KONUX is one of 30 of the World Economic Forum’s most innovative start-ups and scale-ups. Its sustainable solution offers predictive maintenance, network usage, traffic monitoring and planning solutions for railway infrastructure management.
Their innovative team applies data science and machine learning to untapped and disconnected data, to make sure infrastructure managers can prevent failures, optimize their maintenance planning and – ultimately, help railways be more on time. Something we can all appreciate.
One of their solutions, KONUX Switch, is an end-to-end predictive maintenance system for rail switches. It uses IIoT devices to measure the vibration forces from passing trains. Combined with temperature measurements and other operational data , the data is fed into complex AI models – and these AI models must provide trustworthy predictive maintenance insights for infrastructure managers.
KONUX ensures a prediction accuracy of at least 90% or above to their customers. A feat not easy to accomplish given the complexity of AI models which must operate on imperfect labeling and a high amount of sensitive and noise- prone data.
Yet, KONUX’s Data Scientist team does just that – delivers predictive AI models to their customers with at least 90% accuracy! The challenge is, this task is immensely complex and takes a vast amount of resources and effort.
“There are always more things you could do, but for which you do not find sufficient time.” Andrés Hernandez, Principal Data Scientist
Since the data science team has finite resources at their disposal, prioritizing what to test, evaluate, and retrain is of the utmost importance to business outcomes. Predictions cannot be false negatives, as this would pose great reputational and financial risks. Meaning there is a trade-off on speed that must be taken to ensure quality.
Given the growing demand for KONUX’s solution, the KONUX team connected with QuantPi to see if there are ways to streamline and accelerate AI testing to ensure the quality of even more predictions in a scalable way.
Through PiCrystal, QuantPi’s automated assessment engine, KONUX’s AI models were automatically analyzed on four risk likelihood dimensions —
Performance: Which kind of classification mistakes occur most often?
Robustness: When input data is perturbed, what is the decrease in model performance?
Explainability: How good is an explainability technique in capturing the parts of an input which mostly influenced predictions?
Bias: Are there locations and train types where the performance is particularly low?
The assessments offered insights which the team truly appreciated. KONUX could see that with QuantPi, it would be possible to accelerate the validation process of their AI models.
“Having an overview of which cases we have in the respective risk dimensions helps with retraining.” - Andrés Hernandez, Principal Data Scientist
Within the last eight months alone, KONUX’s business has increased immensely. The amount of locations and sensors in use — and hence, the data processed by its AI models — has more than doubled. Scaling resources to meet the demands of their growing business are crucial. Through automated assessments like PiCrystal, the workload is possible.
“Some of our models have huge reputational and financial risks when our predictions are false negatives. The Explainability aspect (of QuantPi) in particular helps us to safely accelerate the confidence in our results.” Andrés Hernandez, Principal Data Scientist
KONUX will continue to explore how to implement these learnings and future insights from QuantPi.
QuantPi is a CISPA-Helmholtz Center spin-off founded in 2020. It enables organisations to validate AI models at an unprecedented speed. Through automated risk and performance assessments, the black boxes of AI are made transparent, explainable and compliant with any regulations or guidelines — for all stakeholders.
For more information, email email@example.com