One of the biggest gondola cableway companies in middle east had the problem of perfectly managing the timeline of it’s huge gondola cableway physical elements to ensure best safety and reliability for it’s visitors. They had a traditional software written in some Jurassic technology which just had the option to define an MTBF for an element and to report close to dead devices based on the time elapsed. The main pitfall of this technique was that they had to hard guess about MTBFs. While in some conditions the devices were provided by an official MTBF, in some occasions the technical team had to change the brand of the device, leading to the necessity of updating the MTBF of the specific element type of the line.
Regardless of the whole messy situation, there were numerous occasions where the device was considered alive on the paper while because of specific conditions like operating in very cold season of the year, that very device should have been changed sooner comparing to its ancestors which were operating in normal seasons before.
The best fit for this problem in order to cut the prices, enhance reliability and safety and also being flexible during different conditions was to use a machine learning approach. Luckily there were tons of data in the accounting system and also the legacy PM software which could be analyzed to deduce the exact nature of what happened, what prices paid, time lines and so on. We analyzed 15 years of past data and processed it to have a clean usable and feature engineered dataset to work on.
We used our own data preprocessing toolset in addition to Google Cloud DataPrep and also Amazon’s AWS Glue as an ETL (Extract, Transform, Load) service to better understand the nature of the data we had, check correlations, finalize the feature engineering and be sure that our dataset is comprehensive enough to be generalized easily.
We have a specific attack plan for machine learning problems, we used that plan and after doing the research on different models alternatives we had the right metrics at hand that random forest and SVM as an ensemble would give us the best results.
The system after only 30 days detected 5400 items out of around 300,000 devices which should have been changed but in the old PM system there were ok and not-surprisingly %40 of them had really horrendous conditions (consider a cabin lock has forgot how to do it’s sole purpose and you lean on to it thinking it’s ok).
Aside from the overall safety increase we could enhance the cost structure and prevent sooner than normal procurements cutting the total costs by %30 each month.
June 20, 2018
AI, Amazon AWS, Cloud, Google GCP, Machine Learning