Blog

 

Latest Stories

Featured Stories

Filter By Categories
Dion Bozec
By
June 15, 2021

Bad Data - Part 2

In our last blog, we looked at some examples of "bad data" and its sources. Now let's take a look at some of the ways that we can deal with bad data so it does not become an excessive burden on our analysis.

 

There are many ways one can deal with data issues but I am just going to look at some of the common ways we can manage these data issues.

 

1. Fix the Aircraft

 

This one is so obvious but it continues to frustrate me. If there is an issue on the aircraft that is causing it to collect erroneous data - FIX IT!

 

This one frustrates me because many of my colleagues will skip this step and go right to Item 2 below.  Those analysts have good intentions as they want to solve the customer's data issue through software, but the solution should be to fix the root cause of the problem - not mask it with software.

 

Consider you have a problem with a car where a bolt broke and caused a part to dangle from the car. You pull over to the side of the road and use a zip-tie or wire to secure the part to the car. The car has not been fixed; you just made a temporary fix so you could drive the car safely to a mechanic to get it properly repaired. Imagine the mechanic then just replaces the zip-tie with an industrial strength zip-tie. It needs a new bolt!

 

This is how I view using software to "mask" data issues that can be fixed on the aircraft. This is fine as a temporary fix if you need to review an incident (or worse), but if there is an issue somewhere on the aircraft, it needs to be properly fixed.Sign up for one of our eLearning courses.

 

2. Software Algorithms

 

As data analysis software becomes more advanced and Machine Learning (ML) and Artificial Intelligence (AI) algorithms become more available, it is easier to program those systems to catch and remove "bad data". 

 

Some data points are simply physically impossible, for example. Consider an altitude parameter that changes from 30000 ft to 35000 ft and back to 30000 ft in the course of a second. It is just not physically possible for the aircraft to manoeuvre this way (and continue to fly). The system can recognize this and flag the data as invalid.

 

These algorithms can be quite useful for Flight Data Monitoring and FOQA programs as they can reduce the analysts' workload when it comes to validating data.

 

Keep in mind, though, that for incident and accident investigation, it might be best to turn these algorithms off completely - or at least make sure you can see what the algorithms have flagged as "bad".

 

These algorithms work by comparing the data to what is normal. During an investigation, it is common to be dealing with very abnormal situations where even the best AI is no substitute for an experienced analyst or investigator.

 

3. Correct the FRED

 

This one is pretty straight forward but if you do have an issue with erroneous data, make sure that your FRED file is programmed according to the documentation for the aircraft. It is quick and easy to check.

 

Modern data frames can store hundreds or even thousands of parameters and these need to be entered by a human being so there is no way to avoid the potential for errors. There can also be errors in the documentation that is used to develop the FRED. In these cases, work with the manufacturer to get the documentation updated so that all operators can benefit.

 

I always recommend using some sort of versioning software for maintaining a history of a FRED file. We use the same system we use for our software repository but your system does not have to be complex. I have known operators to just rename the new file with the version number in the filename. Just pick a system that will work for your.

 

4. Human Validation

 

There are cases where even the best AI or ML cannot determine bad data from good data. In those cases, you will still need the expertise of an analyst to manually review the data and determine whether it is good or not.

 

A great example of this is a hard landing. These can happen so quickly that even the best software algorithms can think the data is just a bad data spike. But you risk missing a potentially serious event. In these cases, it might be best to have an experienced analyst validate this data the old fashioned, manual way.

 

5. Fix the Aircraft!

 

This one is so important (and so frustrating) that I wanted to include it twice!

 

These are just some ways that users can deal with bad data. Aircraft recording systems are getting better with each generation, but there will still be occasions where an analyst will have to be able to identify bad data and decide how to best deal with it.

 

This knowledge only comes with experience, but hopefully, these blogs will help you build up your own expertise. 

 

If you have any questions, though, please leave a comment below and we will be happy to try to help you out.

Subscribe Email

Sign up for one of our eLearning courses.

 

There are many ways one can deal with data issues but I am just going to look at some of the common ways we can manage these data issues.