1 Introduction
FUNCTIONAL ELECTRICAL stimulation (FES) systems for
restoring locomotion in paraplegic patients still have many
basic problems preventing their everyday use
(SOLOMONOW, 1984a). In general the complexity of indu-
cing motions by FES, whether in upper or lower extrem-
ities, requires control strategies at levels ranging from
muscle force regulation
(CRAGO
et
al.,
1980) and joint posi-
tion control
(ALLIN
and
INBAR,
1986) up to the co-
ordination of synergistic and multi-joint muscle
contractions
(ALLIN
and
INBAR,
1986; SOLOMONOW, 1984b).
A typical example of low-level control is the control of
knee angle during standing: guaranteeing reliable locking
of the knee (during quiet standing or in the stand-phase
during walking), without rapidly inducing local muscle
fatigue. Current clinical FES systems are based mainly on
open-loop control of stimulation
(PECKHAM,
1987; JAEGER
et al.,
1989). Locking of the knee is realised by continuous
supramaximal stimulation of knee extensor muscles.
Although providing safe knee locking this method easily
results in ischaemia due to the maintained fused contrac-
tion
(MERTON,
1954; MORTIMER
et al.,
1971) and therefore
induces early quadriceps fatigue and limited standtime.
Minimisation of quadriceps activation through feedback of
kinematical information from the lower limb to the stimu-
lation equipment potentially reduces these problems.
Closed-loop feedback control has been studied exten-
sively to achieve and maintain a desired joint angle or
muscle force
(STANIC
and
TRNKOCZY,
1974; PETROFSKY
and PHILLIPS, 1979; CRAGO, 1983; JAEGER, 1986; WILHERE
et al.,
1985). However, for application in the control of
standing (especially when using surface stimulation) poor
recruitment stability of the stimulated quadriceps muscle is
a problem (TRNKOCZY, 1974). Besides, the mechanical lock
state in the knee joint together with the absence of direct
muscle force information hinders a proper control of con-
traction (see also Section 4: Discussion). PETROFSKY
et al.
(1984) report on one of the few studies concerning closed-
loop multijoint control for standing and walking in para-
plegics. The system was tested in paraplegic subjects and
angle sensor stability is reported to be crucial.
In the current study a control strategy is proposed to
achieve decreased quadriceps force during standing, using
non-numerical or finite-state control. Modified on/off (or
ramp-down) control is used based on feedback of knee
angle and angular velocity. A procedure for repeated cali-
bration of the locked position is incorporated. The control
strategy was applied in paraplegic patients in an experi-
mental setup, in which external parameters (leg load and
movement) could be kept constant. The goal of the experi-
ments was to explore the control scheme for its ability in
stabilising the knee joint by investigating the systems
dynamics and its sensitivity to important control param-
eters.
2 Methods
Correspondence should be addressed to Professor Mulder at the
Roessingh Rehabilitation Centre.
First received 2nd May and in final form 30th October 1989
~) IFMBE: 1990
2.1
Control scheme
Fig. 1 shows the state diagram of the knee joint used in
the control algorithm. Two physical states are considered:
knee-lock and knee-unlock.
Medical & Biological Engineering & Computing September 1990 483
angular velocity
maximum
I
stimulation
L
lower
stimulus
amplitude
knee angle
Fig. 1 State diagram of the control scheme
Fig. 2
nr~ ir~m~ f mr
The stand simulator. The patient is in supine position with
his feet on a moveable carrier
2.1.1 The knee-lock state: The lock state is defined as
the individually and mechanically determined joint end
point of the knee. This includes hyperextended positions or
end points determined by protecting braces. When the
knee joint is locked muscle force cannot be controlled
using feedback of knee angle information. Therefore
muscle activation is minimised by controlling stimulus
pulse amplitude open loop, with the interpulse interval set
at a fixed maximum to avoid fatigue. Amplitude and force
are supposed to have positive correlation. Linear ramp-
down of pulse amplitude (with time-derivative R mA s- 1) is
used to lower muscle activation. As a result, the muscle
force will drop and the knee joint may finally bend. The
dynamics of this bending process depend on external
factors: the degree of (hyper-) extension in the knee joint,
extending or flexing knee moments resulting from the
upper body, passive elasticity and damping of hamstring
or iliopsoas muscles, etc. We call this bending 'breakdown'.
The stimulation amplitude at which it occurs is the break-
down amplitude. Ramp-down always starts at a defined
step level (increment step S mA) above the previous break-
down amplitude. In this way the average stimulus ampli-
tude adapts to changing breakdown amplitudes due to
load and position changes of the leg or to muscle fatigue.
2.1.2 The knee-unlock state: When the knee joint
comes out of the locked position, further bending must be
stopped and the knee joint must be returned to the locked
position. From different possible strategies, we have
chosen to simply switch stimulation to a high level
(MULDER et al. 1988).
2.1.3 State transitions: An accurate detection of tran-
sitions from lock to unlock and vice versa is important for
good controller performance. To signal the transition from
lock to unlock we use knee angle data obtained from a
goniometer. Knee flexion over a defined minimum (the
angular deadband A ~ will switch stimulation on.
To determine the transition from unlock to lock,
angular velocity data are used. Using knee angle data
would make the system highly sensitive to goniometer drift.
Additionally, using angular velocity data allows automatic
calibration of the locked position after each unlock. Lock
will be indicated when angular velocity equals zero for a
minimum time interval (Tms) to eliminate errors due to
sign transition. A velocity deadband (V ~ s -1) is used to
eliminate possible velocity noise.
2.2 Experimental set up
2.2.1 The stand simulator: To test the control strategy
in paraplegic patients under reproducible conditions, we
devised a special stand simulator (Fig. 2). The patient is in
the supine position on a bench with one foot on a horizon-
484
tally moveable carrier. A mass under the bench provides
an artificial inertial leg load. Leg load and movements are
reproducible in this way. It can be shown that for small
movements the dynamics of this system equal that during
real standing. A force transducer under the knee joint both
serves as an artificial kneelock (to obtain an overall flexing
knee moment) and measures the locking force for evalu-
ation. The force transducer consists of a bridge configu-
ration of four strain gauges on a thin steel plate connected
to a bridge amplifier.
2.2.2 Subjects: The experiments were carried out on
two complete T5-T6 level spinal cord injured patients. A
surface cathode and anode (carbon rubber, 4 x 7 cm)
placed over the rectus femoris motor point and near the
patella transcutaneously stimulated the right leg quadri-
ceps mucle. The patients had normal excitability of quadri-
ceps muscles and were well trained during at least a 6
month FES exercise programme. The programme consist-
ed of weight lifting twice a day (30 min, ankle load up to
5 kg) and low-load exercise cycling twice a week (30-60
min). Owing to interpatient differences in training per-
formance and amount of spasticity the muscle contractile
properties may differ between subjects.
2.2.3 Control of stimulation: To control stimulation we
used an IBM-XT computer with AD facilities (Analog
Devices, 12-bit), an electrogoniometer to measure knee
angle (MCB pp27c, 310 ~ nonlinearity 1 per cent), and a
high output impedance current stimulator (monophasic,
rectangular current pulses). The pulse parameters could be
controlled by the computer. Pulse duration was fixed at
300 #s, pulse rate was fixed at the minimum frequency for a
fused contraction: 20Hz. Pulse amplitude was used as
control parameter. The bandwidth of the gonio-signal was
found to be 10 Hz. To determine the knee angle, the gonio-
signal was sampled at 100 Hz and low-pass filtered with a
digital first-order filter at 15 Hz. Angular velocity was cal-
culated digitally from the knee angle intersample difference
and smoothed by a digital third-order low-pass filter at
15Hz.
2.2.4 The experiments: During the experiments knee
angle was locked at 30 ~ This was chosen to create a high
joint load situation to evaluate system dynamics under
what were considered the 'worst-case' circumstances to
occur during real standing. The external leg load (20 kg)
equalled half the trunkload of an average subject. One
experimental session lasted about 60 min and consisted of
a series of about 10 stimulation experiments of 30s each.
Each session started with a short warming-up of the quad-
riceps muscle. During this phase the maximum stimulation
level, defined as the value sufficient to stretch the leg
against the external leg load, was determined. Each stimu-
lation experiment was followed by at least 5 min rest.
Medical & Biological Engineering & Computing September 1990
During the experiments knee angle data, angular velocity
data, stimulation amplitude and locking force data were
recorded and stored on disk for offiine evaluation.
The most important parameters expected to affect the
performance of the controller were: negative slope of the
stimulus ampitude ramp-down (R mA s-'), minimum time
interval after which the knee joint is declared stable after
switch-on of the stimulation (Tms), amplitude increment
step (SmA), angular deadband (A ~ and angular velocity
deadband (V~ To evaluate the controller, the influ-
ence of these parameters on knee angle stability and
extending quadriceps force during knee-lock, as well as the
response of the knee angle to a positive and negative step
change in the leg load, were determined.
Therefore average maximum knee flexion during break-
down
(KF),
average maximum locking velocity of the knee
joint
(LV)
and average knee extension force
(EF)
were cal-
culated from the measurements. This was done after a
session was finished.
KF
and
LV
were calculated over the
full 30 s of an experiment.
EF
was calculated over a 3-10 s
interval to eliminate artefacts from start-up or load
changes.
3 Results
Fig. 3 shows a typical result of a stimulation experiment.
The knee joint continuously switches between the lock and
unlock state. After each unlock stimulation is switched on,
and after the locked position is regained it is recalibrated
and stimulation ramps down again. At t = 11 s the leg load
was increased with 5 kg. The extra load was removed at
160
150
o
.
140
@
c
0
c
<
E
o
80 b
70 ~
0
6o 2
lhO 2'0
'
30
time, s
Fig. 3
Typical example of a stimulation experiment. Upper curve:
knee angle; lower curve: stimulus amplitude. Essential
control parameters: R = 5 mA s- l, T = 200 ms,
S = lOmA, A = 1.8 ~ V = 13.2 ~
t = 24 s. The stimulator responds by instantaneously ter-
minating ramp-down and switching to full stimulation
level. After that, average stimulation is at an increased
level. The first breakdown at 4 s is due to goniometer dis-
turbances. The locked position is recalibrated and the
system remains stable.
3.1
Knee angle
Fig. 4 shows in more detail one stimulus cycle for each
of the two patients studied. After stimulation is switched to
a high level flexing is slowed down. Then the knee returns
to the locked position. The new constant knee angle does
not necessarily equal the angle before bending (Fig. 4b).
This may be due to movements of the goniometer frame,
the hip joint not completely being fixed, or friction in the
mechanical parts of the stand simulator. This emphaslses
the need for repeated calibration of the locked position.
During the experiments average knee flexion during one
stimulation cycle depends on the parameter setting of the
controller and the mechanical parameters of the leg. Figs.
4a and 4b show characteristic angle responses for the two
patients at equal controller setting. Patient 2 exhibits more
Medical & Biological Engineering & Computing
knee bending than patient 1. Obviously, mechanical or
neural response including muscle dynamics are signifi-
cantly different for both patients. During the tests the
average maximum breakdown knee flexion was 3.6 ~
(standard deviation 1.6 ~ for patient 1 and 4.8 ~ (standard
deviation 1.2 ~ ) for patient 2. The time for one stimulus
cycle (knee flexion-extension) ranged from 200 to 400 ms.
a
160
o
J
c
0
@
c
145
150
,, ,
, ',
i
i LI-
'-,_
l
1--11
U
80
g 60
~ 4o
o 2O
>
0
-5 -2O
g -4O
-60
L
: '-l-t_ '
b
...... ]
i
t
I
I LI_
i
d
i ..... "1
t t
i i
i i
i i
8O
75 <
E
7O ~-
65
60 ~
E
55 o
5o
-~
if)
40
Fig. 4
, i
ls
One detailed stimulus cycle for two patients. Solid line:
knee angle (a, b) and angular velocity (c, d) for patient
1 (a, c) and patient 2 (b, d). Dotted: stimulus amplitude.
R, S, A and V: see Fig. 3, T = 400ms (a, c) or T = l s
(b, d)
Angular deadband A must be adjusted to cover the
remaining signal noise due to insufficient filtering of elec-
trical or mechanical disturbances. If this is not performed,
repeated false detection of knee flexion will occur and
stimulation will be continuously at high level. On the other
hand, angular deadband must be minimised to minimise
average knee angle flexion. During the experiments the
remaining angle noise was 0.5 ~ peak-to-peak.
3.2
Angular velocity
During the experiments the average locking velocity was
42 ~ s -1 (standard deviation 15~ -t) for patient 1 and
54~ -1 (standard deviation 15~ -1) for patient 2. Differ-
ences between patients may be caused by different muscle
force levels, or muscle dynamics (Figs. 4c and 4d).
To achieve accurate detection of knee lock both adjust-
ment of angular velocity deadband V and minimum stimu-
lus on-time T is important. During unlock angular
velocity may not exceed the level of noise (Fig. 4c). The use
of T has therefore proven to be useful to avoid false detec-
tion of knee lock due to low unlock velocities. On the
other hand T must be minimal to avoid stimulation at
high level longer than required for locking. Optimally T
equals the maximum time required to reverse the knee
movement from flexion to extension. From velocity data it
was concluded that values of T under lOOms make the
detection procedure critical.
V must be set to cover the remaining velocity noise after
filtering, which was 15 ~ s-t peak-to-peak. Although V can
be reduced when the order of the velocity filter is
September 1990 485
increased, this is not recommendable. It would increase the
time delay between angle and velocity data as well, which
would increase the minimum T which is required, thus
deteriorating the dynamic detection of knee lock.
3.3 Stimulus amplitude
The maximum stimulus amplitude was found to be not
critical with respect to stabilising the knee joint, although
it determines the ultimate level of leg load or allowable
muscle force decrease.
When the leg load increases or the muscle force
decreases the amplitude increment step S increases the
average level of stimulation. When the load increase is
high, repeated switching of the stimulation (as demon-
strated in Fig. 3) may occur until the mean stimulus level is
sufficiently increased. The increment step size S is deter-
mined by the desired maximum number of stimulus cycles
which are needed to increase the muscle force from
minimum to maximum.
For a given leg load, the average interval between
'switch-on stimulation' events is dependent on the stimulus
increment step S, the stimulus amplitude ramp-down R
and the time needed to stabilise the knee after break-down.
This time is dependent on the dynamics of leg and muscles
and was found to be 200-400ms during the experiments.
This means that R and S determine the contraction/
relaxation ratio of the quadriceps muscle. Besides, R and S
determine the response of the muscle to changes in leg
load. Therefore, optimum values depend on both the
requirements for blood support to the muscle and the
patient's individual dynamic load characteristics during
actual standing.
During the experiment of Fig. 5 early fatigue of the
quadriceps was forced using a stimulus frequency of 60 Hz,
and a stimulus on-time T = 1 s. The average stimulus level
is quickly adapted.
7~ f
65
"~ 6o
~ 45
~ I
"~ 40 ~
1
]
I
1'o
i'5
2'o 2%
time. s
Fig. 5. Response of stimulus amplitude to forced fatigue. Refer to
the text for details
The dynamic detection of knee lock from angular veloc-
ity data works satisfactorily when the minimum stimulus
on-time T does not exceed the 200-400 ms needed to
stabilise the knee after breakdown (Fig. 6). When T
exceeds this value actual stimulus on-time equals T in
most cases, so muscle activation is not optimal.
3.4 Locking force
From several experiments nonlinear and nonstationary
recruitment could be seen. Fig. 7a shows intrapatient
variability in muscle gain during linear amplitude ramp-
I q
'-; '-i' BO
:, 1,
: ',
, , , , ,
80 .....
75
, , t J
,
60
~_
,
~_
, 70
, ,- _ ', -,_ ,, ~-~_
o
I '-- ,
'-,- "- ~=
, ,-, ,I ,-, ,I
~-~_ , -,., : -,.~ , ?Z
I '
b
-40
'5
E
g -60
-80 ,
, i
40 6'0 8"0 I0"0
time, s
Fig. 6 Dynamic stimulus on-time (dotted line) resulting from lock
detection by angular velocity (solid line)
down. An explanation may be the movement of the muscle
under the electrode. Fig. 7b shows how both subjects
exhibit different force build-up after switch-on of stimu-
lation. This may be due to interpatient differences in
muscle and leg responses. Obviously during the recordings
of Fig. 7b the minimum on-time T was high with respect
to minimising muscle activity: the locking force stabilises
at a constant level.
I00
90
80
I
7oi
Z
60
2
50
C~
c
40
o 3o
2o
I0
0
Fig. 7
Io0
90
8O
70
Z
6O
5O
40
3O
2O
10
0
70
,_,_ 65 <
E
'-'-'-~ 60 ~"
--I
13
-I- I
-'- ,_ 55
I-1_ r
I--I--i_
50 E
I I--I .~
2
m
4.0 5-0 6.0
hme. s
pat,ent 1 patient 2
ii i' I-- l
I I i
t
I i_ :
I
l
,,
"1 I
- - i -L
I t_
i I-I_
=
I
-t t
i i
ls
b
Locking force for evaluation of muscle response and system
behaviour. (a) Nonlinear recruitment: (b) different force
build-up for both patients
486 Medical & Biological Engineering & Computing September 1990
3.5
Sensitivity
After the exploratory sessions, three sessions of 10
experiments were performed to evaluate the influence of
the control parameters in more detail. Specifically, the
sensitivity of the system's output parameters
KF, LV
and
EF
to changes in the control parameters was investigated.
The experiments were matched to obtain intraindividual
sets of experiments performed under equal environmental
conditions. Within one set only one control parameter dif-
fered (Table 1). The output parameters
KF, LV
and
EF
were calculated for each experiment. Using the simple sign
test (p = 0.5), significance of increase or decrease with
respect to changes in the control parameters was deter-
mined. The results are shown in Table 2. The effect of S
could not be determined due to an insufficient number of
valid experiments.
Table 1 Parameters used in the sensitivity analysis. Refer to the
text for details
Parameter Test values Number of matched experiments
A, ~ 0-9, 1.8, 3'5 2, 2, 3, 3
R, mA s-1 5, 10, 20, 40 2, 2, 3, 4
S, mA 5, 10, 20 3
V, ~ s -1 7.8, 13.2 2, 2, 2, 2
T, ms 100, 200, 400 2, 3, 3, 4, 4
600, 1000
Table 2 Correlation between control and output parameters.
Correlation can be positive (+), negative (-) or non-significant
(0). (ct: level of significance)
Parameter
K F L V E F
A, o + (ct = 0.05) + (~ = 0-125) 0
R, mA s- 1 + (~ = 0.01) 0 - (~ = 0.05)
V,~ -t 0 0 0
T, ms 0 0 + (~ = 0.05)
When controller performance is defined to be optimal
for minimum
KF, LV
and
EF,
Table 2 shows that A and
T must be minimised (positive correlation). R shows both
positive
(KF)
and negative
(EF)
correlation. Therefore its
optimal value depends on the relative weight given to
KF
and
EF.
Besides optimal R depends on the required system
response to dynamic load changes as stated in Section 3.3.
4 Discussion
Important in paraplegic standing and walking is
counteracting gravity without undue stressing of the upper
body. To improve on the effect of walking with long leg
braces stabilising the knee joints additionally requires sta-
bilisation of the hip
(KRALJ
et
al.,
1980; ISAKOV
et al.,
1986). For keeping balance crutches or a walking frame
may be necessary.
Traditionally paraplegic standing requires mechanical
bracing (ROSMAN and
SPIRA,
1974). In conjunction with
FES bracing may still be useful, especially in stabilising the
hip
(PETROFSKY
et al.,
1985; McCLELLAND
et al.,
1987) as
well as for the protection of joints. Considering knee sta-
bilisation, besides being a passive solution mechanical
bracing usually leads to walking with locked knees
requiring increased ann push-up during swing phase. Knee
stabilisation by FES allows walking with unlocked knees
giving more natural gait and limiting upper body loading.
ANDREWS
et al.
(1989) apply finite state control in a
so-called hybrid orthosis, combining low-leg bracing and
stimulation of quadriceps to lock the knee. Stimulation is
on or off depending on force signals derived from the
Medical & Biological Engineering & Computing
brace. The system shows similarity to the work of KRALJ
et al.
(1986) who describe a system which switches between
quadriceps stimulation and soleus stimulation to lock the
knee. This requires feedback of knee and ankle angle. Both
methods offer the opportunity to release the quadriceps
temporarily by taking a specific posture. The method pro-
posed in the present study basically has the advantage to
reduce average quadriceps force even if the knee is not
hyperextended or otherwise locked from ground reaction
force. It is therefore independent of mechanical bracing or
posture. However, when needed (e.g. in patients with
hypermobile knee joints) external bracing may be added to
protect the knee joint or prevent it from hyperextensing,
without affecting the control strategy.
From a model study previously carried out
(MULDER
et
al.,
1987; 1988) we concluded that theoretically a pro-
portional derivative (PD) controller is able to stabilise the
knee joint during standing. However, it was concluded
that the gain robustness of continuous (linear) feedback
control is insufficient to counteract the poor recruitment
stability of the quadriceps muscle when using surface
stimulation
(TRNKOCZY,
1974). Although adaptive
techniques may improve the performance of such a system
(BERNOTAS
et
al.,
1987), its application is still limited to
slowly changing parameters. Besides, continuous feedback
would be highly sensitive to setpoint drifting as this set-
point coincides with the joint end point of the knee, being
a highly nonlinear system element. The present approach
incorporating finite state control has been shown to be
robust with respect to these problems and may be useful
when reliability of system response is more important than
accuracy of (position) control. It therefore is considered
appropriate for joint locking during standing:
The presence of a limit cycle (i.e. a system oscillation
resulting from discontinuity in the control function)
enables repeated calibration of the locked position from
angular velocity data. This makes the requirements for
stability of angle measurement low. Our results indicate
that even large goniometer disturbances do not deteriorate
overall system stability, i.e. the dynamic system response is
not affected, the knee joint does not collapse and muscle
force continues to be minimised.
During the experiments average breakdown knee flexion
rarely exceeded 8 ~ which for an average person would
correspond to a vertical hip movement of 2mm during
standing. In real standing the load situation may be more
favourable than the experimental situation. Therefore gen-
erated knee jerks are expected to be within an acceptable
range of motion, and the prospects for clinical applicability
are good.
In the present study system dynamics were evaluated
under high load conditions compared with during stand-
ing. However, the proposed stimulation strategy clearly
reduces the average muscle force and offers a dynamic
muscle activation which is expected to improve muscle
blood flow over using high level stimulation continuously.
Although in real standing the relative force reduction will
depend on the individual knee load situation, it will be
better than the experimental situation and fatigue is
expected to be reduced significantly
(MARSOLAIS
et
al.,
1988). First results with the artificial reflex controller being
used in real paraplegic standing have shown a standtime
increase of 3.5 times compared with constant supra-
maximal stimulation
(MULDER
et
al.,
1989). A special situ-
ation occurs during longer periods of no flexing knee
moment: the level of stimulation will decrease to zero, and
no muscle fatigue will be induced. The stimulator then acts
as a safety device preventing the knee from inadvertently
buckling. This 'no flexing' situation will certainly occur e.g.
September 1990 487