Monitoring Performance Under Highly Variable Conditions
Case Study
INTRODUCTION
Etalim Inc., a small powerplant manufacturer recently moved one of their prototype units from a highly controlled test environment to a field test site. The field test site default operating conditions were substantially different than the in-house test cell and included fewer controls on system temperatures and pressure. As a result the customer could no longer rely on simply looking at one output metric to determine the condition of the unit. Machinery Analytics was hired to provide a more robust method of detecting changes in actual performance vs. changes in output due to changes in the environment.
.png)
CUSTOMER PROBLEMS
It’s essential to understand if the unit is degrading
Conditions in the field are highly unsteady, making it difficult to discern trends
Data is stored in multiple, files with different naming and formatting
Data set consists of 3.4M rows and cannot be visualized with existing tools
OUR SOLUTION
Fully Scalable Cloud Data Storage
Can collect system telemetry into a database for years and multiple machines
Browser Based Visualization
Allows viewing of million of lines of data from anywhere
State of the Art Machine Learning
Creates a data driven model of the system the model is used to detect real changes in performance vs. changes in operating conditions
HOW WE DID IT
STEP 1
Move the Data Set
Into a Cloud Database
Files with varying formats were parsed and prepared for database upload, ensuring each value conforms to an appropriate data type [eg float, int, string] as needed. All data was then moved into a secure database hosted on Google Cloud.
STEP 2
Implement Conventional Data Visualization Dashboard
A browser-based data visualization dashboard was implemented. The dashboard allows plotting of any variable on either the X or Y axis. A demonstration of the dashboard is shown below:
Create a Data Driven Model of the System
The primary variable of interest is the power output of the system, as can be seen from the figures below there is a significant amount of variability in the operating condition, and this makes it impossible to conclusively comment on how performance is changing with time.
.png)
Files with varying formats were parsed and prepared for database upload, ensuring each value conforms to an appropriate data type [eg float, int, string] as needed. All data was then moved into a secure database hosted on Google Cloud.
The model fits data with a mean absolute error of 1.6% of rated power. Measured and predicted data is shown in normalized units below.
STEP 3
.png)
STEP 4
Data Monitoring
In the final step we continue to collect data from the system and watch for agreement with the previously created model, continued agreement between the model and data shows that machine performance is not changing.
RESULTS
Once moved to a cloud database, the stored size of the database was nearly 50% less than the stored size of the original files. We implemented a web based data viewer to clearly demonstrate the capabilities of a new piece of hardware. We then developed a data driven model of system performance with a mean prediction error of 1.6%.
When we started the customer had no reliable way of knowing if their machine was working to spec. Now the customer can quantify system health by one look at the comparison between model prediction and actual output.
Are you facing similar obstacles in your organization?
At Machinery Analytics, we believe that skilled professionals should always have the best tools for the job. That’s why we are developing a cloud based, deep learning platform which can easily distinguish changes in actual performance from changes in the environment or operating conditions. Our customers are innovators who want to develop quality products faster.