Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

MLSEV Virtual. Anomalies in Oil Temperature Variations in a TBM


Published on

Anomalies in Oil Temperature Variations in a Tunnel Boring Machine, by Guillem Ràfales, Construction Management Product Leader at SENER, and Guillem Vidal, Machine Learning Engineer at BigML.

*MLSEV 2020: Virtual Conference.

Published in: Data & Analytics
  • Be the first to comment

  • Be the first to like this

MLSEV Virtual. Anomalies in Oil Temperature Variations in a TBM

  1. 1. 2nd edition
  2. 2. #MLSEV 2 Tunnel Boring Machines Oil Temperature Variations ML Analysis Guillem Rafales Queralt Infrastructures & Transport Construction Manager Leader SENER Guillem Vidal Machine Learning Engineer BigML
  3. 3. #MLSEV 3 Tunnel Boring Machines
  4. 4. #MLSEV 4 SENER´s Experience
  5. 5. #MLSEV 6 SENER´s Experience
  6. 6. #MLSEV 7 SENER´s Experience
  7. 7. #MLSEV 8 TBM Description Tunnel Boring Machines are those machines used to perform Rock tunneling excavation by mechanical means. Close faced shielded machines also permit the excavation in Soft Ground conditions. There are two basic types of pressurized closed-face tunneling systems: Slurry Tunneling machines and Earth Pressure Balance EPB machines. Courtesy video of an EPB machine, from Herrenknecht Ibérica S.A. and only for academic purposes.
  8. 8. #MLSEV 9 TBM Gear Oil Temperature Main bearing of a TBM is the mechanical core of the machine which enables TBM to turn cutter Head and transmits the torque of motorization to terrain to be excavated. The bearing maintains lubrication due of a full charge of gear oil of about 680 cSt viscosity. About 5000 liters of oil is necessary to full fit the main bearing of a 12m Ø tunneling machine. Oil analysis are performed to monitor wear status and adequate condition of the main bearing
  9. 9. #MLSEV 10 TBM External Factors Impact External factors have a high impact in TBMs outcome, such as difficult geological conditions, urban areas, logistic problems, economical, politic matters and influent stake-holders. Taegu Metro, Corea del Sur (2000) Colonia Metro, Alemania (2009)
  10. 10. #MLSEV 11 Objective Prepare a preliminary study about Tunnel Boring Machines gear oil temperature variations. Analyze relationships between oil temperature changes and the multiple TBM internal parameters. If possible, explore relevant oil temperature variations predictions to anticipate related failures Oil temperature being related to internal TBM matters, it has been picked as the PoC analysis objective to avoid external impact complexity. Oil temperature variations are slow and rare which makes the analysis a challenge
  11. 11. #MLSEV 12 Data
  12. 12. #MLSEV 13 Provided Data • 2GB of historic data from 1 tunnel have been provided • Over 1700 DBF files each one representing a tunnel ring • Each instance represents a TBM data measure with hundreds of attributes at a given instant • Measures are provided every 10 seconds Discarded data: • When the TBM is not advancing • When gear oil temperature is under 46 degrees(C)
  13. 13. #MLSEV 14 Main Features • Torque • Speed • Penetration • Forces • Pressure • Liquid flows and volumes • Chamber material measures • Times and frequencies • Amongst many others…
  14. 14. #MLSEV 15 Feature Engineering Future gear oil temperature variations • Tested with different intervals (3, 5, 10, 15, 30 and 60 minutes) to find relevant variation causes as short as possible • A boolean flag indicating 0.3 degrees (C) temperature raises in the next 10 minutes has been kept Sliding windows data for the past 5 minutes • A 5 minutes past summary has been aggregated for overall features including ranges, standard deviation and averages Resulting dataset 120k rows, 5k temperature raises (4.3%), 171MB
  15. 15. #MLSEV 16 Data Insights
  16. 16. #MLSEV 17 Association Discovery date customer account auth class zip amount Mon Bob 3421 pin clothes 46140 135 Tue Bob 3421 sign food 46140 401 Tue Alice 2456 pin food 12222 234 Wed Sally 6788 pin gas 26339 94 Wed Bob 3421 pin tech 21350 2459 Wed Bob 3421 pin gas 46140 83 Thr Sally 6788 sign food 26339 51 An unsupervised algorithm that looks for coincidences in the data and returns association rules with an antecedent, a consequent and several metrics. It is very useful for data exploration amongst other goals. Example: zip = 46140 amount < 100 Antecedent Consequent Rules: {customer = Bob, account = 3421} {class = gas}
  17. 17. #MLSEV 18 Association Discovery Consequent: Temp_range_10m > 0.3 degrees (temperature range in the next 10m)
  18. 18. #MLSEV 19 Association Rule Example Rule summary • When: • pushing force metric has been high for 5 minutes • temperature has been to an average value for 5 minutes • material is currently being cumulated inside the machine • Then the oil gear temperature increases over 0.3 degrees during the next 10 minutes Rule metrics • 100% confidence (32 instances in 2 different rings)
  19. 19. #MLSEV 20 Data Exploration • Multiple Association Discovery have been trained in the data exploration phase. In each case results were shared with SENER experts to identify irrelevant and interesting rules • After several iterations Association Discovery attributes were sufficiently optimized to produce interesting rules • A final set of rules was sent to SENER in order to select the most potentially interesting rules and proceed with specific analysis for each one • In general a large amount of rules contained negative material deviation features • Data observations showed material accumulations in the chamber were often tied to temperature raises
  20. 20. #MLSEV 21 Data Visualization Grafana screenshot showing a case where negative material deviation (bottom right) happen together with gear oil temperature raise (top left)
  21. 21. #MLSEV 22 Anomaly Detection
  22. 22. #MLSEV date customer account auth class zip amount Mon Bob 3421 pin clothes 46140 135 Tue Bob 3421 sign food 46140 401 Tue Alice 2456 pin food 12222 234 Wed Sally 6788 pin gas 26339 94 Wed Bob 3421 pin tech 21350 2459 Wed Bob 3421 pin gas 46140 83 Thr Sally 6788 sign food 26339 51 23 Anomaly Detection An unsupervised algorithm that looks for unusual instances in a dataset. Anomaly detectors provide an anomaly score to each instance, the higher is the score the most unusual is the instance. Example: • Amount $2,459 is higher than all other transactions • Only transaction • In zip 21350 • For the purchase class “tech"
  23. 23. #MLSEV 24 Anomaly Detector Features: Only material deviation variables
  24. 24. #MLSEV 25 Anomaly Detector Results • Top anomalies often correspond to temperature raises over 0.3 degrees in the next 10 minutes • Filtering anomaly scores over 50% results into 4.5% of the original data including 12.6% of the original temperature raise instances • Resulting dataset is more balanced: 11.8% temperature raises (orange data points in the graph)
  25. 25. #MLSEV 26 Classification
  26. 26. #MLSEV 27 Classification • After filtering high anomaly scores the dataset is more balanced and classification makes sense with supervised learning. The goal is to predict whether the gear oil temperature will raise considerably in the future 10 minutes • Data has been split linearly to evaluate models: • Training dataset 70%: 3687 rows, 273 with high temperature raise • Test dataset 30%: 1804 rows, 377 with high temperature raise • Feature engineering and feature selection has been performed resulting into a subset of features including material deviation, oil temperature and some generic parameters
  27. 27. #MLSEV 28 Decision Tree Decision trees being simple and interpretable have been chosen as the initial classification approach
  28. 28. #MLSEV 29 Decision Tree Evaluation • Precision and Recall are 45.5% and 48.3% respectively. Not specially high but still better than average. This means predictions are possible. Results are much better than in preliminar tests • Ideally current evaluation results should be improved • OptiML has been used to optimize the algorithm and parameters choice
  29. 29. #MLSEV 30 OptiML Model Optimization • An OptiML has been trained to optimize ROC Area Under the Curve metric using the same training and test datasets • OptiML resulting ROC AUC measures appear much higher than the decision tree ones, there is a clear improvement • Best results are achieved by decision tree ensembles
  30. 30. #MLSEV 31 Ensemble Evaluation • With a 39% probability threshold precision and recall values 50.9% and 69.5% • This means model would be able to predict half of the temperature raises and 70% predictions would be correct • The Area Under the ROC Curve (ROC AUC) is almost 80% which is a good overall indicator for the model
  31. 31. #MLSEV 32 Summary
  33. 33. #MLSEV 34 Results • An Anomaly Detector has isolated 12.6% of all temperature raise cases within a smaller dataset (4.6% overall). 650 temperature raises were filtered into a smaller dataset of 5501 instances • Based on Recall in the evaluation, overall 8.8% of temperature raises would be predicted by the Anomaly Detector together with the Ensemble • Based on Precision, the Anomaly Detector together with the Ensemble could predict a temperature raise on 3.8% of the original dataset with 1.9% correct predictions • Based on this numbers the Machine Learning workflow could predict 8.8% temperature raises and provided predictions would have a 50% reliability
  34. 34. #MLSEV 35 Conclusions •Material deviation real time alerts could be implemented in Tunnel Boring Machines using this method •It has been proved Machine Learning means can provide useful TBM insights •Plenty of other Machine Learning analyses and implementations are possible in TBMs. An advanced cockpit could be implemented