Oil filter manufacturing and detecting method based on particle swarm optimization

Abstract

The invention discloses an oil filter manufacturing detection method based on a particle swarm algorithm, which relates to the technical field of mechanical manufacturing quality detection, and comprises the following steps: fusing a stamping dynamic pressure curve and a welding instantaneous current waveform, and carrying out synchronous matching calculation by adopting the particle swarm algorithm to identify an abnormal coupling mode between parameters. And reversely deducing the form and spatial distribution of potential defects based on the mode, and generating a defect prediction map. A microscopic visual probe is guided to carry out self-adaptive zoom imaging on a high-risk area according to the atlas, and the defect microstructure is obtained through cross-scale feature fusion. And quantitatively evaluating potential influence of defects on pressure bearing and sealing performance of the shell by combining a material fatigue model, dividing reliability grades according to a result, and planning a subsequent processing flow. According to the method, early prediction and accurate positioning of defects are realized, and the accuracy of product reliability management and control is improved through quantitative evaluation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mechanical manufacturing quality detection, and specifically relates to an oil filter manufacturing detection method based on a particle swarm algorithm. BACKGROUND

[0002] In the manufacturing process of the oil filter shell, stamping forming and welding are key processes, and the process stability directly determines the quality and reliability of the final product. The existing quality monitoring technology mainly relies on independent threshold judgment of single process parameters. This static monitoring method can capture significant parameter out-of-bounds, but cannot effectively perceive and analyze the complex time sequence correlation and interaction between dynamic parameter sequences in different processes. Due to the lack of deep correlation analysis of cross-process heterogeneous time sequence data, the coupled quality risks induced by fluctuations in the previous process and accumulated or manifested in the subsequent process are difficult to identify.

[0003] The current defect detection method is usually based on appearance inspection or sampling destructive testing of the final product, which belongs to the post-discovery mode. This method cannot predict and locate the defects that may be formed when the product is still in the processing flow. At the same time, the existing technology has not established a quantitative mapping relationship between the abnormal pattern of process parameters and the specific defect morphology and spatial position, resulting in a lack of foresight and pertinence in the detection activity, and low detection efficiency. In addition, for the identified defects, the conventional method also lacks a technical path for quantitatively evaluating their potential functional impact combined with the material mechanics model, so it is difficult to realize accurate classification and differentiated treatment of product reliability. Therefore, a predictive detection method that can early warning, accurate positioning and quantitative evaluation of potential defects in the production process is needed in the manufacturing field. SUMMARY

[0004] The present application aims to at least solve one of the technical problems existing in the prior art; To this end, the present application proposes an oil filter manufacturing detection method based on a particle swarm algorithm, comprising: extracting the multi-batch processing parameter history log recorded when the oil filter shell is circulated on the production line, and fusing the dynamic pressure curve recorded in the stamping forming stage and the instantaneous current waveform recorded in the welding stage; synchronously matching and calculating the fused dynamic pressure curve and instantaneous current waveform by using a particle swarm algorithm, and identifying the abnormal coupling mode between the processing parameters; According to the abnormal coupling mode, the potential defect morphology associated with the oil filter shell and its spatial distribution trend are reversely deduced, and a potential defect prediction map is formed; According to the potential defect prediction map, a plurality of movable microscopic visual probes are deployed to adaptively zoom in on the predicted defect area; Perform cross-scale feature fusion analysis on the collected microscopic image sequence to generate defect micro-morphology and material structure information; Combine the material fatigue model of the oil filter shell and the micro-morphology information to quantitatively evaluate the potential impact of each defect on the pressure-bearing performance and sealing performance of the shell; According to the quantitative evaluation results, the oil filter shell is divided into different reliability levels; According to the reliability level and the beat information of the production line scheduling system, the repair or scrap processing flow of unqualified shells is planned.

[0005] Further, the particle swarm algorithm is used to calculate the synchronization matching of the fused dynamic pressure curve and transient current waveform, and to identify abnormal coupling patterns between processing parameters, specifically including: Segment and align the dynamic pressure curve and transient current waveform according to the time axis to construct signal segment pairs in time window units; Initialize the particle swarm and set the position of each particle to represent a signal segment pair matching offset and similarity threshold combination; Define the fitness function of the particle as the offset and threshold parameters represented by the position of the particle, and calculate the correlation coefficient of the matching signal segment pairs in morphological consistency; Through iterative search of the particle swarm algorithm, find the optimal offset and optimal threshold parameter that maximizes the sum of the overall morphological consistency correlation coefficient; Based on the optimal offset and optimal threshold parameter, the dynamic pressure curve and transient current waveform are re-matched and associated; Analyze the matched and associated signal pairs to identify abnormal coupling segments such as abnormal correspondence between pressure peaks and current valleys, or pressure stable segments accompanied by severe current fluctuations; Cluster all identified abnormal coupling segments to induce abnormal coupling patterns representing different process misalignment types.

[0006] Further, the potential defect morphology and its spatial distribution trend associated with the oil filter shell are inferred based on the abnormal coupling pattern to form a potential defect prediction map, specifically including: Establish a typical defect morphology knowledge base containing crack initiation, micropore aggregation, and insufficient welding penetration. The knowledge base records the historical association rules between each defect morphology and specific processing parameter abnormalities; Match the identified abnormal coupling patterns with the historical association rules in the typical defect morphology knowledge base and calculate the confidence; For the abnormal coupling patterns that match successfully, according to the defect morphology knowledge associated with them, combined with the specific processing sequence of the batch to which the oil filter shell belongs, deduce the process position where the defect may occur; Based on the machining process position and the shell three-dimensional design model, the starting point and the extension direction of the defect on the three-dimensional space of the oil filter shell are mapped out; Comprehensive all matching results, in the surface and internal structure of the shell three-dimensional design model, mark the probability of the existence of specific form defects in different positions in the form of probability cloud map, form the potential defect prediction atlas.

[0007] Further, the potential defect prediction atlas is deployed with multiple movable microscopic visual probes, and adaptive variable imaging is carried out on the predicted defect area, specifically including: Analyzing the potential defect prediction atlas, extracting the center coordinates, area range and predicted defect type of all high probability defect areas; According to the predicted defect type, select the initial optical magnification and illumination scheme from the preset imaging strategy library; Control the six-axis mechanical arm carrying the microscopic visual probe to move to the center coordinates of the target area; Start the probe to carry out low magnification panoramic scanning on the target area, and obtain the area profile image; According to the texture and contrast features in the area profile image, dynamically adjust to the highest magnification suitable for identifying the corresponding microscopic defects of the predicted defect type; Using depth fusion technology, multiple images are collected at different focal planes and synthesized to obtain a clear panoramic depth microscopic image in the entire target area; Record the final imaging parameters, mechanical arm pose and collected image sequence corresponding to each target area.

[0008] Further, the collected microscopic image sequence is subjected to cross-scale feature fusion analysis to generate microscopic morphology and material structure information of the defect, specifically including: Preprocess each microscopic image, including gray scale correction, noise suppression and edge enhancement; Extract the global geometric features of the defect from the low magnification profile image, including the area, perimeter, aspect ratio and main direction of the defect area; Extract the local fine features of the defect from the high magnification clear image, including the roughness of the edge, the directionality of the surface texture, and the deformation or fracture morphology of the material grain; Construct a feature pyramid to align the global geometric features and local fine features in the spatial coordinates and then perform hierarchical fusion; Based on the fused cross-scale feature vector, the microscopic morphology of the defect is structured and described to generate a structured descriptor containing the defect depth, opening width, inner wall inclination angle and bottom morphology; Analyze the deformation characteristics of material grains to infer the type and magnitude of stresses experienced by the material during processing, and generate a report on changes in material structure.

[0009] Furthermore, by combining the material fatigue model of the oil filter housing with the microstructure information, the potential impact of each defect on the pressure-bearing and sealing performance of the housing is quantitatively evaluated, specifically including: The fatigue model of the oil filter housing material, which was established in advance through material experiments, is invoked. The model defines the relationship between the crack propagation rate and life of the material under different stress levels. The structured descriptor of the defect, especially the defect depth and sharp angle, is used as input parameters and substituted into the material fatigue model. The stress concentration factor and crack propagation path at the defect were simulated under cyclic load at the rated working pressure of the oil filter. The estimated number of cycles required to expand from the initial defect to the critical failure size is calculated as a quantification of the durability of the pressure-bearing performance; For defects located on the sealing surface, the degree of interference with the compression and rebound characteristics of the sealing gasket is evaluated based on its opening width and inner wall morphology, and the leakage channel of the sealing medium under compression is simulated. The percentage decrease in theoretical sealing pressure when defects exist is calculated as a quantitative indicator of sealing performance reliability.

[0010] Furthermore, the oil filter housing is classified into different reliability levels based on the quantitative evaluation results, specifically including: Set the pass thresholds for quantitative indicators of pressure-bearing performance durability and quantitative indicators of sealing performance reliability; The quantitative evaluation results of the filter housing for each oil are compared with the qualified threshold. For housings whose quantitative indicators of both pressure bearing capacity and sealing performance are better than the qualified threshold, they are marked as high reliability level; For a housing with only one quantitative indicator that is better than the qualified threshold and another that is close to the qualified threshold, it is marked as a medium reliability level; For any quantitative indicator below the qualified threshold, or for two indicators close to but not reaching the qualified threshold, the shell is marked as low reliability level; For housings with serious defects and whose quantitative assessment results indicate an immediate risk of failure, they are marked as failure levels; Generate a classification list that includes a unique identifier for each housing, specific values ​​for various quantitative indicators, and the final reliability level.

[0011] Furthermore, the process for planning the rework or scrapping of defective shells based on the reliability level and the cycle time information of the production line scheduling system specifically includes: The current production cycle time, the length of the subsequent workstation queue, and the idle status of the rework workstation can be obtained in real time from the production line scheduling system. Read the classification list and filter out oil filter housings marked with low reliability and failure levels; For housings with low reliability levels, appropriate local repair or reinforcement processes are matched from the available rework process library based on the specific location and type of defects. Based on the time required for the matching rework process, and under the premise of meeting the overall production cycle time, the optimal offline rework insertion time and rework path are calculated for each shell to be reworked. For shells with fault levels, a scrapping command is generated directly, and the path and time for removing them from the main line and transporting them to the waste recycling area are planned. All planned rework and scrap instructions, routes, and schedules are synchronized to the production line scheduling system and material handling system.

[0012] Furthermore, the analyzed and matched signal pairs identify abnormal coupling segments where pressure peaks and current troughs do not correspond normally, or where stable pressure periods are accompanied by violent current fluctuations. Specifically, these include: Calculate the local statistical characteristics of each pair of dynamic pressure curve segments and instantaneous current waveform segments after matching, including the mean, variance, and kurtosis of the pressure segment, and the mean, variance, and zero-crossing rate of the current segment. Set a normal time correspondence window for pressure peak and current valley, and check whether the time difference between the pressure peak point and the current valley point in the signal alignment exceeds the normal correspondence window. Set a variance threshold for the pressure stability segment, and check whether the variance of the corresponding current waveform exceeds its normal fluctuation range within the stability segment where the pressure variance is lower than the variance threshold. Signal pairs whose time difference exceeds the window or whose current fluctuations are abnormal are marked as abnormal candidate segments; The abnormal candidate segments are manually verified using process knowledge to exclude reasonable fluctuations caused by normal process adjustments, and finally the true abnormal coupling segments are confirmed.

[0013] Furthermore, the step of dynamically adjusting the magnification to the most suitable high magnification for identifying the microscopic defects corresponding to the predicted defect type based on the texture and contrast features in the region overview image specifically includes: The acquired regional overview image is divided into blocks, the gray-level co-occurrence matrix of each image sub-block is calculated, and contrast and homogeneity features are extracted. Based on the predicted defect type, determine the typical texture representation pattern of the predicted defect type under a microscope, such as the linear low-contrast features of cracks and the circular high-contrast edge features of pores. The similarity between the texture features of image sub-blocks and the typical texture representation patterns of defects is calculated. Select the region containing the image sub-block with the highest similarity as the key observation area; Based on the physical size of the key observation area and the desired level of detail, the required total magnification is calculated by combining the microscope's objective lens magnification and digital magnification capability. Control the microscope vision probe to switch to the objective lens of the corresponding magnification, and adjust the illumination angle to enhance the contrast of the key observation area.

[0014] Compared with the prior art, the beneficial effects of the present invention are: By employing a particle swarm optimization algorithm to perform global synchronous matching calculations on heterogeneous dynamic parameter sequences across processes, the optimal correspondence between stamping pressure curves and welding current waveforms can be adaptively searched. This technology overcomes the limitations of traditional single-parameter threshold monitoring, enabling quantitative identification of nonlinear abnormal coupling modes from complex temporal changes, and effectively improving the early identification capability of process chain interaction faults.

[0015] Based on identified abnormal coupling patterns, the morphology and spatial distribution trends of potential defects are mapped backward, generating a priori defect prediction maps. This technique establishes a mapping relationship from abstract parameter anomalies to specific defect features, enabling detection resources to be precisely targeted at high-risk areas according to the map's guidance, thereby improving the detection efficiency and location accuracy of micro-defects.

[0016] Microscopic imaging of the predicted area and fusion of multi-scale features, combined with a material fatigue model, were used to quantitatively assess the microscopic morphology of the defects. This process enabled quantitative analysis from defect morphology to functional impact, providing quantitative indicators of potential risks to the pressure-bearing and sealing performance of each shell. This provides an objective data basis for product reliability grading and supports the accurate formulation of subsequent processing decisions. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of the oil filter manufacturing and testing method based on particle swarm optimization algorithm described in this invention. Figure 2 A flowchart for identifying anomalous coupling patterns; Figure 3 A flowchart for adaptive zoom imaging; Figure 4 A cross-scale feature fusion analysis diagram of oil filter defects; Figure 5 This is a graph showing the efficiency and frequency of use of oil filter repair processes. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 This invention provides a manufacturing inspection method for oil filters based on particle swarm optimization (PSO). The method first extracts historical logs of processing parameters from multiple batches of the oil filter housing as it flows through the production line. These logs integrate dynamic pressure curves recorded during the stamping stage and instantaneous current waveforms recorded during the welding stage. PSO is then used to perform synchronization matching calculations on the fused dynamic pressure curves and instantaneous current waveforms, aiming to identify abnormal coupling patterns between processing parameters. Based on the identified abnormal coupling patterns, the morphology and spatial distribution trend of potential defects associated with the oil filter housing are deduced, forming a potential defect prediction map. According to this map, multiple sets of movable microscopic vision probes are deployed to perform adaptive zoom imaging on the predicted defect areas in the map. Cross-scale feature fusion analysis is performed on the acquired microscopic image sequences to generate information on the microscopic morphology and material structure of the defects. Combining the material fatigue model of the oil filter housing with the aforementioned microscopic morphology information, the potential impact of each defect on the housing's pressure-bearing and sealing performance is quantitatively assessed. Based on the quantitative assessment results, the oil filter housing is classified into different reliability levels. Based on the identified reliability levels and the real-time cycle information of the production line scheduling system, a rework or scrapping process for defective shells is planned.

[0020] In one embodiment of the present invention, see [reference] Figure 2 The dynamic pressure curve and instantaneous current waveform are segmented and aligned along the production line time axis. Signal segment pairs are constructed with a one-second time window. Each dynamic pressure curve segment is paired with the corresponding instantaneous current waveform segment within the time window. The particle swarm is initialized with a population size of fifty. Each particle's position is a two-dimensional vector. The first dimension represents the matching offset of the signal segment pair, ranging from -500 milliseconds to +500 milliseconds. The second dimension represents the similarity threshold, ranging from 0.6 to 0.95. The fitness function for each particle is defined as the sum of the Pearson correlation coefficients for morphological consistency across all matched signal segment pairs, based on the offset represented by the particle's position and the threshold parameter. The correlation coefficient calculation excludes signal segment pairs that cannot be aligned after offset adjustment.

[0021] In some embodiments, the optimal parameters are found through iterative search using a particle swarm optimization algorithm. The algorithm sets the maximum number of iterations to 200, the inertia weight decreases linearly from 0.9 to 0.4, and both the individual learning factor and the social learning factor are set to 2.0. In each iteration, the velocity and position of each particle are updated, and the fitness value is recalculated. After the iteration terminates, the position coordinates corresponding to the particle with the highest fitness value are selected as the global optimal solution, i.e., the optimal offset and the optimal similarity threshold. Based on the obtained optimal offset, the dynamic pressure curve is time-shifted, and all signal segment pairs are screened based on the optimal similarity threshold. Only segment pairs with a correlation coefficient greater than the threshold are confirmed as valid matching associations.

[0022] Optionally, the matched signal pairs are analyzed, and local statistical characteristics are calculated for each pair of dynamic pressure curve segments and instantaneous current waveform segments. The characteristics of the dynamic pressure curve segments include mean, variance, and kurtosis, while the characteristics of the instantaneous current waveform segments include mean, variance, and zero-crossing rate. A normal time correspondence window of 50 milliseconds is set for the pressure peak and current trough values. The time difference between the peak point within the dynamic pressure curve segment and the trough point within the instantaneous current waveform segment in each signal pair is checked. Signal pairs with an absolute time difference greater than 50 milliseconds are marked as candidate segments for time synchronization anomalies. Simultaneously, a dynamic pressure curve segment variance below 0.1 MPa² is set as a criterion for determining a stable pressure segment. The variance of the instantaneous current waveform segments corresponding to segments meeting this criterion is checked to see if it exceeds 0.5 ampere². Signal pairs with current variance exceeding this limit are marked as candidate segments for fluctuation anomalies.

[0023] Understandably, the marked anomalous candidate segments undergo manual verification using process knowledge. Process engineers, based on the daily production plan and equipment maintenance records, check whether the anomalous time points involved planned mold replacements, welding parameter adjustments, or brief equipment start-ups and shutdowns. This eliminates reasonable signal fluctuations caused by these normal process adjustments. The remaining candidate segments that cannot be explained by known normal events are ultimately confirmed as the true anomalous coupling segments. All confirmed anomalous coupling segments are then subjected to hierarchical clustering based on dynamic time warping distance to summarize anomalous coupling patterns representing different types of process detuning. The particle fitness function is calculated using the following formula: ; in: Indicates the first The fitness value of each particle. Indicates the time window number. Indicates the total number of time windows. Indicates the first Within a given time window, the Pearson correlation coefficient between the dynamic pressure curve segment (adjusted for offset) and the instantaneous current waveform segment. Indicates the first The similarity threshold represented by each particle position. It is a discriminant function, when The function value is 1 when the condition is met, and 0 otherwise.

[0024] In practice, after clustering, an abnormal coupling pattern library is formed. Each pattern is associated with a set of time segments in which it occurs and the mean of the calculated feature statistics. The pattern library is stored in a database for subsequent use by the defect derivation module. In each new detection task, the acquired dynamic pressure curve and instantaneous current waveform data undergo the complete matching and anomaly identification process described above. The identified new abnormal coupling patterns are compared with existing patterns in the pattern library. If the similarity is lower than a set threshold, the new pattern is added to the library, enabling continuous updates to the pattern library. The entire synchronization matching calculation process based on the particle swarm optimization algorithm is run on a dedicated industrial computer, and the calculation time is constrained within the time range allowed by the production cycle.

[0025] In one embodiment of the present invention, a knowledge base of typical defect morphologies including crack initiation, micropore aggregation, and insufficient weld penetration is established. The knowledge base of typical defect morphologies is stored in a relational database. Each row in the database records a historical association rule between a defect morphology and a specific processing parameter anomaly. The historical association rule is stored in a structured manner with fields such as rule number, defect morphology name, description of associated abnormal coupling mode features, number of historical cases, and average confidence level. For example, a historical association rule may be described as "when the dynamic pressure curve shows a plateau period missing in the second stage of stamping, and the instantaneous current waveform shows a double peak phenomenon in the main welding stage, 75% of the cases in the historical statistics correspond to insufficient weld penetration defects at the circumferential weld of the shell".

[0026] In some embodiments, the identified abnormal coupling patterns are matched one by one with historical association rules in the typical defect morphology knowledge base, and confidence scores are calculated. The matching process first compares the textual feature descriptions of the abnormal coupling patterns with the feature description fields in the historical association rules using keywords and semantic similarity. Rules with highly overlapping feature descriptions proceed to the confidence score calculation process. The confidence score calculation comprehensively considers the number of historical cases, the historical average confidence score, and the similarity score of the current match to generate a final confidence score for the current match.

[0027] Optionally, for successfully matched abnormal coupling patterns, based on the associated defect morphology knowledge and the specific processing sequence of the batch to which the oil filter housing belongs, the process location where the defect may occur can be deduced. Specifically, the production execution system log is queried to obtain the precise timestamps of the batch of housings at the stamping and welding stations. The time point of occurrence of the abnormal coupling pattern is mapped to the specific process step. For example, an abnormal coupling pattern of "gradual pressure increase accompanied by intermittent current drop" identified at timestamp T1, if T1 falls within the time interval of the "longitudinal seam welding of the housing" process recorded in the production log, then the potential defect location is deduced to be in the longitudinal seam weld area.

[0028] It is understandable that, based on the location of the processing steps and the three-dimensional design model of the housing, the starting point and direction of the defect in the three-dimensional space of the oil filter housing are mapped. During operation, the computer-aided design model file is called, and the process location information is associated with the feature surfaces and edges of the model. For example, the process of "longitudinal seam welding of the housing cylinder" corresponds to a straight line edge in space on the model. Using this edge as a reference, based on the historical defect statistics showing that such defects usually extend along the weld direction, a linear region extending 10 mm in both the positive and negative directions along this edge is set as the predicted direction and range of defect expansion.

[0029] In practical implementation, by integrating all matching results, the probability of specific morphological defects existing at different locations is marked on the surface and internal structure of the shell 3D design model in the form of probability cloud maps, forming a potential defect prediction map. The calculation of probability values ​​depends on the matching confidence, the spatial distribution probability in historical statistics, and the size distribution model of similar defects. The model surface is triangularly meshed, and each mesh node is assigned a probability value of one or more defect types. The formula for generating the potential defect prediction map is: ; in: Representing vertices on a 3D design model A comprehensive probability index indicating the existence of at least one prediction defect. This represents the total number of historical association rules that successfully matched this shell. Indicates the first The confidence score of the historical association rules that have successfully matched in the current case. Indicates according to the first The defect morphology knowledge associated with each rule, where the defect appears at the vertices. The prior statistical probability of the spatial location. After calculation, it is mapped using color. Value visualization is achieved by rendering a color probability cloud map overlaid on the 3D model, which is the final visualized potential defect prediction map. In the map, the red highlighted areas indicate high-probability defect risk areas, and the blue areas represent low-probability areas.

[0030] In one embodiment of the present invention, see [reference] Figure 3 The system analyzes the potential defect prediction map and extracts information on all defect regions with a probability value higher than 70% from the stored probability cloud map data file. This information includes the 3D coordinates of the center point of each region in the world coordinate system, the size of the circumscribed cuboid region based on the center point, and the predicted main defect type for that region. For example, a record might have center coordinates (x1, y1, z1), a region size of 20 mm x 20 mm x 2 mm, and a predicted defect type of "microcrack cluster". Based on the predicted defect type, an initial optical magnification and illumination scheme are selected from a pre-set imaging strategy library. The imaging strategy library is a configuration file where "microcrack cluster" corresponds to an initial magnification of 50x and an illumination scheme of low-angle ring LED cold light.

[0031] In some embodiments, a six-axis robotic arm equipped with a microscopic vision probe is controlled to move to the center coordinates of the target area. After receiving the coordinate command, the robotic arm controller converts the world coordinates into joint angles through inverse kinematics, driving the robotic arm to move precisely to the target position in a point-to-point motion manner, with a positioning accuracy requirement better than 0.1 millimeters. The probe is then activated to perform a low-magnification panoramic scan of the target area. At a magnification of 50x, the microscopic vision probe performs a raster-style path scan within the area using a built-in two-dimensional translation stage, automatically stitching together a low-magnification overview image covering the entire target area.

[0032] Optionally, the acquired regional overview image is segmented into blocks, uniformly dividing the entire image into 10-pixel by 10-pixel sub-blocks. The gray-level co-occurrence matrix (GLCM) of each sub-block is calculated, and contrast and homogeneity feature values ​​are extracted based on the GLCM. According to the predicted defect type, the typical texture representation pattern of the predicted defect type under a microscope is determined. For example, the typical texture representation pattern of a "microcrack cluster" is defined as a texture feature with a linear orientation and local contrast lower than the background average by 30%. The texture feature vector of each sub-block, i.e., the contrast and homogeneity feature values, is then compared with the feature vector of the typical texture representation pattern of the defect using cosine similarity calculation.

[0033] Understandably, the region containing the image sub-block with the highest similarity is selected as the key observation area. The system records the pixel coordinates of this sub-block in the overview image and maps them back to the actual physical coordinate range of the target area through coordinate transformation. Based on the physical size of the key observation area and the desired level of detail, combined with the microscope's objective magnification and digital magnification capability, the required total magnification is calculated. If the physical size of the key observation area is two millimeters by two millimeters, the desired level of detail is 0.5 micrometers, the microscope's objective magnification is variable, and the digital magnification capability is ten times, then calculations determine that a 200x objective lens needs to be switched to, and five times digital magnification needs to be enabled to achieve a total magnification of 1000 times. This calculation process follows the formula below: ; in: This indicates that the final magnification that the microscope vision probe needs to achieve needs to be controlled. This indicates the physical dimensions and length of the key observation area within the current objective lens's field of view. This represents the physical length of the target detail that you wish to distinguish. This indicates the digital magnification that the system can enable. The result of the formula is used to guide the switching of objectives.

[0034] In practice, the microscope vision probe is switched to the objective lens of the corresponding magnification, and the illumination angle is adjusted to enhance the contrast of key observation areas. The objective lens turret rotates the 200x objective lens into the optical path according to instructions. Simultaneously, the illumination subsystem adjusts the incident angle of the low-angle ring light from 30 degrees to 15 degrees to highlight surface undulations. Depth-of-field fusion technology is employed, acquiring and synthesizing multiple images at different focal length planes. The microscope vision probe, controlled by a piezoelectric ceramic actuator, steps the lens in a direction perpendicular to the object plane, acquiring eleven images in two-micrometer increments from -10 micrometers to +10 micrometers. These eleven images are then synthesized into a single, clear, full-depth-of-field microscope image using a Laplacian pyramid-based image fusion algorithm. The final imaging parameters, robotic arm pose, and acquired image sequence for each target area are recorded. This data is stored in a folder named after the unique identifier of the oil filter housing, containing a log file detailing the objective lens magnification, illumination parameters, robotic arm joint angles, and image file paths.

[0035] In one embodiment of the present invention, each acquired microscopic image is preprocessed. The preprocessing operations are performed sequentially as follows: grayscale correction, noise suppression, and edge enhancement. Grayscale correction uses a linear transformation based on a standard grayscale plate to map the image grayscale values ​​to a standard range of 0 to 255. Noise suppression uses a Gaussian filter with a size of three pixels by three pixels to convolve the image to smooth high-frequency noise. Edge enhancement uses the Laplacian operator to perform a second differentiation on the filtered image and superimpose it back onto the original image to highlight the defect boundary features.

[0036] In some embodiments, global geometric features of defects are extracted from low-magnification overview images. The Otsu method is used to perform binarization segmentation on the preprocessed low-magnification image to identify the defect region and the background. The pixel area of ​​the defect region is calculated and multiplied by the actual physical size calibration coefficient corresponding to each pixel to obtain the area in square millimeters. The pixel perimeter of the defect region outline is calculated and multiplied by the physical calibration coefficient to obtain the perimeter in millimeters. The ratio of the long side to the short side of the smallest bounding rectangle that can completely enclose the defect region is calculated to obtain the aspect ratio. The principal axis direction angle determined by the second central moment of the defect region is calculated to obtain the principal direction.

[0037] Optionally, local fine features of defects are extracted from high-magnification clear images. On the pre-processed high-magnification image, samples are taken at equal intervals along the edge contour of the identified defects. The variance of the change in the direction of the line connecting adjacent sampling points is calculated as the edge roughness feature value. A gray-level gradient direction histogram is calculated for the defect surface area. The main peak direction of the histogram is used as the directional feature of the surface texture. For the metal material area, individual grains are distinguished by corrosion and threshold segmentation. The average equivalent strain of grains or the statistical grain boundary fracture ratio is calculated as the deformation or fracture morphology feature of the material grains.

[0038] It is understandable that constructing a feature pyramid involves aligning global geometric features with local fine features in spatial coordinates and then performing hierarchical fusion. The low-magnification overview image is used as the bottom layer of the pyramid, and the high-magnification clear image is used as the top layer. Image registration technology is used to map the feature point coordinates of the high-magnification image to the coordinate space of the low-magnification image, generating a multi-scale feature vector. The feature vector is organized according to spatial location and includes area, perimeter, aspect ratio, principal direction, edge roughness, texture directionality, and grain deformation morphology feature values ​​from the bottom layer to the top layer.

[0039] In practical implementation, the microscopic morphology of defects is structurally described based on the fused cross-scale feature vectors. An active contour model is used to fit the three-dimensional contour of the defect edges in high-magnification images. The depth of the defect is calculated using focal length variation data. The pixel distance at the narrowest point of the contour is measured and converted to obtain the opening width. The average angle between the defect sidewall and the horizontal plane is calculated to obtain the inner wall tilt angle. The curvature distribution of the bottom region of the contour is analyzed to determine whether the bottom morphology is sharp, smooth, or plateau, generating a structured descriptor containing the above parameters. The deformation characteristics of the material grains are analyzed to infer the type and magnitude of stress experienced by the material during processing. The direction of grain elongation indicates the direction of principal stress, and the density of grain boundary slip bands is related to the magnitude of shear stress, forming a material structure change report describing the local plastic deformation and damage of the material.

[0040] In some embodiments, a fatigue model of the oil filter housing material, pre-established through material experiments, is invoked. This fatigue model is stored as a tabular function, defining the relationship between crack propagation rate and lifespan under different stress levels. The structured descriptor of the defect, particularly its depth and sharpness angle, is substituted into the material fatigue model as input parameters. The variation of stress concentration factor and crack propagation path at the defect are simulated under cyclic loading at the rated operating pressure of the oil filter. The simulation is performed using finite element analysis software, converting the structured descriptor into an initial crack in a three-dimensional solid model, and applying sinusoidal cyclic pressure load boundary conditions for fracture mechanics simulation.

[0041] Optionally, the estimated number of cycles required for the initial defect to expand to the critical failure size is calculated as a quantification of the durability of the pressure-bearing performance. The critical failure size is determined based on the material's fracture toughness calculation, and the estimated number of cycles is obtained using the formula for the crack propagation rate. For defects located on the sealing surface, their interference with the compression and rebound characteristics of the sealing gasket is assessed based on their opening width and inner wall morphology. The leakage path of the sealing medium under compression is simulated, and the flow of the micro-gap between the sealing gasket and the defect after compression is analyzed using fluid dynamics software. The percentage decrease in theoretical sealing pressure when defects are present is calculated as a quantification of the reliability of the sealing performance. The percentage decrease in theoretical sealing pressure is obtained by comparing the difference between the ideal contact area without defects and the actual effective contact area with defects under the same clamping force. Calculated using the following formula: ; in: This indicates the percentage decrease in theoretical sealing pressure. Indicates the complete sealing contact area required by the design. This represents the effective sealing contact area calculated using fluid dynamics simulations when defects are present. See Table 1 for a quantitative assessment of different defects on a specific housing.

[0042] Table 1. Results of Defect Quantitative Assessment; See Figure 4 This is a cross-scale feature fusion analysis chart of oil filter defects. This grouped bar chart compares the standardized feature values ​​of low-ratio global features (blue), high-ratio local features (red), and the fused cross-scale features (green) across seven key dimensions, reflecting the improved defect identification capability of cross-scale feature fusion. In all feature dimensions, the feature values ​​of the fused cross-scale features (green) consistently fall between those of the low-ratio global features and the high-ratio local features, with most dimensions closer to the high-ratio local features. This indicates that the fused features retain global geometric information while enhancing the expression of local details. The fused feature values ​​more comprehensively reflect the morphological and structural information of the defects, providing more reliable input for subsequent assessments of pressure resistance and sealing performance.

[0043] In one embodiment of the present invention, qualified thresholds are set for the quantitative indicators of pressure-bearing performance durability and sealing performance reliability. The qualified threshold for the quantitative indicator of pressure-bearing performance durability is set as no expansion after two million cycles under rated working pressure, and the qualified threshold for the quantitative indicator of sealing performance reliability is set as the percentage decrease in theoretical sealing pressure not exceeding five percent. The quantitative evaluation results of each oil filter housing are compared with the qualified thresholds. The comparison operation is performed automatically by the program. The program reads the defect quantitative evaluation result file corresponding to each housing, which contains data such as defect identification, estimated number of cycles, and percentage decrease in sealing pressure. The program takes the smallest estimated number of cycles among all defects of each housing as the representative value of the pressure-bearing performance of the housing, and the largest percentage decrease in sealing pressure among the defects located on the sealing surface as the representative value of the sealing performance of the housing.

[0044] In some embodiments, housings with both pressure-bearing and sealing performance quantification indicators exceeding the acceptable threshold are marked as high reliability. A quantification indicator exceeding the acceptable threshold means that the representative value of the housing's pressure-bearing performance is greater than two million cycles and the representative value of its sealing performance is less than 5%. Housings with only one quantification indicator exceeding the acceptable threshold and the other close to the acceptable threshold are marked as medium reliability. "Close to the acceptable threshold" is defined as a representative value of pressure-bearing performance between 1.8 million and 2 million cycles or a representative value of sealing performance between 5% and 6%. Housings with any quantification indicator below the acceptable threshold, or both close to but not reaching the acceptable threshold, are marked as low reliability. Housings with a serious defect quantification assessment indicating an immediate risk of failure are marked as faulty. "Immediate risk of failure" is defined as a defect with an estimated cycle count less than 50,000 cycles or a sealing pressure drop percentage greater than 20%. A classification list containing a unique identifier for each housing, specific values ​​for each quantification indicator, and the final reliability level is generated. This classification list is stored in a database in tabular form and published to all terminals on the production line via a network interface.

[0045] Optionally, the production line scheduling system can be used to obtain the current production cycle's subsequent workstation queue length and rework station idle status in real time. The production cycle is in seconds, the workstation queue length is represented by the number of housings to be processed, and the rework station idle status is a Boolean value. The system reads a classification list and filters out oil filter housings marked with low reliability and fault levels. The filtering operation is based on the reliability level field in the classification list, generating a sequence of housings to be processed.

[0046] Understandably, for shells with low reliability levels, suitable local repair or reinforcement processes are matched from the available rework process library based on their specific defect locations and types. The rework process library stores the mapping relationship between various defect types, locations, and available processes such as laser cladding, micro-hammering, and local heat treatment. The matching process involves querying the detailed defect location and type description based on the defect identifier, and then searching for the process with a perfect suitability score in the rework process library as the recommended process. Based on the time required for the matched rework process, the optimal offline rework insertion time and rework path are calculated for each shell to be reworked, while meeting the overall production cycle time. The calculation is based on a heuristic scheduling algorithm, with the goal of minimizing interference with the main production cycle time. The algorithm inputs include the rework process time, the current production cycle time, and the queue length of the subsequent workstations. The output is the time window for each shell to be reworked to enter the offline rework workstation and the material handling path from the main line to its return.

[0047] In practice, for shells with specific fault levels, a scrapping instruction is directly generated, and the path and time for moving them off the main line to the scrap recycling area are planned. The scrapping instruction includes a unique shell identifier and a scrapping reason code. Path planning uses a time window-based path planning algorithm, which avoids conflicts with normal logistics and rework logistics paths, and plans the shortest time path and estimated arrival time from the current inspection station to the scrap recycling area. All planned rework and scrapping instruction paths and schedules are synchronized to the production line scheduling system and the material handling system. This is done synchronously through the Manufacturing Execution System's application programming interface (API) in the form of a message queue. Upon receiving the message, the production line scheduling system updates its scheduling plan, and the material handling system controls the automated guided vehicles (AGVs) to execute the transport tasks. The calculation of the optimal offline rework insertion time point follows the following formula: ; in: This represents the calculated optimal offline rework insertion time. Indicates the current system time. This indicates the current queue length (in terms of the number of workpieces) for subsequent critical workstations on the main line. This indicates the number of parallel processing units at critical workstations. Indicates the production cycle time. This represents the estimated processing time required for the matched rework process. The goal of the formula is to schedule rework during the processing gaps in the mainline critical station queues to reduce congestion.

[0048] See Figure 5 This is a chart analyzing the efficiency and usage frequency of oil filter rework processes. This dual-axis grouped bar chart quantitatively compares the performance of five rework processes in actual production from two dimensions: the number of times the process is used and the process time, providing data support for process selection and scheduling optimization. For high-frequency defects (such as surface cracks and sealing surface damage), micro-hammering or sealing surface grinding is preferred to ensure rework efficiency. For low-frequency severe defects (such as insufficient weld penetration), welding repair is retained as an alternative, but its usage frequency must be strictly controlled to reduce risk. Since the time consumption of all processes is similar, the capacity bottleneck of the rework station mainly depends on the number of equipment and parallel processing capability. It is recommended to prioritize the configuration of dedicated equipment for micro-hammering and sealing surface grinding to match their high-frequency requirements. The usage data of high-frequency processes reveals the main failure modes in production, which can be used to optimize stamping and welding processes, reduce the occurrence of surface damage and sealing defects, and reduce rework requirements from the source.

[0049] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A manufacturing and testing method for oil filters based on particle swarm optimization algorithm, characterized in that, Includes the following steps: Extract the historical logs of multiple batches of processing parameters recorded when the oil filter housing is flowing on the production line, and integrate the dynamic pressure curve recorded in the stamping stage with the instantaneous current waveform recorded in the welding stage; The particle swarm optimization algorithm is used to perform synchronization matching calculations on the fused dynamic pressure curve and instantaneous current waveform to identify abnormal coupling modes between processing parameters. Based on the abnormal coupling mode, the potential defect morphology and spatial distribution trend associated with the oil filter housing are deduced in reverse, forming a potential defect prediction map. Based on the potential defect prediction map, multiple sets of movable microscopic vision probes are deployed to perform adaptive zoom imaging on the predicted defect area. Cross-scale feature fusion analysis is performed on the acquired microscopic image sequence to generate microscopic morphology and material structure information of defects; By combining the material fatigue model of the oil filter housing with the microstructure information, the potential impact of each defect on the pressure-bearing and sealing performance of the housing is quantitatively evaluated. Based on the quantitative assessment results, oil filter housings are classified into different reliability levels; Based on the reliability level and the cycle time information of the production line scheduling system, a rework or scrapping process for defective shells is planned.

2. The oil filter manufacturing and testing method based on particle swarm optimization algorithm according to claim 1, characterized in that, The process employs a particle swarm optimization algorithm to perform synchronous matching calculations between the fused dynamic pressure curve and the instantaneous current waveform, identifying abnormal coupling modes between processing parameters. Specifically, this includes: The dynamic pressure curve and the instantaneous current waveform are segmented and aligned along the time axis to construct signal segment pairs in units of time windows; Initialize the particle swarm, and set the position of each particle to represent a signal segment, matching the offset and similarity threshold combination; The fitness function of a particle is defined as the correlation coefficient of the matched signal fragment pairs in terms of morphological consistency based on the offset represented by the position of the particle and a threshold parameter. The optimal offset and optimal threshold parameters that maximize the sum of correlation coefficients of overall morphological consistency are found through iterative search using the particle swarm optimization algorithm. Based on the optimal offset and optimal threshold parameters, the dynamic pressure curve and the instantaneous current waveform are re-matched and associated. By analyzing the matched signal pairs, abnormal coupling segments were identified where the pressure peak and current valley did not correspond normally, or where the pressure was stable and the current fluctuated violently. All identified anomalous coupling segments are clustered to summarize anomalous coupling patterns representing different types of process detuning.

3. The oil filter manufacturing and testing method based on particle swarm optimization algorithm according to claim 1, characterized in that, The process of reversely deducing the potential defect morphology and spatial distribution trend associated with the oil filter housing based on the abnormal coupling mode to form a potential defect prediction map specifically includes: Establish a knowledge base of typical defect morphologies including crack initiation, micropore aggregation, and insufficient weld penetration. The knowledge base records the historical association rules between each defect morphology and specific processing parameter anomalies. The identified abnormal coupling patterns are matched one by one with the historical association rules in the typical defect morphology knowledge base, and confidence is calculated. For a successfully matched abnormal coupling pattern, based on the associated defect morphology knowledge and the specific processing sequence of the batch to which the oil filter housing belongs, the process location where the defect may occur can be deduced. Based on the location of the processing steps and the three-dimensional design model of the housing, the starting point and expansion direction of the defect in the three-dimensional space of the oil filter housing are mapped. By combining all matching results, the probability of specific morphological defects existing at different locations is marked on the surface and internal structure of the shell three-dimensional design model in the form of probability cloud maps, forming the potential defect prediction map.

4. The oil filter manufacturing and testing method based on particle swarm optimization algorithm according to claim 1, characterized in that, The step of deploying multiple sets of movable microscopic vision probes based on the potential defect prediction map to perform adaptive zoom imaging on the predicted defect area specifically includes: Analyze the potential defect prediction map to extract the center coordinates, area range, and predicted defect type of all high-probability defect regions; Based on the predicted defect type, an initial optical magnification and illumination scheme are selected from a pre-set imaging strategy library; Control the movement of a six-axis robotic arm equipped with a microscopic vision probe to the center coordinates of the target area; The probe is activated to perform a low-magnification panoramic scan of the target area to obtain an overview image of the area. Based on the texture and contrast features in the region overview image, the magnification is dynamically adjusted to the most suitable high magnification for identifying the micro-defects corresponding to the predicted defect type. By employing depth-of-field fusion technology, multiple images are acquired and synthesized at different focal length planes to obtain a clear panoramic depth-of-field microscopic image of the entire target area; Record the final imaging parameters, robotic arm pose, and acquired image sequence for each target area.

5. The oil filter manufacturing and testing method based on particle swarm optimization algorithm according to claim 1, characterized in that, The process of performing cross-scale feature fusion analysis on the acquired microscopic image sequence to generate microscopic morphology and material structure information of defects specifically includes: Each microscopic image is preprocessed, including grayscale correction, noise suppression, and edge enhancement; Extract global geometric features of defects from low-magnification overview images, including the area, perimeter, aspect ratio, and principal orientation of the defect region; Extracting local fine features of defects from high-magnification, high-resolution images, including edge roughness, surface texture directionality, and deformation or fracture morphology of material grains; Construct a feature pyramid, and then perform hierarchical fusion of global geometric features and local fine features after aligning them in spatial coordinates; Based on the fused cross-scale feature vectors, the micromorphology of the defect is described in a structured way, generating a structured descriptor that includes defect depth, opening width, inner wall tilt angle, and bottom morphology. Analyze the deformation characteristics of material grains to infer the type and magnitude of stresses experienced by the material during processing, and generate a report on changes in material structure.

6. The oil filter manufacturing and testing method based on particle swarm optimization algorithm according to claim 1, characterized in that, The material fatigue model of the oil filter housing, combined with the microstructure information, is used to quantitatively assess the potential impact of each defect on the pressure-bearing and sealing performance of the housing, specifically including: The fatigue model of the oil filter housing material, which was established in advance through material experiments, is invoked. The model defines the relationship between the crack propagation rate and life of the material under different stress levels. The structured descriptor of the defect, especially the defect depth and sharp angle, is used as input parameters and substituted into the material fatigue model. The stress concentration factor and crack propagation path at the defect were simulated under cyclic load at the rated working pressure of the oil filter. The estimated number of cycles required to expand from the initial defect to the critical failure size is calculated as a quantification of the durability of the pressure-bearing performance; For defects located on the sealing surface, the degree of interference with the compression and rebound characteristics of the sealing gasket is evaluated based on its opening width and inner wall morphology, and the leakage channel of the sealing medium under compression is simulated. The percentage decrease in theoretical sealing pressure when defects exist is calculated as a quantitative indicator of sealing performance reliability.

7. The oil filter manufacturing and testing method based on particle swarm optimization algorithm according to claim 1, characterized in that, The oil filter housing is classified into different reliability levels based on the quantitative evaluation results, specifically including: Set the pass thresholds for quantitative indicators of pressure-bearing performance durability and quantitative indicators of sealing performance reliability; The quantitative evaluation results of the filter housing for each oil are compared with the qualified threshold. For housings whose quantitative indicators of both pressure bearing capacity and sealing performance are better than the qualified threshold, they are marked as high reliability level; For a housing with only one quantitative indicator that is better than the qualified threshold and another that is close to the qualified threshold, it is marked as a medium reliability level; For any quantitative indicator below the qualified threshold, or for two indicators close to but not reaching the qualified threshold, the shell is marked as low reliability level; For housings with serious defects and whose quantitative assessment results indicate an immediate risk of failure, they are marked as failure levels; Generate a classification list that includes a unique identifier for each housing, specific values ​​for various quantitative indicators, and the final reliability level.

8. The oil filter manufacturing and testing method based on particle swarm optimization algorithm according to claim 1, characterized in that, Based on the reliability level and the cycle time information of the production line scheduling system, the process for reworking or scrapping defective shells is planned, specifically including: The current production cycle time, the length of the subsequent workstation queue, and the idle status of the rework workstation can be obtained in real time from the production line scheduling system. Read the classification list and filter out oil filter housings marked with low reliability and failure levels; For housings with low reliability levels, appropriate local repair or reinforcement processes are matched from the available rework process library based on the specific location and type of defects. Based on the time required for the matching rework process, and under the premise of meeting the overall production cycle time, the optimal offline rework insertion time and rework path are calculated for each shell to be reworked. For shells with fault levels, a scrapping command is generated directly, and the path and time for removing them from the main line and transporting them to the waste recycling area are planned. All planned rework and scrap instructions, routes, and schedules are synchronized to the production line scheduling system and material handling system.

9. The oil filter manufacturing and testing method based on particle swarm optimization algorithm according to claim 2, characterized in that, The analyzed and matched signal pairs identify abnormal coupling segments where the pressure peak and current trough do not correspond normally, or where the pressure plateau is accompanied by violent current fluctuations. Specifically, these include: Calculate the local statistical characteristics of each pair of dynamic pressure curve segments and instantaneous current waveform segments after matching, including the mean, variance, and kurtosis of the pressure segment, and the mean, variance, and zero-crossing rate of the current segment. Set a normal time correspondence window for pressure peak and current valley, and check whether the time difference between the pressure peak point and the current valley point in the signal alignment exceeds the normal correspondence window. Set a variance threshold for the pressure stability segment, and check whether the variance of the corresponding current waveform exceeds its normal fluctuation range within the stability segment where the pressure variance is lower than the variance threshold. Signal pairs whose time difference exceeds the window or whose current fluctuations are abnormal are marked as abnormal candidate segments; The abnormal candidate segments are manually verified using process knowledge to exclude reasonable fluctuations caused by normal process adjustments, and finally the true abnormal coupling segments are confirmed.

10. The oil filter manufacturing and testing method based on particle swarm optimization algorithm according to claim 4, characterized in that, The step of dynamically adjusting the magnification to the most suitable high magnification for identifying the micro-defects corresponding to the predicted defect type based on the texture and contrast features in the region overview image specifically includes: The acquired regional overview image is divided into blocks, the gray-level co-occurrence matrix of each image sub-block is calculated, and contrast and homogeneity features are extracted. Based on the predicted defect type, determine the typical texture representation pattern of the predicted defect type under a microscope, such as the linear low-contrast features of cracks and the circular high-contrast edge features of pores. The similarity between the texture features of image sub-blocks and the typical texture representation patterns of defects is calculated. Select the region containing the image sub-block with the highest similarity as the key observation area; Based on the physical size of the key observation area and the desired level of detail, the required total magnification is calculated by combining the microscope's objective lens magnification and digital magnification capability. Control the microscope vision probe to switch to the objective lens of the corresponding magnification, and adjust the illumination angle to enhance the contrast of the key observation area.

Images