Dynamic compensation method for 3D printing parameters of photosensitive resin based on real-time viscosity monitoring

Abstract

This invention provides a dynamic compensation method for photosensitive resin 3D printing parameters based on real-time viscosity monitoring. It collects time-series viscosity data using a temperature-compensated microfluidic viscosity sensor and optimizes parameters in real-time based on this data through differential analysis, dynamic feature modeling, and a graphical compensation strategy library. The method identifies the nonlinear variation trend of resin viscosity, matches the optimal process compensation path, and combines printing quality feedback to achieve online fine-tuning and knowledge accumulation of the strategy. Finally, it generates and executes multi-dimensional printing control commands, achieving closed-loop adaptive adjustment of printing parameters and significantly improving the process stability and molding consistency of high-viscosity resin printing.

Description

Technical Field

[0001] This invention relates to the field of additive manufacturing control and intelligent parameter compensation technology, and in particular to a method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring. Background Technology

[0002] In the current field of photosensitive resin 3D printing, for the printing process of high viscosity gradient resin materials, mainstream technologies typically employ fixed-parameter models to compensate and adjust printing parameters. These systems usually use preset empirical formulas or simple viscosity-control parameter mappings, based on traditional linear fitting models or simple interval segmentation methods, to coarsely adjust parameters such as exposure time, platform lift-up speed, and interlayer waiting time during the printing process. Viscosity change monitoring mostly relies on offline measurement or periodic sampling, lacking a sufficient, real-time material dynamic state feedback mechanism. These solutions can maintain basic printing consistency and quality when dealing with resin materials with simple structures and relatively stable viscosity changes; With the increasing variety of photosensitive resins and the growing demand for additive manufacturing of functionally graded materials, the industry is placing higher demands on the continuous printing capabilities of high viscosity graded resins. Mainstream research trends are focusing on adaptive parameter adjustment during the printing process and real-time sensing of material states. However, existing general solutions often treat viscosity as a static input, neglecting the complex dynamic changes in viscosity due to external environment, curing reaction, resin flow conditions, and other factors. This is especially problematic for nonlinear, large-amplitude, and abrupt viscosity changes, where mainstream solutions exhibit slow response and low prediction accuracy, leading to significant lag and misjudgment issues in compensation strategies. Some studies have attempted to introduce auxiliary means such as aperture expansion temperature feedback, power regulation, or image recognition to enhance compensation response capabilities. However, these methods increase system integration complexity or are only applicable to specific resin systems. Most existing technologies have not built a reusable strategy library, lack structured archiving and efficient matching capabilities for various actual viscosity evolution modes, and lack long-term knowledge accumulation and transfer learning mechanisms, making it difficult to fundamentally avoid strategy failure phenomena under complex operating conditions. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring.

[0004] The technical solution of this invention is implemented as follows: a dynamic compensation method for photosensitive resin 3D printing parameters based on real-time viscosity monitoring, comprising: S1: Acquire real-time viscosity monitoring data of high viscosity gradient resin during continuous printing. The data is collected by an online viscosity sensor installed in the feeding system at a fixed sampling frequency to form a time series of raw viscosity signals. S2: Perform sliding window segmentation processing on the original viscosity signal, and perform differentiation operation in each window to calculate the viscosity change rate and acceleration between adjacent sampling points, and extract the time-viscosity dynamic feature set including inflection point position, monotonicity transition and local extrema; S3: Construct a local dynamic trend model based on the time-viscosity dynamic feature set. This model characterizes the nonlinear behavior pattern of resin viscosity evolution in the current printing stage and outputs a quantifiable trend feature vector, which is used as the basis for matching subsequent compensation strategies. S4: Input the trend feature vector into the pre-constructed graph-based compensation strategy library for similarity matching. The graph-based compensation strategy library is generated by training on multi-condition experimental data in the offline stage. It contains parameter adjustment paths corresponding to various typical viscosity evolution modes. Each path uses the 'input feature-output action' mapping relationship as a strategy node. The nodes are connected by similarity weight edges to form a topology. S5: Perform nearest neighbor search and confidence evaluation in the graph-based compensation strategy library, identify the subgraph structure with the highest matching degree with the current trend feature vector, and output the initial compensation strategy node and its associated action parameter combination as a candidate set of control instructions to be printed in the current layer. S6: Based on the quality feedback error of the actual printed layer, perform incremental transfer learning update on the initial compensation strategy node, propagate the error information back to the corresponding strategy node, adjust its output parameter weights, generate an optimized compensation strategy that has been fine-tuned online, and realize the knowledge accumulation and adaptive evolution of the strategy graph. S7: Generate specific printing parameter adjustment instructions based on the optimized compensation strategy, including the combination configuration of exposure time, platform lifting speed and interlayer waiting time, and send the instructions to the motion control system and the light source modulation unit to complete the closed-loop parameter compensation execution; S8: Continuously monitor the real-time viscosity signal of the next cycle and return to step S2 to form a dynamic compensation loop; if a sudden change in trend characteristics is detected and exceeds the preset similarity threshold, the strategy map reconstruction mechanism is triggered to start the new working condition label recording and offline retraining process.

[0005] The method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring provided by this invention has the following beneficial effects: (1) By introducing a sliding window piecewise differential analysis and local dynamic trend modeling mechanism, this invention achieves fine capture of real-time viscosity change rate, acceleration and inflection point characteristics, and can identify key turning points in viscosity evolution with higher time resolution, significantly advance the timing of compensation and intervention, and greatly reduce the risk of interlayer curing defects caused by model delay, thereby significantly improving the stability and process robustness of the printing process without increasing the system hardware cost. (2) This invention constructs a graph-based compensation strategy library, organizes the "input feature-output action" mapping relationship corresponding to various typical viscosity change patterns (such as exponential increase, step jump, oscillation decay, etc.) accumulated offline into a strategy node network with a topological structure, and realizes the association expression between patterns through similarity weight edges; during operation, the system quickly locates the optimal matching subgraph and activates the corresponding control strategy by performing nearest neighbor search and confidence evaluation on the local trend feature vector extracted in real time, avoiding the computational burden brought about by solving complex optimization problems from scratch; at the same time, combined with the incremental transfer learning mechanism, the output parameter weights of relevant strategy nodes are updated in reverse using the actual feedback error of the current layer, so that the system has the ability to continuously adapt and evolve, and can realize knowledge reuse and strategy fine-tuning under unknown or gradually changing working conditions, which significantly improves the accuracy and intelligence level of control decision-making; (3) The present invention adopts a closed-loop control architecture based on locally interpretable modeling and structured strategy transfer, which not only reduces the dependence on high-precision long-term viscosity prediction, but also ensures the real-time performance and engineering deployability of the algorithm through lightweight spectrum storage and efficient retrieval mechanism. The execution module dynamically generates key process parameter combinations such as exposure time, lifting speed and waiting time according to the selected strategy to form a complete dynamic compensation closed loop, ensuring that each printing layer can still maintain good molding quality and interface consistency under complex viscosity gradient conditions. This method is particularly suitable for high-end additive manufacturing scenarios such as functional graded materials and multiphase composite resins, which significantly improves the printing success rate and molding accuracy, while having good scalability and process compatibility, providing a new technical path for realizing a high degree of freedom and high reliability intelligent printing control system. Attached Figure Description

[0006] Figure 1 This is a flowchart of the dynamic compensation method for photosensitive resin 3D printing parameters based on real-time viscosity monitoring according to the present invention. Figure 2 This is a sub-flowchart of the photosensitive resin 3D printing parameter dynamic compensation method based on real-time viscosity monitoring of the present invention; Figure 3 This is another sub-flowchart of the dynamic compensation method for photosensitive resin 3D printing parameters based on real-time viscosity monitoring according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which 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.

[0008] 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.

[0009] like Figure 1 As shown, this invention provides a method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring, specifically including: S1: Acquire real-time viscosity monitoring data of high viscosity gradient resin during continuous printing. The data is collected by an online viscosity sensor installed in the feeding system at a fixed sampling frequency to form a time series of raw viscosity signals. S2: Perform sliding window segmentation processing on the original viscosity signal, and perform differentiation operation in each window to calculate the viscosity change rate and acceleration between adjacent sampling points, and extract the time-viscosity dynamic feature set including inflection point position, monotonicity transition and local extrema; S3: Construct a local dynamic trend model based on the time-viscosity dynamic feature set. This model characterizes the nonlinear behavior pattern of resin viscosity evolution in the current printing stage and outputs a quantifiable trend feature vector, which is used as the basis for matching subsequent compensation strategies. S4: Input the trend feature vector into the pre-constructed graph-based compensation strategy library for similarity matching. The graph-based compensation strategy library is generated by training on multi-condition experimental data in the offline stage. It contains parameter adjustment paths corresponding to various typical viscosity evolution modes. Each path uses the 'input feature-output action' mapping relationship as a strategy node. The nodes are connected by similarity weight edges to form a topology. S5: Perform nearest neighbor search and confidence evaluation in the graph-based compensation strategy library, identify the subgraph structure with the highest matching degree with the current trend feature vector, and output the initial compensation strategy node and its associated action parameter combination as a candidate set of control instructions to be printed in the current layer. S6: Based on the quality feedback error of the actual printed layer, perform incremental transfer learning update on the initial compensation strategy node, propagate the error information back to the corresponding strategy node, adjust its output parameter weights, generate an optimized compensation strategy that has been fine-tuned online, and realize the knowledge accumulation and adaptive evolution of the strategy graph. S7: Generate specific printing parameter adjustment instructions based on the optimized compensation strategy, including the combination configuration of exposure time, platform lifting speed and interlayer waiting time, and send the instructions to the motion control system and the light source modulation unit to complete the closed-loop parameter compensation execution; S8: Continuously monitor the real-time viscosity signal of the next cycle and return to step S2 to form a dynamic compensation loop; if a sudden change in trend characteristics is detected and exceeds the preset similarity threshold, the strategy map reconstruction mechanism is triggered to start the new working condition label recording and offline retraining process.

[0010] Step S1: Acquire real-time viscosity monitoring data of the high viscosity gradient resin during continuous printing. This data is collected at a fixed sampling frequency by an online viscosity sensor installed in the feeding system, forming a time-series raw viscosity signal. Specifically, this includes: S1.1: Based on the characteristic that the physical state of high viscosity gradient resin continuously evolves during continuous printing, a rotary microfluidic online viscosity sensor with temperature compensation function is deployed in the resin feeding system pipeline. The built-in differential pressure sensing unit applies a constant shear field to the resin flowing through the channel to obtain the corresponding flow resistance signal as the original sensing input. Based on the nonlinear evolution of the physical state of high viscosity gradient resin over time during continuous printing, a rotary microfluidic online viscosity sensor (parameters: temperature compensation module enabled, shear field strength constant at preset value) is installed at a critical process location in the resin supply system pipeline to achieve real-time acquisition of the shear response of the resin flowing through it. Furthermore, the differential pressure sensing unit (parameter: the resolution of the micro differential pressure sensing chip is better than 0.1 Pa) generates the original signal of the corresponding flow resistance. This signal reflects the instantaneous rheological impedance characteristics of the resin under constant shear field and is output to the signal conditioning module in real time. Furthermore, a temperature compensation algorithm is adopted (parameter: the compensation coefficient is obtained by linear fitting of the resistance response deviation at standard temperature based on the instantaneous temperature measured by the thermistor of the sensing cavity). This achieves temperature drift correction of the sensing signal and obtains the compensated steady-state flow resistance data, thus eliminating the interference of resin temperature changes on viscosity measurement. Furthermore, a stable shear flow field is formed in the microfluidic channel by a micro rotating impeller driven by a constant rotation speed, which ensures uniform stress distribution in the sensing cavity and minimizes secondary flow effects, thereby improving the mapping accuracy between the pressure difference signal and the true viscosity. Furthermore, a high-precision timing sampling mechanism (parameter: sampling frequency fixed at ≥10 Hz) is adopted to continuously sort the temperature-compensated flow resistance signal by time, guiding subsequent analog-to-digital conversion and noise suppression processing, thereby ensuring the synchronization of the acquisition link; By combining a rotary microfluidic online viscosity sensor with a temperature compensation algorithm, the original physical shear resistance signal is transformed into stable and calibrable raw flow resistance data, enabling continuous sensing of the real state of high viscosity gradient resin during the printing process. For example, in a high-functional-gradient photosensitive resin printing task, a rotary microfluidic online viscosity sensor is selected, and its shear field strength is set to... Pa, the calibration coefficient of the built-in temperature compensation module is Pa / °C. The sensor is installed downstream of the photocuring platform. The feed line at cm is equipped with a full-scale differential pressure sensing unit. kPa, accuracy Pa silicon micro differential pressure chip. The real-time acquired differential pressure signal is compensated using the temperature compensation formula: ,in This is the original pressure difference. Instantaneous temperature (unit: °C). The standard temperature is used to obtain the compensated pressure difference. .by As input, at a constant speed A stable shear flow field is formed under rpm conditions, and the pressure difference data is sampled at regular intervals. Accumulate 200 points within a second window and output them in sorted order. Under this configuration, the temperature drift of the output signal is significantly suppressed, and a stable input is provided for subsequent bandpass filtering and viscosity conversion models, thereby achieving high-precision real-time sensing of the resin viscosity state during actual printing. S1.2: Perform analog-to-digital conversion and noise suppression processing on the flow resistance signal, and use a bandpass filtering algorithm (0.1–10 Hz) to filter out mechanical vibration and power supply interference components to generate denoised digital flow resistance time series data, which is used to establish a linear mapping relationship with resin viscosity; For the analog electrical signal of flow resistance output by the rotary microfluidic online viscosity sensor installed in the resin feeding system pipeline, a high-resolution analog-to-digital conversion method (parameters: sampling bit depth ≥ 16 bit, sampling frequency ≥ 1 kHz) is used to realize the digital acquisition of the continuously changing sensor analog voltage signal. Furthermore, by using a bandpass filtering algorithm (parameters: lower limit frequency 0.1 Hz, upper limit frequency 10 Hz, filter order 4), the mechanical vibration and power supply interference frequency bands in the digital signal are suppressed, and a denoised signal data sequence with the spectrum concentrated in the target detection frequency band is obtained. Furthermore, by using the window function weighting method (parameter: the length of the Hanning window is equal to the length of the filter coefficients), the boundary effect of the bandpass filtered signal is weakened, and a smooth signal sequence after window weighting is generated to ensure the continuity of amplitude transition in the time domain; Furthermore, through amplitude normalization (parameters: maximum amplitude mapped to 1, minimum amplitude mapped to...), 1) Achieve consistency in signal amplitude across different acquisition batches to generate amplitude-standardized flow resistance time-series data; Through the above analog-to-digital conversion and filtering, the original analog signal of the rotary microfluidic sensor is converted into digital flow resistance time series data that meets the viscosity mapping requirements in terms of both stability and noise suppression performance, thereby achieving the technical effect of establishing a linear mapping relationship between flow resistance and resin viscosity. For example, the analog voltage signal of flow resistance from a rotary microfluidic sensor is input to a TI ADS1262 high-precision Σ-Δ analog-to-digital converter chip, with a sampling bit depth of 24 bits and a sampling frequency of 1.2 kHz, to obtain the raw digital signal stream. An IIR bandpass filter algorithm is applied to the digital signal, with the lower cutoff frequency set to 0.1 Hz, the upper cutoff frequency set to 10 Hz, and the filter order set to 4th order, to suppress the 8 Hz mechanical vibration component from the stepper motor and the 50 Hz rectified interference from the switching power supply. A Hanning window with a length equal to the filter coefficient length is used for weighting to reduce frequency leakage at the signal endpoints. The weighted signal is then normalized to ensure that the maximum amplitude corresponds to 1.0 and the minimum amplitude corresponds to 1.0. 1.0. After the above processing, the signal-to-noise ratio of the obtained flow resistance time series data in the target frequency band is significantly improved in the spectral analysis, the power spectral density curve has no interference peaks, and the amplitude drift is eliminated, providing high-quality input for the subsequent flow resistance-resin viscosity conversion model based on linear fitting; S1.3: Based on the pre-calibrated flow resistance-viscosity conversion model, the denoised flow resistance time series data is converted into an absolute viscosity value sequence. The conversion model is obtained by fitting the experimental calibration results of multiple sets of standard viscosity liquids under the same shear conditions, and the output is the original viscosity signal with timestamp alignment. For the digital flow resistance time series data processed by S1.2, a pre-established flow resistance-viscosity conversion model (parameters: constant shear rate, temperature correction coefficient) is used to map the resistance signal into an absolute viscosity value. Furthermore, by fitting the transformation function with calibration data of multiple sets of standard viscosity liquids, and using the least squares fitting method (model form: polynomial or power function), the quantitative relationship of viscosity corresponding to different resistance values ​​is calculated, and the fitting coefficient matrix is ​​obtained. Furthermore, by incorporating a temperature compensation term, the fitting coefficients are corrected in real time. Based on the temperature sequence output by the temperature sensing unit of the sensor, a linear correction formula is used to compensate for the temperature drift effect of the viscosity value, and a temperature-corrected viscosity calculation module is generated. Furthermore, by calculating the viscosity value corresponding to the resistance value at each sampling point, a sequence of absolute viscosity values ​​corresponding to the timestamp is generated to ensure that the values ​​are synchronized with the sampling time. The calculation is as follows:

[0011] in, This is the absolute viscosity value. This is the flow resistance value after noise reduction. , The coefficient is obtained through calibration with a standard liquid. For model index; Furthermore, mean smoothing is performed on the sequence (with a window length of 3–5 sampling periods) to reduce the interference of short-term fluctuations on trend analysis and improve the stability of subsequent time-viscosity dynamic feature extraction. Through the above calculations, the digital resistance time series data of S1.2 is transformed into the original viscosity signal aligned with the timestamp, thereby realizing the quantitative characterization of the material state. For example, in a practical photosensitive resin 3D printing system, a shear rate of 50 s is selected. - The rotary microfluidic online viscosity sensor¹ acquired noise-reduced flow resistance ranging from 0.25 to 0.80 Pa. During offline calibration, experiments were conducted using five standard silicone oils of known viscosities (0.5, 1.0, 1.5, 2.0, and 2.5 Pa·s), and the fitted transformation model coefficients were a=2.15, b=0.05, and n=1.12. During operation, the sensor temperature output was 25.8℃, and the system's set temperature correction coefficient was [missing value]. 0.003 Pa·s / ℃, the calculated temperature compensation value is 0.00774 Pa·s, and added to the viscosity calculation results. For example, if the resistance value at a sampling point is 0.62 Pa, substitute it into the conversion formula to calculate:

[0012] The calculated viscosity value is approximately 1.435 Pa·s, which is corrected to 1.427 Pa·s after temperature compensation. This value is then bound to the corresponding timestamp to generate sequence elements. A mean smoothing process with a window length of 3 is performed on the continuous sequence to reduce the interference of instantaneous fluctuations on the next feature extraction step. Ultimately, the original viscosity signal sequence output by this embodiment exhibits a stable trend and can be used for subsequent time synchronization processing in S1.4, significantly improving the viscosity state perception accuracy in the dynamic compensation strategy. S1.4: Perform time synchronization processing on the original viscosity signal, and perform periodic sampling alignment based on the interlayer motion trigger pulse of the printing platform to ensure that each sampling point corresponds to the time node of a specific printing layer, and generate a time-series viscosity data stream with process context correlation; S1.5: The synchronized time-series viscosity data stream is cached in a circular buffer and continuously updated and output at a fixed sampling frequency (≥10Hz) to form the original viscosity signal input set for subsequent differential analysis, ensuring the real-time and continuous nature of the data supply.

[0013] Step S2: Perform sliding window segmentation on the original viscosity signal, and perform differentiation within each window to calculate the viscosity change rate and acceleration between adjacent sampling points, extracting a time-viscosity dynamic feature set including inflection point positions, monotonic transitions, and local extrema. Specifically, this includes: S2.1: Based on the original viscosity signal obtained from the feeding system, a fixed-length sliding time window is set to segment and slice the time series data to isolate the local viscosity behavior of different printing stages, and output a series of time series subsequences as input units for subsequent differential analysis. Based on the raw viscosity signal acquired from the feeding system after time synchronization processing, a fixed-length sliding window slicing method is adopted (parameter: window length). Unit: seconds; Step size: (in seconds) to achieve phased segmentation of time series data, so as to introduce controllable local analysis intervals in the data stream and isolate the viscosity response behavior of different printing stages; Furthermore, through a window boundary alignment algorithm (parameter: trigger marker timestamp array) This enables a precise correspondence between the start and end points of the sliding window and the triggering time between layers of the printing platform, and outputs a time-series slice sequence aligned with the boundary to ensure that subsequent differential analysis is synchronized with the printing process status. Furthermore, a data integrity check method within a window is adopted (parameter: sampling frequency). Threshold of valid data points within the window This allows for the determination of the number of sample points and sampling continuity for each slice, and the generation of a set of window indices marked as "valid" or "invalid". Furthermore, the basic statistics (mean) of the original viscosity signal within the statistical window are used to further analyze the viscosity. ,variance This allows for a preliminary quantitative assessment of noise levels and data stationarity, and generates corresponding statistical feature vectors as predictive indicators before subsequent first-order central difference calculations. Furthermore, a dynamic buffer distribution mechanism is utilized (parameter: window overlap ratio). The tested valid window subsequences are output sequentially to the differential analysis module according to the overlap ratio to form a continuous set of time-series subsequences, which are used to calculate the instantaneous viscosity change rate and acceleration. By using the above-mentioned sliding window slicing and boundary alignment processing method, the original viscosity signal is transformed into a time-series subsequence with process state mapping capability, realizing spatial isolation and temporal alignment of local viscosity behavior, and providing data assurance for the accuracy of inflection point, monotonicity transition and local extremum extraction. For example, in the high viscosity gradient resin DLP printing process, the sampling frequency Configured as Hz, sliding window length Set as Seconds, step size Set as Seconds, window overlap rate for The window boundary alignment algorithm aligns the starting point of each window with the timestamp indicating the completion of the Z-axis lift of the printing platform. The array length matches the number of layers to be printed, ensuring that each window covers the complete data segment before and after curing of a printed layer. In the window integrity check, Set as A number of sampling points are used to eliminate window slices with fewer than a certain number of points. In the statistical evaluation phase, the mean and variance of each slice are calculated, for example, the mean viscosity within a certain window. for Pa·s, variance for The window is marked as "stable and effective". The continuous time-series subsequences output by the buffer distribution mechanism enter the differential algorithm module to calculate the instantaneous rate of change and acceleration sequence. In this scenario, the actual effect is that the inflection point detection error is significantly reduced, the time deviation of monotonic transition recognition is reduced, and the stability of the interlayer curing quality of high viscosity gradient resin is greatly improved. S2.2: Perform a first-order central difference algorithm on the time-series subsequence within each sliding window to calculate the viscosity change rate between adjacent sampling points, generate a local change rate curve to characterize the instantaneous evolution rate of resin viscosity in the current time period, and output a continuous change rate data sequence. S2.3: Based on the viscosity change rate data sequence generated in the previous step, the first-order central difference algorithm is further applied to perform second-order differential processing to calculate its change acceleration, obtain the dynamic acceleration characteristics in the viscosity evolution process, and output as an acceleration time series to identify the nonlinear inflection point and trend change interval of the viscosity response. S2.4: Jointly perform sign change detection and extreme value search on the viscosity change rate and acceleration sequence to identify the zero crossover point of the change rate, the peak point of acceleration and the moment of monotonicity transition, extract the inflection point position, the rising / falling interval transition mark and the local extreme point, and generate a structured time-viscosity dynamic feature set; For the viscosity change rate time series and acceleration time series output by S2.3, a sign change detection algorithm (parameters: detection threshold set to zero crossover tolerance ±0.5%, signal-to-noise ratio lower limit of 20dB) is used to automatically identify zero crossover points in the change rate series; Furthermore, by using an extreme value search algorithm (parameters: local window radius set to 3 sampling points, peak value determination threshold is mean + 2 times standard deviation), positive and negative peak points in the acceleration time series are detected, and peak position and amplitude data are obtained, which are used to identify instantaneous acceleration extreme events in viscosity evolution; Furthermore, based on the joint analysis results of the rate of change symbol sequence and the acceleration symbol sequence, a monotonicity transition discrimination method (parameter: the shortest length of the continuous identical symbol segment is 5 sampling points) is adopted to locate the starting time of the interval when the viscosity changes from rising to falling or from falling to rising within the current window, and generate a monotonicity transition flag; Furthermore, by combining the zero-crossing point sequence and the acceleration peak point sequence, an inflection point extraction function is used to calculate the precise location coordinates of the inflection point on the time axis through timestamp matching and amplitude weight merging, and a structured list of inflection points is output. By using the above-mentioned joint algorithm of sign change detection and extreme value search, the result of the previous step is transformed into a structured time-viscosity dynamic feature set containing inflection point positions, monotonicity transition indicators and local extreme points, so as to achieve the expected technical effect of high-precision characterization of local nonlinear viscosity evolution trend. For example, in a continuous printing task with high viscosity gradient resin, the sampling frequency is set to 20Hz, the window length is 200ms, and the zero-crossing detection threshold for the rate of change sequence is set to ±0.002Pa·s. -1 The lower limit of the signal-to-noise ratio was 20 dB, and the zero-crossing points were detected at sampling points 45, 88, and 132. The acceleration sequence was analyzed using a sliding window extreme value search with a radius of 3 sampling points. The peak value was determined as the mean plus two standard deviations. A positive peak was detected at sampling point 46, with an amplitude of 0.015 Pa·s. -2 The negative peak is located at the 133rd sampling point, with an amplitude of -0.012 Pa·s. -2 Based on the rate of change and acceleration symbol sequences, the monotonicity transition discrimination parameter is set as the symbol consistency threshold of five consecutive sampling points. The interval from rising to falling is determined to start at point 45, and the interval from falling to rising is determined to start at point 132. An inflection point extraction function is used to match the zero-crossing points with the acceleration peak points using timestamps and then weighting them by amplitude to obtain a list of inflection points: point 45 (falling inflection point) and point 132 (rising inflection point). Monotonicity transition markers and local extreme point information are generated as a structured feature set output. In this task, the generated feature set is used as input for subsequent local dynamic trend models, effectively improving the accuracy of trend matching and the response performance of the control strategy. S2.5: The extracted inflection point positions, monotonicity transition markers, and local extreme points are encoded into multi-dimensional feature vectors in time stamp order, and integrated to form a dynamic feature set that can be used to characterize the nonlinear viscosity evolution mode within the current window, serving as the input basis for constructing a local dynamic trend model.

[0014] like Figure 2 As shown, step S3 involves constructing a local dynamic trend model based on the time-viscosity dynamic feature set. This model characterizes the nonlinear behavior pattern of resin viscosity evolution during the current printing stage and outputs a quantifiable trend feature vector, which serves as the basis for matching subsequent compensation strategies. Specifically, this includes: S3.1: Based on the time-viscosity dynamic feature set output by S2, identify key change patterns within each sliding window, including inflection point positions, monotonicity transition intervals, and local extreme points. Use piecewise linear fitting and sign discrimination to mark the sign transformation of viscosity change rate and acceleration, and generate a structured feature sequence containing trend turning semantics to explicitly express the nonlinear dynamic characteristics in the viscosity evolution process. Based on the time-viscosity dynamic feature set output by step S2, the input objects include the timestamps of inflection point positions, monotonicity transition intervals, and local extreme points within each sliding time window, along with the corresponding sets of viscosity change rate and acceleration values. A piecewise linear fitting method is used (parameter: the length of each segment depends on the time difference between adjacent feature points) to achieve a local linear approximation of the viscosity change rate curve within the window, and the slope and intercept values ​​of each segment are used as quantitative indicators of trend strength and initial value. Furthermore, by using a sign discrimination method (parameter: zero value as the sign switching threshold), the sign switching detection of the rate of change and the acceleration of change is realized, and the sign sequence data is obtained to mark the directional characteristics of viscosity change within the current interval; Furthermore, the rate of change symbol sequence and the acceleration symbol sequence are cross-compared to calculate the index position of the symbol transition and the corresponding timestamp, and output the trend inflection point identifier set to characterize the moment when the viscosity curve reverses its trend within the window; Furthermore, the slope values ​​generated by piecewise linear fitting and the time positions of trend inflection points are jointly encoded into a structured feature sequence. Each sequence element includes the interval start time, end time, slope, intercept, direction of change, and acceleration mode category, thereby realizing an explicit expression of the nonlinear viscosity dynamic characteristics within the window. Through the above symbol discrimination and piecewise fitting processing, the time-viscosity feature set is transformed into a structured feature sequence carrying trend reversal semantics, so as to realize the high-fidelity expression of local nonlinear dynamic characteristics in the subsequent normalization and pattern matching stages. For example, in a continuous printing scenario with a high viscosity gradient resin, the sliding window length is set to 0.5 seconds, and the number of sampling points within the window is 10. The input viscosity change rate curve shows a zero-crossing sign change at the 3rd and 7th sampling points. Using a piecewise linear fitting method, with the 1st to 3rd sampling points, the 3rd to 7th sampling points, and the 7th to 10th sampling points as segment intervals, the slopes of each segment are calculated as follows: , , Pa·s / sampling point, intercepts are respectively , , Pa·s. The sign discrimination method is used to detect the rate of change and acceleration, and the sign sequence of the rate of change is [+, +, , , , The acceleration symbol sequence is [+, +, +, +, +]. , , ,+,+,+, , The cross-matching results (+) output trend inflection point identifiers at the 3rd and 7th sampling points. The above segmented slope, intercept, and trend inflection point information are encoded into structured feature sequence elements. For example, element 1 contains the start time (0.0s), end time (0.15s), and slope... ,intercept The direction is "upward" and the acceleration mode is "positive and negative reverse"; element 2 includes the start time 0.15s, end time 0.35s, and slope. ,intercept The direction is "descending" and the acceleration mode is "negative-positive rotation"; element 3 includes the start time 0.35s, end time 0.5s, and slope. ,intercept The sequence, characterized by an upward direction and a mixed positive and negative acceleration pattern, serves as the input for subsequent normalization and trend template matching in S3.2. This enables an explicit characterization of the nonlinear evolution pattern under high gradient viscosity variation scenarios, effectively improving the accuracy of compensation strategy matching. S3.2: Perform pattern normalization processing on the structured feature sequence, map it to a feature space of uniform dimension, use dynamic time warping (DTW) alignment method to eliminate the influence of sampling time offset, generate standardized trend segment templates as the basic unit for constructing local dynamic trend models, and ensure that the features extracted under different working conditions are comparable. S3.3: Based on standardized trend segment templates, a local dynamic trend basis function library is constructed. Each basis function corresponds to a typical local evolution pattern (such as accelerated rise, deceleration, and oscillating transition). The measured feature sequence of the current window is decomposed into a weighted linear combination of basis functions using a sparse coding algorithm to obtain a set of sparse weight coefficients, forming a compact representation of the current viscosity change behavior. Based on the standardized trend segment templates generated by S3.2, the evolution patterns of each segment in the template set are classified by the pattern classification method to form a set of typical local evolution patterns. Each pattern corresponds to a specific viscosity change feature label, which is used to construct the initial index of the local dynamic trend base function library. Furthermore, based on the above set of patterns, basis functions are generated using curve fitting and feature function expansion methods. Each basis function corresponds to a typical local evolution pattern, including accelerated rise, deceleration, and oscillating transition. The parameters are determined by a combination of statistical measures such as the amplitude, slope, and inflection point density of the trend segment template, so that the basis functions can characterize the principal component distribution of the actual observed pattern in the feature space. Furthermore, a sparse coding algorithm (L1 norm constraint, with the sparsity coefficient λ adaptively set according to the noise level) is used to encode the measured feature sequence of the current window. Represented as a set of basis functions The weighted linear combination of the given terms has an objective function that minimizes the sum of the reconstruction error and the sparse penalty term. The sparse weight coefficient vector is then calculated. This enables a compact characterization of local viscosity evolution patterns; Furthermore, the sparse weight coefficients are obtained by solving the following optimization problem:

[0015] in, This is the measured feature sequence for the current window. Let k be the local dynamic trend basis function. Here, N represents the corresponding sparse weight coefficients, and N is the number of sampling points. For sparse constraint coefficients, The L1 norm of the weighting coefficients is used to control the number of non-zero coefficients; By iteratively minimizing the objective function through an optimized solver, a sparse weight coefficient vector w is obtained, ensuring a high-precision approximation of the original feature sequence with a low amount of basis function usage. Through this sparse coding process, the standardized trend segment template is transformed into a set of sparse weight coefficients, realizing a compact representation and pattern condensation of the current viscosity change behavior, and meeting the computational efficiency and interpretability requirements of subsequent matching and migration strategies. For example, in the printing process of a high-viscosity gradient resin, the standardized trend segment template length is 120 sampling points, and the basis function library contains 8 typical patterns, which are respectively fitted by linear increase curves, slow decay curves, and periodic fluctuation curves obtained from previous simulations and experiments. The measured feature sequence of the current window is input into the sparse coding algorithm, and settings are... A value of 0.05 corresponds to a signal-to-noise ratio of 20dB at the noise level. Optimization using the coordinate descent method yields non-zero sparse weight coefficients with only three basis functions, which are: =0.62、 =0.27、 =0.11, reducing the reconstructed mean square error to one-tenth of the original. This sparse weighted coefficient vector serves as an accurate representation of the current viscosity evolution mode. When subsequently input into the graph-based compensation strategy library, it can match the corresponding accelerated rise + slight oscillation composite mode node in a very short time, significantly improving the response speed and matching accuracy of strategy selection. S3.4: The sparse weight coefficients and the corresponding basis function indices are jointly encoded into a multidimensional trend feature vector. This vector not only reflects the intensity and direction of the current viscosity change, but also carries the category information of the evolution mode. As the final output of the local dynamic trend model, it is used to support high-precision semantic matching in the graph-based compensation strategy library. S3.5: Perform confidence assessment on the generated trend feature vector, calculate the reconstruction error between it and the original measured signal, and combine it with the data signal-to-noise ratio index within the sliding window to determine the reliability of the current local model; if the confidence is lower than the preset threshold, trigger the anomaly detection flag and retain the valid model output of the previous period to ensure the stability of the control decision.

[0016] like Figure 3 As shown, step S4 involves inputting the trend feature vector into a pre-built graph-based compensation strategy library for similarity matching. This graph-based compensation strategy library is generated offline based on multi-condition experimental data and includes parameter adjustment paths corresponding to various typical viscosity evolution modes. Each path uses an 'input feature - output action' mapping relationship as a strategy node, and the nodes are connected by similarity weight edges to form a topological structure. Specifically, this includes: S4.1: Based on the multi-condition experimental dataset accumulated in the offline phase, the initial topology of the graph-based compensation strategy library is constructed. The parameter adjustment path corresponding to each typical viscosity evolution mode is abstracted into a strategy node. Each strategy node encapsulates a set of 'input feature-output action' mapping relationship. The input features include viscosity change rate, acceleration and inflection point distribution pattern. The output actions include exposure time increment, lift speed adjustment amount and interlayer waiting time configuration combination, forming a reusable minimum decision unit. Based on the multi-condition experimental dataset collected offline, a condition label grouping method is adopted (parameter: label category is based on the resin viscosity evolution pattern classification standard) to achieve clustering of the original experimental data under different viscosity evolution patterns; Furthermore, through a feature extraction algorithm (parameters: viscosity change rate, first-order acceleration, inflection point time distribution), the input features of each type of viscosity evolution mode are templated, and a structured set of feature vectors is obtained. Furthermore, a parameter mapping construction method is adopted (parameters: exposure time increment range 0.01–0.5 seconds, platform lift speed adjustment range 0.1–2 mm / s, interlayer waiting time configuration combination range 0.5–5 seconds) to generate the mapping relationship between each set of input features and the corresponding optimal output action, and obtain a set of reusable parameter adjustment paths; Furthermore, by utilizing the strategy node encapsulation algorithm (parameter: the node data structure includes the input feature set, the output action set, and the execution record index), each parameter adjustment path is abstracted into a strategy node, and multiple nodes are organized into a set of the smallest decision units; By using the node topology initialization process, the set of strategy nodes generated in the previous step is transformed into the initial topology structure of the graph-based compensation strategy library, thus realizing the structured storage and online retrieval foundation of strategy knowledge in the offline stage. For example, in the printing scenario with high viscosity gradient resin, six different operating conditions were introduced during the offline experimental phase, including an exponential increase (the peak viscosity change rate is...). The viscosity change rate (Pa·s / s) exhibits linear gradual increase, stepwise jump, oscillating decay, stable maintenance, and sharp decline. For the original data of the exponentially increasing condition, the first and second order central difference methods using a sliding window are employed to calculate the peak value of the viscosity change rate time series. Pa·s / s, mean change in acceleration Pa·s / s², and the distribution of inflection point times is concentrated in the window of the first... To the Sampling points. Encode this feature set into an input vector. Based on offline parameter optimization experiments, the optimal output action combination under this condition was determined to be the exposure time increment. Adjustment amount of platform lifting speed per second mm / s, inter-floor waiting time The parameters are mapped in seconds. This mapping path is then encapsulated as a policy node, and the input features, output actions, and execution history index are stored in the node data structure. A similarity labeling method is used to group nodes with relevance ≥ The nodes establish initial connections in the library, resulting in a highly connected subgraph for subsequent online matching. This embodiment can directly invoke this strategy to match nodes with an exponential upward trend during the online phase, achieving coordinated adjustment of exposure time and lift-off speed. This significantly improves the interlayer quality of the printed layers in multiple batch tests and enhances molding stability. S4.2: Normalize the input feature vectors of each policy node, and calculate the similarity weights between nodes based on the Euclidean distance and dynamic time warping (DTW) joint metric method. Use a weighted undirected graph to construct the topological connection relationship between policy nodes, so that policy paths with similar viscosity evolution characteristics form highly connected subgraphs in the graph, thereby supporting smooth policy migration under nonlinear evolution mode. S4.3: The real-time trend feature vector output from the previous step S3 is used as the query vector and input into the constructed graph-based compensation strategy library. Based on the standardized feature space, a preliminary screening is performed to select a set of candidate strategy nodes with feature dimension consistency higher than the preset threshold, so as to narrow the subsequent search range and improve matching efficiency. S4.4: In the candidate strategy node set, the K-Nearest Neighbors (KNN) algorithm combined with the confidence scoring mechanism is used for fine matching. The comprehensive similarity score between the query vector and the central features of each candidate node is calculated. At the same time, the success rate and stability of the matching result in historical verification are evaluated, and the target strategy node with the highest similarity and the confidence score is output. The real-time trend feature vector output by S3 is used as the query vector input to the candidate strategy node set. The K-Nearest Neighbors (KNN) algorithm (parameters: number of neighbors K=5, distance metric is Euclidean distance and dynamic time warping DTW fusion weight w=0.6:0.4) is used to calculate the comprehensive similarity between each candidate node and the query vector. Furthermore, a similarity score function is constructed by fusing distance metric results, using a weighted summation form:

[0017] in, To calculate the overall similarity score, To integrate weight parameters, The result is the DTW distance metric. The result is the Euclidean distance metric. The confidence score is calculated as follows:

[0018] in, Score the confidence level. This represents the number of successful calls in the i-th call. This represents the number of failures in the i-th call. To count the total number of calls within the statistics window; Furthermore, the candidate strategy nodes are sorted in descending order of comprehensive similarity score through a sorting mechanism, and nodes with confidence scores below the threshold are removed to shrink the candidate set. By selecting those with the highest overall similarity score and confidence score... The nodes are used as target strategy nodes to ensure that the matched strategy actions not only closely resemble the current trend characteristics, but also have historically proven execution stability. By combining the KNN algorithm with a confidence scoring mechanism, the candidate node data initially screened in the previous step is transformed into a unique target policy node, thereby improving matching accuracy and enhancing the robustness of policy selection. For example, in a photosensitive resin 3D printing system, the real-time trend feature vector for a certain printing cycle is 12-dimensional. The mean Euclidean distance in the candidate set ranges from 0.15 to 0.32, and the mean DTW distance ranges from 0.12 to 0.28. The system sets the number of neighbors K to 5 and the distance fusion weight w to 0.6. Using the aforementioned comprehensive similarity formula, the S value of each candidate node is calculated, yielding the highest value node S=0.92. Historical call records show that this node achieved 40 successes and 10 failures in the past 50 calls. Substituting these into the confidence formula yields: The result was 18, which, after normalization, was higher than the threshold of 0.85. Based on this, the system selected this node as the target strategy node and used it to generate the action parameter combination of the initial compensation strategy. It was verified that the interlayer transition quality was significantly improved during actual printing, and the response speed during sudden viscosity changes remained within 200ms. S4.5: Extract the optimal parameter adjustment path associated with the target policy node, generate the action parameter combination of the initial compensation policy, and use it as the input basis for the next stage nearest neighbor search and confidence assessment, thus completing the semantic-level mapping transformation from real-time material state to potential control policy.

[0019] Step S5: Perform nearest neighbor search and confidence evaluation in the graph-based compensation strategy library to identify the subgraph structure with the highest matching degree to the current trend feature vector, and output the initial compensation strategy node and its associated action parameter combination as a candidate set of control instructions for printing the current layer. Specifically, this includes: S5.1: Based on a pre-built graph-based compensation strategy library, the strategy nodes and their similarity weight edges stored in its topology are used as the retrieval basis. Each strategy node represents the 'input feature-output action' mapping relationship under a typical viscosity evolution mode. The similarity weight edges are calculated based on multi-condition experimental data in the offline stage, reflecting the degree of pattern closeness between dynamic features of different viscosities. A graph adjacency matrix that supports fast indexing is generated, providing a structured basis for subsequent nearest neighbor search. Based on the complete topology of the pre-built graph-based compensation strategy library, a node data parsing method (parameters: node identifier, input feature dimension, output action composition) is used to extract the 'input feature-output action' mapping relationship of each strategy node in the library into an indexable data record, thereby realizing the structured storage of basic node information; Furthermore, by using the similarity weight calculation method (parameters: Euclidean distance metric result, dynamic time warping (DTW) path length), the similarity weight edge values ​​between nodes formed based on multi-condition experimental data in the offline stage are read and converted into edge attributes of a weighted undirected graph, thereby realizing the quantitative representation of the connection strength between nodes with different strategies. Furthermore, an adjacency matrix generation algorithm (parameters: total number of strategy nodes N, edge weight normalization coefficient) is adopted to perform a structural mapping from a weighted undirected graph to a matrix form, resulting in an adjacency matrix with node indices as row and column coordinates and edge weights as matrix element values, providing an efficient structured retrieval carrier for the subsequent nearest neighbor search process; Furthermore, by using a sparse matrix compression method (parameter: compression threshold ε), weak connection edges with weights lower than ε in the adjacency matrix are removed, and the matrix is ​​converted into a compressed storage format, reducing the participation of invalid edges during the traversal search process and reducing the retrieval computation load. Through the above matrix processing method, the topology of the strategy library is transformed into adjacency matrix data that retains the node feature relationships and has fast indexing capabilities, thereby enabling efficient retrieval support of the graph-based compensation strategy library in real-time matching scenarios. For example, in a graph-based compensation strategy library containing 50 strategy nodes, each node has a 10-dimensional input feature dimension, and its output actions include three types of parameters: exposure time increment, lift speed adjustment, and inter-layer waiting time. The similarity weights between nodes calculated offline range from 0.1 to 0.95. When using the adjacency matrix generation algorithm, the weight normalization coefficient is set to 1.0, standardizing the weights of each edge to the [0,1] interval. Adjacency matrix elements... Represents the similarity between nodes i and j, where The value is This means that the feature patterns of node 5 and node 10 are highly similar. During sparse matrix compression, a compression threshold is set. = To reduce computation, all elements with weights below 0.2 are set to zero. In subsequent searches, this adjacency matrix can return the top K neighbors with the highest weights that are directly connected to the query node in O(K) complexity, significantly improving strategy matching speed while maintaining accurate capture of highly similar patterns. S5.2: Normalize the trend feature vector output in S3 and input it as a query vector into the graph compensation strategy library. Use a similarity measurement algorithm based on Euclidean distance and dynamic time warping (DTW) to perform a local traversal search in the graph adjacency matrix to identify the K candidate strategy nodes with the smallest distance to the current trend feature vector and the highest connection weight, forming an initial matching node set to capture the most similar historical viscosity evolution pattern. S5.3: Based on the matching scores of each candidate strategy node in the initial matching node set and its topological context information in the graph, a confidence evaluation function is constructed. This function comprehensively considers three dimensions: node matching accuracy, neighborhood consistency, and historical successful call frequency. It calculates the comprehensive confidence score of each candidate node, selects high-confidence strategy nodes with confidence scores exceeding a preset threshold, and excludes mismatch results caused by noise or feature drift. Based on the initial set of matching nodes generated in S5.2, a weighted scoring fusion method is adopted (parameter: matching score weight). Neighborhood Consistency Weight Historical call frequency weight This enables quantitative evaluation of the confidence level of candidate nodes; Furthermore, the neighborhood consistency analysis algorithm (parameters: neighborhood depth D=2, similarity threshold σ=0.85) is used to detect the feature consistency of candidate nodes within their local topology range and obtain the neighborhood consistency index as the second component of the confidence assessment. Furthermore, by utilizing the historical successful call frequency normalization function (parameters: time decay coefficient λ=0.05, window period T=500 print layers), the weight of the candidate node call records is corrected, and a time-weighted historical success rate index is generated as the third component of the confidence assessment. Furthermore, a confidence assessment function is constructed based on the above three types of indicators, and the comprehensive score is performed using the following mathematical expression:

[0020] in, To calculate the overall confidence score, Score the node matching accuracy. It is a neighborhood consistency index. A time-weighted historical success rate metric; Furthermore, a confidence threshold comparison method (parameter: threshold θ=0.78) is used to filter high-confidence nodes and obtain a set of nodes that meet the confidence conditions, which serves as the input for subsequent strategy subgraph expansion; The initial matching node results output by S5.2 are transformed into a set of highly reliable strategy nodes through a comprehensive confidence evaluation algorithm, thereby achieving the expected technical effect of eliminating false matches caused by noise and feature drift. For example, in a high-viscosity gradient resin printing task, the initial matching node set size is K=5. The matching accuracy scores M of each node are 0.92, 0.88, 0.81, 0.77, and 0.69, respectively; the neighborhood consistency indices N are 0.95, 0.89, 0.85, 0.80, and 0.72, respectively; and the historical success rates H are 0.91, 0.86, 0.83, 0.78, and 0.70, respectively. The weight parameters are set as α=0.4, β=0.35, and γ=0.25. Substituting the above values ​​into the comprehensive scoring formula, we calculate C1≈0.93, C2≈0.88, C3≈0.83, C4≈0.79, and C5≈0.73. After comparison with a threshold θ=0.78, nodes 1, 2, 3, and 4 are selected as the high-reliability strategy node set. When this set was subsequently expanded into a locally connected subgraph in S5.4, it was verified that the molding quality remained stable in 30 consecutive printing tasks, which greatly improved the success rate of strategy matching under nonlinear viscosity change conditions. S5.4: Taking the strategy node corresponding to the highest confidence score as the core, expand along its similarity weight edge to extract the associated action parameters with collaborative adjustment logic in the adjacent nodes, and generate a locally connected subgraph structure. This subgraph represents a set of compatible and ordered parameter adjustment paths, which serve as the initial compensation strategy subgraph to ensure that the selected strategy not only matches the current trend, but also has the potential for migration to maintain consistency between the previous and subsequent processes. After the target strategy node with the highest confidence score is determined, the graph topology traversal method (parameters: starting node = node with the highest confidence score, traversal depth limit = 3) is used to realize the extended search along the similarity weight edge to capture the neighboring nodes that are highly correlated with the current master node in the viscosity evolution mode. Furthermore, by using a collaborative adjustment logic discrimination algorithm (parameters: output action similarity threshold = 0.85, trend category consistency requirement = true), the adjacent node set is filtered, retaining only nodes that are process-compatible with the master node in terms of output parameter configuration and have successful collaborative records in the historical migration path, and obtaining a list of compatible nodes; Furthermore, a parameter path serialization method (parameter: action parameter sorting rule = timestamp ascending order) is adopted to serialize the output parameter combination of each node in the compatible node list and generate an ordered path data structure containing exposure time increment, lift speed adjustment amount and interlayer waiting time offset, to ensure that the parameters meet the process continuity requirements in execution order; Furthermore, a local connectivity verification algorithm (parameters: connectivity threshold = 0.6, breakpoint tolerance = 1) is adopted to examine the connectivity of the ordered path data structure in the graph subgraph, confirm that the path forms a highly coherent topological link between nodes with similarity weight edges, and generate connectivity verification result indicators. By using a subgraph construction function, the set of ordered paths with qualified connectivity verification results is transformed into a locally connected subgraph structure, which is used as the output of the initial compensation strategy subgraph, realizing the mapping from a single high-confidence node to a strategy cluster with the potential for migration with consistency between front-end and back-end processes. For example, in the 150th layer of a high viscosity gradient resin printing task, the main strategy node obtained by matching the real-time viscosity trend feature vector has an output parameter combination of exposure time increment. Seconds, lifting speed adjustment amount mm / s, interlayer waiting time offset Seconds. Expand outwards along the similarity weight edge of the main node to obtain node A with a similarity weight of 0.92 (parameter: exposure time increment). Seconds, lifting speed adjustment amount mm / s, interlayer wait time offset Node B (parameter: exposure time increment) with a similarity weight of 0.88 and a similarity weight of 0.88. Seconds, lifting speed adjustment amount mm / s, interlayer wait time offset (seconds). Through collaborative adjustment logic, both node A and node B meet the trend category consistency and parameter similarity threshold conditions, and there are print layer sequences that successfully co-executed with the master node in the historical migration records. Their output parameters are arranged in ascending order of timestamps to form a path sequence [(master node parameters), (node ​​A parameters), (node ​​B parameters)]. Connectivity verification is performed in the graph subgraph, and the average connectivity of the path is calculated as follows: greater than the set threshold Furthermore, the breakpoint tolerance meets the requirements. The constructed locally connected subgraph structure includes three strategy nodes and their ordered parameter adjustment paths, which can be directly used as the initial compensation strategy subgraph. This subgraph can significantly improve the process consistency of parameter migration in the subsequent control instruction parsing stage and exhibit stable viscosity adaptability in cross-layer execution, thereby improving the forming quality between printing layers in high viscosity gradient scenarios. S5.5: Parse the main strategy node and its associated action parameter combination in the initial compensation strategy subgraph into a standardized control instruction format, including the exposure time reference value, platform lift speed gain coefficient and inter-layer waiting time offset, to form the control instruction candidate set for the current printing layer, and mark its source strategy label and matching confidence, for subsequent incremental transfer learning module to perform error feedback tracing and parameter fine-tuning.

[0021] Step S6: Based on the quality feedback error of the actual printed layer, incremental transfer learning is performed on the initial compensation strategy node to update it. The error information is backpropagated to the corresponding strategy node, its output parameter weights are adjusted, and an optimized compensation strategy with online fine-tuning is generated, realizing the knowledge accumulation and adaptive evolution of the strategy graph. Specifically, this includes: S6.1: Based on the geometric deviation data between the actual molding quality image of the previous printing layer and the design model, calculate the quality feedback error vector. The error vector includes the inter-layer contour offset, thickness inconsistency index and edge sharpness attenuation rate, which serve as the input supervision signal for incremental transfer learning to quantify the execution deviation of the current initial compensation strategy. S6.2: Input the quality feedback error vector into the local error response model associated with the initial compensation strategy node, use the weighted least squares method to decompose the error source, identify the dominant process parameter dimension with the greatest impact, form the parameter sensitivity weight distribution, and use it as the direction basis for subsequent parameter weight adjustment. S6.3: Based on the parameter sensitivity weight distribution and error gradient direction, perform gradient descent weight update on the 'output action' part in the initial compensation strategy node, use an adaptive learning rate algorithm to adjust the gain coefficient of each parameter output channel, generate a preliminary corrected compensation parameter combination, and reduce the expected modeling residual of the next level; Based on the parameter sensitivity weight distribution and error gradient direction output from step S6.2, the parameter channel set corresponding to the mapping relationship of the "output action" part of the current initial compensation strategy node is selected to determine the target parameter index that needs to be updated. The gradient descent algorithm (parameters: initial learning rate η0, maximum number of iterations T, gradient threshold ε) is used to calculate the corresponding gradient components for each objective parameter channel, and the weight signs are adjusted according to the error gradient direction to achieve parameter updates along the performance optimization direction. Furthermore, the update step size of each parameter channel is dynamically adjusted through an adaptive learning rate algorithm (parameters: β is the historical gradient decay coefficient, δ is the minimum learning rate lower bound, and γ is the learning rate growth coefficient). The new learning rate is calculated using a weighted function of the historical gradient magnitude and the current gradient change rate, achieving the dual effect of fast convergence and oscillation prevention. Using a weighted update formula, the weight of each parameter is calculated as follows:

[0022] in, For the updated parameter weights, The current parameter weights, The adaptively adjusted learning rate, Let the objective function be the quality feedback error. This represents the gradient component of the parameter; Furthermore, the expected modeling residuals for the next layer of printing are calculated using the residual prediction function, as shown in the formula:

[0023] in, To model the residuals as expected, This represents the actual forming quality index of the current layer. To design the target metrics of the model, The number of samples; By combining gradient descent and adaptive learning rate, the weight bias of the previous step is corrected into a new combination of compensation parameters, thereby achieving the technical effect of reducing the expected modeling residual of the next level. For example, in a high-viscosity gradient resin printing scenario, the quality feedback error vector of the previous layer is [contour offset 0.12mm, thickness inconsistency 0.03mm, edge sharpness attenuation rate 0.08], with corresponding parameter sensitivity weight distributions of exposure time 0.65, platform lift-up speed 0.25, interlayer waiting time 0.10, and error gradient direction vector of [-1.2, 0.5, -0.3]. The initial learning rate η0 for gradient descent is set to 0.05, β to 0.9, δ to 0.01, and γ to 1.05, the maximum number of iterations T to 50, and the gradient threshold ε to 1e-4. For the exposure time parameter, the initial gradient is -1.2. After adaptive adjustment, the learning rate increases to 0.052, and the weights are updated. The new weights are approximately 0.7124. For the platform lift speed parameter, the gradient is 0.5, the learning rate is adjusted to 0.051, and the new weights are... Approximately 0.2245. For the inter-layer waiting time parameter, the gradient is -0.3, the learning rate is adjusted to 0.049, and the new weights are... Approximately 0.1147. The corrected parameter combination [0.7124, 0.2245, 0.1147] is input into the residual prediction formula, n=3, yielding the expected modeling residual R approximately... The error was significantly smaller than the initial residual, which verified the role of this step in reducing the printing error of the next layer and improving printing consistency. S6.4: Compare the preliminary modified compensation parameter combination with historical successful strategy fragments, use the dynamic time warping algorithm to match the optimal reference trajectory, and determine whether there is a known stable control mode that can be reused; if so, integrate the smoothing constraint conditions of the mode to generate an optimized compensation strategy with stability guarantee to prevent over-adjustment from causing oscillations. S6.5: Write the optimized compensation strategy back to the corresponding initial compensation strategy node in the original strategy graph, update the similarity weight of its adjacent edges synchronously, and mark the node as 'online fine-tuning' to trigger the knowledge accumulation mechanism and support the enhanced strategy evolution and cross-condition migration capabilities in subsequent printing cycles.

[0024] Step S7: Generate specific printing parameter adjustment instructions based on the optimized compensation strategy, including a combination configuration of exposure time, platform lift-up speed, and interlayer waiting time, and send the instructions to the motion control system and the light source modulation unit to complete the closed-loop parameter compensation execution. Specifically, this includes: S7.1: Based on the optimized compensation strategy node updated through incremental transfer learning, the output action parameter weight vector is parsed. This vector contains three types of control variables: exposure time adjustment coefficient, platform lift speed offset, and inter-layer waiting time correction value, which serve as input conditions for parameter decoding. The weight vector is inversely standardized using a preset linear mapping function to restore its physical dimensions in the original control space, generating a set of basic printing parameters that is executable. S7.2: Input the basic printing parameter set into the multi-objective constraint optimization module, and perform Pareto front search by combining the geometric complexity features corresponding to the current printing layer with the material curing kinetic boundary conditions; wherein the geometric complexity features are obtained by statistical analysis of contour density and overhang angle in the slice data, and the curing kinetic boundary is derived from the resin activation energy and light intensity distribution model; by constructing a weighted penalty term, the parameter combinations that exceed the process feasible domain are corrected to generate a feasible printing parameter set that meets both stability and molding quality requirements; S7.3: Perform timing scheduling planning for each parameter in the feasible printing parameter set: Based on the response delay characteristics of the motion control system and the start / stop cycle of the light source modulation unit, calculate the time alignment offset of each parameter's effective time; use a finite state machine model to arrange the execution sequence of exposure time setting taking priority over platform lifting action and lifting-to-position signal triggering inter-layer waiting time, and generate a structured control instruction sequence with timestamps to ensure the synchronization and logical order correctness of multi-axis coordinated actions; S7.4: The structured control instruction sequence is encapsulated into a data frame format conforming to the industrial bus communication protocol. Specifically, the object dictionary mechanism of the CANopen protocol is used to bind the exposure time (corresponding to sub-index 01 of 0x2001), platform lifting speed (corresponding to sub-index 02 of 0x2001), and inter-layer waiting time (corresponding to sub-index 03 of 0x2001) to the register address. The data frame is sent to the motion control PLC and DLP optomechanical controller through the real-time Ethernet interface to complete the cross-device distribution of control instructions. The structured control instruction sequence output from step S7.3 is used as the input object, and the register address mapping and binding function of each print control variable is implemented by using the CANopen protocol object dictionary binding method (parameter: object index 0x2001, sub-index 01~03). Furthermore, the exposure time parameter is mapped to using the object dictionary configuration tool. Register address, mapping the platform lift-up speed parameter to Register address, mapping inter-layer wait time parameters to The register address is used to generate an address binding table for subsequent cross-device calls; The data frame encapsulation method (parameters: message identifier ID, data field length 8 bytes, protocol type CANopenPDO) is used to realize the binary serialization of control commands, arrange the update values ​​of each register in little-endian byte order in the data field, and attach a check code field to ensure the integrity of message transmission. Furthermore, by using the real-time Ethernet encapsulation method (parameters: transmission protocol type EtherCAT or TCP / IP, port number preset to industrial control standard port), the generated CANopen data frame is embedded in the real-time Ethernet frame payload to generate cross-device transmission data packets that conform to the industrial bus transmission specification, and the destination MAC address and IP address are configured as the network interface of the motion control PLC and DLP optomechanical controller. The Ethernet physical layer transmission module enables the line-driven output of the above industrial bus protocol data packets. The switch port mirroring function is used to monitor the link layer traffic. The protocol parsing module is started at the target device receiving end to decapsulate the data packets, restore them to CANopen data frames and write them to the mapping register, ensuring that each printed parameter takes effect in real time in the control loop. By using the above-mentioned industrial bus protocol encapsulation and transmission processing method, the control command sequence of the previous step is transformed into a standardized data message that can be distributed across devices, thereby achieving the effect of low-latency and synchronized transmission of printing control parameters between the motion control system and the light source modulation unit. For example, in a printing scenario with an exposure time base value of 2.5s, a platform lift-up speed gain coefficient of 1.2mm / s, and an interlayer waiting time offset of 0.8s, the control system binds these parameters to the register addresses corresponding to sub-indices 01, 02, and 03 of object dictionary index 0x2001, respectively, and encodes them in 16-bit unsigned integer format. Using the CANopen PDO data frame format, with a message ID configured as 0x180, the data field encodes the exposure time as 2500ms, the lift-up speed as 1200mm / s, and the waiting time as 800ms, arranged in little-endian order and appended with a CRC checksum of 0x3FA5. The data frame is encapsulated using the EtherCAT protocol, with port number set to 502. The destination IP addresses are 192.168.0.10 for the PLC controller and 192.168.0.20 for the DLP optomechanical controller. The real-time Ethernet driver module sends the data packet on the transmit link port, which is forwarded by the switch and parsed using the EtherCAT protocol at the receiving end. The controller writes the parsed CANopen data into the register and immediately refreshes the corresponding execution unit parameters. The effect is that the exposure time, lift-off speed, and interlayer waiting time parameters in the printing control closed-loop system are updated synchronously within milliseconds, significantly improving the response speed of parameter compensation and cross-device coordination, and reducing delays and error accumulation during the printing process of high-viscosity gradient resins. S7.5: Initiate a feedback confirmation mechanism during instruction execution: Receive position confirmation signals from the motion control system and exposure enable responses from the light source system to verify whether each execution unit has successfully loaded new parameters; if any unit returns a NACK response, initiate a retransmission mechanism and reduce subsequent compensation gain to prevent runaway; finally, form a record of executed print parameter adjustments as historical input for the next layer of viscosity monitoring and error analysis to maintain the continuity of the closed-loop control link.

[0025] Step S8: Continuously monitor the real-time viscosity signal for the next cycle and return to step S2, forming a dynamic compensation loop; if a sudden change in trend characteristics is detected and exceeds a preset similarity threshold, the strategy map reconstruction mechanism is triggered, and the new working condition label recording and offline retraining process is started. Specifically, this includes: S8.1: Based on the real-time viscosity signal of the next cycle after the output of the adjustment module in S7, the new time series original viscosity signal is continuously collected by the online viscosity sensor installed in the feeding system at a fixed sampling frequency, and used as the input data source for the new round of dynamic compensation cycle to maintain the time continuity of the closed-loop control system. S8.2: Perform sliding window segmentation processing on the new round of raw viscosity signal obtained in S8.1, and perform differentiation operation in each window to calculate the viscosity change rate and acceleration between adjacent sampling points. Extract the time-viscosity dynamic feature set containing inflection point position, monotonic transition and local extremum, and generate the trend feature vector of the current period for similarity comparison with historical strategy map nodes. S8.3: Based on the trend feature vector generated in S8.2, perform similarity matching with the input feature template of the existing policy node in the policy graph library, calculate the Euclidean distance and dynamic time warping (DTW) comprehensive similarity index, and determine whether it is lower than the preset similarity threshold; if it is higher than the threshold, it is determined that the trend feature has undergone a significant change, and enter the abnormal working condition identification process. S8.4: When a sudden change in trend features is identified in S8.3 and exceeds the preset similarity threshold, the strategy graph reconstruction mechanism is triggered, and the new working condition label recording process is started: the current trend feature vector and its corresponding printing parameter combination and quality feedback error sequence are packaged and marked as 'unknown high gradient mode' new working condition sample, and stored in the offline retraining database to provide incremental data support for subsequent model expansion; S8.5: Based on the new working condition sample set accumulated in S8.4, the offline retraining process is started during the idle period of the printing task: the clustering algorithm is used to classify the new samples into patterns, and the optimal parameter adjustment path under the corresponding viscosity evolution scenario is generated in combination with the simulation platform. New 'input feature-output action' mapping relationship nodes are constructed, and they are fused into the existing graph-based compensation strategy library through similarity weight edges to complete the topological structure expansion and knowledge update of the strategy graph.

[0026] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring, characterized in that, Includes the following steps: S1: Acquire real-time viscosity monitoring data of high viscosity gradient resin during continuous printing process to form a raw viscosity signal of time series; S2: Perform sliding window segmentation on the original viscosity signal, and perform differentiation operation in each window to calculate the viscosity change rate and acceleration between adjacent sampling points, and extract the time and viscosity dynamic feature set; S3: Construct a local dynamic trend model based on the aforementioned time and viscosity dynamic feature set, and output a trend feature vector; S4: Input the trend feature vector into the pre-built graph-based compensation strategy library for similarity matching, wherein the graph-based compensation strategy library is generated by training based on multi-condition experimental data in the offline stage; S5: Perform nearest neighbor search and confidence evaluation in the graph-based compensation strategy library, identify the subgraph structure with the highest matching degree with the current trend feature vector, and output the initial compensation strategy node and its associated action parameter combination; S6: Based on the quality feedback error of the actual printed layer, perform incremental transfer learning update on the initial compensation strategy node, propagate the error information back to the corresponding strategy node, adjust its output parameter weights, and generate an optimized compensation strategy. S7: Generate specific printing parameter adjustment instructions according to the optimized compensation strategy, and send the printing parameter adjustment instructions to the motion control system and the light source modulation unit to complete the closed-loop parameter compensation execution.

2. The method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring according to claim 1, characterized in that, Following step S7, the following is also included: S8: Continuously monitor the real-time viscosity signal of the next cycle and return to step S2 to form a dynamic compensation loop. If a sudden change in trend characteristics is detected and exceeds the preset similarity threshold, the strategy map reconstruction mechanism is triggered to start the new working condition label recording and offline retraining process.

3. The method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring according to claim 1, characterized in that, The real-time viscosity monitoring data is collected at a fixed sampling frequency by an online viscosity sensor installed in the feeding system.

4. The method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring according to claim 1, characterized in that, The time and viscosity dynamic feature set includes inflection point locations, monotonicity transition markers, and local extrema.

5. The method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring according to claim 1, characterized in that, Step S3 specifically includes: Based on the time and viscosity dynamic feature set output in step S2, key change patterns within each sliding window are identified, the sign conversion of viscosity change rate and change acceleration is marked, and a structured feature sequence is generated. The structured feature sequence is subjected to pattern normalization processing to generate a standardized trend segment template; Based on the standardized trend segment template, a local dynamic trend basis function library is constructed, and the measured feature sequence of the current window is decomposed into a weighted linear combination of basis functions to obtain sparse weight coefficients. The sparse weight coefficients and their corresponding basis function indices are jointly encoded into a multidimensional trend feature vector, which serves as the final output of the local dynamic trend model. A confidence assessment is performed on the multidimensional trend feature vector, the reconstruction error between it and the original measured signal is calculated, and the reliability of the current local model is judged by combining the data signal-to-noise ratio index within the sliding window. If the confidence is lower than the preset threshold, an anomaly detection flag is triggered and the output of the previous period's valid model is retained.

6. The method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring according to claim 1, characterized in that, Step S4 specifically includes: Based on the multi-condition experimental dataset accumulated in the offline phase, the initial topology of the graph-based compensation strategy library is constructed. Normalization is performed on the input feature vectors of each policy node, and the similarity weights between nodes are calculated based on the joint metric method of Euclidean distance and dynamic time warping. The topological connection relationship between policy nodes is constructed using a weighted undirected graph. The real-time trend feature vector output in step S3 is used as the query vector and input into the graph-based compensation strategy library. Based on the standardized feature space, a preliminary screening is performed to select a set of candidate strategy nodes whose feature dimension consistency is higher than a preset threshold. In the candidate strategy node set, the K-nearest neighbor algorithm combined with the confidence score mechanism is used for fine matching. The comprehensive similarity score between the query vector and the central features of each candidate node is calculated. At the same time, the success rate and stability of the matching result in historical verification are evaluated, and the target strategy node with the highest similarity and the confidence score meets the conditions is output. Extract the optimal parameter adjustment path associated with the target strategy node to generate the action parameter combination of the initial compensation strategy.

7. The method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring according to claim 6, characterized in that, In the initial topology of the graph-based compensation strategy library, the parameter adjustment path corresponding to each typical viscosity evolution mode is abstracted into a strategy node. Each strategy node encapsulates a set of input features and output action mapping relationships. The input features include viscosity change rate, acceleration, and inflection point distribution pattern. The output actions include exposure time increment, lift speed adjustment, and interlayer waiting time configuration combination.

8. The method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring according to claim 1, characterized in that, Step S5 specifically includes: Based on a pre-built graph-based compensation strategy library, the strategy nodes and their similarity weight edges stored in its topology are used as the retrieval basis to generate a graph adjacency matrix. Normalization is performed on the trend feature vector output in step S3, and it is input into the graph compensation strategy library as a query vector. A similarity measurement algorithm based on Euclidean distance and dynamic time warping is used to perform a local traversal search in the graph adjacency matrix to identify the K candidate strategy nodes with the smallest distance to the current trend feature vector and the highest connection weight, forming an initial matching node set. Based on the matching scores of each candidate strategy node in the initial matching node set and its topological context information in the graph, a confidence evaluation function is constructed. Taking into account three dimensions—node matching accuracy, neighborhood consistency, and historical successful call frequency—highly reliable strategy nodes with confidence scores exceeding a preset threshold are selected. Taking the policy node corresponding to the highest confidence score as the core, we expand along its similarity weight edge to extract the associated action parameters with collaborative adjustment logic in the adjacent nodes, and generate a locally connected subgraph structure as the initial compensation policy subgraph. The main strategy node and its associated action parameter combination in the initial compensation strategy subgraph are parsed into a standardized control instruction format to form a control instruction candidate set for the current printing layer, and its source strategy label and matching confidence are marked.

9. The method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring according to claim 8, characterized in that, The similarity weight edges are calculated based on multi-condition experimental data in the offline stage, reflecting the degree of pattern similarity between dynamic characteristics of different viscosities.

10. The method for dynamic compensation of photosensitive resin 3D printing parameters based on real-time viscosity monitoring according to claim 8, characterized in that, The standardized control command format includes the exposure time reference value, the platform lift-up speed gain coefficient, and the interlayer waiting time offset.

Images