Acoustic Leak Detection in Water Networks
Robert M
¨
uller
1 a
, Steffen Illium
1 b
, Fabian Ritz
1 c
, Tobias Schr
¨
oder
2
, Christian Platschek
2
,
J
¨
org Ochs
2
and Claudia Linnhoff-Popien
1 d
1
Mobile and Distributed Systems Group, LMU Munich, Germany
2
Stadtwerke M
¨
unchen GmbH, Germany
Keywords:
Leak Detection, Water Networks, Acoustic Anomaly Detection, Applied Machine Learning.
Abstract:
In this work, we present a general procedure for acoustic leak detection in water networks that satisfies multiple
real-world constraints such as energy efficiency and ease of deployment. Based on recordings from seven
contact microphones attached to the water supply network of a municipal suburb, we trained several shallow
and deep anomaly detection models. Inspired by how human experts detect leaks using electronic sounding-
sticks, we use these models to repeatedly listen for leaks over a predefined decision horizon. This way we
avoid constant monitoring of the system. While we found the detection of leaks in close proximity to be a
trivial task for almost all models, neural network based approaches achieve better results at the detection of
distant leaks.
1 INTRODUCTION
Leakage is one of the main causes for water loss in
water supply and distribution networks. Undetected
leaks, which usually occur due to corrosion and soil
movement, may have extensive negative effects on the
surrounding infrastructure, customer convenience and
financial profit. Contributing to this problem is the
fact that it can take a considerable amount of time un-
til a leak is detected, localized and countermeasures
are put into place.
In the UK, approximately 3200 Million liters of
water are wasted due to leakages in water networks
every day (WaterUK, ), a disproportionately high
value with regard to climate change and a shortage
of drinking water in many countries. Greater efforts
are needed to minimize water loss through leakages.
A reliable indicator for substantial water loss is
the deviation of the zero-consumption status in a bal-
ance area at night. But if a leakage is more subtle, it
is usually not detected until water emerges from the
surfaces and residents report the issue.
Locating the source by means of the acoustic emis-
sion from exiting water is one the predominantly used
a
https://orcid.org/0000-0003-3108-713X
b
https://orcid.org/0000-0003-0021-436X
c
https://orcid.org/0000-0001-7707-1358
d
https://orcid.org/0000-0001-6284-9286
Figure 1: Simplified map of the location of each contact
microphone and significant objects in their neighborhood.
For each contact microphone, we depict the distance of the
shortest pipeline link to another contact microphone. Note
that distance is give n in meters of water pipe.
306
Müller, R., Illium, S., Ritz, F., Schröder, T., Platschek, C., Ochs, J. and Linnhoff-Popien, C.
Acoustic Leak Detection in Water Networks.
DOI: 10.5220/0010295403060313
In Proceedings of the 13th International Conference on Agents and Artificial Intelligence (ICAART 2021) - Volume 2, pages 306-313
ISBN: 978-989-758-484-8
Copyright
c
2021 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
approaches (El-Zahab and Zayed, 2019). The process
is as follows:
1) Leak Noise Correlation: Liquid escaping a pipe
creates shock waves that cause the pipe to vibrate, re-
sulting in a characteristic leak sound. At least two
sensors are attached to the pipe around the presumed
leak location. By analyzing the variation in propaga-
tion times of the leak’s acoustic emission between the
sensors, the location is further narrowed down. This
approach requires infrastructural knowledge, such as
pipe material, diameter, length and corresponding
sound velocities.
2) Electro-acoustic Method: To further pinpoint the
location, human leak detection experts search for leak
noises using electronic sounding sticks. A leak can
be localized by exploiting the fact that leak sounds
become more dominant as the expert approaches the
leak. The process relies on trained human experts
and requires a solid estimate of the leakage location
as finding the first evidence might otherwise become
very time consuming.
In this work, we aim to automatize the electro-
acoustic method utilizing machine learning. We
record normal operation data using seven contact mi-
crophones attached on various parts of the water sup-
ply network in a suburban area of Munich. Subse-
quently, we train several unsupervised anomaly de-
tection models and embed these into a broader proce-
dure that satisfies several real world constraints, such
as energy consumption and limited bandwidth. Our
approach can be understood as a first step towards a
fully automatized leak detection system that does not
need to wait for leak aftereffects to appear. More-
over, our approach is data-driven and does not require
a detailed mathematical model of the water network’s
inner workings.
The rest of the paper is structured as follows: In
Section 2 we briefly review related work. Then we
motivate and propose a general procedure for acoustic
leak detection in Section 5. In Section 4 we describe
the data acquisition process followed by the descrip-
tion of the experimental setup and its evaluation. We
close by summarizing our findings in Section 6 and
laying out future work in Section 7
2 RELATED WORK
Related approaches of acoustic leak detection are
mostly based on cross-correlation (Muggleton and
Brennan, 2004; Gao et al., 2017), wavelet trans-
forms (Ni and Iwamoto, 2002; Ting et al., 2019),
Support-Vector-Machines (Kang et al., 2017; Cody
et al., 2017) or neural networks (Kang et al., 2017;
Chuang et al., 2019; Cody et al., 2020). Other meth-
ods combine acoustic data with additional sensory
measurements (Stoianov et al., 2007). Their perfor-
mance depends on the pipe network material (e.g.
polyethylene or metal). However, none of these ap-
proaches directly matches the setup and constraints of
this work as experiments are mostly carried out under
laboratory conditions and only study a single algorith-
mic approach. Other physical phenomena that occur
in presence of a leak can also be taken advantage of,
e.g. using thermography, ground penetration radar or
pressure-based methods. An exhaustive summary can
be found in (Adedeji et al., 2017; Chan et al., 2018).
Apart from leak detection, the broader field of
acoustic anomaly detection has recently gained trac-
tion.
Most of the recent work in the field is based upon
Deep-Autoencoders (AEs). An AE learns to recon-
struct its input from a compressed latent representa-
tion. Unseen, anomalous data is assumed to have a
higher reconstruction error. The approaches mostly
differ in the architecture used.
Duman et al. (Duman et al., 2019) use a deep con-
volutional AE on the spectrograms of sounds from in-
dustrial processes. In (Meire and Karsmakers, 2019)
the authors also conclude that convolutional AEs per-
form well on the task of acoustic anomaly detection
while they have also found the One-Class Support
Vector Machine to be a strong competitor. In the same
vein, M
¨
uller et al. (M
¨
uller et al., 2020) use pretrained
convolutional neural networks for feature extraction
and train various traditional anomaly detection mod-
els. Other work (Marchi et al., 2015; Li et al., 2018;
Nguyen et al., 2019) explicitly takes the sequential
nature of sound into account by training a recurrent
AE based on Long-Short-Term Memory. Koizumi et
al (Koizumi et al., 2017) use a more traditional Feed-
Forward AE in conjunction with a novel loss function
based on statistical hypothesis testing. This approach
requires the simulation of anomalous sounds by us-
ing rejection sampling. Recently, Suefusa et al. (Sue-
fusa et al., 2020) proposed to alter the input an AE
receives to improve the detection of non-stationary
sounds. Instead of predicting all spectrogram frames,
they remove the center frame and use it as prediction
target thereby alleviating the difficulty of predicting
the edge frames.
Another line of work investigates upon methods
that operate directly on the raw waveform (Hayashi
et al., 2018; Rushe and Namee, 2019). WaveNet-
like (Oord et al., 2016) generative architectures that
utilize causal-dialated convolutions are used to pre-
dict the next sample. The prediction error serves as
the anomaly score.
Acoustic Leak Detection in Water Networks
307
3 MOTIVATION AND PROBLEM
DEFINITION
An automated leak detection system should be
energy efficient, easy to deploy and easy to update.
We derive the problem definition from the following
deployment scenario: Small battery driven IoT
devices record sounds from contact microphones
upon request from a central control unit. During a
predefined period, sounds are periodically recorded
for a short time frame (2-5s). Afterwards, the col-
lection of short audio sequences is transmitted to the
central control unit via a low energy, low bandwidth
radio network (SWM, 2018). This reduces the high
energy and bandwidth consumption that constant
monitoring would cause. The central control unit
can then put all received measurements in context to
decide whether a leak is present or not. Moreover,
using a central control unit makes updating the
system (e.g. changing the algorithm or retraining the
model) more feasible. This approach is inspired by
how human leak detection experts make decisions
using electronic sounding sticks. After attaching
the stick on to a hydrant connection, the expert
listens carefully for up to ten seconds. When facing
audible noise emitted by infrastructure, agriculture
or industry, the process is repeated until an accurate
decision can be made. For precise leak localization,
various hydrant or valve connections in the area are
checked accordingly. Increasing leak sounds indicate
a decreased distance to the leak. In this work, we
focus on leak detection and leave localization for
future work. Our approach is formulated as follows:
Problem Definition 1. Let X be some representation
of an h minute recording from a contact microphone
on a water pipe. Further, let F
θ
: Y R
+
be some
trainable function with parameters θ where Y is a
sample of X with a length of t seconds. Apply F
φ
on
m different sections of X to obtain a vector R
m
+
of
positive real valued anomaly scores. To compute a
single anomaly score for X, combine the measure-
ments using an aggregation function φ : R
m
+
R
+
.
Find (F
θ
, φ, h, t, m) such that the anomaly scores for
leak sections are higher than the anomaly-scores for
no-leak sections.
The accompanying algorithm is depicted in Algo-
rithm 1. Note that the sampling timepoints are equally
distributed across the whole recording (Line 3), i.e.
linspace returns the set
{bi
(h t)
m
c|i 1, . . . m}
Algorithm 1: Acoustic Leak Detection.
Input: Audio recording X
Parameters: Score func. F
θ
, Aggregation
func. φ, Number of samples m,
Sample length t, preprocessing
func. ρ
Output: Anomaly Score for X
1 begin
2 h length(X)
3 P linspace(0, h t, m)
4 S [ ]
5 for p in P do
6 x X [p : p +t]
7 ˜x ρ(x)
8 s F
θ
( ˜x)
9 append s to S
10 end
11 return φ(S)
12 end
Figure 2: Visualization of the most important aspects of our
acoustic leak detection approach.
and each sample is preprocessed (Line 7) to fit the do-
main of the score function (e.g. spectrum, raw-audio
or feature vector). We also depict the most important
aspects of the method visually in Figure 2 To obtain
enough leak sound recordings for supervised learning,
one would have to artificially create a vast amount of
leaks. Furthermore, for these recordings to contain
enough diversity, this would have to be done on vari-
ous sections of the water network. Due to the financial
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
308
and environmental issues that would arise, we assume
that F
φ
is trained on normal operation data only.
4 DATA ACQUISITION
For data acquisition, a suburban part of Munich with
little traffic was chosen as a testing ground. Seven
contact microphones were directly attached onto hy-
drants of an iron pipe network ( 100mm). Figure 1
shows a simplified map marking the locations of the
microphones as well as significant objects in the sur-
rounding area that might influence the quality of the
recordings. Sounds were recorded for three months.
To avoid excessively high signals (overdrive), we
kept the pre-amplifier, equalizer and post-amplifier
neutral, effectively recording sounds ”as-is” and dis-
abling the dynamic sensitivity adjustment. This leads
to the recordings having a headroom of 24 dB on av-
erage. All contact microphones are designed to record
sounds reliably between 300Hz and 3000Hz. During
the time of recording, a medium sized water leakage
occurred in close proximity ( 20m) to contact mi-
crophone one (see Figure 1). The leak was present
for 28 days and was fixed thereafter. Furthermore, we
conducted various field tests. We simulated a leakage
by opening a hydrant near contact microphone five
and attached a contact microphone to hydrant connec-
tions between contact microphone four and five every
100 meters. Overall, we were neither able to hear the
leak nor able to observe typical leak patterns in the
frequency spectrograms beyond a distance of 600 me-
ters.
5 EVALUATION
In this section, we conduct various experiments to
find an appropriate setting of (F
θ
, φ, h, t, m). First we
introduce the dataset followed by a brief discussion of
the anomaly detection models that were used. Finally,
we present the experiments and results.
5.1 Methodology and Dataset
The evaluation set up is designed to line up with
Problem Definition 1. We use 600 hours of normal
operation recordings equally sampled from all contact
microphones. Sounds were recorded in mono with
a sampling rate of 16 kHz and a bit depth of 16 bit.
A band-pass filter is applied to remove information
above and below the contact microphone’s supported
frequencies. For training, we split the data into
t = {2s, 5s} long audio samples with no overlap.
These values are at the lower end of how long a
human expert listens using a sounding stick. Samples
are then preprocessed (Algorithm 1, Line 7) by either
computing its mel-spectrogram or extracting eight
spectral features (chromagram, spectral centroid,
spectral bandwidth, spectral contrast, spectral roll
off, spectral flatness, zero-crossing rate and the
root-mean-square value) which we have identified
to be good descriptors of a leak during early stages
of research. Each feature-vector is standardized by
subtracting the mean and dividing by the standard-
deviation of the individual features. Figure 3 provides
more insights into the characteristics of the dataset.
To evaluate the ability of our approach to differentiate
between leak and no-leak, we use the following two
datasets:
Leak in Close Proximity: To measure the
ability to detect leaks in close proximity to a contact
microphone, we select five consecutive days during
which a leak was present in close proximity to contact
microphone 1. For each day, we use the recordings
from contact microphone 1 and one other contact
microphone. This results in a balanced evaluation
dataset with an equal number of leak and no-leak
recordings.
Leak in the Distance (Synthetic): This setting
measures how well a leak can be detected when it is
farther away from a contact microphone. As we were
only able to obtain recordings from a leakage near a
single contact microphone, we mixed normal record-
ings with leak sounds having a high Signal-to-Noise
Ratio of +24dB. Note that here noise stands for the
leak sounds. Doing so yields synthetically generated
recordings that resemble the characteristics of a leak-
age in the distance (Section 4). Synthetically creating
anomalous recordings is a common approach when
anomalous data is scarce (Duman et al., 2019; Puro-
hit et al., 2019; Koizumi et al., 2019; Socor
´
o et al.,
2015; Stowell et al., 2015; Nakajima et al., 2016).
We select 96 hours of no-leak recordings from contact
microphones two and three. 48 hours of these record-
ings were mixed with randomly sampled leak sounds
taken from a time-span between 0a.m. - 4a.m. This
time-span was chosen as it yields pure leak sounds
with very little noise.
5.2 Anomaly Detection Models
To compute anomaly scores for each individual
sample (Algorithm 1, Line 8) we use various density
estimation, ensemble models and deep neural net-
works. Models can be further subdivided according
Acoustic Leak Detection in Water Networks
309
Figure 3: Four Mel-spectrograms depicting a) normal operation, b) a leak in close proximity, c) a synthetically generated
distant leak and d) water withdrawal. Spikes are caused by cars driving over the hydrant cover above the contact microphones.
However, other recordings may also contain footsteps, animal noises or interfering noises due to a nearby transformer house.
A leak is characterized by high energy in the upper frequencies. This pattern becomes less dominant as the distance to the
leak increases and vanishes after 600m. The leak shown in b) has a throughput of 300
ml
s
.
Table 1: Average AUCs with standard deviations (over 5 seeds) for different settings of F , h and t where φ = median, m =
20, h = 30min and m = 40, h = 60min.
Leak in close proximity Leak in the distance (synthetic)
Decision horizon h 30min 60min 30min 60min
Sample length t 2s 5s 2s 5s 2s 5s 2s 5s
GMM 74.8±2.5 88.0±1.3 74.6±2.1 87.6±1.6 64.0±5.5 68.7±0.0 68.4±1.0 70.0±1.2
B-GMM 78.7±3.7 88.1±2.4 78.8±4.0 87.7±2.6 64.0±5.1 70.5±1.8 69.3±1.2 72.5±2.9
IF 98.6±0.0 98.8±0.0 98.8±0.0 100±0.0 41.4±0.0 42.5±1.8 41.0±1.5 42.3±2.3
RealNVP 98.6±2.0 98.0±3.6 99.0±1.3 98.1±3.7 75.2±2.7 76.1±2.5 77.1±2.8 78.3±3.1
DCAE 98.2±0.0 98.9±0.0 99.8±0.0 100±0.0 75.1±1.5 73.4±5.0 76.1±2.1 76.0±3.3
AAE 98.9±0.0 99.0±0.0 99.8±0.0 99.8±0.0 75.0±4.4 69.9±3.2 77.0±5.2 71.6±2.4
AVB 98.1±0.5 99.5±0.1 99.6±0.5 100±0.0 76.9±3.4 75.5±2.7 78.9±3.7 77.5±2.8
to the input they receive.
Mel-spectrogram Input: A mel-spectrogram is
a logarithmically scaled spectrogram to better align
with how humans perceive sound. We compute the
mel-sepctograms with 64 mels, fft window = 2048
frames, hop length = 512 frames and normalize them
to lie in the range [0, 1]. We also experimented with
other settings but have found all methods to be robust
against these parameters. The following deep neural
networks are trained: i) Deep Convolutional Auto
Encoder (DCAE) A neural network that compresses
the input into a low dimensional representation and
then reconstructs the input from this representation.
We use [4, 16, 32] convolution filters with 2 × 2 max-
pooling in-between each layer with ReLU (Agarap,
2018) as activation function. We train for 100 epochs
with a batch size of 128, a L2 weight penalty of
10
6
and optimize the AE using ADAM (Kingma
and Ba, 2014) with a learning rate of 0.0001. For
reconstruction, the inverse operations (deconvolution
and up sampling) are applied in reverse order. ii)
Adversarial Auto Encoder (AAE) (Makhzani et al.,
2015) Adversarially trained DCAE such that the
latent space spanned by the bottleneck features
matches the prior distribution N (0, I). iii) Adver-
sarial Variational Bayes (Mescheder et al., 2017)
Adversarially trained variational DCAE.
Feature-Vector Input: Here we extract eight
spectral features (see Section 5) for each sample. The
resulting feature vectors are used to train i) Gaus-
sian Mixture Model (GMM) A density estimation al-
gorithm that models the underlying probability distri-
bution as a mixture of Gaussians. Parameters are es-
timated using expectation-maximization. We use 16
mixture components with diagonal covariance matri-
ces. ii) Bayesian Gaussian Mixture Model (B-GMM)
In contrast to a GMM, this model is trained using
variational inference. We use the same parameters
as for the GMM. iii) RealNVP (Dinh et al., 2017) A
sequence of invertible transformations modeled by a
neural network that can directly compute the proba-
bility density of the data. The approach is based on
normalizing flows. We use 3 coupling layers, a hid-
den dimension of 150, a batch size of 768 and assume
a normal distribution with zero mean and unit vari-
ance as base distribution. All of the methods above
use the log-probability of a sample as normality score.
iv) Isolation Forest (IF) (Liu et al., 2008) Recursively
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
310
partitions the feature space. The number of splits
needed to isolate a data point is used as normality
score. We use 120 base estimators.
Model performance is measured with the Area
Under the Receiver Operating Characteristics (AUC)
which quantifies how well a model can distinguish be-
tween leak and no-leak across all classification thresh-
olds. The AUC is the standard metric to evaluate
anomaly detection models across many domains (Ag-
garwal, 2015; Chalapathy and Chawla, 2019) as it
yields a complete sensitivity/specificity report. The
evaluation setup follows (Ruff et al., 2018).
5.3 Choosing the Model
The most important aspect of our approach is to
choose a suitable score function F
φ
. The datasets
presented in Section 5 are split into consecutive h =
30min and h = 60min long recordings. On each of
these recordings, we run Algorithm 1 independently
to obtain a single anomaly score. Here we chose the
median as aggregation function φ due to its robustness
against outliers
1
and evaluate across all models from
Section 5.2 with t = 2 and t = 5. We set m = 20 and
m = 40 for h = 30 min and h = 60 min, respectively.
Model parameters were determined using a small de-
velopment set.
Results are depicted in Table 1. In the case of
Leak in close proximity the performance of GMM and
B-GMM is significantly worse compared to all other
models. Interestingly, AUC-scores for GMM and B-
GMM increase by approximately 12% when t = 5.
This finding carries over the second setting as well.
When h = 60 and t = 5 IF, DCAE and AVB reach
perfect scores. Generally, the setting can be consid-
ered trivial for IF, RealNVP, DCAE, AVB and AAE.
Note that in case of the AAE and AVB we also tried
using the log-probability of the latent representation
but the reconstruction error turned out to yield better
results.
Results for Leak in the distance (synthetic) paint
a different picture. In this setting, the differences be-
tween inlier and outlier are more subtle and therefore
considerably harder to detect. IF fails completely on
this task due to the reduced distance between inlier
and outlier in feature space. The number of splits is
the same for almost all samples and mostly depends
on the random splitting points. Moreover, we observe
general superiority of the neural network (NN) based
methods RealNVP, DCAE, AVB and AAE indicating
1
The median showed the best results compared to other
aggregation functions like the mean. Using only the most
normal sample (min-pooling) leads to a comparable, but
worse performance.
that NNs better reveal the more subtle differences. On
2s, Mel-spectrogram based AVB performs best and on
5s RealNVP outperforms all other methods. We ob-
served that auto encoders work best with a small bot-
tleneck as this limits their ability to generalize over
leak patterns based on water withdrawal. Moreover,
we can verify that the extracted feature set is indeed
a good leak indicator as ist shows strong performance
when used in conjunction with NN based RealNVP.
5.4 Choosing the Number of Samples
In this experiment, we evaluate upon the influence
of the number of samples on the model perfor-
mance. In figure 4, we vary the number of sam-
ples m from 5 to 105 in steps of 10 samples. We
evaluate across all combinations of decision horizon
h {30min, 60min} and sample length t {2s, 5s}
on Leak in the distance (synthetic) with φ=median.
For better clarity, we only show results for RealNVP
and DCAE.
In Figure 4, we observe that using less than 15
samples yields significantly worse performance. Re-
sults stabilize thereafter and peak performance is
achieved at m = 65 for most settings. We conclude
that increasing the number of samples has a positive
effect on the performance because it makes the de-
cision less dependent on individual samples. More
samples better represent the underlying distribution as
a leak is characterized by a constantly present acous-
tic pattern. In contrast to a streaming approach, only
a fraction of all possible measurements has to be con-
sidered as we did not observe a substantial increase in
performance by considering even more samples.
Figure 4: Performance comparison for m =
{5, 15, 25, . . . 105} on Leak in the distance (synthetic).
Each value is the average AUC over 5 seeds.
Acoustic Leak Detection in Water Networks
311
6 CONCLUSION
In this work, we presented a general procedure for
leak detection in water networks that satisfies real-
world constraints and provided a thorough evaluation
of different parameter settings. While a leak in close
proximity to a contact microphone is trivial for most
scoring functions, neural network based approaches
yielded superior results with respect to the detection
of a (synthetic) leak in the distance. Additionally, we
found that it is not necessary to constantly monitor
the system. It suffices to consider a fraction of the
recordings during a predefined decision horizon.
7 FUTURE WORK
Future work might investigate further upon differ-
ent scoring functions, e.g. by taking the sequen-
tial nature of the recordings into account. Other av-
enues worth exploring are more sophisticated sam-
pling and aggregation strategies. The extension of
our approach to leak localization (e.g. via trilater-
ation) represents the next logical step. Moreover,
instead of training a single model on data from all
contact-microphones, one might train separate mod-
els. Another possibility would be to use weight shar-
ing and condition (Huang and Belongie, 2017) the
model on the contact-microphone ID or on features
of their surrounding. Another important aspect that
still remains to be investigated, is how to update the
model when new data (possibly collected in another
area) arrives (Koizumi et al., 2020).
ACKNOWLEDGMENTS
This work is part of the research project ErLoWa
which was carried out in cooperation with Stadtwerke
M
¨
unchen GmbH.
REFERENCES
Adedeji, K. B., Hamam, Y., Abe, B. T., and Abu-Mahfouz,
A. M. (2017). Towards achieving a reliable leakage
detection and localization algorithm for application in
water piping networks: An overview. IEEE Access,
5:20272–20285.
Agarap, A. F. (2018). Deep learning using rectified linear
units (relu). arXiv preprint arXiv:1803.08375.
Aggarwal, C. C. (2015). Outlier analysis. In Data mining,
pages 237–263. Springer.
Chalapathy, R. and Chawla, S. (2019). Deep learning
for anomaly detection: A survey. arXiv preprint
arXiv:1901.03407.
Chan, T., Chin, C. S., and Zhong, X. (2018). Review of cur-
rent technologies and proposed intelligent methodolo-
gies for water distributed network leakage detection.
IEEE Access, 6:78846–78867.
Chuang, W.-Y., Tsai, Y.-L., and Wang, L.-H. (2019). Leak
detection in water distribution pipes based on cnn with
mel frequency cepstral coefficients. In Proceedings of
the 2019 3rd International Conference on Innovation
in Artificial Intelligence, pages 83–86.
Cody, R., Narasimhan, S., and Tolson, B. (2017). One
Class SVM Leak Detection in Water Distribution
Systems’.
Cody, R. A., Tolson, B. A., and Orchard, J. (2020). De-
tecting leaks in water distribution pipes using a deep
autoencoder and hydroacoustic spectrograms. Journal
of Computing in Civil Engineering, 34(2).
Dinh, L., Sohl-Dickstein, J., and Bengio, S. (2017). Den-
sity estimation using real NVP. In 5th International
Conference on Learning Representations, ICLR 2017.
Duman, T. B., Bayram, B., and
˙
Ince, G. (2019). Acoustic
anomaly detection using convolutional autoencoders
in industrial processes. In International Workshop on
Soft Computing Models in Industrial and Environmen-
tal Applications, pages 432–442. Springer.
El-Zahab, S. and Zayed, T. (2019). Leak detection in wa-
ter distribution networks: an introductory overview.
Smart Water, 4(1):5.
Gao, Y., Brennan, M. J., Liu, Y., Almeida, F. C., and
Joseph, P. F. (2017). Improving the shape of the cross-
correlation function for leak detection in a plastic wa-
ter distribution pipe using acoustic signals. Applied
Acoustics, 127:24–33.
Hayashi, T., Komatsu, T., Kondo, R., Toda, T., and Takeda,
K. (2018). Anomalous sound event detection based on
wavenet. In 2018 26th European Signal Processing
Conference (EUSIPCO), pages 2494–2498. IEEE.
Huang, X. and Belongie, S. (2017). Arbitrary style transfer
in real-time with adaptive instance normalization. In
Proceedings of the IEEE International Conference on
Computer Vision, pages 1501–1510.
Kang, J., Park, Y.-J., Lee, J., Wang, S.-H., and Eom, D.-
S. (2017). Novel leakage detection by ensemble cnn-
svm and graph-based localization in water distribution
systems. IEEE Transactions on Industrial Electronics,
65(5):4279–4289.
Kingma, D. P. and Ba, J. (2014). Adam: A
method for stochastic optimization. arXiv preprint
arXiv:1412.6980.
Koizumi, Y., Saito, S., Uematsu, H., and Harada, N. (2017).
Optimizing acoustic feature extractor for anomalous
sound detection based on neyman-pearson lemma. In
2017 25th European Signal Processing Conference
(EUSIPCO), pages 698–702. IEEE.
Koizumi, Y., Saito, S., Uematsu, H., Harada, N., and
Imoto, K. (2019). Toyadmos: A dataset of miniature-
machine operating sounds for anomalous sound de-
tection. In 2019 IEEE Workshop on Applications of
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
312
Signal Processing to Audio and Acoustics (WASPAA),
pages 313–317. IEEE.
Koizumi, Y., Yasuda, M., Murata, S., Saito, S., Uematsu,
H., and Harada, N. (2020). Spidernet: Attention
network for one-shot anomaly detection in sounds.
In ICASSP 2020-2020 IEEE International Confer-
ence on Acoustics, Speech and Signal Processing
(ICASSP), pages 281–285. IEEE.
Li, Y., Li, X., Zhang, Y., Liu, M., and Wang, W. (2018).
Anomalous sound detection using deep audio repre-
sentation and a blstm network for audio surveillance
of roads. IEEE Access, 6:58043–58055.
Liu, F. T., Ting, K. M., and Zhou, Z.-H. (2008). Isolation
forest. In 2008 Eighth IEEE International Conference
on Data Mining, pages 413–422. IEEE.
Makhzani, A., Shlens, J., Jaitly, N., Goodfellow, I., and
Frey, B. (2015). Adversarial autoencoders.
Marchi, E., Vesperini, F., Eyben, F., Squartini, S., and
Schuller, B. (2015). A novel approach for automatic
acoustic novelty detection using a denoising autoen-
coder with bidirectional lstm neural networks. In 2015
IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP), pages 1996–2000.
IEEE.
Meire, M. and Karsmakers, P. (2019). Comparison of deep
autoencoder architectures for real-time acoustic based
anomaly detection in assets. In 2019 10th IEEE In-
ternational Conference on Intelligent Data Acquisi-
tion and Advanced Computing Systems: Technology
and Applications (IDAACS), volume 2, pages 786–
790. IEEE.
Mescheder, L., Nowozin, S., and Geiger, A. (2017). Adver-
sarial variational bayes: Unifying variational autoen-
coders and generative adversarial networks. In Pro-
ceedings of the 34th International Conference on Ma-
chine Learning-Volume 70, pages 2391–2400. JMLR.
org.
Muggleton, J. and Brennan, M. (2004). Leak noise propaga-
tion and attenuation in submerged plastic water pipes.
Journal of Sound and Vibration, 278(3):527–537.
M
¨
uller, R., Ritz, F., Illium, S., and Linnhoff-Popien, C.
(2020). Acoustic anomaly detection for machine
sounds based on image transfer learning. arXiv
preprint arXiv:2006.03429.
Nakajima, Y., Naito, T., Sunago, N., Ohshima, T., and
Ono, N. (2016). Dnn-based environmental sound
recognition with real-recorded and artificially-mixed
training data. In INTER-NOISE and NOISE-CON
Congress and Conference Proceedings, volume 253,
pages 1832–1841. Institute of Noise Control Engi-
neering.
Nguyen, D., Kirsebom, O. S., Fraz
˜
ao, F., Fablet, R.,
and Matwin, S. (2019). Recurrent neural networks
with stochastic layers for acoustic novelty detection.
In ICASSP 2019-2019 IEEE International Confer-
ence on Acoustics, Speech and Signal Processing
(ICASSP), pages 765–769. IEEE.
Ni, Q.-Q. and Iwamoto, M. (2002). Wavelet transform of
acoustic emission signals in failure of model com-
posites. Engineering Fracture Mechanics, 69(6):717–
728.
Oord, A. v. d., Dieleman, S., Zen, H., Simonyan, K.,
Vinyals, O., Graves, A., Kalchbrenner, N., Senior,
A., and Kavukcuoglu, K. (2016). Wavenet: A
generative model for raw audio. arXiv preprint
arXiv:1609.03499.
Purohit, H., Tanabe, R., Ichige, K., Endo, T., Nikaido,
Y., Suefusa, K., and Kawaguchi, Y. (2019). Mimii
dataset: Sound dataset for malfunctioning indsutrial
machine investigation and inspection. In Acoustic
Scenes and Events 2019 Workshop (DCASE2019),
page 209.
Ruff, L., Vandermeulen, R., Goernitz, N., Deecke, L., Sid-
diqui, S. A., Binder, A., M
¨
uller, E., and Kloft, M.
(2018). Deep one-class classification. In International
Conference on Machine Learning, pages 4393–4402.
Rushe, E. and Namee, B. M. (2019). Anomaly detec-
tion in raw audio using deep autoregressive networks.
In ICASSP 2019 - 2019 IEEE International Con-
ference on Acoustics, Speech and Signal Processing
(ICASSP), pages 3597–3601.
Socor
´
o, J. C., Ribera, G., Sevillano, X., and Al
´
ıas, F. (2015).
Development of an anomalous noise event detection
algorithm for dynamic road traffic noise mapping. In
Proceedings of the 22nd International Congress on
Sound and Vibration (ICSV22), Florence, Italy, pages
12–16.
Stoianov, I., Nachman, L., Madden, S., Tokmouline, T., and
Csail, M. (2007). Pipenet: A wireless sensor network
for pipeline monitoring. In 2007 6th International
Symposium on Information Processing in Sensor Net-
works, pages 264–273.
Stowell, D., Giannoulis, D., Benetos, E., Lagrange, M., and
Plumbley, M. D. (2015). Detection and classification
of acoustic scenes and events. IEEE Transactions on
Multimedia, 17(10):1733–1746.
Suefusa, K., Nishida, T., Purohit, H., Tanabe, R., Endo,
T., and Kawaguchi, Y. (2020). Anomalous sound de-
tection based on interpolation deep neural network.
In ICASSP 2020-2020 IEEE International Confer-
ence on Acoustics, Speech and Signal Processing
(ICASSP), pages 271–275. IEEE.
SWM, P. (2018). Die swm vernetzen m
¨
unchen:
Lora-netz am start f
¨
ur das internet der dinge.
https://www.swm.de/dam/swm/pressemitteilungen/
2018/06/20180604-swm-bauen-lora-netz-auf.pdf.
Ting, L., Tey, J., Tan, A., King, Y., and Faidz, A. (2019).
Improvement of acoustic water leak detection based
on dual tree complex wavelet transform-correlation
method. In IOP Conference Series: Earth and En-
vironmental Science, volume 268. IOP Publishing.
WaterUK. Leaking pipes. http://https://www.discoverwater.
co.uk/leaking-pipes. Accessed: 2020-01-20.
Acoustic Leak Detection in Water Networks
313