Using Event Data to Build Predictive Engine Failure Models

Question

1. What is event data in the context of this study?

2. How was bias removed from engine failure data?

3. How are positive examples extracted from ESBT events?

4. How are negative examples created in the study?

Accepted Answer

Event data in this study refers to a collection of data items containing at least a time-stamp and failure/event code. It is typically collected to record a change at a point in time and can be gathered from multiple sources, sensors, or devices. The on-board diagnostic system in the DMU trains monitors various vehicle systems and logs events when specific criteria are met, such as thresholds or detected changes. The event data used in this study is sparse, meaning events are not logged continuously for each sensor, adding complexity to the analysis. The dataset contains 14,483,278 records and 379 features related to the operating functions of a DMU passenger train, with a focus on engine failure prediction.

Accepted Answer

To remove bias from engine failure data, genuine ESBT events were identified and false ESBT events were removed. False ESBT events were classified as those that occurred at the servicing depot, when the train was not in motion, or when multiple ESBT events were recorded successively with little time in between. A geofencing polygon was constructed around the servicing depot to identify false ESBT events. The haversine distance was used to determine if the train was in motion during an ESBT event. Additionally, values outside the sensor range were removed and replaced with 'null'. This preprocessing ensured the data were not biased prior to training, which could result in a high false positive incidence.

Accepted Answer

Positive examples are extracted by identifying genuine ESBT events after removing known false positives. For each unique DMU engine, the first ESBT event is found and labeled T 0h. Data 3 h prior (T -3h) is filtered and divided into 15 min intervals. If no other ESBT event is found between T -3h and T 0h, the mean and standard deviation of the features in each interval are calculated. If a second ESBT event is found within the 3 h window, only intervals between the first and second event are used. These instances with mean and standard deviation values are labeled as positive examples, representing an ESBT event.

Accepted Answer

Negative examples are created by selecting a random point in the timeline of data, checking for ESBT events within a 6-hour window. If no ESBT event is found, data between -3 hours and 0 hours is divided into 15-minute intervals, and mean and standard deviation values are calculated for each interval. An arbitrary value of 3 iterations is chosen for each unique unit before moving to the next. These instances are labelled as negative examples, representing no occurrence of an ESBT event.

Accepted Answer

To balance the dataset in feature reduction, a random sample of negative examples is taken to match the number of positive examples. This creates an equal number of positive and negative examples for training and testing classification models. After concatenating binary examples, a low variance filter removes features with constant values. A linear correlation filter eliminates highly correlated features (Person's correlation coefficient >= 0.8). Features with more than 50% missing entries are also removed. This methodology ensures an unbiased dataset for machine learning model training and generalization.

Accepted Answer

C4.5 and random forest algorithms both handle missing values robustly. C4.5 uses a decision tree model that provides explainability for the random forest method. It selects attributes based on their effectiveness in splitting data into respective classes, using the Gini index as the quality measure. Random forest, an ensemble method, combines multiple decision trees and predicts output based on the mode of classes from each tree. It uses a subset of the training set, known as the local set, to grow each tree, with the remaining samples used to estimate goodness of fit. Both algorithms were tested for robustness with 500 iterations and a 70:30 training-testing split.

Accepted Answer

Explainable Artificial Intelligence (XAI) methods are techniques and approaches that aim to provide explanations or justifications for the decisions made by AI models or systems. They enhance transparency, interpretability, and trustworthiness in AI systems, especially in complex and black-box models. XAI methods can be categorized based on functionality, such as model-specific and model-agnostic methods. They are popular for 'black box' machine learning algorithms like neural networks or random forests. In this study, three model-agnostic XAI methods - Skater, Sage, and Shap - are used to gain insight into random forest models. These methods are perturbation-based techniques that simulate the absence of a feature and estimate its contribution to the model's predictions. Skater provides global explanations using cross entropy/F1 score, while Sage and Shap compute mean average from multiple feature coalitions using Shapley values. Local explanations offer insights into individual instances, while global explanations provide a comprehensive understanding of the model's behavior. Global explanations are desirable for understanding the overall decision-making process of the model and the influence of selected features on predictions.

Accepted Answer

The optimal data block window size for predicting engine failures in diesel multiple units is 5 hours (-5 hours to 0 hours). This window size produced the most accurate models, with accuracies improving as the window size increased. Smaller window sizes produced less accurate models due to limited data availability. The initial data curation involved manual selection of relevant features, and the best performing models had 17 features available during training. Features related to coolant level, oil pressure, and coolant temperature were frequently used across iterations. XAI methods showed that the random forest models heavily relied on these features. The authors suggest using a larger window size in future work to process data as far back as possible from the ESBT event, potentially providing a longer warning period for engine failures.

Accepted Answer

Predictive maintenance plays a crucial role in diesel multiple unit engine failures by detecting potential issues before they manifest into failures. This approach helps minimize delays and service cancellations, while avoiding major financial costs. By using remote condition monitoring as a means of preventative maintenance, the rail industry can collect and analyze larger volumes of data about asset condition. The methodology developed in this study explores the opportunity to extract useful information and gain insight from the data already being collected on mid-life vehicles. This approach is not restricted to engine failures and can be adapted to other system failures, making it a valuable tool in the industry's shift towards smart maintenance and industry 4.0 concepts.

Using Event Data to Build Predictive Engine Failure Models

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. What is event data in the context of this study?

2. How was bias removed from engine failure data?

3. How are positive examples extracted from ESBT events?

4. How are negative examples created in the study?

5. How to balance dataset in feature reduction?

6. How do C4.5 and random forest algorithms differ in handling missing values?

7. What are XAI methods?

8. What is the optimal data block window size for predicting engine failures in diesel multiple units?

9. What is the significance of predictive maintenance in diesel multiple unit engine failures?

Citations

A Fault Tree for Reliability Analysis of an Intercity Train Door

References

Random Forests

"Why Should I Trust You?": Explaining the Predictions of Any Classifier

Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI.

Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI

A review on machinery diagnostics and prognostics implementing condition-based maintenance

Related Papers (5)

Development of an AI model for electronic board maintenance decision prediction for railway equipment

A Profile Clustering Based Event Logs Repairing Approach for Process Mining

Machine learning-based methods for TTF estimation with application to APU prognostics

Predictive Maintenance using Machine Learning Based Classification Models

Algorithm for Generating Event Logs Based on Data from Heterogeneous Sources