Experimental wind tunnel study of a smart sensing skin for condition

Experimental wind tunnel study of a smart sensing

skin for condition evaluation of a wind turbine blade

Austin Downey

, Simon Laﬂamme

, Filippo Ubertini

1 Department of Civil, Construction, and Environmental Engineering, Iowa State

University, Ames, IA, USA;

2 Department of Civil and Environmental Engineering, University of Perugia,

Italy;

E-mail: [email protected]

June 2017

Author’s preprint, published in Smart Materials and Structures with

DOI:10.1088/1361-665X/aa9349

Abstract.

Condition evaluation of wind turbine blades is diﬃcult due to their large

size, complex geometry and lack of economic and scalable sensing technologies

capable of detecting, localizing, and quantifying faults over a blade’s global area.

A solution is to deploy inexpensive large area electronics over strategic areas of the

monitored component, analogous to sensing skin. The authors have previously

proposed a large area electronic consisting of a soft elastomeric capacitor (SEC).

The SEC is highly scalable due to its low cost and ease of fabrication, and can,

therefore, be used for monitoring large-scale components. A single SEC is a strain

sensor that measures the additive strain over a surface. Recently, its application

in a hybrid dense sensor network (HDSN) conﬁguration has been studied, where

a network of SECs is augmented with a few oﬀ-the-shelf strain gauges to measure

boundary conditions and decompose the additive strain to obtain unidirectional

surface strain maps. These maps can be analyzed to detect, localize, and quantify

faults. In this work, we study the performance of the proposed sensing skin at

conducting condition evaluation of a wind turbine blade model in an operational

environment. Damage in the form of changing boundary conditions and cuts in the

monitored substrate are induced into the blade. An HDSN is deployed onto the

interior surface of the substrate, and the blade excited in a wind tunnel. Results

demonstrate the capability of the hybrid dense sensor network and associated

algorithms to detect, localize, and quantify damage. These results show promise

for the future deployment of fully integrated sensing skins deployed inside wind

turbine blades for condition evaluation.

Keywords: structural health monitoring, capacitive-based sensor, soft elastomeric

capacitor, ﬂexible membrane sensor, sensor network, damage detection, damage

localization

Smart Mater. Struct.

A Downey et al 2

1. Introduction

The proﬁtability of industrial-scale wind energy

projects is challenging due to their reliance on

public subsidies, unpredictable energy source, and

reliable technology. Additionally, varying operation

and maintenance (O&M) costs add complexity and

uncertainty to the management of wind energy projects

[1]. To achieve an increase in wind turbine system

reliability and therefore decrease costs related to

wind energy production, an O&M approach that

utilizes condition-based maintenance (CBM) should

be implemented [2, 3]. The use of condition based

maintenance is even more important for oﬀshore farms

where O&M costs may be up to three times higher than

that of land-based systems [4], due largely to higher

transportation and site access costs [5]. The current

state of condition monitoring of wind turbine blades

consists mainly of vibrations, and visual analyses [2, 6].

Recently, interest has grown in the use of structural

health monitoring (SHM) for the condition assessment

of wind turbine blades, towers and other structural

components due to their high replacement cost [4, 7],

eﬀect on system availability [5], and maintenance

complexity [8]. Monitoring the mesostructures of wind

turbines (e.g., towers and blades) is diﬃcult due to the

need to distinguish between faults in the structure’s

global (e.g. changing load paths, loss in global

stiﬀness) and local (e.g. crack propagation, composite

delamination) conditions [9]. Recent attempts for

the SHM of wind turbine blades have used a limited

number of sensors and have applied a variety of post-

processing techniques (e.g. statistical and modal-

based) to localize damage [10, 11]. However, this

approach lacks the capability to distinguish local

failures from global events and has demonstrated a

limited eﬀectiveness at damage localization [6, 12].

A solution to this local/global detection problem

is to deploy a dense sensor network (DSN) inside the

component that is capable of detecting local faults.

These integrated sensing skins mimic biological skin

in that they are capable of detecting and localizing

damage over the blade’s global area and with the

objective to enable low-cost, direct sensing of large-

scale structures. Sensing skins can be made of

large area electronics [13] or of rigid or semi-rigid

cells mounted on a ﬂexible sheet [14]. Early work

in the ﬁeld of sensing skins consisted of capacitive-

[15] and resistance- [16] based tactile force sensors.

More recently, sensing skins with piezoceramic (PZT)

transducers and receivers built into a ﬂexible skin have

been proposed [17]. In certain cases, sensing skins

with the integrated electronics for data acquisition

and signal processing mounted directly onto the skin

have been developed [18, 19]. Various researchers have

proposed and experimentally validated sensing skin-

type solutions for wind turbine blades. For instance,

Song et. al. demonstrated through experimental

validation in a wind tunnel that a network of

piezoceramic (PZT) sensors can be used to detect

damage in wind turbine blades [20]. Schulz et al.

proposed the use of series-connected PZT nodes for the

continuous monitoring of wind turbine blades, allowing

for a ﬁner localization of damage [17]. Simulations were

used to show that an array of these sensors, deployed

on a 2D plate, could be used to detect and localize

damage. Ryu et al. demonstrated a self-sensing thin

ﬁlm fabricated from poly(3-hexylthiophene) (P3HT)

and multi-walled carbon nanotubes (MWNTs) that is

capable of monitoring strain through the photocurrent

generated by the photoactive nanocomposite [21].

These sensors are capable of generating their own

power, therefore eliminating their need for external

power sources. Rumsey et al. deployed a number of

SHM systems on the outside of an experimental wind

turbine blade at Sandia National Laboratories [22].

Various sensor technologies were used, including PZT

and strain-based sensors, to monitor the blade during

a fatigue test. In general, successful damage detection

was found to require an optimal sensor placement and

synchronization of sampling between diﬀerent sensor

types.

In this work, the authors present the vision of a

fully integrated DSN for the real-time SHM of wind

turbine blades and experimentally validate a prototype

skin that demonstrates the feasibility of the concept.

This DSN consists of an inexpensive and robust large

area electronic consisting of a highly elastic capacitor

based on a styrene-co-ethylene-co-butylene-co-styrene

(SEBS) block co-polymer. Termed the soft elastomeric

capacitor (SEC), the sensor is customizable in shape

and size [23]. The SEC possesses the unique capability

of measuring the substrate’s additive strain (ε

+ ε

and its static [24] and dynamic [25] behaviors have been

well documented including numerical demonstrations

for damage detection applications in wind turbine

blades [26].

A particularly useful attribute of the SEC is its

capability to measure additive in-plane strain. It

follows that the signal must be decomposed into

orthogonal directions in order to obtain unidirectional

strain maps. A previously developed algorithm is

used in this work to decompose the sensors’ additive

strain into estimated unidirectional strain maps [27].

The algorithm, termed the extended least squares

estimator algorithm, leverages oﬀ-the-shelf sensors

such as resistive strain gauges (RSGs), to form a hybrid

DSN (HDSN). A deﬂection shape function for the

monitored substrate is assumed along with boundary

conditions (assumed or measured through the RSGs)

and uses the least squares estimator (LSE) to solve

A Downey et al 3

dielectric

conductive

plate

contacts

(a) (b)

wires embedded

in flexible substrate

wind turbine

blade

data bus

capacitance-to-digital

converter

low-resolution

DSN/HDSN

high-resolution

DSN/HDSN

control/wireless

transmission node

flexible substrate

Figure 1. Conceptual layout of a fully integrated SEC-based sensing skin for a wind turbine blade: (a) SEC with connectors and

annotated axis; and (b) proposed deployment inside a wind turbine blade.

for the shape function’s coeﬃcients. In this work, the

reconstructed strain maps are inspected to investigate

how damage induced into the monitored substrate

changes the loading path of the blade. Thereafter,

it is shown that damage in the form of leading edge

faults (e.g. changing boundary conditions) can be

localized through changing the assumed boundary

conditions of the plate. Lastly, the quality of these

unidirectional strain maps is measured in the form of

a reconstruction error to develop a damage detecting

feature for a predeﬁned section of the HDSN [28].

This network reconstruction feature (NeRF) algorithm

allows the sensing skin to fuse the high-channel-count

sensing skins data into a single damage detecting

feature, therefore providing a high level of data

compression and increasing the functionality of the

proposed system.

This paper experimentally veriﬁes the HDSN,

deployed inside a model wind turbine blade excited by

aerodynamic loading in a wind tunnel. The reported

results are the ﬁrst use of a large area electronic

for damage detection in a wind turbine blade under

aerodynamic loading. These tests validate the use

of SECs in a wind turbine blade and demonstrate

the potential utility of the concurrently proposed,

fully integrated, SEC-based sensing skin. The

contributions of this work are three-fold: 1) propose

an integrated SEC-based sensing skin for the real-time

structural health monitoring of wind turbine blades; 2)

demonstrate the capability of the SECs to operate in

the electromagnetically noisy environment of a wind

tunnel, showing that the SEC would be capable of

operating inside the similarly noisy environment of

a wind turbine blade; 3) evaluate the HDSN data

through previously developed algorithms showing that

the SEC-based sensing skin is capable of detecting

damage within an HDSN that is not directly monitored

by an SEC.

2. Background on Sensing Skin

The SEC-based sensing skin is illustrated in Figure 1,

with the sketch of an individual SEC shown in Figure

1(a). The fully integrated DSN system, as presented

in Figure 1(b), would consist of SECs of varying

geometries and densities along with the required

electronics for power management, data acquisition,

data processing, and communications, all mounted

onto a ﬂexible substrate (e.g. Kapton). The optimal

placement of RSGs within a grid of SECs has been

previously used by the authors to improve the accuracy

of strain map reconstruction from SEC data [29].

These sensing skins would be deployed inside a wind

turbine blade, either at the factory or in the ﬁeld to

monitor cases of interest, such as repair made at the

root of a blade [8].

Data (capacitance) for a set of SECs in close

proximity would be collected by a centrally located

capacitance-to-digital converter, multiplexed to mea-

sure multiple SECs. These converters are located close

to the SECs to allow for low noise measurements,

while multiplexing allows the sensing skin to function

with a reduced number of converters. Data would

be transferred over a serial bus (e.g. CAN, I2C) to

a control/wireless transmission node, this conﬁgura-

tion allows multiple capacitance-to-digital converters

per transmission node, therefore reducing the number

of wireless channels needed. These control nodes col-

lect, process, and parse the data for wireless transmis-

sion back to a wireless hub mounted inside the rotor

hub. The use of wireless transmission nodes allows for

the easy installation of a sensing skin, particularly in

A Downey et al 4

cases where a sensing skin is added to an in-service

blade such as that needed to monitor a repair. Addi-

tionally, wireless transmission adds redundancy to the

system when compared to a single serial bus being used

to carry data over the entire length of the blade, a use-

ful feature given the long service life of wind turbine

blades. Power can be provided through a variety of

methods, including energy harvesting (for sensing skins

mounted inside a wind turbine blade), ﬂexible solar

cells embedded into the sensing skin (when mounted

on the outside of a wind turbine blade) or batteries

when only short-term monitoring is required.

In the rest of this section, the background on

the SEC sensor is provided, which includes its electro-

mechanical model, followed by a review of the extended

LSE algorithm and the NeRF algorithm for damage

detection, localization, and quantiﬁcation.

2.1. Soft Elastomeric Capacitor

The SEC used in the sensing skin is a robust large

area electronic that is inexpensive, easy to fabricate,

and customizable in shape and size. The sensor’s

fabrication procedure is described in Ref. [23].

Brieﬂy, the sensor’s dielectric is composed of a styrene-

ethylene-butylene-styrene (SEBS) block co-polymer

matrix ﬁlled with titania to increase both its durability

and permittivity. Conductive plates are painted onto

each side of the SEBS matrix using a conductive paint

fabricated from the same SEBS, but ﬁlled with carbon

black particles. Material, equipment and techniques

used in the fabrication are readily available and the

sensor’s fabrication process is relatively simple, making

the technology highly scalable.

The SEC transduces a change in a monitored

substrate’s geometry (i.e., strain) into a measurable

change in capacitance. It is stretched during its

application to enable tensile and compressive strain

measurement and is adhered using commercial epoxy.

Assuming a low sampling rate (< 1 kHz), the SEC can

be modeled as a non-lossy capacitor with capacitance

C deﬁned by the parallel plate capacitor equation,

C = e

(1)

where e

= 8.854 pF/m is the vacuum permittivity,

is the polymer relative permittivity, A = d · l is

the sensor area of width d and length l, and h is

the thickness of the dielectric as annotated in Figure

1(a). Assuming small strain, an expression relating

the sensor’s change in capacitance to its signal can be

expressed as [25]

∆C

= λ(ε

+ ε

) (2)

where λ = 1/(1 − ν) represents the gauge factor of the

sensor, with ν being the sensor material’s Poisson ratio.

For SEBS, ν ≈ 0.49, which yields a gauge factor λ ≈ 2.

Equation (2) shows that the signal of the SEC varies

as a function of the additive strain ε

+ ε

2.2. Strain Decomposition Algorithm

The extended LSE algorithm was designed to decom-

pose the SEC signal’s additive strain measurement, as

expressed in Equation (2), by leveraging an HDSN con-

ﬁguration consisting of SECs and unidirectional strain

sensors (e.g. RSGs). RSGs measure boundary condi-

tions within the HDSN that can be used to increase the

capability of the extended LSE algorithm to decompose

strain maps. Boundary conditions on the edges of the

structure are also introduced into the algorithm as vir-

tual unidirectional sensors at locations where the uni-

directional strain can be assumed within a high level of

conﬁdence. The extended LSE algorithm is presented

in reference [27], diagrammed in the red dashed rect-

angle in Figure 2, and summarized in what follows.

The extended LSE algorithm assumes a p

order

polynomial displacement shape function (w), selected

due to its mathematical simplicity and its capability

to develop a wide range of displacement topographies.

The deﬂection w in the x-y plane can be written

w (x, y) =

i=1,j=1

(3)

where b

i,j

are regression coeﬃcients. Considering

an HDSN with m sensors (SEC and RSGs in this

case), displacement values at sensors locations can be

collected in a vector W ∈ R

. Equation (3) becomes

W =



· · · w



= HB (4)

where the subscript k is associated with the k-th

sensor. Matrix H contains sensor location information,

and B contains the f regression coeﬃcients B =



· · · b



Matrix H is deﬁned as H = [Γ

|Γ

] where

and H

account for the SEC’s additive strain

measurements, with Γ

and Γ

being diagonal weight

matrices holding the scalar sensor weight values γ

x,k

and γ

y ,k

. For instance, an RSG sensor k orientated

so that it measures strain in the x direction will take

the weight values γ

x,k

= 1 and γ

y ,k

= 0. Additionally,

virtual sensors are used to enforce boundary conditions

and are treated as RSG sensors with known signals,

typically ε = 0. These virtual sensors are added into

H at locations where the boundary condition can be

assumed to a high degree of certainty. The components

of matrix H can be developed from Equation (3):

= H





n−1

· · · x

n−1

· · · x

n−1

· · · x

n−1





(5)

A Downey et al 5

extended LSE algorithm

RSG signal

SEC signal

virtual sensing at known

boundary conditions

shape function

least

squares

estimator

(LSE)

mean

squared

error

(MSE)

estimated ε

reconstruction

error

, y)

Figure 2. Network reconstruction feature (NeRF) algorithm, the previously developed extended LSE algorithm for strain map

decomposition is enclosed inside the dashed red box.

Using Kirchhoﬀ’s plate theory, unidirectional strain

functions for ε

and ε

are obtained:

(x, y) = −

∂

w(x, y)

∂x

= Γ

(6)

(x, y) = −

∂

w(x, y)

∂y

= Γ

(7)

where c is the thickness of the plate and B = [B

]

Here, B

and B

hold the regression coeﬃcients

for strain components in the x and y directions,

respectively.

A vector S =



· · · s



contain-

ing the signal for each sensor in the HDSN is con-

structed from measurements with s

= ε

+ ε

for an

SEC and s

= ε

or s

= ε

for an RSG. The regres-

sion coeﬃcient matrix B is estimated using the LSE:

B = (H

−1

S (8)

where the hat denotes an estimation. It follows that

the estimated strain maps can be reconstructed using

= Γ

(9)

where

and

are vectors containing the estimated

strain in the x and y directions for sensors transducing

(x, y) and ε

(x, y), respectively.

Without a suﬃcient number of unidirectional

sensors in an HDSN, H will be multi-collinear because

and H

will share multiple columns. This results

in H

H being non-invertible. This is avoided by

integrating a suﬃcient number of RSGs and virtual

sensors into the HDSN.

2.3. Network Reconstruction Feature (NeRF)

The NeRF algorithm [28] provides a method for

damage detection and localization formulated for

strain map measurements. It works through comparing

the signal measured by an individual sensor with the

estimated strain map (Equation (9)) for a predeﬁned

HDSN. An error function deﬁned as the mean

square error (MSE) between a sensor’s measured

and estimated strains can be used to associate a

feature value with a given increase in the shape

function’s complexity (p in Equation (3)). Consider

an HDSN section similar to that shown in Figure

1(b), consisting of a network of SECs in an array and

a few optimally placed RSGs used at key locations.

To establish the NeRF’s theoretical foundation, we

ﬁrst consider an ideal situation where strain maps are

easily approximated through the use of low order shape

functions. The error in the approximation, calculated

for the m sensors within the HDSN, can be quantiﬁed

as:

V =

k=1

− S

)

(10)

where V is a scalar. For a given sensor location

k, S

is the sensor signal, and S

is the estimated

sensor signal using the reconstructed strain maps. The

estimated sensor signals for RSG sensors measuring

and ε

are taken from

and

, respectively,

while the estimated SEC signals are taken as the

summation of

and

at given locations (Equation

2). The NeRF algorithm is diagrammed in Figure 2,

where the extended LSE algorithm used to develop the

orthogonal strain maps is encapsulated inside the red

dashed rectangle.

For an undamaged area of a structure, the

strain ﬁeld will have a simple strain topology, while

damage will generally represent itself as a discontinuity

in the surface’s strain ﬁeld, which will develop a

more complex strain topology. It follows that in

general areas without damage, the strain ﬁeld can be

accurately estimated with low-order shape functions,

while damaged areas will require higher-order shape

functions to minimize reconstruction error. To

quantify the level of strain map complexity in a section

of the structure, and therefore whether it contains

damage, NeRF uses the section’s reconstruction error

(V ) and how this reconstruction error responds to

adding higher order shape functions. As higher-

order terms are added to the shape function, the

reconstruction error (V ) between the estimated and

measured state will substantially reduce in the case of

damaged sections, allowing the section’s condition to

A Downey et al 6

Figure 3. Experimental setup: (a) wind turbine blade model mounted in the wind tunnel and buﬀeting vanes used for generating

the turbulent airﬂow; (b) wind turbine blade showing the model’s monitored ﬁberglass substrate; and (c) DAQ used for the SEC

sensors.

be evaluated from the changing level of reconstruction

error. This technique is capable of providing damage

detection within an area monitored by an SEC-

based sensing skin even at locations that are not

directly covered by an SEC. Additionally, NeRF adds

versatility to the sensing skin for monitoring wind

turbine blades as it reduces the number and density

of required sensors and is computationally light.

Building the binomial terms used in the NeRF

algorithm, as listed in Table 1, requires starting with

w (x, y) =

i=1,j=1

as the most basic shape

function. To build the following terms of increasing

complexity, shape function components are added in

symmetric pairs from the outside of the Pascal’s

triangle, progressing inwards for a given row. For

example, the value for feature No. 1 becomes the

diﬀerence in reconstruction error, V , between the

baseline shape function w

base

(x, y) =

i=1,j=1

and the baseline shape function with term No. 1 added

(x, y) =

i=1,j=1

+ x

+ y

. Expanding

to feature No. 2, this value becomes the diﬀerence

between w

(x, y) =

i=1,j=1

+ x

+ y

and

(x, y) =

i=1,j=1

+ x

+ y

+ x

y + xy

and so forth. Note that no displacement-deﬁned

boundary conditions are enforced into the shape

functions. Instead, all boundary conditions are

enforced into strain topography through the use

of unidirectional sensors (e.g. RSG) or assumed

boundary conditions. A high level of data compression

is provided through the fusion of all the sensing

channels in the sensing skin into a single parameter,

therefore reducing the computational eﬀort required in

analyzing and storing the extracted data. This level of

compression could oﬀer a great beneﬁt to owners and

operators of wind turbine blades given their complexity

and relatively long design life of 10-30 years [8].

Table 1. Polynomial complexities used for condition assessment

features.

No. term added No. term added

1 x

, y

8 x

, x

2 x

y, xy

9 x

, y

3 x

, y

10 x

y, xy

4 x

y, xy

11 x

, y

5 x

12 x

y, xy

6 x

, y

13 x

, x

7 x

y, xy

14 x

, x

3. Methodology

This section discusses the experimental setup used in

validating the concept of the SEC-based sensing skin

and in verifying the capability of the skin to detect

damage.

3.1. Experimental setup

The SEC-based sensing skin is experimentally vali-

dated using an HDSN consisting of 12, 3 × 3 cm

SECs and 8 unidirectional RSGs, TML model # FCA-

2 deployed onto the inside of a model wind turbine

blade tested in a wind tunnel. The experimental setup,

A Downey et al 7

cut damage

(a) (b) (c)

bolts

RSG

monitored fiberglass

substrate

SEC

substrate edge

SEC

RSG

cables

damage case I

damage case II

2 cm cut

cut damage line

Figure 4. Experimental HDSN conﬁguration: (a) monitored ﬁberglass substrate with labeled bolts along the leading edge (right-

hand side) of the substrate; (b) schematic with labeled SECs and RSGs, where virtual sensors in the x and y directions are denoted

by blue circles and green diamonds, respectively; and (c) interior surface view of the HDSN (RSGs A and D are not shown, as they

were added after the substrate was installed on the model).

shown in Figure 3(a), consisted of a 139 cm wind tur-

bine blade model. It is modeled after the center third

of a 30 m wind turbine blade, designed using NREL

S-series airfoils that are aerodynamically eﬃcient with

high lift to drag ratios that generate low noise dur-

ing operation. The model was 139 cm in length with

airfoil cord lengths at the root and tip of 40 and 15

cm, respectively. Further details on the model’s design

and its experimental setup are presented in Sauder et

al. [30]. The model is mounted vertically with its root

section attached to a 6 degree-of-freedom frame that al-

lowed for the measurement of root forces. The model

(Figure 3(b)) consisted of an aluminum spar ﬁxed at

the root (blade root mounted up) and 10 wood/plastic

airfoil sections mounted onto it [30]. Sections 2 and 3,

if counted from the blade’s root, are used to support

a ﬁberglass substrate that is used in testing of the de-

ployed HDSN. This substrate, shown in Figure 3(b),

could be removed through a series of 24 bolts mounted

around its perimeter. Data acquisition (DAQ) systems

were mounted above the blade model in the mounting

frame. The SEC DAQs are shown in Figure 3(b)-(c).

Each SEC DAQ used a 24-bit capacitance-to-digital

converter multiplexed over 4 channels that sampled at

22 samples/second (S/s). An actively shielded coaxial

cable, used to remove the parasitic capacitance found

in the cables, was used to connect the SEC sensors to

the DAQs. RSG measurements were obtained using a

National Instruments 24-bit 350 Ω quarter-bridge mod-

ule (NI-9236) and sampled at 2000 S/s. Data for the

SECs and RSGs were collected simultaneously through

a LabVIEW code.

Experimental validation was carried out in the

Aerodynamic and Atmospheric Boundary Layer wind

and gust tunnel located in the Wind Simulation

and Testing Laboratory (WiST Lab) at Iowa State

University. The wind tunnel has an aerodynamic

test section of 2.44 × 1.83 m

dimensions and a

design maximum wind speed of 53 m/s. The model

blade was set at a 3-degree angle of attack and air

turbulence was induced into the tunnel by forcing a

set of buﬀeting vanes (Figure 3(a)) to oscillate at

the blade’s characteristic frequency of 3.1 Hz. This

turbulence created an almost sinusoidal buﬀeting load

(lift and moment) along the span of the blade.

The HDSN was mounted onto the inside surface of

the ﬁberglass substrate of dimensions 270 x 220 x 0.8

, shown in Figure 4(a). The deployed HDSN is

sketched in Figure 4(b) and shown in Figure 4(c). Due

to the sectioned geometry of the blade, the majority

of the bending and torsion induced strain developed in

the gap between sections 2 and 3. The 24 bolts used

to fasten the substrate onto the model were used as

boundary conditions for the extended LSE algorithm,

as annotated in Figure 4(b). The thin ﬁberglass

substrate was signiﬁcantly less stiﬀ than the aluminum

frame that formed the backbone of the model. For this

reason, the strain along the axis of the bolts is assumed

to be zero. Thus, ε

= 0 is taken at each bolt location

along the top and bottom of the monitored substrate,

and ε

= 0 is taken at each bolt location along the

vertical edges of the monitored substrate. A picture of

the HDSN before its installation onto the wind turbine

blade model is shown in Figure 4(c). In the picture,

only 4 of the 8 RSGs are shown because the remaining

4 RSGs were installed after the substrate was attached

A Downey et al 8

to the model.

Two forms of damage were induced during the

measurement campaign. Damage case I consisted of

introducing a simulated delamination in the form of

changing boundary conditions through removing the

bolts on the leading edge (facing into the wind ﬂow)

of the blade. The removed bolts are annotated in

Figure 4(a) and their order of removal for 8 diﬀerent

damage steps are listed in Table 2. The section’s

condition is expressed in terms of the length of the

longest unsupported section (damage length) of the

monitored substrate. Experimental data sets were

acquired for the healthy case (where the leading edge

had an unsupported length of 4.2 cm) and following

each damage step, resulting in a total of nine data

sets acquired. Damage case II consisted of cutting

the skin in 1 cm increments after an initial 2 cm cut

through the center of the skin along a predeﬁned path

as shown in Figure 4(a)-(b). The induced cut damage

was approximately 2 mm wide and went completely

through the ﬁberglass substrate. Data was acquired

for the healthy condition (no cut damage) and for the

12 damage steps (2 to 13 cm).

Signal interference between the SEC cables caused

by the active shielding of SEC DAQs required that

only one SEC DAQ was in operation at any given

time. Therefore, experimental data for each test was

obtained over 3 repeated test runs, each test recording

4 SECs and all eight RSGs. This superposition of data

was possible because of the constant load provided by

the buﬀeting machine, which was conﬁrmed through

the similarity of RSG data throughout the repeated

tests. Using the RSG data as a reference, the ﬁnal SEC

experimental data was aligned to provide a complete

data set of 12 SECs and 8 RSGs. To reduce sensor

noise in the SEC and provide a common time stamp to

simplify data analysis, the sensor signals were ﬁltered

as follows. A low pass Weibull ﬁlter with a cutoﬀ

frequency of 10 Hz was applied to remove any high-

frequency noise. Next, a principal component analysis

(PCA) decomposition was applied on the SEC signals

retaining the ﬁrst four eigenvalues. Lastly, the SEC

and RSG signals were resampled to 100 S/s with a

common time stamp using a spline interpolation.

3.2. Veriﬁcation of damage detection capability

The veriﬁcation of the damage detection capability

started with the investigation of the performance of the

SEC to monitor the dynamic buﬀeting-induced strain

in the wind turbine blade, that is investigated through

an analysis in the frequency domain. Thereafter,

unidirectional strain maps decomposed using the

extended LSE algorithm presented in section 2.2 are

used to track the changing load paths between a

healthy state and the fully damaged leading edge case.

Strain maps are computed from data taken when ε

at RSG B was at the maximum compressive strain

(i.e. when the tip of the model was at its maximum

displacement). An empirical damage detection method

is achieved through updating the assumed boundary

conditions and monitoring of the error between the

estimated unidirectional strain maps and the measured

strain. Here, we leverage the concept of updating the

assumed boundary conditions to detect and localize a

damage caused by the change in boundary conditions

for damage step 2. In total, ﬁve possible damage

locations were investigated in an attempt to localize

damage step 2. These attempts were the removal of

boundary conditions (bolts) 2 & 3, 3 & 4, 4 & 5,

5 & 6 and 6 & 7. Assumptions containing bolts 1

and 8 were found to be unfeasible due to the complex

interaction of the monitored substrate’s edge eﬀects

and the assumed shape function. The leading edge

damage consisting of damage step 2 (bolts 4 & 5

removed) was selected because it provided large enough

damage to be trackable with the deployed HDSN, while

still providing a relatively large search space of ﬁve

possible damage locations.

Lastly, the NeRF algorithm is used to track

the damage propagation over the entire section as a

function of the unsupported leading edge (damage case

I) and the length of the induced cut (damage case II).

For damage case I, the features developed from adding

polynomial complexities No. 5 and 7, as listed in Table

1, are used to track the growth of the unsupported

leading edge damage of the monitored section as

presented in Table 2. Thereafter, the extent of the

cut in damage case II is tracked using the features

developed from adding polynomial complexities No. 5

and 6.

4. Validation

The capability of the SEC to track the dynamic

buﬀeting-induced strain in the wind turbine blade is

shown in Figure 5. Data extracted from SEC #5

Table 2. Damage steps for boundary conditions (bolts) removed.

damage step healthy 1 2 3 4 5 6 7 8

bolts removed none 5 4,5 3,4,5 3,4,5,6 3,4,5,6,7 2,3,4,5,6,7 1,2,3,4,5,6,7 1,2,3,4,5,6,7,8

damage length (cm) 4.2 7.7 11.0 14.0 17.0 21.0 23.8 25.5 27.3

A Downey et al 9

SEC

RSG

filtered SEC

SEC

RSG

3.10 Hz

6.2 Hz

SEC maximum power

6.15 Hz

(RSG)

(SEC)

9.25 Hz

(RSG)

Figure 5. Comparison of SEC and RSG signals: frequency domain showing the excitation harmonic as detected by the SEC and

RSG; (insert) time series data for the SEC and RSG signals.

and RSG B (Figure 4(b)) are compared due to their

proximity. It can be observed that the SEC captures

the blade’s excitation frequency of 3.1 Hz and tracks

an additional excitation harmonic at 6.2 Hz. This

compares well with the excitation frequency detected

by the RSG and its additional harmonics, as denoted

in Figure 5. The excitation frequency of 3.1 Hz was

set to the blade’s fundamental frequency, as shown

by Sauder et al. [30]. Time series measurements

for the SEC and the RSG are presented in Figure

5(insert). An approximatively sinusoidal shape can

be seen in both time series, albeit the SEC exhibits

a slower sampling rate and a higher level of noise when

compared to the RSG. Individual SEC strain samples

are shown as black dots, and the ﬁltered SEC signal

is presented as the solid blue line. Overall, the SEC

demonstrates an excellent capability at tracking the

blade’s response and frequency domain components

while operating in the relatively noisy environment of

a wind tunnel. Future deployment of an SEC-based

sensing skin will require an increased precision and

sampling rate of the capacitance-to-digital converter.

The diﬀerence in the amplitude of the measured strain

between the RSG and SEC sensors is a result of their

diﬀerent locations on the substrate, the torsion present

in the substrate, and the the capability of the SEC to

measure additive instead of uni-directional strain (as

expressed in Equation 2).

Next, the performance of the HDSN at developing

full ﬁeld strain maps is experimentally validated.

Results are shown in Figure 6. The decomposed

strain maps ε

and ε

(developed using Equation

9), for the healthy case (Figure 6(a)-(b)) and the

damaged case (Figure 6(c)-(d)), demonstrate that the

HDSN is capable of tracking changes in the monitored

substrate’s strain ﬁelds. For the undamaged test’s

reconstructed strain maps, the enforced boundary

conditions ensure that ε

= 0 along the leading and

trailing edges of the monitored substrate (Figure 6(a)-

(b)). As expected, when the boundary conditions

on the leading edge are removed and the boundary

conditions in the LSE are updated to reﬂect the

monitored substrate’s change in strain, a compressive

strain energy moves into the leading edge due to

the increased bending. Changes in the substrate’s

strain ﬁeld can be related to changes in its load path.

Additionally, results demonstrate that the HDSN can

reconstruct relatively complex strain ﬁelds, such as

that caused by the torsional motion of the blade model,

represented by the diﬀerent parts of the substrate being

under tension and compression. The blade torsion

detected by the strain maps was corroborated through

accelerometers, force transducers, and video captured

during testing [30].

Results from updating the enforced boundary

conditions, as discussed in Section 3.2, to match the

damage state of the system are presented in Figure

7. Here, the error between the estimated strain

A Downey et al 10

no leading edge boundary conditions removed all leading edge boundary conditions removed

Figure 6. Reconstructed strain maps: (a) healthy condition ε

; (b) healthy condition ε

; (c) damage case 8 ε

; (d) damage case 8

maps and the experimental RSG data is measured as

a mean ﬁtting error across all 8 RSGs for the two

orthogonal strain map reconstruction cases. The mean

error is obtained by averaging the error throughout six

full vibration cycles of the model. A comparison in

the measured error between uncorrected strain maps

that maintain a constant set of boundary conditions

throughout all the damage steps and the corrected

strain maps that update the boundary conditions to

match each damage step is presented in Figure 7(a).

Results show that updating the boundary conditions

to match the damage state provides a consistently

better ﬁt than that obtained through the use of

original boundary conditions. In the case of the

damage step 8 (all the leading edge bolts removed), a

44.5% improvement in the measured error is obtained

through updating the boundary conditions to match

measurements. These results further validate the

technique of updating of boundary conditions used to

develop the strain maps presented in Figure 6. Results

presented in Figure 7(b) exhibit the ﬁtting error as a

function of the boundary conditions that are removed,

here shown for damage step 2. Boundary conditions

were removed in pairs to match the known damage

size in damage step 2 (bolts 4 and 5 removed). The

ﬁtting error for the removal of bolts 4 and 5 results in

a lower ﬁtting error, therefore identifying damage step

2 correctly. This demonstrates the capability of the

HDSN to localize damage.

Lastly, we present results obtained from the NeRF

algorithm, presented in Section 2.3, applied to damage

cases I and II. Figure 8 presents the extracted feature

distances as a function of the length of unsupported

leading edge in case I, and as a function of the length

of the induced cut in case II. Figure 8(a) shows that the

feature distance tends to decrease as the length of the

reduction in

error

damage location

Figure 7. Damage localization through updating the monitored substrate’s assumed boundary conditions; (a) improvement in

strain map reconstruction error obtained by updating boundary conditions to match the monitored substrate’s measurements; (b)

damage case 2 localized through updating the assumed boundary conditions of the monitored substrate.

A Downey et al 11

Figure 8. NeRF algorithm results for: (a) changing boundary conditions on the leading edge of the monitored substrate; (b) cut

damage induced into the center of the monitored substrate.

unsupported section of the monitored substrate along

the leading edge increases. This is to be expected as

the removal of discrete boundary conditions (bolts) will

reduce the complexity in the strain map topography,

therefore reducing the error between the estimated

strain maps and the measured strain. This reduction

in strain map complexity manifests itself as a smaller

feature distance, as computed by the NeRF algorithm.

Here, the damage case with an unsupported length

of 27.3 cm is the same damage case presented in

ﬁgure 6(c)-(d). Conversely, the NeRF algorithm

results for damage case II presented in Figure 8(b)

demonstrate that the damage induced into the center of

the monitored substrate in the form of a cut results in

the NeRF’s feature distance increasing with the length

of the cut. This can be justiﬁed by noticing that the

damage introduces a discontinuity into the monitored

substrate’s strain map. These results show that the

HDSN can accurately quantify damage.

5. Conclusion

This paper experimentally investigated the use of a

novel sensing skin for condition evaluation of a wind

turbine blade. The novel sensing skin consists of an

array of soft elastomeric capacitors (SECs), each acting

as a ﬂexible strain gauge. The critical advantage

of the sensing skin is its high scalability to its low

cost and ease of fabrication. It can, therefore, be

used to cover very large surfaces. We presented

a specialized deployment of the sensing skin, which

included a few oﬀ-the-shelf resistive strain gauges

(RSGs) to enable the precise measurement of boundary

conditions, therefore forming a hybrid dense sensor

network (HDSN). The resulting HDSN can be used

to decompose the SEC’s additive strain signal into

unidirectional strain maps based on the previously

developed extended LSE-based algorithm. These

reconstructed strain maps were used with a damage

detection algorithm termed network reconstruction

feature (NeRF), which provided damage detecting

features to detect, localize, and quantify damage.

Experimental validation was conducted by deploy-

ing the HDSN inside a scaled model wind turbine blade

excited in a wind tunnel to simulate an operational

environment. The experimental HDSN consisted of

12 SECs and 8 RSGs. Two diﬀerent damage cases

were investigated: a delamination simulated by the re-

moval of bolts, and a crack simulated by a cut. Re-

sults demonstrated that the HDSN could be used to

track the model wind turbine blade’s global condition

through analysis of SECs outputs in the frequency do-

main, which yielded similar results to the analysis of

the output data of RSGs. Both damage cases were suc-

cessfully detected and quantiﬁed through the use of the

NeRF algorithm. The delamination (bolt removal) was

tracked through an increasingly simpliﬁed strain map

with increasing damage due to the release of restraints

on the boundaries, while the crack (cut) was tracked

through an increasingly complex strain map with in-

creasing damage due to the created discontinuity in

strain. The capability of the HDSN to locate dam-

age was demonstrated with the identiﬁcation of which

bolts were removed. In the case of a crack, localization

would be achieved through proper subdivisions of the

HDSN, which was not possible with the current exper-

imental conﬁguration due to the relatively low number

of SECs. Additionally, the NeRF Algorithm was used

to provide a high level of data compression through

fusing the 20 channel HDSN into a single damage de-

A Downey et al 12

tecting feature.

Results showed the promise of the sensing

skin technology for damage detection, localization,

and quantiﬁcation in a wind turbine blade under

aerodynamic loading in a wind tunnel (i.e., operational

environment). The high level of data fusion provided

by the NeRF algorithm enhances the potential of

the sensing skin through reducing the amount of

data stored for operations. Given the demonstrated

capability of the HDSN at measuring strain maps, the

technology oﬀers potential for updating computational

models in real-time. These high ﬁdelity models could

then be used for the design of structural health

monitoring strategies and research and development

activities. Future work will include development of the

sensing skin hardware and algorithms for updating of

high ﬁdelity models using sensor data collected by a

distributed array of sensing skins.

6. Acknowledgments

The development of the SEC technology was supported

by grant No. 13-02 from the Iowa Energy Center.

This work is also partly supported by the National

Science Foundation Grant No. 1069283, which

supports the activities of the Integrative Graduate

Education and Research Traineeship (IGERT) in Wind

Energy Science, Engineering and Policy (WESEP) at

Iowa State University. Their support is gratefully

acknowledged. The authors would also like to

thank Dr. Heather Sauder and Dr. Partha

Sarkar for their support regarding wind tunnel

testing. Any opinions, ﬁndings, and conclusions or

recommendations expressed in this material are those

of the authors and do not necessarily reﬂect the views

of the National Science Foundation.

References

[1] Svetlana Afanasyeva, Jussi Saari, Martin Kalkofen, Jarmo

Partanen, and Olli Pyrh¨onen. Technical, economic and

uncertainty modelling of a wind farm project. Energy

Conversion and Management, 107:22–33, jan 2016.

[2] Wenxian Yang, Peter J. Tavner, Christopher J. Crabtree,

Y. Feng, and Y. Qiu. Wind turbine condition

monitoring: technical and commercial challenges. Wind

Energy, 17(5):673–693, aug 2012.

[3] Julia Nilsson and Lina Bertling. Maintenance management

of wind power systems using condition monitoring

systems-life cycle cost analysis for two case studies.

IEEE Transactions on Energy Conversion, 22(1):223–

229, mar 2007.

[4] J.K. Kaldellis and M. Kapsali. Shifting towards oﬀshore

wind energy—recent activity and future development.

Energy Policy, 53:136–148, feb 2013.

[5] GJW Van Bussel and MB Zaaijer. Reliability, availability

and maintenance aspects of large-scale oﬀshore wind

farms, a concepts study. Institute of marine engineers,

2003.

[6] Douglas Adams, Jonathan White, Mark Rumsey, and

Charles Farrar. Structural health monitoring of wind

turbines: method and application to a HAWT. Wind

Energy, 14(4):603–623, jan 2011.

[7] Chia Chen Ciang, Jung-Ryul Lee, and Hyung-Joon Bang.

Structural health monitoring for a wind turbine system:

a review of damage detection methods. Measurement

Science and Technology, 19(12):122001, oct 2008.

[8] J.C. Mar´ın, A. Barroso, F. Par´ıs, and J. Ca˜nas. Study of

damage and repair of blades of a 300kw wind turbine.

Energy, 33(7):1068–1083, jul 2008.

[9] Anindya Ghoshal, Mannur J. Sundaresan, Mark J. Schulz,

and P. Frank Pai. Structural health monitoring

techniques for wind turbine blades. Journal of Wind

Engineering and Industrial Aerodynamics, 85(3):309–

324, apr 2000.

[10] Yaowen Ou, Eleni N Chatzi, Vasilis K Dertimanis, and

Minas D Spiridonakos. Vibration-based experimental

damage detection of a small-scale wind turbine blade.

Structural Health Monitoring, 16(1):79–96, jan 2017.

[11] Gustavo Oliveira, Filipe Magalh˜aes,

Alvaro Cunha, and

Elsa Caetano. Development and implementation of

a continuous dynamic monitoring system in a wind

turbine. Journal of Civil Structural Health Monitoring,

6(3):343–353, jun 2016.

[12] Yi Zou, LPSG Tong, and Grant P Steven. Vibration-based

model-dependent damage (delamination) identiﬁcation

and health monitoring for composite structuresa review.

Journal of Sound and Vibration, 230(2):357–378, feb

2000.

[13] Inpil Kang, Mark J Schulz, Jay H Kim, Vesselin Shanov,

and Donglu Shi. A carbon nanotube strain sensor

for structural health monitoring. Smart Materials and

Structures, 15(3):737–748, apr 2006.

[14] Hyung-Kew Lee, Sun-Il Chang, and Euisik Yoon. A

ﬂexible polymer tactile sensor: Fabrication and modular

expandability for large area deployment. Journal of

Microelectromechanical Systems, 15(6):1681–1686, dec

2006.

[15] T.A. Chase and R.C. Luo. A thin-ﬁlm ﬂexible capacitive

tactile normal shear force array sensor. In Proceedings

of IECON 1995 - 21st Annual Conference on IEEE

Industrial Electronics. IEEE, 1995.

[16] Jonathan Engel, Jack Chen, and Chang Liu. Development

of polyimide ﬂexible tactile sensor skin. Journal of

Micromechanics and Microengineering, 13(3):359–366,

feb 2003.

[17] Mark J Schulz and Mannur J Sundaresan. Smart Sensor

System for Structural Condition Monitoring of Wind

Turbines: May 30, 2002-April 30, 2006. National

Renewable Energy Laboratory, 2006.

[18] Yao Yao and Branko Glisic. Detection of steel fatigue cracks

with strain sensing sheets based on large area electronics.

Sensors, 15(4):8088–8108, apr 2015.

[19] AR Burton, JP Lynch, M Kurata, and KH Law. Fully

integrated carbon nanotube composite thin ﬁlm strain

sensors on ﬂexible substrates for structural health

monitoring. Smart Materials and Structures, 26(9),

2017.

[20] Gangbing Song, Hui Li, Bosko Gajic, Wensong Zhou,

Peng Chen, and Haichang Gu. Wind turbine blade

health monitoring with piezoceramic-based wireless

sensor network. International Journal of Smart and

Nano Materials, 4(3):150–166, sep 2013.

[21] Donghyeon Ryu and Kenneth J Loh. Strain sensing

using photocurrent generated by photoactive p3ht-

based nanocomposites. Smart Materials and Structures,

21(6):065016, may 2012.

[22] Mark A. Rumsey and Joshua A. Paquette. Structural

health monitoring of wind turbine blades. In Wolfgang

A Downey et al 13

Ecke, Kara J. Peters, and Norbert G. Meyendorf, editors,

Smart Sensor Phenomena, Technology, Networks, and

Systems 2008. SPIE-Intl Soc Optical Eng, mar 2008.

[23] Simon Laﬂamme, Matthias Kollosche, Jerome J. Connor,

and Guggi Kofod. Robust ﬂexible capacitive surface

sensor for structural health monitoring applications.

Journal of Engineering Mechanics, 139(7):879–885, jul

2013.

[24] Simon Laﬂamme, Hussam S. Saleem, Bharath K. Vasan,

Randall L. Geiger, Degang Chen, Michael R. Kessler,

and Krishna Rajan. Soft elastomeric capacitor network

for strain sensing over large surfaces. IEEE/ASME

Transactions on Mechatronics, 18(6):1647–1654, dec

2013.

[25] Simon Laﬂamme, Filippo Ubertini, Hussam Saleem, An-

tonella D’Alessandro, Austin Downey, Halil Ceylan, and

Annibale Luigi Materazzi. Dynamic characterization of a

soft elastomeric capacitor for structural health monitor-

ing. Journal of Structural Engineering, 141(8):04014186,

aug 2015.

[26] Simon Laﬂamme, Liang Cao, Eleni Chatzi, and Filippo

Ubertini. Damage detection and localization from dense

network of strain sensors. Shock and Vibration, 2016:1–

13, 2016.

[27] Austin Downey, Simon Laﬂamme, and Filippo Ubertini.

Reconstruction of in-plane strain maps using hybrid

dense sensor network composed of sensing skin. Mea-

surement Science and Technology, 27(12):124016, nov

2016.

[28] Austin Downey, Filippo Ubertini, and Simon Laﬂamme.

Algorithm for damage detection in wind turbine blades

using a hybrid dense sensor network with feature level

data fusion. Journal of Wind Engineering and Industrial

Aerodynamics, 168:288–296, sep 2017.

[29] Austin Downey, Chao Hu, and Simon Laﬂamme. Opti-

mal sensor placement within a hybrid dense sensor net-

work using an adaptive genetic algorithm with learn-

ing gene pool. Structural Health Monitoring, page

147592171770253, apr 2017.

[30] Heather Scot Sauder and Partha P. Sarkar. Real-time

prediction of aeroelastic loads of wind turbine blades in

gusty and turbulent wind using an improved load model.

Engineering Structures, 147:103–113, sep 2017.