This was the title of the paper that was published in June 2010, which was about
interpretation of ground movements due to tunnelling for the Stoke Newington to New
River Head London Ring Main extension. The tender was originally won by Amec, but
soon after the job started Amec’s tunnelling business was acquired by Morgan Est. For
most of the project I was the Engineering Manager for Morgan Est, who were later
renamed Morgan Sindall in line with the parent company’s name.
Here you can see the story of the tunnel drive – delivery of the TBM to the Stoke
Newington launch shaft site (you can see the old Victorian pumping house in the
background cunningly disguised as a castle, now an indoor climbing centre – those
Victorians loved a bit of pastiche), lifting it down the shaft, the segmental tunnel lining,
and lifting the machine out at the other end.
1
Click through and read.
2
This plan shows, somewhat schematically, the whole of the ring main. The River Thames
flows from West to East, and most of London is served by the water treatment works in
the West – Ashford Common, Kempton Park, Walton and Hampton. The North-East of
London is served by Coppermills water treatment works on the River Lee, a tributary of
the Thames. The new tunnel from Stoke Newington in Hackney to New River Head in
Islington (near Sadlers Wells and Clerkenwell) would provide resilience by allowing flow
in either direction in case of a breakdown of a water treatment works. Another tunnel
was built from Honor Oak to Brixton, also by Morgan Est, at the same time.
3
At the top you can see a map showing the alignment and the location of boreholes. The
direction of the tunnel drive is from right to left. The tunnel starts near Green Lanes
between the Castle and the West Reservoir, passes under Green Lanes and Brownswood
Road to meet Highbury Grove, then cuts across to go under Highbury Fields, where it
passed under the Canonbury rail tunnel. Then it passed under the CTRL near Corsica
Street and then under the Victoria Line and Northern & City Line tunnels and the North
London Line surface railway at Highbury & Islington Station. The tunnel then went down
the entire length of Upper Street, crossing the Victoria Line again, the 200 year old
Islington tunnel on the Regents Canal, then passing the Northern Line running tunnels
just South of Angel Station before arriving at the reception shaft in a Thames Water
compound just off Amwell Street.
In all 13 other tunnels were underpassed by Cleo. Since Cleo was between 40 and 60m
below the ground surface, many of these tunnels were more than 20m above, and
therefore the potential movements were small compared to the allowable movements.
However, the CTRL (or HS1) tunnels and the Northern Line tunnels were much closer to
the new tunnel, and there were also good reasons why even small movements were not
desirable. For instance if the HS1 track settled 6mm, speed restrictions would be
imposed. There were concerns about settlement reducing the already very tight
clearance envelope for the trains in the Northern Line tunnels or causing problems for
the escalator in Angel Station, which is the longest in Western Europe. In order to predict
movements of these tunnels, it was important to try to calculate trough width and
volume loss from the surface settlements earlier in the tunnel drive.
4
The geological sequence shown on this slide is what was actually encountered during
construction, based on the work of Tim Newman and others, who published a paper in
the same issue of Geotechnical Engineering. Most of the tunnel was constructed in the
Upnor Formation and Thanet Sand, and was therefore perfect for EPB tunnelling. During
the last part of the drive through the very stiff mottled clays of the Lambeth Group,
mucking times were doubled from around 10mins per metre to around 20mins per
metre.
Very little groundwater pressure was encountered – the tunnel horizon being
underdrained by historical and current abstraction from the Chalk with the Reading clays
and London Clay above acting as an aquiclude. Water was encountered briefly at the two
fault zones where there may have been connectivity with the upper aquifer. There was
also a sand channel in the Upper Mottled Clay encountered during shaft construction at
New River Head, which caused some minor problems.
4
This was the TBM, called Cleo. A Lovat 131SE Earth Pressure Balance machine. A small
machine compared to some of the giants being built today, and for a 4.5km drive you
probably couldn’t go much smaller. The total weight of the forward and stationary shells
shown here was 80 tonnes. The excavated diameter was 3.362m, and within the tailskin
a 2.85m ID, 180mm thick steel fibre reinforced precast concrete bolted segmental lining
was erected. At the end of the tunnel drive the tunnel was given a reinforced concrete
secondary lining to take the internal water pressures.
5
Click through and read.
6
Road nails were installed in the road surface or pavement as single points over the
tunnel centreline, or as transverse arrays of points.
Surface settlements were measured using a Leica DNA03 precise laser level with a barcoded Invar staff. Over the length of a levelling loop for a typical surface settlement
monitoring array, this should result in a repeatability of less than ± 0.1mm under
controlled conditions.
However, a range of other factors such as ambient temperature, heavy traffic, the sun
heating the road or pavement and near-surface pore pressure changes due to rainfall or
tree root suctions mean that the repeatability in the field will always be worse than this.
These factors result in background movements, some of which will affect each
monitoring point randomly and independently, and some of which will affect an array of
monitoring points in a similar manner.
7
This figure shows the background readings, i.e. all the readings made before the TBM
was within 50m of any monitoring point, and how they vary from the baseline. If the
measurement system was perfect, there would just be one bar over the zero change
from baseline point.
As we said before, there are two types of error: those that affect single points randomly
and independently, and those that affect a single levelling run of an array of points in a
similar manner (for example misreading of a benchmark – a systematic error, or
movement of a benchmark – an environmental effect). We want the errors to be random
and independent, because then they will tend to cancel each other out during the curvefitting process.
Although the fact that a normal distribution fits the background readings quite well (with
an r2 value of 0.94) seems to imply that these errors are random and independent,
because the normal distribution models random and independent variables, this is not
actually the case because by lumping all the levelling runs together we are hiding
systematic errors and environmental effects that affect a single levelling run in a nonrandom manner.
In order for the variations in the data shown in this histogram to be random and
independent, the mean of the changes from the baseline for each individual levelling run
should be close to zero. In other words, if the errors are random and independent they
should tend to cancel out across an array of points in a single levelling run. In general
this was the case, with the mean of the absolute values of the mean changes equal to
8
0.08 mm, with a standard deviation of 0.05 mm, for all the levelling runs that were within
one week of the baseline readings. However, as more time elapses between the baseline
readings and subsequent readings, the background movements could in some cases grow
larger and the mean absolute changes could be up to 0.5 mm for a single levelling run.
This indicates that errors that are not random and independent tend to grow gradually
over time, and by ensuring that baseline readings are taken immediately prior to the TBM
entering the zone of influence, the errors in the short-term settlement will tend to be
mainly random and independent, and hence the accuracy of the Gaussian curve-fitting
will be much improved.
8
Since in most cases the TBM passed through the zone of influence of an array within 1
week (and in all cases where Gaussian curve parameters were calculated), we can assess
the repeatability of the measurement method by looking only at the background
readings taken within 1 week of a baseline reading. So, the standard deviation of
background readings 1 week from baseline was 0.131mm.
The repeatability is therefore ± 0.26 mm, based on a 95% confidence level.
9
Click through and read.
10
This is a Gaussian settlement trough. Smax is the maximum settlement over the
centreline of the tunnel, y is the lateral offset from the centreline and i is the offset to
the point of inflexion (sometimes called the ‘trough width’.
The settlement is given by this equation [click]. The area under the curve is the volume
loss and this is given by the following equation [click]. This is sometimes expressed as a
percentage of the excavated face area, in which case it would usually be capital-Vsubscript-L. Therefore, the curve is defined by any two of the parameters Smax, i, or Vs.
This means that curve-fitting to real data is not straightforward, and there are several
ways it could be done.
11
Since this paper was published in 2010, I have done a lot of work to try to quantify the
reliability of Gaussian curve-fitting by using Monte Carlo analysis. This will hopefully be
published soon.
There are several ways that a Gaussian curve can be fit to real settlement monitoring
data, and I have listed them here in ascending order of complexity. Fitting a curve by eye
is too subjective when errors are of a similar order of magnitude to the settlements
themselves and won’t be considered further.
Trapezoidal integration can be used to directly calculate volume loss, and this is the most
reliable method. Random independent errors will tend to cancel each other out, so the
reliability is improved if there are more points in an array.
Once trapezoidal integration has been used to calculate volume loss, then a value of i
can be found to fit through any two points. This could be done for several pairs of points
in an array and then the results averaged perhaps.
The method described in the paper used trapezoidal integration to calculate volume loss,
then a direct calculation of the distance to the point of inflexion was done by using the
analogy to the normal distribution standard deviation. Essentially the settlements were
treated as frequencies on a histogram.
The most complicated method would be to use nonlinear regression. This is where
volume loss Vs and distance to the point of inflexion i are both varied until the best fit is
found. The best fit may be defined by the minimum sum of absolute errors (or sum of
residuals), or it may be defined by the minimum value of the squares of residuals. This
can be achieved by creating a large table and finding the minimum. It turns out that
12
nonlinear regression is the most reliable method of calculating the distance to the point
of inflexion i, although occasionally it will not converge on an answer or will find multiple
minima.
The values of volume loss Vs and offset to the point of inflexion i presented in the paper
have since been recalculated using nonlinear regression, but are not significantly different
to the values presented in the paper and certainly do not affect the conclusions.
12
Note that the equation in the paper has a minus 1 on the bottom. This is because the
equation for the ‘population standard deviation’ should have been used rather than the
‘sample standard deviation’. It makes very little difference to the results.
13
There follows a series of graphs showing the short-term surface settlements measured
as soon after the TBM has exited the zone of influence as possible.
The tunnel is approximately 4.5km long, and the chainage is in reverse, so array 4300 is
about 200m from the launch shaft.
On each graph the Gaussian curve has been fitted using the method described in the
paper, denoted DCJ, and also using a nonlinear regression, denoted NRSAE. for this array,
the difference is quite striking as the nonlinear regression appears to want to fit six of
the points as well as possible while ignoring the larger centreline settlement, resulting in
a wider, shallower trough. It’s important to remember that when fitting curves to real
data there is no ‘right’ or ‘wrong’ answer necessarily.
14
These are the array 3975 surface settlements. Very good agreement between the two
curve-fitting methods, and the volume loss is very low at 0.1%.
15
Nearly a mile into the drive, these are the array 3060 surface settlements. The two
curve-fitting methods give different values of distance to the point of inflexion i at 11.6m
and 15.2m, but estimates of volume loss are very similar at around 0.5%.
16
These are the array 2800 surface settlements. The two curve-fitting methods are in good
agreement.
17
These are the array 2470 surface settlements. The two curve-fitting methods are in good
agreement. This settlement monitoring array was in Highbury Fields above the
Canonbury tunnel, a Victorian brick arch tunnel, which to one side of the alignment was
constructed by cut and cover and to the other side was bored. The Canonbury Tunnel
acted like a drain on the ground around it, with significant inflows to the tunnel after
rainfall events. So the ground was historically very disturbed to a depth of 17 m from the
surface.
18
Here are the array 2150 surface settlements, about halfway through the tunnel drive in a
side road just south of Highbury and Islington station. The two curve-fitting methods
disagree on the distance to point of inflexion value, 11.2m vs 15.8m.
19
This table is a summary of the Gaussian curve parameters from the surface settlement
arrays. The volume loss [click] calculated by either curve-fitting method was very similar.
The offset to point of inflection i [click] was a bit more sensitive to the curve-fitting
method used, with the nonlinear regression in some cases fitting a wider trough to the
data. However, the width of trough is still much narrower than expected.
The trough width parameter [click] is sometimes used, where z0 is the depth to axis of
the tunnel under construction and z is the depth to the point in question. Therefore z0 –
z represents the height above the tunnel being constructed. For the particular case of
surface settlements, z is zero, and so the offset to point of inflexion i is given by K times
the depth of the tunnel. This was first proposed by O’Reilly & New (1982), and they
envisaged that K would be a constant dependent only on the geology. The geology above
this tunnel is almost entirely stiff clay, and so one might expect K to be between 0.4 and
0.6. [click] However, all the values of both curve-fitting methods are well below this
range. There is no way to fit curves to this data with a value of K above 0.4 that does not
look entirely wrong.
Having assured ourselves that the repeatability of the data is good enough, that the
curve-fitting method is not a factor, and presented with 6 settlement troughs that show
that surface settlement trough width would be consistently overestimated at these
depths, we have a serious anomaly on our hands. The trough width should be twice as
wide. Could they be freak readings? Luckily there is also the subsurface movements to
look at before we have to make a decision.
20
Now we are going to look at the measured subsurface movements of third party tunnels
as the ring main tunnel passed underneath them.
I’m only going to describe the High Speed One and Northern Line tunnels in the
following slides, as we are focusing on Gaussian curve parameters and they were the
only tunnels for which data suitable for this purpose was obtained.
21
22
The High Speed One, or CTRL, tunnels were underpassed about 200m west of Corsica
Street shaft. There are two 7.15m ID tunnels with a 230 km/h high speed rail track
within them. The HS1 tunnel linings consist of precast reinforced concrete bolted
segments 350mm thick. The rail is mounted on floating slab track.
The HS1 tunnels are in London Clay above axis and in Upper Mottled Clay below axis
level, and the ring main tunnel is in the Upnor Formation, therefore both tunnels could
be considered to be in clay, with London Clay extending upwards above the HS1 tunnels
to within a few metres of the surface.
23
Here is a plan and a cross-section through the HS1 downline crossing. The ring main
tunnel is the little one on the bottom left.
The clear ground distance between ring main tunnel crown extrados and HS1 tunnel
invert extrados was 11.7m.
The CTRL Down line tunnel was underpassed first, and the separation distance between
the axes of the Down and Up lines was 17.3 m. Both tunnels were underpassed at an
angle of approximately 75°.
This was the first underpassing of a high speed rail tunnel in the UK, and at typical TBM
advance rates of 30-40 m/day, the two tunnels would be underpassed in a matter of
hours. Therefore, real time monitoring was required to immediately identify out-oftolerance movements of the track, and if necessary impose speed restrictions until the
track could be reset. Therefore trigger and alarm values with well-defined actions were
agreed in advance between all parties, to ensure there was no risk to the safe operation
of the railway. A designated competent person monitored and evaluated all the data in
real time and as a further level of protection automated text message alerts were sent to
the TWRM site management and Network Rail (CTRL).
24
These were the trigger values used for track movement, based on a standard distance of
35 m. You can see they are much lower than for an ordinary railway or metro system.
The speed restriction value for vertical track movement was 8 mm, but it was reduced to
7 mm to account for the accuracy of the monitoring system, and the possibility of
settlement at the ends of the string of tiltsensors that would not be captured in real
time.
25
The real time monitoring consisted of tiltsensors installed by Datum Monitoring Services.
These were attached to 2m long aluminium square hollow section beams, which were
bolted end-to-end into the concrete trackslab.
11no. 1m long tiltsensor beams placed between the rails at 3m spacing monitored twist
(twist is change in cant)
Multiplexers were installed in the crosspassage and a datalogger was installed in the
headhouse. Readings were recorded every minute and transmitted via Paknet – a
wireless radio communications system with 99.999% availability to a web-based
graphical display. Text messages would be sent to site staff and to Network Rail (CTRL)
staff if trigger levels were exceeded.
[click for photo] Here’s a picture during installation. You can see the longitudinal string of
2m tiltsensor beams to the side of the track and the transverse 1m beams between
sleepers measuring twist.
26
After underpassing, no movement of the Up line was discernible in the data. The Down
line, on the other hand, did experience some settlement. The tilt sensors at each end of
the string showed some rotation, so it was not known what the settlements at the ends
of the string were. Precise levelling of the ends of the string before and after
underpassing showed that no significant movements occurred, but the repeatability was
probably only ± 0.5 mm. A regression analysis varying both trough width and the string
end settlement showed that the best fit to the data was obtained with a string end
settlement of 0.05 mm and a trough width of 7.0 m. The tilt sensor data in the graph
have been resolved to take account of the 75° angle between the TWRM tunnel and the
CTRL tunnel. The volume loss of the Gaussian curve Vl was 0.29 %, the maximum
settlement Smax was 1.4 mm and the trough width parameter K was 0.407.
27
a prediction using Mair, Taylor & Bracegirdle (1993)’s empirical relationship between
trough width parameter and depth using the following equation predicts that the trough
width parameter K should be 0.85
[click] the measured value of K was 0.41 !!!
[click] This appears to be due to the fact that the ring main tunnel is so deep that the
relative depth z/z0 used in the equation is 34.1/51.4 = 0.66. A relative depth of 0.66
would be quite close for normal tunnel depths on which this empirical method was
based. But the HS1 tunnel axis is in this case 17.3m above the ring main tunnel axis, a
considerable distance. So it appears that height above the tunnel being constructed may
be the important factor, not relative depth z/z0.
28
The Northern Line tunnels were underpassed just west of the Angel platforms, under
Pentonville Road near the junction with Islington High Street. The Northern Line is quite
deep here – around 36m below the ground surface. The new ring main tunnel drive is
much deeper still, at 51.8m below the ground surface. The ring main tunnel was being
driven from North to South, so the Southbound tunnel was underpassed first, followed
by the disused tunnel then the Northbound.
29
In section it looks like this [click]. The TBM was excavating a mixed face of Thanet Sand
and Upnor Formation, while the Northern Line tunnels were in the Lambeth Group, with
the London Clay above. Hence in terms of ground movements the ground could be
characterised as predominantly stiff clay.
30
Here is a plan showing the monitoring. Trigonometric levelling techniques were used by
Survey Associates Ltd to level points installed in the concrete invert slab either side of
the tracks. In this way both settlement and twist could be measured manually on a
nightly basis.
31
This graph shows the final short term settlements in the 3 Northern Line tunnels, along
with a Gaussian curve fit to each one. The results for each tunnel are very similar.
32
The Gaussian curve parameters are summarised in this table. Again, as for the surface
settlements, these values are very very slightly different to the ones in the paper
because of that -1 that shouldn’t have been in the curve-fitting equation.
Again, if we compare the measured values of trough width parameter K to the values
predicted using Mair, Taylor & Bracegirdle’s equation, we find that K would have been
comfortably overpredicted. If you’re not sure what the effect of this difference in trough
width parameter K means then [click]
33
Here is a settlement trough with a trough width parameter of 0.91 shown in red, using
the same volume loss as the data.
34
Here is a settlement trough with a trough width parameter of 0.91 shown in red, using
the same maximum settlement as the data.
35
This graph shows the new data in context [highlighted yellow], when compared to the
meta-analysis of subsurface settlement data in clays reported in Figure 20 of Mair &
Taylor (1997) along with more recent measurements made in the Bakerloo and Northern
Line tunnels at Waterloo during construction of the Jubilee Line Extension reported by
Standing & Selman (2001), measurements made in the Piccadilly Line during
construction of the Heathrow CTA station by Cooper et al. (2002), data from
extensometers installed above the tunnel centreline at Heathrow Terminal 4 station, and
settlements measured in the Central Line during underpassing by the City of London
Cable Tunnel by Legge & Bloodworth (2003). In the key, the depths of the tunnels are
shown in square brackets alongside the references.
This graph has relative depth z/z0 on the y-axis. As well as the data from this project, you
can also see a less obvious effect of depth in the other data. For instance, Nyren’s data
[highlighted turquoise], for a relatively deep tunnel at 31m is to the left of the curve as
well.
Mair’s centrifuge test data also shows this pattern; the centrifuge model simulating a
depth of 16.5m [highlighted green] fits the curve really well, but the 9.8m depth tunnel
[highlighted red] is to the right.
So there seems to be trend where the curve seems to fit data from tunnels around 20m
deep, but underpredicts subsurface trough width for shallower tunnels and overpredicts
subsurface trough width for deeper tunnels.
36
An attempt at finding a relationship between trough width parameter K and height
above the tunnel was made. Note that surface and subsurface trough width parameter
values may now be plotted on the same graph. Suddenly the data from this project no
longer appears to be anomalous.
The increasing degree of scatter as the tunnel is approached is still an area that requires
further research, as the effect of tunnel size and construction method are not accounted
for in these simple empirical correlations.
It is therefore recommended that assuming a constant value of K as O’Reilly & New
(1982) suggested or assuming that K is always equal to 0.5 at the surface and that it
varies with relative depth as proposed by Mair et al. (1993) should be done with caution
for tunnels at depths greater than 35 m below the ground surface.
37
[click click] This paper has shown how small values of surface settlements can be
rigorously interpreted to obtain reliable values of volume loss and trough width
parameter to aid in the prediction of, in this case, far more critical subsurface
settlements.
[click] The monitoring systems employed on the Thames Water Ring Main Extension
Stoke Newington to New River Head project were adequate mitigation for the risks, and I
hope this paper is considered a valuable addition to the available literature on surface
and subsurface ground movements in clays, particularly since there are so few data for
tunnels at depths greater than about 35 m. Good predictions of surface and subsurface
ground movements will be critical to the feasibility of future projects, which will tend to
be at greater depth because of the need to avoid existing tunnels. In particular, accurate
prediction of trough width is crucial to the assessment of risk, as it determines not only
the maximum settlement for a fixed value of volume loss, but also the rate of change of
settlement transverse to the tunnel under construction. A narrower settlement trough
will have a higher maximum settlement for a given volume loss as the area under the
curve remains the same. It will also have higher gradients and curvatures, which will
increase the risk of damage to buildings and rail track distortions. On the TWRM project,
believed to be the first tunnel below 35 m depth for which trough widths have been
calculated, the trough width parameter was found to be consistently lower than
predicted using established empirical correlations.
[click] The project was a success in terms of control of volume loss, and it is hoped that a
case study such as this may help to alleviate the concerns of third party stakeholders
38
when faced with similar situations in future.
38
Daljit Dhanda
39