A Smart Quality Inspection Method for PTFE Oil Seal Lip Return Line

Abstract

This invention provides an intelligent quality inspection method for PTFE oil seal lip return lines, belonging to the field of industrial defect detection technology based on computer data processing. Specifically, the method includes: firstly, acquiring images of the oil seal lip return lines from different angles under various operating conditions using an industrial high-definition camera; then, constructing a defect detection dataset with feature labels after expert annotation; subsequently, designing an image defect extraction unit that integrates group convolution and dense connections, using this as the core to construct a multi-path defect feature perception module, and introducing a multi-branch prediction head architecture to implement the defect detection model design; finally, combining the defect detail features and overall defect features calculated by the fault state characterization module, as well as the defect type and hazard level information predicted by the model, utilizing innovative statistical analysis algorithms and long-term window and temporal domain processing methods, completing the overall design of the intelligent quality inspection method for the production process.

Description

Technical Field

[0001] This invention belongs to the field of industrial defect detection technology based on computer data processing, and particularly relates to an intelligent quality detection method for the lip return line of PTFE oil seal. Background Technology

[0002] PTFE oil seals, with their corrosion resistance, aging resistance, and low friction, have become core components in high-end industrial sealing fields, especially irreplaceable in harsh environments such as automotive, aerospace, and chemical industries. The machining precision of their lip return lines directly determines sealing performance and equipment lifespan; however, quality defects can lead to leakage risks, causing equipment downtime or even safety accidents. Therefore, ensuring the quality of PTFE oil seal lip return lines is a crucial link in ensuring supply chain security and enhancing manufacturing competitiveness. However, existing quality inspection technologies suffer from low efficiency and poor accuracy. The core motivation for developing a quality inspection and production process fault early warning method for PTFE oil seal lip return lines lies in the fact that traditional manual sampling inspection methods suffer from three major drawbacks: weak ability to capture microscopic defects, high detection lag, and insufficient coverage. They cannot identify micron-level return line structural anomalies (such as burrs, breaks, and contour deformation) in real time, leading to the transmission of sealing failure risks downstream. By integrating machine vision and statistical data analysis technologies, fully automated quality inspection and dynamic early warning of process anomalies can be achieved, preventing quality defects at the source, improving product reliability and production efficiency, and providing key technological support for intelligent manufacturing of high-precision seals.

[0003] Currently, the main methods for quality inspection of PTFE oil seal lip return lines are as follows:

[0004] (1) Traditional mathematical-physical model quality inspection method: This method is based on the theory of sealing contact mechanics. It calculates the sealing performance by establishing equations such as lip contact pressure distribution and fluid dynamic lubrication, combined with material parameters (such as PTFE elastic modulus and friction coefficient). In actual testing, it is necessary to manually measure key dimensions (such as return line height and inclination angle) using tools such as micrometers and profilometers, and then substitute the data into the model to verify the theoretical leakage rate. Its limitation is that it oversimplifies the micro-morphology (ignores real defects such as burrs and microcracks), and the model relies on empirical parameter adjustment, which cannot directly capture the hidden quality variation caused by process fluctuations. It is only suitable for theoretical verification in the design stage.

[0005] (2) Quality inspection method based on simple image processing: Basic machine vision technology (such as threshold segmentation and edge detection) is used to extract the contour of the statically captured two-dimensional image of the return line, and obvious geometric deviations (such as fractures and large-size deformations) are judged by comparing with a preset template. The system is usually configured with a fixed light source and a low-resolution camera, and the pixel area or perimeter is used as the main criterion. Its drawback is that it can only identify macroscopic defects (>50μm), and its sensitivity to micron-level structural anomalies (such as burrs and shallow scratches) is insufficient, and the false detection rate is high, requiring manual re-judgment and confirmation;

[0006] (3) Confocal Microscopic Dynamic Monitoring Method: This method utilizes an ultra-short depth-of-focus laser beam to scan the cross-section of the PTFE oil seal lip layer by layer, and constructs a high-precision three-dimensional model through axial tomography. Its nanometer-level resolution can clearly capture hidden defects such as surface burrs, microcracks, and internal material delamination, and combined with dynamic analysis algorithms, it can track microscopic deformations caused by process fluctuations in real time. This method has high detection accuracy, but its detection speed and efficiency cannot meet the needs of factories, and the material cost of the detected materials is too high. Summary of the Invention

[0007] To address the above problems, this invention proposes an intelligent quality inspection method for the PTFE oil seal lip return line, comprising the following steps:

[0008] S1, real-time acquisition of oil seal lip return line images and corresponding process parameters during the PTFE oil seal lip return line process;

[0009] S2: Input the data from S1 into the trained oil seal lip return line defect detection model. Extract feature vectors of multiple feature scales based on the oil seal lip return line image, and extract auxiliary feature vectors based on process parameters. Fusion detection of multiple feature vectors, output return line defect detection type result, defect severity result, and small-area cropped image containing defect information.

[0010] S3, input the small-area cropped image obtained in S2 into the production status feature characterization module of the return line, and extract the defect detail features and overall defect features corresponding to the defect image; the defect detail features include the mean of defect grayscale disorder and the defect type similarity coefficient, and the overall defect features include the mean of defect area and the mean of defect perimeter, which are constructed as production status characterization features.

[0011] S4: Construct a dynamic adjustment matrix for defect feature sensitivity based on real-time production conditions. Using the dynamic adjustment matrix for defect feature sensitivity as the weighting coefficient, and based on the return line defect detection type results and defect severity obtained in S2, as well as the production status characterization features obtained in S3, calculate the production process fault characterization factor. Using the time-series feature parameter calibration architecture, calculate the mean and standard deviation of the fault characterization factor to obtain the PTFE oil seal production process fault characterization coefficient.

[0012] S5 determines the fault status based on the fault characterization coefficient and the set threshold.

[0013] Preferably, the process parameters include a mold clamping pressure coefficient. Melt injection temperature Holding time Cooling medium temperature Mold opening speed Melt injection speed .

[0014] Preferably, the defect attributes output by the oil seal lip return line defect detection model include:

[0015] Rectangular box covering defects The information in this rectangle specifically refers to the center coordinates of this matrix frame. and the width of the rectangle and height ;

[0016] Defect types Specifically, the categories of this defect include five types: fracture, deformation, hole, wear, and blurred edges.

[0017] Defect severity This defect is classified into six different severity levels, specifically using numerical values. This is reflected in the numerical value, which indicates a higher degree of defect verification.

[0018] Preferably, the oil seal lip return line defect detection model includes a multimodal defect feature perception module and several defect detection heads;

[0019] The multimodal defect feature perception module includes a flow line image defect feature extraction unit and a process parameter defect feature extraction pathway; the flow line image defect feature extraction unit contains three network pathways with different feature fitting and extraction capabilities, extracting feature vectors of three different scales. , and The process parameter defect feature extraction path extracts auxiliary feature vectors. ;

[0020] Each defect detection head focuses on detecting only one type of return line defect. In each defect detection head, a KAN layer is first used to further weight the input features, followed by two adaptive convolutional layers with variable kernel coefficients for further feature extraction, and finally a KAN layer and a normalization layer are used for processing to output the defect detection result.

[0021] Preferably, the flow line image defect feature extraction unit includes three network pathways with different feature fitting and extraction capabilities;

[0022] The first feature fitting extraction pathway first includes two interconnected CNN layers, each using a kernel size of 3. These layers are followed by four sequentially connected image defect extraction units. Finally, the first pathway is completed by combining two CNN layers with kernel sizes of 3, a BN normalization layer, and a SILU activation function. The input image... Features were obtained after extraction via pathway one ;

[0023] The second feature fitting extraction pathway first includes two interconnected CNN layers with a kernel size of 5. These layers then contain four sequentially connected image defect extraction units. Finally, the second pathway is completed by combining two CNN layers with a kernel size of 5, a BN normalization layer, and a ReLU activation function. The input image... Features were obtained after extraction via pathway 2 ;

[0024] In the third feature fitting extraction pathway, two interconnected CNN layers with a kernel size of 7 are first included. These layers then contain four sequentially connected image defect extraction units. Finally, the third pathway is constructed by combining two CNN layers with a kernel size of 7, a BN normalization layer, and a ReLU activation function. The input image... Features were obtained after extraction via pathway three .

[0025] Preferably, the image defect extraction unit first includes four groups of convolutional blocks with different dilatations to take into account both local and global perception capabilities of defect features; then, residual connections and channel splicing layers are used to integrate the output features of the three convolutional network layers to complete the construction of the group of convolutional blocks; residual connections are inserted for each convolutional layer, that is, the output of each convolutional layer in the group of convolutional blocks is directly connected to the final channel splicing layer;

[0026] The dilation rate of the first group of convolutional blocks is 1, the dilation rate of the second group of convolutional blocks is 2, the dilation rate of the third group of convolutional blocks is 3, and the dilation rate of the last group of convolutional blocks is 4.

[0027] Preferably, the process of extracting the defect detail features includes:

[0028] First, crop the image to a smaller area containing defect information. Perform grayscale conversion with 64 grayscale levels. ,in This represents the total number of defects detected for a single PTFE oil seal lip return line; subsequently, regarding the image... For each pixel in the image, calculate the absolute value of the difference between the grayscale value of this pixel and all its neighboring pixels, and then calculate the sum of their absolute values. This sum of absolute values ​​is then used as the sum of the grayscale differences of the corresponding pixels in the improved local grayscale difference matrix, and the process is repeated for each pixel in the image. After collecting all pixels, we obtain the first... A cropped image The corresponding improved local gray-level difference matrix ;

[0029] Based on the calculated L improved local gray-level difference matrices Calculate each improved local gray-level difference matrix separately. matrix mean To measure the texture disorder of this single defect, the mean of L improved local gray-level difference matrices is then calculated again. The average value is the mean value of the defect grayscale disorder. ;

[0030] Based on the calculated L improved local gray-level difference matrices Compare the L improved local gray-level difference matrices similarity coefficient between This characterizes the stability of the production state; a larger similarity coefficient indicates a higher similarity between defects in a single sample, meaning a more unstable production state for PTFE oil seals, while a smaller similarity coefficient indicates a lower similarity between defects in a single sample, meaning a more stable production state for PTFE oil seals; defect type similarity coefficient Calculated as L improved local gray-level difference matrices The weighted average of the Frobenius norm distances between each pair of pairs.

[0031] Preferably, the extraction process of the global features of the defect includes:

[0032] Based on the obtained small-area cropped image , Image erosion and dilation algorithms are used to further denoise and smooth the edges of the image. Then, the area of ​​each return line defect is calculated mathematically. ,perimeter Among them, area This represents the number of pixels covered by this return line defect, and its perimeter. This indicates the number of pixels contained in the outermost ring of this return line defect;

[0033] Then, the calculated L flow line defect area values ​​were... The mean defect area is obtained by solving the mean value. For the calculated L return line defect area values The mean value of the defect perimeter is obtained by solving the mean value. When the average defect area corresponding to the return line of a PTFE oil seal lip is and the average perimeter of defects The higher the value, the higher the risk of failure in the current reflow production state.

[0034] Preferably, the specific process of S4 includes:

[0035] S41, a dynamic adjustment matrix for defect feature sensitivity was constructed for real-time production conditions. :

[0036] ;

[0037] in, As the benchmark weight value, This represents the traction speed influence coefficient. Indicates the influence coefficient of material viscosity. These represent the standard production traction speed and the standard material viscosity, respectively. These represent the maximum production traction speed and the maximum standard material viscosity, respectively; production traction speed and material viscosity This represents the production coefficient corresponding to the actual manufacturing process of the return line;

[0038] S42, based on dynamic adjustment matrix Fault characterization factors for PTFE oil seal lip return line production process constructed under real-time process parameters ;

[0039] S43, based on the calculation method for production process fault characterization factors determined in S42, when performing intelligent fault detection in the production process, the production process fault characterization factors for 100N PTFE oil seal lip return lines are calculated. Constructed sequence data A comprehensive analysis was conducted to obtain the detection results of production failures on the return line. The sequence was then calculated. The corresponding mean and standard deviation To measure the frequency of defects, the severity level, and the instability of defects in these 100N oil seal lip return line samples, the mean value of the defects is used. Larger, standard deviation The higher the value, the greater the probability of a failure in the production process.

[0040] S44, using the mean calculated in process S43 Standard deviation After weighted summation, the failure characterization coefficient of the PTFE oil seal production process is obtained. .

[0041] Preferably, in step S5, a threshold for judging the production process status is set. And perform the following fault determinations:

[0042] When the calculated fault characterization coefficients Less than the threshold At that time, the actual state of the return line production process is judged as normal.

[0043] when Greater than the threshold Less than the threshold At that time, the actual state of the return line production process was judged as a minor fault;

[0044] when Greater than the threshold Less than the threshold At that time, the actual state of the return line production process was judged as a medium-level fault;

[0045] when Greater than the threshold At that time, the actual state of the return line production process was judged as an emergency fault.

[0046] Compared with the prior art, the innovative points and beneficial effects of this invention include:

[0047] (1) Intelligent quality inspection of PTFE oil seal lip return line: This invention realizes intelligent and unmanned quality inspection of oil seal lip return line by constructing a defect detection model. Based on the multimodal defect feature perception module and multi-defect detection head architecture designed in the model, the accuracy and robustness of PTFE oil seal lip return line quality perception are further increased;

[0048] (2) Design of a multi-feature fusion production process characterization method: This invention innovatively integrates image texture features and morphological geometric features to construct a 4-dimensional feature set. This method comprehensively captures the microscopic features of defects, establishes the correlation between defects and production parameters, and provides fine-grained data support for the system detection of PTFE production process faults. From the perspectives of details and the whole, it provides richer features for subsequent production process quality inspection, ensuring the detection efficiency and accuracy of production process fault detection;

[0049] (3) Design of a production process fault detection system based on dynamic weighting and time-domain analysis: The innovation lies in constructing a dynamic weighting adjustment mechanism for defect features based on real-time production conditions (traction speed, material viscosity), and combining improved Mahalanobis distance and time-domain analysis to quantify the time-series defect feature sequence in multiple dimensions, thereby achieving four-level hierarchical diagnosis of production status. Its beneficial effect is to improve the adaptability and diagnostic accuracy of production process fault detection through dynamic weighting and time-domain analysis, open up the technical link from product quality detection to intelligent monitoring of the production process, and provide an efficient solution for real-time process early warning and defect rate control. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the overall implementation logic of the present invention.

[0051] Figure 2 This is the overall framework design drawing of the oil seal lip return line defect detection model of the present invention.

[0052] Figure 3 This is a comparison chart of the accuracy of defect detection in the return line of the present invention.

[0053] Figure 4 This is a comparison chart showing the accuracy of fault state detection in the production process of the return line of this invention. Detailed Implementation

[0054] This invention proposes an intelligent quality inspection method for the PTFE oil seal lip return line. The overall technical route is as follows: Figure 1 As shown:

[0055] S1, Multi-dimensional oil seal lip return line image acquisition: Using an industrial high-definition camera, oil seal lip return line images are acquired from different angles under multiple working conditions, covering normal return lines and defect samples such as fractures, deformations, and holes. After noise reduction preprocessing, experts label the defect locations and types. Then, combined with specific process parameters, an oil seal lip return line defect detection dataset with feature labels is constructed.

[0056] S2, Design of Oil Seal Lip Return Line Defect Detection Model: In the design of the oil seal lip return line defect detection model, this invention constructs an image defect extraction unit that integrates group convolution and dense connections. Subsequently, the design of the return line multi-path defect feature perception module is completed with the image defect extraction unit as the core. Simultaneously, in the defect prediction head, a multi-branch prediction head architecture design is introduced to achieve efficient feature identification and effective cropping of defect regions.

[0057] S3, Design of the production status characteristic characterization module of the return line: Based on the defect area image obtained by S2, a defect detail feature calculation module and a defect overall feature calculation module were constructed to provide feature data support for subsequent fault detection in the oil seal lip return line manufacturing process;

[0058] S4, Design of Fault Detection System for Oil Seal Lip Return Line Production Process: Based on the defect type and hazard level information obtained from S2 detection and the defect detail features and overall defect features obtained from S3, this invention designs an innovative defect data statistical analysis algorithm, and completes the overall design of the production process fault detection system based on long-term window detection and frequency domain and time domain dual-domain data processing methods.

[0059] The present invention will be further described below with reference to specific embodiments.

[0060] I. Construction of a Defect Detection Dataset for Oil Seal Lip Return Line

[0061] This invention employs image data processing technology to achieve accurate identification of defects in the return flow line of oil seal lips. Therefore, high-quality image data samples are required during the dataset construction phase for subsequent image data processing. To ensure the quality of data acquisition, this invention designs a robust data acquisition method during the dataset acquisition phase to eliminate the noise introduced by environmental factors in the image data, and performs image calibration operations, as detailed below.

[0062] 1. Image Acquisition of Oil Seal Lip Return Line: An industrial-grade line scan camera with a telecentric lens is used, and the camera is deployed on a high-precision robotic arm to achieve multi-angle scanning of the oil seal lip return line within 0° to 180°, thereby eliminating the influence of perspective distortion. Furthermore, to ensure the quantity of defect samples acquired during image data acquisition and to provide data support for subsequent production parameters, five types of oil seal lip return line defect samples—fracture, deformation, hole, wear, and edge blurring—are constructed. In addition, during image acquisition, three sets of duplicate images are acquired for each sample to eliminate the influence of noise during the image acquisition process. Initial images are obtained according to the above image acquisition method. .

[0063] 2. Extraction of core process parameters: This is done after creating the initial image... During image acquisition, it is also necessary to record the core process parameters corresponding to the image acquisition. Specifically, this includes the mold closing pressure coefficient. Melt injection temperature Holding time Cooling medium temperature Mold opening speed Melt injection speed Six core process parameters, namely .

[0064] 3. Image annotation: For the images processed by the preprocessing module, relevant professionals use the LabelMe image annotation tool to annotate the defects and attributes present in the images. Make accurate annotations. Specifically, it includes:

[0065] 1) A rectangular box covering the defect The information in this rectangle specifically refers to the center coordinates of this matrix frame. and the width of the rectangle and height ;

[0066] 2) Defect Types Specifically, the categories of this defect include five types: fracture, deformation, hole, wear, and blurred edges.

[0067] 3) Severity of defects This defect is classified into 6 different severity levels, specifically using numerical values. This is reflected in the numerical value, which indicates a higher degree of defect verification.

[0068] 4. Construction of a standard dataset for detecting defects in the return flow line of oil seal lips:

[0069] Based on the initial image of the PTFE oil seal lip return line obtained from the acquisition and the obtained core process parameters The two types of data are combined to obtain a complete set of input data for detecting defects in the return line of the oil seal lip. After labeling Output data for detecting defects in the return line of the oil seal lip The input and output data are then combined to obtain a complete set of oil seal lip return line defect detection data. In the process of constructing this dataset, N sets of oil seal lip return line defect detection data will be constructed. By combining these data, a standard dataset for detecting defects in the return line of oil seal lip can be completed. The construction process.

[0070] II. Design of Defect Detection Model for Oil Seal Lip Return Line

[0071] The overall framework of the oil seal lip return line defect detection model in this embodiment is as follows: Figure 2 As shown, it includes a multimodal defect feature perception module and five defect detection heads.

[0072] 1. Multimodal Defect Feature Perception Module: This module includes a return line image defect feature extraction unit and a process parameter defect feature extraction path. The input data for the return line image defect feature extraction unit is the initial image of the PTFE oil seal lip return line. The input data for the process parameter defect feature extraction pathway is the core process parameter. ;

[0073] (1) Design of the feature extraction unit architecture for the pipelined image defect: The pipelined image defect feature extraction unit contains three network pathways with different feature fitting and extraction capabilities. The first feature fitting extraction pathway first includes two interconnected CNN layers with a kernel size of 3, followed by four sequentially connected image defect extraction units. Finally, the first feature fitting extraction pathway is completed by combining two CNN layers with a kernel size of 3, a BN normalization layer, and a SILU activation function. The input image is... Features were obtained after extraction via pathway one .

[0074] The second feature fitting extraction pathway first includes two interconnected CNN layers with a kernel size of 5. These layers then contain four sequentially connected image defect extraction units. Finally, the second pathway is completed by combining two CNN layers with a kernel size of 5, a BN normalization layer, and a ReLU activation function. The input image... Features were obtained after extraction via pathway 2 .

[0075] In the third feature fitting extraction pathway, two interconnected CNN layers with a kernel size of 7 are first included. These layers then contain four sequentially connected image defect extraction units. Finally, the third pathway is constructed by combining two CNN layers with a kernel size of 7, a BN normalization layer, and a ReLU activation function. The input image... Features were obtained after extraction via pathway three .

[0076] The image defect extraction unit structure design addresses the issue that the actual coverage area of ​​the oil seal lip return line defect is very small. To ensure accurate perception of this minute defect without excessively increasing computational complexity, this invention utilizes group convolution and tightly connected modules to design the image defect extraction unit architecture.

[0077] The image defect extraction unit first includes four convolutional blocks with different dilation rates to achieve both local and global perception of defect features. Then, residual connections and channel concatenation layers are used to integrate the output features of the three convolutional network layers to complete the construction of the convolutional blocks. It is particularly noteworthy that this invention inserts residual connections for each convolutional layer, meaning that the output of each convolutional layer in the convolutional block is directly connected to the final channel concatenation layer.

[0078] Specifically, the dilation rate of the first group of convolutional blocks is 1, the dilation rate of the second group of convolutional blocks is 2, the dilation rate of the third group of convolutional blocks is 3, and the dilation rate of the last group of convolutional blocks is 4.

[0079] Features extracted from the first group of convolutional blocks The features extracted by the second group of convolutional blocks The features extracted by the third group of convolutional blocks and the features extracted from the fourth group of convolutional blocks This invention achieves tight connection and integration of the output features of four groups of convolutional blocks by constructing a dynamic routing layer. In this dynamic routing layer, an adaptive feature weight allocation is first performed on the features extracted by each group of convolutional blocks, and then the output features of each convolutional block are weighted and summed to obtain the fused defect data features extracted by a single image defect extraction unit.

[0080] (2) Design of process parameter defect feature extraction pathway architecture: This invention innovatively integrates KAN network (Kolmogorov-Arnold Network) into the process parameter defect feature extraction pathway to extract features of core process parameters, thereby further improving the ability to perceive minute defects.

[0081] The process parameter defect feature extraction pathway first includes two interconnected KAN network layers, using spline functions of order 6. Following this, it passes through a normalization layer and a Dropout layer, and then three sequentially connected KAN network layers, using spline functions of order 8. This completes the construction of the process parameter defect feature extraction pathway, focusing on the core process parameters. Features were obtained after processing through the process parameter defect feature extraction pathway. .

[0082] 2. Defect Detection Head Design: Since the five types of defects—fracture, deformation, holes, wear, and blurred edges—differ in both texture and morphological features, this invention, based on the specificity of these flow-line defect features, innovatively implements a multi-branch defect detection head design. Each detection head focuses on detecting only one type of flow-line defect. In each defect detection head, a KAN layer is first used to further weight the input features. Then, two adaptive convolutional layers with variable kernel coefficients are used for further feature extraction. Finally, a KAN layer and a normalization layer are used for further processing to complete the overall design of the defect detection head.

[0083] 3. Implementation of the oil seal lip return line defect detection model: Based on the features extracted by the designed multimodal defect feature perception module. And the described defect detection head design. Input image data. After processing by the multi-path defect feature sensing module, the obtained features are... After being fused into a KAN network layer, the data is fed into five independent defect detection heads to perform defect detection on the branched oil seal lip return line, thus obtaining the defect detection results for this PTFE oil seal lip return line. :

[0084] ;

[0085] Where L represents the total number of defects detected by a PTFE oil seal lip return line. Indicates the detected number Location information of the rectangle containing the defect. Indicates the detected number The defect type of each defect. Indicates the detected number The severity of each defect is determined. Then, based on the obtained defect bounding box location information... Obtain a small-area cropped image containing only defect information. This provides image data support for the S3 return line production status feature characterization module.

[0086] III. Design of the Production Status Characterization Module for the Return Line

[0087] To achieve accurate assessment of the production status of the return line, further detailed and overall defect features can be extracted from the defect images. These detailed and overall defect features are related to the actual production status of the PTFE oil seal; that is, the defect features under fault conditions and those under normal conditions will differ significantly.

[0088] 1. Design of the defect detail feature calculation module:

[0089] (1) Improved calculation of local gray-level difference matrix: When the production status of PTFE oil seal is abnormal, it will directly change the distribution of the material's microstructure. This physical deformation is simultaneously reflected in the variation of gray-level spatial relationship in the defect image. By quantifying the spatial correlation between pixels, the improved local gray-level difference matrix can directly encode such variations into diagnosable texture features, thereby providing richer judgment features for the accurate perception of subsequent production status.

[0090] Small-scale cropped image based on obtained defect information , ,in This represents the total number of defects detected for a single PTFE oil seal lip return line. The image processing procedure first involves... The image was then converted to grayscale with 64 levels. For each pixel in the image, calculate the absolute value of the difference between the grayscale value of this pixel and all its neighboring pixels, and then sum these absolute values. This sum of absolute values ​​is then used as the sum of the grayscale differences of the corresponding pixels in the improved local grayscale difference matrix, and this process is repeated throughout the image. After collecting all pixels, we obtain the first... A cropped image The corresponding improved local gray-level difference matrix .

[0091] (2) The L improved local gray-level difference matrices obtained based on the process in (1) above Calculate each improved local gray-level difference matrix separately. matrix mean To measure the texture disorder of this single defect, the mean of L improved local gray-level difference matrices is then calculated again. The average value is the mean value of the defect grayscale disorder. This indicates the degree of disorder in the defect characteristics of this PTFE oil seal, reflecting its stable performance under production conditions.

[0092] (3) Based on the L improved local gray-level difference matrices calculated in the above (1) process, Compare the L improved local gray-level difference matrices similarity coefficient between This characterizes the stability of the production state. A higher similarity coefficient indicates a higher similarity between defects in a single sample, meaning a more unstable production state for the PTFE oil seal. Conversely, a lower similarity coefficient indicates a lower similarity between defects in a single sample, meaning a more stable production state for the PTFE oil seal. Defect Type Similarity Coefficient Calculated as L improved local gray-level difference matrices The weighted average of the Frobenius norm distances between each pair of pairs.

[0093] Therefore, the final defect detail features include: the mean value of defect grayscale disorder. And the similarity coefficient of defect types .

[0094] 2. Defect Global Feature Calculation Module Design: Based on the obtained small-range cropped image , Image erosion and dilation algorithms are used to further denoise and smooth the edges of the image. Then, the area of ​​each return line defect is calculated mathematically. ,perimeter Among them, area This represents the total number of pixels covered by this return line defect. Its perimeter... This indicates the number of pixels contained in the outermost ring of this return line defect.

[0095] Then, the calculated L flow line defect area values ​​were... The mean defect area is obtained by solving the mean value. For the calculated L return line defect area values The mean value of the defect perimeter is obtained by solving the mean value. When the average defect area corresponding to the return line of a PTFE oil seal lip is... and the average perimeter of defects The higher the value, the higher the risk of failure in the current production return state;

[0096] 3. Construction of Production Status Characteristic Set: Based on the obtained defect detail features, including the improved local gray-level difference matrix mean. And improve the similarity coefficient of the local gray-level difference matrix. The calculated overall characteristics of the defects include the average defect area. and the average perimeter of defects By combining the above four characteristic data, the production status characterization features of the PTFE oil seal lip return line can be obtained. .

[0097] IV. Design of Fault Detection Method for Oil Seal Lip Return Line Production Process

[0098] Based on the detection results of return line defect types Defect severity And the calculated production status characteristics of the PTFE oil seal lip return line. This invention further realizes intelligent fault detection in the PTFE oil seal lip return line production process based on time-series analysis technology. The specific implementation process is as follows:

[0099] Construction of a dynamic adjustment matrix for sensitivity to production status characteristics: In the actual PTFE oil seal production process, under different PTFE oil seal production conditions, the results of defect detection types... Defect severity and characteristics of return line production status The reference weights they occupy in fault detection are different. To ensure the robustness of intelligent fault detection in subsequent production processes, this invention constructs a dynamic adjustment matrix for defect feature sensitivity based on real-time production conditions. :

[0100] ;

[0101] in, As the benchmark weight value, This represents the traction speed influence coefficient. Indicates the influence coefficient of material viscosity. These represent the standard production traction speed and the standard material viscosity, respectively. These represent the maximum production traction speed and the maximum standard material viscosity, respectively; production traction speed and material viscosity This represents the production coefficient corresponding to the actual manufacturing process of the return line.

[0102] Calculation of production process failure characterization factors: Based on the determined defect feature sensitivity dynamic adjustment matrix, this invention further constructs the production process failure characterization factors for the PTFE oil seal lip return line under real-time process parameters. ;

[0103] ;

[0104] in, This represents the vector concatenation operation function. This represents the function for transposing the eigenvector.

[0105] Design of timing characteristic parameter calibration architecture: based on the obtained production process fault characterization factors When conducting intelligent fault detection in actual production processes, a single defect detection on the return line or a very small number of defect detections on the return line cannot accurately reflect the actual state of the return line production process at that time. Therefore, to achieve more accurate and robust fault state perception of the return line production process, this invention designs a time-domain and frequency-domain dual-domain feature judgment algorithm for fault diagnosis of the return line production state, as follows:

[0106] Based on the established calculation method for production process fault characterization factors, the fault characterization factors for 100N PTFE oil seal lip return lines were analyzed during intelligent fault detection in the production process. Constructed sequence data A comprehensive analysis was conducted to obtain the detection results of production failures on the return line. The sequence was then calculated. The corresponding mean and standard deviation To measure the frequency of defects, the severity level, and the instability of defects in these 100N oil seal lip return line samples, the mean value of the defects is used. The larger the value, and the standard deviation The higher the value, the greater the probability of a failure in the production process.

[0107] Using the calculated mean Standard deviation After weighted summation, the failure characterization coefficient of the PTFE oil seal production process is obtained. Set thresholds for judging the production process status. And perform the following fault determinations:

[0108] (1) When the calculated fault characterization coefficient Less than the threshold At that time, the actual state of the return line production process is judged as normal.

[0109] (2) When Greater than the threshold Less than the threshold At that time, the actual state of the return line production process was judged as a minor fault;

[0110] (3) When Greater than the threshold Less than the threshold At that time, the actual state of the return line production process was judged as a medium-level fault;

[0111] (4) When Greater than the threshold At that time, the actual state of the return line production process was judged as an emergency fault.

[0112] V. Experimental Comparison Results

[0113] To further demonstrate the high detection accuracy of the PTFE oil seal lip return line defect detection model proposed in this invention, comparative experiments were conducted with several common image processing algorithms to verify the superior performance of the proposed return line defect detection model. A total of 100 sets of oil seal lip return line defect samples were tested, with each set containing different defect types and numbers. After completing the 100 sets of defect detection experiments, the number of correctly detected defects was divided by the total number of defects to define the defect detection accuracy rate.

[0114] like Figure 3 As shown, to further illustrate that the pixel-level feature extraction method and manufacturing process fault detection system proposed in this invention have a fast and accurate detection rate in the fault detection of the return line production process, this invention is compared with some traditional machine learning algorithms to prove that the method proposed in this invention has good performance in four production process states.

[0115] like Figure 4 As shown in the performance comparison chart, this invention exhibits the best performance in both defect detection and fault detection. This is because, when constructing the return line defect prediction model, this invention adds a KAN-based process parameter feature extraction path, enhancing the process parameter information. In fault detection, the fusion pixel-level feature calculation method designed in this invention can accurately capture defect pixel features, thus providing more complete data support for subsequent fault detection.

[0116] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0117] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for intelligent quality inspection of PTFE oil seal lip return line, characterized in that, Includes the following processes: S1, real-time acquisition of oil seal lip return line images and corresponding process parameters during the PTFE oil seal lip return line process; S2: Input the data from S1 into the trained oil seal lip return line defect detection model. Extract feature vectors of multiple feature scales based on the oil seal lip return line image, and extract auxiliary feature vectors based on process parameters. Fusion detection of multiple feature vectors, output return line defect detection type result, defect severity result, and small-area cropped image containing defect information. S3, input the small-area cropped image obtained in S2 into the production status feature characterization module of the return line, and extract the defect detail features and overall defect features corresponding to the defect image; the defect detail features include the mean of defect grayscale disorder and the defect type similarity coefficient, and the overall defect features include the mean of defect area and the mean of defect perimeter, which are constructed as production status characterization features. S4. Based on the real-time production conditions, construct a dynamic adjustment matrix for defect feature sensitivity. Using the dynamic adjustment matrix for defect feature sensitivity as the weight coefficient, and based on the return line defect detection type results, defect severity, and production status characterization characteristics obtained in S2 and S3, calculate the production process fault characterization factor. By using the time-series characteristic parameter calibration framework, the mean and standard deviation of the fault characterization factors are calculated to obtain the fault characterization coefficient of the PTFE oil seal production process; S5 determines the fault status based on the fault characterization coefficient and the set threshold.

2. The intelligent quality inspection method for the PTFE oil seal lip return line as described in claim 1, characterized in that: The process parameters include the mold clamping pressure coefficient. Melt injection temperature Holding time Cooling medium temperature Mold opening speed Melt injection speed .

3. The intelligent quality inspection method for the PTFE oil seal lip return line as described in claim 1, characterized in that: The defect attributes output by the oil seal lip return line defect detection model include: Rectangular box covering defects This rectangle information specifically refers to the center coordinates of this rectangle. and the width of the rectangle and height ; Defect types Specifically, this defect is categorized into five types: fracture, deformation, hole, wear, and blurred edges. Defect severity This defect is classified into six different severity levels, specifically using numerical values. This is reflected in the numerical value, which indicates the severity of the defect.

4. The intelligent quality inspection method for the PTFE oil seal lip return line as described in claim 1, characterized in that: The oil seal lip return line defect detection model includes a multimodal defect feature perception module and several defect detection heads; The multimodal defect feature perception module includes a flow line image defect feature extraction unit and a process parameter defect feature extraction pathway; the flow line image defect feature extraction unit contains three network pathways with different feature fitting and extraction capabilities, extracting feature vectors of three different scales. , and The process parameter defect feature extraction path extracts auxiliary feature vectors. ; Each defect detection head focuses on detecting only one type of return line defect. In each defect detection head, a KAN layer is first used to further weight the input features, followed by two adaptive convolutional layers with variable kernel coefficients for further feature extraction, and finally a KAN layer and a normalization layer are used for processing to output the defect detection result.

5. The intelligent quality inspection method for the PTFE oil seal lip return line as described in claim 4, characterized in that: The flow line image defect feature extraction unit contains three network pathways with different feature fitting and extraction capabilities. The first feature fitting extraction pathway first includes two interconnected CNN layers, each using a kernel size of 3. These layers are followed by four sequentially connected image defect extraction units. Finally, the first pathway is completed by combining two CNN layers with kernel sizes of 3, a BN normalization layer, and a SILU activation function. The input image... Features were obtained after extraction via pathway one ; The second feature fitting extraction pathway first includes two interconnected CNN layers with a kernel size of 5. These layers then contain four sequentially connected image defect extraction units. Finally, the second pathway is completed by combining two CNN layers with a kernel size of 5, a BN normalization layer, and a ReLU activation function. The input image... Features were obtained after extraction via pathway 2 ; In the third feature fitting extraction pathway, two interconnected CNN layers with a kernel size of 7 are first included. These layers then contain four sequentially connected image defect extraction units. Finally, the third pathway is constructed by combining two CNN layers with a kernel size of 7, a BN normalization layer, and a ReLU activation function. The input image... Features were obtained after extraction via pathway three .

6. The intelligent quality inspection method for the PTFE oil seal lip return line as described in claim 5, characterized in that: The image defect extraction unit first includes four groups of convolutional blocks with different dilatations to take into account both local and global perception capabilities of defect features; then, residual connections and channel splicing layers are used to integrate the output features of the three convolutional network layers to complete the construction of the group of convolutional blocks; residual connections are inserted for each convolutional layer, that is, the output of each convolutional layer in the group of convolutional blocks will be directly connected to the final channel splicing layer. The dilation rate of the first group of convolutional blocks is 1, the dilation rate of the second group of convolutional blocks is 2, the dilation rate of the third group of convolutional blocks is 3, and the dilation rate of the last group of convolutional blocks is 4.

7. The intelligent quality inspection method for the PTFE oil seal lip return line as described in claim 1, characterized in that: The process of extracting the defect details includes: First, crop the image to a smaller area containing defect information. Perform grayscale conversion with 64 grayscale levels. ,in This represents the total number of defects detected for a single PTFE oil seal lip return line; subsequently, regarding the image... For each pixel in the image, calculate the absolute value of the difference between the grayscale value of this pixel and all its neighboring pixels, and then calculate the sum of their absolute values. This sum of absolute values ​​is then used as the sum of the grayscale differences of the corresponding pixels in the improved local grayscale difference matrix, and the process is repeated for each pixel in the image. After collecting all pixels, we obtain the first... A cropped image The corresponding improved local gray-level difference matrix ; Based on the calculated L improved local gray-level difference matrices Calculate each improved local gray-level difference matrix separately. matrix mean To measure the texture disorder of this single defect, the mean of L improved local gray-level difference matrices is then calculated again. The average value is the mean value of the defect grayscale disorder. ; Based on the calculated L improved local gray-level difference matrices Compare the L improved local gray-level difference matrices similarity coefficient between This characterizes the stability of the production state; a larger similarity coefficient indicates a higher similarity between defects in a single sample, meaning a more unstable production state for PTFE oil seals, while a smaller similarity coefficient indicates a lower similarity between defects in a single sample, meaning a more stable production state for PTFE oil seals; defect type similarity coefficient Calculated as L improved local gray-level difference matrices The weighted average of the Frobenius norm distances between each pair of pairs.

8. The intelligent quality inspection method for the PTFE oil seal lip return line as described in claim 1, characterized in that: The process of extracting the global features of the defect includes: Based on the obtained small-area cropped image , Image erosion and dilation algorithms are used to further denoise and smooth the edges of the image. Then, the area of ​​each return line defect is calculated mathematically. ,perimeter Among them, area This represents the number of pixels covered by this return line defect, and its perimeter. This indicates the number of pixels contained in the outermost ring of this return line defect; Then, the calculated L flow line defect area values ​​were... The mean defect area is obtained by solving the mean value. For the calculated L return line defect area values The mean value of the defect perimeter is obtained by solving the mean value. When the average defect area corresponding to the return line of a PTFE oil seal lip is and the average perimeter of defects The higher the value, the higher the risk of failure in the current reflow production state.

9. The intelligent quality inspection method for the PTFE oil seal lip return line as described in claim 1, characterized in that: The specific process of S4 includes: S41, a dynamic adjustment matrix for defect feature sensitivity was constructed for real-time production conditions. : ； in, As the benchmark weight value, This represents the traction speed influence coefficient. Indicates the influence coefficient of material viscosity. These represent the standard production traction speed and the standard material viscosity, respectively. These represent the maximum production traction speed and the maximum standard material viscosity, respectively; production traction speed and material viscosity This represents the production coefficient corresponding to the actual manufacturing process of the return line; S42, based on dynamic adjustment matrix Fault characterization factors for PTFE oil seal lip return line production process constructed under real-time process parameters ; S43, based on the calculation method of the production process fault characterization factor determined in S42, when performing intelligent fault detection in the production process, 100 Fault characterization factors in the production process of PTFE oil seal lip return line Constructed sequence data A comprehensive analysis was conducted to obtain the detection results of production failures on the return line. The sequence was then calculated. The corresponding mean and standard deviation To measure the frequency of defects, the severity level, and the instability of defects in these 100N oil seal lip return line samples, the mean value of the defects is used. Larger, standard deviation The higher the value, the greater the probability of a failure in the production process. S44, using the mean calculated in process S43 Standard deviation After weighted summation, the failure characterization coefficient of the PTFE oil seal production process is obtained. .

10. The intelligent quality inspection method for the PTFE oil seal lip return line as described in claim 1, characterized in that: The S5 setting of a production process status judgment threshold And perform the following fault determinations: When the calculated fault characterization coefficients Less than the threshold At that time, the actual state of the return line production process is judged as normal. when Greater than the threshold Less than the threshold At that time, the actual state of the return line production process was judged as a minor fault; when Greater than the threshold Less than the threshold At that time, the actual state of the return line production process was judged as a medium-level fault; when Greater than the threshold At that time, the actual state of the return line production process was judged as an emergency fault.

Images