Precision self-adaptive identification and calibration method for thickness measurement detector based on industrial vision

Abstract

The invention discloses a thickness measurement detector precision adaptive identification calibration method based on industrial vision, and relates to the technical field of industrial vision detection.The method comprises the specific steps that multi-modal data are synchronously collected according to a preset period, and after preprocessing and feature extraction, a lightweight causal graph and a thickness measurement precision prediction model are constructed; key influence factors and paths are identified through a multi-modal dynamic causal contribution degree algorithm, parameters are adjusted through a directional iterative algorithm, the parameters are updated through two-dimensional verification, and full-process data and scene information and an iterative optimization model are bound to be called by similar scenes; according to the method, a lightweight model is constructed to excavate precision deviation influence factors, parameters are accurately positioned and adjusted by relying on an exclusive algorithm, automatic identification and optimization are realized, a calibration result is ensured to be accurate and stable through a two-dimensional verification mechanism, a reusable scenarized scheme is formed, a data processing flow is standardized, the model can be dynamically adjusted, and stable support is provided for industrial thickness measurement.

Description

Technical Field

[0001] This invention relates to the field of industrial vision inspection technology, specifically to an adaptive recognition and calibration method for thickness measuring machines based on industrial vision. Background Technology

[0002] In industrial production, thickness inspection is a core component in ensuring product quality, directly impacting the stability of subsequent production processes and the final product's conformity. As the manufacturing industry transforms towards intelligent and efficient processes, higher demands are placed on the accuracy and adaptability of thickness inspection. Industrial vision technology, with its non-contact inspection advantages, has become the mainstream technology for thickness inspection. However, in actual production, dynamic changes in production conditions can significantly affect inspection accuracy. These changes involve multiple aspects such as environmental conditions, equipment operating status, and production rhythm. Failure to capture and address these influencing factors in a timely manner will lead to deviations in inspection accuracy, thereby affecting production efficiency and product quality. Therefore, achieving adaptive recognition and calibration of thickness inspection machine accuracy has become a key requirement for solving the problem of inspection accuracy stability in dynamic production scenarios and an important direction for promoting the upgrading of industrial inspection technology.

[0003] Traditional thickness gauge calibration methods often rely on manual experience for parameter adjustments, which not only consumes a lot of manpower but also suffers from low calibration efficiency and strong subjectivity, making it difficult to adapt to rapidly changing production rhythms. Some calibration methods only analyze a single type of data, failing to comprehensively cover the multiple influencing factors in the production process. This results in an inability to accurately identify the root cause of accuracy deviations, and parameter adjustments lack scientific basis, making it difficult to guarantee calibration results. In addition, traditional technologies lack effective analysis and verification of causal relationships, failing to clarify the transmission path between influencing factors and accuracy deviations. This makes the calibrated measuring machine lack accuracy stability when facing changes in production scenarios, and prone to deviations again. Furthermore, the verification methods of traditional calibration methods are relatively simple, focusing only on the accuracy results after calibration and ignoring the impact of dynamic changes in influencing factors on causal logic. This leads to insufficient sustainability and reliability of calibration results, making it difficult to meet the long-term stable high-precision testing requirements. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an adaptive recognition and calibration method for thickness measurement machines based on industrial vision. This method involves synchronously collecting and classifying multiple types of data through a preset acquisition cycle. After preprocessing and feature extraction, a lightweight causal graph model and a thickness measurement accuracy prediction model are constructed. A dedicated algorithm accurately identifies key influencing factors and transmission paths of accuracy deviations. A directional iterative algorithm is used to adjust model parameters, and dual-dimensional verification ensures the calibration effect and the validity of causal logic. By binding full-process data and scenario information, continuous iterative optimization is achieved, forming a reusable scenario-based solution. This enables adaptive calibration without manual intervention, improving the stability and efficiency of detection accuracy, and adapting to various dynamic industrial production scenarios.

[0005] To solve the above-mentioned technical problems, this invention provides the following technical solution: an adaptive recognition and calibration method for thickness measurement machines based on industrial vision, the specific steps of which are as follows: S1, Multimodal Data Acquisition: Based on the production batch, flow rate and detection efficiency, the acquisition cycle is preset, and traditional industrial vision images, multimodal sensing data, temperature and humidity and light source attenuation parameters are collected simultaneously. After being timestamped, the data is classified and stored in the raw data pool. S2, Feature Extraction and Model Building: Denoising, standardization preprocessing and feature extraction are performed on the original data to build a lightweight causal graph model for causal inference and a thickness measurement accuracy prediction model for thickness calculation. S3, Causal Identification and Parameter Adjustment: Based on the analysis results of the lightweight causal graph model, a multimodal dynamic causal contribution algorithm is used to identify the key influencing factors and root cause paths of the accuracy deviation; according to the identification results, the parameters of the thickness measurement accuracy prediction model are adjusted using a causal-driven parameter-oriented iterative algorithm. S4, Dual-dimensional verification: Deploy the adjusted thickness measurement accuracy prediction model on the inspection machine, collect actual thickness measurement data to verify the calibration effect, and at the same time verify whether the causal path of the key influencing factors is valid; update the model parameters according to the verification results and store the process data. S5, Data Storage and Model Iteration: Bind and store full-process data and scene information, iteratively optimize the lightweight causal graph model and thickness measurement accuracy prediction model and parameters, so that they can be directly called by similar scenarios.

[0006] Furthermore, the data acquisition long cycle is preset with a fixed duration or a fixed number of items to be inspected, and can be dynamically adjusted according to the switching of industrial production inspection batches, changes in workpiece turnover speed, and fluctuations in actual inspection efficiency. At the thickness measurement inspection position, traditional industrial vision images, multimodal sensing data, and temperature, humidity, and light source attenuation parameters are acquired simultaneously in the same position and from the same angle. The multimodal sensing data is at least one of hyperspectral imaging data, low coherence interference signals, and active polarized light images. The temperature and humidity parameters are continuously acquired in real time through the sensing module, and the light source attenuation parameter is the ratio of the actual output power of the light source in the industrial vision acquisition component to the rated power.

[0007] Furthermore, the denoising, standardization preprocessing, and feature extraction of the raw data are performed separately according to the data type. The raw data includes traditional industrial visual images, multimodal sensing data, temperature and humidity parameters, and light source attenuation parameters. Noise reduction processing is performed on image data, and targeted noise reduction methods are used for signal data. After all data is standardized and mapped to a unified numerical range, the edge contours, grayscale distribution, and contrast features of traditional industrial visual images are extracted, the specific features corresponding to multimodal sensing data are extracted, and the original parameter features of temperature, humidity, and light source attenuation are directly extracted, ultimately forming a deep scene feature set.

[0008] Furthermore, the specific steps for constructing the lightweight causal graph model include: selecting thickness measurement accuracy deviation data from recent historical production cycles, removing outliers, and integrating it with the depth scene feature set, quantized environmental and equipment parameters; setting the causal graph node types, including depth scene feature nodes, environmental parameter nodes, equipment status parameter nodes, and thickness measurement accuracy deviation nodes, with specific data sub-nodes under each node; establishing the association paths between sub-nodes, and calculating the degree of association between each node and the accuracy deviation node based on historical data, thereby assigning initial association weights to complete the causal graph construction.

[0009] Furthermore, the specific steps for constructing the thickness measurement accuracy prediction model for thickness calculation are as follows: Extract a historical sample set matching the current production scenario from the historical data pool. The historical sample set includes a depth scene feature set aligned in chronological order, a quantized environment and equipment parameter matrix, and the actual thickness value of the corresponding workpiece obtained through a standard measuring device as training labels. Fuse the depth scene feature set with the environment and equipment parameter matrix to form the model input feature vector. Use the actual thickness value as the supervised learning target and train the model using least squares regression. During training, determine the optimal hyperparameters through cross-validation to minimize the root mean square error between the predicted thickness and the actual thickness. Solidify the network weights and structural parameters obtained after training convergence into an initial thickness measurement accuracy prediction model. This model can receive real-time collected similar features and parameter inputs and output the corresponding thickness prediction value.

[0010] Furthermore, the mathematical expression of the multimodal dynamic causal contribution algorithm is: ;in Let be the dynamic causal contribution of the i-th depth scene features at time t to the j-th thickness measurement accuracy deviation. Let be the adaptive weight of the i-th type of multimodal feature at time t. Let be the standardized extracted value of the i-th type of depth scene feature at time t. Let be the interference attenuation coefficient for the j-th type of accuracy deviation at time t. Let be the combined environmental and device interference value corresponding to the j-th type of accuracy deviation at time t, and n be the total number of categories of multimodal depth scene features. It is the minimum value. This is the time-series decay factor.

[0011] Furthermore, when identifying the root cause path and influencing factors of the deviation: based on the causal contribution values ​​of each feature calculated by the multimodal dynamic causal contribution algorithm, key related nodes with contribution values ​​higher than the set contribution threshold are screened. By traversing the directed connections from the key related nodes to the thickness measurement accuracy deviation nodes in the lightweight causal graph model, all possible related transmission links are constructed and evaluated. Based on the weighted accumulation of the contribution values ​​of the upstream nodes in the link and the transmission path with the highest overall contribution value from the link, the root cause path of the deviation is determined. The data objects represented by all feature nodes and parameter nodes involved in the root cause path of the deviation are defined as key influencing factors.

[0012] Furthermore, the mathematical expression for the causal-driven parameter-oriented iterative algorithm is: ;in These are the core parameters of the calibration model for the j-th type of accuracy deviation at time t+1. Let be the initial parameters of the calibration model corresponding to the j-th type of accuracy deviation at time t. Adjust the step size coefficient for the parameters. Let be the causal contribution of the i-th type of depth scene features to the j-th type of bias at time t. Let be the difference in thickness measurement accuracy at time t under the new parameters and the original parameters. The preset standard thickness measurement accuracy threshold, To verify the feedback iteration factor.

[0013] Furthermore, in the calibration effect verification, the adjusted thickness measurement accuracy prediction model is run on the testing machine for a preset verification period. The actual thickness measurement data sequence generated within the verification period is collected, and the statistical distribution index of the actual thickness measurement data sequence is calculated and compared with the preset accuracy tolerance range. The statistical distribution index includes at least the average value, standard deviation and process capability index of the thickness measurement results. When all indicators meet the acceptance criteria defined by the accuracy tolerance range, the calibration effect verification of this model parameter adjustment is deemed to be successful.

[0014] Furthermore, during the verification period, real-time observation data of each key influencing factor determined by the root cause path are collected synchronously. By calculating the multi-dimensional deviation measure between the real-time observation data and the historical benchmark data used in the identification stage of step S3, it is determined whether the deviation value of each dimension is lower than the preset dynamic threshold for maintaining causal stability. If the deviation state of all key influencing factors is effectively controlled, it is determined that the identified causal path still holds in the current production environment, and the causal logic verification is passed.

[0015] Compared with existing technologies, this industrial vision-based thickness measurement machine accuracy adaptive recognition and calibration method has the following advantages: I. This invention utilizes multimodal data synchronous acquisition and classified storage, combined with targeted preprocessing and feature extraction, to construct a lightweight causal graph model and thickness measurement accuracy prediction model. This enables in-depth analysis of factors influencing accuracy deviations, accurately locating key influencing factors and transmission paths based on proprietary algorithms, and then adjusting model parameters through directional iterative algorithms. This breaks through the limitations of traditional calibration relying on experience or single data, making parameter adjustments more targeted and scientific. Automatic identification of influencing factors and model optimization can be achieved without manual intervention, significantly improving calibration efficiency and ensuring the accuracy stability of the testing machine in different production scenarios. It adapts to dynamic scenario requirements such as production batch switching and changes in flow speed, reducing production losses caused by accuracy deviations, providing more flexible and efficient accuracy assurance for industrial thickness measurement, and promoting the development of the testing process towards intelligence and adaptability.

[0016] Second, this invention employs a dual-dimensional verification mechanism to verify whether the calibration effect meets accuracy requirements and confirm the effectiveness of the causal path in the current production environment. This ensures the accuracy of the calibration results and the stability of the causal logic. By binding full-process data and scenario information, it continuously iterates and optimizes the model and parameter rules to form a reusable scenario-based solution, lowering the application threshold for similar scenarios. Through standardized feature extraction and model training processes, it standardizes data processing and calibration processes, avoids human error, and improves the consistency of the testing machine's operation. At the same time, the model can dynamically adjust according to real-time data, adapting to complex situations such as environmental fluctuations and equipment status changes, maintaining a high level of precision detection over the long term. This provides stable and reliable thickness measurement support for industrial production, helping production processes improve quality control capabilities, reduce defect rates, and enhance the controllability and efficiency of the overall production process.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. 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 drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 A flowchart illustrating the steps of an adaptive calibration method for thickness measurement machines based on industrial vision. Figure 2 This is a schematic diagram illustrating the flow of calibration data and model for a thickness gauge. Figure 3 A logic diagram for identifying the cause and effect of thickness measurement accuracy deviation and adjusting parameters. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Example 1: Thickness measurement and inspection scenario in cold-rolled steel sheet production.

[0022] Machinery manufacturing enterprises specialize in the mass production of cold-rolled steel sheets with thicknesses ranging from 2-5mm. Frequent batch changes occur during production, with each batch of 1000 steel sheets having a stable flow rate of 2m / s. The required inspection efficiency is no less than 500 sheets / hour. The workshop's temperature and humidity are easily affected by the external environment, and the light source of the industrial vision inspection components degrades over time, causing fluctuations in the accuracy of the thickness measurement machine. Therefore, the method of this invention is required for adaptive recognition and calibration. Specific steps are as follows: Figure 1 As shown.

[0023] S1, Multimodal Data Acquisition: Based on the current production batch size of 1000 pieces / batch steel plates with a turnover speed of 2m / s and a testing efficiency requirement of 500 pieces / hour, the preset data acquisition cycle is 0.5 seconds. At the thickness measurement and inspection location, a simultaneous acquisition method using the same position and viewing angle is employed, collecting three types of data to ensure comprehensive and synchronized information on the production scenario and equipment operation, laying the foundation for subsequent accurate analysis of precision deviations. The data includes: 1) traditional industrial vision images, used to capture the intuitive features of the steel plate surface and edges; 2) multimodal sensing data, selecting hyperspectral imaging data and low-coherence interference signals to comprehensively acquire in-depth sensing information related to steel plate thickness; and 3) environmental and equipment parameters, continuously acquiring workshop temperature and humidity parameters in real time through sensor modules, while simultaneously calculating the ratio of the actual output power to the rated power of the light source in the industrial vision acquisition component as a light source attenuation parameter, fully recording environmental changes and equipment operating status. All acquired data is timestamped and categorized by data type, stored in the raw data pool for convenient subsequent targeted processing and retrieval.

[0024] S2, Feature Extraction and Model Building: First, the raw data undergoes preprocessing and feature extraction to provide a unified analytical foundation for different data types and reduce the impact of interference factors on feature extraction. Noise reduction is performed on image data such as traditional industrial vision images, while targeted noise reduction methods are used for signal data such as low-coherence interference signals in hyperspectral imaging data to ensure the purity of all data types. All data are standardized and mapped to a unified numerical range to eliminate analytical obstacles caused by differences in data dimensions. Then, edge contour grayscale distribution and contrast features are extracted from traditional industrial vision images, and specific features corresponding to hyperspectral imaging data and low-coherence interference signals are extracted. Simultaneously, original parameter features of temperature, humidity, and light source attenuation are directly extracted. Finally, these features are integrated to form a deep scene feature set, providing comprehensive and effective feature support for subsequent model construction.

[0025] Next, a lightweight causal graph model was constructed to clearly present the potential correlation between various factors and accuracy deviation. Thickness measurement accuracy deviation data from the most recent three production cycles were selected, and outliers caused by sudden equipment failures were removed to avoid interference from causal relationship judgment. This data was then integrated with the quantized temperature, humidity, and light source attenuation data from the aforementioned depth scene feature set to form a complete analysis dataset. Four types of causal graph nodes were defined: depth scene feature nodes (including edge contour hyperspectral features, etc.), environmental parameter nodes (including temperature and humidity, equipment status parameter nodes (including light source attenuation, etc.), and thickness measurement accuracy deviation nodes. The correspondence between various influencing factors and deviation results was clarified. Association paths between each sub-node were established, and the degree of association between each node and the thickness measurement accuracy deviation node was calculated based on historical data. Initial association weights were assigned, completing the construction of the lightweight causal graph model and providing a clear analytical framework for subsequent causal identification.

[0026] Finally, a thickness measurement accuracy prediction model is constructed to achieve accurate thickness prediction based on historical valid data. A historical sample set matching the current cold-rolled steel plate production scenario with a flow rate of 2m / s and a thickness of 2-5mm is extracted from the historical data pool. This sample set includes a quantized matrix of temperature, humidity, and light source attenuation parameters from a depth scene feature set aligned in chronological order, as well as the corresponding actual thickness values ​​of the steel plates obtained through metrological calibration using standard measuring devices, serving as training labels to ensure the validity and accuracy of the training data. The depth scene feature set and the environmental equipment parameter matrix are fused to form the model input feature vector. Using the actual thickness value as the supervised learning target, the model is trained using the least squares regression method. During training, cross-validation is used to determine the optimal hyperparameters, thereby minimizing the root mean square error between the predicted and actual thicknesses and improving the model's prediction accuracy. The network weights and structural parameters obtained after training convergence are solidified to form the initial thickness measurement accuracy prediction model. This model can receive real-time collected similar features and parameter inputs and output the corresponding thickness prediction values, meeting real-time detection requirements.

[0027] S3, Causality Identification and Parameter Adjustment: A multimodal dynamic causal contribution algorithm is adopted. Based on the analysis results of the constructed lightweight causal graph model, the dynamic causal contribution of various depth scene features to different thickness measurement accuracy deviations at time t is calculated. The mathematical expression of the multimodal dynamic causal contribution algorithm is as follows: ;in Let be the dynamic causal contribution of the i-th depth scene features at time t to the j-th thickness measurement accuracy deviation. Let be the adaptive weight of the i-th type of multimodal feature at time t. Let be the standardized extracted value of the i-th type of depth scene feature at time t. Let be the interference attenuation coefficient for the j-th type of accuracy deviation at time t. Let be the combined environmental and device interference value corresponding to the j-th type of accuracy deviation at time t, and n be the total number of categories of multimodal depth scene features. It is the minimum value. As a time-series attenuation factor, it accurately identifies the key factors and transmission paths that play a crucial role in accuracy deviation, avoiding interference from irrelevant factors. Based on the calculation results, key related nodes exceeding a set contribution threshold are selected. By traversing the directed connections from these key related nodes to the thickness measurement accuracy deviation nodes in the lightweight causal graph model, all possible related transmission links are constructed. Based on the weighted sum of the contributions of upstream nodes in each link, the transmission path with the highest overall contribution is determined as the root cause path of the deviation. In this scenario, the thickness measurement accuracy deviation is identified as a change in hyperspectral imaging data characteristics due to the increase in light source attenuation parameters. The light source attenuation parameter nodes and hyperspectral imaging data feature nodes involved in this path are defined as key influencing factors, providing a clear direction for subsequent parameter adjustments.

[0028] Subsequently, the parameters of the thickness measurement accuracy prediction model were adjusted using a causal-driven parameter-oriented iterative algorithm. The mathematical expression of the causal-driven parameter-oriented iterative algorithm is as follows: ;in These are the core parameters of the calibration model for the j-th type of accuracy deviation at time t+1. Let be the initial parameters of the calibration model corresponding to the j-th type of accuracy deviation at time t. Adjust the step size coefficient for the parameters. Let be the causal contribution of the i-th type of depth scene features to the j-th type of bias at time t. Let be the difference in thickness measurement accuracy at time t under the new parameters and the original parameters. The preset standard thickness measurement accuracy threshold, To verify the feedback iteration factor and adapt the model to the accuracy requirements of the current production environment, the initial parameters of the calibration model at time t, corresponding to the current thickness measurement accuracy deviation, are used as a basis. This is combined with the causal contribution of key features at time t to the deviation, the pre-set standard thickness measurement accuracy threshold for the thickness measurement accuracy difference under the adaptation of new and original parameters at time t, and the verification feedback iteration factor. By comprehensively considering the effects of various influencing factors on parameter adjustment, the core parameters of the calibration model at time t+1 for this type of accuracy deviation are calculated. This completes the parameter-oriented adjustment of the thickness measurement accuracy prediction model, ensuring that the model parameters accurately match the current production scenario. Figure 3 As shown.

[0029] S4, Two-Dimensional Verification: The adjusted thickness measurement accuracy prediction model was deployed on a cold-rolled steel plate thickness measurement machine, with a preset verification cycle of 1 hour corresponding to the inspection of 500 products. During the verification cycle, actual thickness measurement data sequences were collected, and the statistical distribution indicators of these sequences, including the mean and standard deviation of the thickness measurement results and the process capability index, were calculated. These indicators were compared with the preset accuracy tolerance range. The accuracy of the adjusted model was comprehensively evaluated through statistical indicators to ensure that production accuracy standards were met. If all indicators met the acceptance criteria defined in the accuracy tolerance range, the calibration effect of this model parameter adjustment was deemed to have passed verification.

[0030] Simultaneously, during the verification period, real-time observation data of the hyperspectral imaging data characteristics of the light source attenuation parameters of key influencing factors are collected. The multi-dimensional deviation measure between this real-time data and the historical benchmark data used in the S3 identification stage is calculated. It is determined whether the deviation values ​​of each dimension are lower than the preset dynamic threshold to maintain causal stability. This confirms that the action path of the key influencing factors is still effective in the current scenario, ensuring the stability of the model. If the deviation state of all key influencing factors is effectively controlled, it is determined that the identified causal path is still valid in the current production environment, and the causal logic verification is passed.

[0031] S5, Data Storage and Model Iteration: The entire process data, including preprocessed feature data and model parameter verification results of the collected raw data, is bound and stored with current cold-rolled steel sheet production scenario information such as a flow rate of 2m / s, a batch size of 1000 pieces, and multimodal sensing data types. This allows subsequent similar scenarios to quickly access effective data and models. Based on the successful verification results, the lightweight causal graph model and thickness measurement accuracy prediction model, along with their parameter rules, are iteratively optimized. This iterative optimization allows the model accuracy to continuously improve with accumulated production experience. The optimized model and rules can be directly accessed in subsequent similar cold-rolled steel sheet production scenarios with the same thickness and specifications, similar flow rates, and inspection efficiencies, thereby improving the efficiency and accuracy of subsequent production inspections.

[0032] In summary, this scenario addresses the issue of fluctuating thickness measurement accuracy caused by frequent batch switching, temperature and humidity fluctuations, and light source attenuation in the mass production of cold-rolled steel sheets. It fully applies an adaptive recognition and calibration method for thickness measurement inspection machines based on industrial vision. By pre-setting the acquisition cycle according to production parameters, multiple types of data are collected simultaneously and stored in a categorized manner. After targeted preprocessing and feature extraction, a lightweight causal graph model and a thickness measurement accuracy prediction model are constructed. A multimodal dynamic causal contribution algorithm is used to identify key influencing factors and paths. Then, a causal-driven parameter-oriented iterative algorithm is used to adjust the model parameters. Dual-dimensional verification ensures the calibration effect and the validity of the causal logic. Finally, the stored data is bound and the model is iteratively optimized. The entire process precisely adapts to production needs, effectively improving thickness measurement accuracy and inspection efficiency. The optimized model can directly serve similar scenarios.

[0033] Example 2: Thickness testing scenario in the production of food-grade PE plastic film.

[0034] A packaging material factory produces food-grade PE plastic film with a thickness range of 0.05-0.1mm. During production, the film turnover speed is relatively fast (5m / s), requiring an inspection efficiency of 1000 pieces / hour, with large batch sizes of 2000 pieces per batch. PE plastic film is susceptible to slight deformation due to environmental humidity. The attenuation of the light source in industrial vision inspection affects image clarity, leading to deviations in the accuracy of the thickness measurement machine. Therefore, the method of this invention is required for adaptive calibration. Specific steps are as follows... Figure 2 As shown.

[0035] S1, Multimodal Data Acquisition: Based on a production batch size of 2000 pieces / batch, a film turnover speed of 5 m / s, and a testing efficiency requirement of 1000 pieces / hour, the preset data acquisition cycle is 0.3 seconds. A simultaneous acquisition method using the same position and viewing angle is employed at the thickness measurement location, simultaneously acquiring four types of data. This simultaneous acquisition of multiple data types comprehensively captures environmental changes, equipment status, and product characteristics during the film production process, providing comprehensive data support for subsequent analysis. Traditional industrial vision images are used to capture film surface flatness and edge features; multimodal sensing data utilizes actively polarized light images and low-coherence interference signals to accurately obtain film thickness-related sensing information; temperature and humidity parameters are continuously acquired in real-time via a sensing module; and light source attenuation parameters are calculated to determine the ratio of the actual output power to the rated power of the light source in the industrial vision acquisition component. All acquired data is timestamped and categorized by data type, stored in the raw data pool, providing an ordered data foundation for subsequent preprocessing and model building.

[0036] S2, Feature Extraction and Model Building: First, the raw data undergoes preprocessing and feature extraction, with targeted processing based on data type to retain effective feature information and remove noise interference, resulting in more accurate feature extraction. Noise reduction is applied to image data such as traditional industrial vision images and actively polarized light images, while targeted noise reduction methods are used for signal data such as low-coherence interference signals, ensuring the quality of different data types. All data is standardized and mapped to a unified numerical range to eliminate analytical obstacles caused by data differences. Then, edge contour grayscale distribution contrast features are extracted from traditional industrial vision images, and specific features from actively polarized light images and low-coherence interference signals are extracted. Original parameter features related to temperature, humidity, and light source attenuation are directly extracted and integrated to form a deep scene feature set, providing comprehensive and high-quality feature input for model construction.

[0037] A lightweight causal graph model was constructed, integrating historical deviation data with various feature parameters to clarify the correlation strength between nodes. Thickness measurement accuracy deviation data from the last five production cycles was selected, outliers were removed, and then integrated with quantized temperature, humidity, and light source attenuation data from the depth scene feature set to ensure the integrity and reliability of the dataset. Node types were defined, including depth scene feature nodes (containing active polarized light feature grayscale distribution, etc.), environmental parameter nodes (containing temperature and humidity, equipment status parameter nodes (containing light source attenuation, etc.), and thickness measurement accuracy deviation nodes. Each node had specific data sub-nodes, clearly identifying various influencing factors. Correlation paths between sub-nodes were established, and the correlation degree between each node and the accuracy deviation node was calculated based on historical data, assigning initial correlation weights to complete the causal graph construction and providing a clear analytical framework for causal identification.

[0038] A thickness measurement accuracy prediction model is constructed and trained based on historical samples matching the current PE plastic film production scenario to ensure the accuracy of thickness prediction for thin products such as films. A historical sample set matching the current PE plastic film production scenario with a flow rate of 5m / s and a thickness specification of 0.05-0.1mm is extracted from the historical data pool. This sample set includes a time-sequential aligned depth scene feature set, a quantized environmental and equipment parameter matrix, and actual thickness values ​​measured by a standard measuring device as training labels. The feature set and parameter matrix are fused to form an input feature vector. Using the actual thickness value as the supervision target, the model is trained using the least squares regression method. Optimal hyperparameters are determined through cross-validation to minimize the root mean square error between the predicted and actual thicknesses, thus improving the model's prediction accuracy. The network weights and structural parameters after training convergence are solidified to form an initial thickness measurement accuracy prediction model that meets the accuracy requirements for real-time film thickness measurement.

[0039] S3, Causality Identification and Parameter Adjustment: A multimodal dynamic causal contribution algorithm is employed to calculate the dynamic causal contribution of various depth scene features to different thickness measurement accuracy deviations. The mathematical expression of the multimodal dynamic causal contribution algorithm is as follows: ;in Let be the dynamic causal contribution of the i-th depth scene features at time t to the j-th thickness measurement accuracy deviation. Let be the adaptive weight of the i-th type of multimodal feature at time t. Let be the standardized extracted value of the i-th type of depth scene feature at time t. Let be the interference attenuation coefficient for the j-th type of accuracy deviation at time t. Let be the combined environmental and device interference value corresponding to the j-th type of accuracy deviation at time t, and n be the total number of categories of multimodal depth scene features. It is the minimum value. As a time-series decay factor, the core factors and transmission paths affecting the accuracy of thin film thickness measurement are precisely screened to avoid interference caused by the deformable characteristics of thin films. Key associated nodes above a threshold are selected, and directed connections from key associated nodes to accuracy deviation nodes are traversed in the lightweight causal graph to construct all possible transmission links. Based on the weighted sum of the contributions of upstream nodes in the link, the root cause path with the highest overall contribution is determined. In this scenario, the thickness measurement accuracy deviation is due to changes in active polarized light image features caused by increased environmental humidity. The environmental humidity parameter node and the active polarized light image feature node are defined as key influencing factors, clarifying the core direction of parameter adjustment.

[0040] A causal-driven parameter-oriented iterative algorithm is used to optimize the parameters of the thickness measurement accuracy prediction model based on the influence of key factors, enabling the model to adapt to the accuracy detection requirements under rapid film flow. The mathematical expression of the causal-driven parameter-oriented iterative algorithm is as follows: ;in These are the core parameters of the calibration model for the j-th type of accuracy deviation at time t+1. Let be the initial parameters of the calibration model corresponding to the j-th type of accuracy deviation at time t. Adjust the step size coefficient for the parameters. Let be the causal contribution of the i-th type of depth scene features to the j-th type of bias at time t. Let be the difference in thickness measurement accuracy at time t under the new parameters and the original parameters. The preset standard thickness measurement accuracy threshold, To verify the feedback iteration factor, based on the initial parameters of the calibration model corresponding to the accuracy deviation at time t, and combined with the causal contribution of the thickness measurement accuracy difference standard thickness measurement accuracy threshold at time t and the verification feedback iteration factor, the core parameters of the calibration model at time t+1 are calculated by comprehensively considering various key influencing factors. This completes the directional adjustment of the thickness measurement accuracy prediction model parameters, ensuring that the model can accurately adapt to the rapid turnover scenario of thin film production.

[0041] S4, Two-Dimensional Verification: The adjusted model was deployed on a PE plastic film thickness measuring machine, with a verification cycle of 1 hour corresponding to the inspection of 1000 products. Actual thickness measurement data sequences were collected within the verification cycle, and statistical distribution indicators such as the mean, standard deviation, and process capability index were calculated. These were compared with the preset accuracy tolerance range. Through multi-statistical evaluation, it was ensured that the adjusted model could meet the requirements for high-precision film thickness measurement. If all indicators met the acceptance criteria, the calibration effect was verified as successful.

[0042] Real-time observation data of key influencing factors such as environmental humidity and active polarized light image features are collected synchronously. Multi-dimensional deviation measures from historical benchmark data are calculated. If all deviation values ​​are below the dynamic threshold, the causal path is determined to be valid and the causal logic is verified. This ensures the stability of the action path of key factors and avoids the impact of fluctuations in environmental humidity and other factors on the reliability of the model.

[0043] S5, Data Storage and Model Iteration: The system integrates end-to-end data storage with current PE plastic film production scenario information, including a 5m / s turnover rate, 2000 pieces / batch, and multimodal sensing data types, eliminating the need for repetitive modeling for similar film production scenarios. Based on validation results, a lightweight causal graph model for thickness measurement accuracy prediction and its parameter rules are iteratively optimized. The iterated model continuously adapts to subtle changes in the production process, maintaining high-precision detection levels. The optimized model and rules can be directly used by subsequent similar food-grade PE plastic film production scenarios.

[0044] In summary, this application addresses the challenges of rapid production of food-grade PE plastic films, including susceptibility to environmental humidity and light source attenuation interference, and strictly adheres to the entire calibration process. The data acquisition cycle is set based on parameters such as production batch and production speed. Multi-dimensional data is collected simultaneously and stored systematically. Preprocessing and feature extraction are performed according to data type. A dual-model adapted for thin product testing is constructed. A multimodal dynamic causal contribution algorithm is used to identify core influencing paths. Model parameters are optimized through a causal-driven parameter-oriented iterative algorithm. Dual-dimensional verification ensures calibration effectiveness and causal stability. Finally, scenario-specific information is bound to the stored data, and the model is iterated. This application is precisely adapted to the rapid film production scenario, effectively solving the accuracy deviation problem and providing a reusable solution for efficient and high-precision thickness measurement of similar films.

[0045] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for adaptive recognition and calibration of thickness measurement machines based on industrial vision, characterized in that, The specific steps of this method are as follows: S1, Multimodal Data Acquisition: Based on the production batch, flow rate and detection efficiency, the acquisition cycle is preset, and traditional industrial vision images, multimodal sensing data, temperature and humidity and light source attenuation parameters are collected simultaneously. After being timestamped, the data is classified and stored in the raw data pool. S2, Feature Extraction and Model Building: Denoising, standardization preprocessing and feature extraction are performed on the original data to build a lightweight causal graph model for causal inference and a thickness measurement accuracy prediction model for thickness calculation. S3, Causal Identification and Parameter Adjustment: Based on the analysis results of the lightweight causal graph model, a multimodal dynamic causal contribution algorithm is used to identify the key influencing factors and root cause paths of the accuracy deviation. Based on the identification results, the parameters of the thickness measurement accuracy prediction model are adjusted using a causal-driven parameter-oriented iterative algorithm. S4, Dual-dimensional verification: Deploy the adjusted thickness measurement accuracy prediction model on the inspection machine, collect actual thickness measurement data to verify the calibration effect, and at the same time verify whether the causal path of the key influencing factors is valid; update the model parameters according to the verification results and store the process data. S5, Data Storage and Model Iteration: Bind and store full-process data and scene information, iteratively optimize the lightweight causal graph model and thickness measurement accuracy prediction model and parameters, so that they can be directly called by similar scenarios.

2. The method for adaptive recognition and calibration of thickness measurement machine accuracy based on industrial vision according to claim 1, characterized in that, In step S2, the long period of data acquisition is preset with a fixed duration or a fixed number of items to be inspected, and can be dynamically adjusted according to the switching of industrial production inspection batches, changes in workpiece turnover speed, and fluctuations in actual inspection efficiency. At the thickness measurement inspection position, traditional industrial vision images, multimodal sensing data, and temperature, humidity, and light source attenuation parameters are acquired simultaneously in the same position and from the same angle. The multimodal sensing data is at least one of hyperspectral imaging data, low coherence interference signals, and active polarized light images. The temperature and humidity parameters are continuously acquired in real time through the sensing module, and the light source attenuation parameter is the ratio of the actual output power of the light source in the industrial vision acquisition component to the rated power.

3. The method for adaptive recognition and calibration of thickness measurement machines based on industrial vision according to claim 1, characterized in that, In step S2, the noise reduction, standardization preprocessing, and feature extraction of the original data are performed separately according to the data type. The original data includes traditional industrial vision images, multimodal sensing data, temperature and humidity parameters, and light source attenuation parameters. Noise reduction processing is performed on image data, and targeted noise reduction methods are used for signal data. After all data is standardized and mapped to a unified numerical range, the edge contours, grayscale distribution, and contrast features of traditional industrial vision images are extracted, the specific features corresponding to multimodal sensing data are extracted, and the original parameter features of temperature, humidity, and light source attenuation are directly extracted, ultimately forming a deep scene feature set.

4. The method for adaptive recognition and calibration of thickness measurement machine accuracy based on industrial vision according to claim 1, characterized in that, In step S2, the specific steps for constructing the lightweight causal graph model include: selecting thickness measurement accuracy deviation data from recent historical production cycles, removing outliers, and integrating it with the depth scene feature set, quantized environmental and equipment parameters; setting the causal graph node types, including depth scene feature nodes, environmental parameter nodes, equipment status parameter nodes, and thickness measurement accuracy deviation nodes, with specific data sub-nodes under each node; establishing the association paths between sub-nodes, and calculating the degree of association between each node and the accuracy deviation node based on historical data, thereby assigning initial association weights to complete the causal graph construction.

5. The method for adaptive recognition and calibration of thickness measurement machine accuracy based on industrial vision according to claim 1, characterized in that, In step S3, the specific steps for constructing the thickness measurement accuracy prediction model for thickness calculation are as follows: Extract a historical sample set matching the current production scenario from the historical data pool. The historical sample set includes a depth scene feature set aligned in chronological order, a quantized environment and equipment parameter matrix, and the actual thickness value of the corresponding workpiece obtained through a standard measuring device as training labels. Fuse the depth scene feature set with the environment and equipment parameter matrix to form the model input feature vector. Use the actual thickness value as the supervised learning target and train the model using the least squares regression method. During training, determine the optimal hyperparameters through cross-validation to minimize the root mean square error between the predicted thickness and the actual thickness. Solidify the network weights and structural parameters obtained after training convergence into an initial thickness measurement accuracy prediction model. This model can receive real-time collected similar features and parameter inputs and output the corresponding thickness prediction value.

6. The method for adaptive recognition and calibration of thickness measurement machine accuracy based on industrial vision according to claim 1, characterized in that, In step S3, the mathematical expression of the multimodal dynamic causal contribution algorithm is: ;in Let be the dynamic causal contribution of the i-th depth scene features at time t to the j-th thickness measurement accuracy deviation. Let be the adaptive weight of the i-th type of multimodal feature at time t. Let be the standardized extracted value of the i-th type of depth scene feature at time t. Let be the interference attenuation coefficient for the j-th type of accuracy deviation at time t. Let be the combined environmental and device interference value corresponding to the j-th type of accuracy deviation at time t, and n be the total number of categories of multimodal depth scene features. It is the minimum value. This is the time-series decay factor.

7. The method for adaptive recognition and calibration of thickness measurement machine accuracy based on industrial vision according to claim 1, characterized in that, In step S3, when identifying the root cause path and influencing factors of the deviation: based on the causal contribution values ​​of each feature calculated by the multimodal dynamic causal contribution algorithm, key related nodes with a contribution value higher than the set threshold are screened. By traversing the directed connections from the key related nodes to the thickness measurement accuracy deviation nodes in the lightweight causal graph model, all possible related transmission links are constructed and evaluated. Based on the weighted summation of the contribution values ​​of the upstream nodes in the link and the transmission path with the highest overall contribution value from the link, the root cause path of the deviation is determined. The data objects represented by all feature nodes and parameter nodes involved in the root cause path of the deviation are defined as key influencing factors.

8. The method for adaptive recognition and calibration of thickness measurement machine accuracy based on industrial vision according to claim 1, characterized in that, In step S3, the mathematical expression of the causal-driven parameter-oriented iterative algorithm is: ;in These are the core parameters of the calibration model for the j-th type of accuracy deviation at time t+1. Let be the initial parameters of the calibration model corresponding to the j-th type of accuracy deviation at time t. Adjust the step size coefficient for the parameters. Let be the causal contribution of the i-th type of depth scene features to the j-th type of bias at time t. Let be the difference in thickness measurement accuracy at time t under the new parameters and the original parameters. The preset standard thickness measurement accuracy threshold, To verify the feedback iteration factor.

9. The method for adaptive recognition and calibration of thickness measurement machine accuracy based on industrial vision according to claim 1, characterized in that, In step S4, during the calibration effect verification, the adjusted thickness measurement accuracy prediction model is run on the testing machine for a preset verification period. The actual thickness measurement data sequence generated within the verification period is collected, and the statistical distribution index of the actual thickness measurement data sequence is calculated and compared with the preset accuracy tolerance range. The statistical distribution index includes at least the average value, standard deviation, and process capability index of the thickness measurement results. When all indicators meet the acceptance criteria defined by the accuracy tolerance range, the calibration effect verification of this model parameter adjustment is deemed to be successful.

10. The method for adaptive recognition and calibration of thickness measurement machine accuracy based on industrial vision according to claim 1, characterized in that, In step S4, real-time observation data of each key influencing factor determined by the root cause path are collected synchronously during the verification period. By calculating the multi-dimensional deviation measure between the real-time observation data and the historical benchmark data used in the identification stage of step S3, it is determined whether the deviation value of each dimension is lower than the preset dynamic threshold for maintaining causal stability. If the deviation state of all key influencing factors is effectively controlled, it is determined that the identified causal path still holds in the current production environment, and the causal logic verification is passed.

Images