SkyVision: APDT Few-Shot Self-Training for Screw Tightening Compliance

DaoAI SkyVision 0-code video surveillance AI platform (on-site hourly training of proprietary models, behavior/event recognition, edge box real-time alerting, 100% local data processing, DaoAI World semantic understanding) leveraged its APDT few-shot self-training capability to elevate a precision manufacturer's screw tightening action compliance from a traditional manual sampling baseline of 92% to over 99.7%, significantly reducing product rework rates and potential quality risks. In industrial SOP/operation specification scenarios, ensuring the normativity and consistency of every critical operation is the cornerstone of product quality and production efficiency. Especially in precision assembly, screw tightening, as one of the most fundamental and critical processes, directly impacts product structural stability, electrical connection reliability, and overall service life. However, traditional manual inspection and sampling methods often struggle to cope with high-beat, multi-station complex production lines.

In industrial manufacturing, particularly in the assembly of precision components, the compliance of screw tightening operations is a critical factor determining product quality. Any insufficient torque, omission, stripped thread, or incorrect operation sequence can lead to product failure during use. Traditionally, companies relied on manual patrols, Standard Operating Procedure (SOP) training, and post-production sampling to ensure operational compliance. However, this method has inherent limitations, especially when dealing with high-beat, multi-station, and complex production lines, where manual monitoring struggles to achieve 100% coverage and is susceptible to factors like fatigue and subjective judgment. As cloud video surveillance systems increasingly demonstrate technical advantages and implementation paths in enhancing enterprise-level security, integrating AI vision technology into production line operation monitoring has become an inevitable trend for improving industrial SOP compliance. DaoAI SkyVision precisely addresses these needs, providing enterprises with efficient, accurate, and intelligent solutions.

Pain Points: Why This Hurdle Is Difficult to Overcome

Monitoring screw tightening action compliance faces multiple challenges. Firstly, in a precision instrument manufacturer's actual production, the compliance rate from manual sampling was only 92%, meaning approximately 8 out of every 100 products might have non-compliant screw tightening, leading to potential product quality risks and high rework costs. Secondly, manually reviewing and correcting these non-compliant operations required an average of 4 additional hours of human labor per day, increasing operational expenses. Furthermore, traditional video surveillance systems could only record for retrospective analysis, unable to identify non-compliant actions in real-time and issue immediate alerts, leading to delayed problem detection and difficulty in immediate correction, resulting in an accumulation of defective products. From a process perspective, subtle differences in screw tightening actions, such as hand posture, tool angle, and tightening duration, can all affect the final quality, and different product models have varying compliance requirements for tightening actions, leading to high model changeover costs. These factors collectively constitute the 'hurdle' of industrial SOP monitoring, urgently requiring an intelligent solution that is real-time, accurate, easy to deploy, and quickly adaptable.

The root cause of these difficulties is that traditional rule-based machine vision systems struggle to capture the dynamic and subtle variations of actions, while deep learning models, despite their powerful capabilities, often require large amounts of annotated data for training. For industrial scenarios with frequent changeovers and varied processes, data collection and annotation costs are prohibitively high. Additionally, traditional solutions typically upload data to the cloud for processing, posing data security risks and limiting real-time performance due to network bandwidth. Addressing the current hot topic of “technical advantages and implementation paths of cloud video surveillance systems in enhancing enterprise-level security,” DaoAI recognizes that localized deployment, edge computing, and few-shot learning are key to overcoming these pain points.

Technical Principles

DaoAI SkyVision's core solution for screw tightening action compliance monitoring lies in its unique APDT (Adaptive Pre-trained Detection & Training) few-shot self-training technology, coupled with real-time alerts from edge boxes and 100% local data processing. APDT technology is based on pre-training with the DaoAI World Model, enabling the model to possess preliminary semantic understanding of general industrial behaviors. In practical applications, users only need to provide a minimal number (1-20) of good or anomalous action samples. The system can then complete rapid self-training and deployment of a proprietary model on-site within hours. For instance, to monitor screw tightening actions, only a few sample images of 'compliant tightening' and 'missed tightening' are needed for SkyVision to quickly learn and recognize these behaviors. This capability significantly lowers the barrier to model development and changeover, reducing the model training cycle, which traditionally takes weeks or even months, to mere hours. This technology achieves an efficient transfer of feature recognition from existing visual foundation models, realizing a paradigm shift from 'big data training' to 'few-shot adaptation,' enabling DaoAI SkyVision to adapt to production line changes with extremely high efficiency.

Compared to traditional manual inspection or fixed-rule-based AOI (Automated Optical Inspection), DaoAI SkyVision's advantages are evident: Firstly, manual inspection is limited by human eye fatigue and subjective judgment, making 100% full inspection difficult and inefficient. Rule-based AOI is sensitive to changes in ambient light, product posture, tool types, etc., requiring extensive manual rule adjustments for new non-compliant actions or product changeovers, which is time-consuming and labor-intensive. In contrast, DaoAI SkyVision, with its behavior/event recognition capability, can precisely capture dynamic operational actions and, through APDT few-shot self-training, quickly adapt to new norms or anomalies without complex rule writing. Furthermore, its edge box real-time alerting mechanism deploys AI inference at the edge of the production line, ensuring millisecond-level response times. Upon detecting a non-compliant action, it immediately triggers an audible and visual alarm or sends notifications, enabling instant problem detection and correction, greatly improving the efficiency of anomaly handling and compliance on the production line. More importantly, DaoAI SkyVision is committed to 100% localized deployment, ensuring all video data and model inference are processed within the client's factory, with data never leaving the premises, fundamentally addressing enterprises' concerns about data security and privacy.

Typical Application Scenarios

• **Screw Tightening Sequence Detection**: In multi-screw fastening scenarios, detecting whether workers follow the preset tightening sequence. The challenge lies in identifying different screw points and tracking the operation path. DaoAI SkyVision identifies the trajectory of the operator's hand or tool at different screw holes, combined with time-series analysis, to determine compliance with the preset sequence.
• **Screw Tightening Tool Usage Compliance**: Monitoring whether workers use the correct tightening tools (e.g., electric screwdrivers, torque wrenches) and if the tool grip posture is standard. The difficulty arises from the variety of tool models and potential hand occlusion affecting recognition. DaoAI SkyVision can identify tool appearance features and operator hand keypoints, combined with the semantic understanding capabilities of the DaoAI World Model, to determine if the tool type and usage posture comply with SOP.
• **Screw Omission/Re-tightening Detection**: In batch screw installation stations, real-time detection of whether any screws are omitted or re-tightened. The challenge is to quickly and accurately determine the state of each screw hole at high production beats. DaoAI SkyVision uses behavior recognition and object tracking technology to continuously monitor the operation status of each screw hole, immediately alerting if an omission or re-tightening occurs.
• **Torque Tool Indicator Light Status Recognition**: For smart torque tools with indicator lights, detecting if the tool's indicator light correctly illuminates after tightening (e.g., green for compliant torque). The difficulty lies in the small size of the indicator light and potential environmental light interference. DaoAI SkyVision combines image enhancement technology and color recognition algorithms to accurately interpret indicator light status, ensuring torque compliance.
• **Poka-Yoke Measure Execution Detection**: In critical processes, detecting whether workers correctly execute Poka-Yoke measures, such as removing a protective cover before tightening. The challenge is that Poka-Yoke steps may involve multiple subtle actions. DaoAI SkyVision identifies a series of continuous action sequences to determine if the Poka-Yoke steps are fully executed.

Case Study

A medium-sized precision electronics manufacturer had long faced challenges with screw tightening process compliance. The factory produced various models of medical-grade power modules, each requiring dozens of screws to be tightened with strict torque and sequence requirements. Previously, they relied primarily on team leader patrols and quality department sampling. However, due to high production beats and staff turnover, the missed detection rate for non-compliant screw tightening was high. In one specific case, the manual sampling compliance rate was only 92%. This directly resulted in approximately 5% of products requiring rework or scrap each quarter, leading to significant production costs and delay risks. After implementing DaoAI SkyVision, the factory deployed edge boxes and industrial cameras at key screw tightening stations. Leveraging SkyVision's APDT few-shot self-training capability, they completed model training for two core abnormal behaviors – 'screw omission' and 'incorrect tightening sequence' – in just 2 hours. Post-deployment, production line data showed that the screw tightening action compliance rate significantly improved to over 99.7%, and the product rework rate decreased by 85%. Simultaneously, because the system could provide real-time alerts and guide workers for immediate correction, manual review time was reduced from 4 hours per day to less than 0.5 hours, freeing up valuable human resources. After its launch, the system successfully built an effective detection model by identifying just 1 anomalous sample of a missed screw, greatly enhancing detection efficiency and accuracy.

DaoAI SkyVision's APDT few-shot self-training capability enabled us to achieve high-precision behavior recognition in an extremely short time, truly realizing refined management of the production process.

DaoAI Solutions and Products

The DaoAI SkyVision platform provides a complete solution for screw tightening action compliance monitoring. Its core advantages include: Firstly, **APDT Few-Shot Self-Training**: Users do not need specialized AI algorithm background. They can simply upload 1-20 good or anomalous sample images using the drag-and-drop interface of SkyVision's 0-code platform, completing proprietary model training and deployment on-site within hours. This rapid iteration capability allows DaoAI SkyVision to easily adapt to multi-variety, small-batch production models and frequent changeover demands. Secondly, **Behavior/Event Recognition**: Based on the powerful semantic understanding capability of the DaoAI World Model, SkyVision not only identifies objects but also understands complex action sequences and events. For screw tightening, it can precisely distinguish compliant operations from various anomalies (e.g., omissions, stripped threads, incorrect sequences, improper tools, etc.). Thirdly, **Edge Box Real-time Alerting**: DaoAI deploys AI inference capabilities in edge boxes at the production line, achieving localized real-time analysis and millisecond-level responses. Upon detecting a non-compliant action, it immediately triggers audible and visual alarms, emails, SMS, or integrates with MES/SCADA systems for instant alerts, ensuring problems are detected and corrected as soon as they occur. Fourthly, **100% Local Data Processing**: All video streams and inference results are processed and stored within the client's factory premises, complying with strict data security and privacy requirements. This deployment method eliminates enterprise concerns about cloud data leaks, ensuring data sovereignty. DaoAI SkyVision offers flexible deployment, supporting various integration methods like SDK / API / Docker, enabling integration with existing production management systems for closed-loop data management. For instance, through API interfaces, non-compliance alerts can be directly pushed to the MES system, triggering work order reviews or quality traceability processes, forming a complete management chain from detection to correction and data analysis.

Through the deployment of DaoAI SkyVision, the precision manufacturing plant's screw tightening action compliance rate increased to over 99.7%, a qualitative leap from the manual sampling baseline of 92%. The product rework rate decreased by 85%, significantly saving production costs. Furthermore, manual review time was reduced by 87.5%, allowing employees to focus on higher-value tasks. In practical application, the system controls the detection time for a single screw tightening action to within 100ms, meeting the demands of high-beat production lines. DaoAI SkyVision not only improved production efficiency and product quality but also provided valuable insights for optimizing operation specifications through real-time monitoring and data analysis, assisting enterprises in their intelligent manufacturing upgrade and transformation.

Full solution for this scenario: SkyVision Video AI industry solutions

This article was generated by AI. Customer cases are simulated scenarios based on real product capabilities and figures are illustrative; see product pages for official benchmarks.