Io T and ML for Predictive Maintenance By

Io. T and ML for Predictive Maintenance By : Mukul Joshi & Mitra Daram Nitor Infotech Pvt. Ltd. Global Summit 2018

Evolution of Maintenance “Change is inevitable – except from a vending machine” – Robert Gallagher Since the invention of first machine, we had to deal with maintenance and repair. Thankfully we have evolved with the approach. • Not to wait till things Fix it when it breaks • Using Io. T to gather data • Use of Vibration analysis break. • Using ML to understand • Infrared imaging • Periodic checks. behavior/patterns Predictive Reactive • Oil analysis. Preventive etc. • Creates Unnecessary • ML models are used for Maintenance To do predictive work predictive maintenance • Labor intensive • Issues arise between schedules Io. T based Predictive Maintenance

Business Problems in predictive maintenance • Detect anomalies in equipment or system performance or functionality. • Predict whether an asset may fail in the near future. • Estimate the remaining useful life of an asset. • Identify the main causes of failure of an asset. • Identify what maintenance actions need to be done, by when, on an asset

Qualification of problem • Target or an outcome to predict. Clear path of action to prevent failures when detected. Problem has to be predictive in nature. • Record of the operational history of the equipment with both good & bad outcomes… Error reports, maintenance logs, repair logs etc. should be available. • The recorded history should be reflected in relevant data - sufficient enough quality to support the use case. • Finally, the business should have domain experts who have a clear understanding of the problem.

Sample use cases • Aviation : Flight delay and cancellation, Engine part failure • Finance : ATM failure • Energy : Wind turbine failure, circuit breaker failure • Transportation and logistics: wheel failure, subway train door failure

Five steps for predictive maintenance solution 1 2 3 4 Identify Assets Gather Data Design Predictive Model Manage Work 5 Learn & Act Smart

Data Preparation • Data should be Relevant, Sufficient & Quality • Relevant data sources are • • Failure history Maintenance/repair history Machine operating conditions Equipment metadata • There are two major types of these • Temporal data • Static data

Static vs. Temporal treatment • Visualize the data as table of records with row as training instance. The columns as “predictor” features. Final column as “target” • Maintenance records: Asset identifier, maintenance action, time etc. • Failure records: Failure reasons as error codes. Correlation between conditions and codes. • Machine & Operation metadata: Merged together to associate with assets. • Missing value handling and normalization • Temporal data is divided into units. Time unit for asset offers distinct information.

Feature engineering • Feature is a “predictive” attribute such as temperature, pressure, vibration etc. To predict the future, decide how much can be looked back. (Lag) Tumbling Aggregates Rolling Aggregates

Examples • Flight delay: count of error codes over the last day/week. • Aircraft engine part failure: rolling means, standard deviation, and sum over the past day, week etc. This metric should be determined along with the business domain expert. • ATM failures: rolling means, median, range, standard deviations, count of outliers beyond three standard deviations • Subway train door failures: Count of events over past day, week, two weeks etc. • Circuit breaker failures: Failure counts over past week, year, three years etc.

Modeling : Binary Classification • Probability that piece of equipment fails within future time period. (Future horizon period X) About to fail as label 1 (for period)

Examples • Flight delays: X may be chosen as 1 day, to predict delays in the next 24 hours. Then all flights that are within 24 hours before failures are labeled as 1. • ATM cash dispense failures: A goal may be to determine failure probability of a transaction in the next one hour. In that case, all transactions that happened within the past hour of the failure are labeled as 1. To predict failure probability over the next N currency notes dispensed, all notes dispensed within the last N notes of a failure are labeled as 1. • Train door failures: X may be chosen as two days. • Wind turbine failures: X may be chosen as two months.

Regression • Compute Remaining Useful Life (RUL) • Assets that have not failed can’t be used in modelling • Survival Analysis

Multi classification • Predict two outcomes • Range of time to failure • Likelihood of failure • What is the probability that asset will fail in next n. Z units of time?

Multi classification • What is the probability that the asset will fail in next X unit due to root cause/problem Pi?

Training, Validation and Testing • Failures are rare • Under-representation in training data • Cost sensitive learning : High cost to mis-classification of minority class. • Sampling techniques

Nitor case study Transformer Monitoring

Transformer Monitoring Business Situation • • 18, 000+ Transformers More than 250 Models / Variants Located Across the Globe (Different Geo Locations with different conditions) An Average of 10+ sensors capturing various readings (Oil temperature, Oil Volume, Load etc. ) Challenges • • • Data Generated from the transformers are in different frequencies Collection of Historic Data including Events (Failure, Breakdown, Down time etc. ) and Data Quality Domain Expertise Integration with External Data and Macro Environmental Factors Data Storage, Processing and Archival Mechanism Validating the Model

Solution Approach Step 1: Model Creation & Training with Historic Dataset Step 2: Production Stage Azure Batch AI 1 2 Anomaly detection & Failure Prediction algorithms Save model per device Real time Data Ingestion Time series data OR Aggregated data at time interval Data Collection Predict

Possible Impact & Benefits Business Impact • • Anticipating at least 30% reduction in Maintenance Costs Transformers Productivity in terms of uptime / availability can go up by 10% Number of Breakdowns / Total Failures can be avoided in at least 50% of the cases Optimization of Spare Parts Inventory Reduce Fuel Costs and Extend Life Worker Safety can be improved Maintenance Actions can be triggered based on the threat level rather than ‘Routine Checks’

Transformer Performance Dashboard

Thank you Mukul Joshi & Mitra Daram Nitor Infotech Pvt. Ltd.