# What kind of model (consists of sequential and non sequential data) to predict failure in devices?

I have a data set of events for each device(sensor)

2019-08-24 Event_1

2019-08-25 Event_2

2019-08-26 Event_3

2019-08-27 Fail

2019-08-28 Event_2

2019-08-29 Event_3


I am recording sequential events of a device. Also I have other factors such as temperature, humidity, data related to manufacturing of each device and how user data (how users are using it). I want to build a model that can take all these thing into account and predict when the next failure will happen. I don't know how to combine this sequential events with other factors. What kind of approach or modelling should I use ?

• Take a look at the questions at the survival-analysis tag. Also, please edit your title to be more specific. Your present one could be used for more than half the posts on this site. – mkt - Reinstate Monica Sep 5 at 17:47

## 3 Answers

Predicting mean time to failure falls under the general heading of Survival Analysis. So first you should read up on that topic.

You should also look at Markov Chains and State Space Models. Markov models are good at representing a system with discrete states and probabilities of transitioning from one state to the other. State space models are a generalization of Markov models, but where the states are continuous instead of discrete.

I know that Markov Models have been used for survival analysis before.

RNNs and LSTMs would definitely be a possibility for this, but they might be overkill compared to Markov Chains.

Microsoft has some great resources on prepping data to predict time to failure as well as predicting failure within a time period 'x' from the last operation.

https://github.com/Azure/PySpark-Predictive-Maintenance

https://gallery.azure.ai/Notebook/Predictive-Maintenance-Implementation-Guide-R-Notebook-2

https://docs.microsoft.com/en-us/azure/machine-learning/team-data-science-process/cortana-analytics-playbook-predictive-maintenance

You should look at RNNs and LSTMs, which are adapted for sequential events because of their recurrent nature

• Thanks. Can I justify it's results to the business team. As, I need to put this model into production so I need it to be reliable and explainable when they ask me why a certain device will fail. – No_Body Sep 5 at 17:23