Cap pressing and detection integrated control method for metal terminal

Abstract

The present application belongs to the technical field of specific computer models, and discloses a metal terminal cap pressing and detection integrated control method, comprising the following steps: S1, collecting the front appearance image of the metal terminal cap, the key size data of the cap installation area and the initial weight data, and synchronously completing the temperature and humidity environment interference compensation and calibration of the weight data through a multi-factor regression algorithm; S2, generating an adaptive cap pressure threshold and pressure retention time through a parameter prediction model; S3, applying pressure to the metal terminal cap installation reference position according to the adaptive parameters output by step S2, and monitoring the dynamic change of the pressure in real time; S4, identifying the newly generated defect features in the cap pressing process through a cap post-product detection model, verifying the overall size and cap fitting accuracy, comparing the actual weight with the standard weight threshold, and bidirectionally determining the screw locking state in combination with the torque data and image texture analysis; and S5, fusing the comprehensive detection results of step S1 and step S4, and outputting three types of determination result rates.

Description

An integrated control method for clamping and detecting metal terminals Technical Field

[0001] This invention relates to the field of specific computer modeling technology, and specifically to an integrated control method for pressing and detecting metal terminals. Background Technology

[0002] As a core component in the field of electrical connections, the assembly quality of metal terminals during the crimping process directly determines the stability, reliability, and service life of electrical connections, and they are widely used in key areas such as automotive, electronics, and industrial control. With the development of end products towards higher precision, miniaturization, and higher reliability, stringent requirements have been placed on the assembly accuracy, defect control, and consistency of metal terminal crimping caps.

[0003] The existing metal terminal capping and inspection process suffers from a core technical pain point: the capping parameters and inspection results lack a closed-loop linkage mechanism based on a specific calculation model, making it difficult to accurately control capping quality. In traditional processes, capping parameters are mostly set manually based on experience, simply referring to the terminal's nominal specifications without considering individual dimensional deviations, material characteristics, and historical quality data for intelligent adaptation. Furthermore, inspections before and after capping often rely on traditional algorithms for defect identification and data processing, making it difficult to deeply extract minute defect features and multi-source data correlation information. The inspection results also cannot drive real-time optimization of capping parameters. This results in low matching between capping pressure and holding time, easily leading to assembly defects such as capping misalignment, loosening, and damage. These similar defects are recurring and difficult to eradicate, failing to meet the consistency requirements of high-precision production for capping quality.

[0004] Although some automated cap-pressing or inspection equipment exists in existing technologies, most fail to deeply integrate inspection data, historical feedback, and cap-pressing parameters through specific computational models, making it difficult to achieve intelligent closed-loop control throughout the entire process. Therefore, an integrated control method is urgently needed. Leveraging the advantages of computer systems based on specific computational models, and through an end-to-end intelligent processing mode, dynamic linkage optimization of cap-pressing parameters and inspection results can be achieved, solving the core problems of poor cap-pressing quality consistency and high defect rate in traditional processes. Summary of the Invention

[0005] The purpose of this invention is to provide an integrated control method for the capping and detection of metal terminals, which solves the problems of poor quality consistency and high defect rate of the capping.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for integrated control of clamping and detection of metal terminals includes the following steps:

[0008] S1. Collect the appearance image of the metal terminal before the cap is pressed, the key dimension data of the cap installation area and the initial weight data. Identify the inherent defect characteristics through the pre-processing detection model before the cap is pressed, accurately measure the reference dimensions of the cap installation area, and simultaneously complete the temperature and humidity environmental interference compensation and calibration of the weight data through a multi-factor regression algorithm.

[0009] S2. Based on the preprocessing detection results of step S1, the preset specifications of the metal terminals, and the historical cap quality feedback data, the appropriate cap pressure threshold and pressure holding time are generated through the parameter prediction model.

[0010] S3. Apply pressure to the mounting reference position of the metal terminal cap according to the adaptation parameters output in step S2, monitor the dynamic changes of pressure in real time, and complete the capping operation when the pressure reaches the preset threshold and is maintained stably for a preset time.

[0011] S4. Collect finished product appearance images, overall key dimension data, actual weight data and screw torque data of the metal terminal after pressing. Identify new defect features generated during the pressing process through the finished product inspection model after pressing. Verify the overall size and pressing accuracy. Compare the actual weight with the standard weight threshold. Combine torque data and image texture analysis to determine the screw tightening status in both directions.

[0012] S5. Combine the preprocessing test results of step S1 with the comprehensive test results of step S4, and output three types of judgment results: qualified, cap parameter unqualified, and defect unqualified. If the cap parameter is unqualified, the parameter optimization instruction is triggered. If the defect is unqualified, the defect type and location are clearly marked.

[0013] S6. Link and store the full process output results of steps S1 to S5 with the metal terminal product number to form traceable historical full process data. Based on the stored historical full process data, periodically iterate and adjust the parameter prediction model, the pre-processing detection model before capping, and the finished product detection model after capping to continuously optimize the capping parameter adaptation accuracy and defect detection accuracy.

[0014] Preferably, the inherent defect features include substrate scratches, deformation, and stains. The pre-processing detection model before capping is based on a lightweight deep learning architecture and enhances the differentiation of minute defect features through an adaptive threshold segmentation algorithm. The specific formula is as follows: ,in, For pixels Adaptive segmentation threshold, This represents the average grayscale value within the neighborhood of that pixel. The standard deviation of the gray level in the neighborhood. This is an adjustment coefficient, with a value range of 1.2-2.0.

[0015] Preferably, the temperature and humidity environmental interference compensation and calibration of the weight data in step S1 specifically involves: real-time acquisition of temperature and humidity data through environmental sensors, and compensation calculation based on a multi-factor regression algorithm. The weight compensation formula of the multi-factor regression algorithm is as follows: ,in, For the calibrated weight data, This is the initial weight measurement. This refers to the actual ambient temperature. The standard calibration temperature is 25°C. This represents the actual ambient humidity. The standard calibration humidity is 50% RH. This is the temperature influence coefficient, with a value range of -0.002 to -0.001 g / ℃. This is the humidity influence coefficient, with a value range of -0.001 to -0.0005 g / %RH. This is a system correction constant, with a value range of -0.001 to 0.001g.

[0016] Preferably, the parameter prediction model in step S2 is constructed based on the LSTM attention mechanism. Its input layer contains 8 feature dimensions, the hidden layer has 3 layers, and the output layer outputs the cap pressure threshold and pressure holding time. The prediction formulas for the cap pressure threshold and pressure holding time are as follows: , ,in, The cap pressure threshold. For the duration of pressure maintenance, The reference dimensions for the cap installation area are as follows: For dimensional deviation, This refers to the hardness parameter of the metal terminal material. Score the quality of historical caps. - , - The weight coefficients obtained during model training. , For bias terms;

[0017] Furthermore, the predicted cap pressure threshold range is 5-50N, the pressure holding time range is 0.5-3s, and the prediction deviations are ≤±0.2N and ±0.2s, respectively.

[0018] Preferably, the specific execution process of step S3 is as follows: a servo press is used to perform the cap-pressing operation, the pressure application speed is set to 0.1-0.5 mm / s, and the pressure dynamic changes are monitored in real time by a pressure sensor with a sampling frequency of 500 Hz; when the pressure reaches the preset threshold, a timer is started and monitoring continues. If the pressure fluctuation within 100 ms is ≤ ±0.05 N, it is determined to be stable and maintained until the preset time is reached, after which pressurization is stopped, completing the cap-pressing operation. A real-time pressure feedback mechanism is set, and the formula for calculating the deviation between the pressure monitoring value and the preset threshold is: ,in, This is the pressure monitoring value. The cap pressure threshold is when When this happens, the operation is automatically paused and a parameter update is triggered. The parameter prediction model is re-invoked, and the adaptation parameters are updated based on the current pressure deviation and real-time terminal size data. The updated pressure threshold is then determined. , The correction factor is set to 0.3-0.5. The job will be restored after the update.

[0019] Preferably, the finished product detection model after cap pressing in step S4 adopts a ResNet-50 architecture with an attention mechanism, and optimizes the defect recognition sensitivity through an anchor frame adaptive adjustment algorithm. The anchor frame scale adjustment formula is as follows: ,in, The adjusted anchor frame dimensions. The basic anchor frame dimensions range from 0.05 to 0.5 mm. This refers to the actual size of the detection area. This is the standard area size.

[0020] Preferably, the screw tightening status determination in step S4 adopts a two-way verification mechanism: First, the tightening torque value of the screw head is collected by a torque sensor, and a preset standard torque is used. When the tightening torque value is less than the preset standard torque, it is directly determined that the tightening is unqualified. When the tightening torque value is greater than or equal to the preset standard torque, further image texture analysis is performed. An enlarged image of the screw and terminal contact surface is collected by an industrial camera, and the grayscale uniformity of the thread engagement area is calculated. ,in, For the uniformity of grayscale in the thread meshing area, This represents the total number of pixels in the thread engagement area. For the first grayscale value of each pixel. The average gray value of the region;

[0021] Preset standard uniformity threshold =0.85, when ≥ If the lock is engaged, it is considered to be properly locked; otherwise, it is considered to be unqualified.

[0022] Preferably, the historical full-process data in step S6 is stored in a distributed database. The stored content includes product number, original detection data of each step, model inference results, capping parameters, judgment conclusions and timestamps. The data storage duration is ≥1 year and supports multi-dimensional query by product number, production date and defect type.

[0023] Preferably, the model iteration adjustment in step S6 adopts a differential update mechanism, triggered once every 24 hours. The preprocessing detection model before cap pressing is incrementally trained based on newly added congenital defect samples, and the loss function update formula is: ,in, This represents the total loss function value after incremental training of the preprocessing detection model before cap pressing. The loss function value of the original training samples of the detection model is used for preprocessing before cap pressing. To add a loss function value for inherently defective samples to the pre-capping inspection model, α is a weighting coefficient with a value of 0.7-0.8. The post-capping finished product inspection model focuses on optimizing the identification of associated defects, adjusting the attention mechanism weights. The parameter prediction model iteratively updates the weighting coefficients based on the adaptation deviation data of products with unqualified capping parameters. The update formula is as follows: ,in, The weight coefficients are the parameters after the model is iteratively updated. Here, represents the original weight coefficients before the parameter prediction model is iteratively updated, and β is the learning rate, ranging from 0.001 to 0.005. This is the weight correction amount calculated based on the deviation data.

[0024] By adopting the above technical solution, the present invention has the following advantages compared with the prior art:

[0025] 1. This invention provides an integrated control method for capping and testing of metal terminals. By establishing a closed-loop mechanism for the entire process of capping and testing through a deep learning model, the individual size deviation of the terminal, material characteristics, historical quality data and capping parameters are deeply integrated to achieve accurate matching of capping pressure and holding time, with small prediction deviation. This effectively solves the defects such as capping offset, loosening and crushing caused by the disconnect between parameters and test results in traditional processes.

[0026] 2. This invention provides an integrated control method for capping and detecting metal terminals. By leveraging the feature extraction and reasoning capabilities based on a specific computational model, it accurately processes multi-source data such as appearance images, dimensions, weight, and torque before and after capping. Through adaptive threshold segmentation and multi-factor regression algorithms, it achieves accurate identification of minor defects and compensation for environmental interference. At the same time, it continuously improves parameter adaptation and defect detection accuracy through model iterative optimization. Attached Figure Description

[0027] Figure 1 is a flowchart of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0029] Example

[0030] Please refer to Figure 1. This invention discloses an integrated control method for the clamping cap and detection of metal terminals, comprising the following steps:

[0031] S1. Collect the appearance image of the metal terminal before the cap is pressed, the key dimension data of the cap installation area and the initial weight data. Identify the inherent defect characteristics through the pre-processing detection model before the cap is pressed, accurately measure the reference dimensions of the cap installation area, and simultaneously complete the temperature and humidity environmental interference compensation and calibration of the weight data through a multi-factor regression algorithm.

[0032] S2. Based on the preprocessing detection results of step S1, the preset specifications of the metal terminals, and the historical cap quality feedback data, the appropriate cap pressure threshold and pressure holding time are generated through the parameter prediction model.

[0033] S3. Apply pressure to the mounting reference position of the metal terminal cap according to the adaptation parameters output in step S2, monitor the dynamic changes of pressure in real time, and complete the capping operation when the pressure reaches the preset threshold and is maintained stably for a preset time.

[0034] S4. Collect finished product appearance images, overall key dimension data, actual weight data and screw torque data of the metal terminal after pressing. Identify new defect features generated during the pressing process through the finished product inspection model after pressing. Verify the overall size and pressing accuracy. Compare the actual weight with the standard weight threshold. Combine torque data and image texture analysis to determine the screw tightening status in both directions.

[0035] S5. Combine the preprocessing test results of step S1 with the comprehensive test results of step S4, and output three types of judgment results: qualified, cap parameter unqualified, and defect unqualified. If the cap parameter is unqualified, the parameter optimization instruction is triggered. If the defect is unqualified, the defect type and location are clearly marked.

[0036] S6. Link and store the full process output results of steps S1 to S5 with the metal terminal product number to form traceable historical full process data. Based on the stored historical full process data, periodically iterate and adjust the parameter prediction model, the pre-processing detection model before capping, and the finished product detection model after capping to continuously optimize the capping parameter adaptation accuracy and defect detection accuracy.

[0037] In step S1, a 5-megapixel industrial camera is used to capture the appearance image before the cap is pressed. The image resolution is set to 2592×1944 pixels. During the acquisition, uniform illumination (illuminance ≥800 lux) is provided by a ring LED fill light. Key dimension data of the cap installation area are collected by a laser displacement sensor. The sampling frequency is 100Hz and the data acquisition time is 0.5s. The average value is taken as the reference dimension. The weight data is collected by an electronic scale with an accuracy of 0.001g. The median of the three acquisitions is taken as the initial weight data.

[0038] Inherent defect features include substrate scratches, deformation, and stains. The pre-processing detection model before capping is based on a lightweight deep learning architecture and enhances the differentiation of minute defect features through an adaptive threshold segmentation algorithm. The specific formula is as follows: ,in, For pixels Adaptive segmentation threshold, This represents the average grayscale value within the neighborhood of that pixel. The standard deviation of the gray level in the neighborhood. This is an adjustment coefficient, with a value range of 1.2-2.0.

[0039] The dimensional measurement adopts sub-pixel level edge detection technology and achieves precise edge positioning through the Zernike moment algorithm, with a positioning error of ≤±0.003mm and a defect identification accuracy of ≥99.3%.

[0040] The temperature and humidity environmental interference compensation and calibration for weight data in step S1 specifically involves: real-time acquisition of temperature and humidity data through environmental sensors, and compensation calculation based on a multi-factor regression algorithm. The weight compensation formula for the multi-factor regression algorithm is as follows: ,in, For the calibrated weight data, This is the initial weight measurement. This refers to the actual ambient temperature. The standard calibration temperature is 25°C. This represents the actual ambient humidity. The standard calibration humidity is 50% RH. This is the temperature influence coefficient, with a value range of -0.002 to -0.001 g / ℃. This is the humidity influence coefficient, with a value range of -0.001 to -0.0005 g / %RH. The system correction constant has a value range of -0.001 to 0.001 g, and the repeatability error of the calibrated weight data is ≤ ±0.002 g.

[0041] The preset specifications for metal terminals include material (copper / aluminum / alloy), nominal size (2-10mm), and hardness (HV50-HV200). Historical capping quality feedback data includes the capping pass rate, defect type statistics, and parameter adaptation deviation of terminals of the same specification within the past 3 months.

[0042] The parameter prediction model described in step S2 is built based on the LSTM attention mechanism. Its input layer contains 8 feature dimensions, and the hidden layer has 3 layers with 64 / 32 / 16 neurons per layer. The output layer outputs the cap-pressing pressure threshold and pressure holding time. The prediction formulas for the cap-pressing pressure threshold and pressure holding time are as follows: , ,in, The cap pressure threshold. For the duration of pressure maintenance, The reference dimensions for the cap installation area are as follows: For dimensional deviation, This refers to the hardness parameter of the metal terminal material. Score the quality of historical caps. - , - The weight coefficients obtained during model training. , For bias terms;

[0043] Furthermore, the predicted cap pressure threshold range is 5-50N, the pressure holding time range is 0.5-3s, and the prediction deviations are ≤±0.2N and ±0.2s, respectively.

[0044] The specific execution process of step S3 is as follows: A servo press is used to perform the cap-pressing operation. The pressure application speed is set to 0.1-0.5 mm / s. The pressure dynamic changes are monitored in real time by a pressure sensor with a sampling frequency of 500 Hz. When the pressure reaches the preset threshold, a timer is started and monitoring continues. If the pressure fluctuation is ≤ ±0.05 N within 100 ms, it is determined to be stable and maintained until the preset time is reached, after which pressurization is stopped, completing the cap-pressing operation. A real-time pressure feedback mechanism is set up, and the formula for calculating the deviation between the pressure monitoring value and the preset threshold is: ,in, This is the pressure monitoring value. The cap pressure threshold is when When this happens, the operation is automatically paused and a parameter update is triggered. The parameter prediction model is re-invoked, and the adaptation parameters are updated based on the current pressure deviation and real-time terminal size data. The updated pressure threshold is then determined. , The correction factor is set to 0.3-0.5. The job will be restored after the update.

[0045] In step S4, an industrial camera and laser displacement sensor of the same specifications as in step S1 are used to collect images of the finished product's appearance and overall key dimensions. An electronic scale collects actual weight data (the median of three collections is taken), and a torque sensor (accuracy ≥ 0.01 N·m) collects screw torque data. The finished product inspection model after the cap is pressed adopts a ResNet-50+ attention mechanism architecture. The newly added defect features include new scratches (length ≥ 0.05 mm), cap offset (≥ 0.01 mm), pressure marks (area ≥ 0.01 mm²), and loosening (gap ≥ 0.005 mm).

[0046] The finished product detection model after cap pressing in step S4 adopts a ResNet-50 architecture with an attention mechanism, and optimizes the defect recognition sensitivity through an anchor frame adaptive adjustment algorithm. The specific formula for anchor frame scale adjustment is as follows: ,in, The adjusted anchor frame dimensions. The basic anchor frame dimensions range from 0.05 to 0.5 mm. This refers to the actual size of the detection area. This is the standard area size.

[0047] The fitting accuracy of the pressure cap is determined by measuring the maximum gap between the edge of the pressure cap and the reference surface of the terminal, with a measurement error of ≤ ±0.005mm; the standard weight threshold is the sum of the calibration weight before pressure cap application and the standard weight of the pressure cap component, with an allowable deviation of ±0.1g.

[0048] In step S4, the screw tightening status determination adopts a two-way verification mechanism: First, the tightening torque value of the screw head is collected by a torque sensor. A preset standard torque is used. When the tightening torque value is less than the preset standard torque, it is directly determined that the tightening is unqualified. When the tightening torque value is greater than or equal to the preset standard torque, further image texture analysis is performed. An industrial camera is used to collect a magnified image of the screw and terminal contact surface, and the grayscale uniformity of the thread engagement area is calculated. ,in, For the uniformity of grayscale in the thread meshing area, This represents the total number of pixels in the thread engagement area. For the first grayscale value of each pixel. The average gray value of the region;

[0049] Preset standard uniformity threshold =0.85, when ≥ If the lock is engaged, it is considered to be properly locked; otherwise, it is considered to be unqualified.

[0050] The specific execution process of step S5 is as follows: A three-level judgment rule is constructed. Level 1 judgment: If step S1 identifies an inherent defect or the weight after calibration exceeds the preset range (±0.2g), it is directly judged as defective and the inherent defect type and location are marked. Level 2 judgment: If step S1 has no inherent defect, but step S4 identifies a new defect, dimensional deviation (≥0.01mm), weight deviation (±0.1g), or locking failure, it is judged as defective and the new defect type and location are marked. Level 3 judgment: If the detection results of step S4 are all qualified, it is judged as qualified; if step S4 detects a cap offset ≥0.02mm and no other defects, it is judged as cap parameter unqualified, triggering a parameter optimization command and recording the current cap parameter and offset. The output delay of the three types of judgment results is ≤1s.

[0051] In step S6, the historical full-process data is stored in a distributed database. The stored content includes product number, original detection data of each step, model inference results, capping parameters, judgment conclusions and timestamps. The data storage time is ≥1 year and supports multi-dimensional query by product number, production date and defect type.

[0052] In step S6, the model iterative adjustment adopts a differential update mechanism, triggered once every 24 hours. The preprocessing detection model before cap pressing is incrementally trained based on newly added congenital defect samples, and the loss function update formula is: ,in, This represents the total loss function value after incremental training of the preprocessing detection model before cap pressing. The loss function value of the original training samples of the detection model is used for preprocessing before cap pressing. To add a loss function value for inherently defective samples to the pre-capping inspection model, α is a weighting coefficient with a value of 0.7-0.8. The post-capping finished product inspection model focuses on optimizing the identification of associated defects, adjusting the attention mechanism weights. The parameter prediction model iteratively updates the weighting coefficients based on the adaptation deviation data of products with unqualified capping parameters. The update formula is as follows: ,in, The weight coefficients are the parameters after the model is iteratively updated. Here, represents the original weight coefficients before the parameter prediction model is iteratively updated, and β is the learning rate, ranging from 0.001 to 0.005. This is the weight correction amount calculated based on the deviation data.

[0053] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for integrated control of capping and detection of metal terminals, characterized in that, Includes the following steps: S1. Collect the appearance image of the terminal before capping, key dimension data of the capping installation area, and initial weight data. Identify inherent defect characteristics through a pre-processing detection model before capping, measure the reference dimensions of the capping installation area, and simultaneously use a multi-factor regression algorithm to compensate for and calibrate the weight data for temperature and humidity environmental interference. S2. Based on the pre-processing detection results of S1, the terminal's preset specification parameters, and historical capping quality feedback data, generate a suitable capping pressure threshold and pressure holding time through a parameter prediction model. The parameter prediction model is built based on an LSTM attention mechanism. Its input layer contains 8 feature dimensions, the hidden layer has 3 layers, and the output layer outputs the capping pressure threshold and pressure holding time. The prediction formulas for the capping pressure threshold and pressure holding time are as follows: ， ,in, The cap pressure threshold. For the duration of pressure maintenance, The reference dimensions for the cap installation area are as follows: For dimensional deviation, This refers to the hardness parameter of the metal terminal material. Score the quality of historical caps. - 、 - The weight coefficients obtained during model training. 、 The bias term is used; and the predicted pressure threshold range is 5-50N, the pressure holding time range is 0.5-3s, and the prediction deviations are ≤±0.2N and ±0.2s respectively; S3, according to the adaptation parameters output by S2, apply pressure to the terminal cap installation reference position, monitor the dynamic changes of pressure in real time, and complete the capping operation when the pressure reaches the preset threshold and is held stably for the preset time; S4, collect the finished appearance image, overall key dimension data, actual weight data, and screw torque data of the cap rear end, identify the new defect features generated during the capping process through the capping finished product detection model, verify the overall size and cap fitting accuracy, compare the actual weight with the standard weight threshold, and determine the screw tightening status by combining torque data and image texture analysis; the capping finished product detection model adopts a ResNet-50 and attention mechanism architecture, and optimizes the defect identification sensitivity through the anchor frame adaptive adjustment algorithm. The anchor frame scale adjustment formula is as follows: ,in, The adjusted anchor frame dimensions. The basic anchor frame dimensions range from 0.05 to 0.5 mm. This refers to the actual size of the detection area. S5: Combine the pre-processing inspection results of S1 with the comprehensive inspection results of S4, and output the judgment result. If the product is unqualified by the capping parameters, the parameter optimization instruction is triggered. If the product is unqualified by defects, the defect type and location are clearly marked. S6: Associate and store the outputs of S1 to S5 with the metal terminal product number to form traceable historical full-process data. The parameter prediction model, the pre-processing inspection model before capping, and the finished product inspection model after capping are periodically iterated and adjusted to continuously optimize the capping parameter adaptation accuracy and defect detection accuracy.

2. The integrated control method for pressing and detecting metal terminals as described in claim 1, characterized in that: The inherent defect features include substrate scratches, deformation, and stains. The pre-processing detection model before capping is based on a lightweight deep learning architecture and enhances the differentiation of minute defect features through an adaptive threshold segmentation algorithm. The specific formula is as follows: ,in, For pixels Adaptive segmentation threshold, This represents the average grayscale value within the neighborhood of that pixel. The standard deviation of the gray level in the neighborhood. This is an adjustment coefficient, with a value range of 1.2-2.

0.

3. The integrated control method for pressing and detecting metal terminals as described in claim 1, characterized in that, The temperature and humidity environmental interference compensation and calibration for weight data in step S1 specifically involves: real-time acquisition of temperature and humidity data through environmental sensors, and compensation calculation based on a multi-factor regression algorithm. The weight compensation formula for the multi-factor regression algorithm is as follows: ,in, For the calibrated weight data, This is the initial weight measurement. This refers to the actual ambient temperature. The standard calibration temperature is 25°C. This represents the actual ambient humidity. The standard calibration humidity is 50% RH. This is the temperature influence coefficient, with a value range of -0.002 to -0.001 g / ℃. This is the humidity influence coefficient, with a value range of -0.001 to -0.0005 g / %RH. This is a system correction constant, with a value range of -0.001 to 0.001g.

4. The integrated control method for pressing and detecting metal terminals as described in claim 1, characterized in that, The specific execution process of step S3 is as follows: A servo press is used to perform the cap-pressing operation. The pressure application speed is set to 0.1-0.5 mm / s. The pressure dynamic changes are monitored in real time by a pressure sensor with a sampling frequency of 500 Hz. When the pressure reaches the preset threshold, a timer is started and monitoring continues. If the pressure fluctuation is ≤ ±0.05 N within 100 ms, it is determined to be stable and maintained until the preset time is reached, after which pressurization is stopped, completing the cap-pressing operation. A real-time pressure feedback mechanism is set up, and the formula for calculating the deviation between the pressure monitoring value and the preset threshold is: ,in, This is the pressure monitoring value. The cap pressure threshold is when When this happens, the operation is automatically paused and a parameter update is triggered. The parameter prediction model is re-invoked, and the adaptation parameters are updated based on the current pressure deviation and real-time terminal size data. The updated pressure threshold is then determined. ， The correction factor is set to 0.3-0.

5. The job will be restored after the update.

5. The integrated control method for pressing and detecting metal terminals as described in claim 1, characterized in that, In step S4, the screw tightening status determination adopts a two-way verification mechanism: First, the tightening torque value of the screw head is collected by a torque sensor. A preset standard torque is used. When the tightening torque value is less than the preset standard torque, it is directly determined that the tightening is unqualified. When the tightening torque value is greater than or equal to the preset standard torque, further image texture analysis is performed. An industrial camera is used to collect a magnified image of the screw and terminal contact surface, and the grayscale uniformity of the thread engagement area is calculated. ,in, For the uniformity of grayscale in the thread meshing area, This represents the total number of pixels in the thread engagement area. For the first grayscale value of each pixel. The average grayscale value of the region; preset standard uniformity threshold. =0.85, when ≥ If the lock is engaged, it is considered to be properly locked; otherwise, it is considered to be unqualified.

6. The integrated control method for pressing and detecting metal terminals as described in claim 1, characterized in that: The historical full-process data mentioned in step S6 is stored in a distributed database. The stored content includes product number, original detection data of each step, model inference results, capping parameters, judgment conclusions and timestamps. The data storage time is ≥1 year and supports multi-dimensional query by product number, production date and defect type.

7. The integrated control method for pressing and detecting metal terminals as described in claim 6, characterized in that: In step S6, the model iterative adjustment adopts a differential update mechanism, triggered once every 24 hours. The preprocessing detection model before cap pressing is incrementally trained based on newly added congenital defect samples, and the loss function update formula is: ,in, This represents the total loss function value after incremental training of the preprocessing detection model before cap pressing. The loss function value of the original training samples of the detection model is used for preprocessing before cap pressing. To add a loss function value for inherently defective samples to the pre-capping inspection model, α is a weighting coefficient with a value of 0.7-0.

8. The post-capping finished product inspection model focuses on optimizing the identification of associated defects, adjusting the attention mechanism weights. The parameter prediction model iteratively updates the weighting coefficients based on the adaptation deviation data of products with unqualified capping parameters. The update formula is as follows: ,in, The weight coefficients are the parameters after the model is iteratively updated. Here, represents the original weight coefficients before the parameter prediction model is iteratively updated, and β is the learning rate, ranging from 0.001 to 0.

005. This is the weight correction amount calculated based on the deviation data.

Images