AI visual defect analysis device for intelligent manufacturing

Abstract

The invention relates to an AI visual defect analysis device for intelligent manufacturing, and belongs to the technical field of intelligent manufacturing. The device comprises a cross-scene data adaptive calibration unit which acquires data and corrects acquisition parameters according to working condition difference to generate a calibration data set; the production link association traceability unit constructs a node mapping system, presets an implicit association rule, tracks a defect path, and positions a key production node; the dynamic game threshold determination unit constructs a model and dynamically adjusts a defect determination boundary, and determines a defect type, a defect grade and a root cause node through double verification; the process deviation optimization unit presets an evaluation index, prejudges a scheme effect and generates a process adjustment instruction in combination with field feedback. According to the method, integrated defect analysis of cross-scene adaptation, full-link traceability, accurate judgment and closed-loop optimization is realized, the problems of poor adaptability, insufficient traceability and the like in the prior art are solved, and the intelligent manufacturing quality control level is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing technology, and specifically relates to an AI visual defect analysis device for intelligent manufacturing. Background Technology

[0002] As the manufacturing industry transforms and upgrades towards intelligent manufacturing, the level of automation and intelligence in production processes is constantly improving. Product quality control, as a core link in intelligent manufacturing, directly determines product competitiveness and production efficiency through its detection accuracy and efficiency. In various product production processes, defect detection is a key means of quality control. Traditional defect detection relies heavily on manual visual inspection, which suffers from low detection efficiency, strong subjectivity, and high rates of false positives and false negatives, and can no longer meet the needs of large-scale, high-precision, and fast-paced intelligent manufacturing production.

[0003] To address the drawbacks of manual inspection, AI visual inspection technology, with its advantages of automation and high efficiency, is gradually penetrating the field of defect detection in smart manufacturing. Its core principle involves capturing product images during the production process using image acquisition equipment, and then relying on pre-set algorithm models to extract and match features from the images, achieving automated identification and preliminary judgment of defects. However, existing AI visual defect analysis technologies still face numerous insurmountable technical bottlenecks in the complex and ever-changing real-world production scenarios of smart manufacturing, severely restricting their application effectiveness and scope of promotion. On the one hand, in intelligent manufacturing scenarios, multiple production lines and processes coexist, and the working conditions of different scenarios vary significantly. This includes not only differences in production rate, equipment vibration amplitude, and processing materials, but also changes in the processing environment caused by process switching (such as cutting fluid residue, dust interference, etc.). These differences lead to obvious heterogeneity in the visual defect data collected in different scenarios. Existing technologies mostly use fixed acquisition parameters and data processing modes, lacking targeted cross-scenario data adaptation and dynamic calibration mechanisms. This makes the collected data prone to noise interference, feature blurring, and other problems, resulting in extremely poor data stability and consistency. This directly leads to inaccurate defect feature extraction, which in turn significantly reduces the accuracy and reliability of defect identification. On the other hand, the core of existing technologies focuses on visual feature recognition of defects themselves, failing to build a deep connection between defects and the entire production chain (from raw material input, various processing steps to semi-finished product flow). It can only determine "whether a defect exists", but cannot associate defects with key information such as process parameters, equipment operating status, and raw material batches at specific production nodes. This makes it impossible to trace the key production nodes and root causes of defects, making it difficult to form targeted process optimization solutions. As a result, defect detection and production improvement are disconnected, and the defect incidence rate cannot be reduced from the source.

[0004] Meanwhile, existing defect judgments mostly use fixed threshold standards, which cannot be dynamically adjusted according to the process benchmarks and equipment operating parameters of the current production scenario, resulting in large deviations in defect level judgments. Furthermore, in the process optimization stage, there is a lack of a closed-loop mechanism of "detection-judgment-optimization-verification", making it difficult to combine actual feedback from the production site to correct the optimization direction in real time and failing to achieve deep linkage between defect detection and process improvement.

[0005] In summary, current AI visual defect analysis technologies in the field of intelligent manufacturing suffer from problems such as poor cross-scenario adaptability, insufficient traceability of production links, rigid threshold judgment, and incomplete process optimization closed loop, making it difficult to meet the needs of intelligent manufacturing for high-quality, high-precision defect detection and full-process quality control. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention provides an AI-based visual defect analysis device for intelligent manufacturing. The objective of this invention can be achieved through the following technical solutions: include: The cross-scenario data adaptive calibration unit acquires defect visual data and process-related data from multiple production scenarios; through a dynamic scenario adaptation calibration algorithm, it corrects the acquired parameters in real time according to differences in working conditions, and generates a cross-scenario calibration dataset. The production link association traceability unit, based on the cross-scenario calibration dataset, constructs a node mapping system between defects and the entire production link; through the production link node association algorithm, it presets implicit association rules; tracks the generation and transmission path of defects in the production process, locates the key production nodes that cause defects, and outputs a link association traceability information table. The dynamic game threshold determination unit, in conjunction with the link-related tracing information table, constructs a dynamic game threshold model; it matches the process benchmark range and equipment operating parameters of the current production scenario in real time, and dynamically adjusts the defect severity determination boundary through the threshold game algorithm; through defect pattern matching verification and link logic verification, it determines the defect type, level and root cause node, and generates a defect tracing determination report. The process deviation optimization unit, based on the defect tracing and judgment report, presets linkage evaluation indicators; analyzes the process data of the corresponding production node, predicts the defect improvement effect of different process adjustment schemes; and, combined with the actual feedback from the production site, corrects the optimization direction and generates process adjustment instructions.

[0007] Specifically, the process of real-time correction of the collected parameters based on differences in operating conditions is as follows: Classify the working conditions and preset the baseline values ​​of the collected parameters for each working condition; Real-time acquisition of current scene operating condition signals, and calculation of the deviation between the actual signal and the reference value; The parameters are adjusted in stages according to the correspondence between the deviation and the preset value. After correction, multiple frames of data are continuously collected. If the data grayscale value fluctuation is within the preset range, the parameters are fixed.

[0008] Specifically, the process of constructing the node mapping system between defects and the entire production chain is as follows: obtain all links in the entire production process from raw material input to finished product output, and determine the key nodes of each link; record the sequential connection order and flow time requirements of each key node; associate the production equipment model, equipment number and core operating parameter items corresponding to each key node; and mark the types of defects that have occurred in the past and the corresponding process influencing factors for each key node.

[0009] Specifically, the process of the preset implicit association rule is as follows: collect defect cases and production data of each production node, extract defect type, production parameters, and equipment status information; count the frequency of occurrence of parameter combinations and defect types to determine basic association pairs; formulate judgment rules for each association pair, and preset the parameter range and equipment status judgment conditions.

[0010] Specifically, the process of tracing the generation and propagation path of defects in the production process is as follows: Extract visual feature information of defects and match initial candidate nodes in the node mapping system; Retrieve the production time-series data of the initial candidate node during the period when the defect occurred; check the subsequent connecting nodes of the initial candidate node in sequence according to the production process, compare the production data of each node with the implicit association rules; record the occurrence time and characteristic changes of the defect at each node to form a complete record of the generation and transmission path.

[0011] Specifically, the process of locating the key production node that causes the defect is as follows: based on the established defect generation and transmission path records, extract the defect feature matching degree and production data conformity of each node; set the judgment criteria for matching degree and conformity; select the node with the highest matching degree and conformity that both meet the criteria requirements; determine the node as the key production node that causes the defect and mark the relevant data basis.

[0012] Specifically, the process of constructing the dynamic game threshold model is as follows: The input data consists of key node parameters, process reference ranges, and equipment operating parameters in the link association traceability information table; the output data is the defect severity judgment boundary. It includes a data input layer, a feature selection layer, a game calculation layer, and a threshold output layer; The data input layer receives input data and performs standardization processing, converting parameters of different dimensions into a unified value range; the feature filtering layer removes redundant data that is irrelevant to defect judgment through correlation analysis; the game calculation layer sets two game subjects, namely the industry standard threshold subject and the on-site production threshold subject; the threshold output layer performs fusion calculation on the thresholds of the two subjects and outputs the final dynamic game threshold model.

[0013] Specifically, the process of dynamically adjusting the boundary for determining the severity of defects using a threshold game algorithm is as follows: Obtain the current scenario's process baseline range and equipment operating parameters, and use the standard range as the initial parameter for the standard threshold; extract the historical defect data of the current batch, and use the statistical parameter deviation range as the initial parameter for the on-site production threshold; Calculate the overlap interval between the initial parameters of the standard threshold and the initial parameters of the on-site production threshold, and set initial candidate values; The parameter weights are iteratively adjusted, the interval and candidate values ​​are recalculated, and the final boundary is determined when the fluctuation of the candidate value is less than the preset range.

[0014] Specifically, the process for determining the defect type, level, and root cause is as follows: Preset various defect patterns, compare defect visual data and features, and match defect types; Based on the boundary threshold of the dynamic game threshold model, the defect parameters are compared and classified with the boundary values. By combining the aforementioned link-related traceability information table, key production nodes can be located, and node process parameters and equipment records can be retrieved. By comparing the implicit association rules, the parameter deviations or equipment anomalies caused by the defects are determined.

[0015] Specifically, the process of setting the preset linkage evaluation indicators is as follows: the preset linkage evaluation indicators include defect incidence rate, process adjustment response time, and parameter stability; specific setting rules are formulated for each indicator, the defect incidence rate indicator is set as the upper limit of the proportion of defective products in a unit production batch to the total number of products; the process adjustment response time indicator is set as the upper limit of the time from the generation of adjustment requirements to the completion of process adjustment; the parameter stability indicator is set as the upper limit of the fluctuation range of the adjusted process parameters within a preset time; weight proportions are assigned to the indicators, and the indicator names, setting standards and weights are organized.

[0016] Specifically, the process for predicting the defect improvement effect of different process adjustment schemes is as follows: Extract the process deviation parameters of key production nodes from the defect tracing and determination report, and calculate the difference between the current value and the standard value of the deviation parameters; Design different process adjustment schemes, each scheme including specific parameter adjustment ranges, adjustment steps and implementation sequence; Retrieve historical process adjustment data from production nodes, establish the correspondence between process parameter adjustment amounts and defect improvement amounts, and calculate the expected defect improvement amount for each scheme.

[0017] Specifically, the process of combining actual feedback from the production site to correct and optimize the direction and generate process adjustment instructions is as follows: collect actual production data after preliminary adjustments on site; compare the actual data with the predicted effect to calculate the deviation value; when the deviation value exceeds the preset range, analyze the reasons and adjust the parameters or sequence of the scheme; repeat the comparison and adjustment steps until the deviation value meets the preset range, and convert the final adjustment scheme into process adjustment instructions.

[0018] The beneficial effects of this invention are as follows: (1) By setting up a cross-scenario data adaptive calibration unit and a production link-related tracing unit, this invention effectively solves the core problems of poor cross-scenario adaptability and insufficient defect tracing capability of existing technologies, laying a solid foundation for defect analysis. Among them, the cross-scenario data adaptive calibration unit can correct the acquisition parameters in real time according to the differences in working conditions. By dividing the working condition type, setting the benchmark value, calculating the deviation, adjusting the parameters in stages, and verifying the data stability, it can eliminate the heterogeneity of defect visual data caused by the differences in working conditions of different production scenarios, improve the stability and consistency of cross-scenario data acquisition, and ensure the data quality of subsequent analysis. The production link-related tracing unit constructs a mapping system covering key nodes of the entire production process, clarifies the node connection sequence, equipment information, and historical defect association factors, and tracks the defect generation and transmission path and locates key production nodes by combining preset implicit association rules. This breaks the limitation of existing technologies that only focus on defect visual recognition and lack production link association, and realizes the breakthrough from defect "recognition" to "tracing". It can lock the fundamental link of defect generation and provide a clear target for subsequent process optimization. (2) By setting up a dynamic game threshold judgment unit and a process deviation optimization unit, this invention realizes defect judgment and process closed-loop optimization, which greatly improves the effect and efficiency of production quality control. The dynamic game threshold judgment unit constructs a dynamic game threshold model with a multi-layer structure, integrates industry standard and on-site production dual subject thresholds for iterative optimization, dynamically adjusts the defect severity judgment boundary, and realizes the judgment of defect type, level and root cause node with defect pattern matching and link logic dual verification. It solves the problem of rigid fixed threshold judgment and large deviation in the existing technology, and avoids the impact of misjudgment and omission on production quality control. The process deviation optimization unit uses preset multi-dimensional linkage evaluation indicators, combines historical data to predict the effect of process adjustment schemes, and iteratively corrects the optimization direction and generates process adjustment instructions based on actual feedback from the production site, forming a complete closed-loop system of "detection-judgment-optimization-verification". It realizes the deep linkage between defect detection and process improvement, can solve the defect generation problem from the source, effectively reduce the defect incidence rate, improve the stability of production process and product quality, and shorten the process adjustment response time and improve production efficiency. Attached Figure Description

[0019] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0020] Figure 1 This is a system architecture diagram of an AI visual defect analysis device for intelligent manufacturing according to the present invention; Figure 2 This is a data flow diagram of an AI visual defect analysis device for intelligent manufacturing according to the present invention. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0022] Please see Figure 1-2 An AI-powered visual defect analysis device for intelligent manufacturing; include: The cross-scenario data adaptive calibration unit acquires defect visual data and process-related data from multiple production scenarios; through a dynamic scenario adaptation calibration algorithm, it corrects the acquired parameters in real time according to differences in working conditions, and generates a cross-scenario calibration dataset. The production link association traceability unit, based on the cross-scenario calibration dataset, constructs a node mapping system between defects and the entire production link; through the production link node association algorithm, it presets implicit association rules; tracks the generation and transmission path of defects in the production process, locates the key production nodes that cause defects, and outputs a link association traceability information table. The dynamic game threshold determination unit, in conjunction with the link-related tracing information table, constructs a dynamic game threshold model; it matches the process benchmark range and equipment operating parameters of the current production scenario in real time, and dynamically adjusts the defect severity determination boundary through the threshold game algorithm; through defect pattern matching verification and link logic verification, it determines the defect type, level and root cause node, and generates a defect tracing determination report. The process deviation optimization unit, based on the defect tracing and judgment report, presets linkage evaluation indicators; analyzes the process data of the corresponding production node, predicts the defect improvement effect of different process adjustment schemes; and, combined with the actual feedback from the production site, corrects the optimization direction and generates process adjustment instructions.

[0023] In this embodiment, the defect visual data refers to images and derived data containing product defects acquired through image acquisition devices such as industrial cameras and vision sensors. Specifically, it covers the image pixel matrix, contour feature data, grayscale value distribution data, texture feature data, etc. of the defect area, which can directly reflect the appearance and location information of the defect. The process-related data refers to time-series data that is directly related to the product manufacturing process and can support defect tracing and process optimization. Specifically, it includes production time, process number, processing parameters (such as cutting speed, feed rate, temperature, pressure, etc.), raw material batch information, equipment operation log fragments, etc., which are used to establish the correlation between defects and the production process.

[0024] In this embodiment, the difference in working conditions specifically refers to the differences in environment and operating status that affect the quality of defect visual data acquisition and the stability of the production process under different production scenarios or different time periods of the same production scenario. It is the core adaptation object of the cross-scenario data adaptive calibration unit, and specifically includes: differences in production operating status: such as differences in production rate (uniform speed / intermittent / high-speed production), differences in equipment operating load (full load / half load operation), fluctuations in operating status caused by process switching, etc.; differences in production environment: such as differences in dust concentration in the processing area, differences in residual cutting fluid, differences in vibration amplitude, etc.; differences in processing objects: such as differences in product processing materials (metal / plastic / ceramic, etc.), differences in product specifications and models, differences in processing process stages (raw material processing / semi-finished product assembly / finished product inspection, etc.).

[0025] In this embodiment, the process reference range refers to the reasonable range of values ​​for key process parameters of each production process, preset based on industry quality standards, enterprise production specifications, and historical qualified production data. It is the standard for judging whether the process parameters are deviated. Specifically, it includes the reference range of cutting temperature and feed rate for a certain process. Each key production node corresponds to a set of exclusive process reference ranges. The equipment operating parameters refer to the status data generated in real time during the operation of the production equipment. It is the core basis for judging whether the equipment is abnormal. Specifically, it includes the spindle speed, motor current, hydraulic pressure, equipment vibration frequency, bearing temperature, etc., which correspond to the process reference ranges and are synchronously associated with each key production node.

[0026] In this embodiment, the key nodes of each stage include: raw material inspection node, first processing node, intermediate inspection node, final processing node, and finished product inspection node.

[0027] In this embodiment, determining the defect type refers to identifying and determining the specific type of defect, and matching it based on a preset defect pattern library. Specific examples include common product defect categories in the production process such as surface scratches, cracks, deformation, missing materials, impurity embedding, and dimensional deviations. Determining the defect level refers to determining the severity level of the defect based on the defect severity determination boundary. Specifically, this is done by comparing the actual characteristic parameters of the defect (such as scratch depth, crack length, deformation, etc.) with the boundary values ​​output by the dynamic game threshold model to classify the defect severity level (such as slight / moderate / severe, etc., with the level classification standard matching the enterprise's quality control needs). Determining the root cause node refers to identifying the key production node and core cause that led to the defect. This includes two parts: first, determining the specific production node where the defect occurred (such as the raw material inspection node, the first processing node, the intermediate inspection node, etc.); second, clarifying the core cause of the defect at that node, such as process parameter deviation (excessive cutting temperature), abnormal equipment status (excessive spindle vibration, etc.). The determination is based on the comparison results between the link association traceability information table and the implicit association rules.

[0028] Specifically, the process of real-time correction of the collected parameters based on differences in operating conditions is as follows: Classify the working conditions and preset the baseline values ​​of the collected parameters for each working condition; Real-time acquisition of current scene operating condition signals, and calculation of the deviation between the actual signal and the reference value; The parameters are adjusted in stages according to the correspondence between the deviation and the preset value. After correction, multiple frames of data are continuously collected. If the data grayscale value fluctuation is within the preset range, the parameters are fixed.

[0029] Specifically, the process of constructing the node mapping system between defects and the entire production chain is as follows: obtain all links in the entire production process from raw material input to finished product output, and determine the key nodes of each link; record the sequential connection order and flow time requirements of each key node; associate the production equipment model, equipment number and core operating parameter items corresponding to each key node; and mark the types of defects that have occurred in the past and the corresponding process influencing factors for each key node.

[0030] Specifically, the process of the preset implicit association rule is as follows: collect defect cases and production data of each production node, extract defect type, production parameters, and equipment status information; count the frequency of occurrence of parameter combinations and defect types to determine basic association pairs; formulate judgment rules for each association pair, and preset the parameter range and equipment status judgment conditions.

[0031] Specifically, the process of tracing the generation and propagation path of defects in the production process is as follows: Extract visual feature information of defects and match initial candidate nodes in the node mapping system; Retrieve the production time-series data of the initial candidate node during the period when the defect occurred; check the subsequent connecting nodes of the initial candidate node in sequence according to the production process, compare the production data of each node with the implicit association rules; record the occurrence time and characteristic changes of the defect at each node to form a complete record of the generation and transmission path.

[0032] Specifically, the process of locating the key production node that causes the defect is as follows: based on the established defect generation and transmission path records, extract the defect feature matching degree and production data conformity of each node; set the judgment criteria for matching degree and conformity; select the node with the highest matching degree and conformity that both meet the criteria requirements; determine the node as the key production node that causes the defect and mark the relevant data basis.

[0033] Specifically, the process of constructing the dynamic game threshold model is as follows: The input data consists of key node parameters, process reference ranges, and equipment operating parameters in the link association traceability information table; the output data is the defect severity judgment boundary. It includes a data input layer, a feature selection layer, a game calculation layer, and a threshold output layer; The data input layer receives input data and performs standardization processing, converting parameters of different dimensions into a unified value range; the feature filtering layer removes redundant data that is irrelevant to defect judgment through correlation analysis; the game calculation layer sets two game subjects, namely the industry standard threshold subject and the on-site production threshold subject; the threshold output layer performs fusion calculation on the thresholds of the two subjects and outputs the final dynamic game threshold model.

[0034] Specifically, the process of dynamically adjusting the boundary for determining the severity of defects using a threshold game algorithm is as follows: Obtain the current scenario's process baseline range and equipment operating parameters, and use the standard range as the initial parameter for the standard threshold; extract the historical defect data of the current batch, and use the statistical parameter deviation range as the initial parameter for the on-site production threshold; Calculate the overlap interval between the initial parameters of the standard threshold and the initial parameters of the on-site production threshold, and set initial candidate values; The parameter weights are iteratively adjusted, the interval and candidate values ​​are recalculated, and the final boundary is determined when the fluctuation of the candidate value is less than the preset range.

[0035] Specifically, the process for determining the defect type, level, and root cause is as follows: Preset various defect patterns, compare defect visual data and features, and match defect types; Based on the boundary threshold of the dynamic game threshold model, the defect parameters are compared and classified with the boundary values. By combining the aforementioned link-related traceability information table, key production nodes can be located, and node process parameters and equipment records can be retrieved. By comparing the implicit association rules, the parameter deviations or equipment anomalies caused by the defects are determined.

[0036] Specifically, the process of setting the preset linkage evaluation indicators is as follows: the preset linkage evaluation indicators include defect incidence rate, process adjustment response time, and parameter stability; specific setting rules are formulated for each indicator, the defect incidence rate indicator is set as the upper limit of the proportion of defective products in a unit production batch to the total number of products; the process adjustment response time indicator is set as the upper limit of the time from the generation of adjustment requirements to the completion of process adjustment; the parameter stability indicator is set as the upper limit of the fluctuation range of the adjusted process parameters within a preset time; weight proportions are assigned to the indicators, and the indicator names, setting standards and weights are organized.

[0037] Specifically, the process for predicting the defect improvement effect of different process adjustment schemes is as follows: Extract the process deviation parameters of key production nodes from the defect tracing and determination report, and calculate the difference between the current value and the standard value of the deviation parameters; Design different process adjustment schemes, each scheme including specific parameter adjustment ranges, adjustment steps and implementation sequence; Retrieve historical process adjustment data from production nodes, establish the correspondence between process parameter adjustment amounts and defect improvement amounts, and calculate the expected defect improvement amount for each scheme.

[0038] Specifically, the process of combining actual feedback from the production site to correct and optimize the direction and generate process adjustment instructions is as follows: collect actual production data after preliminary adjustments on site; compare the actual data with the predicted effect to calculate the deviation value; when the deviation value exceeds the preset range, analyze the reasons and adjust the parameters or sequence of the scheme; repeat the comparison and adjustment steps until the deviation value meets the preset range, and convert the final adjustment scheme into process adjustment instructions.

[0039] In this embodiment, the specific operations for constructing the defect-to-production-chain node mapping system are as follows: First, industrial process modeling tools (such as Visio and ProcessOn) are used to sort out the general industrial production process, clarifying the complete chain from raw material warehousing inspection, raw material processing, intermediate assembly, semi-finished product inspection, finished product processing to finished product outbound inspection, and identifying key nodes in each link, specifically including raw material inspection nodes, processing nodes, assembly nodes, semi-finished product inspection nodes, finished product processing nodes, and finished product inspection nodes. Second, the production equipment control systems (such as PLC and SCADA systems) of each key node are accessed through Industrial Internet of Things (IIoT) data acquisition terminals to collect and extract the connection sequence, flow time threshold, corresponding equipment information (including equipment model, equipment number, and core operating parameters), and historical defect information (including defect type and corresponding influencing factors) of each node. Subsequently, a node information association table was constructed using a MySQL relational database. A unique code was assigned to each key node (the coding rule is "workshop number-process number-node sequence number"). Through the coding, data such as node connection relationship, equipment parameters, and defect information were associated and bound. Finally, a visual node mapping system was generated with the help of data visualization tools (such as ECharts) to realize the intuitive presentation and rapid retrieval of node information in the entire production chain.

[0040] In this embodiment, the implicit association rules are preset through the production link node association algorithm as follows: The Apriori association rule mining algorithm is used as the core algorithm for production link node association, relying on an industrial data acquisition platform (such as Huawei Cloud IoT). Edge collected historical data from the entire production chain over the past 12 months, including production parameters at each node (such as processing speed, processing pressure, processing temperature, and assembly accuracy), equipment operating status data (motor current, equipment vibration frequency, and bearing temperature), and defect record data (defect type, defect occurrence time, and corresponding node). The collected data was imported into a Python data processing environment, and data cleaning was performed using the Pandas library (removing missing and outlier values). Then, the Apriori algorithm was called using the mlxtend library, setting minimum support and minimum confidence to mine the correlation between combinations of production parameters and defect types. For example, the correlation "processing pressure > preset threshold and equipment vibration frequency > preset threshold → surface depression defect" was mined. Such high-frequency, high-confidence correlations were identified as the core content of implicit association rules, and the applicable nodes, parameter judgment range, and corresponding defect types of each rule were clarified. Finally, an implicit association rule library was constructed using XML format, storing the rule content in a structure of "node code - parameter condition - defect type".

[0041] In this embodiment, the specific operation of constructing the dynamic game threshold model is as follows: The model input data consists of six core parameters from the link association traceability information table, including processing node pressure, vibration frequency, and finished product processing temperature. The output data is the severity judgment boundary of surface and structural defects (dents, cracks, scratches, etc.) of industrial products. The specific steps are: First, the data input layer standardizes the input data using the Pandas library and uses the min-max normalization method to convert parameters of different dimensions to the [0,1] interval to eliminate dimension differences. Second, the feature selection layer calculates the correlation coefficient between each input parameter and the severity of the defect through Pearson correlation coefficient analysis, removes redundant parameters (such as environmental humidity), and retains four core parameters, including processing pressure and vibration frequency. The three steps are as follows: First, the game calculation layer sets up two game subjects. The industry standard threshold subject determines the initial threshold based on the relevant industry industrial product quality standards (e.g., the initial threshold for surface indentation depth is 0.2mm). The on-site production threshold subject determines the initial threshold by statistically analyzing the parameter fluctuation range of qualified industrial products over the past three months (e.g., the initial on-site threshold for surface indentation depth is 0.25mm). The Nash equilibrium algorithm is used to calculate the game equilibrium solution of the two subjects to obtain the initial judgment boundary. Second, the threshold output layer smooths the initial boundary using the Sigmoid function, and then uses nearly 300 sets of validation samples to iteratively train the model until the model prediction error is <5%. Finally, a stable dynamic game threshold model is output. The model file is saved in .h5 format for subsequent defect level determination.

[0042] In this embodiment, the specific operations of defect pattern matching verification and link logic verification are as follows: First, defect pattern matching verification is implemented using the SIFT (Scale Invariant Feature Transform) algorithm. Specifically, an industrial camera is used to acquire images of the surface of industrial products. The images are preprocessed (grayscale conversion, Gaussian filtering, edge enhancement) to extract SIFT feature points (128-dimensional feature points) of the defect area. The extracted feature points are matched with a preset defect pattern library (containing standard SIFT feature templates of common industrial defects such as dents, cracks, and scratches). The FLANN matcher is used to calculate the feature point similarity, and a similarity threshold of 0.8 is set. When the matching similarity is ≥0.8, the defect type is determined to be consistent with the corresponding template, and the matching verification is completed. Second, link logic verification uses causal graph analysis combined with time series data. The traceability implementation is as follows: Retrieve the time-series data of key nodes corresponding to the defect type from the link-related traceability information table (e.g., after determining that the dent defect is caused by a processing node, retrieve the processing pressure change curve and equipment vibration record during the defect occurrence period). Use SQL statements to query whether the data for this period meets the judgment condition of "processing pressure > preset threshold and vibration frequency > preset threshold" in the implicit association rule base. Simultaneously, use a cause-effect graph tool to draw the causal link of "parameter deviation - equipment abnormality - defect occurrence," verify the logical sequence of the parameter deviation occurrence time and the defect occurrence time in the time-series data, and after confirming that there is no logical contradiction, complete the link logic verification. Finally, combine the two verification results to output the defect type, level, and root cause node (e.g., "surface dent - moderate - processing node (excessive processing pressure)").

[0043] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An AI visual defect analysis device for intelligent manufacturing, characterized in that, include: Cross-scenario data adaptive calibration unit acquires defect visual data and process correlation data from multiple production scenarios; By using a dynamic scene adaptation calibration algorithm, the collected parameters are corrected in real time according to the differences in working conditions, and a cross-scene calibration dataset is generated. The production chain-related traceability unit constructs a node mapping system between defects and the entire production chain based on the cross-scenario calibration dataset. Implicit association rules are preset through the production link node association algorithm; Track the generation and transmission path of defects in the production process, locate the key production nodes that cause defects, and output a chain-related traceability information table; The dynamic game threshold determination unit, in conjunction with the link association tracing information table, constructs a dynamic game threshold model; Real-time matching of the process baseline range and equipment operating parameters in the current production scenario, and dynamic adjustment of the defect severity determination boundary through threshold game algorithm; By verifying defect patterns and link logic, the defect type, level, and root cause node are determined, and a defect tracing and determination report is generated. The process deviation optimization unit, based on the defect source tracing and judgment report, presets linkage evaluation indicators; analyzes the process data of the corresponding production node, and predicts the defect improvement effect of different process adjustment schemes; Based on actual feedback from the production site, the optimization direction is revised and optimized, and process adjustment instructions are generated.

2. The apparatus according to claim 1, characterized in that, The specific process of real-time correction of the collected parameters based on differences in operating conditions is as follows: Classify the working conditions and preset the baseline values ​​of the collected parameters for each working condition; Real-time acquisition of current scene operating condition signals, and calculation of the deviation between the actual signal and the reference value; The parameters are adjusted in stages according to the correspondence between the deviation and the preset value. After correction, multiple frames of data are continuously collected. If the data grayscale value fluctuation is within the preset range, the parameters are fixed.

3. The apparatus according to claim 1, characterized in that, The specific process of constructing the defect and production chain node mapping system is as follows: obtain all links in the entire production process from raw material input to finished product output, and determine the key nodes of each link; record the sequential connection order and flow time requirements of each key node; associate the production equipment model, equipment number and core operating parameter items corresponding to each key node; and mark the types of defects that have occurred in the past and the corresponding process influencing factors for each key node.

4. The apparatus according to claim 1, characterized in that, The specific process of the preset implicit association rule is as follows: collect defect cases and production data of each production node, extract defect type, production parameters, and equipment status information; count the frequency of occurrence of parameter combinations and defect types to determine basic association pairs; formulate judgment rules for each association pair, and preset parameter range and equipment status judgment conditions.

5. The apparatus according to claim 1, characterized in that, The specific process of tracing the generation and propagation path of defects in the production process is as follows: Extract visual feature information of defects and match initial candidate nodes in the node mapping system; Retrieve the production time series data of the period when the defect occurred in the initial candidate node; check the subsequent connecting nodes of the initial candidate node in the order of the production process, and compare the production data of each node with the implicit association rules; Record the timing and characteristic changes of defects at each node to form a complete record of the generation and transmission path.

6. The apparatus according to claim 1, characterized in that, The specific process for locating the key production node that caused the defect is as follows: Based on the established defect generation and transmission path records, extract the defect feature matching degree and production data conformity of each node; set the judgment criteria for matching degree and conformity; and select the node with the highest matching degree and conformity that both meet the criteria requirements. This node was identified as a critical production node causing the defect, and relevant data was labeled accordingly.

7. The apparatus according to claim 1, characterized in that, The specific process for constructing the dynamic game threshold model is as follows: The input data consists of key node parameters, process reference ranges, and equipment operating parameters in the link association traceability information table; the output data is the defect severity judgment boundary. It includes a data input layer, a feature selection layer, a game calculation layer, and a threshold output layer; The data input layer receives input data and performs standardization processing, converting parameters of different dimensions into a unified value range; the feature filtering layer removes redundant data that is irrelevant to defect judgment through correlation analysis; the game calculation layer sets two game subjects, namely the industry standard threshold subject and the on-site production threshold subject; the threshold output layer performs fusion calculation on the thresholds of the two subjects and outputs the final dynamic game threshold model.

8. The apparatus according to claim 1, characterized in that, The specific process of dynamically adjusting the boundary for determining the severity of defects using the threshold game algorithm is as follows: Obtain the current scenario's process baseline range and equipment operating parameters, and use the standard range as the initial parameter for the standard threshold; extract the historical defect data of the current batch, and use the statistical parameter deviation range as the initial parameter for the on-site production threshold; Calculate the overlap interval between the initial parameters of the standard threshold and the initial parameters of the on-site production threshold, and set initial candidate values; The parameter weights are iteratively adjusted, the interval and candidate values ​​are recalculated, and the final boundary is determined when the fluctuation of the candidate value is less than the preset range.

9. The apparatus according to claim 1, characterized in that, The specific process for determining the defect type, level, and root cause is as follows: Preset various defect patterns, compare defect visual data and features, and match defect types; Based on the boundary threshold of the dynamic game threshold model, the defect parameters are compared and classified with the boundary values. By combining the aforementioned link-related traceability information table, key production nodes can be located, and node process parameters and equipment records can be retrieved. By comparing the implicit association rules, the parameter deviations or equipment anomalies caused by the defects are determined.

10. The apparatus according to claim 1, characterized in that, The specific process of the preset linkage evaluation index is as follows: The preset linkage evaluation index includes defect occurrence rate, process adjustment response time, and parameter stability; specific setting rules are formulated for each index. The defect occurrence rate index is set as the upper limit of the proportion of defective products in a unit production batch to the total number of products; the process adjustment response time index is set as the upper limit of the time from generating the adjustment requirement to completing the process adjustment; the parameter stability index is set as the upper limit of the fluctuation range of the adjusted process parameters within the preset time. Assign weight percentages to indicators, organize indicator names, and set standards and weights.

11. The apparatus according to claim 1, characterized in that, The specific process for predicting the defect improvement effect of different process adjustment schemes is as follows: Extract the process deviation parameters of key production nodes from the defect tracing and determination report, and calculate the difference between the current value and the standard value of the deviation parameters; Design different process adjustment schemes, each scheme including specific parameter adjustment ranges, adjustment steps and implementation sequence; Retrieve historical process adjustment data from production nodes, establish the correspondence between process parameter adjustment amounts and defect improvement amounts, and calculate the expected defect improvement amount for each scheme.

12. The apparatus according to claim 1, characterized in that, The specific process of combining actual feedback from the production site to correct and optimize the direction and generate process adjustment instructions is as follows: collect actual production data after preliminary adjustments on site; compare the actual data with the predicted effect to calculate the deviation value; when the deviation value exceeds the preset range, analyze the reasons and adjust the parameters or sequence of the scheme; repeat the comparison and adjustment steps until the deviation value meets the preset range, and convert the final adjustment scheme into process adjustment instructions.

Images