Predictive Maintenance

DePaul University, Chicago, Illinois, USA

Group #11: Adithya Harsha, Joao Vitor Lira de Carvalho Firmino, Beemnet Desta,
Tejas
About our Dataset
● The dataset we selected is the Predictive Maintenance Dataset, sourced from Kaggle, which contains
10,000 data points and 14 features.
● These features capture various operational conditions, such as air temperature, process temperature,
rotational speed, torque, and tool wear, along with a machine failure label.
● Machine failures are categorized into five types: tool wear failure, heat dissipation failure, power failure,
overstrain failure, and random failures.

Goal:
The goal of this project is to predict machine failures using operational data. By doing so, we hope to improve maintenance
schedules, reduce downtime, and avoid premature replacements.

Use:
This prediction model will assist industries in implementing predictive maintenance strategies, ensuring that machines are serviced
at the most appropriate time. This reduces operational expenses, improves machine reliability, and avoids unexpected
breakdowns. Ultimately, it will result in increased production and cost savings for companies that rely on heavy gear.
Pre processing and Cleaning

● There were no null/missing values.

● Checked for empty strings in the 'Product ID' and 'Type' columns, and found none.
● Detected outliers in 'Rotational Speed [rpm]' (418 outliers) and 'Torque [Nm]' (69 outliers) using the IQR
method.
● Applied log transformation to 'Rotational Speed [rpm]' and 'Torque [Nm]' to reduce the impact of extreme
values and normalize the distribution.
● Converted the categorical 'Type' feature into numerical binary features using One-Hot Encoding.
● Standardized numerical features including 'Air temperature [K]', 'Process temperature [K]', 'Rotational Speed
[rpm]', 'Torque [Nm]', and 'Tool wear [min]' using StandardScaler to ensure equal contribution during
modeling.
● Altogether, we now have transformed and scaled features ready for machine learning models.
Outlier Detection & Handling Feature Scaling

● Outlier Detection: The features in the dataset have different ranges,

○ We applied the Interquartile Range (IQR) which can affect the performance of machine
method to identify outliers in the numerical learning models. To ensure all features contribute
columns: equally, we applied standardization. After
○ Rotational Speed [rpm]: 418 outliers detected. considering several options, we decided on using
○ Torque [Nm]: 69 outliers detected. StandardScaler, which standardizes features by
○ No outliers found in 'Air temperature [K]', giving them: Mean = 0 and Standard Deviation = 1
'Process temperature [K]', or 'Tool wear [min]'.
● Handling Outliers: We made sure that scaling was applied only to the
○ To reduce the impact of extreme values, we numeric features and not to the categorical ones.
applied log transformation to the 'Rotational
Speed [rpm]' and 'Torque [Nm]' columns: ● Features Scaled:
○ Rotational Speed [rpm]: After transformation, ○ 'Air temperature [K]'
the data is concentrated around a tighter ○ 'Process temperature [K]'
range, with fewer extreme outliers. ○ 'Rotational speed [rpm]'
○ Torque [Nm]: The transformed data is more ○ 'Torque [Nm]'
compressed, with smaller impacts from outliers ○ 'Tool wear [min]'
on the lower end.
Exploratory Data Analysis(EDA)

Pie chart illustrates that, within this dataset, 60%

of the manufactured products are of low quality,
30% are of medium quality, and 10% are of high
quality.

Machine failure based on product Quality (Bar Chart): In this graph,

it is possible to see that almost no high-quality products are related to
machine failure. On the other hand, almost all failures occur during the
manufacturing of low-quality products.
Correlation Matrix:

○ Air Temperature [K] & Process Temperature [K] (0.88

correlation): These variables are linked to the machine's
thermal dynamics. Higher air temperatures make cooling
harder, leading to increased process temperatures.

○ Rotational Speed [rpm] & Torque [Nm] (-0.94

correlation): Typically, as rotational speed increases,
torque decreases, indicating an inverse relationship crucial
for maintaining consistent power output.

○ Different Failure Types (HDF, PWF, OSF): Failure modes

like HDF, PWF, and OSF are interconnected. For instance,
poor heat dissipation (HDF) can cause electrical issues,
leading to power failures (PWF)
Histograms of features:
○ Air Temperature Histogram: The distribution is bimodal, with concentrations around 298 K and 301 K.

○ Process Temperature Histogram: Values range from 306 K to 314 K, with most clustered between 309 K and 312
K. A prominent peak between 310 K and 311 K suggests most temperatures are centered here, though higher
temperatures occur less often.

○ Rotational Speed Histogram: Right-skewed with a peak around 1500 rpm, indicating most data points cluster at
lower speeds, with fewer high-speed instances.

○ Torque Histogram: Displays a normal distribution, with a mean torque of approximately 40 Nm.

○ Tool Wear Histogram: Tool wear ranges from 0 to 200 minutes, with a nearly uniform distribution and a sharp
decline after 200 minutes.
Feature Engineering

● New Features Created:

○ Mechanical_Power:
■ Captures the machine’s output using the formula: Mechanical Power=Rotational
Speed×Torque\text{Mechanical Power} = \text{Rotational Speed} \times \
text{Torque}Mechanical Power=Rotational Speed×Torque.
■ Key Insight: Strong inverse correlation between Rotational Speed and Torque (-0.94).
○ Temp_Diff:
■ Highlights potential overheating issues with the formula: Temp_Diff=Process Temperature−Air
Temperature\text{Temp\_Diff} = \text{Process Temperature} - \text{Air
Temperature}Temp_Diff=Process Temperature−Air Temperature.
■ Key Insight: Strong positive correlation between Air and Process Temperatures (0.88).
● Log Transformation Applied After feature creation to avoid distorting Mechanical Power.
○ This maintains the accuracy of features like Mechanical Power, which could become negative or zero,
distorting the data before transformation.
Model Preparation and Training
Scaling Removed for Tree-Based Models:

● Random Forest, Gradient Boosting, and XGBoost don't require scaling, simplifying preprocessing.
● Tree-based models split features based on order, so the magnitude of values (e.g., speed or temperature)
doesn't affect their performance.

Column Renaming:

● Renamed columns to remove special characters like '[', ']', and spaces. This was necessary for XGBoost to
avoid errors during model training, as XGBoost requires column names without special characters.

Train-Test Split:

● Used a train-test split of 80% training data and 20% test data with a random_state=42 for reproducibility.
Scaling Selected Features

● The goal was to Standardize features to improve model performance and convergence.
● Columns Scaled: Air_temperature_K, Process_temperature_K, Rotational_speed_rpm, Torque_Nm,
Tool_wear_min.
● We used StandardScaler to apply scaling.
● This ensures features are on a similar scale, reducing bias toward larger values.
● Post-scaling, mean values are close to 0, and standard deviations are close to 1, indicating successful
normalization.
Train/Test Split
● The purpose is to separate the data for model evaluation and to avoid overfitting.
● Target variable (Any_Failure) indicates any type of machine failure.
● Feature selection:
○ Dropped columns unrelated to model training (Machine_failure and target column).
○ Ensures only predictive features are included in X.

● Outcome:
● The split ensures the model is trained on a larger portion of data while retaining enough for reliable testing.
Model Selection
● Tree-based models were selected due to their ability to handle complex interactions between features.:
○ Random Forest;
○ Gradient Boosting;
○ XGBoost.

● Additionally, other models were applied to evaluate performance across different types of algorithms:
○ AdaBoost;
○ Logistic Regression.
Model Training, Prediction, and Evaluation

● We trained the models on the SMOTE-balanced data using only the selected features.

● This step ensured that the training data was balanced and tailored to the important features, enhancing the model’s
ability to accurately detect instances of the minority class.

● After training, predictions were made on the test set, and the model’s performance was evaluated using metrics such
as accuracy, precision, recall, and F1-score.
Model Training, Prediction, and Evaluation

Random Forest:
● Final Accuracy: 0.965
● Precision: 0.47
● Recall: 0.68
● F1: 0.56
Model Training, Prediction, and Evaluation

Gradient Boosting Model:

● Final Accuracy: 0.938
● Precision: 0.32
● Recall: 0.78
● F1: 0.45
Model Training, Prediction, and Evaluation

XGBoost Model:
● Final Accuracy: 0.978
● Precision: 0.64
● Recall: 0.72
● F1:0.68
Model Training, Prediction, and Evaluation

Threshold Tuning on XGBoost Model:It was employed to balance recall and precision further, enhancing
sensitivity to true failure cases.

● Final Accuracy: 0.967

● Precision: 0.49
● Recall: 0.74
● F1:0.59
Model Training, Prediction, and Evaluation

AdaBoost Model:
● Final Accuracy: 0.904
● Precision: 0.22
● Recall: 0.75
● F1:0.34
Model Training, Prediction, and Evaluation

Logistic Regression Model:

● Final Accuracy: 0.705
● Precision: 0.07
● Recall: 0.63
● F1:0.12
Conclusion

In this predictive maintenance project, multiple models were evaluated to predict failures. In
conclusion, several key points can be highlighted:
● Top Performing Model:
○ XGBoost achieved the highest performance with an accuracy of 0.978.
○ Threshold tuning further improved recall to 0.74, optimizing sensitivity to true failures.

● Model Comparisons:
○ Random Forest also performed well but had slightly lower precision and recall.
○ Gradient Boosting and AdaBoost provided acceptable recall but lacked precision balance.
○ Logistic Regression showed the lowest metrics and was not suited for this task.
Conclusion

● Main Insights:
○ XGBoost, especially with threshold tuning, is the most effective model for this dataset.
○ A balanced approach between recall and precision was essential for reliable failure prediction.

● Practical Implications:
○ The model’s reliability can enhance maintenance scheduling and reduce unexpected downtimes..
Thank You