Avrupa Bilim ve Teknoloji Dergisi European Journal of Science and Technology

Özel Sayı 37, S. 113-118, Haziran 2022 Special Issue 37, pp. 113-118, June 2022
© Telif hakkı EJOSAT’a aittir Copyright © 2022 EJOSAT
Araştırma Makalesi www.ejosat.com ISSN:2148-2683 Research Article

Evaluation and Compensation of Temperature Effects on Ultrasonic

Flow Measurement

Alkım Gökçen1* , Bahadır Yeşil2

1* BAYLAN Measurement Meters, Department of Research and Development, Izmir, Turkey, (ORCID: 0000-0002-8131-388X), [email protected]
2BAYLAN Measurement Meters, Department of Research and Development, Izmir, Turkey, (ORCID: 0000-0002-9622-2593), [email protected]

(5th International Symposium on Innovative Approaches in Smart Technologies – 28-29 May 2022)
(DOI: 10.31590/ejosat.1136816)
ATIF/REFERENCE: Gökçen, A. & Yeşil, B. (2022). Evaluation and Compensation of Temperature Effects on Ultrasonic F low
Measurement. European Journal of Science and Technology, (37), 113-118.

Abstract
This paper presents an evaluation of temperature effects on ultrasonic piezoelectric transducers for electronic flow measurement
devices. Transducers generates ultrasonic wave against electrical signals and electrical signals against ultrasonic waves due to their
bidirectional characteristics. Temperature dynamics of the physical environment is one of the most crucial parameters which a ffects
the electrical dynamics of the ultrasonic transducers. Due to the temperature related false sensor readings, flow me asurement process
for different temperature causes calibration errors. In order to identify the temperature effects on transducers characterist ics and
constitute a generalized solution, a test procedure and data collection process are developed. Initially, two identical transducers are
located reciprocally on a flow meter body. Secondly, bodies are located on a test bench to get signal measurements for differ ent flows.
A wireless communication data acquisition card is employed to collect ultrasonic signal measurements. Test procedure is repeated for
5 different temperatures and 13 flow rates. The created dataset is evaluated and visualized in MATLAB environment. A temperat ure
effect compensation process, which is based on machine learning algorithms, is prop osed. This method considers time domain
information of transducer elements. Experiment temperature value and average values of Time -of-Flight (TOF) signals for each
transducers are considered to predict actual flow velocity. In this manner, machine learnin g algorithms linear regression, suppor vector
regression (SVR), Gaussian process regression (GPR) and artificial neural networks (ANN) are employed to construct the relation
between temperature variation and flow measurement. Compensation performance is in vestigated by considering the 𝑅 2 , root mean
square error (𝑅𝑀𝑆𝐸 ), mean absolute error (𝑀𝐴𝐸 ) and mean square error (𝑀𝑆𝐸) model evaluation metrics. According to the results,
neual network based compensation algorithm gives the best result with 𝑅 2 = 0.95.

Keywords: Ultrasonic transducers, Flow metering, Compensation, Time-of-Flight Measurement.

Ultrasonik Akış Ölçümünde Sıcaklık Etkisinin İncelenmesi ve

Kompenzasyonu

Öz
Bu makale, elektronik akış ölçüm cihazları için ultrasonik piezoelektrik dönüştürücüler üzerindeki sıcaklık etkilerinin bir
değerlendirmesini sunar. Dönüştürücüler, çift yönlü özelliklerinden dolayı elektrik sinyallerine karşı ultrasonik dalga ve ultrasonik
dalgalara karşı elektrik sinyalleri üretir. Fiziksel ortamın sıcaklık dinamiği, ultrasonik dönüştürücülerin elektrik dinamiklerini
etkileyen en önemli parametrelerden biridir. Sıcaklık değişimi kaynaklı yanlış sensör okumaları, farklı sıcaklıklar için akış ölçüm
işlemi sırasında kalibrasyon hatalarına neden olur. Bu nedenle, dönüştürücü özellikleri üzerindeki sıcaklık etkilerini belirlemek ve
genelleştirilmiş bir çözüm oluşturmak için bir test prosedürü ve veri toplama süreci geliştirilmiştir. Başlangıçta , bir akış ölçer gövdesi
üzerinde karşılıklı olarak iki özdeş dönüştürücü konumlandırılmıştır. İkinci olarak, gövdeler, farklı akışlar için sinyal ölç ümleri almak

*
Alkım Gökçen: BAYLAN M easurement M eters, Department of Research and Development, Izmir, Turkey, ORCID: 0000-0002-8131-388X,
a [email protected]

http://dergipark.gov.tr/ejosat 113
 Avrupa Bilim ve Teknoloji Dergisi

üzere bir test masasına yerleştirilmiştir. Ultrasonik sinyal ölçümlerini toplamak için bir ka blosuz iletişim veri toplama kartı
kullanılmıştır. Test işlemi 5 farklı sıcaklık ve 13 debi için tekrarlanmıştır. Veri toplama sonucu elde edilen veri seti MATLAB
ortamında değerlendirilip, çalışma koşulları belirlenmiştir ve makine öğrenmesi algoritmalarına dayalı bir sıcaklık etkisi
kompenzasyon modeli önerilmiştir. Bu yöntem, dönüştürücü elemanlarının zaman ekseni bilgilerini dikkate almaktadır. Gerçek akış
hızını tahmin etmek için her deney sıcaklık değeri ve Uçuş Süresi (TOF) sinyallerinin ortalama değ erleri dikkate alınmaktadır.
Böylece, sıcaklık değişimi ve akış ölçümü arasındaki ilişkiyi oluşturmak için makine öğrenmesi algoritmalarından doğrusal
regresyon, destek vektör regresyonu (SVR), Gaussian süreç regresyonu (GPR) ve yapay sinir ağları (YSA) ku llanılmıştır. Önerilen
modelin kompenzasyon performansı 𝑅 2 , ortalama kare-kök hata (𝑅𝑀𝑆𝐸 ), ortalama mutlak hata (𝑀𝐴𝐸) ve ortalama kare hata (𝑀𝑆𝐸),
gibi hata metriklerinin hesaplanması ile incelenmiştir. Sonuçlara göre, YSA tabanlı kompenzasyon a lgoritmasının 𝑅 2 = 0.95 metriği
ile en iyi sonucu verdiği görülmüştür.

Anahtar Kelimeler: Ultrasonik transduser, Akış ölçümü, Kompanzasyon, Uçuş süresi ölçümü.

measurements, Time Of Flight (TOF) values representing the

1. Introduction time domain features of stream signals and actual flow
measurements are collected using a data acquisition card. During
Parameter change caused by temperature effect is a widely the test process, data are collected for different water
countered disturbance phenomenon in the sensor and temperatures (10𝑜 𝐶, 20𝑜 𝐶, 30𝑜 𝐶, 40𝑜 𝐶, 50𝑜 𝐶) and 13 flow
measurement fields including biomedical (Sarjova et al., 2005), rates between 16 L/h and 5000 L/h.
process control (Mehta et al., 2022), measurement devices (Fang
Remaining parts of the paper is organized as follows: In
et al., 2022) and embedded system designs (Rudnicki, 2020).
Transducers are commonly used cheap and easy to use sensor Section 2. transducer element and its mathematical dynamics are
explained. Problem is detailly defined, collected data are given
elements to measure distance (Balasubramanian et al., 2022),
and proposed compensation method is explained. In Section 3.
liquid flow (Yao et al., 2021), gas flow (Chen et al., 2021) and
pressure (MacAskill et al., 2021). However, change of the results of the experiment and compensation work are
presented. Conclusion and the future work of the study are
temperature has a significant role on transducer electrical
presented in the Section 4.
characteristics which causes false sensor readings and
measurement errors (Zibitsker et al., 2021). Calibration process,
due to the underlying problem, is underwhelmed to converge to 2. Material and Method
actual measurement. Identifying the behavior of the problem and
developing a method to eliminate temperature effects is a 2.1. Ultrasonic Transducer and Flow Metering
crucially significant for the sake of true measurement process. Ultrasonic transducers are piezoelectric components used to
Related literature is investigated to study on different generate and/or receive the ultrasonic sound waves (Jaffe et al.,
perspectives on the problem. Huang and Young (2009) employed 1965). Pulsed ultrasonic transducers use electrical energy to
an external sensor to measure the temperature of a distance generate ultrasonic wave trains into the water medium. Reflected
measurement system to compensate the ultrasound velocity waves, which might be called echo, are transformed into
during the measurement process. Wang and Zhang (2010) electrical energy back by the ultrasonic transducers. The total
proposed to use a neural network model which considers the time during the transmission and reflection is considered to
temperatre sensor data and ultrasonic flow measurements to both compute distance or depth in a water medium. Based upon this
calibrate and compensate the measurements, and proved that the principle, ultrasonic transducers might be employed for flow
measurement error decreased to 3% from 5.2%. Scale transform measurement process. Initially, reciprocal transducers are
and cross-correlation methods are employed by Harley and located on a body (or transducer paths are connected with the
Moura (2012) to find phase delay caused by the temperature mirror reflecting the waves) to generate and receive waves
variations. Herein, they can find the optimal time domain (Figure 1). Transducers are pulsed with a pre-determined
information of the ultrasonic waves. A methodology, that aims to specific period of time. Received signals by the transducers are
model temperature effects on signal amplitute and waveform of employed to compute wave transmission time measure Time-of-
the ultrasound to understand that how temperature affects the Flight (TOF). These waves are called Upstream (In the direction
measurement, is proposed by Jia et al. (2021). Huang et al. of flow ) and/or Downstream (In the reverse direction of flow)
(2021) employed the transducers to predict temperature value of signals. This transmission/receiving operation might be
a specific medium where the ultrasound velocity is known and performed with an ultrasonic Time-to-Digital Conversion (TDC)
used to extract temperature. integrated circuit chip. TOF difference value, which might be
In this study, transducer complex dynamics are investigated computed considering the time difference between downstream
under certain temperature and flow conditions with a data and upstream signals, is considered to measure both direction
acquisition process. The problems defined in the literature focus and amplitute of the flow with an offset compensation and
on the temperature change of the water however we focus on the calibration process.
temperature effect on the transducer and its electrical
characteristics. Upstream and downstream signals, which Transducer A Transducer B
Mirror
represents the electrical signals on transducers caused by
ultrasonic wave transmitted from other transducer, might be
evaluated to determine the working conditions. Thus, Flow L L
D

transducers are placed inside of a brass flow meter body to

measure the stream signals during the flow. Through the Mirror Mirror

instrument of heat test bench, flow and temperature test

conditions are satisfied. Flow measurements, temperature Figure 1. Visual representation of an ultrasonic flow meter.

e-ISSN: 2148-2683 114

 European Journal of Science and Technology

𝑡(𝐵>𝐴) ∗ 𝑡( 𝐴>𝐵)
2.2. Time of Flight Measurement and Data 4(𝐿 + 𝐷) 2
Collection =
(𝐶0 − 𝑣𝑐𝑜𝑠(𝛼))(𝐶0 + 𝑣𝑐𝑜𝑠(𝛼)) (4)
TOF of an ultrasonic wave in the direction of downstream 4𝐷𝑐𝑜𝑠 (𝛼)(𝐿 − 𝐶0𝐿 − 𝐷𝑣𝑐𝑜𝑠(𝛼))
might be given as: + 2
𝐶0 (𝐶0 − 𝑣𝑐𝑜𝑠(𝛼))(𝐶0 + 𝑣𝑐𝑜𝑠(𝛼))
𝐷 𝐿 𝐿 𝐷 (1)
𝑡(𝐵>𝐴) = + + +
𝐶0 𝐶0 + 𝑣𝑐𝑜𝑠(𝛼) 𝐶0 + 𝑣𝑐𝑜𝑠(𝛼) 𝐶0
Left side of the Eq. (4) may be assumed as zero, and substituting
(4) in (3) for 𝐶02 − 𝑣 2 cos 2(𝛼) gives temperature independent
where 𝐿 is the distance between transducers, 𝐷 is the diameter of
flow velocity as:
the pipe, 𝛼 is the degree between pipe and mirror, 𝐶0 is the speed
of ultrasonic wave in water, 𝑣 is the flow velocity, and 𝑡(𝐵 >𝐴) is
Δ𝑇 (𝐿 + 𝐷) 2 (5)
the downstream TOF value. In the same manner, upstream TOF 𝑣= ∗
migh be calculated as: 𝑡(𝐵>𝐴) ∗ 𝑡(𝐴>𝐵) 𝐿𝑐𝑜𝑠(𝛼)

𝐷 𝐿 𝐿 𝐷 (2) Due to the flow calculation is not dependent on 𝐶0

𝑡(𝐴>𝐵) = + + + variations, measurement is not affected by temperature variation.
𝐶0 𝐶0 − 𝑣𝑐𝑜𝑠(𝛼) 𝐶0 − 𝑣𝑐𝑜𝑠(𝛼) 𝐶0
However, the aim of this study is to eliminate temperature
effects on transducer electrical characteristics. To understand the
temperature related measurement behaviors and dynamics of the
transducer, flow measurements are performed for different
temperatures and velocities (Figure 2). TOF values for both

upstream and downstream, temperature readings, flow

measurements and actual flow velocities are recorded to perform
a calibration process which eliminates the temperature variation
effects.

where 𝑡(𝐵>𝐴) is the upstream TOF value. Due to the TOF value
is reverse proportional to the flow velocity, TOF Difference
(TOF DIFF) value migh be computed as:

Δ𝑇 = 𝑡(𝐵>𝐴) − 𝑡(𝐴>𝐵)
2𝐿 2𝐿
= −
𝐶0 − 𝑣𝑐𝑜𝑠(𝛼) 𝐶0 + 𝑣𝑐𝑜𝑠(𝛼) (3)
4𝐿𝑣𝑐𝑜𝑠(𝛼)
= 2
𝐶0 − 𝑣 2 cos 2(𝛼)

Computing the flow velocity is directly affected by the

change of water temperature due to the sound velocity 𝐶0 is
affected by temperature. To eliminate this dependency:

e-ISSN: 2148-2683 115

 Avrupa Bilim ve Teknoloji Dergisi

Figure 2. TOF signal evaluation of collected data during experiments.

2.3. Temperature Compensation Process 4𝐿 (7)

𝐶0 =
Temperature parameter must be know to eliminate its effects
𝑡(𝐵>𝐴) + 𝑡(𝐴>𝐵)
on transducer. Due to the 𝐶0 parameter has a known correlation
with temperature, it migh be employed to estimate temperature Proposed compensation method is based on a machine
value. 𝐶0 might be computed as: learning model which considers the water temperature correlated
parameter 𝐶0, flow measurement 𝑣 to estimate actual velocity 𝑣̂
4𝐿𝐶0 (6) (Figure 3).
𝑡(𝐵>𝐴) + 𝑡(𝐴>𝐵) =
𝐶02 − 𝑣 2 cos 2(𝛼)

due to the 𝐶02 ≫ 𝑣 2 cos 2(𝛼) , equation (6) becomes:

e-ISSN: 2148-2683 116

 European Journal of Science and Technology

Machine
Learning based
In this framework, probabilistic approach to regression
Compensation between input-output relation might be defined as:
Block
𝑦̂ = 𝑓(𝑥) + 𝜖 (11)
Figure 3. Block representation of proposed machine learning
model.
where 𝜖 term represents a Gaussian distribution.
To perform this process, machine learning models are employed. Fundamentals of ANN is based on learning brain neuron
Linear regression is method which defines a linear mathematical cells, and hypostatized in machine learning framework (Eskov et
expression between dependent and independent variables to al., 2019). This model consists of an input layer, hidden layers,
perform a prediction process of the independent variable an output layer and processing elements known as neurons. Each
(Weisberg, 2005). This model might be defined as: neuron node receives its input from previous neuron nodes.
Neurons passes linearly weighted sum of the signal to another
𝑦̂ = 𝜃 𝑇 𝑋 (9) neuron over an activation function. Activation functions within
the hidden layers gain the model its dynamics, and are selected
where 𝑦̂ is the model prediction, 𝜃 represents model regression considering the complexity of the dataset. Multi-Layer
parameters, and 𝑋 represents model inputs. Perceptron (MLP) type of ANN has known number of input and
output layer neurons, and equals to the number of independent
Support Vector Regression (SVR) method employees a and dependent variables, respectively. MLP might be defined as:
mapping kernel function which projects feature space into a
higher dimensional hyperplane (Awad and Khanna, 2015). The (12)
objective of constructing a Support Vector Machine (SVM) is to 𝑦̂ = ℎ 𝑜 (𝑊𝑜𝑇 ∗ ℎ 𝑖 (𝑊𝑖𝑛𝑇 ∗ 𝑋))
map features into a higher dimensional space (𝐹) by employing
a kernel function. Estimation function of a general SVR might where 𝑊𝑖𝑛 represents the hidden-layer neuron parameters, ℎ𝑖 (∙ )
be defined as: represents the hidden-layer activation function, 𝑊𝑜 represents
the output-layer neuron parameters, and ℎ𝑜 (∙ ) is the output-layer
𝑦̂ = 𝛼𝑖 𝐾(𝑥𝑖 ,𝑥𝑗 ) + 𝑏 (10) activation function. In this work, a single hidden-layer MLP type
ANN is employed for compensation process.
where 𝛼𝑖 represents the support vectors with 𝑖 samples, and 𝑏
represents the bias term. Mapping function 𝐾 is employed as 3. Results and Discussion
linear, quadratic and cubic, respectively.
MATLAB environment is employed to study on the
Gaussian process regression is a non-parametric regression collected dataset. Firstly, 𝐶0 and 𝑣 values are computed for each
method based on optimizing the distribution kernel function temperature value. Linear regression, SVR, GPR and ANN are
hyper parameters (Wilson et al., 2011). Kernel function with employed to perform compensation process. Model
optimal parameters define a regression fit which maximize the performances are investigated by considering the metrics
negative log-marginal-likelihood (NLML) of the training set. 𝑅 2 , 𝑅𝑀𝑆𝐸 , 𝑀𝑆𝐸, 𝑀𝐴𝐸 values (Table 1).

Table 1. Model performance metric evaluations.

Linear Linear Kernel Quadratic Kernel Cubic Kernel SVR GPR ANN
Regression SVR SVR

𝑅2 0.82 0.81 0.81 0.84 0.92 0.95

𝑅𝑀𝑆𝐸 0.09876 0.1009 0.1007 0.0932 0.0374 0.0707

𝑀𝑆𝐸 0.0097 0.0101 0.0101 0.0086 0.0014 0.0050

𝑀𝐴𝐸 0.0414 0.0315 0.0352 0.0220 0.0076 0.0065

e-ISSN: 2148-2683 117

 Avrupa Bilim ve Teknoloji Dergisi

According to the results, linear models have close MacAskill, W., Hoffman, B., Johnson, M. A., Sharpe, G. R., &
performance in terms of 𝑅 2 and error metrics. Although the Mills, D. E. (2021). Pressure measurement characteristics of
quadratic kernel SVR is a nonlinear model, it is understood that a micro‐transducer and balloon catheters. Physiological
it provides close performance to linear models in terms of 𝑅 2 Reports, 9(8), e14831.
and error metrics. Cubic kernel SVR gives slightly better Zibitsker, A., Berreby, M., Michaels, D., Shilav, R., & Frisman,
performance on 𝑅 2 , 𝑅𝑀𝑆𝐸 and 𝑀𝑆𝐸 metrics, but it is successful I. (2021). Ultrasonic Temperature Compensating Method for
in terms of 𝑀𝐴𝐸 metric compare to previous models. Thus, it is Tracking Decomposition Front in Silica-Phenolic Thermal
observed that solving the problem with quadratic or qubic Protection Material. Journal of Thermophysics and Heat
approaches is not appropriate in terms of a trade-off between Transfer, 35(4), 770-787.
model complexity and model performance. GPR, as a Huang, Y. S., & Young, M. S. (2009). An accurate ultrasonic
probablistic distribution based model, gives the best distance measurement system with self temperature
performance in terms of 𝑅𝑀𝑆𝐸 and 𝑀𝑆𝐸 metrics due to its compensation. Instrumentation Science and Technology,
stochastic behavior. ANN, which gives the best performance in 37(1), 124-133.
terms of 𝑅 2 , also is successful in terms of 𝑀𝐴𝐸 compare to all Wang, Y. X., Li, Z. H., & Zhang, T. H. (2010, October).
proposed models on predictin the actual flow velocity. Research of ultrasonic flow measurement and temperature
compensation system based on neural network. In 2010
International Conference on Artificial Intelligence and
4. Conclusions and Recommendations Computational Intelligence (Vol. 1, pp. 268-271). IEEE.
This study presents a machine learning based calibration Harley, J. B., & Moura, J. M. (2012). Scale transform signal
method for temperature effects on transducer dynamics on flow processing for optimal ultrasonic temperature compensation.
measurement devices. Herein, the problem is identified and IEEE transactions on ultrasonics, ferroelectrics, and
related works are given. Initially, transducer component and frequency control, 59(10), 2226-2236.
flow measurement process are defined. The data collection Jia, Y., Wu, T., Dou, P., & Yu, M. (2021). Temperature
process for understanding the nature of the problem and adapting compensation strategy for ultrasonic-based measurement of
the proposed method parameters are explained. Afterward, oil film thickness. Wear, 476, 203640.
temperature effect and proposed method are explained. Machine Huang, J., Cegla, F., Wickenden, A., & Coomber, M. (2021).
learning models and machine learning based compensation Simultaneous measurements of temperature and viscosity
process are detailly explained. Finally, model performances on for viscous fluids using an ultrasonic waveguide. Sensors,
compensation process are investigated via performance indexes. 21(16), 5543.
According to the results, neural network based compensation Jaffe, H., & Berlincourt, D. A. (1965). Piezoelectric transducer
block gives the better performance compare to other models in materials. Proceedings of the IEEE, 53(10), 1372-1386.
terms of performance metrix. In the future direction of the study, Weisberg, S. (2005). Applied linear regression (Vol. 528). John
a reinforcement learning based compensation method may be Wiley & Sons.
considered to eliminate disturbance effects. Awad, M., & Khanna, R. (2015). Support vector regression. In
Efficient learning machines (pp. 67-80). Apress, Berkeley,
References CA.
Wilson, A. G., Knowles, D. A., & Ghahramani, Z. (2011).
Sorvoja, H., Kokko, V. M., Myllyla, R., & Miettinen, J. (2005). Gaussian process regression networks. arXiv preprint
Use of EMFi as a blood pressure pulse transducer. IEEE arXiv:1110.4411.
transactions on instrumentation and measurement, 54(6), Eskov, V. M., Pyatin, V. F., Eskov, V. V., & Ilyashenko, L. K.
2505-2512. (2019). The heuristic work of the brain and artificial neural
Mehta, Y., Bhargav, V., & Kumar, R. (2022). Characterization networks. Biophysics, 64(2), 293-299.
and Control of High Temperature Impinging Jet Issued from
a Mach 4 Rocket Nozzle. In AIAA SCITECH 2022 Forum
(p. 0124).
Fang, L., Ma, X., Zhao, J., Faraj, Y., Wei, Z., & Zhu, Y. (2022).
Development of a high-precision and wide-range ultrasonic
water meter. Flow Measurement and Instrumentation,
102118.
Rudnicki, T. (2020). Measurement of the PMSM Current with a
Current Transducer with DSP and FPGA. Energies, 13(1),
209.
Balasubramanian, A. B., Sastry, K. V., Magee, D. P., & Taylor,
D. G. (2022). Transmitter and Receiver Enhancements for
Ultrasonic Distance Sensing Systems . IEEE Sensors
Journal.
Yao, S., Yang, M., Zhang, P., Zhang, K., Fang, J., Huang, J., ... &
Zhao, Y. (2021). A Small Diameter Ultrasonic Water Meter
With Self-Diagnosis Function and Self-Adaptive
Technology. IEEE Access, 9, 80703-80715.
Chen, D., Cao, H., & Cui, B. (2021). Study on flow field and
measurement characteristics of a small-bore ultrasonic gas
flow meter. Measurement and Control, 54(5-6), 554-564.

e-ISSN: 2148-2683 118