CN112733418A - Method for monitoring fluid characteristic change of ejected liquid in pneumatic droplet ejection process - Google Patents
Method for monitoring fluid characteristic change of ejected liquid in pneumatic droplet ejection process Download PDFInfo
- Publication number
- CN112733418A CN112733418A CN202011357907.5A CN202011357907A CN112733418A CN 112733418 A CN112733418 A CN 112733418A CN 202011357907 A CN202011357907 A CN 202011357907A CN 112733418 A CN112733418 A CN 112733418A
- Authority
- CN
- China
- Prior art keywords
- prediction
- prediction model
- droplet
- droplet ejection
- air pressure
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The invention discloses a method for monitoring the characteristic change of the sprayed liquid fluid in the pneumatic droplet spraying process, which is characterized in that the spraying parameters (including the geometrical size of the spraying device, the conduction time delta t of an electromagnetic valve and the air pressure P of an air source at the front end of the electromagnetic valve) are fixed0Equal parameters), a prediction model for the spray state parameter S is established for the liquid whose fluid characteristics are stable, and a prediction error range and a confidence interval of the prediction model are obtained. And if the prediction model fails and the predicted value exceeds the range of the confidence interval, judging that the fluid characteristic is changed. According to the invention, the fluid characteristics of the liquid sample can be effectively judged by collecting the liquid storage cavity air pressure waveform P (t) and the liquid drop spraying state parameter actual value S. The method can be used for real-time monitoring of fluid properties of liquid for pneumatic droplet ejection devices.
Description
Technical Field
The invention belongs to the field of droplet ejection, and particularly relates to a pneumatic droplet ejection system. The method for detecting the change of the fluid characteristics aims at solving the problem that the fluid characteristics of the liquid to be sprayed of the pneumatic micro-drop spraying device can be changed in the working process. Can be used for real-time monitoring of pneumatic droplet ejection devices.
Background
The micro-droplet ejection technology is widely applied to the fields of biomedicine, printing electronics, 3D manufacturing and the like besides being used for ink jet printing. The traditional micro-droplet production method adopts thermal drive or piezoelectric drive and is widely applied to ink-jet printing. In addition to this, many advances have been made in recent years with some non-standard (non-standard) micro-droplet ejection methods. Pneumatic droplet ejection is a representative non-standard droplet ejection technique. Compared with other spraying modes, the method is simple to operate and is suitable for samples with various viscosities and temperature ranges. The pneumatic liquid drop spraying technology is widely applied to the fields of electronic packaging, metal 3D printing, biological medicine and the like. Especially in the field of bioprinting, the shear stress in the liquid during pneumatic ejection is much lower than in the case of piezoelectric actuation. For cell printing, pneumatic droplet ejection achieves near 100% cell viability, making it a technical option for applying cell-containing biological samples.
A common pneumatic droplet ejection device consists of a droplet ejection system and a monitoring system, as shown in fig. 1. The device can obtain controllable droplets, and record the data of the change of the gas pressure in the liquid storage cavity along with the time and the images of the liquid drops and the nozzle at a certain moment.
The droplet ejection system mainly includes: the device comprises a liquid storage cavity, a nozzle, a high-speed electromagnetic valve, an air pressure regulator and a ventilation pipeline (comprising an air inlet channel, an air discharge pipe and a T-shaped joint).
The nozzle is fixed in the bottom of the liquid storage cavity, the vent pipe is positioned at the top of the cavity, three ports of the T-shaped joint are respectively connected to the cavity, the air release pipe and the air inlet channel, and the front end of the air inlet channel is sequentially connected with the electromagnetic valve and the air pressure regulator which are controlled by the control and data processing system.
The monitoring system mainly comprises: the system comprises a high-brightness LED lamp, an industrial camera for machine vision monitoring, and a high-speed pressure sensor controllable and data processing system (comprising an upper computer, control software and a lower computer) for monitoring the gas pressure in a liquid storage device. The control and data processing system will drive the camera to record the droplet status image after a delay (compared to the time zero point defined by the rising edge of the high speed solenoid valve opening signal) according to the user configuration parameters.
The industrial camera is used for shooting and recording images of droplets in a delayed mode, is located below a liquid storage cavity of the droplet jetting system, and is used for ensuring imaging quality, and a lens of the industrial camera is opposite to the high-brightness LED lamp. The high-speed pressure sensor is fixed on the side wall of the liquid storage cavity of the droplet spraying system.
The high-speed electromagnetic valve is briefly conducted by delta t, and the high-pressure air source P at the front end of the high-speed electromagnetic valve0The gas enters the liquid storage cavity and is released through the gas release pipe. A pressure pulse p (t) is generated in the reservoir. Droplet generation is driven by a pressure pulse, i.e. a change in gas pressure over time p (t), in the reservoir. By combining related researches, the geometric dimensions of the air inlet and outlet pipe passages, the volume V of the space above the liquid level in the liquid storage cavity, the conduction time delta t of the electromagnetic valve and the air pressure P of the air source at the front end of the electromagnetic valve can be found0And the like. Will affect the oscillating waveform p (t) of the pressure in the chamber and thus the generation of droplets. In principle, after the above control parameters are determined, p (t) is determined.
However, in practical experiments, the air pressure P of the air source0The air pressure pulse P (t) can have certain random fluctuation even under the normal working state of the device, and further causes the fluctuation of droplet ejection state parameters. The droplet ejection state can be described by a droplet ejection state parameter S. The most common injection state parameters include: number of droplets NdThe relative distance H between the droplet and the nozzle with a certain delayd. The method for collecting the above-mentioned spray state parameters is characterized by that it utilizes the time-delay photographing method to take the pictures of microdroplet and nozzle, and utilizes image processing method to measure the number N of microdropletdAnd the relative distance H of the droplet from the nozzled. The gas pressure pulse P (t) in the liquid storage cavity can pass through the high-speed gas pressure sensorAnd (6) measuring.
The droplet ejection state is mainly influenced by the air pressure pulse p (t), fluid characteristics, nozzle wettability, and the like. In general, the wettability of the nozzle can be kept consistent and stable under the condition that the hydrophobic layer of the nozzle is stable. This is the case for the present patent application.
In addition to this, the fluid properties of the liquid to be ejected in the device may also change over time, such as changes in viscosity, precipitation, etc. Such changes are often difficult to detect. The fluid property of the liquid changes to affect the droplet ejection state as well, so that the droplet ejection state parameter S changes. The patent application of the invention proposes: fixed injection parameters (including the geometry of the injection device, the solenoid valve conduction time delta t, and the front end air pressure P of the solenoid valve0Isoparametric) a prediction model and a prediction error range, i.e., a confidence interval of the prediction model, for the liquid of the stable fluid characteristics are established for the ejection state parameter S. During the actual injection process, if the prediction model fails, the predicted value exceeds the confidence interval range. The fluid characteristic is considered to have changed. Based on the method, different detection effects can be achieved by selecting different injection state parameters. Typically, it is desirable to eject only one droplet at a time when the high speed solenoid valve is open. Number of droplets N only when there is a significant change in fluid propertiesdThe change occurs. Thus, using NdFails to determine that the fluid characteristic is changing insensitive. In contrast, a slight change in fluid properties will result in HdA change occurs.
Systematic changes in the spray pattern parameters due to changes in fluid properties are easily masked by random fluctuations in the spray pattern of the device itself. For the reasons described earlier, even droplet ejection control parameters (including the aforementioned ejection device geometry, solenoid valve on-time Δ t, solenoid valve front end gas pressure P)0Equal parameters) are not changed, the air pressure pulse P (t) will fluctuate. Random fluctuations in the ejection state are mainly due to p (t) fluctuations. It is worth noting that: p (t) slight fluctuation in the spray state due to fluctuation is mainly reflected in HdThe fluctuation of (2). Therefore, the more random P (t) can be represented by the prediction modelCharacteristic, the higher its prediction accuracy will be. The higher the sensitivity of detection of fluid-specific changes in the liquid based on predictive model failure. Two predictive models are described in the present patent application.
Prediction model 1: the most straightforward prediction method is based on statistical methods. Keeping droplet ejection control parameters unchanged, collecting a large number of ejection state sample sets to obtain an average value of S
Then
Naturally becomes the predicted value S for Sp. This prediction model does not take into account the random fluctuation factor of p (t) at all.
Prediction model 2: since droplet ejection is the result of a direct drive of the air pressure pulse p (t), it is feasible to build a predictive model of p (t) versus S by a machine learning method. The BP neural network is a multilayer feedforward neural network trained according to an error back propagation algorithm and is the most widely applied neural network at present. By training the BP neural network, a prediction model which takes the pneumatic oscillation signal P (t) as input and can establish the droplet ejection state parameter S can be constructed. Because the self-interference of the device can be expressed by the air pressure oscillation waveforms P (t) of various forms when the BP neural network is trained, the BP neural network obtained after the training takes the self-fluctuation of the device into account, and the prediction precision of the model on the droplet ejection state parameter S is obviously higher than that of a prediction model based on simple statistical average.
Since the fluctuation of the aforementioned p (t) has randomness, the ejection state parameter follows a normal distribution when the number of droplet ejection samples is sufficiently large according to a statistical method. At droplet relative position HdFor example, HdObey a positive state distribution as shown in fig. 2 and fig. 3. In combination with the confidence interval of normal distribution and its principle, the present patent application proposes: a confidence interval is first set. And a feasible criterion is provided on the basis. Generally, when the injection state parameter prediction error is outside of the range confidence interval in some mannerThen it can be determined that the prediction model has failed. This also means that the fluid properties of the liquid sample are detected as changing, requiring the experimenter to adjust the liquid sample in time to avoid affecting the printing.
Since the change in fluid properties of a liquid sample typically results from changes in the composition and shape of the fluid, the properties and function of the printed product can be affected without intervention. Therefore, the fluid characteristics of the biological ink need to be monitored in real time in the actual printing process, the fluid characteristics of the biological ink need to be monitored to be changed as soon as possible, and an alarm is given to a user. Therefore, the present invention is directed to the above-mentioned problems, and provides a method for monitoring a change in fluid characteristics of a liquid sample, by which a change in fluid characteristics of the liquid sample can be sensitively determined and an alarm can be issued to a user.
Disclosure of Invention
The phenomenon that the fluid properties of a liquid change with time during operation of a pneumatic droplet ejection system is common in personalized applications such as biomedicine. Fluid property changes typically result from liquid behavior, composition changes. Detection of changes in fluid characteristics of the liquid during droplet ejection is difficult.
The invention provides a method for detecting fluid characteristic changes of a liquid. The method is characterized in that the injection parameters (including the geometric dimension of the injection device, the conduction time delta t of the electromagnetic valve and the air pressure P of the air source at the front end of the electromagnetic valve) are fixed0Equal parameters), a prediction model for the spray state parameter S is established for the liquid whose fluid characteristics are stable, and a prediction error range and a confidence interval of the prediction model are obtained. And if the prediction model fails and the prediction value exceeds the range between the confidence zones, judging that the fluid characteristics are changed.
To implement this technique, it is first necessary to establish a predictive model of the droplet ejection state parameter S. The invention provides two prediction models. The first method is based on a simple statistical averaging. I.e. by measuring the injection status parameter data set S for a large number of injection eventsn}, calculating the mean value
WhereinN is the total number of injection events. Will be provided with
As a predicted value S for a future unknown injection state parameterp. I.e. the predicted value is fixed. The second prediction method measures the pressure waveform p (t) in the liquid storage chamber by using the characteristic of high correlation between the injection state parameter and the injection driving pressure p (t). For a large number of injection events, an air pressure waveform data set { P } is collectedn(t) }, and the corresponding injection state parameter { Sn}. Establishing a neural network prediction model of the driving air pressure pulse waveform P (t) to the injection state parameter S by adopting a machine learning method, and obtaining a predicted value Sp，SpDepending on the air pressure waveform. Since P (t) is different for each injection, the predicted value SpAlso differently, this is different from the prediction method based on statistical averaging.
The neural network can be a common BP neural network or other neural networks. Once the prediction model is established, the injection setting parameters are kept unchanged (including the front-end air pressure P of the high-speed electromagnetic valve)0High speed solenoid valve on time Δ t, etc.), the fluid characteristics of the liquid are kept stable, a certain number of injection events are collected, and a test set is formed. For the two prediction methods, the difference between the actual measured value and the predicted value is the prediction error em＝Sm- SpAlso constitute the data set emWhere m is the test set sample number. Normally, one σ is setθE such that θ × 100%mAt + -sigmaθWithin the range, called the confidence interval. In the actual use process of the jetting device, each time of jetting, the micro-droplet monitoring system collects the air pressure signal P in real timel(t) and the actual value S of the droplet ejection state parameterl. The system (upper computer) can calculate a predicted value S through a prediction modelpAnd a prediction error el＝Sp-Sl. l is the number of injection events during actual use of the injection device, independent of the data set number n, and test set number m, from which the predictive model was previously built. For the prediction error elPerforming a smoothing process to obtain
The smoothing operation includes, but is not limited to, a sliding smoothing (smoothing) operation. Error after smoothing
Confidence interval ± σ beyond predictive modelθThen it is determined that the predictive model has failed, i.e., it is determined that a change in fluid properties has occurred. And the system sends out a warning signal.
The specific process of establishing the prediction model by the statistical averaging method is not described herein again. The establishment process of the prediction model based on the BP neural network is specifically described as follows:
keeping the fluid properties of the liquid stable is not within the present patent application. Relevant parameters (front end air pressure P of the electromagnetic valve) of the injection device in a general state are configured through an upper computer control program0Solenoid valve on time Δ t, etc.). And a large number of widely representative { P } s are collected by an acquisition systemn(t) waveform and measured value of droplet ejection state parameter SnLarge data of.
Wherein, the pneumatic oscillation signal P (t) is collected (the rising edge of the electromagnetic valve driving signal is used as reference time) by fsFor a sampling frequency, the sampling interval is
High-speed acquisition. In the actual acquisition process, two groups of signals are acquired simultaneously from the rising edge of the synchronous signal to obtain the air pressure signals P (t) at i discrete momentsi)。
Carrying out basic processing on the acquired data: reading the pressure waveform P (t)i) Samples, filtering and clipping the waveform to obtain high quality waveform sample P (t)i)。
(1) and carrying out global normalization on the obtained waveform samples.
The waveform samples are globally normalized by a normalization function, and the range of the data obtained after normalization is between [ -1, 1 ]. The normalization function is:
k and n are the number of the ejection samples, tiRepresenting the sampling time at the ith discrete time instant.
(2) For the obtained data set { S } of the droplet ejection state parameters SkNormalization is performed.
The droplet state parameters were normalized using a normalization function, with the resulting data ranging between [ -1, 1] after normalization. The normalization function is:
where S represents a droplet state parameter and k and n are numbers of ejected samples. In the present patent application, S generally selects the relative droplet ejection position Hd。
Step 3, establishing a prediction model and adjusting network related parameters
Building a BP neural network prediction model for predicting droplet ejection state, wherein the input variable of the prediction model is the air pressure waveform P (t) in the cavityi) The output variable is a droplet ejection state parameter S (parameters such as the number of droplets and the position).
And configuring parameters such as the number of neurons in an input layer, the number of neurons in an output layer, the number of layers of hidden layers, the number of neurons in each hidden layer and the like. The number of neurons in the input layer is the pressure waveform P (t)i) The number of discrete data points and the number of output layer neurons are 1;
initializing a neural network, selecting a neuron excitation function in the neural network, and starting training the neural network.
And obtaining the trained predictive neural network after the steps are completed. Obtaining P (t)i) A model for predicting S.
Two prediction models for S were obtained, one based on statistical averaging and the other based on machine learning and BP neural networks.
(1) The predictive model described above is tested through a set of test sets. Keeping the jet setting parameters unchanged (including the front end air pressure P of the high-speed electromagnetic valve)0And the conduction time delta t of the high-speed electromagnetic valve and the like), the fluid characteristics of the liquid are kept stable, and a certain number of injection events are collected to form a test set. The test set sample number is m. For a prediction model of a neural network, a waveform signal P of the air pressure in the liquid storage cavity is obtainedm(ti) Inputting the data into a prediction model to obtain a predicted value S of the relative position of the dropletp. And with the actual value S obtained by machine vision and image processingmMaking difference to obtain a prediction difference data set { em}. For the prediction model of the statistical average, the prediction error can be estimated directly according to the predicted value of the statistical average.
(2) From statistical principles and in combination with the prediction error distribution, for a given theta, σ is obtainedθSo that the error of θ × 100% is within ± σθWithin the range. The confidence intervals of the two prediction models are different.
So far, information of the prediction model, prediction error and confidence interval has been obtained.
Step 5, the droplet monitoring system carries out real-time acquisition, and the system automatically judges whether the critical threshold value is exceeded or not
In the actual use process of the jetting device, each time of jetting, the micro-droplet monitoring system collects the air pressure signal P in real timel(ti) And the actual value S of the droplet ejection state parameterl. Simultaneously transmit the air pressure signal Pl(ti) And inputting the data into the trained prediction model. The system (upper computer) can calculate a predicted value SpWith the actual value SlDifference, i.e. prediction error el＝ Sp-Sl. Where l is the number of injection events during actual use of the injection device, the number n of the data set from which the prediction model was previously established, andthe trial set number m is irrelevant. For the prediction error elPerforming smoothing treatment to obtain
Error after smoothing
Confidence interval ± σ beyond predictive modelθThen it is determined that the predictive model has failed, i.e., it is determined that a change in fluid properties has occurred. It is noted that the particular scheme of smoothing is not particularly emphasized here. The stronger the smoothing effect, the lower the sensitivity of the above criterion, but the smaller the probability of erroneous judgment. Similarly, the larger the value of θ, the confidence interval ± σθThe wider the prediction model, the less likely it is to fail, so the lower the sensitivity to fluid property changes, while the less likely the false positives.
Compared with the prior art, the method can effectively judge the fluid characteristics of the liquid sample by collecting the liquid storage cavity air pressure waveform P (t) and the liquid drop spraying state parameter actual value S. The fluid characteristic of the liquid that can be used for pneumatic droplet ejection device is monitored in real time.
Drawings
FIG. 1 is a reference schematic of a pneumatic droplet ejection device;
FIG. 2 is a diagram of droplet placement profiles;
FIG. 3 is a normal distribution diagram of droplet positions;
FIG. 4 is a diagram of a BP network architecture;
FIG. 5 is a prediction error profile and confidence interval for a machine learning based model;
FIG. 6 is an error distribution map and confidence interval for a statistical-based model prediction;
FIG. 7 is a detection of a failure, change in fluid characteristics, of a machine learning based predictive model;
FIG. 8 is a detection of failure of a statistical-based predictive model, a change in fluid characteristics.
Detailed Description
The following are specific examples provided by the inventors to further explain the technical solutions of the present invention.
Example 1:
in this embodiment, the relative droplet position H is established for a liquid whose viscosity remains stable for a pneumatic droplet ejection devicedThe predictive model of (1). And then gradually reducing the viscosity of the liquid, and detecting that the viscosity changes by judging that the prediction model is invalid. The device comprises a droplet ejection system and a droplet ejection state monitoring system.
(1) Configuring relevant parameters of a droplet ejection system and a droplet ejection state monitoring system through an upper computer program: comprises a high-speed electromagnetic valve front-end air pressure P00.03MPa, the conduction time delta t of the high-speed electromagnetic valve is 1.5ms, the injection frequency is 20Hz, and the shooting delay is 5000 mus. The sprayed liquid was a homogeneous 8% glycerol-water mixed solution.
(2) The lower computer control circuit controls the parameters set on the software according to the upper computer; periodically generating a photographing signal, controlling the rising edge of the driving signal of the industrial camera relative to the high-speed electromagnetic valve to delay 5000 mu s to photograph the droplets generated by the device, and using an upper computer to photograph the droplet position HdThe measurement is performed.
(3) While spraying, a widely representative 3000 sets of samples were collected by the collection system, each set containing a barometric pressure signal P (t)i) And the actual value H of the relative position of the dropletd. The process of establishing the prediction model based on the statistical average is not described in detail. The establishment of the prediction model based on the BP neural network will be described in detail below.
(4) Carrying out basic processing on the acquired data: batch reading of the barometric pressure signal P (t) using matlab programi) And a complete waveform corresponding to each injection is intercepted by combining the rising edge of a periodic high-speed electromagnetic valve driving signal.
To obtain high-quality waveform samples P (t)i) Filtering and further clipping the waveform, and generating P (t) after droplet generation based on a priori knowledge of droplet generationi) The impact on the droplet ejection state is negligible. Thus, intercepting and retaining the segment before the moment of fracturePressure waveform signal in the chamber as effective P (t)i) And (4) data.
(1) For the obtained pressure waveform sample Pn(ti) And carrying out global normalization.
The waveform samples are globally normalized by a normalization function, and the range of the data obtained after normalization is between [ -1, 1 ]. The normalization function is:
k and n are the number of the ejection samples, tiRepresenting the sampling time at the ith discrete time instant.
(2) Relative position H of obtained droplet ejectiondAnd (6) carrying out normalization.
Relative position H for droplet ejection using normalization functiondNormalization is carried out, and the obtained data after normalization is in the range of [ -1, 1]In the meantime. The normalization function is:
in the formula HdIs the position of the drop relative to the nozzle with a delay of 5000 mus. k and n are the numbers of the ejection samples.
Step 3, establishing a prediction model and adjusting network related parameters
(1) Building a prediction model, wherein the input variable of the model is the air pressure waveform P (t) in the cavityi) The output variable is the position H of the micro-dropd。
The number of neurons in the input layer is P (t)i) The number of discrete data points, which is 120 in this application example, one hidden layer includes 5 neurons, and the number of output layer units is 1, as shown in fig. 4.
(2) For the present example, the excitation function is chosen to be (but not limited to) "tansig".
(3) Inputting training samples into a neural network for training: first, output layer calculation is performed, and then error calculation is performed. And judging whether the error meets the requirement or the training frequency reaches the maximum, and stopping training if any condition is met.
Otherwise, adjusting the connection weight between each layer of neurons and the threshold value of the offset bias, thereby continuously reducing the error; this process is repeated until the training reaches a set threshold.
After the steps are completed, a trained pass P (t) is obtainedi) Prediction of HdThe BP neural network of (1).
Thus, two pairs of H are obtaineddOne based on statistical averaging and the other based on machine learning and BP neural networks.
(1) The predictive model described above is tested through a set of test sets. Keeping the jet setting parameters unchanged (including the front end air pressure P of the high-speed electromagnetic valve)00.03MPa, the conduction time delta t of a high-speed electromagnetic valve is 1.5ms, the spraying frequency is 20Hz, and the shooting time is delayed by 5000 mu s), and the sprayed liquid is a uniform 8% glycerol-water mixed solution. A test set of 600 injection events was collected. The test set sample number is m. For a prediction model of a neural network, an air pressure waveform signal P in the liquid storage cavity is converted into a pulse signalm(ti) Inputting the data into a prediction model to obtain a predicted value (H) of the relative position of the dropletd)p. And with the actual value (H) obtained by machine vision and image processingd)mMaking difference to obtain a prediction difference data set { em}. For the prediction model of the statistical average, the prediction error can be estimated directly according to the predicted value of the statistical average.
It was verified that for this example, the machine learned predictive model predicted the relative position of the drop HdWhen the error of 95% is within the interval of ± 13 pixels, i.e., when θ is 95%, σ is predictedθ13 pixels (as shown in fig. 5). In contrast, based on a statistical average prediction model, a 95% error is within a ± 70pixel interval, i.e., when θ is 95%, σ isθ70 pixels (as shown in fig. 6). Indicating predictive models based on machine learning and BP neural networksThe prediction precision is obviously improved.
Step 5, the droplet monitoring system carries out real-time acquisition, and the system automatically judges whether the prediction model fails or not
(1) As shown in FIG. 7, after establishing a prediction model for an 8% glycerol-water mixed solution, the glycerol concentration was gradually decreased to 6%, 4%, 2%, and the actual (H) was measuredd)lAnd calculating a prediction error e from the predicted valuel. Using sliding filter (moving average) method to elPerforming smoothing treatment to obtain
Where l is the number of injection events during actual use of the injection device, independent of the data set number n and test set number m for which the predictive model was previously established. It can be seen that the prediction error of the machine learning based prediction model is smoothed when the glycerol-water solution concentration is reduced from 8% to 6%
Significantly exceeds + -sigmaθ(where θ is 0.95) signaling interval, indicating failure of the predictive model. By contrast, when the glycerol-water solution concentration decreased from 8% to 2%, the curve was smoothed based on the prediction error of the statistically averaged prediction model
Is significantly out of + -sigmaθ(where θ is 0.95). This example shows that failure of the predictive model can indeed detect a change in the fluid properties. It can also be seen that failure of the machine learning based predictive model can detect more subtle fluid characteristic changes.
(2) And giving an early warning to the device user.
Claims (2)
1. A method of monitoring changes in characteristics of fluid of liquid ejected during a pneumatic droplet ejection process, comprising: establishing a predictive model for a droplet ejection state parameter S, measuring an ejection state parameter dataset { S ] for an ejection event based on a statistical averagen}, calculating the mean value
Wherein N is the total number of injection events; will be provided with
As a predicted value S for a future unknown injection state parameterpThe predicted value is fixed;
measuring the air pressure waveform P (t) in the liquid storage cavity by utilizing the correlation characteristic between the injection state parameter and the injection driving air pressure P (t); collecting pressure waveform data set { P }n(t) and an injection state parameter Sn}; establishing a prediction model of a BP neural network for driving the air pressure pulse waveform P (t) to the injection state parameter S by adopting a machine learning method, and obtaining a predicted value Sp，SpDependent on the air pressure waveform;
after a prediction model of the BP neural network is established, the injection setting parameters comprise the front-end air pressure P of the high-speed electromagnetic valve0The conduction time delta t of the high-speed electromagnetic valve is unchanged, the fluid characteristics of the liquid are kept stable, and the injection events are collected to form a test set; actual measured value SmDifference from predicted value, i.e. prediction error em＝Sm-SpAlso constitute the data set emM is a test set sample number; setting a sigmaθE such that θ × 100%mAt + -sigmaθWithin a range, referred to as a confidence interval; in the actual use process of the jetting device, the droplet monitoring system for each jetting acquires the air pressure signal P in real timel(t) and the actual value S of the droplet ejection state parameterl(ii) a Calculating a prediction value S through a prediction model of a BP neural networkpAnd a prediction error el＝Sp-Sl(ii) a l is the number of injection events during actual use of the injection device; for the prediction error elPerforming a smoothing process to obtain
Error after smoothing
Confidence interval ± σ beyond predictive modelθAnd if the prediction model of the BP neural network is judged to be invalid, namely the fluid characteristic is judged to be changed, and the system sends out a warning signal.
2. A method of monitoring changes in the characteristics of a fluid of liquid ejected during pneumatic droplet ejection according to claim 1, wherein: the establishment process of the prediction model based on the BP neural network is as follows:
step 1, collecting injection data under conventional conditions:
collection System collects { Pn(t) waveform and measured value of droplet ejection state parameter SnLarge data of }; collecting the gas pressure oscillation signal P (t) by fsFor a sampling frequency, the sampling interval is
High-speed acquisition; in the actual acquisition process, two groups of signals are acquired simultaneously from the rising edge of the synchronous signal to obtain air pressure signals P (t) at i discrete momentsi)；
Processing the acquired data: reading the pressure waveform P (t)i) Samples, filtering and clipping the waveform, thus obtaining a waveform sample P (t)i)；
Step 2, normalizing the samples:
(1) carrying out global normalization on the obtained waveform samples;
carrying out global normalization on the waveform sample by using a normalization function, wherein the range of data obtained after normalization is between [ -1, 1 ]; the normalization function is:
k and n are the number of the ejection samples, tiA sample time representing the ith discrete time instant;
(2) for the obtained data set { S } of the droplet ejection state parameters SkNormalizing;
normalizing the droplet state parameters by using a normalization function, wherein the range of the data obtained after normalization is between [ -1, 1 ]; the normalization function is:
wherein S represents a droplet state parameter, and k and n are numbers of the ejected samples; s selects a droplet ejection relative position Hd；
Step 3, establishing a prediction model and adjusting network related parameters;
building a BP neural network prediction model for predicting droplet ejection state, wherein the input variable of the BP neural network prediction model is the air pressure waveform P (t) in the cavityi) The output variable is a droplet ejection state parameter S;
configuring parameters of the number of neurons in an input layer, the number of neurons in an output layer, the number of layers of hidden layers and the number of neurons in each hidden layer; the number of neurons in the input layer is the pressure waveform P (t)i) The number of discrete data points and the number of output layer neurons are 1;
initializing a neural network, selecting a neuron excitation function in the neural network, and starting training the neural network;
obtaining a trained predictive neural network; obtaining P (t)i) A model of prediction S;
step 4, determining a prediction error and a confidence interval of the prediction model;
(1) testing the predictive model through a set of test sets; collecting injection events to form a test set; the sample number of the test set is m; for the prediction model, the pressure waveform signal P in the liquid storage cavity is usedm(ti) Inputting the data into a prediction model to obtain a predicted value S of the relative position of the dropletp(ii) a And with the actual value S obtained by machine vision and image processingmMaking difference to obtain a prediction difference data set { em}; for a statistically averaged prediction model, the prediction error is estimated based on the predicted value of the statistical averageThe difference is obtained;
(2) from statistical principles and in combination with the prediction error distribution, for a given theta, σ is obtainedθSo that the error of θ × 100% is within ± σθWithin the range; information of the prediction model, prediction error and confidence interval has been obtained;
step 5, the droplet monitoring system collects the droplets in real time, and the droplet monitoring system automatically judges whether the critical threshold value is exceeded or not
In the actual use process of the jetting device, each time of jetting, the micro-droplet monitoring system collects the air pressure signal P in real timel(ti) And the actual value S of the droplet ejection state parameterl(ii) a Simultaneously transmit the air pressure signal Pl(ti) Inputting the data into a trained prediction model; the droplet monitoring system calculates a predicted value SpWith the actual value SlDifference, i.e. prediction error el＝Sp-Sl(ii) a l is the number of injection events during actual use of the injection device; for the prediction error elPerforming smoothing treatment to obtain
Error after smoothing
Confidence interval ± σ beyond predictive modelθThen it is determined that the predictive model has failed, i.e., it is determined that a change in fluid properties has occurred.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011357907.5A CN112733418A (en) | 2020-11-27 | 2020-11-27 | Method for monitoring fluid characteristic change of ejected liquid in pneumatic droplet ejection process |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011357907.5A CN112733418A (en) | 2020-11-27 | 2020-11-27 | Method for monitoring fluid characteristic change of ejected liquid in pneumatic droplet ejection process |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112733418A true CN112733418A (en) | 2021-04-30 |
Family
ID=75597859
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011357907.5A Pending CN112733418A (en) | 2020-11-27 | 2020-11-27 | Method for monitoring fluid characteristic change of ejected liquid in pneumatic droplet ejection process |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112733418A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114083359A (en) * | 2021-11-26 | 2022-02-25 | 青岛理工大学 | Multi-energy-field nano lubricant microscale bone grinding processing measurement system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20150298153A1 (en) * | 2012-12-27 | 2015-10-22 | Kateeva, Inc. | Techniques for Arrayed Printing of a Permanent Layer with Improved Speed and Accuracy |
| CN109871614A (en) * | 2019-02-19 | 2019-06-11 | 北京工业大学 | The method of pneumatic type droplet ejection status predication based on BP neural network |
| WO2020207893A1 (en) * | 2019-04-08 | 2020-10-15 | Mycronic AB | Method using a neural network for generation of jetting control parameters |
-
2020
- 2020-11-27 CN CN202011357907.5A patent/CN112733418A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20150298153A1 (en) * | 2012-12-27 | 2015-10-22 | Kateeva, Inc. | Techniques for Arrayed Printing of a Permanent Layer with Improved Speed and Accuracy |
| CN109871614A (en) * | 2019-02-19 | 2019-06-11 | 北京工业大学 | The method of pneumatic type droplet ejection status predication based on BP neural network |
| WO2020207893A1 (en) * | 2019-04-08 | 2020-10-15 | Mycronic AB | Method using a neural network for generation of jetting control parameters |
Non-Patent Citations (1)
| Title |
|---|
| 王飞: "气动微滴喷射状态的人工神经网络预测模型的建立与应用", 《中国优秀硕士学位论文全文数据库》, pages 15 - 60 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114083359A (en) * | 2021-11-26 | 2022-02-25 | 青岛理工大学 | Multi-energy-field nano lubricant microscale bone grinding processing measurement system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | In-situ droplet inspection and closed-loop control system using machine learning for liquid metal jet printing | |
| EP3429862B1 (en) | Systems and methods for precision inkjet printing | |
| CN109435473B (en) | A kind of flight ink droplet detection device and method suitable for inkjet printing | |
| CN108481749B (en) | Real-time monitoring three-dimensional printing method and device | |
| CN109823050B (en) | Multi-stage positioning error compensation method and device for liquid drop ejection for ink-jet printing | |
| CN107608335B (en) | Data driving method for fault detection and fault separation of unmanned aerial vehicle flight control system | |
| Wang et al. | In-situ droplet inspection and control system for liquid metal jet 3D printing process | |
| CN109823051B (en) | Liquid drop injection fusion overall process volume control method and system and printer | |
| CN109871614B (en) | Pneumatic type droplet ejection state prediction method based on BP neural network | |
| CN112733418A (en) | Method for monitoring fluid characteristic change of ejected liquid in pneumatic droplet ejection process | |
| CN109968671B (en) | 3D printing process fault diagnosis method and device | |
| JP2008539071A (en) | System and method for monitoring the performance of an injector | |
| CN106815639A (en) | The abnormal point detecting method and device of flow data | |
| US7634367B1 (en) | Estimating fluidic properties and using them to improve the precision/accuracy of metered fluids and to improve the sensitivity/specificity in detecting failure modes | |
| CN113433131A (en) | Physical multi-parameter coupled method and device for rapidly monitoring mildew of citrus peel | |
| Mao et al. | Black-box real-time identification of sub-regime of gas-liquid flow using Ultrasound Doppler Velocimetry with deep learning | |
| CN110259616A (en) | Based on can measured data feature common rail for diesel engine system injector failures detection method | |
| US6309571B2 (en) | Method and apparatus for the control of injection molding | |
| CN112879388A (en) | Abnormal leakage detection method for hydraulic system | |
| Yen et al. | Acoustic emission data assisted process monitoring | |
| CN116148801A (en) | Millimeter wave radar-based target detection method and system | |
| CN113324580B (en) | Static droplet jetting state detection method based on induced current and machine vision | |
| CN115008900B (en) | Flexible display jet printing film edge straightness control method and system | |
| CN117634581A (en) | Device and method for predicting pneumatic droplet ejection state based on BP neural network | |
| US11565517B2 (en) | Information processing device, learning device, and information processing method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |