Name:
Aaron Crouch
Title:
Data Analytics Manager
Country:
United States
Organization:
Atlantic Plant Maintenance
Description:
We specialize in the repair and maintenance of Power Plant equipment. We mostly work with GE Coal, Steam, Nuclear and Gas turbines and boilers. It is difficult and dangerous work performed by skilled union labor, often in the elements. Very few laborers work directly for APM on a permanent basis, and are hired out of union halls as needed. Since our work requires taking generators offline, most of our work is done in Spring and Fall when power demand is lowest. Our labor pool has many opportunities outside of APM, especially in the Spring and Fall outage seasons, so it is imperative that our union employees WANT to work for APM. Part of building that loyalty is getting our labor home safely.
Awards Categories:
- Organizational Transformation
- Data Science for Good
Challenge:
The main issue which drove us to Dataiku is the moral imperative to send our workers home safe. Beyond that, there is a financial and business impact for safety too. If our workers go home hurt, they won’t want to come back and work with us again.
Also, the industry, insurance companies, and government regulators track our Injury and Illness (or I&I) rate, which is a formula that takes into account the number of our most serious injures, times 200,000, divided by the number of hours we work. This represents the number of full time employees, per one hundred, that will experience a recordable injury in a year. Many plants require contractors to have an Illness and Injury rate below a certain threshold to work on site. Our I&I rate was close to getting us blocked from certain customers.
So we wanted to combat safety issues, but injuries seemed so random that safety professionals and leadership could only react instead of being proactive. We needed to flag problem jobs, send professionals to problem sites, and train problem employees. It was decided that the data we have could help flag these problem jobs and possibly prevent an incident before it happened.
Solution:
We used Dataiku to combine all the inputs we had on jobs: defects by job site history, superintendent defect history, employee ratings, lines of business, job duration, headcount, turbine type, work scope… Dataiku was able to combine all of these variables and compare them to the past jobs that had safety or quality incidents and calculate the likelihood of a defect on upcoming jobs.
Dataiku’s feature allowing us to see which variables are causing an individual job to be flagged as high risk has allowed management to reduce risk, through being able to identify the most impactful variables (where possible) and make changes in the front end accordingly, leading to fewer high risk jobs to begin with.
The ability of Dataiku to flag which metrics are most important to the model as a whole has allowed our field personnel to see how the model works, and even suggest other metrics that the data team did not consider, increasing field buy in.
We put in place mitigation strategies, such as required twice weekly hazard hunts and leadership site audits to reduce the likelihood of an injury or quality defect. Dataiku has also enabled us to measure how effective our mitigation strategies are after a job has occurred, by looking at the percentage of high risk jobs that used our mitigation strategies and comparing whether those jobs had a safety or quality defect.
Impact:
The results are dramatic. In 2018, before we launched the high risk jobs program, close to 26% of our jobs had some sort of safety or quality defect. That number has declined steadily to less than 11% in YTD 2021.
In 2018, over 86% of the jobs that would have been flagged high risk had defects. In 2021 YTD, only 68% of high risk jobs had a defect.
Using the statistical tools in Dataiku, we were able to see that high risk jobs where a leadership site audit was performed had a 77% chance of having a defect, while high risk jobs that did not have a leadership site audit performed had an 83% chance of having a defect, with a statistically significant p-value of .046.
We obviously can’t point to an injury, or quality event that did not occur, but we can assume from the data that workers who otherwise would have been hurt made it home safely. This keeps our employees safe, our customers happy, and reduces insurance rates.
