A method for intelligent control of a production process of a customized pattern of a mobile phone shell

Abstract

The application relates to a mobile phone shell customized pattern production process intelligent control method, which comprises the following steps: acquiring a micro-distance image sequence of a substrate surface after ink drop ejection; using a pre-trained convolutional neural network to extract an actual diffusion radius of the ink drop, a non-circularity deviation and an edge blur bandwidth; comparing the extracted features with an ideal template to generate a mismatch mode label and a quantitative index of transverse over-diffusion or longitudinal under-wetting; driving a path deformation engine to generate a direction-controllable and amplitude-adaptive normal offset disturbance quantity, performing sub-pixel level dynamic deformation mapping on a subsequent ejection point position, and obtaining a corrected nozzle track. Through in-situ sensing and closed-loop path self-correction of the ink drop deposition behavior, the application effectively suppresses the pattern edge precision decline caused by environmental disturbance, and improves the consistency and fidelity of the customized pattern.

Description

Technical Field

[0001] This invention relates to the field of intelligent printing equipment control technology, and in particular to an intelligent control method for the production process of customized patterns on mobile phone cases. Background Technology

[0002] With the increasing demand for personalized mobile phone cases, digital inkjet printing technology has been widely used in mobile phone case pattern production. Existing intelligent printing equipment typically employs offline path planning methods for printhead motion control, relying on pre-generated vector trajectory files to perform droplet deposition with fixed path parameters. Mainstream technical solutions include path optimization strategies such as rotating printhead structures, multi-nozzle collaborative jetting, printhead acceleration compensation, and surface fitting interpolation, primarily improving printing speed and pattern reproduction capabilities through mechanical structure innovation or path algorithms. Some high-end systems integrate environmental sensors to collect external data such as temperature, humidity, and material properties, achieving a certain degree of environmental compensation in the backend path generation stage. However, these technologies largely rely on indirect measurements of external physical parameters or static analysis of historical data, lacking direct feedback on the actual droplet deposition process and making it difficult to adapt to dynamic changes in the substrate microenvironment and material ink absorption between batches.

[0003] Current printhead path planning typically generates a global trajectory under ideal conditions, failing to consider the real-time effects of ambient temperature and humidity, as well as the ink absorption properties of materials. This leads to uneven ink spread, excessive edge diffusion, or insufficient wetting when local environmental disturbances occur, resulting in blurred edges, distorted lines, and reduced precision in the printed output, impacting the consistency and quality of customized products. Furthermore, existing compensation schemes often employ offline modeling and parameter correction logic, which struggles to handle real-time changes in small-batch, cross-material, and multi-batch production line environments. Their compatibility and response speed fail to meet the demands of flexible production. Various mobile phone case materials, such as glass, PC, and TPU, exhibit significant differences in ink droplet wetting and diffusion behavior, and the ink absorption properties of the same material fluctuate considerably under different temperatures and humidity levels. Traditional compensation measures suffer particularly significant precision degradation when dealing with complex patterns, gradient colors, or fine edge processing, making it difficult to achieve dynamic and precise control under multiple operating conditions.

[0004] In summary, existing printhead paths lack dynamic environmental adaptability and cannot accurately compensate based on real-time feedback from droplet deposition behavior, resulting in limited edge fidelity. External environmental sensors and offline modeling struggle to capture the transient dynamics of droplet spreading, often leading to lag or mismatch in compensation measures. Path optimization logic often focuses on mechanical structure upgrades or software algorithm iterations, failing to fully integrate the physical feedback of the droplet deposition process with the path self-correction mechanism, making it difficult to maintain consistent product quality on flexible customized production lines. Therefore, it is necessary to propose a real-time printhead path deformation compensation method centered on droplet deposition dynamics feedback. Through in-situ sensing of droplet spreading behavior, structured feature extraction, and closed-loop path self-correction, this method systematically addresses the issue of decreased pattern accuracy caused by environmental disturbances without increasing hardware complexity or sacrificing throughput efficiency, thereby improving print quality consistency and edge fidelity. Summary of the Invention

[0005] This application provides an intelligent control method for the production process of customized patterns on mobile phone cases, aiming to solve the problem that existing printhead path planning relies on offline presets or indirect inference from environmental sensors, and cannot provide real-time feedback and path self-correction based on the actual deposition behavior of ink droplets, resulting in loss of control over the edge accuracy of the pattern under environmental disturbances.

[0006] This application provides an intelligent control method for the production process of customized patterns on mobile phone cases, specifically including: Acquire a sequence of macro images of the substrate surface within a preset time window after the printhead assembly performs a single ink droplet ejection action; The macro image sequence is input into a convolutional neural network model pre-trained with multiple materials and multiple environmental conditions to extract and output three key deposition feature parameters that characterize the current ink droplet deposition behavior: actual diffusion radius, non-circularity deviation value, and edge blur band width. Based on the comparison and analysis between the key sedimentary characteristic parameters and the ideal sedimentary templates stored under standard reference conditions, a mismatch pattern type label and a quantitative index of the degree of mismatch are generated to identify whether there is lateral over-diffusion or longitudinal under-wetting in the current local area. Using the mismatch mode type label and mismatch degree quantification index as the driving basis, a path deformation engine is constructed that does not change the geometric topology of the original vector path. The path deformation engine is used to calculate the normal offset disturbance amount that is directionally controllable and amplitude adaptively applied to the original path segment. Based on the normal offset disturbance, subpixel-level dynamic deformation mapping processing is performed on the coordinates of subsequent injection points to generate a corrected nozzle path trajectory that compensates for the influence of environmental disturbances. The intelligent control method for the production process of customized patterns on mobile phone cases provided in this application has the following beneficial effects: By integrating a miniature high-speed optical imaging unit and a lightweight convolutional neural network into the printhead assembly, in-situ perception and precise quantification of deposition features such as droplet diffusion radius, non-circularity deviation, and edge blur bandwidth are achieved, bypassing the lag and mismatch problems caused by the indirect inference of environmental sensors in traditional solutions. By driving the path deformation engine with mismatch mode type labels and quantification indicators, a normal offset perturbation with controllable direction and adaptive amplitude is generated without changing the original vector path topology. Dynamic deformation mapping of the spray point is performed with sub-pixel accuracy, correcting the edge offset caused by temperature and humidity fluctuations and substrate differences from the source, effectively suppressing defects such as edge blur, color block overflow, and detail distortion. This method is applicable to various mobile phone case substrates such as polycarbonate, thermoplastic polyurethane, and glass, without adding additional mechanical complexity or external physical intervention. It is especially suitable for flexible production needs of small batches, cross-materials, and multiple batches, and can significantly improve the consistency and pass rate of customized pattern products. Attached Figure Description

[0007] Figure 1 This is the main flowchart of an intelligent control method for the production process of custom phone case designs.

[0008] Figure 2 This is a sub-flowchart of an intelligent control method for the production process of custom phone case designs.

[0009] Figure 3 This is another sub-flowchart of a smart control method for the production process of custom phone case designs. Detailed Implementation

[0010] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0012] like Figure 1 As shown, this application provides an intelligent control method for the production process of customized patterns on mobile phone cases, specifically including: S1: Acquire a sequence of macro images of the substrate surface within a preset time window after the printhead assembly performs a single ink droplet ejection action. The macro image sequence includes the initial wetting morphology, edge climbing dynamics, and final static contour evolution data of the ink droplet after contacting the substrate.

[0013] S2: Input the macro image sequence into a convolutional neural network model pre-trained with multiple materials and multiple environmental conditions, and extract and output three key deposition feature parameters that characterize the current ink droplet deposition behavior: actual diffusion radius, non-circularity deviation value, and edge blur band width.

[0014] S3: Based on the comparison and analysis of key sedimentary characteristic parameters and ideal sedimentary templates pre-stored under standard reference conditions, generate mismatch pattern type labels and quantitative indicators of mismatch degree that identify whether there is lateral over-diffusion or vertical under-infiltration in the current local area.

[0015] S4: Using the mismatch mode type label and mismatch degree quantification index as the driving basis, a path deformation engine is constructed that does not change the geometric topology of the original vector path. The path deformation engine is used to calculate the normal offset disturbance applied to the original path segment with controllable direction and adaptive amplitude.

[0016] S5: Perform sub-pixel-level dynamic deformation mapping processing on the coordinates of subsequent injection points based on the normal offset disturbance to generate a corrected nozzle path trajectory that compensates for the effects of environmental disturbances.

[0017] Step S1: Acquire a sequence of macro images of the substrate surface within a preset time window after the printhead assembly performs a single ink droplet ejection action. The macro image sequence includes data on the initial wetting morphology, edge climbing dynamics, and final static contour evolution of the ink droplet after contacting the substrate. Specifically, it includes: S1.1: Based on the trigger signal of a single ink droplet ejection action executed by the printhead assembly, the synchronous pulse illumination module is activated to emit a high-frequency strobe light source to perform nanosecond-level exposure control on the ink droplet landing area, so as to eliminate the dynamic blur caused by the printhead movement and generate a high-contrast transient illumination environment.

[0018] It receives a single droplet ejection trigger signal from the printhead controller. This signal serves as a timing synchronization reference, marking the instant the droplet leaves the nozzle orifice.

[0019] The rising edge timestamp of the trigger signal is analyzed, the flight delay time from ink droplet ejection to the expected contact surface of the substrate is calculated, and the start-up time slot of the synchronous pulse illumination module is determined.

[0020] A high-frequency strobe drive command is sent to the synchronous pulse illumination module to control the LED array to emit a high-intensity cold light source within a nanosecond time window, ensuring that the illumination duration is shorter than the time required for the nozzle to move one pixel.

[0021] By using an optical path coupling system, the stroboscopic light source is focused onto the area where the ink droplets fall, forming a local high-brightness transient illumination field and suppressing the interference of ambient background light on the imaging sensor.

[0022] Nanosecond-level exposure control is used to freeze the spatial position of ink droplets in high-speed motion, eliminating motion blur caused by the printhead's lateral scanning speed.

[0023] By adjusting the luminous intensity and incident angle of the pulsed light source, the reflectivity difference between the ink droplet edge and the substrate surface is enhanced, generating a transient lighting environment with high signal-to-noise ratio and high contrast.

[0024] By using the above processing method, the trigger signal from the previous step is transformed into a high-contrast transient illumination condition that eliminates dynamic blur, thereby achieving the expected technical effect of clearly capturing the ink droplet deposition morphology.

[0025] For example, the printhead printing speed is set to 1.5 m / s, and the droplet flight time is approximately 200 microseconds. After detecting the jet trigger signal, pulse illumination is initiated after a 180-microsecond delay, with a pulse width set to 50 nanoseconds and a light intensity of 5000 lumens. At this time, the printhead displacement during exposure is less than 0.075 micrometers, far below the 42-micrometer pixel size corresponding to a 600 dpi printing resolution, effectively eliminating motion blur and significantly improving image edge sharpness.

[0026] S1.2: Using a miniature high-speed optical imaging unit integrated into the printhead assembly, under high-contrast transient illumination, a continuous frame image capture operation is performed on the ink droplet landing area on the substrate surface according to a preset sampling frequency to obtain a sequence of original macro image frames containing timestamp information.

[0027] It receives a high-contrast transient lighting environment signal triggered by the synchronous pulse lighting module, and activates the miniature high-speed CMOS imaging sensor integrated at the end of the nozzle assembly to enter the external trigger acquisition mode.

[0028] The global shutter exposure time of the imaging sensor is strictly matched with the pulse width of the synchronous stroboscopic source, and the exposure time is set to 500 nanoseconds to freeze the microscopic morphological changes of ink droplets during the high-speed spreading process on the substrate surface.

[0029] Based on the preset ink droplet deposition kinetic time constant, the image capture sampling frequency is set to 20kHz to ensure that at least 1000 frames of continuous image data are acquired within the critical 50-millisecond window period after the ink droplet contacts the substrate.

[0030] The on-chip timing controller of the imaging sensor is activated to attach a high-precision hardware timestamp to each frame of the captured raw RAW format image data. The timestamp resolution is better than 1 microsecond, establishing a precise mapping relationship between image frames and physical time.

[0031] The raw macro image frame sequence containing timestamp information is transmitted in real time to the dedicated DMA buffer of the embedded AI coprocessor via the MIPICSI-2 high-speed serial interface, avoiding data frame loss caused by CPU interrupt delay.

[0032] Integrity checks are performed on the transmitted image frame sequence, damaged frames caused by signal interference are marked and removed, and the original macro image frame sequence with continuous temporal sequence and complete metadata is retained as input for subsequent processing.

[0033] By using the high-frequency synchronous capture and timestamp marking processing method described above, the transient illumination environment generated in the previous step is transformed into a sequence of original macro images with extremely high spatiotemporal resolution, realizing a lossless digital record of the entire ink droplet deposition process, and providing a high-fidelity data foundation for subsequent extraction of sub-pixel-level deposition feature parameters.

[0034] For example, a global shutter CMOS sensor with a resolution of 1280x1024 pixels and a frame rate of up to 20,000 fps was selected. When printing a TPU phone case, the light source pulse width was set to 500 ns, and the sensor exposure time was simultaneously set to 500 ns. Within a 50 ms window after ink droplet ejection, 1000 frames of images were continuously acquired at a frequency of 20 kHz. Each frame of image was accompanied by a 64-bit hardware timestamp. Data was transmitted at a rate of 2.5 Gbps via the MIPI interface. After verification, 998 valid original images were successfully acquired, completely recording the entire process of ink droplet contact, spreading, and stillness. The image signal-to-noise ratio was higher than 40 dB, meeting the requirements for subsequent neural network feature extraction.

[0035] S1.3: Based on the timestamp information of the original macro image frame sequence, select the valid image frames located within the preset time window after a single ink droplet ejection action, and remove invalid background frames to construct a time-aligned candidate macro image subsequence.

[0036] The hardware timestamp information embedded in the header of each data frame in the original macro image frame sequence is analyzed to extract the absolute time offset relative to the trigger time of a single ink droplet ejection action.

[0037] Based on the rheological properties of ink on a specific substrate surface and a preset observation window for deposition kinetics, an effective time threshold range for data collection is set, which covers the complete evolution cycle from the moment the ink droplet contacts the substrate to when the profile tends to stabilize.

[0038] Iterate through each frame of the original sequence, compare its time offset with a preset time threshold range, and identify and mark the target image frames that fall within the effective window.

[0039] Background frames, pre-spray frames, and post-steady-state redundant frames with timestamps outside the effective window are removed to eliminate the interference of irrelevant visual data on subsequent feature extraction.

[0040] The selected target image frames are arranged in ascending order of timestamps to construct a temporally aligned and continuous candidate macro image subsequence.

[0041] By using the above-mentioned time-domain filtering method, the original macro image frame sequence containing a large amount of invalid data is transformed into an effective candidate macro image subsequence containing only key depositional evolution stages, thus achieving the expected technical effect of data simplification and precise temporal alignment.

[0042] For example, the ink droplet ejection trigger time is set to T0, and the preset effective observation window is [T0+5ms, T0+45ms]. The original sequence contains 100 frames with a sampling interval of 1ms. The system reads the timestamp Ti of each frame and calculates Δt = Ti - T0. If 5ms ≤ Δt ≤ 45ms, the frame is retained; otherwise, it is discarded. Finally, 41 frames from frame 6 to frame 46 are selected and reassembled into candidate subsequences in chronological order. This process removes more than 90% of invalid background data, ensuring that the subsequent neural network only processes the high-value dynamic wetting process, significantly reducing the computational load and improving the targeting of feature extraction.

[0043] S1.4: Perform multi-frame registration and denoising on the candidate macro image subsequences to correct pixel displacement caused by mechanical vibration and suppress sensor thermal noise, so as to generate a standardized macro image sequence with consistent spatial coordinates and enhanced signal-to-noise ratio.

[0044] Receive candidate macro image subsequences after time-series filtering. These subsequences include mechanical vibration displacement caused by the high-speed reciprocating motion of the nozzle and thermal noise interference introduced by the CMOS sensor under high-frequency sampling.

[0045] A phase-correlation-based frequency domain registration algorithm is adopted to perform Fourier transform on adjacent frames in the subsequence to obtain spectral information, calculate the normalized inverse Fourier transform result of the cross power spectrum, and accurately locate the sub-pixel level translation vector between frames.

[0046] Based on the calculated translation vector, the subsequent frame images are spatially geometrically corrected using bicubic interpolation to eliminate pixel coordinate drift caused by mechanical resonance and ensure that the ink droplet landing point remains spatially aligned in all frames.

[0047] For the corrected image sequence, a nonlocal mean denoising model is constructed. By calculating the weighted average of Euclidean distances between image blocks, random thermal noise is suppressed while preserving high-frequency texture details at the edges of ink droplets.

[0048] A temporal median filtering mechanism is introduced to sort and extract gray values ​​at the same pixel location along the time axis, thereby eliminating isolated noise points caused by transient fluctuations in illumination and enhancing the temporal continuity of the deposition process.

[0049] Through the above multi-frame registration and joint denoising process, the original candidate subsequences with spatial misalignment and noise pollution are transformed into standardized macro image sequences with strictly consistent spatial coordinates and significantly enhanced signal-to-noise ratio, providing a clean data foundation for subsequent high-precision deposition feature extraction.

[0050] For example, the image resolution is set to 512×512 pixels, and the sampling frequency is 1000fps. The maximum inter-frame vibration displacement is detected to be 3.5 pixels. After phase correlation registration, the residual displacement is reduced to within 0.05 pixels. For thermal noise with a variance of 15, nonlocal mean denoising with a window size of 5×5 is applied, improving the peak signal-to-noise ratio from 28dB to 42dB. Temporal median filtering effectively eliminates abnormal frames with brightness abrupt changes exceeding 20 gray levels. In the final output normalized sequence, the clarity of ink droplet edge gradients is significantly improved, with no visible artifacts, meeting the stringent requirements of lightweight convolutional neural networks for input data consistency.

[0051] S1.5: Based on the standardized macro image sequence, the initial wetting morphology data, edge climbing dynamic trajectory data and final static contour evolution data of the ink droplet after contacting the substrate are extracted and encapsulated to form a macro image sequence characterizing the complete deposition process as the output result.

[0052] The system receives a normalized macro image sequence processed by S1.4, containing timestamp-aligned and denoised multi-frame data of droplet deposition. Binarization thresholding is performed on the first frame to extract the initial pixel connected components at the droplet-substrate interface, calculating their centroid coordinates and equivalent circle diameter to establish the initial wetting morphology reference point. For the intermediate frames, optical flow is used to track the displacement vector field of the droplet edge pixels, constructing a time-varying edge-climbing dynamic trajectory matrix to quantify the radial spreading rate of the droplet on the substrate surface. The Canny edge detection operator is applied to the last frame to extract the closed contour line of the droplet in its stationary state, and the geometric parameters of the final stationary contour are fitted. The initial wetting morphology reference point, the edge-climbing dynamic trajectory matrix, and the final stationary contour geometric parameters are spatiotemporally correlated and encapsulated into a structured data object containing complete deposition evolution information. Through the aforementioned image processing and feature extraction chain, the standardized image data obtained in the previous step is transformed into a macro image sequence characterizing the physical deposition behavior of ink droplets. This achieves high-precision digital reconstruction of the entire process of ink droplets from contact and spreading to stillness, providing temporal input data with clear physical meaning for subsequent neural network feature extraction.

[0053] For example, a normalized grayscale image sequence with a resolution of 1024x1024 pixels is processed. The first frame image has a grayscale threshold of 128, and the initial connected component centroid coordinates (512, 512) are extracted, with an equivalent circle diameter of 0.05 mm. The middle 10 frames use the Lucas-Kanade optical flow method, setting a window size of 5x5 and iterating 3 times to obtain an average radial velocity of 2 m / s for the edge pixels. The last frame image uses Canny detection (low threshold 50, high threshold 150) to extract the closed contour, fitting an ellipse with a major axis of 0.12 mm and a minor axis of 0.11 mm. The above data is encapsulated into a JSON object containing "timestamp", "initial_wetting", "spreading_trajectory", and "final_contour" fields, and output to the S2 module.

[0054] like Figure 2 As shown, step S2 involves inputting the macro image sequence into a convolutional neural network model pre-trained for multiple materials and environmental conditions. This model extracts and outputs three key depositional feature parameters characterizing the current ink droplet deposition behavior: the actual diffusion radius, the non-circularity deviation value, and the width of the edge blur band. Specifically, this includes: S2.1: Perform pixel-level normalization and noise suppression preprocessing on the macro image sequence to eliminate ambient light fluctuations and sensor thermal noise interference, and generate a standardized grayscale image matrix as clean input data for subsequent feature extraction.

[0055] The system receives a sequence of macro images output from S1.5, which contains spatiotemporal evolution data of the entire ink droplet deposition process. Color space conversion is performed on the original image frames, mapping the RGB three-channel data to a single-channel grayscale space to eliminate brightness inconsistencies caused by differences in the spectral response of the color sensor.

[0056] A histogram equalization algorithm is used to enhance the global contrast of the grayscale image, stretching the pixel value distribution range and highlighting the grayscale gradient difference between the ink droplet edge and the substrate background. To address sensor thermal noise and ambient light shot noise, a nonlocal mean denoising algorithm is applied. This algorithm uses weighted average filtering based on the similarity weights between image blocks to suppress high-frequency random noise interference while preserving the sharpness of the ink droplet edges. Min-Max normalization is then performed on the denoised image, linearly mapping the pixel grayscale values ​​to the [0,1] interval to unify the data dimensions under different illumination intensities.

[0057] It should be noted that the convolutional neural network model pre-trained under multiple materials and environmental conditions is a deep convolutional neural network used to regress and predict ink droplet deposition feature parameters from macro images. This model can employ an encoder-regressor architecture. The encoder part consists of multiple stacked convolutional layers, pooling layers, and spatial attention layers, used to extract high-frequency texture information and low-frequency contour information of the ink droplet wetting edge from the input image. The regressor part consists of flattening layers, dropout layers, and multiple fully connected layers, used to map the high-dimensional deposition feature tensor output by the encoder to a three-dimensional physical space metric domain. The model's input is a pre-processed, normalized grayscale image matrix, and the output consists of three scalar values: the actual diffusion radius, the non-circularity deviation value, and the width of the edge blur band.

[0058] Through the pixel-level normalization and noise suppression preprocessing described above, the original image sequence from the previous step is transformed into a standardized grayscale image matrix, achieving illumination invariance and high signal-to-noise ratio of the input data, thus providing a clean feature extraction foundation for the subsequent lightweight convolutional neural network.

[0059] For example, a macro image sequence with a resolution of 640×480 is first converted into an 8-bit single-channel grayscale image. The grayscale histogram of the entire image is calculated, and histogram equalization is performed, stretching the pixel values ​​originally concentrated in the 50-150 range to the full range of 0-255, significantly enhancing the contrast of the ink droplet edges. Non-local mean denoising is used, with the search window size set to 21×21, the similarity block size to 7×7, and the filter intensity parameter h set to 10, effectively removing thermal noise generated under ISO800, and improving the peak signal-to-noise ratio to over 35dB. Finally, Min-Max normalization is performed, linearly mapping the pixel grayscale values ​​to the [0,1] range. The processed standardized grayscale matrix eliminates the influence of workshop lighting fluctuations, and the ink droplet outlines are clearly visible, meeting the input requirements of the neural network.

[0060] S2.2: Based on the standardized grayscale image matrix, perform multi-scale feature mapping operation on the convolutional neural network model pre-trained with multiple materials and multiple environmental conditions. Use deep convolutional kernels to extract high-frequency texture information and low-frequency contour information of ink droplet wetting edges to generate a deposition feature tensor containing spatial location encoding.

[0061] It should be noted that the multiple materials include at least three commonly used base materials for mobile phone cases: polycarbonate, thermoplastic polyurethane, and glass; the multiple environmental conditions cover at least the printing environment range of temperature from 15℃ to 35℃ and relative humidity from 30% to 80%, as well as two printhead movement modes: high-speed printing and low-speed printing; the pre-training process involves simultaneously acquiring macro image sequences of the ink droplet deposition process under a combination of multiple materials and multiple environmental conditions, and manually annotating the true values ​​of the three feature parameters: actual diffusion radius, non-circularity deviation value, and edge blur width, forming an annotated dataset for supervised training, so that the network has the ability to extract features across materials and across conditions.

[0062] Receive the normalized grayscale image matrix generated by S2.1 preprocessing and use it as the input tensor of the convolutional neural network model.

[0063] By using the shallow convolution kernel group of the first layer of the model to perform high-frequency edge detection operation on the input matrix, the gradient abrupt change features at the interface between the ink droplet and the substrate are extracted, and a high-frequency feature map containing the micro-texture of the wetting front is generated.

[0064] In parallel, deep dilated convolution kernels are used to perform low-frequency semantic extraction operations on the same input matrix to capture the overall spreading shape of ink droplets and the color saturation distribution in the central region, generating a low-frequency feature map that represents the macroscopic contour.

[0065] A spatial attention mechanism module is introduced to calculate the correlation weights between high-frequency and low-frequency feature maps in the channel dimension, thereby suppressing background noise interference and enhancing the feature response intensity of the ink droplet edge region.

[0066] The weighted high-frequency feature map and the low-frequency feature map are spliced ​​and fused in the depth dimension to construct an intermediate feature cube containing multi-scale spatial information.

[0067] The pixel coordinate information is embedded into the intermediate feature cube through the position encoding layer to generate a deposited feature tensor with absolute spatial location identification.

[0068] By using a multi-scale feature mapping process, the clean image data from the previous step is transformed into a deposition feature tensor containing spatial location encoding, thereby achieving a joint representation of the micro-texture and macro-contour of ink droplets and providing a highly robust feature foundation for subsequent regression prediction.

[0069] For example, for an ink droplet image of a TPU phone case at 25℃ and 60% humidity, the input size is 64x64 pixels. A shallow convolutional kernel of size 3x3 extracts edge gradient features with a width of 2 pixels; a deep dilated convolutional kernel with an inflation rate of 2 extracts overall contour features with a diameter of approximately 15 pixels. A spatial attention mechanism assigns a weight of 0.85 to the edge region and a weight of 0.1 to the background region. The fused deposition feature tensor has a dimension of 16x64x64, containing precise coordinate information of the ink droplet wetting front, significantly improving the extraction accuracy of non-circularity deviation values.

[0070] S2.3: Perform fully connected layer regression prediction calculation based on the deposition feature tensor, map the high-dimensional abstract features to the physical space metric domain, and respectively calculate the actual diffusion radius value representing the ink droplet spreading range, the non-circularity deviation value representing the shape regularity, and the edge blur band width value representing the boundary clarity.

[0071] The system receives a deposition feature tensor containing spatial location encoding generated by S2.2. This tensor integrates high-frequency texture information and low-frequency contour information of the ink droplet wetting edge, and serves as input data for regression prediction of the fully connected layer.

[0072] The high-dimensional deposition feature tensor is flattened into a one-dimensional feature vector, and some neuron connections are randomly dropped through the Dropout layer to suppress overfitting and enhance the model's ability to generalize to the differences in microstructure of different material surfaces.

[0073] The first fully connected layer is used to perform nonlinear mapping on the one-dimensional feature vector, and the ReLU activation function is used to extract deep abstract features, thus establishing the implicit correlation between image pixel distribution and physical deposition morphology.

[0074] The abstract features are projected onto the three-dimensional physical metric space through the second fully connected layer, corresponding to the three output nodes of actual diffusion radius, non-circularity deviation value and edge blur width, respectively, thus achieving decoupling from visual features to physical parameters.

[0075] The mean squared error loss function is used to constrain the deviation between the predicted value and the true labeled value, and the network weights are updated through the backpropagation algorithm to ensure the accuracy of the output parameters at the sub-pixel level.

[0076] The values ​​of the three nodes in the output layer are linearly inversely normalized, and the dimensionless network output is converted into a specific value with the physical unit micrometer based on the pre-trained calibration coefficients.

[0077] By using regression prediction calculations of fully connected layers, the high-dimensional abstract features of the previous step are transformed into the actual diffusion radius value representing the spread range of ink droplets, the non-circularity deviation value representing the regularity of shape, and the edge blur band width value representing the clarity of the boundary, thereby achieving a quantitative characterization of deposition behavior.

[0078] For example, an input deposition feature tensor of size 7x7x64 is flattened to obtain a 3136-dimensional vector. The first fully connected layer contains 512 neurons and outputs 512-dimensional features; the second fully connected layer contains 3 neurons and directly outputs the predicted value. Assuming the network outputs are 0.85, 0.12, and 0.05, based on the calibration coefficients (radius coefficient 10μm / unit, non-circularity coefficient 1.0, blur band coefficient 5μm / unit), the actual diffusion radius is calculated to be 8.5μm, the non-circularity deviation is 0.12, and the edge blur band width is 0.25μm. This process takes less than 2ms on an embedded AI coprocessor, significantly improving the real-time performance and accuracy of feature extraction.

[0079] S2.4: Construct a structured sedimentary feature vector using the actual diffusion radius, non-circularity deviation, and edge blur band width. Then, perform confidence verification and outlier filtering on the structured sedimentary feature vector to output the final set of key sedimentary feature parameters for mismatch pattern comparison.

[0080] The system receives three independent scalar values ​​from the convolutional neural network model: the actual diffusion radius, the non-circularity deviation, and the width of the edge blur band. A three-dimensional depositional feature vector space is constructed, and these three scalar values ​​are mapped into structured vector elements according to a preset physical dimension order to form the initial depositional feature vector. A confidence verification mechanism based on Mahalanobis distance is introduced to calculate the statistical distance between the current initial depositional feature vector and the cluster centers of historical normal operating conditions. The confidence score is calculated using the following formula: in, Score the confidence level. This is the current initial sedimentation feature vector. This is the vector of historical normal operating conditions' average values. Let covariance matrix be the variance matrix. This is the maximum permissible Mahalanobis distance threshold.

[0081] The system determines whether the confidence score is lower than a preset minimum confidence threshold. If the confidence score is lower than the minimum confidence threshold, the current data is considered an outlier, triggering outlier filtering logic. A sliding window mid-range filtering algorithm is executed, using depositional feature parameters from the previous N valid periods to replace the current outlier data, generating a corrected depositional feature vector. If the confidence score is higher than or equal to the minimum confidence threshold, the current initial depositional feature vector is directly retained as valid data. The valid data undergoes dimensional normalization to eliminate the weighting interference of different physical quantities on subsequent comparison analyses, outputting the final set of key depositional feature parameters for mismatch pattern comparison. Through confidence verification and outlier filtering, the original prediction results from the previous step are transformed into a highly reliable set of key depositional feature parameters, achieving the expected technical effect of eliminating instantaneous sensor noise interference and ensuring the accuracy of path compensation.

[0082] For example, the actual diffusion radius is set to 0.15 mm, the non-circularity deviation value is 0.02, and the edge blur width is 0.03 mm. An initial vector x = [0.15, 0.02, 0.03] is constructed. The historical mean vector μ = [0.14, 0.015, 0.025] is used, and the covariance matrix Σ is a unit diagonal matrix. max Set the threshold to 3.0. The calculated Mahalanobis distance is approximately 1.12, and the confidence level C is approximately 0.627. If the preset minimum confidence threshold is 0.6, then C is greater than the threshold, and the data is valid. Z-score normalization is performed on the data, outputting a standardized set of key depositional feature parameters. If the edge blurring band width is mismeasured to be 0.1 mm due to strong light interference, the calculated C will significantly decrease and trigger median filtering. The effective value of 0.025 mm from the previous period will be used instead to ensure the stability of the output parameters.

[0083] Step S3: Based on the key sedimentary characteristic parameters and the ideal sedimentary templates pre-stored under standard reference conditions, a comparative analysis is performed to generate mismatch pattern type labels and quantitative indicators of the degree of mismatch, identifying whether the current local area exhibits lateral over-diffusion or vertical under-wetting. Specifically, this includes: S3.1: Obtain three key depositional feature parameters from the output of the convolutional neural network model: actual diffusion radius, non-circularity deviation value, and edge blur band width. Then, call the ideal depositional template data under the standard reference conditions pre-stored in the storage unit to construct a comparison dataset containing measured data and benchmark data.

[0084] It should be noted that the ideal deposition template is a set of ink droplet deposition benchmark data pre-collected and calibrated for a specific material substrate under standard reference conditions. The process of establishing this template includes: in a constant temperature and humidity laboratory environment, controlling the ambient temperature fluctuation to no more than ±1℃ and the relative humidity fluctuation to no more than ±5%, using the same printhead assembly and ink system as mass production, performing multiple standard ink droplet ejection actions on the surface of the substrate to be tested; capturing a sequence of macro images of the entire ink droplet deposition process using the same imaging system as S1; after manual screening and multi-frame averaging, extracting three benchmark parameters for the ink droplets on the material: standard diffusion radius, standard non-circularity threshold, and standard sharpness bandwidth; binding the above three benchmark parameters with the corresponding material identifier and temperature and humidity range to form a structured ideal deposition template data entry and storing it in the storage unit.

[0085] It should be noted that the standard reference conditions refer to a printing environment that meets the cleanliness standards of a conventional inkjet printing workshop. The temperature setting is 25°C, the relative humidity setting is 50%, the printing substrate undergoes surface cleaning and electrostatic elimination treatment before ink droplet deposition, the printhead assembly is at its nominal operating temperature, and the droplet flight distance is the calibrated value. The specific numerical range of the standard reference conditions can be adaptively adjusted by those skilled in the art based on the type of mobile phone case substrate and the ink system.

[0086] The system receives the structured deposition feature vector output by the convolutional neural network model and extracts three physical quantities: actual diffusion radius, non-circularity deviation, and edge blur width. Based on the substrate material identifier of the current printing task, it retrieves the corresponding ideal deposition template dataset from the index table of the storage unit.

[0087] Call the standard reference condition parameters that match the current ambient temperature and humidity range, and load the pre-stored baseline diffusion radius, standard non-circularity threshold, and ideal clarity bandwidth values.

[0088] A two-column comparison matrix containing measured dynamic data and static benchmark data is constructed to establish the mapping correspondence of each feature dimension.

[0089] Perform a dimension consistency check between the measured data and the benchmark data to ensure that both are under the same pixel-to-physical length conversion ratio.

[0090] Data alignment processing transforms the feature extraction results from the previous step into a standardized comparison dataset with comparability, thus providing the data foundation for accurate determination of mismatch patterns in the subsequent process.

[0091] For example, in a printing scenario of a TPU phone case at 25℃ and 50% humidity, the neural network outputs an actual diffusion radius of 0.12mm, a non-circularity deviation of 0.08, and an edge blur width of 0.03mm. The system retrieves data from an ideal deposition template under this condition, showing a standard diffusion radius of 0.10mm, a standard non-circularity threshold of 0.02, and an ideal clear bandwidth of 0.01mm. The system encapsulates these six values ​​into a comparison dataset, where the measured radius corresponds to the standard radius, the measured non-circularity corresponds to the standard threshold, and the measured blur bandwidth corresponds to the ideal bandwidth. After dimension verification confirms that all numerical units are uniformly set to millimeters or dimensionless ratios, a comparison dataset containing three sets of differential input variables is generated. This provides accurate data support for calculating the radial diffusion deviation, significantly improving the targeting of subsequent path compensation.

[0092] S3.2: Perform multidimensional feature difference operation processing based on the comparison dataset, subtract the actual diffusion radius from the standard diffusion radius in the ideal deposition template to obtain the radial diffusion deviation, compare the non-circularity deviation value with the standard non-circularity threshold to obtain the morphological distortion coefficient, and subtract the edge blur band width from the standard clear bandwidth to obtain the edge diffusion increment.

[0093] The actual diffusion radius values ​​in the comparison dataset are compared with the standard diffusion radius reference values ​​in the ideal deposition template. A subtraction operation is performed to calculate the radial diffusion deviation, which characterizes the degree to which the actual radial spread of the ink droplets on the substrate surface deviates from the ideal state.

[0094] The non-circularity deviation value is read and logically compared with a pre-stored standard non-circularity threshold. The difference between the actual non-circularity and the standard threshold is calculated to generate a morphological distortion coefficient, which quantifies the magnitude of geometric regularity distortion of the ink droplet deposition profile.

[0095] The width of the blurred edge band is extracted and subtracted from the standard clear bandwidth baseline. The edge diffusion increment is obtained through differential calculation, which reflects the pixel-level blur diffusion range of the ink droplet edge caused by capillary action or environmental disturbances.

[0096] The radial diffusion deviation, morphological distortion coefficient, and edge dispersion increment are encapsulated into a multidimensional feature difference vector. This vector serves as the input basis for mismatch mode classification, accurately describing the comprehensive deviation state of the current droplet deposition behavior in three dimensions: size, shape, and sharpness.

[0097] Through multidimensional feature difference operation processing, the key deposition feature parameters of the previous step are transformed into three quantitative indicators: radial diffusion deviation, morphological distortion coefficient and edge diffusion increment. This realizes the digital mapping from the original visual features to the physical mismatch state, providing high-precision data support for subsequent mismatch mode determination.

[0098] For example, the standard diffusion radius is set to 0.15 mm, the standard non-circularity threshold to 0.02, and the standard sharpness bandwidth to 0.01 mm. When the actual diffusion radius is detected to be 0.18 mm, the radial diffusion deviation is calculated to be 0.03 mm. If the actual non-circularity is 0.05, the morphological distortion coefficient is 0.03. If the actual edge blur width is 0.025 mm, the edge diffusion increment is 0.015 mm. The system outputs a feature difference vector containing the above three values, which is used to identify the lateral over-diffusion trend and significantly improve the targeting of path compensation.

[0099] S3.3: Using radial diffusion deviation, morphological distortion coefficient and edge diffusion increment as input variables, execute the preset mismatch mode classification logic judgment. If the radial diffusion deviation is greater than zero and the edge diffusion increment exceeds the first threshold, it is judged as a lateral over-diffusion state. If the radial diffusion deviation is less than zero and the morphological distortion coefficient shows longitudinal stretching, it is judged as a longitudinal under-wetting state, thereby generating a mismatch mode type label that represents the physical state of the current local area.

[0100] It should be noted that lateral over-diffusion refers to the phenomenon where the spread of ink droplets on the substrate surface along the direction perpendicular to the printing travel significantly exceeds the ideal state. Physically, this manifests as the ink droplet edges extending outward beyond the expected pattern outline, causing color bleeding or blurred boundaries between adjacent pixels or pattern areas. This phenomenon is usually caused by the ink surface tension being lower than the critical surface energy of the substrate, or the ambient humidity being higher than the reference value. Longitudinal under-wetting refers to the phenomenon where the spread length of ink droplets on the substrate surface along the printing travel direction is significantly shorter than the ideal state. Physically, this manifests as ink droplets not spreading sufficiently on the substrate surface, forming an elliptical or understretched shape, resulting in gaps between adjacent ink droplets or blank lines within the pattern. This phenomenon is usually caused by excessively high substrate surface roughness or the ambient temperature being lower than the reference value.

[0101] The radial diffusion deviation, morphological distortion coefficient, and edge diffusion increment are received as input variables for mismatch mode classification. A logical decision matrix based on multidimensional threshold discrimination is constructed, with the first threshold set as the critical value for edge diffusion due to lateral over-diffusion and the second threshold set as the critical value for morphological distortion due to longitudinal under-wetting.

[0102] A sign determination operation is performed on the radial diffusion deviation. If the value is greater than zero, it is marked as an expansion trend; if it is less than zero, it is marked as a contraction trend. Under the expansion trend label, the relationship between the edge diffusion increment and a first threshold is further compared. When the edge diffusion increment exceeds the first threshold, the current droplet deposition state is determined to be lateral over-diffusion, and a corresponding mismatch mode type label is generated. Under the contraction trend label, the anisotropic component of the morphological distortion coefficient is analyzed. If the morphological distortion coefficient shows stretching characteristics along the printing direction and exceeds a second threshold, the current droplet deposition state is determined to be longitudinal under-wetting, and a corresponding mismatch mode type label is generated.

[0103] Through the above multi-level logical judgment, continuous physical deviation values ​​are transformed into discrete state identifiers, realizing accurate classification of the physical state of local areas and providing a clear compensation direction basis for the subsequent path deformation engine.

[0104] For example, a first threshold of 5 micrometers and a second threshold of 0.15 are set. When the radial diffusion deviation is 8 micrometers and the edge diffusion increment is 6 micrometers, it is determined to be lateral over-diffusion because 8 > 0 and 6 > 5. When the radial diffusion deviation is -3 micrometers and the longitudinal component of the morphological distortion coefficient is 0.2, it is determined to be longitudinal under-wetting because -3 < 0 and 0.2 > 0.15. This classification logic significantly improves the accuracy of state recognition and ensures the pertinence of the compensation strategy.

[0105] S3.4: Based on the feature dimension corresponding to the mismatch mode type label, select the dominant deviation value among radial diffusion deviation, morphological distortion coefficient or edge dispersion increment, and convert the dominant deviation value into a dimensionless mismatch degree quantification index through a normalization mapping algorithm to quantitatively describe the specific extent to which the current ink droplet deposition behavior deviates from the ideal state.

[0106] It receives the mismatch mode type label determined by the mismatch mode classification logic, as well as the original deviation values ​​of the three dimensions of radial diffusion deviation, morphological distortion coefficient and edge diffusion increment calculated by the previous steps.

[0107] Based on the mismatch mode type label, the dominant feature dimension is determined. If the label indicates a horizontal over-diffusion state, the radial diffusion deviation is selected as the dominant deviation value; if the label indicates a vertical under-wetting state, the morphological distortion coefficient is selected as the dominant deviation value.

[0108] The pre-stored upper limits of the physical range of each feature dimension are called, and a linear normalization mapping operation is performed on the selected dominant deviation value to eliminate the numerical scale difference between different physical dimensions.

[0109] The dimensionless quantitative index of mismatch is calculated using the following formula: in, As a quantitative indicator of the degree of mismatch, The dominant deviation value, This is the maximum permissible deviation threshold for this feature dimension under standard reference conditions.

[0110] The calculated mismatch quantification index is subjected to boundary truncation, and its value range is strictly limited to a closed interval of 0 to 1 to prevent numerical overflow caused by extreme environmental disturbances from affecting the stability of subsequent control.

[0111] By using a normalized mapping algorithm, the dominant deviation values ​​with physical units are converted into dimensionless quantitative indicators of mismatch, thus achieving a standardized quantitative description of the deviation of the current ink droplet deposition behavior from the ideal state, and providing a unified driving basis for the path deformation engine.

[0112] For example, when the mismatch mode type label indicates lateral over-diffusion, a radial diffusion deviation of 0.05mm is selected as the dominant deviation value. A query of the pre-stored range reveals that the maximum permissible radial diffusion deviation threshold for this material is 0.2mm. Substituting this into the formula, the mismatch degree quantification index is calculated to be 0.25. This index is directly input into the path deformation engine, indicating that the printhead path needs to undergo moderate inward contraction compensation, effectively suppressing pattern blurring caused by lateral ink diffusion and significantly improving the sharpness of printed edges.

[0113] S3.5: The generated mismatch mode type label and the calculated mismatch degree quantification index are structurally encapsulated to form a complete mismatch analysis result data packet containing status identifiers and quantification values. The mismatch analysis result data packet is then output to the path deformation engine to trigger the subsequent normal offset disturbance calculation process.

[0114] like Figure 3 As shown, step S4: Using the mismatch mode type label and mismatch degree quantification index as the driving basis, a path deformation engine is constructed that does not change the geometric topology of the original vector path. The path deformation engine is used to calculate the normal offset perturbation amount with controllable direction and adaptive amplitude applied to the original path segment. Specifically, it includes: S4.1: Start the path deformation engine, obtain the mismatch mode type label and mismatch degree quantification index generated in the previous steps as input parameters, call the preset perturbation strategy mapping table for index matching, extract the reference compensation coefficient vector and direction determination flag corresponding to the lateral over-diffusion or longitudinal under-wetting state, and generate a primary perturbation instruction set containing the initial compensation amplitude and correction direction.

[0115] The path deformation engine is a path correction algorithm module running on an embedded controller. Its core function is to calculate the normal offset perturbation applied to the original path segments based on the mismatch analysis results, without changing the original geometric topology of the vector path. The engine's inputs are mismatch mode type labels and mismatch degree quantification indicators, and its output is a sequence of normal offset perturbations with controllable direction and adaptive amplitude.

[0116] The path deformation engine receives the mismatch mode type label and mismatch degree quantification index from the preceding steps as the core input parameters driving path correction. It calls a perturbation strategy mapping table pre-stored in the embedded storage unit, which establishes a static association logic between mismatch states and compensation strategies. Based on the mismatch mode type label, it performs an index matching operation to retrieve the baseline compensation coefficient vector corresponding to the lateral over-diffusion or longitudinal under-wetting states.

[0117] A direction determination flag is extracted that matches the current mismatch quantification index. This flag defines the positive or negative polarity of the normal offset to indicate the physical direction of suppressing diffusion or promoting infiltration. Combining the baseline compensation coefficient vector and the direction determination flag, a primary perturbation instruction set is constructed, which includes the initial compensation amplitude value and the correction direction.

[0118] By using table lookup matching and parameter extraction, the abstract mismatch status identifier is transformed into concrete vector control command data, realizing a logical closed-loop connection from perception diagnosis to execution control.

[0119] For example, when the mismatch mode type label is "lateral over-diffusion" and the mismatch degree quantification index is 0.85, the path deformation engine retrieves the corresponding baseline compensation coefficient vector [0.12, 0.05] from the perturbation strategy mapping table, with the direction determination flag set to "-1". The system extracts an initial compensation amplitude of 0.12 pixels and corrects the direction to point inward to counteract ink overflow. The generated primary perturbation instruction set contains an amplitude value of 0.12 and a direction code of -1, which is directly used for subsequent normal vector calculation, ensuring that the printhead trajectory is finely adjusted inward to compress the ink droplet spreading space and significantly improve the sharpness of the pattern edges.

[0120] S4.2: Using the direction determination flag in the primary disturbance instruction set, perform normal vector calculation on the discrete control point sequence on the original vector path segment, calculate the unit normal vector coordinates at each control point, and generate a normal direction reference dataset that represents the local geometric orientation of the path.

[0121] Receive the direction determination flag bit in the primary disturbance instruction set, parse and obtain the normal offset polarity identifier corresponding to the current mismatch mode, and determine whether the compensation vector points to the left or right relative to the original path tangent direction.

[0122] Extract the discrete control point sequence from the original vector path's geometric topology, traverse each intermediate control point in the sequence except for the first and last endpoints, and locate its precise position coordinates in the two-dimensional printing coordinate system.

[0123] For the i-th control point to be processed, retrieve the coordinate data of its predecessor and successor nodes, and construct two adjacent path segment vectors from the predecessor to the current point and from the current point to the successor point.

[0124] Calculate the unit tangent vectors of the two adjacent line segment vectors, and then calculate the local smooth tangent direction vector at the control point by performing a weighted average operation or angle bisector algorithm on the two unit tangent vectors.

[0125] Based on the local smooth tangent direction vector, an orthogonal transformation operation of 90 degrees counterclockwise is performed to generate an initial normal vector. The sign of the initial normal vector is checked and flipped according to the direction determination flag to ensure that the normal direction meets the physical requirements of suppressing diffusion or promoting wetting.

[0126] The normal vector after verification is normalized to eliminate the influence of vector magnitude difference on subsequent magnitude mapping, and the standardized unit normal vector coordinates are obtained.

[0127] The generated unit normal vector coordinates are bound to the control point index and stored in the normal direction reference dataset to complete the geometric orientation representation of all key nodes of the entire path segment.

[0128] Through the above chain derivation process, the primary perturbation command from the previous step is transformed into a normal direction reference dataset that characterizes the local geometric orientation of the path, thus achieving the expected technical effect of providing a precise spatial direction reference for subsequent amplitude adaptive correction.

[0129] For example, in a lateral over-diffusion state, the direction determination flag indicates an outward offset. Take control point P. i (10.5, 20.3), precursor point P i-1 (10.0, 20.0), successor point P i+1 (11.0, 20.6). Calculate the tangent vector T. avg After normalization, 90-degree counterclockwise rotation, and sign verification, the unit normal vector N is obtained. i (0.8, -0.6). This vector points precisely in the opposite direction of ink diffusion and is used for subsequent superposition of compensation amplitude.

[0130] S4.3: Based on the initial compensation amplitude in the normal direction reference dataset and the primary perturbation instruction set, and combined with the current printing speed parameters and ink viscosity constant, a nonlinear gain adjustment algorithm is executed to perform amplitude adaptive correction calculation, so as to generate a dynamic compensation amplitude sequence that eliminates the influence of speed lag and adapts to the rheological properties of the material.

[0131] Read the unit normal vector coordinates from the normal direction reference dataset and the initial compensation magnitude from the primary perturbation instruction set as input variables for the nonlinear gain adjustment algorithm.

[0132] The current printhead real-time printing speed parameter is obtained from the motion control unit. This parameter represents the dynamic coupling state between the ink droplet ejection frequency and the relative displacement rate of the substrate.

[0133] The viscosity characteristic constant in the ink property database is used to retrieve the ink’s rheological response at different shear rates and its inhibition effect on diffusion behavior.

[0134] A nonlinear gain function based on velocity hysteresis compensation and rheological adaptation is constructed to calculate dynamic correction coefficients to eliminate phase delay caused by mechanical inertia.

[0135] The dynamic compensation amplitude is calculated using the following formula: in, To dynamically compensate for the amplitude, This is the initial compensation range. For speed-dependent gain factors, For real-time printing speed, This is the viscosity correction factor. This represents the viscosity characteristic constant of the ink.

[0136] Based on the real-time printing speed v, a preset speed-gain mapping table is queried, and the corresponding speed-dependent gain factor k(v) is obtained through linear interpolation to compensate for the lag in ink droplet flight time under high-speed motion.

[0137] The viscosity correction coefficient f(η) is calculated based on the ink viscosity characteristic constant η. This coefficient adopts an inverse proportional decay model to suppress the risk of excessive penetration of high viscosity ink at low printing speed.

[0138] Substitute the initial compensation amplitude A0, the velocity-dependent gain factor k(v), and the viscosity correction coefficient f(η) into the above formula, and perform a multiplication operation to generate the dynamic compensation amplitude for the current control point.

[0139] Traverse all discrete control point sequences on the original vector path and repeat the above gain adjustment operation to generate a dynamic compensation amplitude sequence that strictly matches the path geometry.

[0140] By using a nonlinear gain adjustment processing method, the static initial compensation amplitude of the previous step is transformed into a dynamic compensation amplitude sequence that adapts to the real-time printing speed and ink rheological properties, thereby achieving the expected technical effect of eliminating the influence of speed lag and adapting to the material rheological properties.

[0141] For example, setting the initial compensation amplitude A0 to 5 micrometers, the real-time printing speed v to 600 mm / s, the speed-dependent gain factor k(v) is found to be 1.2 from the table, and the ink viscosity characteristic constant η is 15 mPa·s. The viscosity correction coefficient f(η) is calculated to be 0.9. Substituting these values ​​into the formula, the dynamic compensation amplitude A = 5 × 1.2 × 0.9 = 5.4 micrometers is obtained. This amplitude effectively compensates for the ink droplet lag under high-speed printing, while avoiding excessive diffusion of high-viscosity ink in low-speed regions, significantly improving the sharpness and consistency of the pattern edges.

[0142] S4.4: Based on the dynamic compensation amplitude sequence and normal direction reference dataset, perform vector synthesis operation to project the scalar amplitude to the corresponding unit normal vector direction, and generate a preliminary normal offset vector set with spatial coordinate attributes to realize the conversion of compensation amount from one-dimensional numerical value to two-dimensional spatial displacement.

[0143] Obtain the dynamic compensation amplitude sequence and normal direction reference dataset output by the path deformation engine. The dynamic compensation amplitude sequence contains scalar displacement values ​​after nonlinear gain adjustment, and the normal direction reference dataset contains the unit normal vector coordinate components at each control point of the original vector path.

[0144] Each scalar compensation value in the dynamic compensation amplitude sequence is spatially expanded and mapped to a displacement modulus in a two-dimensional plane, ensuring that the scalar value has a clear physical length dimension to meet the needs of subsequent vector operations.

[0145] Read the unit normal vector coordinates of the corresponding control points in the normal direction reference dataset, extract their horizontal and vertical components, construct a basis vector matrix representing the local geometric vertical direction, and ensure that the displacement direction is strictly perpendicular to the tangent direction of the printing path.

[0146] Perform vector composition operation, multiply the expanded displacement modulus with the transverse and longitudinal components of the unit normal vector respectively, and calculate the projected displacement components of each control point in the X and Y axes.

[0147] The calculated X-axis and Y-axis projected displacement components are combined and encapsulated to generate a preliminary normal offset vector with clear spatial coordinate attributes. This vector fully describes the displacement direction and magnitude of the single-point compensation in the two-dimensional printing plane.

[0148] Traverse all control points to be processed, repeat the above vector synthesis and component calculation process, and arrange all the initial normal offset vectors of single points in the original path topology order to form a set of initial normal offset vectors.

[0149] By using vector synthesis operations, the one-dimensional scalar compensation amplitude from the previous step is transformed into two-dimensional displacement vector data with spatial orientation, realizing the accurate mapping of the compensation quantity from the numerical domain to the geometric spatial domain, and providing structured input for subsequent path smoothing filtering.

[0150] S4.5: Apply topological constraint filtering to the initial normal offset vector set, limit the offset difference threshold between adjacent control points to ensure path smoothness, and output the final normal offset disturbance with controllable direction and adaptive amplitude, thus completing the non-destructive correction data preparation of the original vector path by the path deformation engine.

[0151] The system receives a preliminary set of normal offset vectors generated by the path deformation engine. This set contains the initial compensation displacement data for each discrete control point on the original vector path. Differential operations are performed on the normal offsets between adjacent control points to calculate the gradient rate of change of the offset amplitude between consecutive points, quantifying the degree of curvature abrupt changes in the path. A maximum allowable offset difference threshold is set, determined based on the balance between the printhead mechanical response bandwidth and the ink surface tension, to define the physical boundary of path smoothness. When the absolute value of the offset difference between any two adjacent points exceeds the preset threshold, a topology constraint filtering mechanism is triggered to identify high-frequency noise components that may cause printing jitter or ink droplet accumulation. A sliding window mean filtering algorithm is used to locally smooth the offset vector sequence in the exceeding region, suppressing high-frequency oscillation interference while preserving the overall compensation trend. Quadratic spline interpolation reconstruction is performed on the filtered offset vector sequence to ensure that the change in normal offset between adjacent control points satisfies the first derivative continuity condition, eliminating sharp corners or discontinuities in the path's geometric topology. The reconstructed smooth offset vector sequence is mapped back to the original coordinate system to generate the final normal offset perturbation with controllable direction and adaptive amplitude.

[0152] By using topological constraint filtering and spline reconstruction, the initial offset vector from the previous step is transformed into the final normal offset disturbance data that meets the requirements of smooth mechanical motion, thereby achieving the expected technical effects of ensuring geometric continuity and improving printing stability during the path correction process.

[0153] For example, in a printing scenario for TPU phone cases at 25°C, the maximum allowable offset difference threshold is set to 0.8 micrometers. When adjacent control points P in the initial normal offset vector set... i With P i+1When the offsets are 3.2 micrometers and 4.5 micrometers respectively, the calculated difference is 1.3 micrometers, exceeding the preset threshold. The system triggers sliding window filtering, with the window size set to 5 points, for P i-2 To P i+2 The offsets within the range are weighted and averaged, with the weighting coefficients following a Gaussian distribution. After filtering, P... i The correction offset was adjusted to 3.6 micrometers, P i+1 The value was adjusted to 3.9 micrometers, and the differential value was reduced to 0.3 micrometers to meet the smoothness requirements. Subsequently, quadratic spline interpolation was performed to generate a smooth offset curve for intermediate transition points, ensuring the continuity of the first derivative of the path. The final output sequence of normal offset perturbations drives the nozzle movement, eliminating mechanical vibrations caused by abrupt path changes, significantly reducing jagged defects at the pattern edges, and greatly improving the flatness of the printed surface.

[0154] Step S5: Perform sub-pixel-level dynamic deformation mapping processing on the subsequent ejection point coordinates based on the normal offset disturbance to generate a corrected printhead path trajectory that compensates for the effects of environmental disturbances, so that the ink droplet automatically fine-tunes in the direction that inhibits diffusion or promotes wetting. Specifically, this includes: S5.1: Obtain the sequence of line segments to be processed and the corresponding set of initial injection point coordinates in the original vector path geometric topology, and read the normal offset disturbance data with controllable direction and adaptive amplitude output by the path deformation engine to construct a path correction input dataset containing spatial location information and dynamic compensation parameters.

[0155] It should be noted that subpixel level refers to coordinate offset increments that are less than the physical size of a single pixel, meaning that point adjustment is performed with an accuracy lower than the minimum single-step displacement of the printhead motion control system. At a printing resolution of 600 dpi, the accuracy of subpixel level dynamic deformation mapping is a fraction of the size of a single pixel, and the minimum resolvable displacement is better than 1 micrometer.

[0156] The system receives directionally controllable and amplitude-adaptive normal offset disturbance data from the path deformation engine. This data contains compensation vector information for each discrete control point. The original vector path geometry is analyzed to extract the sequence of line segments to be processed and the corresponding initial injection point coordinate set, and a spatial position index is established.

[0157] The normal offset disturbance data is time-stamped and spatially mapped to the initial injection point coordinate set to ensure a one-to-one correspondence between compensation parameters and physical points. A path correction input dataset containing spatial location information and dynamic compensation parameters is constructed, which integrates the original geometric coordinates, local normal reference, and dynamic compensation amplitude.

[0158] By using a data structuring and encapsulation process, the independent perturbation quantities generated in the previous step are transformed into a path correction input dataset with spatiotemporal correlation, thus preparing the data foundation for subsequent subpixel-level dynamic deformation mapping.

[0159] For example, in a TPU phone case printing scenario, a 10mm long Bézier curve path segment is obtained, containing the coordinates of 20 initial injection points. The normal offset perturbation output by the path deformation engine is read, with a maximum compensation amplitude of 5 micrometers and a direction labeled "inward" (contracting inward to suppress diffusion). The 20 coordinate points (x... i ,y i ) and the corresponding offset (delta) d_i ,direction i Pair by index. Construct a path correction input dataset containing 40 data elements, each containing the original coordinates, normal unit vector, and compensation scalar. This dataset is directly used for normal vector calculation in S5.2, ensuring that the compensation command is accurately applied to specific physical points and eliminating the risk of trajectory distortion caused by asynchronous data.

[0160] S5.2: Based on the direction identifier of the normal offset perturbation in the path correction input dataset, perform unit normal vector calculation on each line segment to be processed in the original vector path geometric topology to generate a local coordinate system normal reference vector sequence that strictly corresponds to the current mismatch mode type label.

[0161] Read the direction identifier of the normal offset perturbation in the path correction input dataset, parse the lateral over-diffusion or longitudinal under-wetting state bits contained therein, and determine the compensation polarity required for the current line segment to be processed.

[0162] Traverse each line segment to be processed in the geometric topology of the original vector path, extract the coordinate sequence of the endpoints of the line segments, and construct a local Cartesian coordinate system to define the spatial orientation reference of the line segments.

[0163] For each discrete control point, select the control points before and after it to form a local tangent vector, and use vector subtraction to calculate the tangent vector representing the instantaneous motion direction of the path.

[0164] An orthogonal transformation operation is performed based on the tangent vector to generate a candidate set of initial normal vectors pointing to the outside or inside of the path. The initial normal vectors are then oriented according to the mismatch mode type label. If it is determined to be lateral over-diffusion, the normal vector is kept pointing to the outside of the pattern contour. If it is determined to be longitudinal under-wetting, the normal vector is reversed to point to the inside of the pattern contour, ensuring that the compensation direction strictly corresponds to the physical requirements of suppressing diffusion or promoting wetting.

[0165] The normal vector after verification is normalized to eliminate the difference in vector length caused by changes in path curvature, and a local coordinate system normal reference vector of unit length is generated.

[0166] Arrange the unit normal reference vectors of all control points in chronological order to form a normal reference vector sequence that maps one-to-one with the original path topology.

[0167] Through the above-mentioned orthogonal solution and direction verification processing method, the path geometry data of the previous step is transformed into a local coordinate system normal reference vector sequence that strictly corresponds to the mismatch mode, so as to achieve the expected technical effect of establishing the direction reference for subsequent sub-pixel level displacement mapping.

[0168] For example, considering a straight path segment consisting of three control points P1(10.0,20.0), P2(10.5,20.5), and P3(11.0,21.0), the system detects lateral over-diffusion. First, the tangent vector T at point P2 is calculated as T = P3 - P1 = (1.0, 1.0). Then, the initial normal vector N = (-1.0, 1.0). Since it is lateral over-diffusion, compensation needs to be applied outwards while maintaining the normal vector direction. The magnitude |N| is calculated as sqrt((-1.0)). 2 +1.0 2 =1.414. After modulus normalization, the unit normal vector is obtained as (-0.707, 0.707). This unit vector will serve as the directional reference for subsequent calculations of the subpixel displacement of point P2, ensuring that the ink droplet landing point is finely adjusted along this direction to counteract ink diffusion.

[0169] S5.3: Using the local coordinate system normal reference vector sequence and normal offset disturbance amplitude data, vector synthesis operation is performed to perform sub-pixel level displacement vector mapping processing on the initial injection point coordinate set to generate a discrete coordinate offset increment matrix characterizing the single-point compensation effect.

[0170] Read the unit normal vector coordinate components corresponding to each control point in the local coordinate system normal reference vector sequence, and extract the scalar displacement values ​​from the normal offset disturbance magnitude data.

[0171] A vector synthesis operation model in a two-dimensional plane is constructed, using the scalar magnitude as the modulus and the unit normal vector as the direction basis, and point-by-point multiplication is performed to generate a displacement vector with spatial orientation.

[0172] The coordinate offset increment of the i-th control point is calculated using the following formula: in, For the first The coordinate offset increment vector of each control point For the first Normal offset disturbance amplitude at each control point For the first The unit normal vector of each control point.

[0173] The calculated coordinate offset increment vector is quantized with subpixel precision, and the continuous floating-point numbers are mapped to the minimum resolution grid of the printer motion control system.

[0174] After quantizing all control points, the offset increments are arranged according to the point order of the original path and assembled into a discrete coordinate offset increment matrix.

[0175] By using vector synthesis and sub-pixel quantization, the normal reference and amplitude data from the previous step are transformed into a discrete coordinate offset increment matrix, achieving the expected technical effect of high-precision spatial fine-tuning of the nozzle position.

[0176] For example, in a TPU phone case printing scenario, the printing resolution is set to 600 dpi, corresponding to a minimum physical step size of 42.3 micrometers. When a lateral over-diffusion mismatch mode is detected, the path deformation engine outputs a normal offset disturbance amplitude of 15.6 micrometers at a certain control point, with the unit normal vector at that point being (0.707, 0.707). Substituting this into the formula, the coordinate offset increment vector is calculated to be (11.03, 11.03) micrometers. The system maps this floating-point value to the minimum resolution grid inside the controller, quantizing it into a (0.26, 0.26) pixel step size. In the final generated discrete coordinate offset increment matrix, this point is recorded as a positive X-axis offset of 0.26 steps and a positive Y-axis offset of 0.26 steps. This process ensures that while suppressing ink diffusion, the sub-pixel positioning accuracy of the printhead movement is maintained, significantly improving the sharpness of the pattern edges.

[0177] S5.4: The initial set of spray point coordinates is corrected point by point by superimposing the discrete coordinate offset increment matrix, and the corrected discrete points are smoothly connected by spline interpolation algorithm to generate a continuous and smooth corrected nozzle path trajectory to compensate for the influence of environmental disturbances.

[0178] Read the displacement vector data of each control point in the discrete coordinate offset increment matrix, and align it with the initial set of injection point coordinates to construct a point mapping pair to be corrected.

[0179] For each mapping pair, the two-dimensional plane coordinates of the initial injection point and the corresponding normal offset disturbance components are extracted, and vector addition is performed to calculate the corrected target coordinate position.

[0180] A cubic B-spline interpolation algorithm is used to smoothly connect the corrected discrete point sequence to ensure that the second derivative of the path is continuous at the connection point.

[0181] Curvature constraint checks are performed on the generated continuous trajectory to eliminate abnormal inflection points with curvature radii smaller than the nozzle mechanical response limit, and motion smoothness is ensured through local linearization correction.

[0182] Through the above vector superposition and spline smoothing process, the discrete offset of the previous step is transformed into continuous and smooth corrected nozzle path trajectory data, realizing dynamic deformation mapping with sub-pixel accuracy, and ensuring precise fine-tuning of the ink drop point in the direction of suppressing diffusion.

[0183] For example, in a PC material phone case printing scenario, the initial inkjet point coordinates are (10.00mm, 20.00mm), and the normal offset perturbation is (-0.05mm, 0.02mm). Vector addition is performed to obtain the corrected point (9.95mm, 20.02mm). Cubic B-spline interpolation is applied to 10 consecutive corrected points, including this point, with the node vectors set to uniform distribution to generate a smooth trajectory. Curvature checking shows a maximum curvature value of 0.15mm. -1 Less than the nozzle's allowable threshold of 0.2mm -1 No additional corrections are required. The final output correction trajectory is sent to the motion control unit, reducing the width of the blurred edge band from 0.08mm to 0.03mm, significantly improving the clarity of the pattern.

[0184] S5.5: The generated corrected printhead path trajectory, which compensates for the effects of environmental disturbances, is sent to the motion control unit to drive the printhead assembly to perform subsequent ink droplet ejection actions according to the corrected trajectory. This allows the ink landing point to be automatically fine-tuned in the direction of suppressing diffusion or promoting wetting, ultimately achieving dynamic closed-loop compensation for the accuracy of the pattern edge.

[0185] The system receives continuous and smooth corrected nozzle path trajectory data generated by S5.4 to compensate for environmental disturbances. This data includes a discrete control point coordinate sequence processed by spline interpolation and the corresponding timing trigger commands. Inverse kinematics calculation is performed on the corrected nozzle path trajectory data to convert the spatial points in the Cartesian coordinate system into pulse equivalent sequences for the X / Y axis servo motors of the printing platform and the Z-axis piezoelectric actuators of the nozzle.

[0186] A look-ahead buffer algorithm is used to perform velocity planning on the pulse sequence to ensure that the rate of change of acceleration of each axis remains within a preset threshold during sub-pixel-level path fine-tuning, thus avoiding mechanical vibration caused by sudden changes in path curvature.

[0187] The planned motion control commands are sent to the underlying motion control unit in real time via a high-speed fieldbus, establishing a hardware-level synchronization mechanism between the nozzle spray trigger signal and the platform position signal. The motion control unit drives the actuator to move according to the corrected trajectory, causing the physical landing point of the nozzle to deviate normally from the original design coordinates. This deviation direction strictly corresponds to the mismatch mode identified by S3.

[0188] When lateral over-diffusion is detected, the printhead tip contracts inward to counteract ink smudging; when longitudinal under-wetting is detected, the printhead tip expands outward to fill the edge gaps. Through the above closed-loop control process, the correction trajectory generated in the previous step is transformed into actual physical deposition position deviation compensation, achieving self-correction and high-fidelity restoration of pattern edge accuracy under dynamic environmental disturbances.

[0189] For example, in a printing scenario involving TPU phone cases at 28℃ and 65% humidity, the corrected trajectory output by S5.4 includes a set of control points distributed along a curve. The original coordinates of a key point are (100.000mm, 50.000mm), and the calculated normal offset is (-0.015mm, +0.008mm). The motion control unit converts these coordinates into stepper motor pulses, reducing the X-axis by 15 micro-steps and increasing the Y-axis by 8 micro-steps. The system operates at 600dpi resolution, with the printhead movement speed remaining constant at 0.5m / s. Actual measurements show that after correction, the distance error between the ink droplet center and the ideal edge is reduced from 45μm to less than 8μm, and the width of the blurred edge band is reduced from 120μm to 35μm. This significantly eliminates the lateral ink diffusion caused by ambient humidity, ensuring the clarity of the customized pattern's outline.

[0190] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0191] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0192] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent control of the production process of a customized pattern of a mobile phone case, characterized in that, include: Acquire a sequence of macro images of the substrate surface within a preset time window after the printhead assembly performs a single ink droplet ejection action; The macro image sequence is input into a convolutional neural network model pre-trained with multiple materials and multiple environmental conditions to extract and output three key deposition feature parameters that characterize the current ink droplet deposition behavior: actual diffusion radius, non-circularity deviation value, and edge blur band width. Based on the comparison and analysis between the key sedimentary characteristic parameters and the ideal sedimentary templates stored under standard reference conditions, a mismatch pattern type label and a quantitative index of the degree of mismatch are generated to identify whether there is lateral over-diffusion or longitudinal under-wetting in the current local area. Using the mismatch mode type label and mismatch degree quantification index as the driving basis, a path deformation engine is constructed that does not change the geometric topology of the original vector path. The path deformation engine is used to calculate the normal offset disturbance amount that is directionally controllable and amplitude adaptively applied to the original path segment. Based on the normal offset disturbance, subpixel-level dynamic deformation mapping processing is performed on the coordinates of subsequent injection points to generate a corrected nozzle path trajectory that compensates for the effects of environmental disturbances.

2. The intelligent control method for the production process of customized patterns on mobile phone cases according to claim 1, characterized in that, The acquisition of a sequence of macro images of the substrate surface within a preset time window after the printhead assembly performs a single ink droplet ejection action includes: Based on the trigger signal of the single ink droplet ejection action, the synchronous pulse illumination module is activated to perform nanosecond-level exposure control. The original macro image frame sequence of the substrate surface is continuously captured by the miniature high-speed camera integrated in the printhead assembly. After time window filtering and multi-frame registration and noise reduction, the macro image sequence is generated.

3. The intelligent control method for the production process of customized patterns on mobile phone cases according to claim 1, characterized in that, The step of inputting the macro image sequence into a convolutional neural network model pre-trained for multiple materials and environmental conditions to extract key depositional feature parameters includes: The macro image sequence is preprocessed with pixel-level normalization and noise suppression to eliminate ambient light fluctuations and sensor thermal noise interference, generating a standardized grayscale image matrix as clean input data for subsequent feature extraction. Multi-scale feature mapping operations are performed on a convolutional neural network model pre-trained with multiple materials and multiple environmental conditions based on a standardized grayscale image matrix. High-frequency texture information and low-frequency contour information of ink droplet wetting edges are extracted using deep convolutional kernels to generate a deposition feature tensor containing spatial location encoding. Regression prediction calculations of fully connected layers are performed based on the deposition feature tensor, mapping high-dimensional abstract features to the physical space metric domain, and respectively solving the actual diffusion radius value representing the ink droplet spreading range, the non-circularity deviation value representing the shape regularity, and the edge blur band width value representing the boundary clarity. A structured depositional feature vector is constructed using the actual diffusion radius, non-circularity deviation, and edge blur band width. The confidence level of this structured depositional feature vector is then verified and outlier filtering is performed. The three key depositional feature parameters are then output for mismatch pattern comparison.

4. The intelligent control method for the production process of customized patterns on mobile phone cases according to claim 1, characterized in that, The method of constructing a path deformation engine that does not alter the original vector path's geometric topology by using the mismatch mode type label and mismatch degree quantification index as the driving basis includes: Obtain the mismatch mode type label and the mismatch degree quantification index, and call the perturbation strategy mapping table to generate a primary perturbation instruction set containing the initial compensation amplitude and direction determination flag. Based on the direction determination flag, normal vector calculation is performed on the discrete control point sequence on the original vector path segment to generate a normal direction reference dataset. Based on the normal direction reference dataset and the initial compensation amplitude, nonlinear gain adjustment is performed to generate a dynamic compensation amplitude sequence; Based on the dynamic compensation amplitude sequence and the normal direction reference dataset, vector synthesis is performed to generate a preliminary normal offset vector set; A topology constraint filtering process is applied to the preliminary normal offset vector set to output the normal offset perturbation amount with controllable direction and adaptive amplitude.

5. The intelligent control method for the production process of customized patterns on mobile phone cases according to claim 1, characterized in that, The macro image sequence contains data on the initial wetting morphology, edge climbing dynamics, and final static contour evolution of the ink droplet after it contacts the substrate.

6. The intelligent control method for the production process of customized patterns on mobile phone cases according to claim 1, characterized in that, Based on the trigger signal of a single ink droplet ejection action performed by the printhead assembly, the synchronous pulse illumination module is activated to perform nanosecond-level exposure control in order to eliminate dynamic blur caused by printhead movement.

7. The intelligent control method for the production process of customized patterns on mobile phone cases according to claim 1, characterized in that, The comparative analysis includes: performing a differential operation between the actual diffusion radius and the standard diffusion radius in the ideal deposition template to obtain the radial diffusion deviation; if the radial diffusion deviation is greater than zero and the width of the edge blurring band exceeds a first threshold, it is determined to be a transverse over-diffusion state; if the radial diffusion deviation is less than zero and the non-circularity deviation value shows longitudinal stretching, it is determined to be a longitudinal under-wetting state.

8. The intelligent control method for the production process of customized patterns on mobile phone cases according to claim 1, characterized in that, Based on the feature dimension corresponding to the mismatch mode type label, the dominant deviation value is converted into a dimensionless quantitative index of the mismatch degree through a normalization mapping algorithm.

9. The intelligent control method for the manufacturing process of customized patterns on mobile phone cases according to claim 1, characterized in that, The generation of the corrected nozzle path trajectory includes: determining the normal offset direction based on the mismatch mode type label, determining the normal offset amplitude based on the mismatch degree quantification index, applying the normal offset disturbance to the coordinates of the subsequent injection points, and performing spline interpolation to smoothly connect the corrected discrete points.

10. The intelligent control method for the manufacturing process of customized patterns on mobile phone cases according to claim 1, characterized in that, Before inputting the macro image sequence into the convolutional neural network model, the macro image sequence is subjected to pixel-level normalization, color space conversion, histogram equalization, and nonlocal mean denoising.

Images