Ejection state prediction for a pneumatic micro droplet ...
Transcript of Ejection state prediction for a pneumatic micro droplet ...
Bulletin of the JSME
Journal of Advanced Mechanical Design, Systems, and ManufacturingVol.14, No.1, 2020
Paper No.19-00268© 2020 The Japan Society of Mechanical Engineers[DOI: 10.1299/jamdsm.2020jamdsm0001]
Ejection state prediction for a pneumatic micro-droplet
generator by BP neural networks
Fei WANG*, Weijie BAO*,**, Yiwei WANG*,**, Xiaoyi WANG*,***, Keyan REN*, Zhihai WANG*,**
and Jiangeng LI* * Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China
E-mail: [email protected] ** Key Laboratory of Optoelectronics Technology, Beijing University of Technology, Beijing 100124, China
*** Beijing Engineering Research Center for IoT software and Systems, Beijing University of Technology, Beijing 100124, China
Received: 23 May 2019; Revised: 23 October 2019; Accepted: 6 December 2019
1. Introduction
Micro-droplet ejection technology, beside of being used in inkjet printing, is also widely used in many fields such as printed electronics (Raut, et al., 2018), and 3D additive manufacturing (Hoath, 2018). Micro-droplet ejection is related to micro-sample dispensing technique and is used in the field of combinatorial chemistry and biomedicine (Basaran, et al, 2013). In recent years, biological cell printing technology, also based on micro-droplet ejection, shows potentials to solve problems that are difficult in traditional tissue engineering. People have achieved a variety of cell printing by modifying commercially available printers (thermal, piezoelectric, and electrostatic types, etc.) (Gudupati, et al., 2016), but still suffered from cell precipitation, clogging of nozzle, reservoir cleaning difficulties etc. Some companies (Ex. Microfab) have developed piezoelectric micro-droplet generator that are specifically designed for cell printing (Christensen, et al., 2015) Micro-droplet generators based on micro-valves are also widely used for cell printing (Ng, et al., 2017). However, at the nozzle, the bio-ink is impacted repeatedly by a tiny plunger (driven by a solenoid actuator) at high speed, which may cause potential damage to the cells. In our previous experiment, a home-build pneumatic micro-droplet generator (shown in Fig. 1(a)) was tested. Unlike the micro-valve systems, here, the high-speed solenoid valve stays outside of the reservoir, so there is no moving part in contact with the liquid. The high-speed solenoid valve is briefly turned on, and the high-pressure gas enters the reservoir. A pressure pulse P(t) is generated in the reservoir, forcing the liquid out through a nozzle to form a micro-droplet. The pneumatic micro-droplet generator is simple to operate, good for liquids of wide ranges of viscosity (Goghari, et al., 2008) and temperature (Luo, et al., 2012). Initially used for cell printing, it achieved nearly 100% cell viability (Wang, et al., 2018a), making it a potential choice for dispensing cell laden biomedical samples.
1
Abstract
Micro-droplet generation is related to liquid dispensing technology that has potential applications in many fields.
Specifically, pneumatic micro-droplet generation is controlled by a solenoid valve being briefly turned on, so
that high pressure gas enters the liquid reservoir, forming a gas pressure pulse waveform , forcing the liquid
out through a tiny nozzle to form a micro-droplet. For each ejection, is acquired by a high speed pressure
sensor, and the ejection state is obtained by machine vision methods. A prediction model based on BP neural
network is established, with as input and the droplet ejection state as output. Experiments show that the BP
neural network can predict the number of droplets with an accuracy higher than 99%. It is also shown that the
BP neural network can improve the prediction accuracy for the position of droplets relative to the nozzle, at a
given moment. Under typical working conditions, is not consistent. As a result, the ejection state is not
consistent either. These prediction models may be used for real time monitoring and control of the pneumatic
micro-droplet generator.
Keywords : Micro-droplet, Pneumatic, BP neural network, Ejection state, Machine vision
2© 2020 The Japan Society of Mechanical Engineers[DOI: 10.1299/jamdsm.2020jamdsm0001]
Fei Wang, Bao, Yiwei Wang, Xiaoyi Wang, Ren, Zhihai Wang and Li, Journal of Advanced Mechanical Design, Systems, and Manufacturing, Vol.14, No.1 (2020)
For commercial printers, inks are uniform and stable. Therefore, the printing parameters can be fixed; and the ejection
state is highly consistent over a long period of time. But for the pneumatic micro-droplet generator, as will be described
later, the pressure pulse waveform is far from being consistent. Therefore, for each pressure pulse, the state the
droplet ejection (including number of droplets, break time and position of the liquid flow, initial speed of the droplet,
etc.) is not consistent either. In addition, for many situations, the ink is non-standard and its flow properties may even
change over time, making the situation more complex (For example, biological cells contained in the bio-ink tends to
precipitate, leading to changes of ink composition and its flow properties. As another example, sodium alginate, a popular
bio-ink material, may slowly react with calcium ions in the ink, resulting in changes of flow properties over time.).
Therefore, real-time monitoring and control of the micro-droplet ejection is necessary.
Pneumatic micro-droplet ejection is a complex multi-phase flow process driven by the pressure pulse . And the
formation of the droplet needs considering instability caused by the surface tension effects (Eggers, et al., 2008).
Theoretically, it is necessary to solve the Navier-Stokes equation and the level set (or phase field) equation in
combination. Although the finite element method can give a result, the simulation process is time-consuming, and its
reliability still need to be verified. High speed photography is the most direct and effective method to study the process
of micro-droplet ejection (Pongrác, et al., 2014). But the experimental equipment is expensive, and real-time processing
of a large amount of images is not so realistic. Optical measurements, such as light scattering, can achieve high-speed
measurement of some geometrical parameters (such as the droplet diameter) (Gao, et al.,2016) , but the system needs to
work in the dark, which limits its practicability. A predictive model through some readily available experimental data
may be helpful in solving the problem. With the development of machine learning, some scholars have begun to apply it
to the studies of complex fluids. A machine learning algorithm is used to construct a complex relationship between flow
/ energy distributions and driving / boundary conditions, for the flow field monitoring around the boundary layer (Debien,
et al., 2016). In the field of micro-droplet generation, a neural network was used to predict the droplet sizes in a
microfluidic system (Mahdi, et al., 2016). Recently, machine learning was applied to drop-on-demand systems for
optimizing their operation parameters (Wang, et al. 2018b; Shi, et al., 2018). In both works, neural networks are used to
established corresponding relation among main operation parameters (including voltage amplitude and pulse width),
liquid parameters (viscosity and surface tension) and ejection parameters (number of droplet etc.). For the pneumatic
ejection, as will be described later, the opening and closing of the high speed solenoid valve shows large inconsistency,
especially for ejection at higher rate. Therefore, even with operation parameters fixed, the ejection may still be quite
different. As the pressure pulse waveform in the reservoir is more directly related to the droplet ejection state, it is
proposed in this paper to use deep neural network training to obtain a nonlinear prediction model with as the input
and the characteristics of the ejection state as the output (Bengio, et al. ,2016).
Fig. 1 (a) photo of the home-build pneumatic micro-droplet generator, (b) Schematic representation of a pneumatic micro-
droplet generator. It consist of a reservoir (1), a nozzle (2), a high speed solenoid valve (3), a pressure regulator (4),
a T-joint (5), a venting tube (6). The micro-droplet ejection is monitored by a camera (7) with LED illumination (8).
The gas pressure in the reservoir is monitored by a high speed pressure sensor (9). The micro-droplet ejection and
monitoring are automatically operated by a controller (10).
2
2© 2020 The Japan Society of Mechanical Engineers[DOI: 10.1299/jamdsm.2020jamdsm0001]
Fei Wang, Bao, Yiwei Wang, Xiaoyi Wang, Ren, Zhihai Wang and Li, Journal of Advanced Mechanical Design, Systems, and Manufacturing, Vol.14, No.1 (2020)
The pressure pulse waveform can be collected in real time and can be analyzed quickly. The establishment of
such prediction model may have two potential applications. Firstly, by direct predicting the ejection state through the
pressure waveform, it can replace the machine vision based monitoring method to some extent. Secondly, failure of the
prediction model may be a more sensitive indication that the flow properties of the ink have changed.
2. Pneumatic micro-droplet generator and its working principles
Referring to the previous literature (Cheng, et al., 2003), the home-build pneumatic micro-droplet generator is shown
schematically in Fig. 1(b). The setup includes a micro-droplet ejection system and an attached monitoring module. The
ejection system includes a liquid reservoir, a nozzle, a high-speed solenoid valve, a pressure regulator, and air path
(mainly, a venting tube and a T-joint); the monitoring module includes an industrial CCD Camera with LED illumination
for machine vision monitoring, and a high-speed pressure sensor for monitoring the gas pressure in the reservoir. Both
the ejection system and the monitoring module are operated automatically by a controller.
2.1 The Generation of Pressure Pulse P(t) in the Reservoir
The generation of the pressure pulse in the reservoir may be understood by making an electro-acoustic analogy
(Kinsler, et al., 1999). Specifically, the pressure in the reservoir in Fig. 2(a) and the voltage across the capacitor in the
RLC circuit shown in Fig. 2(b) satisfy the same differential equation, where capacitance, inductance, and electric
resistance are the electric counterparts for compliance, inertance, and acoustic resistance. The ideal switch in series of a
resistance is equivalent to the solenoid valve, the constant voltage source corresponds to the constant pressure at the
front end of the solenoid valve. The values in Fig. 2(b) are estimated compliance, inertance, and acoustic resistance in SI
units. In the equivalent electric circuit, when the switch is turned on, the voltage source charges the capacitor that is
shunted by the inductor and the resistor. After the switch is turned off, the capacitor is discharged through the inductor
and the resistor. Through a simulation (via Multisim) of the charging and discharging process, the voltage across the
capacitance, or equivalently the pressure inside the reservoir is calculated.
Fig. 2 (a) the micro-droplet generator comprising mainly the solenoid valve, the reservoir, and the venting tube, where “V”
is the volume of gas (volume above the liquid). (b): acoustic-electric analogy, in which compliance corresponds to
capacitance, inertance corresponds to inductance, acoustic resistance corresponds to electric resistance. The ideal
switch in series with an electric resistance is equivalent to the solenoid valve, the constant voltage source corresponds
to the constant pressure at the front end of the solenoid valve. (c) and (d) are measured and simulated , where
the solenoid valve is kept “ON” for 10 ms.
3
2© 2020 The Japan Society of Mechanical Engineers[DOI: 10.1299/jamdsm.2020jamdsm0001]
Fei Wang, Bao, Yiwei Wang, Xiaoyi Wang, Ren, Zhihai Wang and Li, Journal of Advanced Mechanical Design, Systems, and Manufacturing, Vol.14, No.1 (2020)
By comparing the actual pressure waveform with the simulation result, as shown respectively in Fig. 2(c) and (d),
the effectiveness of the electro-acoustic analogy model is verified. Based on the above analysis, ∆t, as marked in Fig.
2(c) and (d), is roughly the period of time that the solenoid valve is “ON”. The amplitude of the waveform is largely
determined by . The pressure oscillation is associated with the well-known Helmholtz cavity, with its frequency
determined by the volume of gas in the reservoir (“V” in Fig. 2(a)), and the geometry of the venting tube, and has been
discussed in many literatures. Fluctuation of the and is very obvious in practical experiments. Meanwhile, the
geometry of the Helmholtz cavity changes versus shift of liquid level in the reservoir.
2.2. Measuring Pressure Pulse Waveform P(t) in the Reservoir
A high-speed pressure sensor (acquisition frequency 1MHz, measuring range -20kPa~+20kPa) is used to collect the
gas pressure waveform in the reservoir. A virtual oscilloscope (sampling frequency 25kHz) simultaneously
acquires the analog output signal of the pressure sensor and a synchronization signal generated by the controller. The
rising edge of the synchronization signal is used as trigger for the high speed solenoid valve. Fig. 3(a) shows the GUI of
the virtual oscilloscope, where both the pressure waveform (upper curve) and the synchronization signal (lower
curve) are displayed.
Starting from the rising edge of the synchronization signal, corresponding pressure waveform is recorded as ,
at 256 discrete moments (i = 0 - 255, or 0 - 10.2 ms). After that, the pressure is negligibly small. By smoothing, high
quality pressure waveform data are obtained. Figure 3(b) shows the pressure waveforms of two measurements for the
same operation parameters. Although they look similar, the distance between the generated droplet and the nozzle at
certain delay (defined as in Fig. 4) is not consistent. The vertical solid line in the graph will be described latter.
2.3. Machine-vision Based Ejection State Monitoring
The machine vision monitoring uses a monochromatic industrial CCD camera (Daheng Optoelectronics, Model
MER-125-30UM) with a resolution of . For the imaging optics, a Navitar 6.5× zoom lenses (magnification of
0.7 - 4.5×) is used. As the information obtained from the machine vision method is sensitive to the magnification, the
imaging system is always calibrated based on the fact that the outer diameter of the nozzle is 600 μm. In the rest of this
paper, one pixel corresponds to 20 μm. The camera is triggered relative to the synchronization signal, via a delay function.
Therefore, the acquired pressure waveforms are in one-to-one correspondence with the captured droplet images.
The image processing is shown in Fig. 4. Usually, the nozzle and droplet only occupy a small region of the whole
image. Therefore, the region of interest (ROI) of the original image is specified first to improve image processing
efficiency (as in Fig. 4(a)). In the second step, binarization of the ROI is performed by ultilizing an automatic threshold
segmentation method (Otsu algorithm) for grayscale images (as in Fig. 4(b)). Due to refraction effect, there is tiny dark
Fig. 3 (a) the pressure waveform, together with the synchronization signal, acquired by a virtual oscilloscope (b) two
truncated pressure waveforms (after smoothing). The generated droplets shows different (displacement of
droplet relative to the nozzle at certain delay, see Fig.4), 133 pixels (red) and 196 pixels (black) respectively.
4
2© 2020 The Japan Society of Mechanical Engineers[DOI: 10.1299/jamdsm.2020jamdsm0001]
Fei Wang, Bao, Yiwei Wang, Xiaoyi Wang, Ren, Zhihai Wang and Li, Journal of Advanced Mechanical Design, Systems, and Manufacturing, Vol.14, No.1 (2020)
region within the droplet. For not to affect the subsequent extraction of the droplet information, in the 3rd step, an image
filling algorithm may be used to have the dark region removed, as shown in Fig. 4(c). The complete droplet area is then
labeled by a connected region analysis. On this basis, the number 𝑁𝑑 of droplets, and the geometrical parameters of
each droplet, including the centroid coordinates ( ), the size of the droplet, are obtained, shown in Fig. 4(d).
Typical size of the droplet is 40 pixel (800 μm ) in diameter. With the coordinate of the droplet, 𝐻𝑑 is defined as its
displacement relative to the nozzle, also shown in the figure.
3. Prediction of ejection state from pressure waveform P(t) via BP neural networks
3.1. Micro-droplet Experiment and the Taking of the “Big Data”
The ejection experiment is performed with DI water, using a 300 μm internal diameter nozzle, with ejection frequency
set to 20 Hz. A rather diverse set of ejection parameters are used. Specifically, varies in the range of 1.5 - 3 ms;
solenoid valve front end pressure 𝑃0 is tuned in the range of 40 - 60 kPa. The liquid level changes during the experiment,
so the oscillation period, as seen in Fig. 2(c) and (d), varies from 3 to 4 ms. This produces rich varieties of pressure pulse
waveforms and the corresponding ejection states, serving as the “Big Data” for training of the neural networks.
Each sample of database consists of a waveform, as well as parameters for characterizing the ejection state.
In this paper, two parameters are taken, the number of the droplets ejected, referred as ; the location of the droplet
relative to the nozzle at a given moment, referred as 𝐻𝑑 . It is found that the droplets appear 3 - 4 ms after the
synchronization signal. After detached from the nozzle, the droplet should not be affected much by . Based on this
prior knowledge, valid portion of can be reduced to < 4.7 ms (or < 119), the region on the left side of the
solid line in Fig. 3(b).
The data should be normalized in favor of numerically stable training for the neural networks. Traditionally, the
samples of pressure are normalized independently, as shown by Eq. (1), where s is the sample number, and is the
infinity norm operator.
(1)
However, since the magnitude of the press plays a key role in the generation of the droplets, we should preserve this
information by normalizing the samples globally, which is shown by Eq. (2).
(2)
For data such as droplet number 𝑁𝑑 and location 𝐻𝑑, the normalization is done via the following Eq. (3)
(3)
Fig. 4 Machine vision processing steps for acquiring the micro-droplet ejection states, including positions, sizes, and
numbers of the micro-droplets.
5
2© 2020 The Japan Society of Mechanical Engineers[DOI: 10.1299/jamdsm.2020jamdsm0001]
Fei Wang, Bao, Yiwei Wang, Xiaoyi Wang, Ren, Zhihai Wang and Li, Journal of Advanced Mechanical Design, Systems, and Manufacturing, Vol.14, No.1 (2020)
It should be mentioned that, the output of the neural network need to be de-normalized to produce the final prediction.
3.2. Principle of BP Neural Network
Neurons, the basic units of human brain, form network, where each neuron takes combined signals from some
neurons, and produces an output for some other neurons. An artificial neural network replicates this by having layers of
nodes, with each connected to nodes in the preceding and subsequent layers. The first layer of nodes is the input layer,
simply for taking input signals. For each node in the hidden layers and output layer, as shown in Fig. 5(b), the combined
input is from the nodes in the preceding layer, moderated by link weights that represent possible damping effects
associated with signal transmission between neurons. Then a node bias and an activation function, which mimic threshold
behavior of the neuron, are applied to generate the final output of that node. A neural network with only one hidden layer
is shown in Fig. 5(a). Typically, it is the link weights and node biases that are iteratively refined to give a better and better
neural network.
The BP neural network is a multi-layer feed-forward network trained by the error back propagation algorithm. Its
learning rule is to use the gradient descent method to continuously adjust the link weights and node biases by back
propagation so that the squared error sum between the actual output and the expected output of the neural network is the
smallest. BP neural network has self-learning, self-organization, self-adaptive ability and strong nonlinear mapping
ability, and is often applied to the study of non-deterministic problems with complex causality. Therefore, we use BP
neural network to construct the relationship between the pressure waveform and the droplet ejection state.
3.3. Predicting the Number of Droplets from P(t) Via BP Neural Networks
The number of droplets generated by each ejection is an important parameter. For predicting from , BP
neural networks with one hidden layer are used. There are 120 nodes in the input layer, and one node in the output layer.
The activation functions of the hidden layer and output layer nodes are the “tansig” function.
Fig. 5 (a) A neural network with one hidden layer. The output layer has only one node. The input layer nodes are only for
taking input signals. (b) For the nodes in the hidden layer and the output layer, the combined input is from the nodes
in the preceding layer, moderated by link weights. The combined input is biased, and then a “tansig” activation
function is applied.
6
2© 2020 The Japan Society of Mechanical Engineers[DOI: 10.1299/jamdsm.2020jamdsm0001]
Fei Wang, Bao, Yiwei Wang, Xiaoyi Wang, Ren, Zhihai Wang and Li, Journal of Advanced Mechanical Design, Systems, and Manufacturing, Vol.14, No.1 (2020)
Our database consists of 10000 samples, with either none, one droplet, or two droplets per ejection. 80% of the
samples were randomly selected as the training set, and the remaining 20% of the samples are left for testing. Fig. 6(a)
shows the predicted droplet number for a neural network with only one node in the hidden layer ( ). To get the
number of droplets, the predicted values need to be rounded off to integers. The prediction is 97.8% correct. For the
network that has two nodes in the hidden layer , the prediction is 99.9% correct, even without performing a
round-off, quite different from the case. Similar prediction accuracy is also realized for larger than two,
as shown in Table 1.
Further studies are conducted on the neural network with . As shown in Table 2, for the three ejection states
(with 0, 1, or 2 droplets), the outputs of the two hidden layer nodes are displayed. For the node in the output layer, its
link weights with the two hidden layer nodes are 1.93 and -4.01 respectively, and its node bias is 2.08. Therefore, the
combined input and hence the output of this node can be calculated. Finally, the predicted number of droplets is obtained
by so-called de-normalization.
Although with “tansig” activation function, the outputs of these hidden layer nodes have only two options, 1 and -1.
Therefore, for the output layer node, there are options of input, hence options of output. In this experiment,
there are only 3 categories of ejection states (0 droplet, 1 droplet, and 2 droplets), which can be well represented by those
options as long as It makes sense that (under the condition of ) the prediction can be accurate even
before round-off to integers.
0 500 1000 1500 2000
0
1
2
(a)
num
ber
of
dro
ple
ts
test set0 500 1000 1500 2000
0
1
2
num
ber
of
dro
ple
ts
test set
(b)
Fig. 6 Prediction of droplet number for a test set (with 2000 samples). (a) neural network with only one hidden layer node,
and (b) neural network with two hidden layer nodes. The blue circle represents the actual number of the droplets,
and the red dot represents the predicted value of the neural network (before round-off to integers). The neural
networks are trained with a training set of 8000 samples.
Table 1 Statistics of prediction for the droplet number per ejection, by neural networks of
different .
Number of Nodes in
the Hidden Layer
(𝒏𝑯)
Statistics of Prediction Results
Correct Prediction Total number of
samples tested Accuracy
1 1956 2000 97.8%
2 1999 2000 99.95%
3 1999 2000 99.95%
4 1998 2000 99.9%
5 1999 2000 99.95%
7
2© 2020 The Japan Society of Mechanical Engineers[DOI: 10.1299/jamdsm.2020jamdsm0001]
Fei Wang, Bao, Yiwei Wang, Xiaoyi Wang, Ren, Zhihai Wang and Li, Journal of Advanced Mechanical Design, Systems, and Manufacturing, Vol.14, No.1 (2020)
3.4. Predicting the droplet location from P(t) via BP neural networks
As described earlier, the pressure waveforms are inconsistent. Therefore, at a given moment (delayed relative to the
synchronization signal, see Fig. 3), the distance 𝐻𝑑 between the generated droplet and the nozzle distributes in a wide
range. In this paper, we focus on the situation of single droplet per ejection. It should be pointed out that 𝐻𝑑 is not the
most suitable parameter to characterize an ejection state, because it is collectively determined by break time and position
of the liquid flow, as well as initial velocity of the droplet. However, measurement of those physical quantities needs
ultrahigh-speed photography and image processing, so is prohibitively expensive. The parameter 𝐻𝑑 becomes the easily
available alternative.
Our database consists of 3000 samples, all with one droplet per ejection. The training set takes 70% of the total
samples (the rest of samples are left for testing). A straightforward and reasonable prediction of 𝐻𝑑 for the test set is the
statistical average of 𝐻𝑑 data for the training set, referred as 𝐻𝑑̅̅ ̅̅ . The prediction errors are shown as blue circles in Fig.
7. And the standard error of predictions reaches 30 pixels.
0 200 400 600-70
-35
0
35
70
Err
or
(pix
el)
test set
Table 2 For the three ejection states (with 0, 1, or 2 droplets), the outputs of the two hidden
layer nodes, the input and output of the output layer node, and the predicted droplet numbers
(before round-off), are displayed.
Target Number of Droplets per Ejection
0 droplet 1 droplet 2 droplets
Outputs of the
Hidden Layer Nodes
Node 1 -1 1 1
Node 2 1 1 -1
Input to the Output Layer Node -3.851 0 8.0128
Output of the Output Layer Node -1 0 1
Predicted Droplet Number (After De-
normalization) 0 1 2
Fig. 7 Prediction errors of droplet positions (relative to the nozzle), by two methods, which are taking statistical average
of droplet positions (blue circles), and using BP neural networks (red stars). The prediction errors are measured in
pixel.
8
2© 2020 The Japan Society of Mechanical Engineers[DOI: 10.1299/jamdsm.2020jamdsm0001]
Fei Wang, Bao, Yiwei Wang, Xiaoyi Wang, Ren, Zhihai Wang and Li, Journal of Advanced Mechanical Design, Systems, and Manufacturing, Vol.14, No.1 (2020)
For predicting from , BP neural networks with one hidden layer are used. There are 120 nodes in the input
layer. There is one node in the output layer because 𝐻𝑑 is the only output. The activation functions of the hidden layer
and output layer nodes are the “tansig” function. As shown in Fig. 7, the red dots indicate the prediction errors based on
waveforms via a neural network with 5 nodes in the hidden layer ( ). It can be seen from the Figure, and
also from Table 3 (see next paragraph) that the neural network improves the prediction accuracy by nearly 3 times
compared with the prediction based on statistical average. Typical neural network performance is shown in Fig. 8.
For networks with different 𝑛𝐻, two groups of training and testing are conducted. In the first group, training set
accounts for 70% of the total samples, and is used to roughly determine the optimal . The standard error of prediction
decreases with 𝑛𝐻, and tends to be stable for 𝑛𝐻 larger than 5. For the second group, the training set accounts for 30%
of the total samples. The comparison of the two groups indicates the neural network prediction model possesses good
generalization ability.
The time domain waveform can be reasonably converted into its spectrum , in the frequency domain.
It is instructive to establish a prediction model for 𝐻𝑑, but with as inputs. A similar BP neural network with one
hidden layer is used. for a typical pressure waveform is shown in Fig. 9. is negligibly small with 𝑖 > 19.
Fig. 8 Typical neural network performance. Specifically in this figure, the best validation performance is 0.028791 at
epoch 17.
Fig. 9 Discrete Fourier transformation (DFT) of a typical , where = , = 0 - 119. The transformation is
performed using FFT technique. Only the amplitude of the DFT is shown in the Figure. The inset shows
the same data of low frequency region.
9
2© 2020 The Japan Society of Mechanical Engineers[DOI: 10.1299/jamdsm.2020jamdsm0001]
Fei Wang, Bao, Yiwei Wang, Xiaoyi Wang, Ren, Zhihai Wang and Li, Journal of Advanced Mechanical Design, Systems, and Manufacturing, Vol.14, No.1 (2020)
So for the first trial, to be safe, the input layer takes the lowest 20 frequency components of . Significantly
improved prediction accuracy, similar to that shown in Fig. 7 is obtained.
Then, as shown in Table 4, BP neural networks with different are examined in two groups of training and testing.
Similar to what we have done earlier, in the first group, the training set accounts for 70% of the total samples, which is
used to roughly determine the optimal . The standard error of the prediction decreases on increasing , reaches the
minimum at = 5. Keep increasing leads to gradual and minor deterioration of prediction accuracy. In the second
group, the training set accounts for 30% of the total samples. Comparison of results from these two groups indicates the
prediction model has good generalization ability.
The neural networks with a hidden layer of 5 nodes will be used for the rest of the paper. Our next goal is to determine
which frequency components are decisive for the prediction of . As shown in Table 5, selected are taken as
inputs to the neural network. Two sets of experiments are carried out, all with the training set accounting for 70% of the
total samples. In the first set, shown in the left part of Table 5, the high frequency components are gradually deleted from
the inputs of the neural network. It is shown that the prediction error is rather small as long as the lowest four frequency
components ( ) are kept as inputs. Drop-off of more low frequency components leads to abrupt increasing
of prediction error.
Table 3 Standard errors of predictions based on neural networks, with as inputs.
Networks with different numbers of hidden layer nodes are examined in two groups of
experiments, with the training sets take 70% and 30% of total samples, respectively.
Number of Hidden Layer
Nodes ( )
Standard Error (in pixel) of the Prediction
70% training set 30% training set
1 12.71 12.78
2 11.75 12.44
3 11.45 12.02
4 11.13 11.49
5 10.79 11.49
6 10.99 11.48
7 10.91 11.65
8 10.61 11.40
9 10.91 11.40
10 10.90 11.57
Table 4 Standard errors of predictions based on neural networks, with as inputs.
Networks with different numbers of hidden layer nodes are examined in two groups of
experiments.
Number of Hidden Layer
Nodes ( )
Standard Error (in pixel) of the Prediction
70% training set 30% training set
1 11.86 12.84
2 11.56 11.71
3 11.07 11.65
4 10.89 11.43
5 10.57 11.35
6 10.84 11.27
7 10.76 11.25
8 10.71 11.48
9 11.00 11.41
10 11.15 11.23
10
2© 2020 The Japan Society of Mechanical Engineers[DOI: 10.1299/jamdsm.2020jamdsm0001]
Fei Wang, Bao, Yiwei Wang, Xiaoyi Wang, Ren, Zhihai Wang and Li, Journal of Advanced Mechanical Design, Systems, and Manufacturing, Vol.14, No.1 (2020)
For the second set of experiments, shown in the right side of Table 5, the prediction accuracy gets worse rapidly as
the lowest frequency components are gradually removed from the neural network. The results of these two sets of
experiments are mutually corroborated. That the lower frequency components are more relevant to is also consistent
with our common knowledge of spectra for acoustic signals. In particular, represents the DC component of the
pressure waveform, which is relevant to the speed at which the liquid is extruded. Therefore, it is not surprised that
should be included for good prediction of . Although not shown here, simulation based on finite element
method (COMSOL 5.3) also shows that the impact of high frequency components (higher than ) on 𝐻𝑑 can be
ignored. In addition, for completeness, standard errors for prediction models based on 1-3 and 2-3 frequency components
are also presented in Table 5.
The pressure waveforms corresponding to different number of droplets are usually quite different in shape. Therefore,
predicting the number of droplets by the pressure waveforms is more like to demonstrate that the BP neural network can
be applied to such nonlinear ejection process. In contrast, the technique of predicting by pressure waveform may
have more practical applications. As described earlier, in biomedical applications, detection of fluid property change is
highly expecting. Taking the viscosity change as an example, as a demonstration of this scheme, glycerin water mixture
is used for different viscosity. Under the same operation parameters ( and ), a data set is collected using 8% glycerin
concentration solution. The BP neural network prediction model and the vision based model are established and
tested (80% of this data set is for training and 20% for testing). The left side of Fig. 10 shows the prediction errors, and
a standard error is obtained, similar to those shown in figure 7 and table 3. Then, glycerin concentration is reduced to
6%, 4% and 2%, the ejection states are collected and used as test data sets. With the operation parameters fixed, it is
found that 𝐻𝑑 increases with the decrease of viscosity (glycerin concentration). The prediction error will get larger for
Table 5 Standard errors of for the neural network prediction, where different frequency
components of spectrum are taken as inputs.
components
(taken as inputs)
Standard
Error
components
(taken as inputs)
Standard
Error
components
(taken as inputs)
Standard
Error
0~19 10.57 0~19 10.57 1~3 13.58
0~9 10.78 1~19 13.65 2~3 15.72
0~4 10.89 2~19 15.68
0~3 10.86 3~19 16.10
0~2 11.65 4~19 17.27
0~1 13.32 9~19 20.58
0 25.42
Fig. 10 Prediction errors of vision based method ( ) in blue and BP neural network model in red, measured for several
glycerin water solutions (with glycerin concentration marked). Both prediction models are obtained, based on the
training data set taken with 8% glycerin solution. The dashed line for each figure defines the region of ,
where σ is standard error, derived from the test data set, also taken with 8% glycerin solution.
11
2© 2020 The Japan Society of Mechanical Engineers[DOI: 10.1299/jamdsm.2020jamdsm0001]
Fei Wang, Bao, Yiwei Wang, Xiaoyi Wang, Ren, Zhihai Wang and Li, Journal of Advanced Mechanical Design, Systems, and Manufacturing, Vol.14, No.1 (2020)
both prediction models, shown in Fig. 10. Once the prediction error (after 10-point moving average) is greater than ,
it is determined that the viscosity of the liquid has changed. Compared with the vision based prediction model, the
standard error σ of the BP neural network prediction model is much smaller. Therefore, based on the above criterion,
failure of the BP neural network model is a more sensitive indication of viscosity change.
4. Conclusions
In summary, prediction of ejection states for a pneumatic micro-droplet generator based on measured pressure
waveform in the reservoir is realized through BP neural networks. The number of droplets , and the position
of droplet (at a certain delay time relative to the synchronization signal) 𝐻𝑑 are important parameters for evaluating the
ejection states. In the first study, a model with two or more nodes in the hidden layer can predict with an accuracy
as high as 99%. In the second study, compared with the prediction solely from statistical average, a neural network
with 5 nodes in the hidden layer, reduces the standard error of the prediction by a factor of 3. The inputs of the neural
network can be the pressure waveform (in the time domain), or this spectrum (in the frequency domain). These prediction
methods based on BP neural networks may be in the future used for monitoring and controlling of the pneumatic micro-
droplet generator.
References
Basaran, O.A., Gao, H. and Bhat, P.P., Nonstandard inkjets, Annual Review of Fluid Mechanics, Vol.45, No.1 (2013),
pp.85-113
Goodfellow, I., Bengio, Y. and Courville, A. Deep Learning (2016), MIT Press
Cheng, S. and Chandra, S. A pneumatic droplet-on-demand generator, Experiments in Fluids, Vol.34, No.6 (2003),
pp.755-762
Christensen, K., Xu, C., Chai, W., Zhang, Z., Fu, J. and Huang, Y, Freeform inkjet printing of cellular structures with
bifurcations, Biotechnology & Bioengineering, Vol.112, No.5 (2015), pp.1047-1055
Debien, A., Von Krbek, K. A., Mazellier, N., Duriez, T., Cordier, L., Noack, B. R. and Kourta, A., t losed-loop separation
control over a sharp edge ramp using genetic programming, Experits in Fluids, Vol.57, No.3 (2016), pp.40
Gao, R. S., Telg, H., McLaughlin, R. J., Ciciora, S. J., Watts, L. A., Richardson, M. S. and Markovic, M. Z.., A light-
weight, high-sensitivity particle spectrometer for PM2.5 aerosol measurements, Aerosol Science and Technology,
Vol.50, No.1 (2016), pp.88-99
Goghari, A.A. and Chandra, S., Producing droplets smaller than the nozzle diameter by using a pneumatic drop-on-
demand droplet generator, Experiments in Fluids, Vol.44, No.1 (2008), pp.105-114
Gudupati, H.; Dey, M. and Ozbolat, I., A comprehensive review on droplet-based bioprinting: past, present and future,
Biomaterials, Vol.102 (2016), pp.20-42
Harris, C.M. Handbook of Acoustical Measurements and Noise Control (1991), McGraw-Hill
Hoath, S.D. Fundamentals of inkjet printing: the science of inkjet and droplets (2016), Wiley
Kinsler, L. E., Frey, A. R., Coppens, A. B. and Sanders, J. V., Fundamentals of acoustics 4th edition (1999), Wiley
Luo, J., Qi, L. H., Zhong, S. Y., Zhou, J. M. and Li, H. J., Printing solder droplets for micro devices packages using
pneumatic drop-on-demand (dod) technique, Journal of Materials Processing Technology, Vol.212, No.10 (2012),
pp.2066-2073
Mahdi, Y. and Daoud, K., Microdroplet size prediction in microfluidic systems via artificial neural network modelling
for water-in-oil emulsion formulation, Journal of Dispersion Science and Technology, Vol.38, No.10 (2016),
pp.1501-1508,
Microfab Technologies Inc. < http://www.microfab.com >, (accessed on 24 April, 2019).
Ng, W. L., Lee, J. M., Yeong, W. Y. and Naing, M. W., Microvalve-based bioprinting-process, bio-inks and applications,
Biomater. Sci, Vol.5, No.4 (2017), pp.632-647
Pongrác, B., Kim, H. H., Janda, M., Martišovitš, V. and Machala, Z., Fast imaging of intermittent electrospraying of
water with positive corona discharge, Journal of Physics D: Applied Physics, Vol.47, No.31 (2014), 315202
Rashid, T. Make Your Own Neural Network (2016), CreateSpace Independent Publishing Platform
Raut, N.C. and Al-Shamery, K., Inkjet printing metals on flexible materials for plastic and paper electronics, Journal of
12
2© 2020 The Japan Society of Mechanical Engineers[DOI: 10.1299/jamdsm.2020jamdsm0001]
Fei Wang, Bao, Yiwei Wang, Xiaoyi Wang, Ren, Zhihai Wang and Li, Journal of Advanced Mechanical Design, Systems, and Manufacturing, Vol.14, No.1 (2020)
Materials Chemistry C, Vol.6, No.7 (2018), pp.1618–1641
Shi, J., Wu, B., Song, B., Song, J., Li, S. and Trau, D., Learning-Based Cell Injection Control for Precise Drop-on-
Demand Cell Printing, Annals of Biomedical Engineering, Vol.46, No.9 (2018), pp.1267-1279
Wang, M.; Wang, C. and Wang, F., Bioprinting: The Influence of Pneumatic Microdrop-on-Demand Printing Process on
Cell Viability, Proceedings of 8th International Conference on Biomedical Engineering and Technology, (2018),
pp.89-95
Wang, T., Kwok, T. H., Zhou, C. and Vader, S., In-situ droplet inspection and closed-loop control system using machine
learning for liquid metal jet printing, Journal of Manufacturing Systems, Vol.47 (2018), pp.83–92
Wei, H. Wei, H., Xiao, X., Yin, Z., Yi, M .and Zou, H., A waveform design method for high DPI piezoelectric inkjet
print-head based on numerical simulation, Microsystem Technologies, Vol.23, No.12 (2017), pp.5365–5373
13