An elevator car defect diagnosis analysis method and system based on industrial vision

Abstract

The present application relates to the technical field of elevator car defect diagnosis analysis, and particularly relates to an elevator car defect diagnosis analysis method and system based on industrial vision, which comprises the following steps: performing local feature analysis on a low-definition image stream of a weld to identify potential abnormal areas; performing motion correction on a high-definition local image; extracting a plurality of visual features from the high-definition local image after enhancement processing; outputting a defect judgment confidence, and adaptively adjusting the standard of defect judgment according to material information and process parameter information in the production process; pushing the identified sample to an artificial review queue to obtain labeling, incrementally updating the standard of defect judgment after adaptive adjustment according to the artificially labeled sample, and identifying a new normal process texture mode, and writing the new normal process texture mode into a normal sample library. The above can reduce false positives and missed detection rates, and improve the adaptability of the system to changes in the production environment.

Description

Technical Field

[0001] This invention relates to the technical field of elevator car defect diagnosis and analysis, specifically to an elevator car defect diagnosis and analysis method and system based on industrial vision. Background Technology

[0002] On automated production lines for elevator cars, to ensure product quality and operational safety, a defect diagnosis system based on industrial vision is typically deployed to automate the inspection of welded seams in the elevator car. However, in actual production, such systems often face challenges in identifying micron-level defects. For example, the weld surface contains many normal process traces that visually resemble tiny defects, easily leading to false alarms from the system.

[0003] The purpose of this invention is to address the aforementioned shortcomings by proposing a method and system for diagnosing and analyzing elevator car defects based on industrial vision. The present invention adopts the following technical solution: An industrial vision-based method for diagnosing and analyzing elevator car defects includes the following steps: Acquire a low-resolution image stream of the weld and perform local feature analysis on the low-resolution image stream of the weld to identify potential anomaly areas; When a potential abnormal area is identified, a high-resolution local image of the potential abnormal area is acquired, and the device motion information during the high-resolution local image acquisition process is collected. Based on the device motion information, motion correction is performed on the high-resolution local image. Enhancement processing is performed on the motion-corrected high-resolution local image, and various visual features are extracted from the enhanced high-resolution local image; Based on the extracted visual features, defects are judged and classified in the local areas corresponding to potential abnormal areas, and the confidence level of defect judgment is output. The standard for defect judgment is adaptively adjusted according to the material information and process parameter information in the production process. Based on the confidence level of defect judgment, identify samples with low confidence level of defect judgment, push the identified samples to the manual review queue to obtain annotation, incrementally update the adaptively adjusted defect judgment standard based on the manually annotated samples, identify new normal process texture patterns, and write the new normal process texture patterns into the normal sample library. Through this technical solution, this application can effectively distinguish between normal process textures and minor defects by multi-stage image processing and adaptive adjustment of defect judgment criteria, thereby reducing false alarms and missed detection rates. At the same time, through incremental updates and recognition of new normal process texture patterns, the system's adaptability to changes in the production environment is improved, thus solving the problem of difficulty in balancing defect recognition accuracy and efficiency in the prior art. This application also discloses an elevator car defect diagnosis and analysis system based on industrial vision, applied to the aforementioned elevator car defect diagnosis and analysis method based on industrial vision. The system includes: The analysis module acquires a low-resolution image stream of the weld and performs local feature analysis on the low-resolution image stream of the weld to identify potential abnormal areas. The correction module, when a potential abnormal area is identified, triggers the acquisition of a high-resolution local image of the potential abnormal area, and collects the device motion information during the acquisition process. Based on the device motion information, motion correction is performed on the high-resolution local image. The processing module enhances the motion-corrected high-resolution local image and extracts various visual features from the enhanced high-resolution local image. The adjustment module, based on the extracted visual features, performs defect judgment and classification on the local areas corresponding to potential abnormal areas, outputs the defect judgment confidence level, and adaptively adjusts the defect judgment standard according to the material information and process parameter information in the production process. The identification module identifies samples with low defect confidence based on the defect judgment confidence level, pushes the identified samples to the manual review queue to obtain annotations, incrementally updates the adaptively adjusted defect judgment standards based on the manually annotated samples, and identifies new normal process texture patterns and writes the new normal process texture patterns into the normal sample library. Through modular design, image analysis, correction, processing and adjustment functions are effectively integrated, thereby providing hardware and software support for the automated and intelligent diagnosis of elevator car weld defects, effectively solving the problems of low system integration and single function in existing technologies. This application effectively improves the accuracy, robustness, and intelligence level of elevator car weld defect diagnosis through a multi-stage, adaptive, and incremental learning strategy, which is significantly better than existing technologies. To further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are for reference and illustration only and are not intended to limit the present invention. Attached Figure Description

[0004] Figure 1 This is a flowchart of the elevator car defect diagnosis and analysis method based on industrial vision according to the present invention; Figure 2 This is a schematic diagram of the elevator car defect diagnosis and analysis system based on industrial vision according to the present invention. Detailed Implementation

[0005] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. Furthermore, the accompanying drawings of the present invention are for simple illustrative purposes only and are not depictions of actual dimensions; this is stated in advance. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of protection of the present invention. This embodiment provides a method and system for diagnosing and analyzing elevator car defects based on industrial vision, combined with... Figure 1 and Figure 2 As shown. refer to Figure 1 An elevator car defect diagnosis and analysis method based on industrial vision, the method includes the following steps: Acquire a low-resolution image stream of the weld and perform local feature analysis on the low-resolution image stream of the weld to identify potential anomaly areas; When a potential abnormal area is identified, a high-resolution local image of the potential abnormal area is acquired, and the device motion information during the high-resolution local image acquisition process is collected. Based on the device motion information, motion correction is performed on the high-resolution local image. Enhancement processing is performed on the motion-corrected high-resolution local image, and various visual features are extracted from the enhanced high-resolution local image; Based on the extracted visual features, defects are judged and classified in the local areas corresponding to potential abnormal areas, and the confidence level of defect judgment is output. The standard for defect judgment is adaptively adjusted according to the material information and process parameter information in the production process. Based on the confidence level of defect judgment, identify samples with low confidence level of defect judgment, push the identified samples to the manual review queue to obtain annotation, incrementally update the adaptively adjusted defect judgment standard based on the manually annotated samples, identify new normal process texture patterns, and write the new normal process texture patterns into the normal sample library. In this application, "low-resolution image stream of welds" refers to a sequence of relatively low-resolution weld images continuously acquired on a production line using industrial cameras. Its primary purpose is to perform rapid preliminary scanning to identify areas that may exhibit anomalies. "Potential anomaly areas" refer to local areas in the low-resolution image stream that are suspected of having defects or abnormal textures through preliminary analysis. "High-resolution local images" refer to images with richer details acquired using higher-resolution cameras or zoom lenses for potential anomaly areas, used for refined analysis. "Equipment motion information" refers to the motion parameters such as displacement, speed, and attitude of the image acquisition equipment (e.g., cameras, robotic arms) during the acquisition of high-resolution local images; this information is crucial for subsequent image motion correction. "Defect judgment confidence" is the probability or credibility of the system's judgment on whether a region is a defect and its defect type, reflecting the certainty of the system's judgment. "Human review queue" is a set of samples awaiting secondary confirmation and annotation by human experts, used to process samples with low system judgment confidence or those subject to controversy. The "Normal Sample Library" is a database that stores known normal weld texture patterns for reference and comparison when the system makes defect judgments. "New Normal Process Texture Patterns" refer to new normal weld textures that appear during the production process due to process improvements or material changes. These textures may be initially misjudged as abnormal by the system, but should be included in the normal sample library after verification. Firstly, various methods can be employed to acquire low-resolution image streams of weld seams. For example, one or more linear scan cameras can be deployed above the weld seam inspection station in an elevator car to continuously scan the passing weld seam at a constant speed, generating a low-resolution image stream. These cameras can be configured with lower resolution to ensure that the image acquisition speed meets the high-speed requirements of the production line. Another approach is to use area scan cameras to capture images at a fixed frame rate and combine consecutive frames into an image stream using image stitching technology. When performing local feature analysis on the low-resolution image stream of weld seams to identify potential anomaly areas, algorithms based on image grayscale changes, edge detection, or texture analysis can be used. For example, the grayscale gradient of a local area of ​​the image can be calculated, and when the gradient value exceeds a preset threshold, it is marked as a potential anomaly area. Alternatively, texture analysis methods such as Gabor filters can be used to identify areas that significantly differ from the texture of normal weld seams. When a potential anomaly region is identified, it is necessary to trigger the acquisition of a high-resolution local image of that region. This can be achieved by controlling a robotic arm equipped with a high-resolution area scan camera and a motorized zoom lens. After the low-resolution image analysis module transmits the coordinates of the potential anomaly region, the robotic arm quickly moves above that region and adjusts its focus to acquire a high-resolution local image. Simultaneously, to ensure image quality, it is necessary to collect the device's motion information during the high-resolution local image acquisition process. This can be achieved by integrating an inertial measurement unit or optical encoder onto the robotic arm to record its displacement, velocity, and angular velocity along the X, Y, and Z axes in real time. Based on this motion information, motion correction can be performed on the high-resolution local image. For example, an image registration-based algorithm can be used to predict the direction and degree of blurring during image acquisition using motion information, and then a deconvolution filter can be applied for image deblurring. Alternatively, a motion compensation algorithm can be used to correct image distortion caused by motion through geometric transformations. Enhancement processing of motion-corrected high-resolution local images is a crucial step in improving defect visibility. This can include operations such as contrast enhancement, noise suppression, and sharpening. For example, adaptive histogram equalization can be used to enhance local image contrast, making the difference between defects and the background more apparent. To suppress random noise in the image, nonlocal mean filtering or bilateral filtering can be applied. Sharpening can be achieved through the Laplacian operator or unsharpened masking to highlight image edges and details. Extracting various visual features from the enhanced high-resolution local images can employ a variety of advanced image processing techniques. For example, texture features based on local binary patterns, keypoint features based on scale-invariant feature transformation or accelerated robust features, and high-level semantic features automatically learned by deep learning convolutional neural networks can be extracted. These features can describe the local region of the weld from different dimensions, providing rich information for subsequent defect assessment. Based on extracted visual features, defects are identified and classified in local areas corresponding to potential abnormal regions, and the confidence score of the defect judgment is output. This can be achieved by training a multi-classifier model, such as a support vector machine, random forest, or deep neural network. The model determines the type of defect (e.g., crack, porosity, incomplete penetration) or normal texture of the region based on the input visual features and outputs a confidence score. For example, a crack judgment with a confidence score of 0.95 indicates that the system has high confidence in the judgment. Furthermore, the defect judgment criteria are adaptively adjusted based on material and process parameter information during production. For example, when the system detects that the current batch uses a specific alloy material, or when process parameters such as welding current and speed change, the threshold or weight of the defect judgment model can be dynamically adjusted to adapt to the new production conditions. This can be achieved by establishing a mapping relationship between parameters and defect features, or by allowing the system to autonomously learn and adjust strategies through reinforcement learning. Based on the confidence level of defect judgment, samples with low confidence levels are identified. For example, a confidence threshold can be set; samples below this threshold (e.g., confidence level below 0.7) are considered low-confidence samples. These identified samples are pushed to a manual review queue for labeling. These low-confidence samples are sent to a manual interface for visual inspection and accurate labeling by experienced quality inspectors. The adaptively adjusted defect judgment criteria are incrementally updated based on the manually labeled samples. Once the samples in the manual review queue are labeled, this new labeled data is used for small-batch, continuous retraining of the defect judgment model, enabling the model to learn new defect patterns or correct previous erroneous judgments. Simultaneously, new normal process texture patterns are identified and added to the normal sample library. For example, if manual review reveals that an area marked as abnormal by the system is actually a new, harmless, normal weld texture, the features of this texture are extracted and added to the normal sample library to prevent future misjudgments. This application further proposes a method for diagnosing and analyzing elevator car defects based on industrial vision, which also includes the following steps: A network of miniature environmental sensors is deployed near the elevator car weld inspection station to continuously collect environmental data at a frequency synchronized with the image acquisition device. The environmental data is preprocessed, and a timestamp is added to each environmental data point; The timestamps and spatial location information of visually abnormal events corresponding to potential abnormal regions are correlated with preprocessed environmental data, and the presence of specific patterns in the preprocessed environmental data related to the formation of normal microstructures is analyzed. Visually anomalous events exhibiting specific patterns are labeled as context-dependent, environment-related anomalies. Based on the labeling of contextual environmental association anomalies, visual abnormal events are judged for defects. When the confidence of defect judgment is in the high uncertainty range and there are labels of contextual environmental association anomalies, the possibility of visual abnormal events being judged as real defects is reduced, and visual abnormal events are classified as contextual normal textures. Contextual normal texture samples are not pushed to the manual review queue; For samples that are identified as defective or have high uncertainty but no contextual environmental anomalies, the corresponding samples will be pushed to the manual review queue. Statistical analysis is performed on samples that are consistently classified as contextual normal textures. Visual features that appear simultaneously with specific patterns but are not judged as defects are clustered, automatically included in the normal sample library, and the criteria for defect judgment are updated. Specifically, a network of miniature environmental sensors is deployed near the elevator car weld inspection station to capture real-time, precise changes in the microscopic environment of the inspection area. This sensor network can include, but is not limited to, temperature sensors, humidity sensors, light sensors, and vibration sensors. These sensors are configured to continuously collect environmental data at a frequency synchronized with the image acquisition device, ensuring a high degree of temporal consistency between environmental data and image data, providing an accurate basis for subsequent correlation analysis. Preprocessing environmental data can be understood as cleaning, denoising, and formatting the raw collected environmental data to eliminate sensor errors and environmental interference, thereby improving data quality and usability. Adding a timestamp to each environmental data point aims to accurately record the time of data collection so as to accurately match it with the timestamps of visual anomalies. In practical applications, the timestamps and spatial location information of visually abnormal events corresponding to potential abnormal areas are correlated with preprocessed environmental data. This can be achieved, for example, through database queries or data fusion algorithms. After correlation, the system analyzes whether the preprocessed environmental data contains specific patterns related to the formation of normal microstructures. These specific patterns refer to features on the weld surface that, under certain environmental conditions, visually resemble defects but are actually normal process textures. For example, under specific humidity levels, slight water vapor condensation textures may appear on the weld surface, which could be visually misjudged as defects. When this specific pattern is identified, the corresponding visual anomaly event will be marked as a context-dependent anomaly. This means that the occurrence of the visual anomaly is highly correlated with specific environmental conditions. Based on this marking, the system will make a defect judgment on the visual anomaly event. Specifically, when the confidence level of the defect judgment is in the high uncertainty range (i.e., the system has difficulty in clearly determining whether it is a defect or normal), and the event is marked as a context-dependent anomaly, the system will reduce the likelihood of it being judged as a real defect and classify it as a context-dependent normal texture. A context-dependent normal texture refers to a normal visual manifestation that does not belong to a defect and appears under specific environmental conditions. This application's solution introduces a miniature environmental sensor network to collect and analyze environmental data synchronized with image acquisition in real time, enabling the association of potential visual anomalies with specific environmental patterns. When the confidence level of a visual anomaly is low, by combining the labeling of environmentally associated anomalies, the system can intelligently identify contextually normal textures caused by environmental factors, avoiding misclassification as real defects. Thus, the system can more accurately classify visual anomalies, reduce false alarms caused by environmental factors, and avoid pushing contextually normal texture samples to the manual review queue, significantly reducing the burden of manual review. Furthermore, by statistically analyzing and clustering samples continuously classified as contextually normal textures, the system can automatically learn and incorporate new normal process texture patterns, further improving the normal sample library and enhancing the adaptability and robustness of the defect judgment criteria. This application further proposes a method for diagnosing and analyzing elevator car defects based on industrial vision, which also includes the following steps: Calculate the feature distance between the new contextual normal texture sample and all known normal process texture patterns in the normal sample library. When the feature distance exceeds the preset similarity threshold, the new contextual normal texture sample is marked as a potential new normal pattern sample. The activation pattern evolution tracking mechanism dynamically clusters the visual features and environmental association patterns of potential new normal pattern samples within a preset observation period, resulting in multiple clusters and corresponding cluster centers. A pattern separation evaluation method is introduced to calculate the feature distance between different cluster centers. Based on the feature distance between different cluster centers, the separation degree of each cluster relative to the cluster centers of known normal process texture patterns is calculated, and the compactness of samples within each cluster is evaluated. When the number of samples within a cluster reaches a preset value and the separation degree exceeds a preset separation threshold, the cluster is identified as a new normal process texture pattern. Continuously monitor the frequency of occurrence of new normal process texture patterns and the stability of the environmental association patterns corresponding to the new normal process texture patterns. When the frequency of occurrence of new normal process texture patterns is stable during long-term operation and the corresponding environmental association patterns remain consistent, the new normal process texture patterns will be automatically included in the normal sample library and the defect judgment criteria will be updated. Periodic pattern consistency checks are performed on the normal sample library. When it is found that the matching degree between the visual features corresponding to the included patterns and the current environment-related patterns has decreased over a long period of time, or when the included patterns overlap with newly emerging defective patterns in the feature space, a normal pattern drift warning is issued and pushed to the manual review queue. Specifically, calculating the feature distance between a new contextual normal texture sample and all known normal process texture patterns in the normal sample library involves obtaining the visual feature vector of the new contextual normal texture sample using a predefined feature extraction algorithm and comparing it with the feature vector of each known normal process texture pattern stored in the normal sample library. For example, Euclidean distance or cosine similarity can be used to calculate the feature distance. When all these feature distances exceed a preset similarity threshold, it indicates that the sample differs significantly from existing normal patterns and is therefore marked as a potential new normal pattern sample. The purpose is to initially screen samples that may represent new processes or new environmental conditions. The activation pattern evolution tracking mechanism dynamically clusters the visual features and environmental association patterns of potential new normal pattern samples within a preset observation period, resulting in multiple clusters and corresponding cluster centers. This can be understood as the system continuously collecting these potential new normal pattern samples over a period of time (e.g., several hours, days, or weeks), and combining their visual features with collected environmental data (such as temperature, humidity, equipment parameters, etc.), using unsupervised learning algorithms (such as K-means, DBSCAN, etc.) for dynamic cluster analysis to discover whether there are stable, recurring patterns. Its purpose is to identify truly representative new patterns from noise and random fluctuations. In practical applications, a pattern separation evaluation method is introduced. This method calculates the feature distances between different cluster centers, and based on these distances, calculates the separation degree of each cluster relative to the cluster centers of known normal process texture patterns. It also evaluates the density of samples within each cluster. When the number of samples within a cluster reaches a preset value and the separation degree exceeds a preset threshold, the cluster is identified as a new normal process texture pattern. Specifically, after dynamic clustering, the system calculates the distances between the newly formed cluster centers and the distances between these new cluster centers and the cluster centers of known normal process texture patterns in the normal sample library to quantify their "separation degree." Simultaneously, it evaluates the density of samples within each cluster (e.g., by calculating the average distance from samples within the cluster to the cluster center) to ensure the effectiveness of the clustering. When a cluster meets both the conditions of the number of internal samples and the separation degree, it is confirmed as a stable new normal process texture pattern. The purpose is to ensure that newly identified patterns are independent and representative, avoiding misclassification of abnormal or unstable patterns as normal. Furthermore, the system continuously monitors the frequency of new normal process texture patterns and the stability of their corresponding environmental association patterns. When a new normal process texture pattern exhibits a stable frequency of occurrence over long-term operation and its corresponding environmental association patterns remain consistent, the new normal process texture pattern is automatically added to the normal sample library, and the defect judgment criteria are updated. This means the system will track identified new normal process texture patterns over a long period, observing whether their frequency of occurrence in production is stable and whether their associated environmental conditions remain consistent. Once these stability conditions are met, the new pattern will be automatically added to the normal sample library and used to update the defect judgment criteria. The purpose is to enable the system to adaptively learn and adapt to changes in the production environment, maintaining the timeliness and accuracy of the defect judgment criteria. In addition, the system performs periodic pattern consistency checks on the normal sample library. When it detects a long-term decline in the matching degree between the visual features of an included pattern and the current environment's associated patterns, or when an included pattern overlaps with a newly identified defective pattern in the feature space, a normal pattern drift warning is issued and pushed to a manual review queue. Specifically, the system periodically checks the patterns already in the normal sample library, comparing their visual features with the currently collected real-time environmental data to assess their matching degree. If the matching degree declines over a long period, or if the feature space of a normal pattern begins to overlap with the feature space of a newly identified defective pattern, the system issues a warning indicating possible normal pattern drift and submits it for manual review. The purpose is to promptly detect the degradation or change of normal patterns and prevent misjudgments or missed detections due to pattern drift. This application's solution, by introducing a pattern evolution tracking mechanism and a pattern separation evaluation method, can proactively identify and confirm new normal process texture patterns. This solves the problem of existing methods lagging in updating the normal sample library and failing to adapt to new "normal" states in a timely manner when facing changes in production processes or environments. Simultaneously, by periodically verifying the consistency of the normal sample library and issuing warnings about normal pattern drift, this application effectively addresses the limitation that included normal patterns may gradually deviate from their original definitions due to minor changes in the production environment or process, or even overlap with defect patterns, leading to system misjudgments or missed detections. It is precisely because of these mechanisms that the system can dynamically maintain and update its understanding of "normal," ensuring the accuracy and robustness of defect diagnosis. This application further proposes the following steps before performing periodic pattern consistency verification on the normal sample library: Continuously monitor and record the visual features of each pattern that has been included in the normal sample library, the environmental associated pattern parameters corresponding to the pattern, and the physical location information of the pattern on the production line to form a historical drift record for each pattern. Based on the historical drift records and current production line load for each mode, calculate the mode stability index for each mode; When the pattern stability index is lower than the preset stability threshold, the pattern verification frequency is increased; When the pattern stability index is higher than the preset stability threshold, the pattern verification frequency is reduced. Based on the adjusted pattern verification frequency, periodic pattern consistency verification is performed on the included normal sample library. Specifically, before periodically verifying the consistency of patterns in the normal sample library, it is necessary to continuously monitor and record the visual characteristics, environmentally relevant pattern parameters, and physical location information of each pattern included in the normal sample library. "Visual characteristics" can be understood as quantitative data describing the pattern's appearance, texture, color, etc.; "environmentally relevant pattern parameters" refer to environmental sensor data related to the pattern's formation, such as temperature, humidity, and air pressure; and "physical location information" refers to the specific location or area of ​​the pattern on the elevator car weld. By continuously recording this information, a "historical drift record" for each pattern can be created, reflecting the pattern's changing trend over time, environment, and location. Furthermore, based on the historical drift records of each pattern and the current production line load, a "pattern stability index" is calculated for each pattern. The pattern stability index is a quantitative indicator used to assess the degree of change and reliability of the pattern over a period of time. For example, this index can be calculated by analyzing the fluctuation range and rate of change of visual features, environmentally related pattern parameters, and physical location information in the historical drift records. Production line load can be considered as an external factor affecting pattern stability; for example, high load may cause fluctuations in process parameters, thus affecting pattern stability. Based on this, the verification frequency of the model is dynamically adjusted according to the calculated model stability index. Specifically, when the model stability index is lower than the preset stability threshold, it indicates that the model may have a significant risk of change or drift. In this case, the verification frequency of the model needs to be increased to detect potential model drift more promptly. Conversely, when the model stability index is higher than the preset stability threshold, it indicates that the model is relatively stable under the current conditions. In this case, the verification frequency of the model can be appropriately reduced to save computing resources and improve the overall system efficiency. Finally, based on the adjusted pattern verification frequency, periodic pattern consistency checks are performed on the included normal sample library. This means that different normal process texture patterns can have different verification cycles, thereby achieving more refined and adaptive pattern management. This application's solution effectively addresses the limitations of traditional fixed-cycle verification by introducing a pattern stability index and dynamically adjusting the verification frequency. Specifically, by continuously monitoring and recording historical drift data of patterns, the system can comprehensively grasp the evolution trajectory of each normal process texture pattern and the degree to which it is affected by the environment and production line load. Based on this historical data and real-time production line load, the calculated pattern stability index can objectively reflect the current reliability and changing trend of the pattern. It is precisely because of this dynamic evaluation mechanism that the system can intelligently determine which patterns need more frequent attention and which patterns can be appropriately relaxed in monitoring. When the pattern stability index is low, increasing the verification frequency can ensure that the pattern is detected in the early stages of significant drift, thereby avoiding misjudgment or missed judgment caused by pattern drift; while when the pattern stability index is high, reducing the verification frequency can avoid unnecessary resource consumption and improve the system's operating efficiency. This adaptive verification strategy enables the entire defect diagnosis and analysis method to cope more flexibly and efficiently with the complexity and dynamism of the production environment. In some preferred embodiments, it is assumed that two main normal process texture patterns exist during elevator car weld inspection: Pattern A and Pattern B. Pattern A is typically formed in a stable production environment, and its visual characteristics and environmentally related pattern parameters show minimal fluctuations in historical records, meaning its pattern stability index remains consistently high. Pattern B, on the other hand, is more sensitive to changes in production line load and ambient temperature. Its historical drift records show slight characteristic fluctuations under specific conditions, and its pattern stability index sometimes falls below a preset stability threshold. According to the scheme of this application, the system continuously monitors and records the visual characteristics, environmental associated mode parameters, and physical location information of Mode A and Mode B, forming their respective historical drift records. For example, for Mode A, its historical drift record shows that its stability index is consistently higher than the preset stability threshold of 0.8. At this time, the system will reduce the verification frequency of Mode A, for example, from once per hour to once every four hours. For Mode B, when the system detects an increase in production line load or fluctuations in ambient temperature, its mode stability index may drop to 0.6, lower than the preset stability threshold of 0.7. At this time, the system will automatically increase the verification frequency of Mode B, for example, from once per hour to once every 15 minutes. Through this dynamic adjustment, the system can concentrate more computing resources on mode B, which is prone to drift, ensuring timely monitoring of its consistency and issuing early warnings at the initial stage of potential drift in mode B. Simultaneously, for the highly stable mode A, unnecessary verifications are reduced, effectively conserving system resources. When the stability index of mode B returns to a high level, the verification frequency is automatically lowered. This adaptive verification mechanism enables the entire defect diagnosis system to operate more intelligently and efficiently, ensuring accurate management of normal modes under different production conditions. This application further proposes a method for priority management of normal mode drift warnings, which includes the following steps: For each normal mode drift warning issued, calculate the potential impact of each warning on product quality. Based on the calculated potential impact, multiple simultaneous normal pattern drift warnings are prioritized. The highest priority normal mode drift warning will be pushed to the manual review queue, along with the corresponding potential impact information. Specifically, calculating the potential impact of each normal pattern drift warning on product quality involves conducting a risk assessment for each identified normal pattern drift warning to quantify its potential negative impact on the final product quality. This impact calculation can be based on various factors, such as the magnitude of the pattern drift, the similarity between the drifted pattern and known defect patterns, and the correlation between the pattern drift and product quality issues (such as rework rates and scrap rates) in historical data. The purpose is to provide a quantitative basis for subsequent prioritization. This process prioritizes multiple simultaneous normal pattern drift warnings based on their calculated potential impact. This can be understood as follows: when the system detects multiple normal pattern drift warnings, it arranges them according to the calculated potential impact value of each warning, following preset rules (e.g., higher impact, higher priority). In practical applications, numerical comparison, threshold division, or machine learning models can be used for prioritization to ensure that the most critical warnings receive priority attention. Furthermore, pushing the highest-priority normal mode drift alert to the manual review queue, along with corresponding potential impact information, means that after prioritizing alerts, the system only sends the alerts currently determined to be of the highest priority to manual reviewers. Simultaneously, to assist reviewers in making decisions, the alert also includes its corresponding potential impact information, allowing reviewers to intuitively understand the severity of the alert and thus more effectively allocate review resources and formulate response strategies. This application's solution effectively addresses the issues of information overload and low response efficiency inherent in basic solutions by introducing a mechanism for calculating and prioritizing the potential impact of normal mode drift warnings. When the system issues multiple normal mode drift warnings, it no longer simply pushes all warnings to a manual review queue. Instead, it first performs a refined potential impact assessment on each warning. This assessment allows the system to distinguish the severity and urgency of different warnings. Subsequently, based on these quantified impact levels, the system intelligently prioritizes the warnings, ensuring that those warnings with the greatest impact on product quality are identified first. Finally, only the highest-priority warnings, along with their impact information, are pushed to manual reviewers. This not only reduces the workload of manual reviewers, allowing them to focus on the most critical issues, but also, by providing information on the potential impact level, enables reviewers to understand the severity of the problem more quickly and accurately, thus making more timely and effective decisions. In some preferred embodiments, it is assumed that during the elevator car weld inspection process, the system simultaneously detects three normal mode drift warnings: warning A, warning B, and warning C. First, the system calculates the potential impact of each warning on product quality. For example, by analyzing the high degree of overlap between the visual features of warning A and known severe defect patterns in the feature space, and its association with high rework rates in historical production, warning A is calculated to have a "high" potential impact. Warning B has a small drift amplitude and a low historical correlation with product quality issues, resulting in a "medium" potential impact. Warning C exhibits a slight texture change and has almost no historical correlation with product quality, resulting in a "low" potential impact. Next, the system prioritizes the three warnings based on their potential impact. In this example, warning A has the highest priority, followed by warning B, and warning C has the lowest priority. Ultimately, the system pushes the highest-priority alert A to the manual review queue, along with information about its "high" potential impact. Upon receiving this alert, the manual reviewer can immediately identify its high risk and prioritize its review and handling, such as initiating more detailed inspections or adjusting production parameters to avoid potential quality issues. Alerts B and C, on the other hand, may be temporarily shelved or processed later when their priority is lower, thus optimizing the allocation of limited review resources. In some embodiments of this application described above, for each issued normal pattern drift warning, it is necessary to calculate the potential impact of each warning on product quality and prioritize the warnings based on this potential impact. Specifically, the potential impact is quantified by assessing the degree of overlap between the visual features of the pattern and the feature space of known severe defect patterns, as well as the frequency with which the pattern has been associated with product quality problems in historical production. The "potential impact level" refers to the quantitative assessment of the potential negative impact of normal mode drift warnings on the final product quality. This assessment aims to provide an objective basis for subsequent warning prioritization, ensuring that resources are allocated preferentially to potential issues that have the greatest impact on product quality. Specifically, quantifying the potential impact involves two main aspects. The first aspect is "the degree of overlap between the visual features of the assessment pattern and the feature space of known severely defective patterns." This means comparing the visual features of the drifting normal pattern with the visual features of patterns known to the system and identified as severely defective. For example, this overlap can be quantified by calculating the distance, similarity, or area of ​​overlap between the two patterns in the feature space. The higher the degree of overlap, the closer the drifting pattern is to a known severe defect, and the greater its potential impact. The second aspect is "the frequency with which patterns are associated with product quality problems in historical production." This requires the system to record and analyze the association between each pattern and actual product quality problems (e.g., rework, scrap, customer complaints) in past production processes. For example, it can be used to statistically analyze the number or proportion of product quality problems that occurred after a specific pattern emerged. The higher the association frequency, the more likely the pattern has historically been to cause quality problems, and the greater its potential impact. By combining these two aspects of evaluation, a more comprehensive and accurate quantitative value of the potential impact can be obtained. This application's solution quantifies the potential impact of normal pattern drift warnings, enabling the system to more objectively and accurately assess the severity of each warning. Specifically, by evaluating the degree of overlap between the visual features of a drifting pattern and the feature space of a known severe defect pattern, the likelihood and severity of the drifting pattern transforming into an actual defect can be directly reflected. For example, if the visual features of a normal pattern begin to approach the feature space of a known defect pattern that leads to structural failure, its overlap will significantly increase, indicating a higher potential impact. Simultaneously, by analyzing the frequency with which patterns are associated with product quality issues in historical production, historical data and experience can be used to predict future risks. If a pattern has repeatedly caused product rework or scrapping in the past, even if its current visual features do not overlap with severe defects, its historical association frequency will increase its potential impact. This dual assessment mechanism ensures that the quantification of potential impact considers both current technical feature similarities and historical actual production experience, thus providing solid data support for subsequent warning prioritization. This application further proposes that when the highest priority normal mode drift warning is pushed to the manual review queue, the visual alert level of the warning is adjusted according to the corresponding potential impact information; the core data summary related to the warning is filtered and displayed according to the potential impact level; based on the core data summary related to the warning and the adjusted visual alert level of the warning, an interactive review interface is provided, allowing reviewers to adjust the display order of the warnings or filter the warnings, and a preset review template is provided. Specifically, adjusting the visual alert level of a warning refers to dynamically changing its visual presentation on the review interface based on the potential impact of a normal mode drift warning on product quality. For example, when the potential impact is high, the warning can be set to red, flashing, or displayed in a larger font to immediately attract the reviewer's attention; when the potential impact is low, a softer color or smaller icon can be used. The aim is to help reviewers quickly distinguish the urgency and importance of warnings through intuitive visual differences, thereby prioritizing high-risk warnings. The filtering and display of core data summaries related to the early warning can be understood as the system automatically extracting the most directly relevant and valuable information from massive amounts of data based on the potential impact of the warning, and presenting it to reviewers in a concise and clear manner. These core data summaries may include specific parameters of pattern drift, affected production batches, statistics of similar historical events, changing trends in environmental correlation patterns, and potential types of product quality problems. The aim is to avoid reviewers spending time searching for key information in large amounts of raw data, improving information acquisition efficiency, and providing data support for decision-making. In practical applications, providing an interactive review interface refers to building a user-friendly operating platform. This platform not only displays alert information and core data summaries but also grants reviewers certain operational permissions. For example, reviewers can adjust the display order of alerts in the list according to their work habits or current task priorities, placing the most urgent or concerning alerts at the top. They can also set filtering conditions to view only alerts of specific types, impact levels, or time periods for focused processing. Furthermore, the interface can provide preset review templates, which may include standardized review processes, solutions to common problems, and structured fields for recording review results and decisions. The aim is to optimize the reviewer's work experience and improve review efficiency and decision accuracy by providing flexible interactive tools and standardized review processes. This application's solution effectively addresses the inefficiencies and inaccurate decisions that may arise in traditional early warning push methods during manual review by combining the highest-priority normal pattern drift warning with information on potential impact levels, and further optimizing multi-dimensional information presentation and interaction. Specifically, by adjusting the visual alert level of the warning, the system leverages human visual sensitivity to color, size, and dynamic changes to deliver the most critical warning information to reviewers in the most prominent way, enabling them to quickly identify and prioritize high-risk events. Simultaneously, by filtering and displaying core data summaries related to the warning, reviewers can obtain the crucial contextual information needed for decision-making without manually searching and analyzing large amounts of data, significantly reducing the time required for information understanding and analysis. Furthermore, an interactive review interface is provided, allowing reviewers to adjust the display order of warnings or filter them according to actual needs, and providing preset review templates. This not only enhances the flexibility and personalization of the review process but also guides reviewers to engage in systematic thinking and recording through standardized templates, ensuring the standardization and traceability of review decisions. It is precisely because of these synergistic effects that this application's solution significantly improves the efficiency and quality of manual review. In some preferred embodiments, it is assumed that on the elevator car weld production line, the system detects a normal pattern drift warning, which indicates that a certain weld texture pattern is gradually deviating from its normal range. The system first calculates, based on historical data of the pattern drift and current production parameters, that its potential impact on product quality is "moderately high". When the highest priority normal mode drift alert is pushed to the manual review queue, the system will automatically adjust the visual warning level of the alert based on the "moderate to high" potential impact information. For example, on the review interface, the alert item may be highlighted with an orange background and accompanied by a medium-sized exclamation mark icon to remind the reviewer of its importance. Simultaneously, the system intelligently filters and displays core data summaries related to the warning based on the potential impact. These summaries may include: the visual characteristic change curve of the drift pattern, the affected weld area number, the ambient temperature and humidity change trends for the last three shifts, historical product rework rate data related to the drift pattern, and the possible causes initially determined by the system (e.g., welding torch wear or minor batch differences in materials). This information is presented in the form of charts and concise text on the warning details page. After entering the queue through the interactive review interface, reviewers can see all pending alerts. This interface allows reviewers to drag "moderately high" alerts to the top of the list for priority processing. If reviewers wish to view only alerts related to environmental factors, they can use the interface's filtering function to display only those related to environmental pattern changes. Furthermore, the interface provides a preset "pattern drift review template" containing standardized fields such as "drift cause analysis," "recommended measures," and "reviewer's signature," guiding reviewers to systematically complete the review process and record their decisions and analysis results. In this way, reviewers can efficiently obtain information, operate flexibly, and complete the review of normal pattern drift alerts in a standardized manner. The above method also includes the following steps: The system analyzes the decision-making content of the reviewer regarding the highest priority normal mode drift warning in the interactive review interface, and generates production line control instructions or quality control instructions. Transmit production line control commands or quality control commands to the production line controller; Receive the execution results and current production status feedback from the production line controller; Based on the execution results and current production status feedback, update the alert status and record the decision results and execution feedback; Based on the execution feedback from the production line controller, assess the potential impact of the decision on product quality after its implementation. Specifically, analyzing the decision-making content of the reviewer regarding the highest priority normal mode drift warning in the interactive review interface refers to the system performing semantic analysis and structured processing on the selections, inputs, or confirmations made by the reviewer in the interactive review interface. For example, the reviewer might select operations such as "adjust welding current," "change batch materials," "recalibrate image acquisition device," or "mark as false alarm." These decisions are then translated into specific production line control instructions or quality control instructions that can be understood and executed by automated equipment. Production line control instructions may include adjusting equipment parameters (such as welding speed, temperature, and pressure), initiating specific maintenance procedures, or modifying production processes; quality control instructions may involve additional testing of specific batches of products, isolating problematic products, or updating quality standards. Transmitting production line control or quality control commands to the production line controller can be understood as sending the generated commands to the corresponding production line controller via industrial communication protocols (such as Modbus TCP, EtherNet / IP, PROFINET, etc.). The production line controller is the core unit responsible for managing and coordinating the operation of various devices (such as robots, sensors, actuators, etc.) on the production line, and its purpose is to ensure that commands are accurately and timely transmitted and executed. In practical applications, receiving the execution results and current production status feedback from the production line controller refers to the system continuously listening to and acquiring the status information returned by the production line controller after receiving and executing instructions. This feedback information may include whether the instruction was successfully executed, any anomalies encountered during execution, the current operating parameters of the equipment, and real-time output, energy consumption, and other status data of the production line. Furthermore, based on the execution results and current production status feedback, the system updates the status of the warning and records the decision results and execution feedback. This means that the system updates the status of previously issued normal mode drift warnings based on the received feedback information, for example, marking them as "processed," "processing," or "processing failed." Simultaneously, the system records in detail all relevant information, including the reviewer's decision-making content, generated instructions, the production line controller's execution results, and current production status feedback, forming a complete decision chain and execution log. The purpose of this is to provide data support for subsequent traceability, analysis, and optimization. Finally, based on the execution feedback from the production line controller, the potential impact of the decision execution on product quality is assessed. This means that after the instruction is executed, the system combines the data from the production line feedback to predict and evaluate possible changes in product quality. For example, if the instruction is to adjust welding parameters, the system will analyze the potential impact of the adjusted welding parameters on weld quality indicators (such as strength, toughness, and appearance). The purpose is to quantify the actual effect of manual decision-making and provide a basis for subsequent decision optimization. This application's solution parses the decisions made by reviewers in an interactive review interface and transforms them into executable production line control or quality control instructions, thus achieving a deep integration of artificial intelligence and production line automation control. Because these instructions are transmitted to the production line controller and executed, manual review is no longer merely at the information level but can directly intervene in and influence the actual production process. Furthermore, by receiving execution results and current production status feedback from the production line controller, the system can understand the implementation of decisions in real time and update warning status and record detailed decision and execution logs accordingly, providing a foundation for subsequent traceability and analysis. Moreover, by evaluating the potential impact of decision execution on product quality based on the production line controller's execution feedback, the system can quantify the actual effect of manual decisions, forming a complete closed-loop feedback mechanism. This effectively solves the limitations of simply providing a review interface without a decision execution and effect evaluation mechanism. In some preferred embodiments, suppose the system issues a high-priority normal pattern drift warning, indicating that a specific texture pattern of a batch of elevator car welds has slightly drifted from the pattern in the normal sample library, and the potential impact is high. The reviewer examines the warning through an interactive review interface and, based on the core data summary and visual alert level provided by the system, determines that the drift may be related to slight fluctuations in the welding current. The reviewer selects the "Adjust Welding Current" option in the interface and enters the suggested adjustment range. The system then parses this decision and generates a production line control instruction, such as "Adjust the current parameter of the welding equipment from X Amps to Y Amps." This instruction is transmitted to the production line controller of the welding equipment via an industrial Ethernet network. Upon receiving the instruction, the production line controller immediately executes the current adjustment operation and feeds back the execution result (e.g., "Current adjustment successful") and the current welding parameters, equipment operating status, etc., to the system. Based on this feedback, the system updates the status of the normal pattern drift warning to "Processed" and records the reviewer's decision, the generated instruction, the execution result, and related production status data. Subsequently, the system will assess the potential impact of this current adjustment on weld quality based on the execution feedback from the welding equipment and the weld quality index data collected later. For example, it will analyze whether the strength and appearance of the weld have returned to normal levels after the adjustment, thereby verifying the effectiveness of human decision-making. Specifically, the steps for assessing the potential impact of decision implementation on product quality include the following. The steps for assessing the potential impact of a decision on product quality after its implementation include: Obtain various quality index data of elevator car welds; Based on the data of each quality indicator, independent data processing and preliminary evaluation are carried out according to the corresponding evaluation method and data characteristics to obtain preliminary evaluation results of multiple quality indicators. Set the weight parameters for each quality indicator; By comprehensively analyzing the preliminary evaluation results of multiple quality indicators and the weight parameters of each quality indicator, a potential impact index on product quality is obtained. Based on the product quality potential impact index, an overall assessment is conducted on the potential impact on product quality after the implementation of the decision. Obtaining various quality index data of elevator car welds refers to extracting various quality indicators of the weld from high-definition local images or obtaining them through other testing equipment using high-precision sensors, laser scanners, or specialized image analysis algorithms. These quality indicators may include, but are not limited to, weld width, weld height, penetration depth, weld width, reinforcement height, undercut depth, incomplete penetration length, number and size of pores, and slag inclusion distribution. Furthermore, based on the data for each quality indicator, independent data processing and preliminary evaluation are conducted according to the corresponding evaluation methods and data characteristics to obtain preliminary evaluation results for multiple quality indicators. Specifically, the data types and distributions of different quality indicators may vary. For example, for continuous indicators (such as weld width), statistical analysis methods (such as mean, standard deviation, skewness, and kurtosis) can be used for evaluation; for discrete indicators (such as porosity), counting or frequency analysis can be used. The evaluation method must also fully consider the physical meaning of the indicator and its potential impact on product performance. The preliminary evaluation results can be the pass rate, deviation value, or risk level for each indicator. Based on this, weight parameters are set for each quality indicator. These weight parameters reflect the relative importance of different quality indicators in the overall product quality assessment. For example, indicators critical to safety (such as penetration depth and incomplete penetration) can have higher weights, while indicators with a greater impact on appearance (such as weld reinforcement and undercut) may have relatively lower weights. These weight parameters can be set and adjusted based on expert experience, historical defect data analysis, or the requirements of the quality management system. Subsequently, the preliminary evaluation results of various quality indicators are comprehensively analyzed along with the weighting parameters of each quality indicator to obtain the product quality potential impact index. This comprehensive analysis can be achieved through weighted summation, multi-indicator decision analysis (such as the Analytic Hierarchy Process (AHP) or TOPSIS), or other multivariate statistical methods. The product quality potential impact index is a quantitative indicator used to comprehensively reflect the overall potential impact on product quality after the implementation of current decisions; a higher value indicates a greater potential impact. Finally, based on the product quality potential impact index, a comprehensive assessment is conducted on the potential impact on product quality after the decision is implemented. This comprehensive assessment involves comparing the calculated product quality potential impact index with preset risk thresholds or level standards to arrive at a final assessment conclusion, such as "low risk," "medium risk," "high risk," or a specific quality level. This provides a basis for subsequent quality control and production adjustments. This application's solution acquires multiple quality indicator data of elevator car welds and, based on each quality indicator, performs independent data processing and preliminary evaluation according to the corresponding assessment method and data characteristics, thereby obtaining preliminary evaluation results for multiple quality indicators. Subsequently, by setting weight parameters for each quality indicator and comprehensively analyzing the preliminary evaluation results with these weight parameters, a product quality potential impact index is obtained. Finally, the potential impact on product quality after decision implementation is comprehensively evaluated based on this index. This multi-dimensional, weighted, and comprehensive evaluation mechanism can comprehensively and objectively quantify the actual impact of decision implementation on product quality, avoiding the one-sidedness of single-indicator evaluation, and providing a more refined and reliable basis for subsequent quality control and production optimization. refer to Figure 2 This application proposes an elevator car defect diagnosis and analysis system based on industrial vision, which is applied to an elevator car defect diagnosis and analysis method based on industrial vision. The system includes: The analysis module is used to acquire a low-resolution image stream of the weld and perform local feature analysis on the low-resolution image stream of the weld to identify potential abnormal areas. The correction module is used to trigger the acquisition of high-resolution local images of potential abnormal areas when they are identified, and to collect equipment motion information during the acquisition process. Based on the equipment motion information, motion correction is performed on the high-resolution local images. The processing module is used to enhance the motion-corrected high-resolution local image and extract various visual features from the enhanced high-resolution local image. The adjustment module is used to judge and classify defects in local areas corresponding to potential abnormal areas based on multiple extracted visual features, output the confidence level of defect judgment, and adaptively adjust the standard of defect judgment according to material information and process parameter information in the production process. The identification module identifies samples with low defect confidence based on the defect judgment confidence level, pushes the identified samples to the manual review queue to obtain annotations, incrementally updates the adaptively adjusted defect judgment standards based on the manually annotated samples, and identifies new normal process texture patterns and writes the new normal process texture patterns into the normal sample library. Specifically, the analysis module can be understood as the system's front-end data acquisition and preliminary analysis unit. Its core function is to perform preliminary, large-scale scanning and anomaly warning of the elevator car welds. The low-resolution image stream of the welds refers to the continuous acquisition of weld surface image data through wide-angle or low-resolution cameras. Its purpose is to cover the largest possible detection range and quickly locate potentially problematic areas. Local feature analysis can employ various image processing algorithms, such as edge detection, texture analysis, and grayscale histogram analysis, to identify areas in the image that significantly differ from the texture of normal welds; these areas are then marked as potential anomaly regions. The correction module is activated after the analysis module identifies a potential anomaly region. Its main responsibility is to acquire more detailed image information and eliminate motion blur. High-resolution local image acquisition typically involves focusing on the potential anomaly region with a high-resolution or zoom camera to capture finer defect details. Equipment motion information can be acquired by encoders, inertial measurement units, or other position sensors to accurately record the relative motion trajectory of the camera or elevator car during image acquisition. Motion correction techniques, such as algorithms based on optical flow, feature matching, or geometric transformation, can deblur and align the high-resolution local image based on the acquired motion information, ensuring the accuracy of subsequent analysis. The processing module is responsible for optimizing and extracting features from the corrected image. Enhancement processing aims to improve image quality, such as through contrast enhancement, noise suppression, and sharpening techniques, to make defect features more prominent. The extraction of various visual features is fundamental to defect identification. These features can include, but are not limited to, geometric features (such as the shape, size, and location of the defect), texture features (such as roughness and directionality), color or grayscale features, and abstract features extracted by deep learning models. The richness of these features helps to describe potential defects more comprehensively and accurately. The adjustment module is the core decision-making and learning unit of the system. Based on various extracted visual features, this module uses machine learning or deep learning models to judge and classify defects in local areas corresponding to potential abnormal regions, such as cracks, porosity, and incomplete penetration, and outputs the defect judgment confidence score, which reflects the reliability of the judgment result. Furthermore, this module can adaptively adjust the defect judgment criteria based on material information (such as steel type and batch) and process parameter information (such as welding current, speed, and shielding gas flow rate) during the production process. This means that the system can dynamically optimize its defect identification logic according to changes in actual production conditions, improving the robustness of the judgment. Furthermore, the adjustment module is also responsible for continuous learning and optimization. When the confidence level of a defect judgment is low, the system identifies the corresponding sample and pushes it to the manual review queue. After the manual reviewers professionally annotate these samples, the system incrementally updates the adaptively adjusted defect judgment criteria based on these annotated samples, thereby continuously improving the accuracy of the model. At the same time, the system can also identify new normal process texture patterns that appear during the production process, such as new normal weld appearances caused by process improvements or material changes, and write these new normal process texture patterns into the normal sample library to avoid misjudging them as defects. The identification module is a key component of the elevator car defect diagnosis and analysis system. Its core function is to accurately identify and classify potential abnormal areas based on images and feature data acquired from the front end, combined with machine learning or deep learning models, and output the confidence level of the defect judgment. Through in-depth analysis and classification of weld images, this module can promptly detect welding quality problems and perform subsequent processing and feedback based on the defect judgment results. This application's solution visualizes each step of the elevator car weld defect diagnosis method as a functional module, realizing a complete automated process from preliminary analysis of low-resolution image streams to fine processing of high-resolution images, and then to intelligent judgment based on multiple features and adaptive standard adjustment. The analysis module can quickly screen out suspicious areas, avoiding costly fine processing of all data; the correction module ensures the quality of high-precision images, providing reliable input for subsequent analysis; the processing module provides rich and effective visual information for defect judgment; the adjustment module not only performs intelligent judgment, but more importantly, its adaptive adjustment and continuous learning capabilities enable the system to adapt to changes in the production environment and continuously improve diagnostic performance. Through this modular system design, the complex diagnostic process that originally required manual intervention or multi-stage collaboration can be executed efficiently and accurately automatically. The content disclosed above is only a preferred and feasible embodiment of the present invention, and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent technical changes made based on the content of the present invention specification and drawings are included within the scope of protection of the present invention. Furthermore, the elements therein can be updated as technology develops.

Claims

1. A method for diagnosing and analyzing elevator car defects based on industrial vision, characterized in that, The method includes the following steps: Acquire a low-resolution image stream of the weld and perform local feature analysis on the low-resolution image stream of the weld to identify potential anomaly areas; When a potential abnormal area is identified, a high-resolution local image of the potential abnormal area is acquired, and the device motion information during the high-resolution local image acquisition process is collected. Based on the device motion information, motion correction is performed on the high-resolution local image. Enhancement processing is performed on the motion-corrected high-resolution local image, and various visual features are extracted from the enhanced high-resolution local image; Based on the extracted visual features, defects are judged and classified in the local areas corresponding to potential abnormal areas, and the confidence level of defect judgment is output. The standard for defect judgment is adaptively adjusted according to the material information and process parameter information in the production process. Based on the confidence level of defect judgment, identify samples with low confidence level of defect judgment, push the identified samples to the manual review queue to obtain annotation, incrementally update the adaptively adjusted defect judgment standard based on the manually annotated samples, identify new normal process texture patterns, and write the new normal process texture patterns into the normal sample library.

2. The elevator car defect diagnosis and analysis method based on industrial vision as described in claim 1, characterized in that, The method also includes the following steps: A network of miniature environmental sensors is deployed near the elevator car weld inspection station to continuously collect environmental data at a frequency synchronized with the image acquisition device. The environmental data is preprocessed, and a timestamp is added to each environmental data point; The timestamps and spatial location information of visually abnormal events corresponding to potential abnormal regions are correlated with preprocessed environmental data, and the presence of specific patterns in the preprocessed environmental data related to the formation of normal microstructures is analyzed. Visually anomalous events exhibiting specific patterns are labeled as context-dependent, environment-related anomalies. Based on the labeling of contextual environmental association anomalies, visual abnormal events are judged for defects. When the confidence of defect judgment is in the high uncertainty range and there are labels of contextual environmental association anomalies, the possibility of visual abnormal events being judged as real defects is reduced, and visual abnormal events are classified as contextual normal textures. Contextual normal texture samples are not pushed to the manual review queue; For samples that are identified as defective or have high uncertainty but no contextual environmental anomalies, the corresponding samples will be pushed to the manual review queue. Statistical analysis is performed on samples that are consistently classified as contextual normal textures. Visual features that appear simultaneously with specific patterns but are not judged as defects are clustered, automatically included in the normal sample library, and the criteria for defect judgment are updated.

3. The elevator car defect diagnosis and analysis method based on industrial vision as described in claim 2, characterized in that, The method also includes the following steps: Calculate the feature distance between the new contextual normal texture sample and all known normal process texture patterns in the normal sample library. When the feature distance exceeds the preset similarity threshold, the new contextual normal texture sample is marked as a potential new normal pattern sample. The activation pattern evolution tracking mechanism dynamically clusters the visual features and environmental association patterns of potential new normal pattern samples within a preset observation period, resulting in multiple clusters and corresponding cluster centers. A pattern separation evaluation method is introduced to calculate the feature distance between different cluster centers. Based on the feature distance between different cluster centers, the separation degree of each cluster relative to the cluster centers of known normal process texture patterns is calculated, and the compactness of samples within each cluster is evaluated. When the number of samples within a cluster reaches a preset value and the separation degree exceeds a preset separation threshold, the cluster is identified as a new normal process texture pattern. Continuously monitor the frequency of occurrence of new normal process texture patterns and the stability of the environmental association patterns corresponding to the new normal process texture patterns. When the frequency of occurrence of new normal process texture patterns is stable during long-term operation and the corresponding environmental association patterns remain consistent, the new normal process texture patterns will be automatically included in the normal sample library and the defect judgment criteria will be updated. Periodic pattern consistency checks are performed on the normal sample library. When it is found that the matching degree between the visual features corresponding to the included patterns and the current environment-related patterns has decreased over a long period of time, or when the included patterns overlap with newly emerging defective patterns in the feature space, a normal pattern drift warning is issued and pushed to the manual review queue.

4. The elevator car defect diagnosis and analysis method based on industrial vision as described in claim 3, characterized in that, Before performing periodic pattern consistency checks on the normal sample library, the following steps are also included: Continuously monitor and record the visual features of each pattern that has been included in the normal sample library, the environmental associated pattern parameters corresponding to the pattern, and the physical location information of the pattern on the production line to form a historical drift record for each pattern. Based on the historical drift records and current production line load for each mode, calculate the mode stability index for each mode; When the pattern stability index is lower than the preset stability threshold, the pattern verification frequency is increased; When the pattern stability index is higher than the preset stability threshold, the pattern verification frequency is reduced. Based on the adjusted pattern verification frequency, periodic pattern consistency verification is performed on the included normal sample library.

5. The elevator car defect diagnosis and analysis method based on industrial vision as described in claim 3, characterized in that, The method also includes the following steps: For each normal mode drift warning issued, calculate the potential impact of each warning on product quality. Based on the calculated potential impact, multiple simultaneous normal pattern drift warnings are prioritized. The highest priority normal mode drift warning will be pushed to the manual review queue, along with the corresponding potential impact information.

6. The elevator car defect diagnosis and analysis method based on industrial vision as described in claim 5, characterized in that, The potential impact is quantified by assessing the degree of overlap between the visual features of the pattern and the feature space of known serious defect patterns, as well as the frequency with which the pattern has been associated with product quality problems in historical production.

7. The elevator car defect diagnosis and analysis method based on industrial vision as described in claim 5, characterized in that, The method also includes the following steps: When the highest priority normal mode drift warning is pushed to the manual review queue, the visual alert level of the warning is adjusted according to the corresponding potential impact information. Based on the degree of potential impact, the core data summary related to the early warning is selected and displayed; Based on the core data summary related to the early warning and the adjusted visual alert level of the early warning, an interactive review interface is provided, allowing reviewers to adjust the display order of the early warnings or filter the early warnings, and providing preset review templates.

8. The elevator car defect diagnosis and analysis method based on industrial vision as described in claim 7, characterized in that, The method also includes the following steps: The system analyzes the decision-making content of the reviewer regarding the highest priority normal mode drift warning in the interactive review interface, and generates production line control instructions or quality control instructions. Transmit production line control commands or quality control commands to the production line controller; Receive the execution results and current production status feedback from the production line controller; Based on the execution results and current production status feedback, update the alert status and record the decision results and execution feedback; Based on the execution feedback from the production line controller, assess the potential impact of the decision on product quality after its implementation.

9. The elevator car defect diagnosis and analysis method based on industrial vision as described in claim 8, characterized in that, The steps for assessing the potential impact of a decision on product quality after its implementation include: Obtain various quality index data of elevator car welds; Based on the data of each quality indicator, independent data processing and preliminary evaluation are carried out according to the corresponding evaluation method and data characteristics to obtain preliminary evaluation results of multiple quality indicators. Set the weight parameters for each quality indicator; By comprehensively analyzing the preliminary evaluation results of multiple quality indicators and the weight parameters of each quality indicator, a potential impact index on product quality is obtained. Based on the product quality potential impact index, an overall assessment is conducted on the potential impact on product quality after the implementation of the decision.

10. An elevator car defect diagnosis and analysis system based on industrial vision, applied to the elevator car defect diagnosis and analysis method based on industrial vision as described in claim 1, characterized in that, The system includes: The analysis module acquires a low-resolution image stream of the weld and performs local feature analysis on the low-resolution image stream of the weld to identify potential abnormal areas. The correction module, when a potential abnormal area is identified, triggers the acquisition of a high-resolution local image of the potential abnormal area, and collects the device motion information during the acquisition process. Based on the device motion information, motion correction is performed on the high-resolution local image. The processing module enhances the motion-corrected high-resolution local image and extracts various visual features from the enhanced high-resolution local image. The adjustment module, based on the extracted visual features, performs defect judgment and classification on the local areas corresponding to potential abnormal areas, outputs the defect judgment confidence level, and adaptively adjusts the defect judgment standard according to the material information and process parameter information in the production process. The identification module identifies samples with low defect confidence based on the defect judgment confidence level, pushes the identified samples to the manual review queue to obtain annotations, incrementally updates the adaptively adjusted defect judgment standards based on the manually annotated samples, and identifies new normal process texture patterns and writes the new normal process texture patterns into the normal sample library.

Images