A method and system for detecting nozzle clogging based on machine learning

Abstract

The present invention relates to the field of inkjet printing technology, and specifically discloses a nozzle clogging detection method and system based on machine learning: 1) cleaning the nozzle, obtaining standard inkjet printing parameters, moving the nozzle to a trial printing area for trial spraying, and detecting the corresponding ink droplet landing points of all nozzle holes of the nozzle; 2) data acquisition and annotation, collecting piezoelectric sensor signal images of the nozzle under standard inkjet printing parameters and different nozzle clogging distribution conditions, and performing data annotation; 3) data preprocessing and feature extraction, preprocessing the collected piezoelectric sensor signal images, and extracting key features in the piezoelectric sensor signal images; 4) constructing a machine learning model The invention adopts a machine learning-based nozzle clogging state evaluation model, which combines the image features of the piezoelectric sensor signal as input and the distribution of clogged nozzles of different nozzles as output, and uses the transfer learning method to establish a nozzle clogging distribution state prediction model, and performs model training, verification and testing; the present invention utilizes a piezoelectric actuator as a sensor for actively detecting channel acoustics. When the piezoelectric actuator is no longer driven by a waveform, it is used to detect the internal channel pressure fluctuation; with the image features of the piezoelectric sensor signal as input and the distribution of different nozzle clogging as output, a nozzle clogging state evaluation model based on machine learning is constructed, which is used to detect the clogging state of inkjet nozzles in real time, effectively improving the detection efficiency.

Description

Technical Field

[0001] The present invention relates to the field of inkjet printing technology, and in particular to a method and system for detecting nozzle clogging based on machine learning. Background Art

[0002] Inkjet printing technology, due to its advantages of high precision, high speed, and low cost, is widely used in fields such as chemistry, biology, and industrial production. However, during the inkjet printing process, nozzle clogging can be caused by impurities in the ink, or by chemical reactions and precipitation after mixing inks from different manufacturers and brands, which can negatively impact print quality.

[0003] Currently, the main method for detecting whether a nozzle is clogged is to process the image printed by the nozzle on the substrate and compare it with the image printed by a normal nozzle. Based on the obtained image information, the nozzle is judged to be clogged and the approximate location of the nozzle where the blockage occurs is determined. This method is time-consuming and lacks real-time performance. It requires the nozzle to stop working and print a test image to detect nozzle blockage, which is cumbersome to operate. At the same time, high-speed cameras are used to capture images of ink droplets ejected from single-row, multi-column or multi-row, single-column nozzles in real time to determine the blockage status of the nozzle. However, it is difficult to observe the falling images of ink droplets from multi-row, multi-column nozzles, and the ink droplets ejected from a single nozzle are small in size and high in speed. This requires higher specifications for the camera equipment and is expensive. In actual production, it is difficult to equip every inkjet printer with a high-speed camera to detect the blockage status of the nozzle in real time.

[0004] Therefore, in order to solve the problems that the existing methods for detecting nozzle clogging status are difficult to apply to large-scale real-time detection scenarios and have high costs and low efficiency, a nozzle clogging detection method and system based on machine learning is designed, which is suitable for large-scale real-time detection of nozzle clogging. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a nozzle clogging detection method and system based on machine learning, which can obtain the distribution of nozzle clogging holes through the piezoelectric sensor signal of the nozzle and the constructed deep learning model, and detect the nozzle clogging situation in real time.

[0006] To achieve the above objectives, the present invention discloses the following technical solutions:

[0007] A method for detecting nozzle clogging based on machine learning, the method comprising the following steps:

[0008] S1. Clean the print head, obtain standard inkjet printing parameters, move the print head to the test printing area for test jetting, and detect the corresponding ink droplet landing points of all nozzles of the print head to ensure that each nozzle is working normally;

[0009] S2. Data collection and annotation: collecting piezoelectric sensor signal data of the printhead under standard inkjet printing parameters and different nozzle blockage distribution conditions, and annotating the data;

[0010] S3, data preprocessing and feature extraction, preprocessing the collected piezoelectric sensor signal data, extracting feature information from the piezoelectric sensor signal data (referring to the different feature parts in the time-frequency acoustic texture image when the nozzle is blocked in different states), for subsequent classifier training and prediction;

[0011] S4. Build a machine learning model, combine the piezoelectric sensor signal data characteristics as input, and the distribution of nozzle blockage as output. Use the transfer learning method to establish a nozzle blockage distribution state prediction model, and perform model training, verification, and testing.

[0012] Among them, in step S1, all the nozzle holes of the nozzle are cleaned and unobstructed to reduce the possibility of nozzle hole blockage, and the nozzle is wiped to remove residual ink on the surface of the nozzle to avoid subsequent clogging of the nozzle holes by residual ink, thereby ensuring subsequent normal testing.

[0013] Wherein, step S1, obtaining standard inkjet printing parameters, specifically includes:

[0014] The nozzle driving waveform data and the standard speed and standard volume of ink droplets ejected by the preset nozzle; the nozzle driving waveform data includes delay time, rise time, hold time, fall time and hold voltage.

[0015] Wherein, step S1, detecting whether all nozzles can normally eject ink droplets, includes:

[0016] Generate printing pattern data that matches the printing of all nozzles to print and solidify the ink droplets, move the printing plate from the printing area to the detection area, take photos of the printed pattern on the printing plate for detection, and check whether all nozzles can spray ink droplets normally.

[0017] Among them, step S2, collecting the piezoelectric sensor signal data of the nozzle under standard inkjet printing parameters and different nozzle blockage distribution conditions, and performing data annotation, specifically includes the following steps:

[0018] As shown in the figure, all the nozzle holes of the print head are divided and blocked in sequence, so that the print head prints according to the printing pattern data corresponding to the injection status of all the nozzle holes. The printed image on the printing plate is photographed and detected, and the image algorithm is processed to check whether the actual nozzle hole blockage distribution status is consistent with the preset nozzle hole blockage distribution status. If they are consistent, the piezoelectric sensor signal data during the printing process of the print head is collected and the data is annotated with the corresponding nozzle hole blockage distribution status.

[0019] Among them, step S3, data preprocessing specific steps include:

[0020] Data preprocessing uses the wavelet threshold noise reduction method to filter out interference while retaining signal details, facilitating subsequent signal feature extraction.

[0021] The ultrasonic amplitude frequency domain data is processed, the signal characteristic information is extracted by wavelet packet analysis, and the appropriate wavelet basis function is selected to perform wavelet packet transform on the signal.

[0022] Wavelet transform uses the wavelet basis function ψ a，b (t)

[0023] Transform the signal. For any signal f(t), its continuous wavelet transform formula is as follows:

[0024]

[0025] Where a>0 is called the scale factor, which is used to adjust the basic wavelet function ψ a，b (t) expands and contracts, and b reflects the displacement, which can be positive or negative.

[0026] The specific operations of step S3, data feature processing, include:

[0027] The piezoelectric sensor signal of the pressure fluctuation in the ink channel of the nozzle after the wavelet threshold denoising method is processed using the time-frequency analysis method to obtain a pressure fluctuation time-frequency acoustic texture image; the time-frequency acoustic texture image is compared with the single time domain image and frequency domain image, and has the dynamic characteristics of the pressure fluctuation state in the ink channel of the nozzle.

[0028] Specifically, the time-frequency analysis method uses the Synchrosqueezing Wavelet Transform (SWT). By using the Synchrosqueezing operator to rearrange the time-frequency coefficients, the time-frequency distribution of the signal at any point in the time-frequency plane is moved to the energy center of gravity, resulting in a more concentrated time-frequency transform, which can better resolve the time-frequency ambiguity problem. The specific steps are as follows:

[0029] The continuous wavelet transform formula W(t) for any signal f(t) is as follows:

[0030]

[0031] Where a>0 is called the scale factor, b is the displacement factor, which reflects the displacement and its value can be positive or negative. ψ is an appropriately selected wavelet and redistributes W f (a, b) A centralized time-frequency image is obtained, from which the instantaneous frequency lines are extracted.

[0032] For any W fFor any (a, b) where (a, b) ≠ 0, calculate the instantaneous frequency W of the signal f(t) f (a, b) can be expressed as follows:

[0033]

[0034] Synchronous compression redistributes the wavelet transform "time-scale" into "time-frequency", suppresses blurring along the scale and maps the scale to frequency, thereby improving the resolution of the time-frequency sound texture to a limited extent and fully characterizing the dynamic characteristics of the pressure fluctuation state in the ink channel of the nozzle. The calculation formula is as follows:

[0035]

[0036] Where a k is the value of the kth discretization scale factor a; ω l is the lth discrete angular frequency; a k -a k-1 =(Δa) k ;ω l -ω l-1 =Δω.

[0037] Among them, step S4, building a machine learning model specifically includes the following steps:

[0038] Machine learning uses the time-frequency, acoustic, texture, and image features of piezoelectric sensor signals as input, and the extracted high-level features to accurately identify different types of images, transforming the nozzle blockage status identification problem into a nozzle orifice blockage distribution classification problem.

[0039] Among them, the time-frequency sound texture data set of various different nozzle nozzle blockage distribution conditions is used for model training. Through the fine-tuning method of transfer learning, some parameters of the training model are frozen, and a small number of parameters in the latter part are modified and trained, so as to efficiently realize model training for different classification problems.

[0040] A nozzle clogging detection system based on machine learning, characterized by comprising:

[0041] Printhead module, used to repeat inkjet test according to different settings of nozzle blockage distribution;

[0042] The image detection module is used to observe the ink droplet landing point corresponding to the nozzle during the inkjet process and detect whether the nozzle blockage condition is consistent with the actual ink discharge condition of the nozzle;

[0043] The acoustic signal acquisition module is used to collect the piezoelectric sensor signal from the nozzle ignition to the ink droplet falling;

[0044] A motion control module, used to control the motion state of the nozzle and substrate according to the inkjet state;

[0045] The data processing module is used to process piezoelectric sensor signal data and build a machine learning model.

[0046] Furthermore, the image detection module shown includes an optical detection camera and a stroboscopic light source; the optical detection camera is installed on one side of the nozzle, and the stroboscopic light source is installed at a coaxial position of the optical detection camera, which is responsible for detecting the corresponding ink droplet landing point position under different nozzle blockage distribution conditions.

[0047] Beneficial effects of the present invention: The present invention utilizes a piezoelectric actuator as a sensor for actively detecting channel acoustics (the electrical signal collected by the piezoelectric measurement is used as acoustic signal data). When the piezoelectric actuator is no longer driven by a waveform, it is used to detect internal channel pressure fluctuations; the piezoelectric sensor signal data characteristics are input, and different nozzle blockage distribution conditions are output. A nozzle blockage status assessment model based on machine learning is constructed to detect the inkjet nozzle blockage status in real time, effectively improving detection efficiency. Compared with the method of detecting nozzle blockage by printing images, the nozzle blockage status can be detected without stopping the nozzle; compared with the method of detecting nozzle blockage by taking a camera to capture an image of ink droplets ejected from a single-row and multiple-column or multiple-row and single-column nozzle, the application scenario can be expanded to multiple rows and multiple columns of nozzles. The present invention has the advantages of low cost and high efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Flowchart of the present invention;

[0049] Figure 2 A schematic diagram of the waveform of the piezoelectric sensor of the present invention;

[0050] Figure 3 Schematic diagram of the inkjet ink droplet landing point detection method of the present invention;

[0051] Figure 4 Schematic diagram of different blockage situations of the nozzle orifice of the present invention;

[0052] Figure 5 A schematic diagram of the structure of the machine learning nozzle clogging detection system of the present invention;

[0053] In the figure: 1. Printhead; 2. Printhead driver board; 3. Printhead control board; 4. Power supply; 5. Printhead nozzle blockage setting assembly; 6. Host computer; 7. Cleaning and wiping assembly; 8. Optical detection camera; 9. Stroboscopic light source; 10. X-axis linear motor; 11. X-axis sensor; 12. Y-axis linear motor; 13. Y-axis sensor; 14. Z-axis servo motor; 15. Z-axis displacement sensor; 17. Control board; 18. Motion control module. DETAILED DESCRIPTION

[0054] The exemplary embodiments, features, and aspects of the present invention will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise indicated.

[0055] Please refer to Figure 1 , Figure 1 A method for detecting nozzle blockage based on machine learning in an embodiment of the present application is used to detect nozzle blockage, comprising the following steps:

[0056] S1. Clean the print head, obtain standard inkjet printing parameters, move the print head to the test printing area for test jetting, and detect the corresponding ink droplet landing points of all nozzles of the print head to ensure that each nozzle is working normally;

[0057] S2. Data collection and annotation: collecting piezoelectric sensor signal data of the printhead under standard inkjet printing parameters and different nozzle blockage distribution conditions, and annotating the data;

[0058] S3, data preprocessing and feature extraction, preprocessing the collected piezoelectric sensor signal data and extracting key features from the piezoelectric sensor signal data;

[0059] S4. Build a machine learning model, combine the piezoelectric sensor signal data characteristics as input, and the distribution of nozzle blockage as output. Use the transfer learning method to establish a nozzle blockage distribution state prediction model, and perform model training and testing.

[0060] Among them, in step S1, all the nozzle holes of the nozzle are cleaned and unobstructed to reduce the possibility of nozzle hole blockage, and the nozzle is wiped to remove residual ink on the surface of the nozzle to avoid subsequent clogging of the nozzle holes by residual ink, thereby ensuring subsequent normal testing.

[0061] Wherein, step S1, obtaining standard inkjet printing parameters, specifically includes:

[0062] The nozzle drive waveform data and the preset nozzle ejection standard speed and standard volume of ink droplets; the nozzle drive waveform data is usually a trapezoidal wave, such as Figure 2 As shown, including the delay time t d , rise time t r , hold time t h , fall time t f And holding voltage V h During the inkjet printing process, any change in the above parameters will affect the inkjet quality of the printhead. In order to ensure the stability of the printhead inkjet quality in subsequent tests, the inkjet printing parameters need to be standardized after the printhead structure and ink characteristics are determined.

[0063] Wherein, step S1, detecting whether all nozzles can normally eject ink droplets, includes:

[0064] Generate printing pattern data that matches the printing of all nozzle holes of the nozzle to print and solidify the ink droplets, and move the substrate from the printing area to the detection area, such as Figure 3 As shown, the printed pattern on the substrate is photographed and detected, and the ink droplets in the detection image are counted and compared with the number of all nozzles of the printhead. Normally, the distance between adjacent ink droplets in the x direction is Δx, and the distance between adjacent ink droplets in the y direction is Δy. If the distance between each droplet and the adjacent droplet is greater than 1.1 times Δx, Δy or less than 0.9 times the detection Δx, Δy, as shown in the figure, Figure 3 In case 1, the nozzle corresponding to the droplet is sprayed obliquely, which may be caused by partial blockage of the nozzle; if the number of ink drop counts is less than the number of all nozzles in the printhead, the nozzle corresponding to the missing droplet position is blocked, such as Figure 3 Case 2: The nozzle sprays obliquely or is blocked and needs to be cleaned again to ensure that all nozzle holes can work normally.

[0065] Among them, step S2, collecting the piezoelectric sensor signal data of the nozzle under standard inkjet printing parameters and different nozzle blockage distribution conditions, and annotating the data, specifically includes the following steps:

[0066] The nozzle holes are divided and blocked sequentially, so that the nozzle prints according to the printing pattern data corresponding to the ejection of all nozzle holes. The printed image on the substrate is photographed and detected repeatedly each time, and the image algorithm is processed to determine whether the actual nozzle hole blockage distribution state is consistent with the preset nozzle hole blockage distribution state. If consistent, the piezoelectric sensor signal data during the nozzle printing process is collected and the data is annotated with the corresponding nozzle hole blockage distribution state.

[0067] Among them, such as Figure 2 As shown in the figure, the process of collecting the piezoelectric sensor signal data of the nozzle is to use the piezoelectric sensor to detect the driving waveform of different pressure pulses (delay time t d ) is used as a pressure sensor to monitor the pressure fluctuations in the ink chamber caused by the previous drive waveform of the nozzle and the current nozzle blockage. The electrical signal collected by the piezoelectric measurement is used as acoustic signal data. When the piezoelectric sensor is used as a pressure sensor, the mode switching from pressure to electrical signal and back again takes approximately several milliseconds. The driving waveform of the pressure pulse needs to be as short as possible, and the interval between the two driving waveforms should be as long as possible.

[0068] Specifically, for a nozzle with an ignition frequency of 20 kHz, the single drive waveform maintenance time (holding time t h ) In 15μs, the complete time of a single drive waveform (rise time t r, hold time t h , fall time t f ) is 20μs, then the interval between adjacent driving waveforms (delay time t d ) is 30μs, the piezoelectric sensor mode switching time is 5μs, the effective signal acquisition time is 20μs, and the signal resolution can reach 0.1μs.

[0069] The specific operations of dividing the nozzle holes and blocking them in sequence include:

[0070] like Figure 4 As shown, the nozzle positions are divided into nine sections. Each section is blocked one by one, and the blocked positions of the individual nozzles in each section are arranged and combined to form a complete blockage distribution for each section. After traversing all blockage distributions for each section, if the detected actual blockage distribution matches the preset blockage distribution, the corresponding piezoelectric sensor signal data is collected, and the blocking operation is repeated for the next section.

[0071] Among them, step S3, data preprocessing, uses a wavelet threshold noise reduction method to filter out interference while retaining signal details, so as to facilitate subsequent feature extraction of the signal.

[0072] Wavelet transform replaces the infinite-length trigonometric function basis with a wavelet basis whose effective length will decay, and realizes the local amplification analysis of non-stationary signals through the translation and expansion of the wavelet basis.

[0073] Wavelet transform uses the wavelet basis function ψ a，b (t)

[0074] Transform the signal. For any signal f(t), its continuous wavelet transform formula is as follows:

[0075]

[0076] Where a>0 is called the scale factor, which is used to adjust the basic wavelet function ψ a，b (t) expands and contracts, and b reflects the displacement, which can be positive or negative.

[0077] The wavelet threshold denoising method includes the following specific steps:

[0078] After wavelet decomposition, a series of wavelet decomposition coefficients are obtained. The wavelet function and the number of decomposition layers are determined according to the signal, and the signal is divided into low-frequency components and high-frequency components. Signals with correlation greater than the set threshold are concentrated in the high-amplitude wavelet decomposition coefficients. The noise of the high-frequency component is in the wavelet domain, and the wavelet decomposition coefficient value is low. According to the set threshold, the signal below the threshold is regarded as a noise signal and removed. By processing the high-frequency and low-frequency components of the threshold function, the signal is reconstructed using the inverse wavelet transform to achieve noise reduction.

[0079] Correlation is the correlation coefficient, which is defined as the covariance of X and Y divided by the standard deviation of X and Y, representing the similarity of two signals;

[0080]

[0081] The result range of the correlation coefficient is in [0,1]. A correlation coefficient of 1 means that the two signals are most similar, and a correlation coefficient of 0 means that the two signals are completely opposite in similarity. The high-frequency components that need to be reconstructed are determined by the correlation coefficient.

[0082] When the correlation coefficient value is greater than the set threshold, the relevant component needs to be denoised; when it is less than the set threshold, it is regarded as a noise component and the noise signal is removed.

[0083] Specifically, a fixed threshold value selection method is adopted, as shown in the formula below:

[0084] n represents the number of wavelet coefficients in the three-layer decomposition; σ represents the mean square error of the noise signal.

[0085] Select the soft threshold function to perform wavelet decomposition coefficient threshold quantization, as shown in the formula below:

[0086]

[0087] thr(j) is the threshold; C j,k is the decomposition coefficient; are the estimated wavelet coefficients.

[0088] in, The selection of threshold and threshold function has a greater impact on the wavelet threshold algorithm.

[0089] Step S3, data feature processing, uses time-frequency analysis to process the piezoelectric sensor signal data of pressure fluctuations within the ink channel of the printhead, after wavelet threshold noise reduction, to generate a time-frequency acoustic texture image of the pressure fluctuations. This time-frequency acoustic texture image, compared to the single time-domain and frequency-domain images, captures the dynamic characteristics of pressure fluctuations within the ink channel of the printhead.

[0090] Specifically, the time-frequency analysis method uses the Synchrosqueezing Wavelet Transform (SWT). By using the synchronous compression operator to rearrange the time-frequency coefficients, the time-frequency distribution of the signal at any point in the time-frequency plane is moved to the energy center of gravity, resulting in a more concentrated time-frequency transform, which can better resolve the time-frequency ambiguity problem. The specific steps are as follows:

[0091] The continuous wavelet transform formula W(t) for any signal f(t) is as follows:

[0092]

[0093] Where a>0 is called the scale factor, b is the displacement factor, which reflects the displacement and its value can be positive or negative. ψ is an appropriately selected wavelet and redistributes W f (a, b) A centralized time-frequency image is obtained, from which the instantaneous frequency lines are extracted.

[0094] For any W f For any (a, b) where (a, b) ≠ 0, calculate the instantaneous frequency W of the signal f(t) f (a, b) can be expressed as follows:

[0095]

[0096] Synchronous compression redistributes the wavelet transform "time-scale" into "time-frequency", suppresses blurring along the scale and maps the scale to frequency, thereby improving the resolution of the time-frequency sound texture to a limited extent and fully characterizing the dynamic characteristics of the pressure fluctuation state in the ink channel of the nozzle. The calculation formula is as follows:

[0097]

[0098] Where a k is the value of the kth discretization scale factor a; ω l is the lth discrete angular frequency; a k -a k-1 =(Δa) k ;ω l -ω l-1 =Δω.

[0099] Among them, in step S4, the piezoelectric sensor signal data feature in the machine learning model is the time-frequency sound texture image obtained in step S3; the machine learning model is built with the deep residual network model as the basic framework, the time-frequency sound texture images of different nozzle nozzle orifice blockage distribution conditions are divided into training set and test set, the deep residual network is used to perform feature extraction of the time-frequency sound texture image, the machine learning model is trained using the training set, and a deep residual network model that has been preliminarily trained is obtained, which is verified on the test set. Through the fine-tuning method of transfer learning, some parameters of the preliminarily trained deep residual network model are frozen, a small number of parameters are modified and trained and optimized, and a nozzle blockage distribution state prediction model is obtained, which is verified on the test set; the time-frequency sound texture image of the unknown nozzle nozzle orifice blockage distribution condition is input into the nozzle nozzle orifice blockage distribution state prediction model to obtain a model recognition result of the nozzle nozzle orifice blockage distribution condition type.

[0100] Specifically, the final fully connected layer of the deep residual network model was modified to recognize 10 categories (9 partial blockage conditions at the nozzle orifice location in step S2 + 1 unblocked nozzle orifice condition). The parameters of the previous layers of the deep residual network model were all frozen, and training continued using the time-frequency sound texture dataset. The time-frequency sound texture image of the unknown nozzle orifice blockage distribution was transferred to the nozzle orifice blockage distribution prediction model, which output a 1×10 vector representing the probability of each of the ten categories.

[0101] See also Figure 5 The machine learning-based nozzle clogging detection system for inkjet printing includes a nozzle module, an image detection module, an acoustic signal acquisition module, a motion control module, and a data processing module. The nozzle module is used to repeat inkjet tests based on different nozzle clogging distribution settings; the image detection module is used to observe the ink droplet landing point position corresponding to the nozzle during the inkjet process and detect whether the set nozzle clogging situation is consistent with the actual ink discharge situation of the nozzle; the acoustic signal acquisition module is used to collect piezoelectric sensor signal data from the nozzle ignition to the ink droplet landing stage; the motion control module is used to control the motion state of the nozzle and substrate according to the inkjet state; and the data processing module is used to process the piezoelectric sensor signal data and construct a machine learning model.

[0102] The nozzle module includes: a nozzle 1, a nozzle driving board 2, a nozzle control board 3, a power supply 4, a nozzle nozzle hole blocking plate 5 (the nozzle nozzle hole blocking plate is connected to the nozzle, and is a thin metal flat plate slightly larger than the size of the nozzle. Holes are precisely machined to meet different nozzle hole blocking conditions. The diameter of the holes is slightly larger than the nozzle hole diameter, simulating the nozzle hole blocking conditions in actual production), a host computer 6, and a cleaning and wiping assembly 7 (the cleaning and wiping assembly includes a fine cloth containing a cleaning liquid and a reciprocating device. Before printing, the reciprocating device drives the fine cloth and the nozzle to move adaptively relative to each other, wipe the nozzle surface, and reduce the possibility of nozzle hole blocking other than the set nozzle hole blocking).

[0103] The nozzle control board 3 is connected to the host computer 6, receives the printing data transmitted from the host computer 6, and generates corresponding waveform signals and negative pressure signals; the power supply 4 is connected to the nozzle control board 3, receives the control signal of the nozzle control board 3, generates the high and low levels required for the nozzle ignition sequence and transmits them to the nozzle drive board 2; the nozzle drive board 2 is connected to the nozzle control board 3, and is responsible for the injection state of the nozzle nozzle; the cleaning and wiping component 7 is connected to the motion control module 18, and the motion control module 18 receives the signal from the host computer 6 and drives the cleaning and wiping component 7 to flush and wipe the nozzle after printing the printing data; the nozzle nozzle hole blockage setting component is connected to the nozzle to set the blockage condition of the nozzle nozzle.

[0104] The image detection module includes an optical detection camera 8 and a stroboscopic light source 9; the optical detection camera is installed on one side of the nozzle, and the stroboscopic light source is installed at a coaxial position of the optical detection camera, which is responsible for detecting the corresponding ink droplet landing point position under different nozzle blockage distribution conditions.

[0105] The motion control module includes an X-axis linear motor 10, an X-axis sensor 11, a Y-axis linear motor 12, a Y-axis sensor 13, a Z-axis servo motor 14, a Z-axis displacement sensor 15, etc.; the X-axis linear motor 10 controls the substrate 17 to complete the reciprocating motion of acceleration, uniform speed, and deceleration in the x-direction; the Y-axis linear motor 12 controls the substrate 17 to move a preset step distance along the y-direction after the nozzle completes the printing task to start the next stage of the printing task; the X-axis sensor 11 and the Y-axis sensor 13 are used to measure the feedback of the motion state of the substrate 17; the Z-axis servo motor 14 is responsible for controlling the height adjustment between the nozzle and the substrate 17; and the Z-axis displacement sensor 15 is responsible for transmitting the measurement feedback of the z-axis position information.

[0106] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A nozzle clogging detection method based on machine learning, characterized in that: The method comprises the following steps: Step 1) Clean the nozzle, obtain standard inkjet printing parameters, move the nozzle to the test printing area for test spraying, and detect the corresponding ink droplet landing points of all nozzles of the nozzle to ensure that each nozzle is working properly; Step 2) Data collection and annotation: collecting piezoelectric sensor signal data of the nozzle under standard inkjet printing parameters and different nozzle blockage distribution conditions, and annotating the data; Step 3) Data preprocessing and feature extraction: preprocessing the collected piezoelectric sensor signal data to extract feature information from the piezoelectric sensor signal data for subsequent classifier training and prediction; The process of step 3) is as follows: The preprocessing operation includes using wavelet threshold to reduce noise and filter out interference while retaining signal details: processing the amplitude and frequency domain data of the piezoelectric sensor signal, using wavelet packet analysis to extract signal feature information, and selecting wavelet basis functions to perform wavelet packet transform on the signal; Extracting signal feature information includes generating a time-frequency acoustic texture image using a time-frequency analysis method: using the time-frequency analysis method to process the piezoelectric sensor signal data of the pressure fluctuation in the ink channel of the nozzle after the wavelet threshold noise reduction method to obtain a pressure fluctuation time-frequency acoustic texture image, as follows: The time-frequency analysis method is to use the synchronous compression wavelet transform (SWT). The time-frequency coefficients are rearranged by the synchronous compression operator, and the time-frequency distribution of the signal at any point in the time-frequency plane is moved to the energy center of gravity, obtaining a time-frequency transform with a more concentrated frequency band. The continuous wavelet transform formula W(t) for any signal f(t) is as follows: Where a>0 is the scale factor, b is the shift factor, ψ is the selected wavelet, and W is redistributed f (a, b) Obtain a centralized time-frequency image and extract the instantaneous frequency line from it, where t represents time and f(t) is a function that changes with time; For any W f For any (a,b) where (a,b)≠0, calculate the instantaneous frequency W of the signal f(t) f (a,b) is expressed as follows: Indicates differential, upper and lower pairs b and W f (a,b) is differentiated; represents an imaginary number; Synchrocompression redistributes the wavelet transform time-scale into time-frequency, suppresses blurring along the scale and maps the scale to frequency, as calculated by: Where a k is the value of the kth discretization scale factor a; ω l is the lth discrete angular frequency; ω represents the discrete angular frequency; a k -a k-1 =(Δa) k ;ω l -ω l-1 =Δω; The wavelet threshold denoising process is as follows: After wavelet decomposition processing, a series of wavelet decomposition coefficients are obtained. The wavelet function and the number of decomposition layers are determined according to the signal, and the signal is divided into low-frequency components and high-frequency components. Signals with correlation greater than the set threshold are concentrated in the wavelet decomposition coefficients with high amplitude. The noise of the high-frequency component is in the wavelet domain, and the wavelet decomposition coefficient value is low. According to the set threshold, the signal less than the threshold is regarded as a noise signal and removed. By processing the high-frequency and low-frequency components of the threshold function, the signal is reconstructed using the inverse wavelet transform to achieve noise reduction processing. Correlation is the correlation coefficient, which is defined as the covariance of X and Y divided by the standard deviation of X and Y, representing the similarity of two signals; The result range of the correlation coefficient is in [0,1]. A correlation coefficient of 1 means that the two signals are most similar, and a correlation coefficient of 0 means that the two signals are completely opposite in similarity. The high-frequency components that need to be reconstructed are determined by the correlation coefficient. When the correlation coefficient value is greater than the set threshold, the relevant component needs to be denoised; when it is less than the set threshold, it is regarded as a noise component and the noise signal is removed; Step 4) Build a machine learning model, take the piezoelectric sensor signal data features as input, and the distribution of blocked nozzles of the nozzle as output. Use the transfer learning method to establish a nozzle blockage distribution state prediction model, and perform model training, verification and testing.

2. The nozzle clogging detection method based on machine learning according to claim 1 is characterized in that: The step 1) obtains standard inkjet printing parameters, including: nozzle drive waveform data and the speed and volume of ink droplets ejected by the preset nozzle; the nozzle drive waveform data includes delay time, rise time, hold time, fall time and hold voltage.

3. The nozzle clogging detection method based on machine learning according to claim 1 is characterized in that: The process of step 2) is as follows: All nozzles of the printhead are divided and blocked sequentially, so that the printhead prints according to the print pattern data corresponding to the spray status of all nozzles. The printed image on the printing plate is photographed and detected, and image algorithm processing is performed to determine whether the actual nozzle blockage distribution state is consistent with the preset nozzle blockage distribution state. If consistent, the piezoelectric sensor signal data during the printhead printing process is collected and labeled with the corresponding nozzle blockage distribution state. The operation of dividing all the nozzle holes and blocking them in sequence is as follows: Divide all nozzle positions of the nozzle into multiple regions, perform blocking operations on each region one by one, and arrange and combine the blocked positions of each nozzle in each region to form all situations of nozzle blockage distribution in each region; traverse all situations of nozzle blockage distribution in each region, and when the detected actual nozzle blockage distribution state is consistent with the preset nozzle blockage distribution state, collect the corresponding piezoelectric sensor signal data and repeat the blocking operation for the next region; The nozzle blockage distribution state includes: the number and position distribution of all blocked nozzles in the nozzle; Among them, the piezoelectric sensor signal is used by the piezoelectric actuator to actively detect channel acoustics. The electrical signal collected by the piezoelectric measurement is used as acoustic signal data. When the piezoelectric actuator is not driven by an external waveform, the piezoelectric actuator is used to detect internal channel pressure fluctuations.

4. The nozzle clogging detection method based on machine learning according to claim 1 is characterized in that: The process of step 4) is as follows: The piezoelectric sensor signal data feature in the machine learning model is the time-frequency acoustic texture image obtained in step 3); the machine learning model is built with a deep residual network model as the basic framework, the time-frequency acoustic texture images of different nozzle nozzle orifice blockage distribution conditions are divided into a training set and a test set, the deep residual network is used to perform feature extraction of the time-frequency acoustic texture image, and the machine learning model is trained using the training set to obtain a preliminarily trained deep residual network model, which is then verified by the test set; Through the fine-tuning method of transfer learning, some parameters of the deep residual network model that has completed the preliminary training are frozen, and the other parameters are modified and trained and tuned to obtain a nozzle clogging distribution state prediction model, which is verified on the test set; the time-frequency acoustic texture image of the unknown nozzle orifice clogging distribution is input into the nozzle orifice clogging distribution state prediction model to obtain the model recognition result of the nozzle orifice clogging distribution type.

5. The nozzle clogging detection system according to the machine learning-based nozzle clogging detection method according to claim 1, characterized in that: include: Printhead module, used to repeat inkjet test according to different settings of nozzle blockage distribution; The image detection module is used to observe the ink droplet landing point corresponding to the nozzle during the inkjet process and detect whether the nozzle blockage condition is consistent with the actual ink discharge condition of the nozzle; The acoustic signal acquisition module is used to collect the piezoelectric sensor signal data from the ignition of the printhead to the droplet falling of the ink; A motion control module, used to control the motion state of the nozzle and substrate according to the inkjet state; The data processing module is used to process piezoelectric sensor signal data and build a machine learning model.

6. The nozzle clogging detection system according to claim 5, characterized in that: The image detection module includes an optical detection camera and a stroboscopic light source; the optical detection camera is installed on one side of the nozzle, and the stroboscopic light source is installed coaxially with the optical detection camera, which is responsible for detecting the corresponding ink droplet landing point positions under different nozzle blockage distribution conditions.

Images