Predictive Modelling of assets

[Sandy1]

Our company collects the condition of many asset types e.g. stormwater assets. I would like to train the model to determine the future condition of assets based on data such as material, diameter etc. The model should use the condition based on similar pipes such as 150mm uPVC. For simplicity the condition would be 1 - Excellent to 5 requires replacement and plotted against percentage effective life (0 to 100). The predictive model could be set for condition in 1 year, 5 years, 10 years and vary with the condition of all similar pipes in this case to provide a state of the assets report based on time for the network. So an update all function would be required to update all the predicted condition for all similar pipes in this case

The same process could be used for other asset types e.g. building assets, watermains, sewermains, park assets, roads, etc.

[Aleksi]

@tony any thoughts?

[Douglas_Thom]

I will rephrase to test my understanding. Let me know if I misunderstand the challenge.

In a predictive model, one picks some variable to predict, and inputs that tend to correlate to the prediction.

It sounds like you are predicting:

1. Condition (1 (Excellent) to 5 (Needs Replacement)).

It sounds like your inputs are:

1. Age category (1, 5, 10 years)
2. Material category (Stormwater, Building, Watermain, Park, Road)
3. Material subcategory (150mm uPVC and similar, other categories, etc.)

My guess is then you would be able to query which assets are in a condition more advanced than predicted to prioritize early failure investigations.

Does that sound about right?

[Sandy1]

Nearly right.

Yes predicting condition and relating it to % effective life. As an example. Condition 1 = 100%, 2=75%, 3=50%, 4=25%, 5=10% (user defined) Therefore Condition 4 = 25% of Asset Life remaining.

Age or asset life actually is related to material which is a factor of the asset class. At the highest level Stormwater is the asset class, Pipes = Asset Type (correct).

Subcategory input = pipe material = uPVC, Pipe Diameter = uPVC, Asset Life of uPVC pipes = 80 years.

The future prediction of 1, 5, 10 or even 20 years will then provide future prediction of pipe material remaining life at pipe level. E.g. Condition 4 = 25% effective life so remaining life = 20 years for uPVC pipe. From there the future type of treatment for the pipe can be selected. As an example this will then determine that all uPVC pipes in Condition 4 will require say a pipe relining in 10 years time, Condition 4.5 will require replacement in 5 years time etc… This then forms the basis of the capital works program.

Hope this helps. You were close.

[tony1]

Hi @Sandy. Your problem sounds very interesting and promising. You should be able to use an AppSheet predictive model to make this happen.

I’d first recommend reading through this article to make sure you’ve got the basic idea of what AppSheet can do with predictive models: https://help.appsheet.com/intelligence/predictive-models

Next, I’d examine your training data. I’d make your Condition column an Enum type with values Excellent to Needs Replacement as the values it takes. Then, I’d make sure that your input columns (Age, Material Category, Subcategory) are all Enum columns as well. Then, I’d define a predictive model with Condition as your output column and (Age, Category, Subcategory) as your input columns.

That would be a good starting point. Does this help, or am I misunderstanding something?

[Sandy1]

Tony, I think you might be spot on there. You haven’t misunderstood anything. Thanks

[tony1]

Please feel free to ask for help or advice or give feedback along the way. This is a beta feature and we want to learn about your use cases.

[Mahesh]

@Sandy, very interested in your experience. So would greatly appreciate if you could post your experience and lesson learnt. This can have some really useful uses.

@tony, It is awesome you have this functionality.
Reading the page you referred to, sounds like the functionality right now relies on columns within one table as input.

The little I understand from my reading on this subject, I hear creating these model is iterative. In this case, one would try with the most likely set of columns, add/remove some columns and retry and keep repeating until one get a reasonably good predictive result? Is that how one would proceed?

[tony1]

@Mahesh Yes, you can create as many predictive models as you like and compare the performance across models. You might add another column to your dataset and see how that changes the results.

[tony1]

@Mahesh If you run into problems or have feedback, please let me know.

[Sandy1]

Would love to send some info on me but can’t upload PDF file. Do you have an email I can send it to. You can probably google my name and asset management to get some idea of my background.