A wheel hub coating off-line time prediction method

Abstract

The application discloses a kind of wheel hub coating off-line time prediction method, to solve the problems of poor adaptability of existing technology to special-shaped wheel hub, insufficient response to process fluctuation, low prediction accuracy.The core scheme includes: visual identification and partition, sub-regional dynamic time consumption prediction, adaptive disturbance compensation and multi-modal fusion iterative prediction.By machine vision, the coating sub-region is divided and the structural features are identified.Based on the structural features, the sub-regional coating time consumption is dynamically corrected.According to the coating type, the dedicated disturbance compensation model is called to calculate the compensation time consumption.Fusion of multi-dimensional data and iterative correction according to fixed cycle, output the off-line time prediction value.The application is suitable for various wheel hub structures and coating types, with high prediction accuracy and strong real-time responsiveness, which can improve the efficiency and scheduling accuracy of production line.

Description

Technical Field

[0001] This invention relates to the field of wheel hub coating production scheduling technology, specifically to a method, system, and equipment for predicting wheel hub coating off-line time. Background Technology

[0002] With the increasing personalization of the automotive industry, the variety of wheel products has surged, encompassing multiple sizes and irregular shapes. The wheel painting process has shifted from "mass production of single specifications" to "mixed-line painting of multiple specifications and irregular shapes of wheels." The accuracy of predicting the coating off-line time directly impacts production line scheduling, subsequent curing processes, and delivery cycles; however, existing technologies have significant limitations.

[0003] 1. Irregular structure leads to large prediction deviation and lag in element preparation: Irregular wheel hubs need to be subdivided into multiple painting areas. Due to differences in masking and spraying angle, the time taken for each area is different. The existing DMS system uses "average time" to estimate and cannot provide feedback on the progress of each area, resulting in lag in the preparation of special fixtures and color powder, which leads to delays in process connection or rework.

[0004] II. Process fluctuations are not included in the forecast, and real-time deviations are difficult to correct: Fluctuations in parameters such as electrostatic voltage and fluidizing gas pressure directly affect coating time. Existing forecasts are based only on static parameters and do not incorporate real-time fluctuation data. The actual deviation from the forecast often exceeds 15%, affecting the accuracy of scheduling.

[0005] 3. Difficulty in accurately predicting the time required for mixing production lines and causing gaps in scheduling: When mixing multiple specifications, it is necessary to adjust the spray gun trajectory and change the powder. The time required for changing production lines varies greatly. Existing technology relies on manual experience for estimation and has not established a scientific model, which can easily lead to equipment idleness or coating interruption.

[0006] IV. Lack of partitioning and parameter association logic makes it difficult to trace abnormal time consumption: The time consumption of each area of ​​the irregular wheel hub is strongly correlated with the exclusive process parameters. The existing system only records the total time consumption deviation and cannot associate abnormal regional parameters, making it difficult to locate the cause of delay and optimize the prediction model.

[0007] Therefore, there is an urgent need for a method to predict the off-line time of a multi-specification, non-standard wheel hub painting production line, in order to solve the above-mentioned technical pain points. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a method, system, and equipment for predicting wheel hub painting off-line time. The core of this invention lies in achieving high-precision prediction of off-line time through visual partitioning dynamic correction, coating type adaptive disturbance compensation, and multi-modal fusion iteration collaborative design. This allows for adaptation to various specifications of irregularly shaped wheels and different coating types, thereby improving production line collaboration efficiency and real-time response capabilities.

[0009] The core of the technical solution of this invention lies in constructing a full-link prediction system of "visual perception-regional correction-adaptive compensation-multimodal iteration". The generalized solution includes the following steps:

[0010] S1. Visual Recognition and Segmentation: Acquire wheel hub images, identify their structural features through machine vision, and divide the painted area into multiple sub-regions;

[0011] S2. Regional Dynamic Time Prediction: Based on the structural characteristics of each sub-region and combined with the spraying process parameters, the coating time of each sub-region is predicted, and dynamic correction is performed based on real-time structural characteristics.

[0012] S3. Adaptive Disturbance Compensation: Real-time acquisition of coating process disturbance parameters, and calculation of disturbance compensation time based on the current wheel hub coating type by calling the corresponding disturbance compensation model;

[0013] S4. Multimodal fusion iterative prediction: The painting time of each sub-region after dynamic correction and the disturbance compensation time are fused together, and combined with historical production data, iterative calculation and correction are performed at fixed intervals to output the final offline time prediction value.

[0014] Preferably, the present invention can further improve prediction accuracy and practicality through the following refinement schemes:

[0015] I. Detailed Implementation of Visual Recognition and Regionalization: An industrial camera with a resolution of 1920×1080 and a frame rate of 25fps is used to acquire RGB color images of the wheel hub. Combined with order information from the Manufacturing Execution System (MES) (including the number of spokes, wheel hub specifications, etc.), the images are preprocessed using weighted average grayscale conversion, 5×5 convolution kernel Gaussian filtering, and Canny operator edge detection. Then, Hough transform is used to detect the number of spokes (removing interference lines <50mm), and the U-Net semantic segmentation model is used to label the spoke regions and calculate the occlusion area. Simultaneously, key dimensions of the rim and flange are detected. If the discrepancy between the identified spoke count and the MES order is greater than 2, a second acquisition of 3 images is triggered, and the majority result is taken to ensure that the structural parameter accuracy is ≥99%. Finally, the structural feature parameters of each region (accurate to 0.1mm / integer) and the boundary coordinates of each region are output. The grayscale conversion formula is as follows: The standard deviation of the Gaussian filter is 1.2, the Canny operator threshold is low = 50 and high = 150, and the camera calibration conversion factor is 1 pixel = 0.1 mm.

[0016] II. Detailed Implementation of Regional Dynamic Time Prediction: Based on the regional structure parameters output in step S1 (number of spokes N, occlusion area) outer diameter of wheel rim (etc.), combined with spraying process parameters (target coating thickness) Spray gun movement speed etc.), spray gun configuration parameters (number of spray guns M, effective spray width, etc.). The actual spraying area of ​​the spoke area, rim area, and flange surface is calculated using specific formulas. The basic spraying time is then derived using the formula for the total mass of the paint. Based on the structural parameters and historical deviation data of the same specification wheel hubs over the past three months, dynamic correction factors (K1, K2, K3) are calculated for each area. The basic spraying time is corrected using these correction factors to obtain the final spraying time and total time for each area. The K1 and K2 coefficients are refitted monthly to ensure a correction accuracy of ≥98%.

[0017] III. Refined Implementation of Adaptive Disturbance Compensation: Real-time acquisition of raw data for electrostatic spraying voltage (50-90kV), powder fluidizing gas pressure (0.1-0.3MPa), and liquid coating viscosity (20-50mPa·s, specific to liquid coating) at a frequency of 10Hz. A moving average filter with 10 data point windows and outlier removal within ±30% of the baseline value are used to calculate the fluctuation value. , , The maximum fluctuation within 10 seconds is used to determine the disturbance level from 1 to 3 based on the fluctuation amplitude; according to the coating type configured in the MES order (1=powder, 2=liquid), a differentiated compensation logic is adopted, with the core formula being: ( For coating type weighting, powder ,liquid Powder coating focuses on voltage and pressure disturbances, while liquid coating additionally incorporates viscosity disturbances. The time for multi-dimensional disturbance compensation and the total disturbance compensation time are calculated using dedicated compensation coefficients.

[0018] IV. Detailed Implementation of Multimodal Fusion Iterative Prediction: Scores for four modalities—computer vision structure, process parameters, real-time perturbation, and historical feedback—are used, based on modal scores and priority coefficients (default). , , , When the sensor fails ) through formula Dynamically allocate weights and fuse the relevant time-consuming components of each mode to obtain a cross-modal fusion prediction value; update the cumulative disturbance compensation time in 10-second cycles, and combine the initial prediction value, total clock cycle correction, historical calibration correction, and prediction deviation from the previous cycle through an iterative formula. Correct the predicted off-line time and calculate the remaining painting time. A Level 3 disturbance warning is triggered. The formula for calculating the score for each modality is as follows:

[0019] Visual structural modality score ,constraint ;

[0020] Modal score of process parameters ,constraint ;

[0021] Real-time perturbation mode score Level 1 = 0.9, Level 2 = 0.8, Level 3 = 0.7;

[0022] Historical feedback modal score ,constraint ;

[0023] The cross-modal fusion prediction formula is: ,in , This represents the total time consumed by each region. , .

[0024] V. Detailed Implementation of Additional Optimization Modules:

[0025] 1. Changeover time prediction: based on the difference between the previous and current wheel hub diameters. Difference in coating thickness (From MES + visual recognition), paint type, number of spray guns The historical model change datasets of different specifications were used to calculate the cleaning time and parameter adjustment time respectively, and then the results were combined with historical correction coefficients. The total replacement time is obtained, and a replacement warning signal is output (0=normal, 1=pay attention, 2=abnormal).

[0026] 2. Production line cycle time linkage correction: Queue quantity is collected via sensors and MES (Manufacturing Execution System). Number of people queuing for curing ovens Idle time of subsequent assembly processes and production line benchmark cycle time The pretreatment correction amount, curing oven correction amount, and subsequent idle correction amount are calculated separately according to specific rules, and then summed to obtain the total cycle time correction amount. Simultaneously, it outputs the cycle stability coefficient. Among them, : The time is -12s, 3 < The time is 0s. At that time ; ; : The time is -18s, 2≤ The time is -6s. The time is 0s;

[0027] 3. Specification-Specific Historical Data Adaptation: Match the current wheel hub specification with historical samples, use a linear regression model to train and obtain historical correction coefficients, calculate the matching degree between the current working condition and historical samples (constraint ≥70%), and output specification-specific correction coefficients. (0.9-1.1), matching degree and historical calibration correction amount The matching degree is calculated using the following formula: Historical calibration correction amount .

[0028] The beneficial effects of this invention are as follows:

[0029] I. Significantly Improved Prediction Accuracy: By integrating multi-module and multi-modal data, through regional calculation, dynamic correction, multi-dimensional compensation and real-time iteration, the prediction error is <1% and the accuracy is ≥98.5%, which is 3-5 percentage points higher than that of a single module, effectively solving the prediction deviation problem in mixed-line production.

[0030] II. Significantly enhanced adaptability: Through regional calculation, historical adaptation by specification, and optimization of dynamic correction factors, it adapts to 10-24 inch different sizes and irregularly shaped wheel hubs such as multi-spoke / hollowed-out designs, while also being compatible with both powder and liquid coating types, reducing changeover errors.

[0031] III. Production Collaboration Optimization: Link the queuing status of upstream and downstream processes on the production line, output production scheduling suggestions, reduce process disconnection, and improve production line collaboration efficiency;

[0032] IV. Outstanding real-time response capability: Iterative correction is performed in 10-second cycles to accurately capture process disturbances, dynamically correct predicted values, adapt to real-time production changes, and improve the accuracy of prediction by more than 30% compared with static compensation. Attached Figure Description

[0033] Appendix Figure 1 This is a design diagram of the multimodal adaptive fusion architecture of the present invention. Detailed Implementation

[0034] As attached Figure 1 As shown,

[0035] I. Building a Data and Hardware Platform

[0036] 1. Data preparation: Organize 10,000 labeled wheel hub images (including spokes, rims, and flange surfaces) and 6 months of historical process time data from the production line to construct a historical dataset by specification;

[0037] 2. Hardware configuration: Calibrate the industrial camera, determine the pixel-to-actual size conversion factor (1 pixel = 0.1 mm), and establish a data transmission link between the sensors (voltage, pressure, viscosity) and the PLC;

[0038] 3. Determination of process parameters: Jointly with the process department, set the baseline values ​​for powder / liquid coating (such as powder voltage 70kV, fluidization pressure 0.2MPa);

[0039] 4. Development Environment Setup: Set up a Python (OpenCV / PyTorch) development environment and deploy a time-series database to store real-time and historical data.

[0040] II. Module Development and System Integration

[0041] 1. Development of the visual feature extraction module:

[0042] (1) Write code for grayscale conversion, Gaussian filtering, and Canny edge detection, and use Hough transform to count the number of spokes and remove short lines <50mm;

[0043] (2) Train the U-Net model to achieve hub region segmentation and calculate parameters such as occlusion area;

[0044] (3) Develop data verification logic. When the deviation between the number of identified spokes and the MES order is greater than 2, trigger secondary data collection and take the majority result.

[0045] 2. Development of the basic time consumption and correction module:

[0046] (1) Write the formula for calculating the basic time consumption of each region, input the structural parameters and process parameters, and output the basic time consumption;

[0047] (2) Develop the calculation logic for dynamic correction factors (K1 / K2 / K3) and add numerical constraints. , Output the time consumed after correction and the total time consumed.

[0048] 3. Development of disturbance compensation module:

[0049] (1) Write code for moving average filtering and outlier removal to process real-time disturbance data and output. , , and disturbance level;

[0050] (2) Develop a dedicated compensation time calculation logic for powder / liquid coating, substitute historical compensation coefficients, and output the total disturbance compensation time.

[0051] 4. Development of a module linking changeover time and cycle time:

[0052] (1) Write code to calculate the time consumed by the changeover cleaning and the time consumed by parameter adjustment, and output the total changeover time and early warning signal in combination with historical changeover data;

[0053] (2) Write code to calculate the cycle correction amount, and output the total cycle correction amount based on the queuing status of the upstream and downstream processes and the idle time of the subsequent process.

[0054] 5. Development of a specification-specific history adaptation module:

[0055] (1) Write code for historical data screening and linear regression model training, and calculate the sub-sizing correction coefficient and matching degree;

[0056] (2) Output historical calibration correction amount.

[0057] 6. Development of Multimodal Fusion and Real-time Iteration Module:

[0058] (1) Write modal score calculation functions (visual, process, disturbance, history) and calculate adaptive weights according to the formula;

[0059] (2) Integrate the time consumption components of each mode, write a cross-modal fusion prediction formula, superimpose a 10-second cycle iterative correction logic, and output the final offline time prediction value, remaining painting time and production scheduling suggestions.

[0060] III. System Testing and Optimization

[0061] The developed modules were integrated and deployed, and powder / liquid coating tests were conducted on 1,000 sets of different specifications of irregular-shaped wheels (including 10-24 inches, multi-spoke / hollow structure) to verify that the prediction accuracy was ≥98.5%, the iteration cycle was stable at 10 seconds, the prediction error of changeover time was <2%, and the response to process disturbances was timely, meeting the requirements of production line scheduling and delivery cycle. Based on the test results, the dynamic correction factor coefficient and modal priority coefficient were optimized monthly to continuously improve the prediction performance.

[0062] Example 1:

[0063] 1. Test conditions

[0064] • Wheel specifications: 10 inches (minimum size), number of spokes N=10 (minimum number of spokes), actual outer diameter of rim D r =400mm (minimum outer diameter of rim), rim height H r =80mm, flange outer diameter D f =200mm, flange face center hole diameter D hole =80mm, spoke shading area A occ =50cm², center distance between adjacent spokes S s =30mm;

[0065] • Coating type: Powder coating (W1=1.0);

[0066] • Process parameters: Target coating thickness T=50μm (minimum coating thickness), spray flow rate Q=100ml / min, powder coating density ρ=2.8g / cm³, coating adhesion efficiency η=0.8, number of spray guns M=4, electrostatic voltage reference value U reference=50kV (minimum voltage), fluidizing gas pressure reference value P reference=0.1MPa (minimum pressure).

[0067] • Change parameters: difference between previous and current wheel hub diameter ΔD = 10mm (minimum diameter difference), difference in coating thickness Δd = 10μm (minimum thickness difference), historical correction coefficient for each specification K specification = 0.9 (minimum correction coefficient).

[0068] • Iteration parameters: Iteration period = 5 seconds (minimum period), iteration correction coefficient α = 0.85 (minimum value);

[0069] • Production line status: Pre-processing queue quantity Q_pre-processing = 2, curing oven queue quantity Q_curing = 2, and subsequent assembly process idle time T_idle = 6 minutes.

[0070] 2. Calculation process

[0071] (1) Calculation of actual spraying area by region

[0072] •Spoke area: A s =[π×(400 / 2)²-π×(200 / 2)²-50×100]×1.2×10^(-2)=(125600-31400-5000)×0.012=89200×0.012=1070.4cm²;

[0073] •Rim area: A r =π×400×80×10^(-2)=10048×0.01=100.48cm²;

[0074] • Flange face: A p =π×[(200 / 2)²-(80 / 2)²]×10^(-2)=(31400-5024)×0.01=263.76cm².

[0075] (2) Calculation of basic spraying time

[0076] • Total mass of coating (taking the spoke area as an example): m = 1070.4 × 50 × 10^(-4) × 2.8 = 1070.4 × 0.005 × 2.8 = 14.9856 g;

[0077] • Basic time consumed in the spoke area: t sBaseline = (14.9856 × 1000) / (100 × 2.8 × 0.8 × 4) × 60 = 14985.6 / 896 × 60 ≈ 16.725 × 60 = 1003.5 s;

[0078] Similarly, calculate t for the rim area. r Reference point = 120.58s, flange face t p Baseline = 316.51s.

[0079] (3) Dynamic correction factor and time consumed after correction

[0080] • The spoke area K1 = 1.0 + 0.02 × 10 - 0.001 × 30 = 1.0 + 0.2 - 0.03 = 1.17 (located in the range [1.1, 1.5]).

[0081] • Rim area K2 = 0.95 + 0.0001 × (400 - 400) = 0.95 (within the range [0.92, 1.0]).

[0082] • Correction time: t s Final value = 1003.5 × 1.17 ≈ 1174.10 s, t r Final value = 120.58 × 0.95 ≈ 114.55 s, t p Final value = 316.51 × 1.02 ≈ 322.84s, total time ttotal = 1174.10 + 114.55 + 322.84 ≈ 1611.49s.

[0083] (4) Disturbance compensation time

[0084] • After real-time acquisition and processing of disturbance data, ΔU=0kV, ΔP=0MPa, disturbance level 1;

[0085] • Powder coating compensation: Δt voltage1 = 1611.49 × 0.006 × 0.2 × 1.2 × 1.0 ≈ 2.32s, Δt pressure1 = 1611.49 × 0.2 × 0.2 × 1.2 × 1.0 ≈ 77.35s, Δt = 2.32 + 77.35 + 0 = 79.67s.

[0086] (5) Calculation of changeover time

[0087] • Cleaning time: T_cleaning powder = 0.5 + (10 / 100) × 0.2 = 0.52 min = 31.2 s, after parallel correction T_cleaning = 31.2 / (4 / 2) = 15.6 s;

[0088] • Parameter adjustment time: T_adjustment = 0.3 + (ΔD / 100) × 0.1 + (Δd / 50) × 0.15 = 0.3 + 0.01 + 0.03 = 0.34 min = 20.4 s;

[0089] • Total changeover time: T_changeover = (15.6 + 20.4) × 0.9 = 36 × 0.9 = 32.4 s.

[0090] (6) Production line cycle time correction

[0091] • Pre-treatment correction amount T_pre-treatment correction = -12s, curing oven correction amount T_curing correction = 0s, subsequent idle correction amount T_idle correction = -18s, total correction amount T_cycle correction = -12 + 0 - 18 = -30s.

[0092] (7) Multimodal fusion and iterative correction

[0093] • Modal score calculation: Svisual = 0.95, Sprocess = 0.92, Sperturbation = 0.9, Shistory = 0.93, adaptive weights Wvisual ≈ 0.41, Wprocess ≈ 0.29, Wperturbation ≈ 0.18, Whistory ≈ 0.12;

[0094] • The final value of the cross-modal fusion prediction T = 0.41 × 1611.49 + 0.29 × (1003.5 + 120.58 + 316.51) + 0.18 × 79.67 + 0.12 × (1611.49 + 20) ≈ 1580.3s;

[0095] • Correction of final offline time T after iterative correction n =(1611.49+32.4+79.67-30+1580.3×0.1)×0.85+0×0.15≈(1693.56+158.03)×0.85≈1577.35s.

[0096] 3. Test Results

[0097] The final predicted offline time was 1577.35s, while the actual offline time was 1568.2s, with a prediction error of ≈0.58%, which meets the requirement of accuracy ≥98.5%.

[0098] Example 2

[0099] 1. Test conditions

[0100] • Wheel specifications: 18 inches (center size), number of spokes N=12 (number of center spokes), actual outer diameter of rim D r =600mm (intermediate outer diameter), rim height H r =120mm, flange outer diameter D f =300mm, flange face center hole diameter D hole =100mm, spoke blocking area A occ =100cm², center distance between adjacent spokes S s=50mm;

[0101] • Coating type: Liquid coating (W2=0.9);

[0102] • Process parameters: Target coating thickness T = 100 μm (intermediate coating thickness), spray flow rate Q = 200 ml / min, powder coating density ρ = 2.8 g / cm³, coating adhesion efficiency η = 0.85, number of spray guns M = 6, electrostatic voltage reference value U reference = 70 kV (intermediate voltage), fluidizing gas pressure reference value P reference = 0.2 MPa (intermediate pressure), coating viscosity reference value μ reference = 35 mPa·s;

[0103] • Change parameters: difference between previous and current wheel hub diameter ΔD = 30mm (intermediate diameter difference), difference in coating thickness Δd = 30μm (intermediate thickness difference), historical correction coefficient for specifications K specification = 1.0 (intermediate correction coefficient).

[0104] • Iteration parameters: Iteration period = 10 seconds (intermediate period), iteration correction coefficient α = 0.9 (intermediate value);

[0105] • Production line status: Pre-processing queue quantity Q_pre-processing = 6, curing oven queue quantity Q_curing = 4, and subsequent assembly process idle time T_idle = 3 minutes.

[0106] 2. Calculation process

[0107] (1) Calculation of actual spraying area by region

[0108] •Spoke area: A s =[π×(600 / 2)²-π×(300 / 2)²-100×100]×1.2×10^(-2)=(282600-70650-10000)×0.012=201950×0.012=2423.4cm²;

[0109] •Rim area: A r =π×600×120×10^(-2)=226080×0.01=2260.8cm²;

[0110] • Flange face: A p =π×[(300 / 2)²-(100 / 2)²]×10^(-2)=(70650-7850)×0.01=628cm².

[0111] (2) Calculation of basic spraying time

[0112] • Total mass of coating (spoke area): m = 2423.4 × 100 × 10^(-4) × 2.8 = 2423.4 × 0.01 × 2.8 = 67.8552 g;

[0113] • Basic time consumed in the spoke area: t s Baseline = (67.8552 × 1000) / (200 × 2.8 × 0.85 × 6) × 60 = 67855.2 / 2856 × 60 ≈ 23.76 × 60 = 1425.6 s;

[0114] Similarly, calculate t for the rim area. r Reference point = 1356.48s, flange face t p Baseline = 753.6s.

[0115] (3) Dynamic correction factor and time consumed after correction

[0116] • The spoke area K1 = 1.0 + 0.02 × 12 - 0.001 × 50 = 1.0 + 0.24 - 0.05 = 1.19;

[0117] • Rim area K2 = 0.95 + 0.0001 × (600 - 400) = 0.97;

[0118] • Correction time: t s Final value = 1425.6 × 1.19 ≈ 1696.46 s, t r Final value = 1356.48 × 0.97 ≈ 1315.79 s, t p Final value = 753.6 × 1.02 ≈ 768.67s, total time ttotal = 1696.46 + 1315.79 + 768.67 ≈ 3780.92s.

[0119] (4) Disturbance compensation time

[0120] • After real-time acquisition and processing of disturbance data, ΔU=2kV, ΔP=0.02MPa, Δμ=3mPa·s, disturbance level 2;

[0121] • Liquid coating compensation: Δt voltage2 = 3780.92 × 0.005 × 0.2 × 1.0 × 0.9 ≈ 3.40s, Δt pressure2 = 3780.92 × 0.2 × 0.2 × 1.0 × 0.9 ≈ 136.11s, Δt viscosity2 = 3780.92 × 0.003 × 0.9 ≈ 10.21s, Δt = 3.40 + 136.11 + 10.21 = 149.72s.

[0122] (5) Calculation of changeover time

[0123] • Cleaning time: T cleaning liquid = 1.0 + (30 / 100) × 0.3 + (30 / 50) × 0.2 = 1.0 + 0.09 + 0.12 = 1.21 min = 72.6 s, after parallel correction T cleaning = 72.6 / (6 / 2) = 24.2 s;

[0124] • Parameter adjustment time: T_adjustment = 0.3 + (ΔD / 100) × 0.1 + (Δd / 50) × 0.15 = 0.3 + 0.03 + 0.09 = 0.42 min = 25.2 s;

[0125] • Total changeover time: T_changeover = (24.2 + 25.2) × 1.0 = 49.4 s.

[0126] (6) Production line cycle time correction

[0127] • Pre-treatment correction amount T_pre-treatment correction = 0s, curing oven correction amount T_curing correction = (4-3)×90 = 90s, subsequent idle correction amount T_idle correction = -6s, total correction amount T_cycle correction = 0 + 90 - 6 = 84s.

[0128] (7) Multimodal fusion and iterative correction

[0129] • Modal score calculation: Svisual = 0.93, Sprocess = 0.90, Sperturbation = 0.8, Shistory = 0.91, adaptive weights Wvisual ≈ 0.42, Wprocess ≈ 0.30, Wperturbation ≈ 0.17, Whistory ≈ 0.11;

[0130] • The final value of the cross-modal fusion prediction T = 0.42 × 3780.92 + 0.30 × (1425.6 + 1356.48 + 753.6) + 0.17 × 149.72 + 0.11 × (3780.92 + 35) ≈ 3720.5s;

[0131] • Correction of final offline time T after iterative correction n =(3780.92+49.4+149.72+84+3720.5×0.1)×0.9+0×0.1≈(4064.04+372.05)×0.9≈3992.48s.

[0132] 3. Test Results

[0133] The final predicted offline time was 3992.48s, while the actual offline time was 3968.5s, with a prediction error of ≈0.60%, which meets the requirement of accuracy ≥98.5%.

[0134] Example 3

[0135] 1. Test conditions

[0136] • Wheel specifications: 24 inches (maximum size), number of spokes N=15 (maximum number of spokes), actual outer diameter of rim D r =800mm (maximum rim outer diameter), rim height H r =160mm, flange outer diameter Df =400mm, flange face center hole diameter D hole =120mm, spoke shading area A occ =150cm², center distance between adjacent spokes S s =70mm;

[0137] • Coating type: Powder coating (W1=1.0);

[0138] • Process parameters: Target coating thickness T=150μm (maximum coating thickness), spray flow rate Q=300ml / min, powder coating density ρ=2.8g / cm³, coating adhesion efficiency η=0.9, number of spray guns M=8, electrostatic voltage reference value U reference=90kV (maximum voltage), fluidizing gas pressure reference value P reference=0.3MPa (maximum pressure).

[0139] • Change parameters: Previous and current wheel hub diameter difference ΔD = 50mm (maximum diameter difference), coating thickness difference Δd = 50μm (maximum thickness difference), specification-specification historical correction coefficient K specification = 1.1 (maximum correction coefficient).

[0140] • Iteration parameters: Iteration period = 30 seconds (maximum period), iteration correction coefficient α = 0.95 (maximum value);

[0141] • Production line status: Pre-processing queue quantity Q_pre-processing = 15, curing oven queue quantity Q_curing = 8, and subsequent assembly process idle time T_idle = 1 minute.

[0142] 2. Calculation process

[0143] (1) Calculation of actual spraying area by region

[0144] •Spoke area: A s =[π×(800 / 2)²-π×(400 / 2)²-150×100]×1.2×10^(-2)=(502400-125600-15000)×0.012=361800×0.012=4341.6cm²;

[0145] •Rim area: A r =π×800×160×10^(-2)=401920×0.01=4019.2cm²;

[0146] • Flange face: A p =π×[(400 / 2)²-(120 / 2)²]×10^(-2)=(125600-11304)×0.01=1142.96cm².

[0147] (2) Calculation of basic spraying time

[0148] • Total mass of coating (spoke area): m = 4341.6 × 150 × 10^(-4) × 2.8 = 4341.6 × 0.015 × 2.8 = 182.3472 g;

[0149] • Basic time consumed in the spoke area: t s Baseline = (182.3472 × 1000) / (300 × 2.8 × 0.9 × 8) × 60 = 182347.2 / 6048 × 60 ≈ 30.15 × 60 = 1809 s;

[0150] Similarly, calculate t for the rim area. r Reference point = 2411.52s, flange face t p Baseline = 1371.55s.

[0151] (3) Dynamic correction factor and time consumed after correction

[0152] • The spoke area K1 = 1.0 + 0.02 × 15 - 0.001 × 70 = 1.0 + 0.3 - 0.07 = 1.23;

[0153] • Rim area K2 = 0.95 + 0.0001 × (800 - 400) = 0.99;

[0154] • Correction time: t s Final value = 1809 × 1.23 ≈ 2225.07 s, t r Final value = 2411.52 × 0.99 ≈ 2387.40 s, t p Final value = 1371.55 × 1.02 ≈ 1398.98s, total time ttotal = 2225.07 + 2387.40 + 1398.98 ≈ 6011.45s.

[0155] (4) Disturbance compensation time

[0156] • After real-time acquisition and processing of disturbance data, ΔU=3kV, ΔP=0.03MPa, disturbance level 3;

[0157] • Powder coating compensation: Δt voltage1 = 6011.45 × 0.006 × 0.2 × 1.2 × 1.0 ≈ 8.65s, Δt pressure1 = 6011.45 × 0.2 × 0.2 × 1.2 × 1.0 ≈ 288.55s, Δt = 8.65 + 288.55 + 0 = 297.2s.

[0158] (5) Calculation of changeover time

[0159] • Cleaning time: T cleaning powder = 0.5 + (50 / 100) × 0.2 = 0.6 min = 36 s, after parallel correction T cleaning = 36 / (8 / 2) = 9 s;

[0160] • Parameter adjustment time: T_adjustment = 0.3 + (ΔD / 100) × 0.1 + (Δd / 50) × 0.15 = 0.3 + 0.05 + 0.15 = 0.5 min = 30 s;

[0161] • Total changeover time: T_changeover = (9 + 30) × 1.1 = 39 × 1.1 = 42.9 s.

[0162] (6) Production line cycle time correction

[0163] • Pre-treatment correction amount T_pre-treatment correction = 6 × (15-10) = 30s, curing oven correction amount T_curing correction = (8-3) × 90 = 450s, subsequent idle correction amount T_idle correction = 0s, total correction amount T_cycle correction = 30 + 450 + 0 = 480s.

[0164] (7) Multimodal fusion and iterative correction

[0165] • Modal score calculation: Svisual = 0.90, Sprocess = 0.88, Sperturbation = 0.7, Shistory = 0.89, adaptive weights Wvisual ≈ 0.43, Wprocess ≈ 0.31, Wperturbation ≈ 0.15, Whistory ≈ 0.11;

[0166] • The final value of the cross-modal fusion prediction T = 0.43 × 6011.45 + 0.31 × (1809 + 2411.52 + 1371.55) + 0.15 × 297.2 + 0.11 × (6011.45 + 50) ≈ 5980.2s;

[0167] • Correction of final offline time T after iterative correction n =(6011.45+42.9+297.2+480+5980.2×0.1)×0.95+0×0.05≈(6831.55+598.02)×0.95≈7088.52s.

[0168] 3. Test Results

[0169] The final predicted offline time was 7088.52s, while the actual offline time was 7032.6s, with a prediction error of ≈0.79%, which meets the requirement of accuracy ≥98.5%.

Claims

1. A method for predicting the off-line time of wheel hub painting, characterized in that, include: S1. Visual Recognition and Segmentation: Acquire wheel hub images, identify their structural features through machine vision, and divide the painted area into multiple sub-regions; S2. Regional Dynamic Time Prediction: Based on the structural characteristics of each sub-region and combined with the spraying process parameters, the coating time of each sub-region is predicted, and dynamic correction is performed based on real-time structural characteristics. S3. Adaptive Disturbance Compensation: Real-time acquisition of coating process disturbance parameters, and calculation of disturbance compensation time based on the current wheel hub coating type by calling the corresponding disturbance compensation model; S4. Multimodal fusion iterative prediction: The painting time of each sub-region after dynamic correction and the disturbance compensation time are fused together, and combined with historical production data, iterative calculation and correction are performed at fixed intervals to output the final offline time prediction value. In step S1, the plurality of sub-regions include at least a spoke area, a rim area, and a flange surface; the identification of its structural features by machine vision includes: identifying at least one of the following: the number of spokes, the area of ​​spoke obstruction, the outer diameter of the rim, the rim height, the outer diameter of the flange surface, and the diameter of the center hole; the machine vision identification process includes image preprocessing, feature extraction, and data verification, wherein the image preprocessing employs at least one of the following: weighted average grayscale conversion, Gaussian filtering, and edge detection; In step S2, the dynamic correction based on real-time structural features is achieved by multiplying the base time of each sub-region by a dynamic correction factor. The dynamic correction factor is calculated based on at least one parameter among the number of spokes, spoke spacing, and rim outer diameter. The base time of each sub-region is derived from the actual spraying area and spraying process parameters of the corresponding sub-region.

2. The method for predicting the off-line time of wheel hub painting according to claim 1, characterized in that, The formula for calculating the actual sprayed area of ​​each region is as follows: Spoke area: ， in This is the actual outer diameter of the wheel rim. The outer diameter of the flange face. 1.2 is the spoke shading area, 1.2 is the spoke side coating compensation coefficient, and 100 is... change Conversion factor; Rim area: ， The height of the wheel rim. for change Conversion factor; Flange face: ， The diameter of the center hole on the flange face; The general formula for calculating the basic spraying time is: ， in , This represents the actual sprayed area. For the target coating thickness, For the density of powder coating, For coating adhesion efficiency, For spraying flow rate, For the number of spray guns, for change Conversion factor, 1000 is change Conversion factor, 60 is change Conversion factor.

3. The method for predicting the off-line time of wheel hub painting according to claim 2, characterized in that, The calculation logic for the dynamic correction factor is as follows: Spoke area correction factor , constraint , For the number of spokes, The center-to-center distance between adjacent spokes; Rim area correction factor , constraint Flange face correction factor ; The final time consumed is , , Total time Furthermore, it refits the data monthly based on historical deviation data of the same wheel specifications over the past three months. , coefficient.

4. The method for predicting the off-line time of wheel hub painting according to claim 3, characterized in that, In step S3, the coating type includes at least powder coating and liquid coating; for powder coating, the disturbance compensation model is mainly based on electrostatic voltage fluctuations and fluidizing gas pressure fluctuations for compensation; for liquid coating, the disturbance compensation model is also based on coating viscosity fluctuations for compensation; the core formula of the disturbance compensation model is... ,in Weighting by coating type, powder coating Liquid coating .

5. The method for predicting the off-line time of wheel hub painting according to claim 4, characterized in that, The process of acquiring and processing the process disturbance parameters includes: acquiring raw disturbance data in real time at a frequency of 10Hz; using a moving average filter with a window of 10 data points; removing outliers within a ±30% range of the baseline value; calculating the fluctuation value and the maximum fluctuation amplitude within a fixed time period; and determining a disturbance level of 1-3 based on the fluctuation amplitude. The moving average filter formula is as follows: The outlier removal rule is that when the original data exceeds the baseline value by ±30%, it is replaced with the filtered data from the previous time point; the fluctuation value includes... , , .

6. The method for predicting the off-line time of wheel hub painting according to claim 5, characterized in that, The calculation logic for disturbance compensation for different coating types is as follows: Powder coating: ， ， ； Liquid coating: ， ， 。 7. The method for predicting the off-line time of wheel hub painting according to claim 6, characterized in that, In step S4, the fusion process adopts an adaptive weight allocation method, and the weights are dynamically determined based on at least one of the following factors: visual recognition confidence, process parameter stability, real-time disturbance level, and historical prediction accuracy. The adaptive weight allocation formula is as follows: , This is the priority coefficient, default value. , , , When the sensor fails .

8. The method for predicting the off-line time of wheel hub painting according to claim 7, characterized in that, In step S4, the fixed period is 5 seconds to 30 seconds; The formula for iterative calculation and correction is as follows: ， in , The cumulative disturbance compensation time for the nth period is... This is the iterative correction coefficient, with a value range of 0.85-0.

95. The prediction bias is for the first n-1 periods; when A Level 3 disturbance warning was triggered.

9. The method for predicting the off-line time of wheel hub painting according to claim 8, characterized in that, The method further includes step S5: Step S5: Based on the specification differences between the previous wheel hub and the current wheel hub, the adjustment time required for the model change is predicted and incorporated into the final production time prediction; the calculation logic for the model change time is as follows: Cleaning time: powder , liquid , After parallel correction , The difference between the previous and current wheel hub diameters. Due to differences in coating thickness, This refers to the number of spray guns; Parameter adjustment time: ; Total time spent on model changeover , This refers to the historical correction factor for different specifications.

Images