In the Manufacturing sector machines represent a fundamental asset, and a “run to failure” approach often results in the purchase of new machinery, prohibitive repair costs and downtime that can slow down the production chain. The application of the Artificial Intelligence allows to anticipate the failure safeguarding the line and extending the life of the machine.
Today we present a brief case study implementing a predictive maintenance model of a Blue BI customer.
Case study of preventive maintenance in manufacturing industry
Customer and context
The customer is a multinational company operating in the Fashion & Luxury sector that has more than a thousand employees and a turnover of about 1 billion euros. The business model is both B2C and B2B thanks to a network of more than a thousand stores and boutiques.
Business areas involved:
- Supply Chain
As a result of the cost analysis, Finance was concerned that the maintenance expense item would increase by 10% and this was complemented by the investment that the company would have to make to replace some machinery.
Why Blue BI
The objectives of the Blue BI team consultancy were identified in:
- estimate when an equipment component may fail;
- planning maintenance at the most convenient and cost-efficient time;
- optimize the life of the equipment thanks to the provision of corrective maintenance that extend the life of the machine.
Activities carried out
What are the most common factors behind failures? Which machines are more likely to fail?
To answer these questions following a data driven approach, first it was necessary to verify and evaluate the data available to the company that, in this case, referred to past failures, as well as the use and maintenance of machinery.
In particular, we focused on three data sets:
- Usage: hours worked;
- Maintenance: records of when and which parts of the machines have been maintained, the reason for the service and the amount of parts replaced;
- Fault: if a machine has had a registered fault (not all cases are labeled).
Once the necessary data was identified, these were imported into an advanced AI platform that made the whole process more efficient; in fact, it was much easier to process them and achieve the same level of granularity in order to be able to combine them and obtain more information.
Next, before training predictive models we decided to create two separate data sets:
- Training dataset to indicate whether or not a failure event occurred, which was used to tow the models;
- Scoring dataset that contained no fault data, and was used to predict whether machinery had a high probability of failure.
Once we had all the datasets ready, we applied the algorithms considered better than the target variable and then compared them to evaluate the performance of one over the other. In this case, the Random Forest model obtained better results than several metrics taken into account and based on three possible events: probability of failure, probability of non-failure, failure prediction model or not.
Thanks to the probabilities generated by the model, one step that we thought it useful to implement was to identify two levels of failure risk (High and Medium) that allowed us to refine the dataset even more.
The last step was to put the predictive model into production by creating a scenario that would automatically rebuild the flow with new data to get new assets every day.
At the same time, we implemented a new maintenance model dictated by the algorithm by verifying the results compared to the old management, which the customer could easily monitor through dashboards.
By integrating the historical data with those of real-time monitoring, we were able to answer the customer’s questions and provide a monitoring tool able to predict the moment and risk of the possibility of the occurrence of the fault, reduce costs by scheduling the most convenient time to perform the interventions, lengthen the life of the machinery.
We could say that in numerical terms we have obtained:
- approximately 27% of reduction of the costs of maintenance
- the cases of malfunction came down of 60%
- the productive line has reduced the downtime of 30%.