synthesis and refinement of artificial hvac sensor data intended for supervised learning in...

20
IEECB&SC Efficiency in Commercial Buildings and Smart Communities Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques Presenter: David McCabe Wednesday, 14 January 2016

Upload: ies-ve

Post on 09-Feb-2017

1.583 views

Category:

Technology


0 download

TRANSCRIPT

Page 1: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

IEECB&SC Efficiency in Commercial Buildings and Smart Communities

Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Presenter: David McCabe

Wednesday, 14 January 2016

Page 2: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

IES Building Design Tools

Page 3: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

IES Future: Building Operation Tools

profiles

SCAN:

Page 4: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

EINSTEIN:

Einstein Overview:

- Marie Curie Grant funded FP7 project- Partnered with Trinity College Dublin- Four year duration- Secondments between academia and industry

Objectives:

- Forge partnerships between Industry and Academia- Develop a prototype smart building control framework- Use performance prediction and control optimisation- Exploit framework further with an AFDD capability

Page 5: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Prevalence and Effect of Building Faults

“Poorly maintained, degraded, and improperly controlled equipment wastes an estimated 15% to 30% of energy used in commercial buildings.”

- [Brambley et al 2005]

“estimated 25% to 45% of energy consumption in HVAC plant serving commercial buildings is wastage due faults”

- [Akinci et al 2011]

“Studies have indicated that 20–30 % HVAC system energy savings are achievable by recommissioning air handling units (AHU) to rectify faulty operation.”

- [Bruton et al 2013]

Page 6: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Building Fault Detection & Diagnostics

1) *Fault Detectionidentify whether or not a fault is present

2) *Fault Diagnosisdetermine the precise source of the problem

3) Impact Evaluation

cost, comfort, environmental or safety impact

4) Decision/Actiontolerate or shutdown for repair etc.

3) Impact Evaluation

4) Decision/Action

System e.g. building HVAC

Fault Detection & Diagnostics (FDD)

1) Fault Detection

2) Fault Diagnostics

BEMS Data

*Fault Detection & Diagnostics

Page 7: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Automated Fault Detection & Diagnostics

Desirable to detect and diagnose faults as quickly as possible, automated approach is favourable over a traditional manual approach.

Diligence and round the clock monitoring Immediately interprets high volumes Risks false alarms without proper guidance Takes time to develop Require high volumes of real-time data

- commonly used within mass manufactured units such as consumer electronics and vehicles, through the utilisation of their on-board electrical architecture.

- noticeably absent from commercial buildings perhaps due to both their lack of design homogeneity and in many cases the low availability of sensor data. Further more AFDD in sectors such as aerospace and automotive sectors also fulfils a vitally critical safety function.

Page 10: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Data Driven AFDD: Neural Networks

Forward propagating sensor readings through a trained MLP returns fault predictions…

The ANN’s parameter values are trained using labelled fault data (Supervised Learning).

Return Temp

Flow Rate

Elec.

Fan [SF005] breakdown pred.

Valve [#142] breakdown pred.

ALERT: Check Valve - No. 142

Page 11: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

- regions of the feature space:- non-faulty system operation- faulty system operation

- fault type 1, type 2, type 3 …

AFDD: Classification Problem

Sensors variables are known as features e.g.

Sensor variables form a feature space:

- the complete set of sensor time series data represents a feature space trajectory

Fault Detection:(Binary classification problem)

Fault Classification & Diagnosis:(Multi-Variable classification problem)

An Artificial Neural Network is a type of classifier. It maps or categorises the feature space into different regions.

Page 12: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Neural Networks & Feature Space

decision boundary takes the form of a parameterised function represented by a network:

… network topology determines the complexity of the prediction boundaryParameterised weights in the network determine the exact positions of the decision boundaries.

Page 13: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Classifier Training

supervised learning technique involves teaching the classifier the predication boundaries using:- some labelled training data- a learning algorithm

Fault Detection: Fault Diagnostics:

Page 14: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Faulty Operation Data Procurement Issues

Data-driven Approach:Here are some of the problems with fault operation BEMS data:

- insufficient datanot enough data

- unbalanced data (Skewed Classes)one class is over-represented in the data

- incorrect labellinge.g. periodic problem is labelled continuously

- unknown faultsfaults other than the known types may be present

Effect: Extreme mismatch between decision boundaries and fault regions

Page 15: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Fault Data Synthesis Overview

Synthesis of Artificial Building Fault Data using IESVEtechnique for the quick production of high volumes well labelled training data

- Simulation ensures saturation-Anomaly Detection ensures observability-Balancing ensures uniform distribution

Harvesting (Threshing ) of Data from Fault Simulations

Fault Type IISimulation

Fault Type IVSimulation

Training,Cross

Validation& Testing

Data

DataRefinement

DataBalancing

NotRequired

Synthetic Fault DataLabelling(AnomalyDetection)

No AnomalyDetected

Anomaly DetectedRequired

DataBalancing

Synthetic Fault DataLabelling(AnomalyDetection)

Anomaly DetectedRequired

RawSyntheticFault DataAcquisition

No AnomalyDetected

RawSyntheticFault DataAcquisition

NotRequired

Page 16: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Synthesis of Artificial Fault Data (1 of 2)Fault Simulation UtilitySimulating faulty operation using IESVE software to produce unlabelled results.

Programmatic perturbations to operational and design profiles allow us to programmatically :- introduce faults- diversify results

Higher coverage of thebuildings operationalEnvelope in feature space

Data labelling, volume, and balancing now require refinement…

Page 17: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Synthesis of Artificial Fault Data (2 of 2)Anomaly detection using Gaussian Kernel Density Estimation

Example using a single feature (sensor) example…

Page 18: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Use of pure Synthetic Data for Classifier

System has only been evaluated using simulated data - not on a real building!

99.8% of faults successfully detected 100% of detected faults successfully diagnosed No fault alarms AFDD only performed on simulation data

Page 19: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Future WorkConcerns:- model is not a precise representation of the building

– region boundaries for model and building may be quite different

- accurate modelling can be as time consuming as designing an expert rule set and calibration requires actual data – commissioning needs to be done first!

- Kernel Density Estimation (pictured) does not scale well

Page 20: Synthesis and Refinement of Artificial HVAC Sensor Data Intended for Supervised Learning in Data-Driven AFDD Techniques

Q&A