Predictive Maintenance in HVAC industry

Transcript

1. Predictive Maintenance For HVAC
   Arnab Biswas
   Data Scientist, EcoEnergy, Carrier Global Corporation
   @arnabbiswas1
   

   View Slide

2. HVAC (Heating, Ventilation, Air Conditioning)
   Controls temperature, humidity & air quality for indoor environment
   

   View Slide

3. HVAC Equipment
   Accounts for 40% of energy usage in commercial buildings (US)
   Reduction in energy, cost & greenhouse gas emission can be
   achieved via proper design, installation & maintenance
   Heavily dependent on Annual Maintenance
   

   View Slide

4. Types of Maintenance
   Type Part
   Replacement
   Part’s Life
   Utilization
   Downtime
   Reactive On Failure Complete Unplanned
   Proactive Periodic,
   Before Failure
   Pre-determined Scheduled
   Predictive Just Before
   Failure
   Almost complete Scheduled
   

   View Slide

5. Predictive Maintenance
   Predict
   reason
   for failure
   Predict
   days to
   failure
   Predict a
   failure in
   future
   Predicting equipment’s failure over a future time period, based on its
   historical behavior
   

   View Slide

6. PdM Architecture
   

   View Slide

7. Data Science For PdM Understanding
   Business Problem
   Defining ML
   Problem
   Understand Data
   Prepare Data
   Build & Evaluate
   Model
   Deploy Model
   Maintain Model
   

   View Slide

8. Business Problem into ML Problem
   • Predict if HVAC is going to fail within next 2 weeks
   Binary Classification
   • Predict amount of time before next failure
   Regression
   • Predict if HVAC is going to fail in next 3w, 2w, 1w
   Multi-class
   Classification
   • Predict if HVAC is going to fail within next 2 weeks
   for a particular reason
   Multi-class
   Classification
   

   View Slide

9. Regression vs Classification?
   • Prefer Classification over Regression if business allows
   • Regression is equivalent to Classification with infinite number of classes
   • Single Stage Model
   • Multistage Model
   • Stage 1: Predict assets going to fail in next 2 weeks
   • Stage 2: Predict reason for failure for assets from Stage 1
   

   View Slide

10. Data
    Temporal
    Health
    Return Air
    Temperature
    Supply Air
    Temperature
    Run Time
    Set Point
    Temperature
    Power
    Consumption
    Zone
    Temperature
    Weather
    Outside Air
    Temperature
    Relative
    Humidity
    Pressure
    Maintenance
    Failure
    History
    Repair
    History
    Root Cause
    Parts
    Replaced
    Others
    Building
    Occupancy
    Static
    Equipment
    Metadata
    Climate Zone
    Geographic
    Location
    Building
    Metadata
    Maintenance
    Agency
    Data Collection
    

    View Slide

11. Prepare Data
    Structure data from various sources into tabular format
    

    View Slide

12. • Standard: Forward Filling, Interpolation
    • Domain Specific: Fill missing value of vibration on 1 PM, Tuesday
    • with last Tuesday 1 PM’s value
    • with Tuesday 1 PM’s value averaged over last 1 month
    • Strategy should be validated using cross-validation
    Missing Value Handling (TS Data)
    • Maintenance Data needs attention
    Duplicate Handling
    • TS Data: Data with higher deviation from normal may indicate degradation
    • Maintenance data: May indicate human error
    Outlier Handling
    Data Preprocessing
    

    View Slide

13. Feature Engineering (TS Data)
    • Rolling/Tumbling Aggregate
    • Need of Aggregation
    • Data Collection & Prediction frequency may not match
    • Data over individual time unit is noisy
    • Needs to be smoothened by aggregating over time windows
    • Lag Window
    • “How far in future the model has to predict” influences
    “how far in past the model has to look back”
    • Window Size should be validated using cross-validation
    Aggregation
    Average
    Maximum
    Minimum
    Median
    Standard
    Deviation
    Variance
    Count
    Sum
    Cumulative Sum
    Derivate
    2nd Derivate
    Count of outliers
    

    View Slide

14. Feature Engineering
    • Date
    • Day
    • Week
    • Weekday/Weekend
    • Month
    • Quarter
    • Year
    • …
    • Maintenance Data
    • Days since last failure
    • Days since last failure because of
    specific reason
    • Days since specific part replacement
    • Days since last maintenance
    • Static Data
    • Age of the equipment
    • Building Type
    

    View Slide

15. • Cross Validation
    • Split Data in Time Dependent Manner
    • On Test Data
    • Split by Time
    • Separate Train & Test data by the window size
    • Split by Equipment
    • Better performance with new equipment
    • Metrics
    • Imbalanced Data
    • High Cost of False Positive
    Measure Model Performance
    

    View Slide

16. Algorithms
    Phase I
    • Simpler Model
    • Simpler Features
    Phase II
    • Complex Features
    Phase III
    • Complex Models
    Multi-class
    Classification
    Regression
    RNN, LSTM RNN, LSTM
    DNN DNN
    GBM
    Random Forest
    SVM
    Hidden Markov Chain
    …
    GBM
    RF Regression
    …
    

    View Slide

17. Model Serving & Monitoring
    • Serving : Batch Scoring
    • Monitoring
    • Compare predictions with Ground Truth (Maintenance Data from Future)
    • Degradation of model’s performance
    • Change in the incoming data
    • Bug in Data Pipeline
    • Drift in data
    

    View Slide

18. References
    • Azure AI guide for predictive maintenance solutions
    • A process for implementing industrial predictive maintenance
    • A Survey of Predictive Maintenance: Systems, Purposes and Approaches
    • Rules of ML
    

    View Slide

19. https://speakerdeck.com/arnabbiswas1/predictive-maintenance-in-hvac-industry
    @arnabbiswas1
    

    View Slide