Automobile foot mat cutting and defect detection method and system based on machine vision

Abstract

The application discloses a kind of based on machine vision's automobile foot pad cutting and defect detection method and system, belong to the intelligent manufacturing technical field of automotive interior parts, this method includes: obtaining multi-source perception data, executing material identification and the soft coupling feedback of deformation compensation;Perform spatial alignment and online feedback learning of defect detection before and after cutting;Perform collaborative early retirement of edge and cloud, and the decision of local output or uploading cloud is executed according to inference confidence and dynamic upload threshold comparison by edge;Perform hybrid deformation compensation of offline simulation and online force feedback, and the amount of compensation is corrected in real time by force feedback based on offline simulation results.Through four cross-link collaborative feedback loops, the information silos of traditional independent operation of each module are broken, and soft coupling and continuous self-optimization of material identification, deformation compensation, defect detection and cutting control are realized, which significantly improves cutting precision, defect detection accuracy and system adaptive capacity.

Description

Technical Field

[0001] This invention relates to a machine vision-based method and system for cutting and detecting defects in automotive floor mats, belonging to the field of intelligent manufacturing technology for automotive interior parts. Background Technology

[0002] Car floor mats come in a variety of materials (foam, coiled yarn, leather, etc.), requiring cutting precision on the order of ±0.1mm. Furthermore, flexible materials are prone to deformation during cutting, making compensation difficult. Currently, the industry generally designs material identification, defect detection, cutting control, and deformation compensation as independent processes, with each module connected via a "hard connection"—once the upstream module outputs a command, it disconnects from the downstream, and the downstream module cannot provide feedback to the upstream.

[0003] A search revealed that Chinese invention patent CN120620301A discloses a flexible material cutting robot and its cutting method. This system includes a cutting system, a vision positioning module, a feeding platform, a quality monitoring module, and an intelligent nested material arrangement system. However, the modules within this system still operate independently with a "hard connection": pre-cutting defect detection and post-cutting quality inspection are completely independent. The defect location map output by the pre-detection system has limited accuracy, and its direct use for post-detection traceability introduces spatial errors. Furthermore, there is no data correlation between the pre- and post-detection results, preventing the pre-detection system from continuously learning and improving from the post-detection system. Regarding deformation compensation, the patent only mentions a conceptual description of a "dynamic compensation system," without disclosing the specific implementation combining offline simulation benchmarks and online force feedback correction.

[0004] Furthermore, Chinese invention patent CN121259345A discloses a method and device for material recognition based on channel switching and visual-touch fusion. This method acquires paired data samples composed of visual and tactile images to construct a visual-touch fusion network for material recognition. However, in this scheme, material recognition only outputs a single discrete category label. Errors in the recognition result directly lead to the complete failure of subsequent compensation, and the recognition result cannot be corrected during the cropping process. Furthermore, no information transmission mechanism is established with the downstream deformation compensation module.

[0005] The information silos between these stages limit the accuracy, yield, and operational efficiency of the entire production system, preventing continuous improvement with increased operating time and data volume. Therefore, there is an urgent need in this field for an adaptive cutting and defect detection method that can break down information barriers between stages and achieve cross-stage collaborative feedback. Summary of the Invention

[0006] To achieve the above objectives, this invention provides a machine vision-based method for cutting and detecting defects in automotive floor mats. This method does not rely on improvements to the algorithms of individual modules, but rather on constructing four cross-stage collaborative feedback loops, each loop resolving a new technical contradiction arising from the combination of modules. Overall, the method includes the following steps:

[0007] The system acquires multi-source sensing data of the car floor mat material; performs soft-coupled feedback for material identification and deformation compensation, where material identification outputs the probability distribution of material categories, deformation compensation determines initial cutting compensation parameters based on the probability distribution, and uses force feedback data during the cutting process to update the probability distribution in reverse, and corrects the cutting compensation amount in real time; performs spatial alignment and online feedback learning for defect detection before and after cutting, where the defect positions detected before cutting and the defect positions detected after cutting are mapped to the same physical coordinate system for matching, and training data is generated based on the matching results to fine-tune the defect detection model before cutting; performs collaborative early termination between the edge and the cloud, where the cloud dynamically adjusts the upload threshold based on the load, and the edge makes a decision on local output or upload to the cloud based on the comparison between the inference confidence and the dynamic upload threshold; and performs hybrid deformation compensation of offline simulation and online force feedback, where the compensation amount is corrected in real time through force feedback based on the offline simulation results, and the simulation model parameters are calibrated in reverse based on the statistical deviation accumulated in batches.

[0008] Furthermore, in the soft-coupled feedback: the probability distribution is generated based on multimodal data, which includes at least visible light images and force sensor signals; based on the likelihood between the force feedback data and the predicted force data of each type of material, the probability distribution is updated in reverse using the Bayesian formula; the initial trimming compensation parameter is a weighted sum of the preset compensation parameters corresponding to each category according to the probability distribution, and the compensation variance is output; the proportional coefficient of the real-time correction compensation amount is dynamically adjusted according to the compensation variance.

[0009] Furthermore, in the spatial alignment and online feedback learning: the perspective transformation matrix of image pixel coordinates and physical coordinates is calculated by a calibration board to realize defect location mapping; the matching adopts nearest neighbor matching, the matching radius is less than a preset threshold, and the matching result includes three situations: successful matching, no matching, and the defect detected in the previous detection does not appear in the subsequent detection; among them, the case of no matching automatically generates weak labels and adds them to the active learning queue, and the elastic weight consolidation method is used to fine-tune the defect detection model before pruning online.

[0010] Furthermore, in the collaborative early termination: the dynamic upload threshold is positively correlated with the cloud load status; the edge performs hierarchical decision-making based on the comparison result between the inference confidence and the dynamic upload threshold: when the confidence is higher than the first threshold, local output is performed; when the confidence is between the first threshold and the second threshold, the data to be inferred is compressed and uploaded to the cloud, along with a local low-confidence feature vector to accelerate cloud inference; when the confidence is lower than the second threshold, manual review is triggered; the cloud inference results are periodically used for knowledge distillation to update the edge model.

[0011] Furthermore, in the hybrid deformation compensation: the offline simulation result is a compensation amount lookup table generated based on the hyperelastic model; the force feedback real-time correction compensation amount includes looking up the predicted force according to the current material properties, calculating the deviation between the actual cutting force and the predicted force, and inputting the deviation into the PID controller to output the compensation correction amount; the batch-accumulated statistical deviation reverse calibration includes statistically analyzing the mean and standard deviation of a batch of cutting force deviations, and when the mean deviates from zero by more than a threshold, adjusting the hyperelastic parameters of the simulation model through Bayesian optimization and updating the compensation amount lookup table.

[0012] Furthermore, the method also includes cross-loop information sharing: using the probability distribution of the material identification output for parameter adjustment of the PID controller in the hybrid deformation compensation; using the matching results of the defect detection before and after cutting as samples for training the meta-model in the collaborative early termination; and using the compensation amount data of the force feedback real-time correction for reverse updating of the probability distribution.

[0013] This invention also provides a machine vision-based automotive floor mat cutting and defect detection system, comprising: a material identification module configured to receive multi-source sensing data and output a probability distribution of material categories; a deformation compensation module configured to determine initial cutting compensation parameters and compensation variance based on the probability distribution; a force feedback module configured to collect real-time force feedback data and update the probability distribution in reverse based on the force feedback data; a pre-cutting detection module configured to perform defect detection on the material and map the defect location to physical coordinates; and a post-cutting detection module configured to perform defect detection on the finished product and map the defect location to the physical coordinates. The system includes: a physical coordinate system; a feedback learning module configured to match the defect locations before and after trimming, and generate training data based on the matching results to fine-tune the pre-trimming detection module; an edge-cloud collaboration module configured to dynamically adjust the upload threshold based on cloud load, and make decisions on local output or cloud upload based on a comparison of inference confidence and the dynamic upload threshold; and a hybrid deformation compensator configured to use offline simulation results as a benchmark, correct the compensation amount in real time through force feedback, and calibrate the simulation model parameters in reverse based on the statistical deviation accumulated in batches. These modules form a cross-stage information sharing and collaborative feedback loop.

[0014] Furthermore, the hybrid deformation compensator includes: a PID controller whose control parameters are dynamically adjusted according to the variance of the compensation amount; and a simulation parameter calibration unit for calibrating the hyperelastic model parameters through Bayesian optimization when batch deviation is triggered.

[0015] Furthermore, the edge-cloud collaboration module includes: a dynamic threshold calculation unit for calculating upload thresholds based on cloud load and network latency; a multi-exit early termination network for outputting inference confidence; and a knowledge distillation unit for periodically distilling cloud inference results to the edge model.

[0016] In addition, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: By constructing a soft-coupled feedback loop between material identification and deformation compensation, material identification is output in the form of a probability distribution, deformation compensation determines the initial parameters based on probability weighting, and the probability distribution is corrected in reverse and the compensation amount is adjusted in real time using force feedback in cutting. This solves the technical problem in the prior art where material identification is completed in one go and cannot be corrected downstream, resulting in the scrapping of the entire material. It realizes the transfer of discrete tags to continuous parameters and bidirectional closed-loop update. By establishing a spatial alignment and online feedback learning mechanism for pre- and post-cutting detection, using a perspective transformation matrix to realize the physical coordinate mapping of the defect location, generating positive and negative samples through nearest neighbor matching, and using an elastic weight consolidation method to fine-tune the pre-detection model, the technical problems of existing pre- and post-detection running independently and the pre-detection being unable to continuously learn and improve from the post-detection are solved, thus achieving continuous optimization of detection accuracy. By introducing an edge-cloud collaborative early termination mechanism, the cloud dynamically adjusts the upload threshold based on the load, and the edge performs hierarchical decision-making based on the comparison between confidence and threshold. The cloud inference results are updated to update the edge model through knowledge distillation, which solves the technical problem that the existing fixed threshold scheme cannot adapt to dynamic network conditions and significantly reduces cloud bandwidth consumption while maintaining high accuracy. By constructing a hybrid deformation compensator combining offline simulation and online force feedback, using a lookup table generated by the Abaqus hyperelastic model as a benchmark, and combining it with a PID controller to correct the compensation amount in real time, and using Bayesian optimization to calibrate the simulation model parameters based on batch statistical deviations, the contradictions of existing offline simulation being unable to adapt to batch differences and pure online feedback response lag are resolved, resulting in a significant improvement in cutting accuracy. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the overall architecture of the car floor mat cutting and defect detection system in an embodiment of the present invention; Figure 2 This is a schematic diagram of the spatial alignment method for pre-cutting and post-cutting detection of the present invention, illustrating the calibration plate coordinate mapping and nearest neighbor matching process; Figure 3 The flowchart of the uncertainty-aware edge-cloud collaborative early termination mechanism of the present invention is shown, illustrating the three-layer threshold partitioning processing logic. Figure 4 This is a flowchart of the PID control loop and batch parameter calibration of the simulation-feedback hybrid deformation compensator of the present invention; Figure 5 This is a flowchart illustrating the overall process of cutting and detecting defects in car floor mats in this embodiment of the invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that in the optional embodiments of the present invention, the object information, industrial data, and other related data involved require the permission or consent of the object when the embodiments of the present invention are applied to specific products or technologies, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0021] Terminology Explanation: Soft coupling feedback: The material identification module outputs the probability distribution of each category (instead of a single discrete label), the deformation compensation module calculates the weighted compensation amount based on the probability distribution, and uses the force feedback data during the cutting process to update the probability distribution in reverse, thereby forming a two-way information closed loop between the identification result and the compensation decision.

[0022] Spatial alignment: The perspective transformation matrix between the image pixel coordinates and the physical coordinates of the cropping platform is calculated by the calibration board, and the defect positions detected before and after cropping are mapped to the same coordinate system to eliminate the spatial deviation between the camera and the platform.

[0023] Dynamic upload threshold: The cloud calculates the upload decision threshold in real time based on its own load (CPU / GPU utilization) and network latency. The higher the load, the larger the threshold, in order to reduce the number of uploaded samples; the lower the load, the smaller the threshold, in order to accept more samples for model optimization.

[0024] Hybrid Deformation Compensator: A compensation device that combines an offline simulation-generated benchmark compensation lookup table with online force feedback PID real-time correction, taking into account both the adaptability to material differences between batches and the speed of real-time response.

[0025] Knowledge distillation: The process of using the inference results of large cloud models as "soft labels" to update the parameters of small edge models, so that the edge models gradually approach the performance of cloud models.

[0026] Example 1 Please see Figure 5 This is a flowchart illustrating the overall process of a method for cutting and detecting defects in automotive floor mats, as provided in Embodiment 1. It includes the following steps: Step S1: Acquire multi-source sensing data and perform soft-coupled feedback for material identification and deformation compensation.

[0027] It should be noted that the acquisition of multi-source sensing data for car floor mat materials includes at least visible light images, ultraviolet supplemental light images, and contact force signals collected by force sensors. Image features and force signal features are concatenated into a high-dimensional feature vector, which is then input into a pre-trained material recognition neural network with a ResNet18 framework. The output is a probability distribution of the material belonging to each category, for example, P(foam) = 0.7, P(silk loop) = 0.2, and P(leather) = 0.1.

[0028] Optionally, step S1 may include the following sub-steps: S11. Determine initial cutting compensation parameters based on the probability distribution. Specifically, pre-store compensation parameters for each material. The initial compensation amount is calculated using the following formula, along with the compensation confidence matrix:

[0029]

[0030] Meanwhile, the variance of the compensation amount is calculated using the following formula:

[0031] in, For the first The probability of similar materials, , These are the preset compensation amounts for this type of material in the x and y directions, respectively. The variance of these compensation amounts is used for the adaptive adjustment of subsequent PID control parameters.

[0032] S12. During the cutting process, the probability distribution is updated in reverse using force feedback data. Specifically, the force sensor collects the actual cutting force in real time. For each material, the predicted force is calculated using a physical model or lookup table based on the current cutting parameters. Calculate the likelihood of the actual force versus the predicted force:

[0033] in, To measure the standard deviation of the noise, we then use Bayes' theorem to update the material probability distribution in reverse:

[0034] The updated probability distribution is fed back to the deformation compensation module to recalculate the compensation amount, forming a closed loop.

[0035] S13. Real-time correction of cutting compensation. Based on the updated probability distribution and real-time force feedback data, the baseline compensation and predicted force are obtained by looking up a table according to the current material properties (weighted by the real-time probability distribution), and the force deviation is calculated. .Will The compensation correction is obtained by inputting the PID controller. . Among them, the proportional coefficient of the PID controller According to the variance of the compensation amount Dynamic adjustment: The larger the variance, The larger the value, the greater the weight of the feedback response.

[0036] It is worth noting that S11 to S13 together constitute a soft-coupled feedback loop between material identification and deformation compensation, realizing the transfer of discrete tags to continuous parameters and bidirectional closed-loop updates. This step solves the technical bias in the prior art where material identification is completed in one go and force feedback is only used for monitoring.

[0037] Step S2: Perform spatial alignment and online feedback learning for pre- and post-cropping detection.

[0038] It should be noted that a high-precision checkerboard calibration board with a grid side length of 10mm is placed on the cutting platform, the image of the calibration board is acquired, the corner pixel coordinates are extracted, and the corresponding physical coordinates are calculated to calculate the perspective transformation matrix H, which is used to map the image pixel coordinates (u,v) to the physical coordinates (x,y) of the cutting platform.

[0039] Furthermore, step S2 may include the following sub-steps: S21. Before cutting, the camera scans the material surface to detect defects such as bubbles, scratches, holes, and stains, outputs a defect location map, maps it to physical coordinates through the perspective transformation matrix H, and stores it in the database, along with the defect type and severity.

[0040] S22. After cutting, image the finished product, extract the defect location, and map it to physical coordinates. For each post-inspection defect, search for the nearest neighbor pre-inspection defect in the pre-cutting database, with a matching radius of 2mm. The value of the matching radius can be adjusted according to the actual accuracy requirements.

[0041] S23. Generate training data based on the matching results. Specifically, the matching results are processed in three ways: If a match is successful, the pre-detection result is correct, and positive samples are accumulated. If no match is found in the subsequent defect detection, it is determined to be a missed detection. The image of the defect area is cropped to automatically generate a weak label and added to the active learning queue. If a defect in the pre-detection does not appear in the post-detection, it is correctly ignored and not included in the training.

[0042] S24. Based on the accumulated positive and negative samples from the matching results each day, the elastic weight consolidation (EWC) method is used to fine-tune the pre-pruning detection model online: calculate the Fisher information matrix of the model parameters, apply strong constraints to important parameters, prevent forgetting existing knowledge, and achieve continuous optimization of pre-detection performance.

[0043] Step S3: Implement the edge-cloud collaborative early termination mechanism.

[0044] Optionally, step S3 may include the following sub-steps: S31. Deploy a multi-exit early termination network at the edge (set up 3 exits), and output the predicted category and confidence level for each exit. The cloud server monitors its own CPU / GPU load and network latency (RTT) in real time, and calculates the dynamic upload threshold using the following formula. :

[0045] in As the baseline threshold, Sensitivity coefficient. Under low load. Reduce the load to accommodate more uploaded samples when the load is high. Increase the upload speed to reduce the upload speed.

[0046] S32. The edge device performs hierarchical decision-making based on inference confidence and dynamic upload threshold: like The result is output directly at the edge; like ,in To reduce bandwidth latency, images are compressed and uploaded to the cloud, along with local low-confidence feature vectors to accelerate cloud inference. like If so, manual review will be triggered.

[0047] S33. The cloud-based inference results for uploaded samples are periodically updated to the edge model through knowledge distillation. Simultaneously, the upload necessity label for each sample is recorded, and a meta-model is trained to predict which samples most need to be uploaded, thus optimizing the early termination strategy.

[0048] Step S4: Perform hybrid deformation compensation using offline simulation and online force feedback.

[0049] Specifically, step S4 includes the following sub-steps: S41. Using the Abaqus hyperelastic model, perform offline simulations for different combinations of materials, hardness, thickness, and cutting speed. Generate a baseline compensation lookup table (LUT) to determine the material type, hardness, thickness, and speed, and store the corresponding predicted forces. .

[0050] S42. During the cutting process, the baseline compensation amount is obtained by looking up a table based on the current material properties obtained by weighting the probability distribution from step S1 and the cutting parameters. and predictive power The force sensor collects the actual cutting force in real time. Calculation of force deviation . Input PID controller output compensation correction amount PID parameters are proportional parameters. ,integral ,differential The compensation amount is dynamically adjusted based on the variance output in step S1: when the variance is large, the adjustment is increased. To increase the weight of feedback response. .

[0051] S43. After each batch is completed, compile all statistics. mean of values and standard deviation .like Significant deviation from zero, i.e. This triggers the calibration of simulation model parameters. By adjusting the coefficients such as C10 and C01 of the Mooney-Rivlin model through Bayesian optimization, the predictive power of the regenerated LUT is made closer to the measured power, the lookup table is updated, and a positive feedback loop of simulation, measurement, correction and re-simulation is formed in sequence.

[0052] Cross-loop information sharing: The four loops mentioned above generate a synergistic effect through information sharing. Specifically: the variance of the compensation output in step S1. The PID parameters are adaptively adjusted in step S4; the matching results from step S2 are positive and negative samples, used for meta-model training in step S3; the force feedback deviation in step S4... The Bayesian update is used in step S1. Ablation experiment data shows that when all four loops are enabled, the material identification accuracy is 99%, the cutting accuracy is ±0.08mm, and the defect detection recall is 99%, which is significantly higher than the linear superposition of any single loop or two-loop combination, proving that cross-loop information sharing produces unexpected technical effects.

[0053] Example 2 Please see Figure 1 This is a structural block diagram of an automotive floor mat cutting and defect detection system provided in Embodiment 2.

[0054] This embodiment provides a machine vision-based automotive floor mat cutting and defect detection system, including: The material identification module is configured to receive multi-source sensing data, which includes at least visible light images, ultraviolet supplemental light images, and force sensor signals. The multi-modal features are then stitched together and input into a deep neural network to output the probability distribution of material categories.

[0055] The deformation compensation module is configured to calculate the initial cutting compensation amount and the compensation amount variance based on the probability distribution, wherein the compensation amount is a weighted sum of preset compensation parameters corresponding to each category according to the probability distribution. This module also receives an updated probability distribution from the force feedback module and dynamically adjusts the compensation amount.

[0056] The force feedback module is configured to collect cutting force data in real time during the cutting process, and based on the likelihood between the cutting force data and the predicted force data of each type of material, use Bayes' formula to update the probability distribution of the material type in reverse, and return the updated probability distribution to the deformation compensation module.

[0057] The pre-cutting inspection module is configured to detect surface defects in the material, map the detected defect locations to the physical coordinates of the cutting platform using a perspective transformation matrix, and store them in the database.

[0058] The post-cutting inspection module is configured to detect surface defects in the finished product after cutting. It also maps the defect location to physical coordinates and performs nearest neighbor matching with the defects in the pre-cutting inspection module database.

[0059] The feedback learning module is configured to generate training data based on the matching results of the detection before and after cropping: when a match is successful, positive samples are accumulated; when there is no match, weak labels are generated and added to the active learning queue, and the elastic weight consolidation method is used to fine-tune the detection model before cropping online.

[0060] The edge-cloud collaboration module is configured to dynamically calculate the upload threshold based on cloud load and network latency, and perform hierarchical decision-making based on a comparison between the edge inference confidence and the dynamic upload threshold. Hierarchical decision-making includes local output, cloud upload, and manual review. This module internally includes a dynamic threshold calculation unit, a multi-exit early termination network, and a knowledge distillation unit.

[0061] The hybrid deformation compensator is configured to use a compensation lookup table generated from offline simulation as a reference, and a PID controller to correct the compensation amount in real time based on the force feedback deviation. Furthermore, it uses batch-accumulated statistical deviations to back-calibrate the simulation model parameters, i.e., adjusting the hyperelastic model coefficients through Bayesian optimization. The PID control parameters of this compensator are dynamically adjusted according to the variance of the compensation amount output by the deformation compensation module.

[0062] The modules mentioned above are connected via a data bus or communication network, forming a cross-stage information sharing and collaborative feedback loop. Specifically, the output of the material identification module is connected to the input of the deformation compensation module; the output of the force feedback module is connected to both the material identification module and the hybrid deformation compensator, whereby the material identification module is used for Bayesian updates and the hybrid deformation compensator is used for PID correction; the outputs of both the pre-cutting detection module and the post-cutting detection module are connected to the feedback learning module; the output of the feedback learning module is connected to the pre-cutting detection module for model fine-tuning; the edge-cloud collaboration module communicates with the cloud server and interacts with the edge model to exchange knowledge distillation results; the hybrid deformation compensator communicates bidirectionally with both the force feedback module and the deformation compensation module.

[0063] Example 3 Two typical configuration options are provided, suitable for production scenarios with different budgets and accuracy requirements.

[0064] Hardware Configuration Comparison Table Industrial cameras 2 megapixels 2-megapixel and line laser profilometer light source 365nm UV LED supplemental lighting and 400nm long-pass filter backlight transmission LED surface light source Same as left and supplement uniform light field Preload rollers Silicone preload roller Same as left and online pressure sensor edge nodes Raspberry Pi 4B and Google Coral Accelerator Stick NVIDIA Jetson Orin Nano cloud Lightweight GPU server (available for rent) Tongzuo Cutting execution Vibration knife force sensor Same as left and ultrasonic flaw detection module Total investment (excluding the cutting host) Approximately 15,000 yuan Approximately 45,000 yuan Software / Algorithm Configuration Comparison Table Material identification model RGB + UV + Force Feedback Fusion Tongzuo Defect detection model PatchCore Anomaly Detection and Synthetic Data Augmentation Small Sample Classifier Same as left and online thickness measurement closed loop Multiple export early retirement network 3 exits, threshold Tongzuo Deformation Compensation Lookup Table Thickness step size: 0.5mm; Speed ​​step size: 25mm / s Tongzuo Simulation model Abaqus hyperelasticity model Tongzuo Under the above configuration, the actual implementation effects of the two schemes are as follows: Material identification accuracy 97.3%±1.2% 99% Defect detection accuracy 96.5%～98.8% 99.5% Cutting precision (soft materials) ±0.12mm ±0.10mm Internal defect detection rate — 95% Cutting accuracy distribution (within ±0.08mm) — 82% Cutting accuracy distribution (±0.08~0.12mm) — 15% Cutting accuracy distribution (±0.12~0.15mm) — 3% Single piece beat <60 seconds <60 seconds Total investment (excluding the cutting host) Approximately 15,000 yuan Approximately 45,000 yuan

[0065] It should be noted that the low-cost basic configuration is suitable for small and medium-sized car floor mat manufacturers with limited budgets, primarily producing foam and coil materials; the high-precision reinforced configuration is suitable for high-end car floor mat production, where materials include rigid composites and leather, requiring extremely high precision. The precision distribution data in Table 3 is based on the sampling inspection results of 100 batches. Analysis of variance shows that material thickness fluctuation is the main factor contributing to the precision differences, and further optimization can be achieved by introducing an online thickness measurement closed loop.

[0066] Implementation Notes: The UV LED module needs to be equipped with a constant current drive and a cooling fan to prevent light decay. That is, when working continuously for 16 hours a day, the power can drop from 30mW / cm² to 22mW / cm². The silicone pre-pressure roller needs to be replaced every 2 months and equipped with an online pressure sensor for monitoring.

[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine vision-based method for cutting and detecting defects in automotive floor mats, characterized in that, include: Acquire multi-source sensing data on car floor mat materials; The soft-coupled feedback of material identification and deformation compensation is implemented, wherein the material identification outputs the probability distribution of material categories, the deformation compensation determines the initial cutting compensation parameters based on the probability distribution, and uses the force feedback data during the cutting process to update the probability distribution in reverse, and corrects the cutting compensation amount in real time. Spatial alignment and online feedback learning for defect detection before and after trimming are performed. Specifically, the defect positions detected before trimming and the defect positions detected after trimming are mapped to the same physical coordinate system for matching. Training data is generated based on the matching results to fine-tune the defect detection model before trimming. The system performs collaborative early termination between the edge and the cloud, wherein the cloud dynamically adjusts the upload threshold based on the load, and the edge makes a decision on local output or upload to the cloud based on the comparison between the inference confidence and the dynamic upload threshold. Hybrid deformation compensation combining offline simulation and online force feedback is performed. The offline simulation results are used as a benchmark, and the compensation amount is corrected in real time through force feedback. The simulation model parameters are then calibrated in reverse based on the statistical deviation accumulated from batches.

2. The method for cutting and detecting defects in car floor mats based on machine vision according to claim 1, characterized in that, In the soft-coupled feedback: the probability distribution is generated based on multimodal data, which includes at least visible light images and force sensor signals; the probability distribution is updated in reverse using Bayes' theorem based on the likelihood between the force feedback data and the predicted force data of each type of material. The initial trimming compensation parameter is a weighted sum of the preset compensation parameters corresponding to each category according to the probability distribution, and the compensation variance is output. The proportional coefficient of the real-time correction compensation amount is dynamically adjusted according to the compensation variance.

3. The method for cutting and detecting defects in car floor mats based on machine vision according to claim 1, characterized in that, In the spatial alignment and online feedback learning: the perspective transformation matrix between image pixel coordinates and physical coordinates is calculated by a calibration board to realize defect location mapping; the matching adopts nearest neighbor matching, the matching radius is less than a preset threshold, and the matching result includes three cases: successful matching, no matching, and the defect detected in the previous detection does not appear in the subsequent detection; among them, the case of no matching automatically generates weak labels and adds them to the active learning queue, and the elastic weight consolidation method is used to fine-tune the defect detection model before cropping online.

4. The method for cutting and detecting defects in car floor mats based on machine vision according to claim 1, characterized in that, In the collaborative early termination process: the dynamic upload threshold is positively correlated with the cloud load status; the edge performs hierarchical decision-making based on the comparison result between the inference confidence and the dynamic upload threshold: when the confidence is higher than the first threshold, local output is performed; when the confidence is between the first threshold and the second threshold, the data to be inferred is compressed and uploaded to the cloud, along with a local low-confidence feature vector to accelerate cloud inference; when the confidence is lower than the second threshold, manual review is triggered; the cloud inference results are periodically used for knowledge distillation to update the edge model.

5. The method for cutting and detecting defects in car floor mats based on machine vision according to claim 1, characterized in that, In the hybrid deformation compensation: the offline simulation result is a compensation amount lookup table generated based on the hyperelastic model; the force feedback real-time correction compensation amount includes looking up the predicted force according to the current material properties, calculating the deviation between the actual cutting force and the predicted force, and inputting the deviation into the PID controller to output the compensation correction amount; the batch accumulated statistical deviation reverse calibration includes statistically analyzing the mean and standard deviation of the cutting force deviation of a batch, and when the mean deviates from zero by more than a threshold, adjusting the hyperelastic parameters of the simulation model through Bayesian optimization and updating the compensation amount lookup table.

6. The method for cutting and detecting defects in car floor mats based on machine vision according to claim 1, characterized in that, It also includes cross-loop information sharing: using the probability distribution of the material identification output for parameter adjustment of the PID controller in the hybrid deformation compensation; using the matching results of the defect detection before and after cutting as samples for training the meta-model in the collaborative early termination; and using the compensation amount data of the force feedback real-time correction for reverse updating of the probability distribution.

7. A machine vision-based system for cutting and detecting defects in automotive floor mats, characterized in that, include: The material identification module is configured to receive multi-source sensing data and output the probability distribution of material categories; The deformation compensation module is configured to determine the initial cutting compensation parameters and the compensation amount variance based on the probability distribution. The force feedback module is configured to collect real-time force feedback data and update the probability distribution in reverse based on the force feedback data. The pre-cutting inspection module is configured to detect defects in the material and map the defect locations to physical coordinates; The post-cutting inspection module is configured to detect defects in the finished product and map the defect locations to the physical coordinates. The feedback learning module is configured to match the defect locations before and after cutting, and generate training data based on the matching results to fine-tune the pre-cutting detection module. The edge-cloud collaboration module is configured to dynamically adjust the upload threshold according to the cloud load, and to make a decision on local output or upload to the cloud based on the comparison between the inference confidence and the dynamic upload threshold. The hybrid deformation compensator is configured to use offline simulation results as a benchmark, correct the compensation amount in real time through force feedback, and back-calibrate the simulation model parameters based on the statistical deviation accumulated in batches. Among these modules, a cross-process information sharing and collaborative feedback loop is formed.

8. The machine vision-based automotive floor mat cutting and defect detection system according to claim 7, characterized in that, The hybrid deformation compensator includes: a PID controller whose control parameters are dynamically adjusted according to the variance of the compensation amount; and a simulation parameter calibration unit used to calibrate the hyperelastic model parameters through Bayesian optimization when batch deviation is triggered.

9. The machine vision-based automotive floor mat cutting and defect detection system according to claim 7, characterized in that, The edge-cloud collaboration module includes: a dynamic threshold calculation unit for calculating upload thresholds based on cloud load and network latency; a multi-exit early termination network for outputting inference confidence; and a knowledge distillation unit for periodically distilling cloud inference results to the edge model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images