Search

CN-121980355-A - Industrial vision-based thickness measuring detector precision self-adaptive identification and calibration method

CN121980355ACN 121980355 ACN121980355 ACN 121980355ACN-121980355-A

Abstract

The invention discloses a thickness measuring detector precision self-adaptive identification and calibration method based on industrial vision, which relates to the technical field of industrial vision detection and comprises the following specific steps of synchronously acquiring multi-mode data according to a preset period, constructing a lightweight causal chart and a thickness measuring precision prediction model after preprocessing and feature extraction, identifying key influence factors and paths through a multi-mode dynamic causal contribution algorithm, adjusting parameters by utilizing a directional iterative algorithm, updating parameters through two-dimensional verification, binding full-flow data and scene information, and iteratively optimizing the model for similar scene call; according to the method, the light weight model is constructed to excavate the precision deviation influencing factors, the parameters are accurately positioned and adjusted by means of the exclusive algorithm, automatic identification and optimization are realized, the accuracy and stability of the calibration result are ensured by a two-dimensional verification mechanism, a reusable scene scheme is formed, the data processing flow is standardized, the model can be dynamically adjusted, and stable support is provided for industrial thickness measurement.

Inventors

  • ZHANG WENQUAN

Assignees

  • 青岛新宝嘉扬精工机械有限公司

Dates

Publication Date
20260505
Application Date
20260205

Claims (10)

  1. 1. The method for adaptively identifying and calibrating the accuracy of the thickness measuring detector based on industrial vision is characterized by comprising the following specific steps of: S1, multi-mode data acquisition, namely presetting an acquisition period according to production batches, circulation speed and detection efficiency, synchronously acquiring traditional industrial visual images, multi-mode sensing data, temperature and humidity and light source attenuation parameters, and classifying and storing the marked data into an original data pool after time stamping; S2, feature extraction and model construction, namely carrying out noise reduction, standardization pretreatment and feature extraction on original data, and constructing a lightweight causal graph model for causal reasoning and a thickness measurement precision prediction model for thickness calculation; S3, causal identification and parameter adjustment, namely identifying key influence factors and deviation root cause paths which cause precision deviation by adopting a multi-mode dynamic causal contribution algorithm based on the analysis result of the lightweight causal graph model; S4, two-dimensional verification, namely deploying the adjusted thickness measurement precision prediction model on a detector, collecting actual thickness measurement data to perform calibration effect verification, and verifying whether a causal path of the key influence factors is established or not; And S5, data storage and model iteration, namely binding and storing full-flow data and scene information, and iteratively optimizing the lightweight causal graph model and thickness measurement precision prediction model and parameters for direct call of similar scenes.
  2. 2. The method is characterized in that in the step S2, the data acquisition long period is preset according to fixed time length or fixed detection number, and can be dynamically adjusted according to the switching of industrial production detection batches, the change of workpiece circulation speed and the fluctuation of actual detection efficiency, traditional industrial visual images, multi-mode sensing data and at least one of temperature and humidity and light source attenuation parameters are synchronously acquired at the thickness detection position in an acquisition mode of the same viewing angle at the same position, the multi-mode 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 a sensing module, and the light source attenuation parameters are the ratio of the actual output power to the rated power of a light source in an industrial visual acquisition assembly.
  3. 3. The method is characterized in that in the step S2, the noise reduction, standardization pretreatment and feature extraction are respectively carried out on original data, the original data comprise traditional industrial visual images, multi-mode sensing data, temperature and humidity parameters and light source attenuation parameters, the noise reduction treatment is carried out on image data, the targeted noise reduction method is adopted on signal data, after all the data are mapped to a unified numerical range in a standardized mode, edge contours, gray scale distribution and contrast features of the traditional industrial visual images are respectively extracted, exclusive features corresponding to the multi-mode sensing data are extracted, the original parameter features of temperature and humidity and light source attenuation are directly extracted, and finally a depth scene feature set is formed.
  4. 4. The method is characterized in that in the step S2, the specific steps of constructing a lightweight causal graph model comprise the steps of selecting thickness measurement precision deviation data in a recent historical production period, integrating the thickness measurement precision deviation data with a depth scene feature set and quantized environment and equipment parameters after abnormal values are removed, setting causal graph node types, including depth scene feature nodes, environment parameter nodes, equipment state parameter nodes and thickness measurement precision deviation nodes, setting specific data sub-nodes under each node, establishing a correlation path between the sub-nodes, calculating the correlation degree of each node and the precision deviation nodes based on the historical data, giving initial correlation weights, and completing causal graph construction.
  5. 5. The method is characterized in that in the step S3, the specific steps of constructing the thickness measurement precision prediction model for thickness calculation are that a historical sample set matched with a current production scene is extracted from a historical data pool, the historical sample set comprises a depth scene feature set, quantized environment and equipment parameter matrixes and real thickness values of corresponding workpieces obtained through a standard measuring device, which are aligned in time sequence, are used as training labels, the depth scene feature set and the environment equipment parameter matrixes are fused to form model input feature vectors, the real thickness values are used as supervision learning targets, model training is conducted through a least square regression method, optimal super parameters are determined through cross verification in the training process, root mean square errors between predicted thickness and real thickness are minimized, network weights and structural parameters obtained after training convergence are solidified into an initial thickness measurement precision prediction model, and the model can receive similar features and parameter inputs acquired in real time and output corresponding thickness prediction values.
  6. 6. The method for calibrating precision adaptive recognition of a thickness measuring detector based on industrial vision according to claim 1, wherein in step S3, the mathematical expression of the multi-modal dynamic causal contribution algorithm is: Wherein The dynamic causal contribution degree of the ith depth scene feature to the jth thickness measurement precision deviation at the moment t, As the adaptive weight of the i-th type multi-modal feature at the time t, The normalized extracted value of the i-th class depth scene feature at the moment t, The interference attenuation coefficient of the j-th type precision deviation at the t moment, For the environment-equipment comprehensive interference value corresponding to the j-th type precision deviation at the t moment, n is the total type number of the multi-mode depth scene characteristics, At the level of the minimum value of the total number of the components, Is a timing decay factor.
  7. 7. The method for adaptively identifying and calibrating accuracy of thickness measuring detector based on industrial vision according to claim 1, wherein in step S3, when the deviation root cause path and the influence factors are identified, key associated nodes higher than a set contribution threshold are screened based on each characteristic causal contribution value calculated by a multi-mode dynamic causal contribution algorithm, all possible associated conductive links are constructed and evaluated by traversing the directed connection from the key associated nodes to the thickness measuring accuracy deviation nodes in the lightweight causal graph model, a conductive path with the highest overall contribution is determined from the links according to the weighted accumulation of the contribution of the nodes upstream of the links as the deviation root cause path, and all the characteristic nodes and the data objects characterized by the parameter nodes involved in the deviation root cause path are defined as key influence factors.
  8. 8. The method for adaptively identifying and calibrating precision of a thickness measuring detector based on industrial vision according to claim 1, wherein in step S3, the mathematical expression of the causal driving parameter orientation iterative algorithm is: Wherein The core parameters of the calibration model aiming at the j-th class precision deviation at the time t+1, The initial parameters of the calibration model corresponding to the j-th type precision deviation at the t moment, The step size coefficients are adjusted for the parameters, The causal contribution degree of the ith class depth scene characteristic to the jth class deviation at the moment t, The thickness measurement precision difference value between the new parameter at the time t and the original parameter is adopted, Is a preset standard thickness measurement accuracy threshold value, To verify the feedback iteration factor.
  9. 9. The method according to claim 1, wherein in step S4, in the calibration effect verification, by running the adjusted thickness measurement accuracy prediction model on the detector for a preset verification period, collecting an actual thickness measurement data sequence generated in the verification period, calculating a statistical distribution index of the actual thickness measurement data sequence, and comparing the statistical distribution index with a preset accuracy tolerance interval, wherein the statistical distribution index at least comprises an average value, a standard deviation and a process capability index of a thickness measurement result, and when all indexes meet an acceptance criterion defined by the accuracy tolerance interval, determining that the calibration effect verification of the current model parameter adjustment is passed.
  10. 10. The method according to claim 1, wherein in step S4, real-time observation data of each key influencing factor determined by the root cause path is collected synchronously in the verification period, a multi-dimensional deviation metric between the real-time observation data and the historical reference data based on the identification stage in step S3 is calculated, whether the deviation value of each dimension is lower than a dynamic threshold preset for maintaining causal stability is determined, and if the deviation states of all key influencing factors are effectively controlled, it is determined that the identified causal path still holds in the current production environment, and causal logic verification is passed.

Description

Industrial vision-based thickness measuring detector precision self-adaptive identification and calibration method Technical Field The invention relates to the technical field of industrial vision detection, in particular to a precision self-adaptive identification and calibration method of a thickness measuring detector based on industrial vision. Background In industrial production, thickness detection is one of the key links for guaranteeing product quality, the stability of a subsequent production process and the qualification of a final product are directly related, along with the change of the manufacturing industry to intellectualization and high efficiency, higher requirements are provided for the accuracy and the adaptability of thickness detection, industrial vision technology has become the main technical direction of thickness detection by virtue of non-contact detection, but in the actual production process, dynamic changes of production conditions can have significant influence on the detection accuracy, and the changes relate to multiple aspects such as environmental conditions, equipment running states, production rhythm and the like, and if the influence factors cannot be captured and dealt with in time, the detection accuracy is deviated, so that the production efficiency and the product quality are influenced, therefore, the self-adaptive identification and the calibration of the accuracy of a thickness detection machine are realized, the key requirements for solving the problem of the stability of the detection accuracy in a dynamic production scene are also important directions for promoting the upgrade of the industrial detection technology. The traditional thickness measuring detector has the advantages that the traditional thickness measuring detector is easy to calibrate by relying on manual experience to carry out parameter adjustment, a large amount of labor cost is consumed, the problems of low calibration efficiency and strong subjectivity exist, the quick-change production rhythm is difficult to adapt, the partial calibration method is only analyzed based on single type data, multiple influencing factors in the production process cannot be comprehensively covered, the root cause of precision deviation cannot be accurately identified, the parameter adjustment lacks scientific basis, the calibration effect is difficult to guarantee, in addition, the traditional technology lacks effective analysis and verification of causality, the conduction path between influencing factors and the precision deviation cannot be clearly influenced, the calibrated detector is insufficient in precision stability when facing to production scene change, deviation is easy to appear again, meanwhile, the traditional calibration method is relatively single in verification mode, only focuses on the calibrated precision result, the influence of dynamic change of influencing factors on causal logic is ignored, the durability and reliability of the calibration effect are insufficient, and long-term stable high-precision detection requirements are difficult to meet. Disclosure of Invention The invention aims to make up the defects of the prior art, and provides an industrial vision-based thickness measuring detector precision self-adaptive identification calibration method, which is characterized in that a lightweight causal graph model and a thickness measuring precision prediction model are constructed after preprocessing and feature extraction by synchronously acquiring multiple types of data through a preset acquisition period, key influencing factors and conducting paths of precision deviation are accurately identified by means of a proprietary algorithm, model parameters are adjusted by means of a directional iterative algorithm, calibration effect and causal logic effectiveness are ensured by means of two-dimensional verification, continuous iterative optimization of full-flow data and scene information is bound, a reusable scene scheme is formed, self-adaptive calibration without manual intervention is realized, the stability and efficiency of detection precision are improved, and various dynamic industrial production scenes are adapted. The invention provides a method for adaptively identifying and calibrating the accuracy of a thickness measuring detector based on industrial vision, which comprises the following specific steps of: S1, multi-mode data acquisition, namely presetting an acquisition period according to production batches, circulation speed and detection efficiency, synchronously acquiring traditional industrial visual images, multi-mode sensing data, temperature and humidity and light source attenuation parameters, and classifying and storing the marked data into an original data pool after time stamping; S2, feature extraction and model construction, namely carrying out noise reduction, standardization pretreatment and feature extraction on original data, and constructing a lightweight