The document discusses United Technologies Corporation's application of deep learning techniques to problems in aerospace and building systems. Specifically, it discusses using deep belief networks for aircraft sensor diagnostics at Pratt & Whitney and Otis elevators prognostic health monitoring. It also discusses using deep autoencoders for chiller power estimation at Carrier Climate Control systems. The approaches analyzed sensor data using deep learning models to provide diagnostics, predict health issues, and estimate power usage.
Deep Learning for industrial Prognostics & Health Management (PHM)
1. Applying Deep Learning to Aerospace and Building System Applications at UTC
VivekVenugopalan, Kishore Reddy and Michael Giering
• Deep Learning is an evolving area of research in
neural networks and it has been adopted by UTC for
tackling various problems in aerospace and building
systems.
•Three different use cases discussed here: (1) Aircraft
sensor diagnostics for UTAS, Pratt & Whitney, (2)
Prognostic Health Monitoring for Otis Elevators, (3)
Chiller power estimation for Carrier Climate
Control systems
• Aircraft sensors provide huge amount of data that
needs to be tracked such as air data systems, fuel
measurement and management systems, health and
usage systems and mission data recorders.
[1]Y. Bengio, P. Lamblin, D. Popovici, H. Larochelle, et al.,“Greedy layer-wise training of deep networks,” Advances in neural information processing systems, vol. 19, p. 153, 2007.
[2] P.Vincent, H. Larochelle,Y. Bengio, and P.A. Manzagol,“Extracting and composing robust features with denoising autoencoders,” in ICML, 2008
[3] F. Bastien, P. Lamblin, R. Pascanu, J. Bergstra, I. Goodfellow,A. Bergeron, N. Bouchard, D.Warde-Farley, andY. Bengio,“Theano: new features and speed improvements,” arXiv preprint arXiv:1211.5590, 2012
[4] M. Giering,V.Venugopalan, and K. Reddy.“Multi-modal sensor registration for vehicle perception via deep neural networks”. In IEEE High Performance Extreme Computing Conference (HPEC), 2015.
Implementation and Results
Introduction
Conclusion
References
Deep Auto-Encoders
• 4xNvidia K40 GPUs with with 2880 cores and 12
GB device RAM each in Ubuntu OS workstation
•Theano based toolchain for Deep Learning
• Nvidia K40 with 12 GB device RAM - driving factor
for large dataset inhalation, caching and computation -
especially the pre-training stage for DBNs
Email:{venugov, gierinmj, reddykk}@utrc.utc.com
Deep Belief Nets
Layer 1
Layer 2
Bottleneck layer
Input layer
W2
T
Layer 1
Layer 2
RBM
RBM
RBM
Recursive pre-training
W1
T
W3
T
• Successful adoption of Deep Learning
methodologies to UTC applications in
aerospace and building systems as shown
in the timeline.
• Deep Belief Nets (DBN) consist of using a
probabilistic Restricted Boltzmann Machine
(RBM) approach, trying to reconstruct noisy
inputs.
• Training involves the reconstruction of a clean
sensor input from a partially destroyed/missing
sensor.
•Depending on the application, a final layer can
be added after the bottleneck layer.
• Deep Auto-Encoders (DAE) performs the fine-
tuning by generating the layers mirroring the initial
network upto the bottleneck layer after the pre-
training using the DBNs.
• The weights and the bias of the upper and lower
hidden layers for the DAE are updated in the fine-
tuning stage.
• The main objective of the DAE is to minimize the
reconstruction error.
utcaerospacesystems.com
MRO & Support Services
Features more than 6,000 customer service
employees across 16 countries dedicated to the
operation of nearly 60 MRO service and support
facilities. Customer Response Center available
for a range of needs – from AOG to spare parts
and technical support. Offers customized support
agreements to help operators achieve optimal
aircraft utilization.
+1 877 808 7575 crc@utas.utc.com utascrc.com
150004001.indd 05/27/2015
Actuation & Propeller Systems
Designs and manufactures actuation and propeller
systems for commercial and military aircraft. Products
range from single actuators to complete flight control
systems for the fixed wing, rotorcraft and missile
segments as well as fly-by-wire cockpit controls,
cabin equipment, trimmable horizontal stabilizer
actuators and flight safety parts for helicopters.
Engine & Environmental
Control Systems
Provides engine controls, accessories and solutions
for turbofan, turboprop and turboshaft engines
and environmental control systems for aerospace
and defense applications. Engine products include
electronic engine controllers, fuel systems, engine
actuation, thermal management systems, accessory
drive gearboxes and transmissions, drive shafts
and flexible couplings, engine start systems, turbine
blades and vanes. Environmental control systems
include air conditioning, liquid cooling, engine
bleed air, pressurization control, ventilation control,
humidification and fuel tank inerting.
Landing Systems
Designs, manufactures and services fully integrated
landing systems such as main and nose gear
structures, electric and hydraulically actuated
brakes with steel or carbon friction material, and
brake control systems. Innovative solutions include
more electric technologies, DURACARB®
carbon
friction material, EDL®
extended life configurations,
and lighter-weight, high-strength materials.
Sensors & Integrated Systems
Provides cutting-edge sensors and sensor-based
systems for the commercial aerospace, ground
vehicle and defense industries including electronic
flight bags, air data systems, ice detection and
protection systems, fire protection systems, fuel
measurement and management systems, guidance
navigation and control systems, health and usage
management systems, rescue hoists, mission data
recorders, and sensing suites for aircraft engines.
Interiors
Designs, manufactures and supports advanced
systems that enhance safety, performance and
aesthetics across a wide range of commercial,
business jet and military aircraft. Provides
WINSLOW life rafts, interior and exterior lighting
systems, aircraft evacuation systems, cargo
systems, pyrotechnic egress systems, VIP and
specialty seating systems including Advanced
Concept Ejection Seats (ACES II and ACES 5), and
cabin systems featuring custom-crafted artisan
Booth Veneers, cabin management systems and
in-flight entertainment products.
ISR & Space Systems
Provides products and services to global
government and commercial markets that enable
mission success in space, in the air, at sea and
on the ground. Manufactures products providing
actionable intelligence through surveillance and
reconnaissance solutions; products for small
unmanned airborne systems; state-of-the-art
Shortwave Infrared (SWIR) products to support
warfighters; and environmental control and life
support systems that enable humans to safely
operate in space and under the sea.
Electric Systems
Provides electric power systems for commercial,
regional, business, and military aircraft. Products
include main and emergency power generation,
power conversion and motor control, power
distribution, and aircraft utilities management. A
complete range of electric power generation options
is provided, including constant and variable frequency
AC and high-voltage DC.
Aerostructures
Designs, manufactures, and integrates nacelles,
thrust reversers, pylons and flight control surfaces
for commercial and military aircraft. Aerostructures
includes the Engineered Polymer Products business,
which designs, tests and manufactures composite
components for ships, submarines and commercial
airplanes.
UTC Aerospace Systems A range of capabilities
On-board sensor diagnostics and data
collection (FAST box)
e.g. fuel measurement and
management systems, mission data
recorders, etc.
Integrated sensor management and
real-time analysis for variety of sensing
suites for aircraft engines
Aircraft sensors
Layer 1
Layer N
Bottleneck layer
Input layer
Layer 1
Layer N
Output layer
DBN pre-training
Bank of elevators
Sensors embedded
for prognostic
health monitoring
Diagnostic
and decision
• Sensors embedded in Otis
Elevator systems mainly
used for collecting data
about the health of the
system
• Chillers used with Carrier
HVAC units - understanding
energy requirements
Carrier Chillers in HVAC units- understanding
more about optimizing the energy utilization
Chiller power output prediction based on the inputs to DBN
GIVEN --->
PREDICT --->
Watts
Chiller power output reconstruction
Blue – Original
Red – Predicted
Watts
Sensor estimation from the FAST box
Algorithm Reconstruction error
Discrete Bayesian Network 17192.63
Continuous Bayesian Network 17966.18
Structured Learning 14921.63
Koopman 16823
Deep Learning 10819.55
• Benchmarked Carrier Chiller energy
utilization using variety of Machine
Learning algorithms.
• Deep Learning approach provided the
lowest reconstruction error enhancing the
energy prediction capability.
Deep Auto-Encoders
Elevator data streamed
using smartphone app Damage information:
Good cab door
Moderate or severe
damage to cab door
• Sensors embedded in the elevators streamed
using smartphone app and then fed to the Deep
Auto-Encoder.
• Performance metric measured in terms of the
health of the elevator cab door
Timeline of Deep Learning adoption and application to UTC
•Variety of use cases - sensor estimation from onboard sensing suites on aircraft
engines using DBN, chiller power prediction for building systems using DBN, PHM in
elevator systems using DAE.
• Huge amount of data generated - offline training using Nvidia GPUs.
• Online diagnostics and decision using Nvidia’s Jetson GPUs - future.
ECCN: 9E991 - This information is subject to the export control laws of the United States, specifically including the Export Administration Regulations (EAR), 15 C.F.R. Part 730 et seq.Transfer, retransfer or disclosure of this data
by any means to a non-US person (individual or company), whether in the U.S. or abroad, without any required export license or other approval from the U.S. Govt. is prohibited.
File: PWC_400057_0787331054_1
PSNR: 27.69 dB
NRMS: 4.46 %
File: PWC_400057_0788706605_1
PSNR: 34.22 dB
NRMS: 2.11 %
File: PWC_400057_0839520402_1
PSNR: 25.47 dB
NRMS: 5.72 %
Y-axisvaluesconfidential
Y-axisvaluesconfidentialY-axisvaluesconfidential
2015 Q1 2015 Q2
Problem formulation and
capability development
Algorithm fine-tuning and
technology demonstration
Data collection on-field,
experimental setup
Infrastructure development,
toolchain selection