Adaptive adjustment method for 3D printing parameters of workpieces based on dual-mode monitoring

Abstract

This invention relates to the field of 3D printing, specifically to an adaptive adjustment method for 3D printing parameters of manufactured parts based on dual-modal monitoring. The method includes: collecting current time-series data of a servo motor; extracting current statistical features from the current time-series data to determine viscosity and material breakage risk; capturing images of the part forming area; extracting high-dimensional semantic features of the forming area; quantifying filament length and collapse area ratio to determine degradation degree and defect risk; determining the rPET strip grade by comprehensively considering the mean current, viscosity, material breakage risk, filament length, collapse area ratio, degradation degree, and defect risk; inputting the rPET strip grade, current time-series data, current statistical features, high-dimensional semantic features, and bottle recycled material identification information into a pre-trained model; and outputting fine-tuning parameters under the adjustment direction determined by the rPET strip grade; and adjusting the 3D printing parameters of the manufactured part. This invention can reduce the material breakage rate and defect rate while leveraging the advantages of low cost and environmental friendliness of rPET strips.

Description

Technical Field

[0001] This invention relates to the field of 3D printing, and more specifically to an adaptive adjustment method for 3D printing parameters of parts based on dual-modal monitoring. Background Technology

[0002] Large 3D printed parts (such as industrial molds and large structural parts) require a large amount of consumables, and virgin PET is expensive and resource-intensive. Recycled PET strips (rPET strips) made from cut plastic bottles are widely available and cost only 40%-60% of virgin materials, while also being environmentally friendly, making them an ideal consumable for large parts. However, rPET strips face several problems in actual printing: due to differences in recycling processes, their melt viscosity fluctuates greatly, and their thermal stability is poor, making them prone to thermal degradation during printing, resulting in a high risk of material breakage and high rates of defects such as stringing and collapse; at the same time, existing monitoring methods are limited and lagging, failing to correlate the "viscosity-degradation-parameter" relationship in real time; in addition, large parts require the use of recycled bottle materials from different brands, batches, and years of use, which vary significantly in terms of additive residues, crystallinity, and mechanical strength, further exacerbating the fluctuation of filament feeding resistance and the difficulty of parameter adaptation, severely restricting the large-scale application of rPET strips.

[0003] Therefore, there is an urgent need for a method that can monitor the melt viscosity and degradation state of rPET strips in real time, accurately adapt to the differences in recycled materials for different bottles, adaptively adjust printing parameters, and reduce the risk of material breakage and defects, so as to promote the large-scale application of rPET strips in the 3D printing of large parts. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide an adaptive adjustment method for 3D printing parameters of parts based on dual-modal monitoring. It can reduce the breakage rate and defect rate while taking advantage of the low cost and green environmental protection of rPET strips.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is: an adaptive adjustment method for 3D printing parameters of a part based on dual-modal monitoring, comprising:

[0006] During the 3D printing process, the current timing data of the servo motor is collected, and images of the forming area of ​​the part are captured.

[0007] A sliding window is used to extract current statistical features from current time series data, including at least the current mean and variance; based on a pre-calibrated mapping relationship between current, viscosity and material breakage risk, viscosity and material breakage risk are determined according to the current mean.

[0008] Based on the high-dimensional semantic features of the forming area of ​​the part, the drawing length and the proportion of the collapsed area are extracted from the image of the forming area of ​​the part. Based on the pre-calibrated mapping relationship between drawing length / collapse-degradation degree-defect risk, the degradation degree and defect risk are determined according to the obtained drawing length and the proportion of collapsed area.

[0009] The rPET strip grade is determined by considering factors such as average current, viscosity, material breakage risk, fiber length, collapse area ratio, degradation degree, and defect risk.

[0010] The rPET strip level, current time series data, current statistical features, high-dimensional semantic features, and bottle recycled material identification information are input into the pre-trained model. The model outputs fine adjustment parameters under the adjustment direction determined by the rPET strip level.

[0011] Adjust the 3D printing parameters of the part based on fine-tuning parameters.

[0012] Furthermore, the mean and variance of the current are after compensation based on the strip size error compensation strategy; wherein, the strip size error compensation strategy is specifically as follows:

[0013] Calculate the ratio of the standard size to the actual size of the rPET strip;

[0014] If the ratio is within the preset compensation coefficient range, the ratio shall be used as the compensation coefficient.

[0015] If the ratio is less than the lower limit of the compensation coefficient interval, the lower limit of the compensation coefficient interval shall be used as the compensation coefficient.

[0016] If the ratio is greater than the upper limit of the compensation coefficient range, the upper limit of the compensation coefficient range shall be used as the compensation coefficient.

[0017] Multiply the compensation coefficient by the statistical mean current and variance current respectively to obtain the compensated mean current and variance current.

[0018] Furthermore, high-dimensional semantic features of the forming region are extracted from the image of the forming region of the part, and the wire drawing length and the proportion of collapsed area are quantified. A CNN model is used for this purpose. The CNN model includes:

[0019] The preprocessing module is used to scale the image of the forming area of ​​the part to a preset size and perform normalization and data augmentation;

[0020] The MobileNetV3-small encoder contains multiple depthwise separable convolutional modules for extracting multi-scale feature maps from preprocessed images;

[0021] The CBAM attention module is used to sequentially apply channel attention weighting and spatial attention weighting to the multi-scale feature map to obtain an enhanced feature map as a high-dimensional semantic feature.

[0022] The U-Net decoder is used to perform multiple upsampling on the enhanced feature map and fuse the feature map of the corresponding scale in the MobileNetV3-small encoder with the upsampled feature map through skip connections;

[0023] The dual-branch output layer includes a parallel segmentation branch and a quantization branch. The segmentation branch outputs a 3-channel segmentation map representing the background, wire drawing, and collapse. The quantization branch outputs a 2-channel quantization map representing the wire drawing length and the proportion of the collapse area.

[0024] Furthermore, the rPET strip grade is determined by comprehensively considering factors such as average current, viscosity, breakage risk, fiber length, collapse area ratio, degradation degree, and defect risk; specifically including:

[0025] First, calculate the normalized score for each indicator;

[0026] Then, weights are assigned to each indicator, and the normalized scores of each indicator are summed in weights to obtain the initial total score.

[0027] Further verify whether the key indicators, such as average current, wire length, and collapse, are all within the grade threshold range corresponding to the initial total score;

[0028] Finally, the initial score is adjusted according to the type of recycled bottle material, additive residue, and mechanical strength to obtain the final score, and the rPET grade is determined based on the final score.

[0029] Further, the normalized scores of each indicator are calculated; specifically:

[0030] For the positive indicators of current and viscosity, the score = corresponding value / corresponding quality threshold. If the score is greater than 1, the value is 1.

[0031] For the reverse indicators of material breakage risk, wire drawing length, and collapse ratio, the following rules apply: When the actual

[0032] When the value is greater than or equal to the qualified threshold, the score is 1 - (corresponding value - corresponding qualified threshold) / (corresponding poor quality threshold - corresponding qualified threshold). If the score is less than 0, the value is 0. When the actual value is less than the qualified threshold, the base score is 0.8, and the excess performance bonus is added. The final score is 0.8 + (qualified threshold - corresponding value) / qualified threshold × 0.2. If the score is greater than 1, the value is 1.

[0033] For neutral indicators such as degradation degree and defect risk, the following rules apply: When the actual value is greater than or equal to the warning threshold, the score is 1 - (corresponding value / corresponding warning threshold). If the score is less than 0.5, the value is 0.5. When the actual value is less than the warning threshold, a base score of 0.5 is used as the passing score, and additional performance bonuses are added. The final score is 0.5 + (warning threshold - corresponding value) / warning threshold × 0.5. If the score is greater than 1, the value is 1. The degradation degree warning threshold is a 30% decrease in intrinsic viscosity, and the defect risk warning threshold is 15%.

[0034] In the above-mentioned neutral indicator calculation rules, 0.5 is the base score for the pass line. Its core functions are: first, to be compatible with the minimum score requirement of the original rules and ensure that the judgment logic remains unchanged under the deterioration state; second, to unify the scoring logic of the neutral indicator with the positive and negative indicators, with the better the indicator, the higher the score; and third, to accurately quantify the safety margin of the indicator being better than the warning threshold through the excess bonus space of 0.5.

[0035] For core evaluation indicators such as mean current and viscosity, if the associated stability indicators exceed the corresponding high-quality threshold, the normalized score of the core indicators needs to be corrected downwards. The correction range is determined according to the degree of exceeding the standard of the stability indicator, with the maximum correction range for a single indicator not exceeding 0.2, and the lowest corrected score not lower than 0.5. Among them, the associated stability indicator for mean current and viscosity is the current variance, with the high-quality threshold for current variance being ≤0.8A², the qualified threshold being 0.8-1.2A², and the poor-quality threshold being >1.5A².

[0036] Furthermore, the adjustment direction determined by the rPET strip grade is specifically as follows:

[0037] If the rPET strip is rated as excellent, adjust the printing temperature and speed accordingly.

[0038] If the rPET strip grade is qualified, the adjustment direction is to maintain or fine-tune the printing temperature and printing speed;

[0039] If the rPET strip is of poor quality, adjust the printing temperature and printing speed to reduce them, or trigger a material change reminder.

[0040] Furthermore, the 3D printing parameters of the part are adjusted based on fine-tuning parameters and configured to be executed in the order of priority: preventing material breakage, reducing defects, and printing efficiency. When the adjustment instructions output by the analytical model find that there is a conflict between the adjustment directions of different parameters, the priority determination mechanism is triggered.

[0041] Furthermore, when the adjustment instructions output by the analytical model detect conflicts in the adjustment directions of different parameters, a priority determination mechanism is triggered; specifically, this includes:

[0042] If the risk of material shortage is ≥50%, then speed increase will be rejected and speed will be reduced by 5-10 mm / s. When the servo current is ≤ current quality threshold, the extrusion speed will be increased and the printing speed will be reduced. When the risk of material shortage is ≥80%, printing will be stopped directly, an alarm will be triggered and the progress will be saved.

[0043] If the risk of material breakage is less than 50%, then when the wire drawing length is a moderate defect, limit the temperature rise and maintain or slightly increase the printing speed. When the collapse area ratio is in a qualified state, only fine adjust the printing temperature or cooling air speed, and prohibit any adjustment that increases the risk of defects.

[0044] Full heating and speed increase will be implemented only when the risk of material shortage is less than 20% and the risk of defect is less than 15%. Additional speed increase will be implemented when the material is qualified, the current variance is ≤0.5A² and there are no obvious defects, but the speed will not exceed the preset maximum threshold. If the risk exceeds the limit, the parameters will be adjusted back.

[0045] If the rPET strip is in a poor quality state for more than 5 seconds and there is no improvement after adjustment, a material change reminder is triggered. The system will automatically pause and save the current layer coordinates. After the material change, an empty wire feed calibration will be performed and the reference current will be updated before printing resumes.

[0046] Furthermore, the model includes:

[0047] The input layer receives rPET strip level, current timing data, current statistical features, high-dimensional semantic features, and bottle recycling material identification information.

[0048] The data preprocessing layer is used to perform Z-Score normalization and temporal alignment on the input features;

[0049] The LSTM time series feature extraction layer uses a 3-layer stacked bidirectional LSTM network to extract the dynamic dependency features of current time series data fluctuations.

[0050] The Transformer global association layer adopts a 6-layer encoder architecture, with each layer containing an 8-head multi-head self-attention mechanism and a feedforward neural network. It takes the LSTM temporal feature vector, high-dimensional semantic features and bottle recycling material identification information as input, calculates the global attention weights between multi-source features, and outputs a multi-dimensional global association feature vector.

[0051] The large component adaptation layer generates a 64-dimensional adaptation vector based on the component size parameters. It assigns a weight coefficient of 1.2-1.5 times to four core features: total printing time, single-layer printing area, interlayer cooling time, and long-sequence current variance threshold. The vector is then fused with the multi-dimensional global correlation feature vector.

[0052] The status level and recycled material adaptation layer selects a pre-trained multi-dimensional exclusive adaptation vector based on the rPET strip level and bottle recycled material identification information, and multiplies it element-wise with the fusion features from the large component adaptation layer to obtain the enhanced features;

[0053] The fully connected output layer uses a two-layer fully connected network to map the enhanced features into process parameter adjustment schemes and early warning signals, and corrects the printing temperature for ambient temperature.

[0054] Furthermore, the bottle body recycled material identification information includes the crystallinity of the recycled material, the type and content of additive residues, and the mechanical strength grade;

[0055] The crystallinity is obtained through offline preset, online correction, and automatic calculation. Offline preset associates the crystallinity range with the recycled material identifier. Online correction dynamically corrects the crystallinity based on the servo motor current through the current-crystallinity mapping relationship. Automatic calculation matches the default value with the identifier and initial current and makes online fine adjustments when no crystallinity is entered. With 28% as the boundary, recycled material with a crystallinity ≥28% is considered high-crystallinity recycled material, and recycled material with a crystallinity <28% is considered low-crystallinity recycled material. For high-crystallinity recycled material, the maximum torque and current of the motor are limited to no more than 120% of the rated value. For low-crystallinity recycled material, the pressure of the feeding wheel is reduced to 5N and a speed priority control mode is adopted. When the current is continuously lower than the reference current -0.4A for more than 0.3s, the anti-material interruption mode is triggered, automatically reducing the speed, increasing the extrusion speed, and issuing a warning.

[0056] Additive Residue: Offline entry of the residual content levels of antioxidants, colorants, and plasticizers (low <0.1%, medium 0.1%-0.5%, high ≥0.5%); for recycled materials containing antioxidant residues, the upper limit of printing temperature is increased by 5℃, and the maximum allowable temperature rise is increased by 2℃; for recycled materials with colorant residues ≥0.5%, the extrusion pressure is increased by 0.1-0.2MPa, while the upper limit of printing speed is reduced by 5mm / s; for recycled materials with plasticizer residues ≥3%, the upper limit of printing temperature is reduced by 3℃, and printing above 250℃ is prohibited;

[0057] Mechanical strength: The tensile strength grade of the recycled material is entered offline (≥50MPa is high strength, 30-50MPa is medium strength, <30MPa is low strength); for medium strength recycled material, the upper limit of printing speed is reduced by 10mm / s; for low strength recycled material, the filament feeding speed is reduced by 15%, the extrusion pressure is increased by 0.2MPa, and any speed-up operation is prohibited. The material breakage risk warning threshold is reduced from 20% to 15%.

[0058] By adopting the above technical solution, this invention simultaneously collects dual-modal data of servo motor current and forming area image, extracts statistical features such as mean and variance from current time series data, and determines the melting state in real time based on the current-viscosity-material breakage risk mapping relationship. At the same time, it uses a CNN model to extract high-dimensional semantic features from the image and quantifies the wire drawing length and collapse area ratio. Then, it assesses the material degradation degree based on the wire drawing / collapse-degradation degree-defect risk mapping relationship, and then comprehensively weights seven indicators to determine the rPET strip grade. The LSTM-Transformer hybrid model outputs fine adjustment parameters in combination with the bottle recycled material identification information under the adjustment direction determined by the grade, thereby realizing accurate perception and closed-loop adaptive adjustment of the rPET strip printing state. This method can reduce the rate of material breakage, control the product defect rate, and has a response speed much faster than manual adjustment. It can also adapt to the differences in crystallinity, additive residue, and mechanical strength of recycled bottle materials from different brands and batches, ensuring the stability of continuous printing of large parts. Using rPET strips cut from plastic bottles as consumables reduces costs by 40%-60% compared to virgin materials and realizes the resource utilization of plastic waste. Moreover, it does not require modification of the core structure of existing FDM printers, only the addition of sensors and software integration, resulting in low modification costs and easy large-scale promotion. Attached Figure Description

[0059] Figure 1 The flowchart shows the adaptive adjustment method for part 3D printing parameters based on dual-modal monitoring according to the present invention.

[0060] Figure 2 This is a framework diagram of the CNN model of the present invention;

[0061] Figure 3 This is a framework diagram of the LSTM-Transformer hybrid architecture of the present invention. Detailed Implementation

[0062] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0063] like Figure 1 As shown, an adaptive adjustment method for 3D printing parameters of a part based on dual-modal monitoring includes:

[0064] Step S1, data acquisition.

[0065] During the 3D printing process, current timing data of the servo motor is collected, and images of the forming area of ​​the part are captured.

[0066] The servo motor refers to the motor that drives the printing platform or gantry to perform large-format movements. A current sensor with a sampling frequency ≥100Hz is used to collect the servo motor current timing data; the current sensor must be equipped with an anti-electromagnetic interference shielding shell. An industrial camera is used to capture images of the part forming area, with a frame rate ≥30fps, and a ring LED fill light is used to photograph the printed part forming area. The camera lens is equipped with an anti-fog coating (to cope with printhead heat dissipation vapor), and the telescopic bracket has an obstacle avoidance travel of ≥5cm to prevent collisions with the printhead.

[0067] Step S2: Feature extraction and association analysis.

[0068] A sliding window is used to extract current statistical features from current time series data, including at least the current mean and variance; based on a pre-calibrated mapping relationship between current, viscosity and material breakage risk, viscosity and material breakage risk are determined according to the current mean.

[0069] Based on the high-dimensional semantic features of the part forming area extracted from the image of the part forming area, and the wire drawing length and the proportion of collapsed area quantified, the degree of degradation and the defect risk are determined based on the pre-calibrated mapping relationship of wire drawing length / collapse-degradation degree-defect risk, and the obtained wire drawing length and the proportion of collapsed area are used to determine the degree of degradation and the defect risk.

[0070] The steps for establishing the mapping relationship are as follows: First, based on offline experimental calibration and multiple linear regression modeling, the mean current is quantitatively correlated with melt viscosity and material breakage risk (the correlation process includes two steps: ① Offline modeling stage: based on measured data, a multiple linear regression model of 'mean current - melt viscosity - material breakage risk' is established (e.g., viscosity = 0.2 × mean current + 0.05 × current variance + 0.1, material breakage risk = 0.3 × (0.8 - viscosity) + 0.02); ② Online application stage: the mean current is substituted into the model to calculate the melt viscosity and material breakage risk of the current rPET strip in real time); Second, by using the degree of degradation (quantified by the rate of decrease in intrinsic viscosity) as a bridge, the correlation between the fiber length / collapse ratio and the degree of degradation and defect risk is established to adapt to the degradation characteristics of different recycled materials.

[0071] Step S3, rPET strip status level determination.

[0072] The grade of recycled PET strips used for printing is determined by considering factors such as average current, viscosity, material breakage risk, filament length, collapse area ratio, degradation degree, and defect risk. This is known as the rPET strip grade. The rPET strip grades are categorized as high-quality, acceptable, and low-quality. Generally, high-quality rPET strips (new bottle recycling, 30%-35% crystallinity): high melt viscosity (≥0.8 Pa·s), high filament feeding resistance, and motor output current reaching the high-quality threshold; acceptable rPET strips (routine recycling): medium melt viscosity (0.5-0.8 Pa·s), with current stable within the acceptable threshold range; low-quality rPET strips (recycled multiple times, severely degraded): low melt viscosity (<0.5 Pa·s), current dropping to the low-quality threshold or below, low filament feeding resistance, and prone to filament pulling / collapse defects.

[0073] Step S4: Model generates adjustment scheme.

[0074] The rPET strip level, current time series data, current statistical features, high-dimensional semantic features, and bottle recycled material identification information are input into the pre-trained model. The model outputs fine adjustment parameters under the adjustment direction determined by the rPET strip level.

[0075] The adjustment direction determined by the rPET strip grade is specifically as follows:

[0076] If the rPET strip is of high quality (i.e., high viscosity, low stringiness, and low degradation), the adjustment direction is to increase the printing temperature (increase by 3-5℃) and increase the printing speed (increase by 5-10mm / s) to improve the printing efficiency of large parts;

[0077] If the rPET strip grade is qualified (i.e., medium viscosity, no obvious stringing / collapse), the adjustment direction is to maintain or fine-tune the printing temperature and printing speed;

[0078] If the rPET strip is of poor quality (i.e., low viscosity, severe stringing / collapse, severe degradation), the adjustment direction is to reduce the printing temperature (reduce by 5-8℃) and the printing speed (reduce by 10-15mm / s). If there is still no improvement after adjustment, trigger the material shortage warning or material replacement reminder.

[0079] Taking into account the antioxidant properties of Material A, the model automatically and temporarily raised the upper limit of the printing temperature from 250℃ to 254℃, allowing a temperature increase of 4℃ within the safe range; at the same time, since Material A is a high-strength recycled material, the upper limit of the printing speed remains unchanged at 75mm / s, and the speed increase of 8mm / s meets the requirements.

[0080] Step S5, parameter execution.

[0081] The 3D printing parameters for the parts are adjusted based on fine-tuning parameters. It is configured to execute adjustments in the order of priority: preventing material breakage, reducing defects, and improving printing efficiency. When the adjustment instructions output by the analytical model detect conflicts in the adjustment directions of different parameters, a priority determination mechanism is triggered. Additionally, the printing temperature is controlled between 225-255℃ (to avoid thermal degradation), the printing speed is 30-80mm / s, and printing is immediately stopped and an early warning is issued when the risk of material breakage is ≥80%. The early warning signals are divided into three levels: Emergency Warning; Important Warning; and General Warning.

[0082] Step S6: Feedback and closed-loop iteration.

[0083] Within 0.5 seconds of adjustment, recheck the dual-modal data to verify whether risks and defects have been mitigated. After printing, record the adjustment parameters, product defect rate, and consumable usage. Summarize ≥30 sets of valid data daily, and ≥100 sets cumulatively over 7 days. Valid data is preferentially stored on the local hard drive (SD card expansion supported) and synchronously backed up to the encrypted cloud. In case of abnormal power failure, automatically restore the most recent set of adjustment parameters and printing progress to avoid data loss. Incremental learning is used to optimize model parameters, enhance adaptability to large-part printing scenarios, form a closed-loop iteration, and add an optional "cross-device data sharing" function (which users can disable), sharing anonymized "recycled material type - parameters - effect" data.

[0084] In this embodiment, preferably, in step S2, the mean and variance of the current are after compensation based on the strip size error compensation strategy; wherein, the strip size error compensation strategy is specifically as follows:

[0085] Calculate the ratio of the standard size to the actual size of the rPET strip;

[0086] If the ratio is within the preset compensation coefficient range, the ratio shall be used as the compensation coefficient.

[0087] If the ratio is less than the lower limit of the compensation coefficient interval, the lower limit of the compensation coefficient interval shall be used as the compensation coefficient.

[0088] If the ratio is greater than the upper limit of the compensation coefficient range, the upper limit of the compensation coefficient range shall be used as the compensation coefficient.

[0089] Multiply the compensation coefficient by the statistical mean current and variance current respectively to obtain the compensated mean current and variance current.

[0090] The specific instructions are as follows: The standard size is based on experimental calibration of the optimal fit size for common manual cutting. In this embodiment, 4mm × 8mm is the standard size, and the upper limit of the compensation coefficient is set to 1.5 and the lower limit is set to 0.7. If the calculation result exceeds this range, it is corrected according to the boundary value (e.g., if the cutting size is 3.5mm × 4mm, the compensation coefficient = standard cross-sectional area / actual cross-sectional area = 2.286). If the entered cutting size exceeds the range of 3-5mm × 5-10mm, the system will first pop up a window prompting "The cutting size exceeds the recommended range (3-5mm × 5-10mm), which may affect the printing stability. Do you want to continue?". After the user selects "continue", the system will then correct according to the boundary value of the compensation coefficient, and at the same time, it will continuously prompt "Size abnormal, it is recommended to optimize the cutting size" during the printing process.

[0091] In this embodiment, preferably, in step S2, high-dimensional semantic features of the forming region of the part are extracted based on the image of the forming region of the part, and the wire drawing length and the proportion of collapsed area are quantified, using a CNN model. Figure 2 As shown, the CNN model includes:

[0092] The preprocessing module is used to scale the image of the forming area of ​​the part to a preset size and perform normalization and data augmentation;

[0093] The MobileNetV3-small encoder contains multiple depthwise separable convolutional modules for extracting multi-scale feature maps from preprocessed images;

[0094] The CBAM attention module is used to sequentially apply channel attention weighting and spatial attention weighting to the multi-scale feature map to obtain an enhanced feature map as a high-dimensional semantic feature.

[0095] The U-Net decoder is used to perform multiple upsampling on the enhanced feature map and fuse the feature map of the corresponding scale in the MobileNetV3-small encoder with the upsampled feature map through skip connections;

[0096] The dual-branch output layer includes a parallel segmentation branch and a quantization branch. The segmentation branch outputs a 3-channel segmentation map representing the background, wire drawing, and collapse. The quantization branch outputs a 2-channel quantization map representing the wire drawing length and the proportion of the collapse area.

[0097] Among them, while quantifying the length of the filament and the proportion of the collapsed area, the last convolutional layer of the encoder (MobileNetV3-small) of the CNN model outputs 128-dimensional high-dimensional semantic features (including deep information such as target edges, texture, and shape). These features are not only used for defect quantification, but also serve as core image features reflecting the printing status.

[0098] Specifically, CNN is a deep learning model specifically designed for grid-structured data (such as images and time-series signals). Its core functionality involves extracting local features and achieving global correlation through convolution and pooling operations. Leveraging its parameter sharing and sparse connectivity, it significantly outperforms traditional algorithms in feature extraction efficiency and accuracy, making it the core technology supporting the quantification of filament length and collapse area ratio in this system. Its core design philosophy simulates the "local perception-feature integration" logic of the human visual system: first, low-level neurons perceive simple features such as edges and textures in the image; then, high-level neurons fuse these simple features into complex features (such as the linear shape of filaments and the contour of collapsed areas), ultimately achieving accurate target recognition and quantification. In this 3D printing system, the CNN model is used to quantify the filament length and collapse area ratio of the printed part, extracting key visual features from images of the formed area captured by an industrial camera, providing data support for determining the degradation degree and defect risk of rPET strips. The CNN model architecture adopts an improved U-Net+MobileNetV3 hybrid structure, which combines depthwise separable convolution to achieve lightweighting while balancing accuracy and real-time performance (single frame processing ≤0.1s). The CBAM attention mechanism is introduced to enhance the recognition of string / collapsed target features and suppress background interference in industrial scenes. The output layer is adapted to quantization requirements, directly outputting the target segmentation map and associating it with pixel-physical size calibration coefficients to accurately quantize the length / area ratio, rather than just performing classification.

[0099] In this embodiment, preferably, in step S3, the rPET strip grade is determined by comprehensively considering the average current, viscosity, material breakage risk, fiber length, collapse area ratio, degradation degree, and defect risk; specifically including:

[0100] First, calculate the normalized score for each indicator:

[0101] For the positive indicators of current and viscosity, the score = corresponding value / corresponding quality threshold. If the score is greater than 1, the value is 1.

[0102] For the reverse indicators of material breakage risk, wire drawing length, and collapse ratio, the following rules apply: When the actual

[0103] When the actual value is greater than or equal to the qualified threshold, the score is calculated as follows: 1 - (corresponding value - corresponding qualified threshold) / (corresponding substandard threshold - corresponding qualified threshold). If the score is less than 0, the value is 0. When the actual value is less than the qualified threshold, a base score of 0.8 is used as the qualified threshold, and additional performance bonuses are added. The final score is calculated as follows: 0.8 + (qualified threshold - corresponding value) / qualified threshold × 0.2. If the score is greater than 1, the value is 1. Among these, the qualified threshold for material breakage risk is 20%, and the substandard threshold is 50%; the qualified threshold for wire drawing length is 2mm, and the substandard threshold is 5mm; the qualified threshold for collapse area ratio is 3%, and the substandard threshold is 5%.

[0104] In the above-mentioned reverse indicator scenario calculation rules, 0.8 is the base score for the pass line. Its core functions are: first, to fill the scenario loopholes in the original formula and ensure that the scores of the reverse indicator in the entire numerical range are all within the normalized reasonable range of 0-1; second, to match the judgment threshold of ≥0.8 for the PET strip of this method, so as to achieve the unity of the scoring logic of positive and negative indicators; and third, to distinguish the performance difference between the pass minimum and the high-quality excess, and to accurately quantify the performance advantage of the indicator that is better than the pass requirement through the excess bonus space of 0.2.

[0105] For neutral indicators such as degradation degree and defect risk, the following rules apply: When the actual value is greater than or equal to the warning threshold, the score is 1 - (corresponding value / corresponding warning threshold). If the score is less than 0.5, the value is 0.5. When the actual value is less than the warning threshold, a base score of 0.5 is used as the passing score, and additional performance bonuses are added. The final score is 0.5 + (warning threshold - corresponding value) / warning threshold × 0.5. If the score is greater than 1, the value is 1. The degradation degree warning threshold is a 30% decrease in intrinsic viscosity, and the defect risk warning threshold is 15%.

[0106] In the above-mentioned neutral indicator calculation rules, 0.5 is the base score for the pass line. Its core functions are: first, to be compatible with the minimum score requirement of the original rules and ensure that the judgment logic remains unchanged under the deterioration state; second, to unify the scoring logic of the neutral indicator with the positive and negative indicators, with the better the indicator, the higher the score; and third, to accurately quantify the safety margin of the indicator being better than the warning threshold through the excess bonus space of 0.5.

[0107] For core evaluation indicators such as mean current and viscosity, if the associated stability indicators exceed the corresponding high-quality threshold, the normalized score of the core indicators needs to be corrected downwards. The correction range is determined according to the degree of exceeding the standard of the stability indicator, with the maximum correction range for a single indicator not exceeding 0.2, and the lowest corrected score not lower than 0.5. Among them, the associated stability indicator for mean current and viscosity is the current variance, with the high-quality threshold for current variance being ≤0.8A², the qualified threshold being 0.8-1.2A², and the poor-quality threshold being >1.5A².

[0108] In this method, the normalized scores of core performance indicators need to simultaneously consider both the absolute value of the indicator meeting the target and the fluctuation of related stability indicators. This avoids misjudgments caused by a single mean indicator meeting the target but actual printing stability being extremely poor. The specific correction rules are as follows:

[0109] 1. Definition of core correlation: The mean current and viscosity are the core positive indicators of the melting performance of rPET strips, and the current variance is the correlation stability indicator between the two. The current variance reflects the viscosity uniformity and feeding resistance fluctuation of the rPET strip along its length. If the variance exceeds the standard, it means that even if the average viscosity / mean current meets the standard, there are still local viscosity changes, material breakage / stallization risks in the material, and the core indicator scores need to be corrected downward.

[0110] 2. Correction trigger conditions: When the current variance exceeds the excellent threshold of 0.8A², the viscosity and current mean scores are corrected; when the variance exceeds the qualified threshold of 1.2A², a forced correction is performed; when the variance reaches the poor threshold of 1.5A², the qualification for excellent grade is directly rejected.

[0111] 3. Correction range rules: Current variance 0.8-1.2A² (qualified range): correction range 0.1-0.2, core indicator score corrected down by no more than 0.2; Current variance 1.2-1.5A² (warning range): correction range 0.2, core indicator score corrected down to a minimum of 0.5;

[0112] 4. General Adaptation Rules: In addition to the correlation correction of current variance-viscosity, anomalies in filament length / collapse ratio can be simultaneously corrected for the neutral index scores of degradation degree and defect risk. The correction rules follow the above-mentioned range limits to ensure cross-validation of multi-source indicators.

[0113] Then, weights were assigned to each indicator (current 0.3, viscosity 0.2, material breakage risk 0.15, fiber drawing 0.075, collapse 0.075, degradation degree 0.1, defect risk 0.1), and the normalized scores of each indicator were weighted and summed to obtain the initial total score;

[0114] The weighting is centered on preventing material breakage. Current (0.3) receives the highest weight due to its real-time nature of advanced early warning and high-frequency sampling; melt viscosity (0.2), derived from current, serves as an intrinsic state bridge connecting electrical and visual signals; and material breakage risk (0.15) directly quantifies the probability of material breakage. The three factors combined total 0.65, ensuring that preventing material breakage plays an absolutely dominant role in the grade determination. Visual indicators such as wire length and collapse area ratio each account for 0.075, forming the second priority dimension for defect reduction; comprehensive evaluation indicators such as degradation degree and defect risk each account for 0.1, serving as a supplement to indirect derivation and improving the overall judgment. The overall weighting reflects the principles of "prioritizing intrinsic characteristics, prioritizing direct signals over indirect derivation, and prioritizing prevention over post-event remediation."

[0115] Further verify whether the key indicators, such as average current, wire length, and collapse, are all within the grade threshold range corresponding to the initial total score;

[0116] The method for obtaining the grade threshold of the average current is as follows: Each time the rPET strip batch is changed or every 2 hours of printing, the system automatically performs "empty wire feeding calibration" (feeding wire at 5mm / s for 3s to collect current, and retracting at 3mm / s for 2s to reduce waste). The average motor current is collected as the "baseline current" for that batch / stage, and the dynamic threshold is derived (high-quality threshold = baseline current + 0.3A, qualified threshold = baseline current ± 0.0A, poor-quality threshold = baseline current - 0.3A). This adapts to the differences in the basic viscosity of different recycled materials. The motor current is the key basis for determining the state grade (the material characteristics of the rPET strip determine the wire feeding resistance, which in turn affects the current required for the motor to maintain the set speed. The system infers the state grade through current and other data, rather than adjusting the current according to the grade). Specifically, for high-quality rPET strips (new bottle recycling, crystallinity 30%-35%): high melt viscosity (≥0.8Pa・s), high wire feeding resistance, and the motor output current reaches the high-quality threshold (baseline current + 0.3A); qualified rPET... Standard strips (regular recycling): medium melt viscosity (0.5-0.8 Pa·s), current stable within the acceptable threshold range (reference current ± 0.3 A); Inferior rPET strips (recycled multiple times, severely degraded): low melt viscosity (< 0.5 Pa·s), low wire feeding resistance, current drops to the inferior threshold (reference current - 0.3 A) and below, prone to wire pulling / collapse defects.

[0117] Finally, the initial score is adjusted according to the type of recycled bottle material, additive residue, and mechanical strength (individual adjustment range ±0.03, cumulative adjustment range not exceeding ±0.05): recycled material containing antioxidants adds 0.02-0.03, recycled material with high colorant / plasticizer residue subtracts 0.02-0.03; high-strength recycled material adds 0.01-0.02, low-strength recycled material subtracts 0.02-0.03; the final score is obtained, and the rPET grade is determined based on the final score.

[0118] Specifically, by differentiating normalized scoring rules for positive, negative, and neutral indicators, the problem of inconsistent dimensions and directions of multi-source heterogeneous indicators is solved, making the scoring more scientifically comparable. After obtaining the initial total score through weighted summation, the grade threshold is verified by key indicators such as average current, wire length, and collapse ratio, forming a two-dimensional verification matrix of quantitative scoring and key hard constraints, effectively eliminating the risk of misjudging the overall grade due to anomalies in a single indicator. Finally, the type of recycled bottle material is introduced to correct the score with material weights, enabling the grade determination to adapt to the inherent performance differences of recycled materials from different sources, improving the accuracy and robustness of state determination in scenarios involving the mixing of multiple batches and brands of recycled materials.

[0119] In this embodiment, preferably, in step S5, when the adjustment instructions output by the analytical model detect a conflict in the adjustment directions of different parameters, a priority determination mechanism is triggered; specifically, this includes:

[0120] Priority weight allocation and conflict determination: Quantitative weights are assigned to three types of objectives (0.5 for preventing material breakage, 0.3 for reducing defects, and 0.2 for printing efficiency). The system first analyzes the adjustment scheme output by the model. If there is a conflict between the adjustment directions of different parameters (e.g., speeding up can improve efficiency, but may increase the risk of material breakage; cooling can reduce defects, but may reduce efficiency), then the priority determination mechanism is triggered.

[0121] Material breakage prevention priority logic: Based on the "material breakage risk threshold" as the core decision-making basis—If the current material breakage risk is ≥50%, prioritize the implementation of material breakage prevention related adjustments: ① If the model outputs a "speed up" command, immediately reject it and change to maintain the current speed or reduce the speed by 5-10mm / s; ② If the servo motor current is ≤1.3A (characteristic of inferior material), prioritize increasing the extrusion speed by 1-2mm / s, while reducing the printing speed by 8-12mm / s to increase the material flow filling density and avoid material breakage; ③ When the material breakage risk is ≥80%, directly terminate parameter adjustment, immediately stop printing and issue an audible and visual warning, save the current printing progress, and prioritize ensuring that the parts are not scrapped due to material breakage.

[0122] Defect reduction as a secondary priority: If the risk of material breakage is less than 50% (the material breakage prevention target has been met), focus on optimizing defect risk: ① If the model outputs a "heat up and improve efficiency" instruction, but CNN quantization shows that the wire drawing length is 2-5mm (moderate defect), then reduce the temperature increase (from 3-5℃ to 1-2℃), while maintaining the printing speed or only slightly increasing it (2-3mm / s), balancing efficiency and defect control; ② If the collapse area accounts for 1%-3% (defect in acceptable condition), prioritize fine-tuning the printing temperature (±2℃) or cooling air velocity (±0.5m / s), rather than adjusting the speed to improve efficiency; ③ Prohibit any adjustments that may increase the risk of defects (e.g., in the case of inferior material, even if the model suggests increasing the speed, force a decrease in speed and a decrease in temperature).

[0123] Printing efficiency fallback logic: Efficiency-related adjustments are only fully executed when the risk of material shortage is <20% and the defect risk is <15% (both risks are at low levels): ① In high-quality material condition, the system fully increases the temperature (3-5℃) and speed (5-10mm / s) according to the model instructions; ② In qualified material condition, if the dual-modal data shows a stable state (current variance ≤0.5A², no obvious stringing / collapse), the printing speed can be slightly increased by 3-5mm / s. The increased speed should be ≤ the maximum printer speed threshold entered in step S1. If it exceeds the threshold, the system will execute according to the threshold and will not increase the speed further to maximize efficiency; ③ Efficiency adjustments are based on the premise of not exceeding the safety threshold of "preventing material shortage and reducing defects" (e.g., the risk of material shortage after speed increase should not exceed 20%, and the defect risk should not exceed 15%). Once the risk exceeds the standard, the parameters will be immediately reverted. When the poor quality state lasts for more than 5 seconds and there is no improvement after adjustment, a material change reminder is triggered. After the material change reminder is triggered, the system automatically pauses printing and saves the current layer printing coordinates; after the material change is completed, step 1 is automatically executed. After a blank wire feed calibration (5s), printing resumes after updating the reference current.

[0124] In this embodiment, as Figure 3 As shown, preferably, the model in step S4 includes:

[0125] The input layer receives rPET strip grade, current timing data, current statistical features, high-dimensional semantic features, and bottle recycling material identification information (extracting 16 types of statistical features such as mean, variance, peak value, and kurtosis within each window; bottle recycling material identification information: 1 = new bottle recycling material, 2 = regular recycling material, 3 = multiple recycling material, 4 = containing antioxidants, etc.); the input layer also receives strip size, ambient temperature, real-time printing parameters such as printing temperature, extrusion speed, cooling air velocity, etc., and part size parameters such as length, width, height, and number of printing layers).

[0126] The data preprocessing layer is used to perform Z-Score standardization and temporal alignment on the input features (feature standardization and temporal alignment solve the problems of inconsistent dimensions and mismatched temporal lengths in multi-source data; numerical features are standardized using Z-Score, and the mean (μ) and standard deviation (σ) are dynamically updated based on 100,000+ sets of training data; temporal data are uniformly interpolated to a 50Hz sampling frequency, aligned with the image acquisition frequency; forward padding + temporal interpolation is used to handle missing data).

[0127] The LSTM time series feature extraction layer uses a 3-layer stacked bidirectional LSTM network (each hidden layer has a dimension of 256) to extract the dynamic dependency features of current time series data fluctuations; the time step is 64 (corresponding to a 3.2s time series window), the forget gate threshold is 0.7, the dropout rate is 0.2, and the output is a 256-dimensional time series feature vector.

[0128] The Transformer global association layer adopts a 6-layer encoder architecture, with each layer containing an 8-head multi-head self-attention mechanism and a feedforward neural network. It takes the LSTM temporal feature vector, high-dimensional semantic features and bottle recycling material identification information as input, calculates the global attention weights between multi-source features, and outputs a multi-dimensional global association feature vector.

[0129] The large component adaptation layer generates a 64-dimensional adaptation vector based on the component size parameters. It assigns a weight coefficient of 1.2-1.5 times to four core features: total component printing time, single-layer printing area, interlayer cooling time, and long-sequence current variance threshold. This vector is then fused with the multi-dimensional globally associated feature vector. The "continuous printing stability" related features include four core features: total component printing time, single-layer printing area, interlayer cooling time, and long-sequence current variance threshold.

[0130] The state level and recycled material adaptation layer selects a pre-trained multi-dimensional exclusive adaptation vector based on the rPET strip level and bottle recycled material identification information, and multiplies it element-wise with the fusion features from the large component adaptation layer to enhance the differences in parameter adjustment logic under different scenarios and obtain enhanced features.

[0131] The fully connected output layer uses a two-layer fully connected network to map the enhanced features into process parameter adjustment schemes and early warning signals, and corrects the printing temperature for ambient temperature.

[0132] The printing temperature is constrained between 225-255℃, and the temperature parameters are corrected according to the ambient temperature: the ambient temperature correction baseline is 25℃. When the ambient temperature is below 25℃, the temperature is increased by 0.5℃ for every 1℃ decrease (cumulative 1-2℃); when the ambient temperature is above 30℃, the temperature is decreased by 0.5℃ for every 1℃ increase (cumulative 1-2℃); when the ambient temperature is above 30℃, the printing temperature is decreased by 1-2℃. After the ambient temperature correction, the printing temperature must not be lower than 225℃ or higher than 255℃. If the calculated temperature exceeds the range, a safety boundary value is used (e.g., if the ambient temperature is 20℃ and the initial temperature is 225℃, the corrected temperature will still be 225℃). After the ambient temperature correction and the model adjustment range are combined, the printing temperature must meet the 225-255℃ requirement. If the combined temperature exceeds the upper limit, the model adjustment range is applied first, and then the ambient temperature correction is applied to the boundary value (e.g., if the model temperature rises by 5℃ to 257℃, it is first lowered to 255℃, and then the ambient temperature correction is applied).

[0133] The core reasons for using the LSTM-Transformer architecture are: ① To achieve precise quantitative adjustment: the state level only provides direction, and the model needs to combine current fluctuation trends, image feature details, and characteristics of recycled bottle material (such as the need for less temperature increase due to antioxidant residue) to output specific adjustment values ​​(such as a temperature increase of 3℃ instead of 5℃); ② To capture dynamic change patterns: the LSTM layer can extract the temporal dependence of current and defects (such as a continuous increase in current indicating a gradual change in viscosity, requiring a gradual increase in speed), avoiding adjustment lag caused by static level judgment; ③ To adapt to complex scenario coupling: when multiple factors conflict (such as a good state but poor heat resistance of recycled bottle material), the Transformer layer can mine global correlations and output a balanced solution (such as appropriately reducing the temperature increase while fine-tuning the extrusion speed); ④ To support model iterative optimization: multi-source input data provides rich samples for incremental learning, strengthening the model's adaptability to different recycled bottle materials and different part sizes.

[0134] In this embodiment, step S4 optimizes the parameter adaptation logic based on the differences in melting characteristics of different recycled bottle materials. When printing large parts, it is necessary to mix bottle rPET strips from different brands / batches / recycling cycles. The differences in their materials (additive residues, crystallinity, degree of degradation) will cause fluctuations in the wire feeding resistance, requiring targeted optimization of the servo motor control strategy.

[0135] Torque Limiting and Anti-Stall Protection: Crystallinity Acquisition: ① Offline Preset: When entering the bottle recycling material identification information, the system simultaneously associates the basic crystallinity data of this type of material (e.g., 32% for new bottles, 28% for regular recycling, 23% for multiple recycling, 26% for containing antioxidants), derived from material reports provided by recycling material suppliers or previous experimental calibration; ② Online Correction: Combining the collected servo motor current characteristics (the average current is positively correlated with crystallinity), the system dynamically corrects the crystallinity value through a "current-crystallinity" mapping model (based on fitting of 100,000+ sets of training data). For example, when the average current is 0.1A higher than the benchmark value of the same type of material, the crystallinity is increased by 2%-3%; ③ Automatic Calculation Mode: If the user does not enter the crystallinity, the system automatically matches the default crystallinity (32% for new bottles, 28% for regular recycling, 23% for multiple recycling) based on "bottle recycling material identification (e.g., 1 = new bottle) + initial current characteristics". Online correction is still based on dynamic fine-tuning of current characteristics, taking into account the needs of both professional and ordinary users.

[0136] Using 28% as the dividing line, materials with a crystallinity ≥28% are considered high-crystallinity recycled materials (such as recycled fresh bottle material), while materials with a crystallinity <28% are considered low-crystallinity recycled materials (such as bottle material that has been recycled multiple times or is severely degraded). Rated torque range: The rated torque of the servo motor is set at 1.27N・m-2.39N (compatible with 400W-750W motors), therefore the maximum output torque range is 1.52N・m-2.87N・m; the torque detection frequency is specified as 10Hz for real-time monitoring of load status; Rated current range: compatible with 400W-750W motors, the rated current of the servo motor is 3.5A-6.5A. Therefore, for high-crystallinity recycled materials (crystallinity ≥28%, high wire feeding resistance), the maximum output torque of the motor is set to 120% of the rated value (1.52N・m-2.87N・m), and the corresponding current does not exceed 120% of the rated current (4.2A-7.8A) to avoid stalling and burnout. For low-crystallinity / degraded materials (crystallinity <28%, prone to slippage), the wire feeding wheel pressure is reduced (from 8N ​​to 5N), and the motor adopts a "speed priority" control mode (prioritizing stable wire feeding speed, widening the current fluctuation tolerance range to ±0.1A, and dynamically adjusting PID parameters to improve speed response), allowing small current fluctuations to maintain continuous wire feeding. A "current lower limit trigger mechanism" is added to the "speed priority" mode: when the current is lower than "reference current -0.4A" for 0.3s, it immediately switches to "anti-material interruption mode", automatically reducing the speed by 5-8mm / s and increasing the extrusion speed by 1-2mm / s, while issuing a material interruption risk warning.

[0137] Differentiated parameter control rules should be established for different types of additive residues:

[0138] Antioxidant residue compatibility: Antioxidants can delay the thermal oxidative degradation of PET. For recycled materials with antioxidant residue ≥0.1%, the upper limit of printing temperature is increased from 255℃ to 260℃, the maximum single temperature rise is increased from 5℃ to 7℃, and the upper limit of printing speed is increased by 10mm / s, making full use of its thermal stability advantage to improve printing efficiency.

[0139] Colorant Residue Adaptation: Inorganic colorants increase the melt viscosity of the material. For recycled materials with 5%-10% colorant residue, the extrusion pressure is increased by 0.1MPa; for recycled materials with >10% colorant residue, the extrusion pressure is increased by 0.2MPa, while the upper limit of printing speed is reduced by 5mm / s to avoid poor filament feeding due to increased viscosity.

[0140] Plasticizer Residue Compatibility: Plasticizers lower the glass transition temperature of PET. For recycled materials with plasticizer residue ≥3%, the upper limit of printing temperature is reduced from 255℃ to 252℃. Printing at temperatures above 250℃ for extended periods is prohibited to prevent collapse and deformation caused by excessive softening of the material.

[0141] Mechanical strength adaptation strategy: Optimize filament feeding and printing control logic based on the tensile strength grade of the recycled material.

[0142] Medium-strength recycled material (30-50MPa): Printing speed limit reduced from 80mm / s to 70mm / s, filament feed roller pressure remains unchanged at 5N, and material breakage risk warning threshold remains at 20%;

[0143] Low-strength recycled material (<30MPa): The filament feeding speed is reduced by 15%, the extrusion pressure is increased by 0.2MPa to reduce the tensile stress of the material flow, the upper limit of printing speed is reduced to 60mm / s, and the material breakage risk warning threshold is reduced from 20% to 15% to trigger the anti-material breakage protection in advance.

[0144] If the current variance is greater than 1.0A² during printing and is accompanied by frequent small current fluctuations, it is determined that the material has uneven mechanical strength, and the printing speed is automatically reduced by 10mm / s until the current fluctuations stabilize.

[0145] The solutions involved in the above embodiments will be described in detail below with reference to specific examples.

[0146] An adaptive adjustment method for 3D printing parameters of manufactured parts based on dual-modal monitoring is described in the following steps.

[0147] Step S1: Data Acquisition.

[0148] The following data were collected synchronously: servo motor current data (sampling frequency 100Hz). After segmented empty wire feeding calibration (5mm / s wire feeding for 3s, 3mm / s retraction for 2s), the reference current for this batch of material was 1.4A. Therefore, the high-quality threshold was 1.7A, the qualified threshold was 1.4A, and the poor-quality threshold was 1.1A. The mean and variance of the current were recorded. Product surface condition data (camera frame rate 30fps) was collected, capturing features such as wire length and collapse area. Basic parameters of recycled rPET strips: strip size 4mm×8mm (mixed from recycled material from new bottles of brand A and recycled material from secondary recycling of brand B). Initial printing parameters: printing temperature 235℃, extrusion speed 15mm / s, printing speed 60mm / s. Dimensions: 600mm×600mm×350mm (large structural component); Ambient temperature is 28℃ (accuracy ±0.5℃) collected by an external digital temperature and humidity sensor; Maximum printer parameters: Temperature 250℃, Speed ​​75mm / s.

[0149] Step S2: Feature extraction and association analysis.

[0150] Data processing: Feature extraction; current variance calculated using a 0.5s sliding window = 0.9A² (exceeding threshold 0.8A²), current mean = 1.7A; after compensation for strip size error (strip size 4mm×8mm, compensation coefficient = 1.0, the corrected current is still 1.7A, corrected current variance = 0.9A²×1.0 = 0.9A²); quantification of collapse area percentage using a CNN model = 4% (exceeding the acceptable threshold 3%, not exceeding the severe threshold 5%), wire length = 1.8mm (≤ the excellent threshold 2mm).

[0151] Step S3, rPET strip status level determination.

[0152] The dynamic thresholds derived from pre-calibration (1.7A for high-quality, 1.4A for qualified, and 1.1A for poor-quality) are used as the current judgment standard. Combined with indicators such as viscosity (derived from current data as 0.85 Pa·s) and material breakage risk (15%), the judgment is completed using a "dual-dimensional threshold matrix + material weight correction" mechanism. The weights of the seven core indicators are calculated based on the normalized scoring rules, as follows: Current: High-quality threshold 1.7A, actual value 1.7A, normalized score = 1.7A / 1.7A = 1.0 (positive indicator rule), weight score = 0.3 × 1.0 = 0.3; Viscosity: derived from current data as 0.85 Pa·s (≥0.8 Pa·s high-quality threshold), normalized score = 0.85 / 0.8 ≈ 1.06 (capped at 1.0, positive indicator rule), weight score = 0.2 × 0.9 = 0.18 (due to current variance of 0.9A²). For materials exceeding the 0.8A² threshold, the viscosity score is slightly adjusted from 1.0 to 0.9 to balance the impact of fluctuations; For material breakage risk: 15% < 20% of the acceptable threshold, the normalized score = (20% - 15%) / 20% + 0.8 = 0.95 (according to the normalization rule for the reverse indicator scenario, the score = 0.8 + (20% - 15%) / 20% × 0.2 = 0.85, and the weighted score = 0.15 × 0.85 = 0.1275; For stringing: the acceptable threshold is 2 mm, the actual value is 1.8 mm, the normalized score = (2 mm - 1.8 mm) / 2 mm + 0.8 = 0.9 (according to the reverse indicator rule, with the base score of 0.8 added), and the weighted score = 0.075 × 0.9 = 0.0675; For collapse: the acceptable threshold is 3%, the actual value is 4%, and the normalized score = 1 - (4% - 3%) / (5% - 3%) = 0.5 (reverse indicator rule), weight score = 0.075 × 0.5 = 0.0375; Degradation degree: Material A contains antioxidants, Material B is recycled twice, the overall degradation degree is moderate, normalized score = 0.9 (neutral indicator rule), weight score = 0.1 × 0.9 = 0.09; Defect risk: Combining the stringing / collapse data, the defect risk is 12% < 15% warning threshold. According to the neutral indicator high-quality scenario normalization rule (0.5 qualified line base score + excess bonus points), score = 0.5 + (15% - 12%) / 15% × 0.5 = 0.6, weight score = 0.1 × 0.6 = 0.06; Preliminary weighted total score = 0.3 + 0.18 + 0.1275 + 0.0675 + 0.0375 + 0.09 + 0.06 = 0.8625; Dual-dimensional threshold matrix verification: Current meets standard (high quality), wire drawing meets standard (high quality), collapse does not reach inferior quality, other indicators have no deterioration, meeting the basic conditions for high quality level; Material weight correction: Material A (new bottle + antioxidant, antioxidant residue 0.15%, high strength) + 0.03 (type correction + 0).02 + Antioxidant Correction + 0.01), Material B (secondary recycling, medium strength, no obvious additive residue) - 0.01 (Type Correction - 0.01), Overall Correction + 0.02, Final Total Score = 0.8625 + 0.02 = 0.8825, the current rPET strip is ultimately determined to be in excellent condition (total score ≥ 0.8 is the excellent threshold, current reaches the excellent threshold of 1.7A, wire length ≤ 2mm, material breakage risk 15%), and a status report is output.

[0153] Step S4: Model generates adjustment scheme.

[0154] The operation is based on the LSTM-Transformer architecture: The LSTM layer takes three consecutive time windows (192 sampling points) as input to extract the dynamic dependence features of current fluctuations; the Transformer layer mines the global correlation between "current-collapse-recycled material characteristics"; combined with the new bottle material of brand A (containing antioxidants, pre-trained adaptation vector V = dedicated sub-model output) and the ambient temperature of 28℃ (no correction required), the optimal adjustment scheme is output: the printing temperature is increased from 235℃ to 239℃ (a 4℃ increase, not exceeding the 250℃ upper limit), and the model temperature increase plus the ambient temperature is 239℃ (not exceeding the 255℃ upper limit), no correction to the boundary value is required; the printing speed is increased from 60mm / s to 68mm / s (an 8mm / s increase, not exceeding the 75mm / s upper limit); according to the "speed-current" correlation formula, the initial speed of 60mm / s corresponds to a high-quality current range of 1.4-1.7A, and after increasing the speed to 68mm / s, the adjusted high-quality current range = The current ranges from 1.4×(68 / 60)≈1.59A to 1.7×(68 / 60)≈1.91A. The actual current is between 1.7-1.8A, which is within this range, indicating effective adaptation. The extrusion speed is maintained at 15mm / s, and the extrusion speed adaptation logic is optimized for the viscosity difference between A and B materials (dynamically adjusting the extrusion pressure ±0.1MPa).

[0155] Step S5: Parameter execution and priority control.

[0156] The adjustment plan was sent to the printer in real time, and parameter adjustments were executed according to the priority of "preventing material breakage > reducing defects > printing efficiency". The printing temperature was ensured to be within the range of 225-255℃. The adjustment plan was analyzed to be conflict-free (the risk of material breakage is still 15% < 20% after increasing the speed by 8mm / s), and the "printing efficiency fallback logic" was followed. The current was monitored in real time to be stable at 1.7-1.8A, the printing temperature at 239℃, and the motor torque at 1.6N・m (within the safe range of 1.52-2.87N・m). The stall protection was not triggered when using new bottle material of brand A (high crystallinity 32%), and the feed roller pressure of recycled material of brand B (low crystallinity 23%) was maintained at 5N. No warnings were triggered (emergency warning thresholds: material breakage risk ≥ 80%, temperature above 250℃ / below 225℃), ensuring that the parameters operated stably within the safe range.

[0157] Step S6: Feedback and closed-loop iteration.

[0158] Real-time re-inspection showed that within 0.5 seconds after adjustment, the current variance decreased to 0.4A², the wire drawing length decreased to 1.2mm, and the material breakage risk remained at 12%, meeting the effective adjustment standard. After printing, the product defect rate was 12% (automatically quantified by the CNN model, with defects ≥0.5mm² included), a 65% reduction compared to before adjustment, and material costs were reduced by 52% compared to new materials. The data on "adjustment parameters - defect changes - adaptation effect of different recycled materials" was entered into the database, recording the core data: recycled material type (A new bottle + B secondary), adjustment parameters (heating 4℃, speed increase 8mm / s, ambient temperature 28℃), and defect rate of 12% (defect rate = total area of ​​defects ≥0.5mm² / total surface area of ​​the part × 100%, consistent with the defect quantification logic). This data was included in the effective samples of the day (35 groups in total) and participated in the incremental learning of the AI ​​model to dynamically strengthen the adaptation weight of mixed A / B materials. This sample will be included in the historical test set, and will be updated after 30... The model's performance was enhanced by the participation of the test set update team. The current offline verification shows that the model's state determination accuracy is 95% and the effective adjustment rate is 88%, which meets the performance requirements and completes the closed-loop iteration.

[0159] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for adaptive adjustment of 3D printing parameters for parts based on dual-modal monitoring, characterized in that, include: During the 3D printing process, the current timing data of the servo motor is collected, and images of the forming area of ​​the part are captured. A sliding window is used to extract current statistical features from current time series data, including at least the current mean and variance; based on a pre-calibrated mapping relationship between current, viscosity and material breakage risk, viscosity and material breakage risk are determined according to the current mean. Based on the high-dimensional semantic features of the forming area of ​​the part, the drawing length and the proportion of the collapsed area are extracted from the image of the forming area of ​​the part. Based on the pre-calibrated mapping relationship between drawing length / collapse-degradation degree-defect risk, the degradation degree and defect risk are determined according to the obtained drawing length and the proportion of collapsed area. The rPET strip grade is determined by considering factors such as average current, viscosity, material breakage risk, fiber length, collapse area ratio, degradation degree, and defect risk. The rPET strip level, current time series data, current statistical features, high-dimensional semantic features, and bottle recycled material identification information are input into the pre-trained model. The model outputs fine adjustment parameters under the adjustment direction determined by the rPET strip level. Adjust the 3D printing parameters of the part based on fine-tuning parameters.

2. The adaptive adjustment method for 3D printing parameters of a part based on dual-modal monitoring according to claim 1, characterized in that, The mean and variance of the current are calculated based on a cutting size error compensation strategy; specifically, the cutting size error compensation strategy is as follows: Calculate the ratio of the standard size to the actual size of the rPET strip; If the ratio is within the preset compensation coefficient range, the ratio shall be used as the compensation coefficient. If the ratio is less than the lower limit of the compensation coefficient interval, the lower limit of the compensation coefficient interval shall be used as the compensation coefficient. If the ratio is greater than the upper limit of the compensation coefficient range, the upper limit of the compensation coefficient range shall be used as the compensation coefficient. Multiply the compensation coefficient by the statistical mean current and variance current respectively to obtain the compensated mean current and variance current.

3. The adaptive adjustment method for 3D printing parameters of a part based on dual-modal monitoring according to claim 1, characterized in that, Based on the image of the forming area of ​​the part, high-dimensional semantic features of the forming area are extracted, and the wire drawing length and the proportion of collapsed area are quantified. A CNN model is used for this purpose. The CNN model includes: The preprocessing module is used to scale the image of the forming area of ​​the part to a preset size and perform normalization and data augmentation; The MobileNetV3-small encoder contains multiple depthwise separable convolutional modules for extracting multi-scale feature maps from preprocessed images; The CBAM attention module is used to sequentially apply channel attention weighting and spatial attention weighting to the multi-scale feature map to obtain an enhanced feature map as a high-dimensional semantic feature. The U-Net decoder is used to perform multiple upsampling on the enhanced feature map and fuse the feature map of the corresponding scale in the MobileNetV3-small encoder with the upsampled feature map through skip connections; The dual-branch output layer includes a parallel segmentation branch and a quantization branch. The segmentation branch outputs a 3-channel segmentation map representing the background, wire drawing, and collapse. The quantization branch outputs a 2-channel quantization map representing the wire drawing length and the proportion of the collapse area.

4. The adaptive adjustment method for 3D printing parameters of a part based on dual-modal monitoring according to claim 1 or 2, characterized in that, The rPET strip grade is determined by considering factors such as average current, viscosity, breakage risk, fiber length, collapse area percentage, degradation degree, and defect risk; specifically including: First, calculate the normalized score for each indicator; Then, weights are assigned to each indicator, and the normalized scores of each indicator are summed in weights to obtain the initial total score. Further verify whether the key indicators, such as average current, wire length, and collapse, are all within the grade threshold range corresponding to the initial total score; Finally, the initial score is adjusted according to the type of recycled bottle material, additive residue, and mechanical strength to obtain the final score, and the rPET grade is determined based on the final score.

5. The adaptive adjustment method for 3D printing parameters of a part based on dual-modal monitoring according to claim 4, characterized in that, Calculate the normalized score for each indicator, specifically as follows: For the positive indicators of current and viscosity, the score = corresponding value / corresponding quality threshold. If the score is greater than 1, the value is 1. For the reverse indicators of material breakage risk, wire drawing length, and collapse ratio, when the actual value is greater than or equal to the qualified threshold, the score is calculated as follows: 1 - (corresponding value - corresponding qualified threshold) / (corresponding substandard threshold - corresponding qualified threshold). If the score is less than 0, the value is 0. When the actual value is less than the qualified threshold, a base score of 0.8 is used as the qualified line, and additional performance bonuses are added. The final score is calculated as follows: 0.8 + (qualified threshold - corresponding value) / qualified threshold × 0.

2. If the score is greater than 1, the value is 1. For neutral index degradation degree and defect risk, the following rules apply: When the actual value ≥ the warning threshold, the score = 1 - (corresponding value / corresponding warning threshold). If the score is less than 0.5, the value is 0.

5. When the actual value < the warning threshold, a base score of 0.5 is used as the passing score, and excess performance bonus points are added. The final score = 0.5 + (warning threshold - corresponding value) / warning threshold × 0.

5. If the score is greater than 1, the value is 1. The degradation degree warning threshold is a 30% decrease in intrinsic viscosity, and the defect risk warning threshold is 15%. For the core judgment indicators, mean current and viscosity, if the related stability indicator, variance current, exceeds the corresponding quality threshold, the normalized score of the core judgment indicator will be corrected downward. The correction range is determined according to the degree of exceeding the standard of the stability indicator, with the maximum correction range for a single indicator not exceeding 0.2 and the lowest score after correction not lower than 0.

5.

6. The adaptive adjustment method for 3D printing parameters of a part based on dual-modal monitoring according to claim 1, characterized in that, The adjustment direction determined by the rPET strip grade is as follows: If the rPET strip is rated as excellent, adjust the printing temperature and speed accordingly. If the rPET strip grade is qualified, the adjustment direction is to maintain or fine-tune the printing temperature and printing speed; If the rPET strip is of poor quality, adjust the printing temperature and speed to reduce them, or trigger a material change reminder.

7. The adaptive adjustment method for 3D printing parameters of a part based on dual-modal monitoring according to claim 6, characterized in that, The 3D printing parameters of the part are adjusted based on fine-tuning parameters and are configured to be executed in the order of priority: preventing material breakage, reducing defects, and printing efficiency. When the adjustment instructions output by the analytical model find that there is a conflict between the adjustment directions of different parameters, the priority determination mechanism is triggered.

8. The adaptive adjustment method for 3D printing parameters of a part based on dual-modal monitoring according to claim 7, characterized in that, When the adjustment instructions output by the analytical model detect conflicting adjustment directions for different parameters, a priority determination mechanism is triggered; specifically, this includes: If the risk of material shortage is ≥50%, then speed increase will be rejected and speed will be reduced. When the servo current is ≤ current quality threshold, the extrusion speed will be increased and the printing speed will be reduced. When the risk of material shortage is ≥80%, printing will be stopped directly, an alarm will be triggered and the progress will be saved. If the risk of material breakage is less than 50%, then when the wire drawing length is a moderate defect, limit the temperature rise and maintain or slightly increase the printing speed. When the collapse area ratio is in a qualified state, only fine adjust the printing temperature or cooling air speed, and prohibit any adjustment that increases the risk of defects. Full heating and speed increase will be implemented only when the risk of material shortage is less than 20% and the risk of defect is less than 15%. Additional speed increase will be implemented when the material is qualified, the current variance is ≤0.5A² and there are no obvious defects, but the speed will not exceed the preset maximum threshold. If the risk exceeds the limit, the parameters will be adjusted back. If the rPET strip is in a poor quality state for more than 5 seconds and there is no improvement after adjustment, a material change reminder is triggered. The system will automatically pause and save the current layer coordinates. After the material change, an empty wire feed calibration will be performed and the reference current will be updated before printing resumes.

9. The adaptive adjustment method for 3D printing parameters of a part based on dual-modal monitoring according to claim 1, characterized in that, The model includes: The input layer receives rPET strip level, current timing data, current statistical features, high-dimensional semantic features, and bottle recycling material identification information. The data preprocessing layer is used to perform Z-Score normalization and temporal alignment on the input features; The LSTM time series feature extraction layer uses a 3-layer stacked bidirectional LSTM network to extract the dynamic dependency features of current time series data fluctuations. The Transformer global association layer adopts a 6-layer encoder architecture, with each layer containing an 8-head multi-head self-attention mechanism and a feedforward neural network. It takes the LSTM temporal feature vector, high-dimensional semantic features and bottle recycling material identification information as input, calculates the global attention weights between multi-source features, and outputs a multi-dimensional global association feature vector. A large component adaptation layer generates a multi-dimensional adaptation vector based on the component size parameters, assigns weight coefficients to the total printing time of the component, single-layer printing area, interlayer cooling time, and long-sequence current variance threshold, and fuses it with the multi-dimensional global correlation feature vector. The status level and recycled material adaptation layer selects a pre-trained multi-dimensional exclusive adaptation vector based on the rPET strip level and bottle recycled material identification information, and multiplies it element-wise with the fusion features from the large component adaptation layer to obtain the enhanced features; The fully connected output layer uses a two-layer fully connected network to map the enhanced features into process parameter adjustment schemes and early warning signals, and corrects the printing temperature for ambient temperature.

10. The adaptive adjustment method for 3D printing parameters of a part based on dual-modal monitoring according to claim 1, characterized in that, The bottle body recycled material identification information includes the crystallinity of the recycled material, the type and content of residual additives, and the mechanical strength grade; The crystallinity is obtained through offline preset, online correction, and automatic calculation. Offline preset associates the crystallinity range with the recycled material identifier. Online correction dynamically corrects the crystallinity based on the servo motor current through the current-crystallinity mapping relationship. Automatic calculation matches the default value with the identifier and initial current and makes online fine adjustments when no crystallinity is entered. With 28% as the boundary, crystallinity ≥28% is considered high crystallinity recycled material, and <28% is considered low crystallinity recycled material. For high crystallinity recycled material, the maximum torque and current of the motor are limited to no more than 120% of the rated value. For low crystallinity recycled material, the pressure of the feeding wheel is reduced to the preset value and a speed priority control mode is adopted. When the current is continuously lower than the difference between the reference current and the preset deviation value for more than a preset time, the anti-material interruption mode is triggered, automatically reducing the speed, increasing the extrusion speed, and issuing a warning. Additive Residue: Offline entry of antioxidant, colorant, and plasticizer residue levels; for recycled materials containing antioxidant residue, the upper limit of printing temperature is increased by 5℃, and the maximum allowable temperature rise is increased by 2℃; for recycled materials with colorant residue ≥0.5%, the extrusion pressure is increased by 0.1-0.2MPa, while the upper limit of printing speed is reduced by 5mm / s; for recycled materials with plasticizer residue ≥3%, the upper limit of printing temperature is reduced by 3℃, and printing above 250℃ is prohibited. Mechanical strength: Enter the tensile strength grade of the recycled material offline; For medium-strength recycled materials, the maximum printing speed is reduced by 10 mm / s; For low-strength recycled materials, the feeding speed is reduced by 15%, the extrusion pressure is increased by 0.2 MPa, and any speed-up operation is prohibited. The material breakage risk warning threshold is lowered from 20% to 15%.

Images