Yet Another Predictive Maintenance Article -2
In part-1, I discussed about basics of Predictive Maintenance (pdM) and overview of the procedure for building pdM systems. This article contains the steps for implementing the pdM system along with the results.
Before we directly make jump towards the implementation, we will look at what kind of data we are working on.
DATASET
Predictive Maintenance requires the machine with sensors which are capable of gathering the data at a fixed time interval. Time-series or sequential dataset can be used for predicting the life condition of the components. I have implemented pdM using following dataset:
NASA - Turbo Fan Degrade Engine Dataset
Turbo Fan Engine dataset by NASA — This dataset consists of engine’s data at different operational conditions along with different sensors readings at each operating cycle of the engine.
- Three operations settings: Altitude, Mach Number and Throttle Resolver.
- 21 sensors for examining the condition of the engine.
Train Data consists aircraft engine’s run-to-failure data.
Testing Data consists of aircraft engine’s operating data without failure events recorded.
- DATA CLEANING AND PRE-PROCESSING
We will first make sure our data must not contain any missing or NULL values. There are several techniques in data science to handle missing data. Turbo Fan Engine dataset doesn’t contain any missing value and hence we will simply skip this step.
After data cleaning, we will apply Feature Engineering or Data pre-processing before feeding to our model. Three basics pre-processing techniques are applied before prediction.
- Zero Variance features -Any feature or column containing same value for every timestamp or every sample, we can simply drop it.
- Strongly related features -If the correlation between two columns is very high (greater than certain threshold), we can drop any one of them.
- Normalization -For non-uniform values for different features, we will scale the dataset between 0 and 1 using min-max normalization for better performance of ML model.
2. PROBLEM DEFINITION (Revisited)
I have taken into consideration the two problems I have discussed in the previous blog. Technically we have two definitions:
Regression -predicting the exact value of the breaking point.
Classification — classifies whether the engine is going to fail or not.
3. MACHINE LEARNING MODEL
Regression: I have applied several supervised ML models for predicting the Remaining Useful Lifetime (RUL) of the engine.
The results are shown in the above table. Random Forest gave the best fitting of the model for test data.
Classification: For classification, we bifurcated the train set into three categories. Two-layer LSTM network with sequence length of 50 was used and performed well for the test set.
The results for the Bi-directional LSTM are as follows:
The complete implementation was done in python 3.6 and jupyter notebook.
CONCLUSION
In this article, we looked how to approach to predictive maintenance problem and steps for making a predictive model. Several other techniques can be used for making pdM systems. The same methodology can also be applied for other time-series dataset and can be expanded to larger scale. This article gives a simplified methodology for actual implementation of predictive models.
FUTURE WORK
- Automate Maintenance: Predictive maintenance is an essential part of the factory of the future. Manufacturers who automate not only manufacturing processes, but also equipment maintenance, can benefit from a whole new level of production efficiency.
- Extending to different domains: Once this type of dataset having different sensor information of a machine is available, we can use these models for different machines.
- Dynamic Prediction: Predicting about the breaking point of the machine as and when the data from sensors is recorded.
REFERENCES
Dataset : https://archive.ics.uci.edu/ml/datasets/diabetes
https://towardsdatascience.com/remaining-life-estimation-with-keras-2334514f9c61
