Artificial intelligence driven inspection and detection post skill reconstruction and quality control integrated workbench

Abstract

The application discloses an inspection and detection post skill reconstruction and quality control integrated workbench based on artificial intelligence driving, which comprises a multi-modal data perception and convergence layer, a vertical field detection large model module and an interaction unit. The inspection and detection post skill reconstruction and quality control integrated workbench based on artificial intelligence driving, the multi-modal data perception and convergence layer is used for acquiring multi-source heterogeneous data related to detection tasks and completing structured processing, the vertical field detection large model module is used for generating an inspection scheme or a draft detection report, constructing a business, quality control and talent integrated collaborative system, and solving the pain point of multi-system fragmentation. The application breaks through the whole-link data through the four-layer architecture, realizes the deep integration of detection business automation, quality control real-time and skill dynamic, forms a linkage mechanism of business data driven skill evaluation, quality control performance into ability examination and personnel experience feedback model upgrading, and constructs a sustainable evolution of man-machine collaborative quality management ecology.

Description

Technical Field

[0001] This invention relates to the field of intelligent quality management technology for inspection and testing, specifically to an integrated workbench based on artificial intelligence-driven skills reconstruction and quality control for inspection and testing positions. Background Technology

[0002] With testing volume increasing year by year and industry regulations and quality requirements becoming increasingly stringent, the demand for testing efficiency, result accuracy, and the comprehensive competence of personnel is constantly rising. Currently, the industry's digital transformation is deepening, resulting in three mainstream technical solutions: laboratory information management systems, AI-assisted testing tools, and online training systems. These solutions are widely used in various fields such as manufacturing quality inspection, third-party testing laboratories, and metrology calibration.

[0003] In existing technologies, laboratory information management systems and quality management systems can achieve online control over sample registration, task assignment, data storage, and process approval; AI-assisted equipment such as automated optical inspection can replace some manual processes in the testing execution stage and output single-sample conformity judgment results; online training systems can provide standardized on-demand courses to support personnel in learning basic skills. These solutions play a role in process management, testing execution, and knowledge transfer, respectively, and to a certain extent promote the digital upgrade of the industry.

[0004] However, existing technologies still suffer from significant technical bottlenecks: First, various systems operate independently, with business execution, quality control, and personnel skills development fragmented, failing to create a synergistic improvement cycle. Second, quality control relies primarily on post-inspection and non-conformity tracing, lacking real-time monitoring and pre-emptive warning mechanisms for the testing process, resulting in a significant lag in quality risk prevention and control. Third, personnel skills assessment depends on static manual evaluation, with limited assessment dimensions and a disconnect from real-world business scenarios, failing to meet the job skills transformation needs of human-machine collaborative models. Fourth, massive amounts of historical testing data and personnel operation data are merely archived and stored, failing to be transformed into effective digital assets for quality control optimization, talent development, and technological iteration. Based on these shortcomings, the industry urgently needs an integrated technical solution to achieve synergistic improvement in efficiency, quality, and personnel capabilities across the entire testing process. Summary of the Invention

[0005] The purpose of this invention is to provide an integrated workbench for the reconstruction of inspection and testing job skills and quality control based on artificial intelligence, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This AI-driven integrated workbench for reconstructing inspection and testing job skills and controlling quality includes a multimodal data perception and aggregation layer, a vertical domain inspection model module, and an interaction unit. The multimodal data perception and aggregation layer acquires multi-source heterogeneous data related to inspection tasks and performs structured processing. The vertical domain inspection model module generates inspection plans or draft inspection reports. The interaction unit displays business data and receives user interaction operations.

[0008] It also includes a dynamic skill profile engine, a PBRTQC and SPC integrated quality control module, and the interactive unit is an integrated interactive interface;

[0009] The dynamic skill profile engine is pre-set with a five-dimensional skill evaluation system, including human-machine collaboration capability, AI calibration capability, data analysis capability, risk assessment capability, and cross-domain integration capability. Its data input end is respectively connected to the manual operation log of the vertical domain detection large model module and the abnormal handling record of the PBRTQC and SPC integrated quality control module, which is used to dynamically calculate and update the capability scores of each dimension based on the user's full-process operation behavior.

[0010] The data input terminal of the PBRTQC and SPC fusion quality control module is connected to the real-time operating data and environmental parameters of the multimodal data sensing and aggregation layer, which is used to integrate sample result statistical monitoring and process parameter statistical monitoring to realize real-time abnormal early warning of the detection process;

[0011] The integrated interactive interface is connected to the vertical domain detection large model module, the dynamic skill profile engine, and the PBRTQC and SPC fusion quality control module, respectively, and is used to display business operation content, real-time quality control status and user skill profile on the same screen.

[0012] As a further aspect of the present invention: the multimodal data perception and aggregation layer includes an RPA intelligent document parsing unit, which integrates OCR recognition and large language model semantic understanding to parse multi-format unstructured technical documents and extract detection parameters to convert them into structured data. It has built-in document format verification and knowledge base logic verification sub-units. The multimodal data perception and aggregation layer includes a standard and knowledge base structured interface unit, which adopts a standard digital slicing mechanism to split standard text into five-tuple structured data of detection items, judgment limits, detection methods, applicable scope, and precautions.

[0013] As a further aspect of the present invention: the vertical domain detection large model module adopts a technical architecture combining a base large model with LoRA lightweight fine-tuning and RAG retrieval enhancement. The RAG retrieval enhancement unit connects to the structured standard library and the historical case library. The vertical domain detection large model module is equipped with a confidence scoring unit. The confidence scoring unit calculates the confidence score of the generated content by weighting four dimensions: standard matching degree, case similarity, parameter completeness, and logical consistency. The scores are divided into three confidence levels: high, medium, and low. Low-confidence content is subject to mandatory manual review.

[0014] As a further aspect of the present invention: the dynamic skill profile engine calculates the comprehensive skill score using a weighted statistical algorithm, with the calculation formula S=w1S1+w2S2+w3S3+w4S4+w5S5, where S1-S5 correspond to the scores of the five dimensions, and w1-w5 are the weights of the corresponding dimensions, with initial default weights of 0.25, 0.25, 0.2, 0.2, and 0.1 respectively. The dynamic skill profile engine is configured with a skill score decay mechanism and real-time scoring rules, performing monthly score decay on the dimensions corresponding to task types that have not been used for a long time, and adjusting the score in real time based on the value of user operation behavior.

[0015] As a further aspect of the present invention: the PBRTQC and SPC integrated quality control module is equipped with a three-level early warning mechanism, including general early warning, important early warning and emergency early warning. According to the level of abnormality, it triggers workbench pop-ups, message notifications and responsible person alarms. The module has a built-in large model root cause analysis unit to generate abnormality handling suggestions. The PBRTQC and SPC integrated quality control module is configured with dual-dimensional linkage analysis logic: when PBRTQC identifies a result deviation, it automatically calls the process parameters of the same period SPC to perform correlation analysis to locate the root cause; when SPC identifies a process abnormality but the result has not deviated, it triggers an early warning.

[0016] As a further aspect of the present invention, it also includes a skills reshaping training sandbox module, which comprises a simulated practical training unit, a guided teaching unit, a competency assessment unit, and a growth incentive unit. This module is used to adaptively push training tasks based on the skill deficiencies identified by the dynamic skills profile. Upon passing the assessment, the skills score is updated and the corresponding business permissions are unlocked.

[0017] As a further aspect of the present invention: the integrated interactive interface adopts a three-column single-page layout, with the left side being the AI ​​assistant area for natural language interaction, the middle being the main business operation area, and the right side being a dynamic information area that dynamically switches between skill profiles and quality control dashboards.

[0018] As a further aspect of the present invention, the invention also includes a data security and compliance management module, which comprises a full-process operation tracking unit, a data hierarchical desensitization unit, a fine-grained permission management unit, and an automatic compliance rule update unit.

[0019] An AI-driven method for reconstructing skills and controlling quality in inspection and testing positions, applied to the entire process of inspection and testing operations, includes the following steps:

[0020] Step 1: Receive the detection task, parse the associated multi-source technical documents, extract key detection parameters, and generate structured task data;

[0021] Step 2: Based on the structured task data, combined with the pre-built structured standard library and historical case library, a draft inspection plan or test report is generated through a large language model combined with retrieval enhancement, and low-confidence risk points are marked.

[0022] Step 3: Receive user feedback and modification requests for the initial draft, and record the entire process in a logbook.

[0023] Step 4: Based on the operation logs and task processing quality, update the five-dimensional dynamic skill profile of the corresponding user, identify skill gaps and push personalized training tasks. The five dimensions include human-machine collaboration ability, AI calibration ability, data analysis ability, risk assessment ability, and cross-domain integration ability.

[0024] Step 5: During the detection process, the PBRTQC and SPC algorithms are integrated to monitor the data fluctuations of sample results and process parameters in real time, identify anomalies and trigger graded early warnings, and record the user's abnormal handling behavior;

[0025] Step Six: After de-identifying the confirmed final version of the solution, report, and handling cases, store them in the knowledge base for regular use in large model iteration and optimization.

[0026] As a further aspect of this invention: before step one, a task intelligent assignment step is included: based on the difficulty level, category attributes, and risk level of the inspection task, personnel with corresponding skill levels and ability dimensions are matched for task assignment. The specific process for generating the inspection plan in step two is as follows: identifying product drawings and technical requirement features → retrieving and matching corresponding standard judgment clauses → recalling historical similar cases → generating sampling strategies and inspection methods → estimating measurement uncertainty → calculating the confidence level of each part and marking risk points. The update logic of the dynamic skill profile in step four is as follows: based on the type and frequency of user modification operations, the accuracy and compliance of the modified results, and... The response time and accuracy of anomaly handling are adjusted accordingly to improve the scores of each skill dimension. When a user's score in a certain dimension is consistently below the threshold, corresponding simulated training and teaching content is automatically pushed. The specific logic of real-time quality control in step five is as follows: the moving average and exponentially weighted moving average algorithms are used to monitor the deviation trend of continuous sample test results, and the statistical process control chart is used to monitor the fluctuation trend of equipment operation and environmental parameters. The dual-dimensional data linkage analysis locates the root cause of the anomaly and triggers corresponding warnings according to the anomaly level. In step six, the large model iterative optimization adopts the LoRA lightweight fine-tuning method and updates the model parameters regularly based on the newly added high-quality knowledge base data.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] 1. This invention constructs an integrated collaborative system encompassing business operations, quality control, and talent, overcoming the pain points of fragmented multi-system structures. Addressing the shortcomings of existing technologies where business execution, quality control, and personnel training are independent and unable to form a positive feedback loop, this invention utilizes a four-layer architecture to connect data across the entire chain. This achieves deep integration of automated testing, real-time quality control, and dynamic skills, forming a linkage mechanism where business data drives skills assessment, quality control performance is incorporated into competency evaluation, and personnel experience feeds back into model upgrades. This constructs a sustainable and evolving human-machine collaborative quality management ecosystem.

[0029] 2. This invention enables proactive quality control and enhances the ability to prevent and control quality risks. Addressing the shortcomings of traditional post-inspection methods due to their inherent lag, this invention employs a dual-dimensional quality control system integrating PBRTQC and SPC. Through a sample grouping mechanism, it adapts to multi-category, small-batch scenarios. Combined with the linked analysis of result deviations and process fluctuations, it can identify anomalies and pinpoint root causes within seconds, triggering tiered early warnings and pre-inspection risk alerts. This effectively reduces the probability of missed detections and misjudgments, improving the reliability and compliance of test results.

[0030] 3. This invention enables dynamic quantitative assessment and on-the-job restructuring of personnel skills, shortening the talent development cycle. Addressing the shortcomings of existing static and outdated skills assessments that are disconnected from business scenarios, this invention utilizes a five-dimensional dynamic skills profiling engine to update competency scores in real-time based on actual operational data. Combined with an adaptive training sandbox to precisely push personalized training tasks, it constructs a closed-loop training mechanism of "testing-learning-practice-examination," supporting a smooth transition for personnel to human-machine collaborative roles and significantly shortening the onboarding and competency enhancement cycle for new employees.

[0031] 4. This invention improves the efficiency of core business operations and promotes the reuse of testing knowledge assets. Addressing the shortcomings of reliance on manual labor and low utilization of historical data in core business processes, this invention employs a triple-protection mechanism of "LoRA fine-tuning + RAG retrieval enhancement + standard digital slicing" to intelligently generate compliant testing plans and reports. Simultaneously, confirmed results are de-identified and deposited into the enterprise knowledge base, feeding back into model iterations. This significantly improves business processing efficiency while ensuring compliance, realizing the transformation of personal experience into reusable digital assets for the enterprise. Attached Figure Description

[0032] Figure 1 This is a system architecture diagram of an integrated workbench for the reconstruction of skills and quality control of inspection and testing positions based on artificial intelligence.

[0033] Figure 2 This is a flowchart illustrating the closed-loop process of AI-based testing business and skills feedback within an integrated workbench for skills restructuring and quality control of inspection and testing positions driven by artificial intelligence. Detailed Implementation

[0034] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0035] It should be noted that when a component is referred to as being "fixed to" or "set on" another component, it can be directly on or indirectly on that other component. When a component is referred to as being "connected to" another component, it can be directly connected to or indirectly connected to that other component.

[0036] It should be understood that the terms "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation in the specification, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0037] Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0038] Please see Figures 1-2 In this embodiment of the invention, an AI-driven integrated workbench for the reconstruction of inspection and testing job skills and quality control adopts a four-layer technical architecture of perception convergence, cognitive decision-making, application interaction, and security assurance. It constructs a core logic integrating business automation, real-time quality control, and dynamic skills, achieving integrated management of automated execution of the entire inspection process, real-time process quality control, and dynamic upgrading of personnel skills. The overall system architecture is as follows: Figure 1 As shown, from bottom to top, the architecture consists of a multimodal data perception and aggregation layer, a cognition and decision-making center layer, and a business application and interaction layer. The data security and compliance management module runs through all layers of the architecture, and the integrated interactive interface serves as a unified human-computer interaction entry point to achieve full-process collaborative operation on the same screen.

[0039] The multimodal data perception and aggregation layer serves as the system's data foundation. It communicates with the upper-level cognitive and decision-making central layer via a data bus, uniformly accessing multi-source heterogeneous data across the entire detection business chain. This layer performs data cleaning, normalization, and structured storage, providing standardized input for upper-level models and applications. This layer includes an RPA intelligent document parsing unit (RPA: Robotic Process Automation, which simulates manual operations to complete repetitive processes such as cross-system data capture and document processing), an IoT device interconnection interface unit, and a standard and knowledge base structured interface unit.

[0040] The RPA Intelligent Document Parsing Unit: The data input end connects to external business systems such as ERP, MES, PLM, SRM, and supplier portals via a communication interface. It integrates an OCR text recognition module and a large language model semantic parsing module, capable of parsing unstructured documents such as PDF standards, Word technical requirements, Excel bills of materials, DWG engineering drawings, and COA quality certification documents. The output end connects to the system's data platform, converting extracted inspection items, technical indicators, tolerance limits, and sampling requirements into standardized structured data output. This unit has built-in document format verification and knowledge base logic verification submodules, which automatically mark and reject scanned documents that are blurry, missing pages, or incompletely signed, and automatically correct the parsing results based on a standard knowledge base.

[0041] The IoT device interconnection interface unit supports mainstream industrial protocols such as OPCUA, Modbus, HTTP / HTTPS, and MQTT. Its input end connects to various testing instruments, AI vision inspection equipment, and environmental monitoring sensors to collect raw testing data, equipment operating parameters, ambient temperature and humidity, AI recognition results, and other time-series data. The output end connects to a time-series database, automatically performing data cleaning, outlier removal, and timestamp alignment upon data input. This unit supports the issuance of reverse control commands to the device, automatically triggering operations such as device pause and calibration process activation based on quality control early warning results. It also features a device health assessment submodule that predicts device failure cycles based on runtime time-series data and proactively pushes maintenance reminders.

[0042] The Standards and Knowledge Base Structured Interface Unit: Internally, it interfaces with the enterprise's internal testing specifications, work instructions, and historical case database; externally, it interfaces with regulatory databases such as ISO17025, GB / T industry standards, pharmacopoeias, and metrological verification procedures. It uses a standard digital slicing mechanism to process standard text and outputs five-tuple structured data to the system's knowledge base. This unit sequentially includes a standard text preprocessing submodule, a semantic entity recognition submodule, a relation mapping submodule, a tag system construction submodule, and a version management submodule, which respectively complete non-core content removal, extraction of five types of core entities, establishment of entity relationships, multi-level tag annotation, version difference comparison, and change marking.

[0043] The cognitive and decision-making central layer is the core computing and decision-making layer of the system. It integrates a vertical domain detection large model module, a dynamic skill profiling engine, and a PBRTQC and SPC fusion quality control module (PBRTQC: real-time quality control based on sample results, which monitors the stability of the detection system through statistical analysis of continuous sample data; SPC: statistical process control, which monitors process parameter fluctuations and identifies process anomalies through control charts) through internal data bus. The modules are linked and called together through internal data bus.

[0044] Vertical Domain Detection Large Model Module: It adopts a base large model + LoRA lightweight fine-tuning + RAG retrieval enhancement technology architecture (LoRA: Low-rank adaptation, a large model lightweight fine-tuning technology that injects domain capabilities with a very small number of parameters; RAG: Retrieval Enhancement Generation, which improves the accuracy of large model generation by retrieving knowledge base content in real time and reduces illusions). The input end connects to the structured task data and knowledge base data of the multimodal data perception and aggregation layer, and the output end connects to the corresponding functional units of the business application and interaction layer. The RAG retrieval enhancement subunit connects to a structured standard library, retrieving matching standard clauses and historical cases in real time during content generation, and using the search results as context input to the model. A confidence score unit is configured, which calculates the confidence score of the generated content based on four dimensions: standard matching degree, case similarity, parameter completeness, and logical consistency, with corresponding weights of 40%, 25%, 20%, and 15%, respectively. The scores are divided into three confidence levels: high, medium, and low. A total score below 60 is marked as low confidence and requires mandatory manual review; 60-80 is marked as medium confidence and prompts for focused manual review; and 80 and above is marked as high confidence and can be set to pass without review.

[0045] Dynamic Skill Profile Engine: Inputs are fed into four data sources: business operation logs, task quality results, quality control records, and training assessment data. Outputs are fed into a skills reshaping training sandbox module and a task assignment module. This engine pre-defines a five-dimensional skill evaluation system encompassing human-machine collaboration, AI calibration, data analysis, risk assessment, and cross-domain integration capabilities. A weighted statistical algorithm is used to calculate a comprehensive skill score; the calculation formula is as follows:

[0046] S = w1S1 + w2S2 + w3S3 + w4S4 + w5S5

[0047] S1-S5 correspond to the scores of the five dimensions, and w1-w5 are the weights of the corresponding dimensions. The initial default weights are 0.25, 0.25, 0.2, 0.2, and 0.1 respectively, and can be customized by enterprises. The engine configures a scoring rule submodule and a skill decay submodule, which add or subtract points in real time based on the value of user operation behavior; for task types that have not been used for a long time, the corresponding dimension scores are slightly decayed on a monthly basis. The engine can output skill radar charts and ability level assessment results. The ability levels are divided into four levels: beginner, intermediate, advanced, and expert. The level is automatically associated with the user's business processing permissions and task assignment rules.

[0048] The PBRTQC and SPC integrated quality control module: Its input interface connects to the real-time time-series data from the IoT device interconnection interface unit, and its output interface connects to the integrated interactive interface quality control dashboard and equipment control unit. This module includes a PBRTQC result quality control unit and an SPC process control unit: The PBRTQC result quality control unit uses moving mean, moving range, and exponentially weighted moving average (EWMA) algorithms (EWMA: Exponentially weighted moving average, a statistical control algorithm that assigns decreasing weights to time-series data, more sensitive to small process shifts) to monitor the overall shift and dispersion trend of continuous sample test results in real time; the SPC process control unit uses control charts such as XR charts and XS charts to monitor abnormal fluctuations in equipment operating parameters, environmental parameters, and operational process parameters.

[0049] The module incorporates a fusion analysis engine. When PBRTQC detects a systematic shift in results, it automatically calculates the Pearson correlation coefficient using process parameters from the concurrent SPC, identifying the process factor with the highest correlation as the primary root cause. When SPC detects abnormal process parameters but no shift in results has yet occurred, it triggers a pre-emptive warning. The module employs a three-tiered warning mechanism: a general warning corresponds to a single-dimensional monitoring indicator exceeding the warning line or three consecutive points showing the same trend, triggering a pop-up notification on the workbench; a critical warning corresponds to a single-dimensional indicator exceeding the control limit or five consecutive points showing the same trend, triggering message notifications and alarms from the team leader; and an emergency warning corresponds to multiple indicators simultaneously exceeding tolerances or a critical quality accident risk, triggering alarms from multiple levels of management and automatic equipment shutdown. Upon triggering a warning, the large model is simultaneously invoked to generate preliminary root cause analysis and handling suggestions.

[0050] At the business application and interaction level, scenario-based application functions are provided to users in different positions, including an intelligent inspection solution generator, an AI collaborative report writing module, and a skills reshaping training sandbox module. All functional units are uniformly connected to an integrated interactive interface.

[0051] Intelligent Inspection Solution Generator: The input end interfaces with structured task data and vertical domain testing model modules, supporting the input of sample information, uploading drawings and technical documents, and generating a complete inspection solution with one click; it supports multi-solution comparison, and can simultaneously generate multiple solutions based on different standards and sampling levels; it can automatically associate with equipment calibration status and automatically mark and prompt instruments that have exceeded their calibration validity period. This unit is suitable for various scenarios such as incoming inspection, process inspection, finished product inspection, and third-party commissioned inspection.

[0052] The AI-powered collaborative report writing module: Inputting raw test data, environmental data, and judgment results, it automatically generates a draft test report based on predefined templates. It supports automatic generation of data charts, automatic conclusion determination, and automatic formatting. Natural language interactive editing is supported; users can adjust the report by inputting commands through a dialog box. The module has a built-in compliance verification engine that automatically checks the report's data integrity, standard applicability, signature permissions, and format specifications. Only after successful verification can the report enter the approval process. It supports one-click export in multiple formats and has built-in electronic signatures linked to the approval workflow. After the report is approved, it is automatically stamped with an electronic signature and archived in the LIMS system (LIMS: Laboratory Information Management System, whose core functions are sample registration, task assignment, data storage, and report archiving).

[0053] The Skills Reshaping Training Sandbox module: The input side connects to the skill gap identification results of the dynamic skills profiling engine. It includes simulated practical training units, guided teaching units, competency assessment units, and growth incentive units. It can adaptively push simulated practical training, guided tutorials, and specialized assessment tasks. After completing the assessment, the user's skill score is updated and corresponding business permissions are unlocked. The module also features a team competency dashboard, allowing managers to view the team's overall competency distribution, gap statistics, and growth trends.

[0054] The data security and compliance management module permeates all levels of the system, including a full-process operation tracking unit, a data classification and anonymization unit, a fine-grained access control unit, and an automatic compliance rule update unit. The full-process operation tracking unit immutably stores all user operations, AI-generated records, data modification records, and quality control handling records, supporting end-to-end auditing and traceability. The data classification and anonymization unit automatically anonymizes sensitive data such as core technical parameters and customer information based on data sensitivity levels, removing corporate identifiers and personal privacy information. The fine-grained access control unit implements dual access control based on roles and skill levels, strictly differentiating the data access scope and functional operation permissions for employees at different levels. The automatic compliance rule update unit connects to the regulatory department's standard update database, automatically verifying the compliance of rules within the system and pushing change notifications.

[0055] The integrated interactive interface adopts a single-page, three-column layout: the left side is the AI ​​assistant area, the middle is the main business operation area, and the right side is the dynamic information area. The AI ​​assistant area is a natural language interactive dialog box, supporting business consultation, command operation, Q&A, and knowledge retrieval. The main business operation area is the core function area, automatically switching between pages for scheme editing, report writing, and data viewing based on the current task. The dynamic information area supports dynamic switching between two panels; by default, it displays a personal skill radar chart and growth path, automatically switching to a real-time quality control dashboard when a quality control warning is triggered, displaying monitoring curves, warning information, and handling suggestions. The interface supports customizable shortcut commands, operation rollback, and version restoration, adapting to the usage needs of different roles such as inspectors, quality supervisors, and laboratory directors.

[0056] This invention also provides an integrated method for reconstructing inspection and testing job skills and quality control based on the above-mentioned workbench, the workflow of which is as follows: Figure 2 As shown, the specific steps are as follows:

[0057] Step 1: Intelligent Task Assignment: After receiving a detection task, the system extracts the task's difficulty level, category attributes, and risk level characteristics, and matches the task with personnel possessing the corresponding skill level and ability dimensions to assign the task, thereby achieving intelligent matching between tasks and personnel capabilities.

[0058] Step 2: Task Reception and Multimodal Analysis: The workbench automatically connects with the business system to receive the inspection task, triggering the RPA robot to pull related technical documents, drawings, and standard requirements; it extracts the text and image information of the documents through OCR, and extracts key inspection parameters, technical indicators, and judgment requirements by combining large model semantic analysis; after data verification and standardization alignment, it generates a standardized structured inspection task sheet.

[0059] Step 3: AI Intelligent Draft Generation: Based on structured task parameters, the system calls upon a large-scale detection model in the vertical domain, uses RAG to search and match standard clauses and historical similar cases, and automatically generates a draft inspection plan or inspection report; it calculates the confidence level of each segment of the generated content, marks low-confidence risk points and items requiring manual confirmation, and pushes them to the workbench to-do list.

[0060] The sub-process for generating the inspection plan is as follows: Identify the characteristics of product drawings and technical requirements → Search for and match the corresponding standard judgment clauses → Recall similar historical cases → Generate sampling strategies and testing methods → Estimate measurement uncertainty → Calculate the confidence level of each part and mark risk points (measurement uncertainty: characterizes the dispersion of the reasonably assigned value of the measurand, a parameter associated with the measurement result, used to measure the reliability of the test result).

[0061] Step 4: Human-Machine Collaborative Verification and Operation Traceability: The testing personnel review, modify, supplement, and confirm the AI ​​draft on the workbench; the system records the modification location, modification content, modification duration, confirmation result, remarks, and other operation behaviors throughout the process, forming a complete operation log, while automatically saving multi-version comparison records and outputting the final solution or report.

[0062] Step 5: Real-time updates of dynamic skill profiles: Based on the operation logs, completion quality, time spent, and accuracy data of this task, the backend updates the user's score and level for the corresponding skill dimension according to the preset scoring rules; if a persistent skill deficiency is identified, the corresponding special training task will be automatically triggered.

[0063] The update logic is as follows: based on the type and frequency of user modification operations, the accuracy and compliance of the modified results, as well as the response time and accuracy of exception handling, the weight scores of each skill dimension are adjusted accordingly; when a user's score in a certain dimension is consistently below the threshold, the corresponding special simulation training and teaching content is automatically pushed.

[0064] Step Six: Real-time Quality Control Monitoring and Early Warning Handling: During the testing process, IoT device data flows into the PBRTQC and SPC integrated quality control module in real time. The system performs millisecond-level real-time analysis on sample results and process parameters. After identifying abnormal trends, an early warning is displayed in a pop-up window on the workbench according to the warning level, and preliminary root cause analysis and handling suggestions are pushed. The personnel handling process and results are recorded simultaneously and included in the risk assessment capability assessment.

[0065] The quality control logic is as follows: the moving average and exponentially weighted moving average algorithms are used to monitor the deviation trend of continuous sample test results, and the statistical process control chart is used to monitor the fluctuation trend of equipment operation and environmental parameters; the dual-dimensional data linkage analysis locates the root cause of the abnormality and triggers the corresponding early warning according to the abnormality level.

[0066] Step 7: Knowledge Accumulation and Model Iteration: Final versions of solutions, reports, and anomaly handling cases that have been manually verified are automatically anonymized and stored in the enterprise knowledge base; high-quality data are regularly extracted to build fine-tuning datasets, and LoRA lightweight fine-tuning is used to iteratively optimize large models in vertical domains without the need for full retraining, forming a continuous positive cycle of "manual verification - knowledge accumulation - model upgrade".

[0067] The high-precision parsing mechanism for unstructured documents adopts a three-level parsing mechanism: OCR text recognition + large-scale model semantic understanding + industry knowledge verification. The first level completes the recognition and extraction of text and graphic annotations through OCR. The second level performs semantic understanding and parameter extraction through a domain-specific large-scale model. The third level performs logical verification and error correction through a standard knowledge base. This solves the problem that traditional OCR cannot understand technical semantics and achieves high-precision structured conversion of unstructured technical documents.

[0068] The compliance generation control mechanism for large models adopts a triple guarantee mechanism of standard digital slicing + RAG precise retrieval + confidence level. The standard is broken down into a five-tuple structured data of detection items, limits, methods, scope of application and precautions. The precise standard clauses are used as the generation context through RAG. Combined with the confidence score, low-confidence content is subject to mandatory manual review, which greatly reduces the risk of illusion in professional scenarios of large models and ensures the compliance and accuracy of generated content.

[0069] Dynamic Quantification Mechanism of Personnel Skills: Construct a five-dimensional behavior-driven skills profile system, directly based on the real operation data of employees throughout the entire process, and dynamically score through weighted statistical algorithms combined with scoring rules and attenuation mechanisms, establish a mapping model between operational behavior and ability level, and realize objective, real-time and quantitative assessment of personnel capabilities in the human-machine collaboration mode.

[0070] Dual-dimensional quality control fusion adaptation mechanism: The traditional PBRTQC algorithm introduces a sample grouping mechanism, which automatically groups samples according to the test items, sample types, and equipment numbers. Statistical analysis is performed on continuous samples within the same group to solve the adaptation problem of small batch and multi-category general testing scenarios. Combined with SPC process data, cross-validation and correlation analysis are performed to realize the linkage identification of result deviation and process anomalies, and achieve quality control upgrade of in-process monitoring + pre-event early warning.

[0071] This embodiment is applied to the incoming quality inspection process of automotive parts manufacturing companies, specifically for the incoming inspection of power battery packs delivered by suppliers. The specific implementation process is as follows:

[0072] Task Triggering and Document Parsing: After the purchasing department issues a receiving notice in the SRM system, the workbench automatically retrieves the arrival inspection task for the new batch of battery packs. The RPA intelligent document parsing unit automatically logs into the supplier portal and downloads the COA quality certificate, battery pack engineering drawings, and technical specifications for that batch. The system uses OCR to recognize the dimensions and technical requirements in the drawings, calls the large model to parse and extract 12 key inspection indicators: terminal flatness tolerance ≤ 0.05mm, insulation withstand voltage ≥ 2V / 1min, appearance scratch length ≤ 5mm and depth ≤ 0.1mm, etc. After verification and error correction by the standard knowledge base, a structured inspection requirement sheet is automatically generated.

[0073] Intelligent Solution Generation: Based on testing requirements, the system retrieves three historical inspection plans for similar battery packs from the company's internal "Lithium Battery Pack Incoming Inspection Specifications" via RAG for reference. It automatically generates a complete inspection work instruction document including a sampling plan (GB / T2828.1 General Inspection Level II), testing steps, instrument configuration, judgment criteria, and uncertainty assessment. The appearance defect grading judgment, due to its ambiguous standard boundaries and a confidence score of 58, is marked as a low-confidence item and pushed to the to-do list for manual review. Simultaneously, the system automatically schedules the workshop's AOI (Automated Optical Inspection) vision inspection equipment (AOI: Automated Optical Inspection, equipment for automatically detecting surface defects on workpieces based on optical principles) through the IoT device interconnection interface unit, generating the corresponding scanning path and inspection parameters and sending them to the equipment.

[0074] Human-machine collaborative inspection execution: The inspector reviews the plan through an integrated interactive interface, supplements and refines the appearance judgment clauses, and then confirms and issues the plan. After clicking to start the inspection, the AOI equipment automatically performs a scan and finds a 4.2mm long surface scratch. The AI ​​system initially judges it as a suspected NG (Not Acceptable) with a confidence level of 72 points, and pushes the defect image and result to the workbench's pending confirmation list. Based on product safety requirements, the inspector determines that the scratch is a minor scratch on a non-critical surface, does not affect performance, clicks to release, and notes that it is not on the mounting surface and has no functional impact. The system records the entire modification operation and confirmation result, generating an unalterable operation log.

[0075] Skill profile update and model optimization: The dynamic skill profile engine records the AI ​​false alarm correction operation and adds 2 points to the employee's AI calibration ability dimension. The skill radar chart is updated in real time. At the same time, the defect image and the manual judgment result are processed by the data classification and desensitization unit and added to the training sample set of the AI ​​vision model.

[0076] Real-time quality control and report generation: During the testing process, the IoT device interconnection interface unit collects test data from the withstand voltage tester and workshop temperature and humidity data in real time, transmitting them to the PBRTQC and SPC integrated quality control module. SPC monitors equipment stability, and no abnormalities are detected. After all tests are completed, the AI ​​collaborative report writing module automatically captures all test data, generates a standardized inspection report, and pushes it for review after compliance verification. Upon approval, it is automatically archived to the LIMS system, synchronously updating the supplier's quality profile.

[0077] The application effects of this embodiment are as follows: the processing time for a single batch of battery packs entering the factory for inspection has been reduced from 4 hours to 45 minutes, the time for preparing inspection plans has been reduced from 8 hours to 20 minutes, the average daily batch processing volume per person has increased by 4 times, and the misjudgment rate of appearance defects has decreased from 12% to 4%.

[0078] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-driven integrated workbench for reconstructing inspection and testing job skills and quality control, comprising a multimodal data perception and aggregation layer, a vertical domain inspection large model module, and an interaction unit. The multimodal data perception and aggregation layer is used to acquire multi-source heterogeneous data related to inspection tasks and perform structured processing. The vertical domain inspection large model module is used to generate inspection plans or draft inspection reports. The interaction unit is used to display business data and receive user interaction operations. Its features include: It also includes a dynamic skill profile engine, a PBRTQC and SPC integrated quality control module, and the interactive unit is an integrated interactive interface; The dynamic skill profile engine is pre-set with a five-dimensional skill evaluation system, including human-machine collaboration capability, AI calibration capability, data analysis capability, risk assessment capability, and cross-domain integration capability. Its data input end is respectively connected to the manual operation log of the vertical domain detection large model module and the abnormal handling record of the PBRTQC and SPC integrated quality control module, which is used to dynamically calculate and update the capability scores of each dimension based on the user's full-process operation behavior. The data input terminal of the PBRTQC and SPC fusion quality control module is connected to the real-time operating data and environmental parameters of the multimodal data sensing and aggregation layer, which is used to integrate sample result statistical monitoring and process parameter statistical monitoring to realize real-time abnormal early warning of the detection process; The integrated interactive interface is connected to the vertical domain detection large model module, the dynamic skill profile engine, and the PBRTQC and SPC fusion quality control module, respectively, and is used to display business operation content, real-time quality control status and user skill profile on the same screen.

2. The integrated workbench for skill reconstruction and quality control of inspection and testing positions based on artificial intelligence as described in claim 1, characterized in that: The multimodal data perception and aggregation layer includes an RPA intelligent document parsing unit, which integrates OCR recognition and large language model semantic understanding to parse unstructured technical documents in multiple formats and extract detection parameters to convert them into structured data. It also includes a built-in document format verification and knowledge base logic verification subunit. The multimodal data perception and aggregation layer includes a standard and knowledge base structured interface unit, which uses a standard digital slicing mechanism to split standard text into five-tuple structured data consisting of detection items, judgment limits, detection methods, applicable scope, and precautions.

3. The integrated workbench for skill reconstruction and quality control of inspection and testing positions based on artificial intelligence as described in claim 1, characterized in that: The vertical domain detection large model module adopts a technical architecture that combines a base large model with LoRA lightweight fine-tuning and RAG retrieval enhancement. The RAG retrieval enhancement unit connects to the structured standard library and the historical case library. The vertical domain detection large model module is equipped with a confidence scoring unit. The confidence scoring unit calculates the confidence score of the generated content by weighting four dimensions: standard matching degree, case similarity, parameter completeness, and logical consistency. The scores are divided into three confidence levels: high, medium, and low. Low-confidence content is subject to mandatory manual review.

4. The integrated workbench for skill reconstruction and quality control of inspection and testing positions based on artificial intelligence as described in claim 1, characterized in that: The dynamic skill profile engine calculates a comprehensive skill score using a weighted statistical algorithm. The calculation formula is S=w1S1+w2S2+w3S3+w4S4+w5S5, where S1-S5 correspond to the scores of the five dimensions, and w1-w5 are the weights of the corresponding dimensions. The initial default weights are 0.25, 0.25, 0.2, 0.2, and 0.1, respectively. The dynamic skill profile engine is configured with a skill score decay mechanism and real-time score addition and subtraction rules. Monthly score decay is performed on the dimensions corresponding to task types that have not been used for a long time, and real-time score adjustments are made based on the value of user operation behavior.

5. The integrated workbench for skill reconstruction and quality control of inspection and testing positions based on artificial intelligence as described in claim 1, characterized in that: The PBRTQC and SPC integrated quality control module is equipped with a three-level early warning mechanism, including general early warning, important early warning, and emergency early warning. According to the level of abnormality, it triggers workbench pop-ups, message notifications, and responsible person alarms. The module has a built-in large model root cause analysis unit to generate abnormality handling suggestions. The PBRTQC and SPC integrated quality control module is configured with dual-dimensional linkage analysis logic: when PBRTQC identifies a result deviation, it automatically calls the process parameters of the same period's SPC to perform correlation analysis to locate the root cause; when SPC identifies a process abnormality but the result has not deviated, it triggers an early warning.

6. The integrated workbench for skill reconstruction and quality control of inspection and testing positions based on artificial intelligence as described in claim 1, characterized in that: It also includes a skills reshaping training sandbox module, which contains a simulated practical training unit, a guided teaching unit, a competency assessment unit, and a growth incentive unit. This module is used to adaptively push training tasks based on the skills deficiencies identified by the dynamic skills profile. After passing the assessment, the skills score is updated and the corresponding business permissions are unlocked.

7. The integrated workbench for skill reconstruction and quality control of inspection and testing positions based on artificial intelligence as described in claim 1, characterized in that: The integrated interactive interface adopts a three-column single-page layout, with the left side being the AI ​​assistant area for natural language interaction, the middle being the main business operation area, and the right side being a dynamic information area that dynamically switches between skill profiles and quality control dashboards.

8. The integrated workbench for skill reconstruction and quality control of inspection and testing positions based on artificial intelligence as described in claim 1, characterized in that: It also includes a data security and compliance management module, which includes a full-process operation tracking unit, a data hierarchical desensitization unit, a fine-grained permission management unit, and an automatic compliance rule update unit.

9. A method for reconstructing and controlling the skills and quality of inspection and testing positions based on artificial intelligence, applied to the entire process management of inspection and testing operations, characterized in that... Includes the following steps: Step 1: Receive the detection task, parse the associated multi-source technical documents, extract key detection parameters, and generate structured task data; Step 2: Based on the structured task data, combined with the pre-built structured standard library and historical case library, a draft inspection plan or test report is generated through a large language model combined with retrieval enhancement, and low-confidence risk points are marked. Step 3: Receive user feedback and modification requests for the initial draft, and record the entire process in a logbook. Step 4: Based on the operation logs and task processing quality, update the five-dimensional dynamic skill profile of the corresponding user, identify skill gaps and push personalized training tasks. The five dimensions include human-machine collaboration ability, AI calibration ability, data analysis ability, risk assessment ability, and cross-domain integration ability. Step 5: During the detection process, the PBRTQC and SPC algorithms are integrated to monitor the data fluctuations of sample results and process parameters in real time, identify anomalies and trigger graded early warnings, and record the user's abnormal handling behavior; Step Six: After de-identifying the confirmed final version of the solution, report, and handling cases, store them in the knowledge base for regular use in large model iteration and optimization.

10. The integrated workbench for skill reconstruction and quality control of inspection and testing positions based on artificial intelligence driven according to claim 9, characterized in that: Before step one, there is also a task intelligent assignment step: according to the difficulty level, category attributes and risk level of the testing task, personnel with corresponding skill levels and ability dimensions are matched for task assignment. The specific process of generating the inspection plan in step two is as follows: identify the characteristics of product drawings and technical requirements → search and match the corresponding standard judgment clauses → recall historical similar cases → generate sampling strategies and testing methods → estimate measurement uncertainty → calculate the confidence level of each part and mark risk points. The update logic of the dynamic skill profile in step four is as follows: according to the type and frequency of user modification operations, the accuracy and compliance of the modified results, and the response time and accuracy of abnormal handling, the scores of each skill dimension are adjusted accordingly; when it is identified that the user's score in a certain dimension is continuously lower than the threshold, the corresponding special simulation training and teaching content is automatically pushed. The specific logic of real-time quality control in step five is as follows: the moving average and exponential weighted moving average algorithms are used to monitor the deviation trend of continuous sample test results, and the statistical process control chart is used to monitor the fluctuation trend of equipment operation and environmental parameters. The two-dimensional data linkage analysis locates the root cause of the anomaly and triggers corresponding warnings according to the anomaly level. In step six, the large model iterative optimization adopts the LoRA lightweight fine-tuning method and updates the model parameters regularly based on the newly added high-quality knowledge base data.

Images