Predicting future failures in turbofan engines with Machine Learning

La Paz. 2021

We apply Machine Learning (a type of recurrent neural network (RNN) called LSTM (Long short-term memory)) to predict motor failures and thus help reduce one of the biggest problems at an industrial level, which is corrective maintenance (due to its high costs of both repair and production). On the other hand, preventive maintenance is planned through an agenda to carry out a scheduled stoppage, where it does not matter if the equipment can still work or if its parts are still good; maintenance is carried out because it was scheduled that way.

That is why, as the years go by, technology allows us to invest in new solutions that allow us to know when a piece of equipment will fail, this is called predictive maintenance (PdM). Currently, through different sensors we can have a complete dataset of many variables, such as: vibration, temperature, oil or grease analysis, etc.

Fig. 1. Difference in production profit between traditional preventive maintenance and predictive maintenance, where only production is stopped based on conditions.

PROBLEM DESCRIPTION

The problem is the time and financial resources invested to be able to solve a corrective maintenance failure in the shortest time possible. At the same time, preventive maintenance could be optimized by doing it based on equipment conditions. Both take time in their maintenance and also in the production of the product.

DATASET

The datasets used are from the degradation of NASA turbo-fan engines generated by C-MAPSS, which is applicable to any other rotating machine, such as: industrial transport engines, hydraulic pumps, etc.

Fig. 2. Image of the turbofan simulated by C-MAPSS.

Table 1. Detail of the columns of the test and training datasets.

PROPOSED SOLUTION

For this reason, the project will help estimate the time in which turbofans will fail. Giving enough time to schedule a maintenance stop and search for the necessary spare parts at the best price.

MODEL SELECTION

To detect motor failures we will see the following Machine Learning models:

RNN

They are a type of recurrent neural network (RNN) and these have recurrent connections between hidden states, with a time delay. Ideally, RNNs are able to discover temporal correlations between sequential events that are very far from each other.

Fig. 3. Architecture of a Recurrent Neural Network (RNN)

Fig. 4. Architecture of a Long Short-Term Memory (LSTM)

LSTM

However, in practice, it is difficult to train RNN due to Vanishing Gradient and Exploding Gradient problems. So RNNs face difficulties in handling dependencies in long-term sequences. So, the most effective RNN-based model is called LSTM. Briefly, its structure consists of memory cells. Which its main function is to store a value and determine the time it should be stored. Additionally, these cells select which entries are stored and also decide whether they will be remembered, deleted, or output to a network.

Fig. 5. RNN programming with LSTM

The network has two layers of LSTM, the first with 100 units and the second with 50 units. At the output of each one there is a DropOut to avoid overfitting, and finally there is the Dense, which is activated by a sigmoid, which gives us the final probability.

PCA

Principal component analysis (PCA) is one of the most widely used unsupervised machine learning algorithms. For dimension reduction and data pre-processing.

Fig. 6. PCA applied to three dimensions to be able to graph it.

Fig. 7. Data from the complete dataset, with the 21 sensors, and the 3 user configurations.

Fig. 7. Data from the dataset with PCA applied. They are reduced to 8 main components.

METRICS

The analyzed metrics will be seen below. Where we can see that the accuracy is 95.11%, and that the F1-Score are very good, the one that has the least value is the macro. Furthermore, comparing the predicted data with the test data, we have an accuracy of 97%.

Fig. 8. Calculated metrics: Accuracy Score, F1-Score Macro, F1-Score Micro, F1-Score Weighted.

ANALYSIS OF RESULTS

The results we analyze are the confusion matrix, which shows us that its accuracy is 92.77%, and on the other hand the ROC Curve, where we can observe a good discrimination threshold of the model.

Fig. 9. Confusion Matrix

Fig. 10. ROC CURVE

RESULT

Therefore, we can predict the probability with which the turbo fan motor would fail in 30 days. Choosing the machine ID, we would have the following result; where we can see that the probability in which the Engine 16 fails, it is 1.8%, so we can continue using this engine in a calm manner.

Ideally, this analysis should be done periodically, when implemented. Because it gives maintenance supervisors a good survey of the state of the machines to plan the most critical things for the next stop.

Fig. 11. Final screen, where the program tells us the probability that the selected motor has of failing within 30 days.

CONCLUSION

Through this Machine Learning application we can determine the failure time in rotary motors, thanks to the data analysis of said equipment. Therefore, there is enough time to purchase spare parts and plan scheduled maintenance, mitigating the company's maintenance costs.

The accuracy of the implemented model is 95%, making it reliable when analyzing the machines. This program can be implemented in multiple industrial plants, and through the Internet Of Things, we can collect information from all the necessary sensors, which will be registered in the cloud for subsequent analysis.

Implementing this type of technology also has an impact on the environmental area, since fewer industrial spare parts would be changed or the use of oils or greases would be maximized according to their degradation; translates into less industrial waste for the environment.

CODE

https://github.com/albmarale/SaturdaysAIMachineLearning/blob/main/detecci-n-de-fallas-futuras-en-motores-de-turboven.ipynb

BIBLIOGRAPHY

[1] D. Bruneo and F. De Vita, “On the use of LSTM networks for predictive maintenance in smart industries,” in Proceedings — 2019 IEEE International Conference on Smart Computing, SMARTCOMP 2019, 2019, pp. 241–248, doi: 10.1109/SMARTCOMP.2019.00059.

[2] S. Guldamlasioglu, O. Aydin, and D. Scientist, “Using LSTM networks to predict engine condition on large scale data processing framework,” 2017, doi: 10.1109/ICEEE2.2017.7935834.

[3] L. Swanson, “Linking maintenance strategies to performance,” Int. J. Prod. Econ., vol. 70, no. 3, pp. 237–244, Apr. 2001, doi: 10.1016/S0925–5273(00)00067–0

[4] A. Martínez, “Recurrent Neural Networks with LSTM applied to Predictive Maintenance, Case: Degradation of turbofan engines”, 2020, Universidad Católica Boliviana “San Pablo”.

• Machine Learning
• Predictive Maintenance
• Predictive Maintenance
• Engine
• Bolivia

Repository

In the following repository you will find the code used to develop this application:https://github.com/SaturdaysAI/Projects/tree/master/Lapaz/detecci-n-de-fallas-futuras-en-motores-de-turboven-main

More artificial intelligence!

The mission of Saturdays.ai is to make artificial intelligence more accessible (#ai4all) through intensive courses and programs where projects for good are carried out (#ai4good).

Explore our AI master’s programme at https://saturdays.ai/master-ia-online/

If you want to learn more about artificial intelligence join our community at community.saturdays.ai or visit us on our website www.saturdays.ai Join us!