A simplistic LSTM model for predicting if failure occurs for Seagate Exos X14 ST12000NM0008 Hard Drives within the next two weeks using SMART data aggregated by the drive from the past two weeks.
For a detailed walkthrough of problem formulation, model development and evaluation please refer to the .ipynb file in this repository.
Credit goes to the Backblaze Hard Drive Dataset, which was used to train and evaluate the LSTM in this repository.
Included in this repository are the model weights and the scaler used for the input data. The model expects two weeks worth of data input, in the form of a tensor of shape (1, 14, 5)
It is recommended to automate SMART data readings from your drive(s) daily or more frequently, scale them using the scaler provided, remove your oldest readings in "sliding window" fashion and input into the model. The model raw output will be a value between 0-1.0, indicating likelihood of drive failure within the next two weeks. You may wish to set a threshold to determine whether this drive should be discarded. Furthermore, a per-drive cooldown should be set so that marked drives do not alert failure several days in a row.
I recently fried my own drives (pictured below) by plugging in the wrong SATA power cables into them. A little bit of smoke came out and data recovery is looking to set me back a few hundred :(
