Context
As a world leader in the design of special devices for harsh operating conditions (particularly in the nuclear industry) and industrial applications, we design valves for which we want to be able to ensure their proper operation
These valves play an essential role in various industrial sectors, including energy, oil & gas, chemicals, shipbuilding and nuclear power. Their main purpose is to control the flow of fluids (liquids, gases, vapors) through complex piping and systems.
Valves are crucial to the safety of these installations. They are designed to open or close automatically in the event of a system failure, in order to prevent dangerous incidents.
For this project, and given the criticality of the applications, it was impossible to carry out tests under real conditions in order to build a model that could identify valve malfunctions. For this reason, the valve was instrumented and tested on a test bench, stimulating several openings and closures under different operating conditions (temperature, pressure, etc.).
Faults were also artificially created by increasing the stress on certain parts of the valve.
Need
In order to offer a premium-quality product and, above all, to provide new services to its end-user customers. The manufacturer wanted to use Artificial Intelligence (AI) and Machine Learning (ML) to detect faulty valve operating cycles or determine the sensors indicating these failures.
Thanks to these indications, the customer user would be able to determine the root cause of the malfunction more quickly, and also secure his installations by not reacting to a faulty valve.
Solution
Using our DiagFit software, a model was built to differentiate between faulty and healthy cycles.
By learning about valve normality in different contexts and comparing these results to a large number of healthy and unhealthy cycles, the model was not only able to determine unhealthy cycles, but also to give the contribution of the sensors from which the anomaly originated.
Thanks to the visualization of the raw data and advanced data mining functionalities, the engineer assigned to the project was able to observe unexpected valve behavior, and was also able to improve their knowledge of the valve under certain conditions.
Results
By comparing the results of the automatically generated model with reference values (truth known to the user from his simulated tests), the model achieved the following metrics:
* 83% TNR (True Negative Rate): the model’s ability to detect healthy cycles.
* 92% TPR (True Positive Rate): the model’s ability to detect unhealthy cycles.
The compromise between these results can easily be improved by modifying the model threshold. This option allows the user to choose the best compromise for his application.
In this case, the user would like to lower the threshold to improve the detection of unsound cycles, even if this means increasing the number of “false positives”.
