LNG processing plants.Until then our client was relying on no more than outdated power points. This created an opportunity to vastly improve the value from each sales engagement.
Solution
It’s expected that our work will transform our client’s sales presentation to their customers and will be attributed with helping our client close the sale for some of their LNG plants. The solution BKO Services developed was to build and deploy a modern 3D Sales Model application. The model starts with a zoomed-out aerial view of the customer site, and project an image of LNG equipment configuration on the aerial image. The application allows the user to zoom in on the aerial image and arrive to “street view” and walk the site.
Approach
Our approach was split into 3 branches.
-
Sales Models:
Customer Application to deliver a polished 3D Sales Model presentation to the customer.
-
Custom Design:
User Interface that allows customizing LNG plants to the respective users and configurations.
-
Solid Infrastructure:
Back-end infrastructure to provide the infrastructure, integration, security & documentation needed to effectively manage the User Interface and Customer Application.
Customer Application
The goal of the Customer Application is to deliver a polished 3D Sales Model presentation to the customer. The basic model will start with a zoomed-out aerial view of the customer site. The application will project an image of client’s equipment configuration on the aerial image. Users can zoom in on the aerial image and get to 3D sales model image “street view”. Users will be able to continue to zoom in, and navigate 360 degrees around the 3D template.
User Interface
The goal of the User Interface is to customize the Customer Application so it is tailored to each customer’s configuration. The User Interface will enable client sales employees to enter all relevant data into a display. Once information is entered and submitted, the application will essentially build a new plant model each time. Once created, the plant model will exist for use throughout the customer sales process.
Back-end
The goal of the Back-end deliverables is to provide the infrastructure, integration, security & documentation needed to effectively manage the User Interface and Customer Application.
- Hosted Azure Platform on Client’s Azure account
- Application and UI built with Unity WebGL
- Development platform
- Map API Integration
- SQL Server Database
- API/Middleware Integration
- Role-based Security using AD Authentication
Conclusion
BKO Services built an innovative digital solution that optimized business development for an LNG Client. Gone are the days when the sales team have to explain to their customers what this new multimillion dollar LNG investment is going to provide. They simply now show their customers what their new plant will look like, where it will be located and take them on a journey to interact with their equipment.
Hydrocarbon Prospecting with Machine Learning
Challenge
Our Client, an Oil &Gas E&P company, wanted to maximize their development investments in their unconventional onshore basins. This meant improving their capabilities for hydrocarbon potential across any basin they have interest in.
Solution
BKO Services provided a predictive model for hydrocarbon potential. This was intended to serve 3 principal purposes:
- Quantify any interaction between the reservoir and completions.
- Provide probabilistic guidance to optimize new completions.
- Estimate productivity away from known production, with optimal completions design showing P10, P50 and P90 scenarios.
- To achieve those goals, BKO applied Machine Learning and public data augmentation to deliver a map of expected reservoir performance per formation. The solution also included up to 20 P10, P50, P90 completion designs for each formation.
Approach
To help deliver this solution to our client we followed a workflow that relied on best practices.
-
Data Exploration
Begin by carefully exploring the data, We used series of visualization techniques to understand data ranges, variable relationships, and potentially problematic data.
-
Data Cleansing
Production Preprocessing: For a variety of reasons, not all data is meaningful. Armed with the understanding in step 1 (data exploration), we identified the highest quality data needed for the analysis. We also Computed an economic BOE = Oil + 1.1*Gas / 20, Accumulated volumes and “days online” and Interpolated data to a consistent “Cum Oil At 365 Days”
We then proceeded to augment the dataset to improve the predictive power of our models.
-
Predictive Modeling
Holdout: As part of the modeling work, we performed a Holdout Process. One can think of ML as an automatic hypothesis generator. To get some confidence that the model can predict production at a new location, we “hide” some data and try and predict. We then test the hypotheses 10s, 100s of times and look at prediction performance.
Model evaluation on holdout:
Scenario testing: Our machine learning algorithms produced maps based on P10, P50 and P90 completions designs. A suite of possible completions parameters was then ran to highlight what completions may be most effective in different areas.
If completions were held constant, what would my expected production be?
If I pumped a 10% larger job?
What uplift should I expect?
Is the uplift the same everywhere?
This gave the client the ability to interrogate the algorithm and ask questions like:
Conclusion
To help deliver this solution to our client we followed a workflow that relied on best practices.
Data Exploration
BKO Services developed a predictive model for hydrocarbon potential across the client’s Basin. The model relied on client’s subsurface, engineering, and well performance data. This enabled the client to reliably iterate over well economics and optimize well completions, therefore improving the development commercial valuation.
Optimizing Power Generation in Deregulated Markets
Challenge
Our Client, a large Independent Power Producer, wanted to improve the power generation marketing for the plants. This would optimize the future commercial value of the power plants. The challenge is that a typical plant might have 288 different generation points, each has its own revenue and cost. Those 288 daily offers are tied to the operational profile of the equipment and vary with weather and equipment changes. Until now our client was relying on no more than outdated curve on a spreadsheet. This created an opportunity to vastly improve the overall profit from power generation.
Solution
Modern Power Plants are very efficient so are the de-regulated markets they operate in. With all that efficiency, complexity has come along for the ride. In a deregulated power market, most plants can offer their generation capabilities and their costs to a clearing house called Day Ahead Market, DAM.To address the modeling complexity, we created a power generation marketing and planning system for each plant centered around machine learning model representations of the plants.The end result was a robust process that improved the enterprise financial goals through optimized power plant dispatching and generation planning.
Approach
Our approach was to create a two- step solution.
-
Use Machine Learning and Thermodynamics Simulations to create a digital twin representation of each plant capabilities.
-
Build an interface for plant operators to forecast future generation capabilities based on local weather forecast.
Results are then shared with energy dispatchers to monetize those capabilities.
Data Cleansing
Production Preprocessing: For a variety of reasons, not all data is meaningful. Armed with the understanding in step 1 (data exploration), we identified the highest quality data needed for the analysis. We also Computed an economic BOE = Oil + 1.1*Gas / 20, Accumulated volumes and “days online” and Interpolated data to a consistent “Cum Oil At 365 Days”
User Interface
The goal of the User Interface is to customize the Customer Application so it is tailored to each customer’s configuration. The User Interface will enable client sales employees to enter all relevant data into a display. Once information is entered and submitted, the application will essentially build a new plant model each time. Once created, the plant model will exist for use throughout the customer sales process.
For most generators this sort of forecasting of the plant capabilities can be out of reach. To make this scalable we designed a turnkey solution that went from plant data collection all the way to dispatch interface and automation of generation planning process . We implemented a monitoring process that made those models adaptive and therefore increased the fidelity of the generation forecasting.
Conclusion
BKO Services designed and built a machine learning based solution for an Independent Power Producer that will enable the IPP to improve profitability by several millions of dollars annually.