A design risk control review method, system and device

Abstract

The application discloses a design risk control review method, system and device, and belongs to the technical field of business processing. First, the selected area data of the design to be reviewed and the explicit business data input by the user are acquired, multi-modal implicit feature extraction is carried out on the selected area data, the selected area portrait structured data package is generated, and intelligent pre-filling and correction of the explicit business data are completed. A dynamic weight matrix is constructed based on the mapping relationship between business intent and cognitive dimension, a page five-dimensional cognitive radar chart is generated through a dynamic offset algorithm, and a dynamic evaluation standard adaptable to a scene is formed. Then, the interactive elements in the design and the corresponding risk levels are identified, the radar chart, the selected area portrait data and the defect type library are combined, the defect risk score is accurately calculated through a multi-factor product logic, the defect grade is determined and output, and the review feedback data is recorded to update the historical feedback matrix. The application effectively makes up for the defects of traditional review, improves the accuracy and integrity of review data, and reduces the misjudgment and wrong judgment probability of design risk.

Description

Technical Field

[0001] This application relates to the technical field of business processing, and in particular to a design risk control review method, system and equipment. Background Technology

[0002] In the field of UI design for financial products, design risk control review and page walkthrough are core links in controlling product design quality, mitigating business risks, and ensuring compliant operations. Financial businesses are characterized by stringent rules, high sensitivity to funds, standardized compliance requirements, and strong hidden risks. Different business pages have significantly different functional positioning, business intentions, and interaction logic, corresponding to varying risk types and levels. This places extremely high demands on UI design risk control reviews, requiring the review process to be scenario-adaptable, comprehensively identify risks, and accurately quantify defects to effectively avoid business problems such as financial losses and compliance violations caused by design flaws. This is a crucial line of defense for quality control before financial products go live.

[0003] Currently, traditional financial UI design reviews and risk control assessments in the industry heavily rely on fixed, static review rules, which have serious technical limitations. The core problem is the universal review model of "one set of criteria for all pages." Traditional review systems rely solely on single, explicit business data manually entered by users to complete the review judgment. The single source of review data makes it extremely easy for misjudgments to occur due to missing information and data bias, reducing the accuracy and completeness of the review data. Furthermore, it lacks the ability to identify risks specific to financial business scenarios, resulting in a very high probability of missed or incorrect risk assessments. Summary of the Invention

[0004] To improve the accuracy and completeness of the review and reduce the probability of misjudgment, this application provides a method, system and equipment for designing risk control review.

[0005] Firstly, this application provides a design risk control review method, which adopts the following technical solution: A design risk control review method includes: Acquire the selection area data of the design draft to be reviewed, and receive explicit business data input by the user; Multimodal implicit feature extraction is performed on the selected area data of the design draft to generate a structured data package of selected area profiles, and the explicit business data is pre-filled and corrected based on the structured data package of selected area profiles; Based on the corrected explicit business data and dynamic weight matrix, a five-dimensional cognitive radar chart of the current page is generated through a dynamic offset algorithm as a dynamic evaluation standard. The dynamic weight matrix is ​​constructed based on the mapping relationship between business intent and cognitive dimension. The design draft selection area data is scanned to identify interactive elements and corresponding risk levels. Based on the five-dimensional cognitive radar chart, the risk levels of the interactive elements, the structured data package of the selection area profile, and the defect type library, the risk score of the design defect is calculated through multi-factor product logic. The defect level is determined and output based on the risk score, and review feedback data is recorded to update the historical feedback matrix.

[0006] By adopting the above technical solutions, the limitations of traditional reviews relying on single explicit business data are broken. Multimodal implicit feature extraction technology deeply mines the potential information in the selected areas of the design draft. The generated structured data package of selected area profiles not only intelligently pre-fills and corrects explicit business data, but also makes the review basis more aligned with the actual design, avoiding misjudgments caused by incomplete or biased information, and improving the accuracy and completeness of the review data. The dynamic offset algorithm, combined with the business intent-cognitive dimension mapping matrix to generate a five-dimensional cognitive radar chart, is no longer a fixed evaluation standard, but can be flexibly adjusted according to the business intent of the current page, making the evaluation dimensions more targeted and achieving dynamic adaptation of review standards. This effectively solves the problem of insufficient universality of evaluation standards in different design scenarios. The introduction of multi-factor product logic deeply integrates and calculates multi-dimensional information such as the risk level of interactive elements, selected area profile data, and defect type library, which can more accurately quantify the risk score of design defects, avoid bias caused by subjective judgment, and more clearly locate the root cause of risks. The recording of review feedback data and the updating of the historical feedback matrix form a complete closed-loop iterative mechanism. As review cases accumulate, the system's risk identification capability and evaluation accuracy will continue to improve, gradually realizing the self-evolution of review capabilities, reducing business losses caused by design risks as a whole, and improving the operational efficiency and quality control level of the design process.

[0007] Optionally, the step of performing multimodal implicit feature extraction on the selected area data of the design draft to generate a structured data package of the selected area profile includes: Recursively traverse the node tree of the design draft to extract text characters, node types, and size and style metadata; Export the canvas preview image and call the visual big data model to identify page profiles, layout, and key business actions; Monitor the recognition confidence level of the output of the large visual model. When the recognition confidence level is lower than a preset threshold, automatically switch to the local recognition stream based on the design draft node tree for recognition. The extracted metadata and recognition results are cleaned and denoised to generate a structured data package containing page title, key action semantics, business field summary, canvas size and structural signals.

[0008] By employing the aforementioned technical solutions, recursively traversing the node tree of the design draft to extract metadata accurately captures the underlying structural information of the design draft, ensuring comprehensive coverage of basic data such as text, node types, and styles. Furthermore, calling the visual large-scale model to identify page profiles, layouts, and key business actions breaks through the limitations of traditional structured data, uncovering deep business intent and interaction logic at the visual level of the design draft, achieving a multi-dimensional understanding of the design content. The recognition confidence monitoring and local recognition stream switching mechanism effectively ensures the reliability of the recognition results. When the accuracy of the visual large-scale model recognition is insufficient, it automatically switches to local recognition based on the node tree, avoiding data distortion caused by model recognition errors and making the entire recognition process more stable and fault-tolerant. The cleaning, denoising, and aggregation processing of metadata and recognition results integrates scattered multi-source data into standardized selected area profile structured data packages. This not only removes redundant and invalid information but also systematically presents core information such as page titles, key action semantics, and business field summaries, improving the operational efficiency of the entire review process.

[0009] Optionally, the explicit business data includes fixed-dimensional data and deep business driver data; the steps for generating the five-dimensional cognitive radar chart include: Based on the mapping rules defined in the dynamic weight matrix, the options in the corrected fixed-dimensional data and deep business motivation data are discretized into weight vectors on five cognitive dimensions: rigor, certainty, efficiency, clarity, and fault tolerance. The algorithm uses a dynamic offset algorithm to perform weighted summation and offset calculation on the weight vector using the dynamic weight matrix to obtain the comprehensive score of each cognitive dimension. The dynamic weight matrix assigns a base offset weight to the deep business motivation data that is greater than that to the fixed dimension data, so as to ensure that the business intent plays a dominant role in the radar chart offset. The fixed dimension weight summation lays the foundation for the audit, and the deep business motivation weight summation represents the business intent. The comprehensive score is normalized to generate the basic five-dimensional cognitive radar chart of the current page; Obtain the environment multiplier and page type features, apply additional weighting to the review rules corresponding to the mobile terminal based on the environment multiplier, adjust the weights of the corresponding cognitive dimensions based on the page type features, and output the final five-dimensional cognitive radar chart.

[0010] By adopting the above technical solution, the corrected fixed-dimensional data and deep business motivation data are discretized into weight vectors of five cognitive dimensions. This transforms abstract business requirements into quantifiable evaluation indicators, making the review criteria clear and focused on core cognitive dimensions such as rigor and certainty, ensuring that the evaluation direction is highly aligned with business objectives. By setting differentiated weight coefficients, the role of fixed-dimensional data in ensuring the bottom line of the review is preserved, while highlighting the dominant role of deep business motivation in the review criteria. This allows the radar chart to accurately reflect the core business intent, avoiding the problem of review criteria deviating from the actual business. The integration of environment multipliers and page type characteristics further enhances the scenario adaptability of the review criteria. The additional weighting for mobile devices and the adjustment of dimension weights corresponding to page types allow the radar chart to flexibly adjust the evaluation focus according to different usage environments and page functions. For example, in mobile scenarios, the weight of efficiency and fault tolerance is strengthened, while the proportion of rigor and certainty is increased in transaction-type pages, truly realizing the dynamic and personalized nature of the review criteria.

[0011] Optionally, the specific process for identifying interactive elements and their corresponding risk levels is as follows: Scan the interactive components in the node tree of the design draft, and classify buttons, links, input boxes, and text fields into atomic levels. The atomic level is associated with common interactive risks. The amount input component, date picker, and upload control are categorized as components, and the associated logic inputs risk at the component level. The issuance button, endorsement button, risk warning label, and review status indicator are classified as business-sensitive, and the business-sensitive level is associated with asset security risks.

[0012] Optionally, the steps for calculating the risk score of design defects using multi-factor product logic include: Obtain the basic risk value from the defect type library; Extract business impact factors from the structured data package of the selected area profile; The element level factor is determined based on the risk level of the interactive element, and the element level factor increases with the atomic level, component level, and business sensitivity level. The scores of the corresponding dimensions are extracted from the five-dimensional cognitive radar chart to determine the cognitive dimension deviation factor; Obtain the output of the large visual model or the local recognition stream, and determine the recognition confidence factor; The risk score of the design defect is obtained by multiplying the basic risk value, business impact factor, element level factor, cognitive dimension deviation factor, and identification confidence factor.

[0013] By adopting the above technical solutions, basic risk values ​​are obtained from the defect type library, providing a standardized benchmark for risk assessment. This ensures that risk evaluation of different types of design defects has a unified starting point, avoiding evaluation biases caused by differences in defect types. Extracting business impact factors from the structured data package of the selected area profile deeply binds design defects to their actual business impact, extending risk assessment beyond the design itself to the business level. For example, design defects involving core business fields receive higher risk weights, truly reflecting the potential threat of defects to business operations. The element-level factors determined by the risk level of interactive elements further refine the granularity of risk assessment. The incremental factor settings from atomic level, component level to business-sensitive level accurately distinguish the risk degree of defects at different levels of interactive elements, highlighting the risk priority of business-sensitive elements and making risk assessment more aligned with the actual structure and business value of the design. The cognitive dimension deviation factor extracted from the five-dimensional cognitive radar chart enables the linkage between risk assessment and dynamic review standards. When a design defect causes a cognitive dimension to deviate from the standard, the deviation factor amplifies the risk score of that defect, ensuring that the risk assessment closely adheres to the business intent and evaluation requirements of the current page. The introduction of confidence factor adds a layer of reliability assurance to risk assessment. When the confidence of the identification result is low, the weight of the risk score will be appropriately reduced to avoid over-risk judgment due to identification error, making the risk assessment result more credible.

[0014] Optionally, the specific logic for determining the defect level is as follows: Set a first preset threshold and a second preset threshold, wherein the first preset threshold is greater than the second preset threshold; When the risk score exceeds the first preset threshold, the design defect is marked as a P0 blocking level design problem; When the risk score exceeds the second preset threshold but does not exceed the first preset threshold, a manual review is required. Monitor the risk score changes across thresholds. When preset rule conditions are met, trigger the level jump mechanism to adjust the defect priority and indicate the jump conditions in the report.

[0015] Optionally, before the step of generating the five-dimensional cognitive radar chart of the current page using a dynamic offset algorithm based on the corrected explicit business data and dynamic weight matrix, the step of constructing the dynamic weight matrix is ​​also included: Based on the page title, key actions and terminal type in the structured data package of the selected area profile, the corresponding expert skill package is automatically matched and loaded from the expert skill center. The expert skill package includes supply chain invoice skill package, payment risk control skill package and mobile terminal usability skill package. Obtain a pre-set basic matrix, an expert matrix corresponding to the expert skill package, and a historical feedback matrix constructed from past review feedback data. The basic matrix includes standard weights for general audit rules and questionnaires, and the expert matrix includes business rule weights for vertical fields. A dynamic weight matrix is ​​formed by linearly combining the base matrix, expert matrix, and historical feedback matrix using fusion coefficients. The process of generating a five-dimensional cognitive radar chart for the current page using a dynamic offset algorithm specifically involves generating the five-dimensional cognitive radar chart for the current page based on the corrected explicit business data and the dynamic weight matrix.

[0016] By adopting the above technical solution, the core information in the structured data package of the selected area profile is automatically matched and loaded with the corresponding expert skill package. This enables precise matching of professional review needs in vertical fields. Whether it is supply chain invoicing, payment risk control, or mobile usability scenarios, professional rule support tailored to business characteristics can be obtained, avoiding the problem of insufficient adaptation of general review standards in vertical fields, making the review standards more targeted and professional. The integration of the basic matrix, expert matrix, and historical feedback matrix constructs a multi-level, multi-dimensional weighting system. The basic matrix ensures the standard baseline of general audit rules and questionnaires, ensuring that the review does not deviate from basic compliance and quality requirements. The expert matrix injects professional experience in vertical fields, making the review standards fit the actual business needs. The historical feedback matrix incorporates past review experience, realizing the self-iteration and optimization of review standards. As review cases accumulate, the matrix will be continuously revised and improved, making the review standards more in line with actual review scenarios. The dynamic weight matrix, formed by linearly combining fusion coefficients, can flexibly adjust the weight ratio of each matrix according to different business scenarios and review needs. For example, in payment risk control scenarios with extremely high professional requirements, the fusion coefficient of the expert matrix can be increased to highlight the leading role of professional rules. In general page review scenarios, the weight of the basic matrix can be increased to ensure the universality and standardization of the review. The five-dimensional cognitive radar chart generated based on this dynamic weight matrix is ​​a dynamic evaluation benchmark that fully integrates general rules, professional experience, and historical practices, and can accurately match the business characteristics and review needs of the current page.

[0017] Optionally, the steps of recording review feedback data to update the historical feedback matrix include: For each review, the following are recorded: a snapshot of the selected area profile, explicit business data options, a generated five-dimensional cognitive radar chart, identified defects and their corresponding risk scores and levels, and user feedback on the review results. The feedback includes the acceptance, ignoring, or false alarm flags for defects. Based on the type of the operation feedback, the operation feedback is converted into a historical correction vector for the five-dimensional cognitive dimension. The historical correction vector reflects the weight offset direction and offset amount of the current five-dimensional cognitive radar chart. The historical correction vector is aggregated with the current historical feedback matrix to update the historical feedback matrix; The updated historical feedback matrix is ​​used as an input for the linear combination in subsequent reviews to adaptively adjust the dynamic weight matrix.

[0018] By adopting the above technical solution, a complete record of data from the entire review process is achieved, covering the entire chain of information from area profiling and business data to review results and user feedback. This not only provides a rich and comprehensive data source for updating the historical feedback matrix but also provides a complete basis for subsequent review debriefing and issue tracing, ensuring that the experience of each review is fully preserved. Based on the historical correction vector generated from user operation feedback, the deviation between actual business needs and current review standards is captured. Whether it's the adoption, ignoring, or false alarm identification of defects, it can be translated into the direction and magnitude of adjustments to the weights of the five-dimensional cognitive dimensions. This ensures that the correction of review standards is no longer blind but closely aligned with actual business feedback, guaranteeing that the corrected standards better fit real business scenarios. Aggregating the historical correction vector with the current historical feedback matrix achieves the accumulation and sedimentation of review experience. Each feedback correction is integrated into the matrix, allowing it to continuously absorb new needs and experiences from business practice. As the number of review cases increases, the historical feedback matrix will increasingly accurately reflect the actual review requirements of the business. The updated historical feedback matrix then serves as input to the dynamic weight matrix, participating in the standard setting for subsequent reviews. This forms a complete closed loop of "review-feedback-correction-re-review," enabling the design risk control review system to have adaptive evolution capabilities.

[0019] Secondly, this application provides a risk control review system, which adopts the following technical solution: A risk control review system is designed, comprising: The data acquisition module is used to acquire the selection area data of the design draft to be reviewed and to receive explicit business data input by the user. The feature extraction and profile building module is used to perform multimodal implicit feature extraction on the design draft selection area data, generate a selection area profile structured data package, and prefill and correct the explicit business data based on the selection area profile structured data package; The dynamic evaluation standard generation module is used to generate a five-dimensional cognitive radar chart of the current page based on the corrected explicit business data and dynamic weight matrix, using a dynamic offset algorithm as a dynamic evaluation standard. The dynamic weight matrix is ​​constructed based on the mapping relationship between business intent and cognitive dimension. The risk quantification calculation module is used to scan the design draft selection area data to identify interactive elements and corresponding risk levels, and calculate the risk score of design defects through multi-factor product logic based on the five-dimensional cognitive radar chart, the risk level of the interactive elements, the structured data package of the selection area profile, and the defect type library. The defect classification and feedback update module is used to determine the defect level based on the risk score and output it, while recording review feedback data to update the historical feedback matrix.

[0020] Thirdly, this application provides a computer device that adopts the following technical solution: A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the design risk control review method as described in the first aspect.

[0021] In summary, this application includes at least one of the following beneficial technical effects: This application breaks through the technical bottleneck of traditional UI risk control review relying on single explicit data and static fixed criteria. By extracting implicit features from multiple modalities, it mines deep implicit information in the design draft, realizes intelligent pre-filling and correction of business data, and ensures comprehensive and accurate review data. At the same time, it integrates a multi-matrix construction dynamic weight system and combines dynamic offset algorithm to generate a five-dimensional cognitive radar chart adapted to different business scenarios. This solves the audit inaccuracy problem caused by the traditional "one set of criteria for general review". It can accurately identify implicit asset losses and compliance risks in financial scenarios and improve the adaptability of review scenarios and the accuracy of risk identification.

[0022] This application establishes a multi-factor product-based quantitative risk assessment system, which combines multiple factors such as risk levels of interactive elements, profile data, and cognitive dimension bias to accurately quantify defect risks. It abandons the traditional subjective judgment mode, eliminates review bias, and accurately locates the root causes of risks. At the same time, it builds a closed-loop iterative mechanism for full-process review feedback, continuously accumulates review data, iteratively optimizes the review matrix and evaluation standards, and enables the system to have adaptive evolution capabilities. This effectively improves the quality and efficiency of risk control review for financial UI design and consolidates the effect of business risk management. Attached Figure Description

[0023] Figure 1 This is a first flowchart of an embodiment of the method of this application; Figure 2 This is a second flowchart of an embodiment of the method of this application; Figure 3 This is a third flowchart of an embodiment of the method of this application; Figure 4 This is the fourth flowchart of an embodiment of the method of this application; Figure 5 This is the fifth flowchart of an embodiment of the method of this application; Figure 6 This is the sixth flowchart of an embodiment of the method of this application. Detailed Implementation

[0024] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-6 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0025] The first embodiment of this application discloses a design risk control review method. (Refer to...) Figure 1 The method includes S110-S150: S110: Obtain the design draft selection area data to be reviewed, and receive explicit business data input by the user. The explicit business data includes fixed-dimensional data and deep business motivation data. S120: Perform multimodal implicit feature extraction on the selected area data of the design draft, generate a structured data package of the selected area profile, and prefill and correct the explicit business data based on the structured data package of the selected area profile. S130, based on the corrected explicit business data and dynamic weight matrix, generates a five-dimensional cognitive radar chart of the current page through a dynamic offset algorithm as a dynamic evaluation standard. The dynamic weight matrix is ​​constructed based on the mapping relationship between business intent and cognitive dimensions. S140, scan the design draft selection area data to identify interactive elements and corresponding risk levels, and calculate the risk score of design defects through multi-factor product logic based on the five-dimensional cognitive radar chart, the risk level of interactive elements, the structured data package of selection area profiles, and the defect type library. S150 determines and outputs the defect level based on the risk score, while recording review feedback data to update the historical feedback matrix.

[0026] Specifically, the implementation of step S110 involves the system relying on the Figma plugin's front-end interactive module to capture in real-time the selected area data of the design draft awaiting review currently being operated by the designer. This locks the range of the user-selected canvas, page, and corresponding design nodes. Simultaneously, an interactive business data collection interface is built to receive explicit business data manually entered by the user. This explicit business data is strictly divided into two modules: fixed-dimensional data and deep business motivation data. The fixed-dimensional data corresponds to the six core dimensions D1-D6, specifically including design stage, interaction type, operation frequency, business impact scope, page information density, and terminal type. All dimensions are equipped with standardized optional parameters and numerical ranges to ensure the standardization of entered data. The in-depth business motivation data corresponds to the five core dimensions of Q1-Q5, focusing on five in-depth business attributes: product compliance goals, core user operation paths, business loss risk points, core business demands, and scenario risk levels. The system will uniquely bind the collected selected area data with two types of explicit business data, and associate it with the current design engineering document number and review batch, ensuring that each set of review data corresponds to an independent design review scenario. This ensures the accuracy and uniqueness of data matching from the source, providing basic data support for subsequent weight calculation and risk assessment.

[0027] Reference Figure 2 In S120, the steps of extracting multimodal implicit features from the selected area data of the design draft and generating a structured data package of the selected area portrait include S210-S240: S210, recursively traverse the design draft node tree to extract text characters, node types, and size and style metadata; S220: Export canvas preview image, call visual big model to recognize page profile, layout and key business actions; S230 monitors the recognition confidence level of the visual large model output. When the recognition confidence level is lower than the preset threshold, it automatically switches to the local recognition stream based on the design draft node tree for recognition. S240 performs character cleaning and noise reduction on the extracted metadata and recognition results, and aggregates them to generate a selected area portrait structured data package containing page title, key action semantics, business field summary, canvas size and structural signals.

[0028] Specifically, the S120 step is implemented by calling the dedicated Figma plugin application interface. Unlike the traditional browser DOM traversal mode, it adopts a deep recursive traversal algorithm to traverse the root node, parent node, child node, and nested secondary nodes of the original node tree of the design draft from top to bottom. It captures the core metadata of all valid design nodes without omission. The extracted content covers all text characters on the page, the unique type tags of each node, the pixel size of the node, the color scheme, the margin layout, the hierarchical nesting relationship, and other full-dimensional physical attribute data. At the same time, invalid nodes are initially filtered during the traversal process, and redundant nodes that do not participate in business interaction and information display, such as blank placeholder nodes, purely decorative icon nodes, and transparent invalid layers, are automatically removed. Only valid node data that carries business information, interactive functions, and risk control display is retained. This simplifies the data volume from the bottom layer, improves the computational efficiency of subsequent feature extraction and recognition, and ensures the effectiveness and relevance of the metadata.

[0029] The system uses Figma's native rendering interface to perform lossless high-definition rendering of the current canvas to be reviewed, and exports a high-definition PNG format canvas preview image that conforms to the visual large model input standard, avoiding the interference of image blur and proportion distortion on visual recognition.

[0030] Subsequently, the system calls the inference interface of a dedicated multimodal visual large model (such as GPT-4V or a visual language large model of the same level) adapted to financial design risk control scenarios. The standardized preview image is input into the model, and pre-built structured prompts specific to financial UI review are injected to guide the model to accurately complete multi-dimensional recognition work. The prompt template defines the model's role boundaries, recognition task list and output format constraints. Specifically, it sets the model role as "financial risk control design review expert", instructs it to perform three core tasks: overall business profile classification of the page, page layout structure judgment and semantic recognition of key business actions of the page, and forces the model to accurately distinguish financial core business actions such as bill issuance, endorsement, acceptance, fund payment, business review and risk withdrawal, while outputting the confidence assessment of the corresponding recognition results.

[0031] After performing multimodal inference based on images and prompts, the model returns results in a pre-defined JSON structured format. This includes image classification labels, layout type, a list of key actions and their coordinate regions, and confidence scores for each recognition item (uniformly locked between 0.6 and 1.0), providing accurate judgment criteria for the subsequent dual-recognition stream switching mechanism. Simultaneously, the system uses the postMessage communication protocol at its core to achieve asynchronous synchronization of structured data between the Figma main thread and the UI plugin thread. This ensures that node tree metadata and canvas rendering image data obtained through recursive traversal can be efficiently and non-blockingly transmitted across threads to the review plugin's computation module.

[0032] The system presets a fixed recognition confidence threshold of 0.8. After the large visual model completes image recognition, it captures and monitors the confidence value output by the model in real time. When the confidence value is detected to be lower than the preset threshold of 0.8, the system automatically determines that the current visual recognition flow is affected by factors such as complex page layout, element occlusion, special color scheme, and niche control styles, which may lead to inaccurate recognition, missed detection, or false detection. It then triggers a multimodal redundancy fallback mechanism, automatically switching to a local recognition flow based on the design draft node tree without manual intervention or awareness. Relying on the complete node tree metadata obtained by recursive traversal, the local rule engine completes page structure parsing, interactive element positioning, business action recognition, and risk node screening. The entire recognition operation is completed using local computing power, without relying on AI models or network environment, effectively ensuring the availability and stability of the review process in extreme scenarios and achieving intelligent complementarity between the two recognition flows.

[0033] The system first standardizes and cleans the raw metadata extracted earlier using a rule-based character cleaning and denoising algorithm. Specifically, in terms of text cleaning, the system uses regular expressions to match and remove formatting errors such as newline characters, tabs, and residual HTML tags. It also traverses the node tree's visibility attributes to automatically remove redundant notes and placeholders (such as "Lorem ipsum") within hidden layers. In terms of structural denoising, the system reads the size and coordinate parameters of nodes, removes canvas overflow nodes and transparent masking nodes with zero area, and filters invisible watermark characters with a font size smaller than 8px based on financial UI design standards. It also unifies the encoding and numerical format of all valid metadata, thus completing the data denoising optimization.

[0034] Subsequently, the cleaned and standardized metadata is aggregated in a structured manner to generate a dedicated SelectionProfile structured data package. This data package serves as the underlying core data foundation of the entire review system, fully containing core structured information such as the semantic parsing results of standardized page titles, semantic tags of core key actions, summaries of core business fields such as amount, invoice number, and account information, standard pixel size of the canvas, count of form list controls, and page hierarchy signals.

[0035] Ultimately, the system relies on the objective physical and business characteristics of the selected area profile to intelligently prefill and accurately correct the explicit business data entered by the user in the early stage. For example, it automatically distinguishes between mobile and PC terminal types by canvas pixel size and prefills D6 terminal dimension data, automatically corrects business asset loss risk level parameters by core business fields on the page, and calibrates operation frequency data by density of page interactive controls, thereby reducing subjective errors in manual input and realizing the automatic conversion of physical attributes of design drafts into business layer review signals.

[0036] Reference Figure 3Before generating the five-dimensional cognitive radar chart, step S130 includes the construction of the dynamic weight matrix, steps S310-S330. S310 automatically matches and loads the corresponding expert skill packages from the expert skill center based on the page title, key actions and terminal type in the selected area profile structured data package. The expert skill packages include supply chain invoice skill package, payment risk control skill package and mobile terminal usability skill package. S320: Obtain the pre-set basic matrix, the expert matrix corresponding to the expert skill package, and the historical feedback matrix constructed from past review feedback data. The basic matrix contains the standard weights of general audit rules and questionnaires, and the expert matrix contains the business rule weights of the vertical field. S330 uses fusion coefficients to linearly combine the base matrix, expert matrix, and historical feedback matrix to form a dynamic weight matrix.

[0037] Specifically, the implementation of step S130 involves the system pre-building a modular, pluggable expert skill center, which includes three standardized expert skill packages for three vertical fields: supply chain bill skill package, payment risk control skill package, and mobile usability skill package. Each skill package deeply encapsulates the corresponding field's financial regulatory compliance rules, business risk standards, design risk control priorities, and exclusive review weight rules, forming an independent vertical field review capability system.

[0038] During the matching and loading phase, the system extracts features such as page title, key business actions, terminal type, and core business field attributes from the structured data package of the selected area generated by S120 to construct the scene feature vector of the current page. It then calculates the cosine similarity with the domain feature vector sets preset by each skill package in the expert skill center. When the similarity exceeds the preset matching threshold, it is determined that the automatic matching is successful and the corresponding one or more expert skill packages are loaded. For example, when the page is identified to contain core elements such as bill issuance, endorsement, and acceptance, the supply chain bill skill package is automatically loaded, focusing on the review of high-risk design defects related to the transfer of bill ownership. When the page is identified to contain payment amount, payment and receipt accounts, and fund confirmation operations, the payment risk control skill package is automatically loaded, focusing on the verification and review of financial transaction-related asset loss risks. When the mobile terminal canvas size and touch-based interactive layout are identified, the mobile terminal usability skill package is automatically loaded, conducting special reviews on dimensions such as mobile terminal operation adaptation and touch security. It supports the superposition and empowerment of multiple skill packages, realizing the deep integration of general review capabilities and vertical domain risk control capabilities.

[0039] Subsequently, the system synchronously retrieves three types of core weight matrix data from the local persistent database: the Mbase basic matrix is ​​the system's initial pre-set standardized benchmark matrix, built on a standard questionnaire system based on general UX audit rules and fixed dimensions D1-D6 and deep business drivers Q1-Q5. Each questionnaire answer option in the matrix corresponds to a standardized weight vector of five cognitive dimensions: rigor, certainty, efficiency, clarity, and fault tolerance, forming the basic weight background for all review scenarios; the Mexpert expert matrix is ​​a vertical domain-specific matrix corresponding to the loaded expert skill packages, embedding exclusive compliance requirements, business risk rules, and industry design specifications for supply chain invoices, payment risk control, and mobile scenarios, supplementing the general review rules with differentiated weight offset parameters for vertical scenarios; the Mhistory historical feedback matrix is ​​a dynamic matrix built by the system in real time based on user operation feedback data from all previous review batches, accumulating experience in correcting weight deviations in historical reviews.

[0040] After obtaining the three types of matrices mentioned above, the system uses a linear combination algorithm for weighted fusion. The construction of the dynamic weight matrix strictly follows a progressive evolutionary logic: in the initial stage of the system, before the expert skill pack is loaded and there is no historical data, the system mainly relies on the basic matrix. The evolution formula is Mdynamic≈Mbase, ensuring the output of the system's basic and general review capabilities. When a user enables an expert skill pack but historical data is insufficient, the system introduces an expert matrix to enhance vertical domain capabilities. The fusion coefficients are dynamically allocated based on the domain confidence of the expert package (default). The evolution formula is: Once the system has accumulated sufficient historical review feedback data, a historical feedback matrix is ​​introduced for adaptive correction. The three sets of fusion coefficients are dynamically allocated based on the historical data size and variance (default). The evolution formula is: .

[0041] During the fusion mapping process, the system strengthens the internalized connection between the Security rigor dimension and financial regulatory compliance requirements by setting differentiated base values. In the initial construction of the Mbase basic matrix, when mapping compliance rules such as anonymization requirements and strong risk warnings to the weights of the Security dimension, the initial weight base value is set to be 20% higher than that of the other four dimensions (certainty, efficiency, clarity, and fault tolerance). This ensures that the financial compliance bottom line occupies an absolute numerical dominance in the radar chart evaluation standard and subsequent risk calculation, completely abandoning the traditional fixed static review criteria and realizing the adaptive optimization of the dynamic weight matrix according to business scenarios, vertical fields, and historical review results.

[0042] Reference Figure 4S130, The steps for generating a five-dimensional cognitive radar map include S410-S440: S410, based on the mapping rules defined in the dynamic weight matrix, discretizes the options in the corrected fixed-dimensional data and deep business motivation data into weight vectors on five cognitive dimensions: rigor, certainty, efficiency, clarity and fault tolerance. S420 uses a dynamic offset algorithm to perform weighted summation and offset calculation on the weight vector using a dynamic weight matrix to obtain a comprehensive score for each cognitive dimension. The dynamic weight matrix assigns a greater basic offset weight to the deep business motivation data than to the fixed dimension data to ensure that business intent plays a dominant role in the radar chart offset. The fixed dimension weight summation lays the foundation for the audit, while the deep business motivation weight summation represents the business intent. S430, normalizes the comprehensive score and generates the basic five-dimensional cognitive radar chart for the current page; S440: Obtain the environment multiplier and page type features; apply additional weighting to the review rules for mobile devices based on the environment multiplier; adjust the weights of the corresponding cognitive dimensions based on the page type features; and output the final five-dimensional cognitive radar chart.

[0043] Specifically, relying on the standardized mapping rules preset in the dynamic weight matrix, the system quantifies and discretizes all user selections in the D1-D6 fixed-dimensional data and Q1-Q5 deep business motivation data after the selected area profile data has been pre-filled and corrected. The system accurately decomposes and maps the business attributes, scenario characteristics, and risk levels corresponding to each dimension data into independent standardized weight vectors on five cognitive dimensions: rigor, certainty, efficiency, clarity, and fault tolerance. All weight vectors are uniformly standardized to a standard numerical range, ensuring that the vectors of each dimension are calculable and comparable. This provides a standardized vector data source for subsequent dynamic weighted aggregation and weight offset calculation. The entire discretization process is executed entirely based on the scenario adaptation rules of the dynamic weight matrix. Unlike the fixed mapping mode, it can be synchronously adapted and adjusted with the iterative update of the dynamic weight matrix, ensuring the scenario accuracy of the vector mapping.

[0044] The system is based on a dynamic offset algorithm. It calls the generated dynamic weight matrix to perform global weighted summation and dimensional offset calculation on all discretized weight vectors. It then uses a standard dimensional weighted normalization formula to calculate the comprehensive score for each cognitive dimension. The core formula is... ,in This represents the weight vector of the i-th fixed-dimensional data in the k-th cognitive dimension. This represents the weight vector of the j-th deep business motivation data in the k-th cognitive dimension. The number of fixed-dimensional questions that can be effectively included in the calculation. The number of deep business motivation questions to be effectively included in the calculation.

[0045] The system has preset fixed coefficients. This ratio is based on risk preference criteria within a financial risk control scenario: since the fixed dimension (Question D) only represents the objective context of the design stage and terminal type, its impact on risk offset exhibits a linear basic characteristic, hence it is assigned a weight of 1.0; while the deep business motivation (Question Q) is directly related to compliance objectives and risk sensitivity, and the harm caused by its trigger defects exhibits an exponential amplification characteristic, requiring it to dominate the radar chart offset. Verified through fitting with historical financial defect data, a 1.5 multiplier is the critical empirical threshold for ensuring that business intent dominates the offset while preventing the complete failure of general rules. This setting ensures the dominant role of business intent in radar chart dimension offsets from the algorithm's underlying layer. The weighted aggregation result of the fixed dimension data is used to lay the foundation for the basic audit context of the page review, while the weighted aggregation result of the deep business motivation data is used to accurately represent the core business risk intent of the page, ultimately completing the accurate calculation of the original comprehensive score for the five cognitive dimensions.

[0046] The system uses a nested Normalize algorithm within the formula to perform full-domain standardization calibration on the raw comprehensive scores of each cognitive dimension. Using the weighted sum of coefficients of fixed dimensions and deep business drivers as the baseline, it eliminates numerical deviations caused by differences in the number of options and weight bases of different page dimensions. The comprehensive scores of the five cognitive dimensions are uniformly converted into standardized scores in the 0-1 range. Then, using the standardized scores of each dimension as the vertex coordinates of the radar chart, a basic five-dimensional cognitive radar chart adapted to the current page's business scenario is generated. This radar chart is generated entirely based on the adaptive offset of the dynamic weight matrix. Each page has its own exclusive review criteria, completely solving the audit inaccuracy problem caused by the traditional static review rules of "one set of criteria for all pages".

[0047] The system introduces a standardized environment multiplier as a global correction factor, with a uniform value range of 1.0-1.3. First, it uses the terminal type identified by the selected area profile to complete the adaptive weighting of the environment multiplier. For mobile design drafts, the system automatically enables the environment multiplier weighting mechanism, directly multiplying the original weight values ​​of the cognitive dimensions corresponding to mobile-specific review rules such as mobile touch hotspot size, page font size specifications, and bottom main operation button layout by the environment multiplier (default value is 1.2), thereby achieving a quantitative upward correction of the weights and strengthening the review of mobile usability and operational security.

[0048] Meanwhile, the system accurately identifies page types by analyzing page title semantics, business field density, number of interactive controls, and node structure characteristics. It distinguishes between different scenarios such as bill business pages, fund payment pages, audit management pages, and information display pages. Based on the identified page type characteristics, the system adjusts the weights of the corresponding cognitive dimensions. For high-risk funds and bill core business pages, a weight increment of 0.2 is added to the original scores of the rigor and fault tolerance dimensions before normalization. For high-frequency daily operation pages, a weight increment of 0.2 is added to the efficiency and certainty dimensions. After completing the dual quantitative weight adjustment based on the terminal environment and page scenario, the final five-dimensional cognitive radar chart is output, serving as the sole dynamic and authoritative evaluation standard for this design risk control review.

[0049] In S140, the specific process for identifying interactive elements and their corresponding risk levels is as follows: Scan the interactive components in the node tree of the design draft, and classify buttons, links, input boxes, and text fields into atomic levels, and associate common interactive risks at the atomic level; The amount input component, date picker, and upload control are categorized as component-level components, and the associated logic input at the component level poses a risk. The issuance button, endorsement button, risk warning label, and review status indicator are classified as business-sensitive, and the business-sensitive level is associated with asset security risks.

[0050] Specifically, the implementation of step S140 is as follows: the system relies on the node tree traversal results to complete a fully automatic and refined hierarchical classification of all interactive components in the design draft, and binds them to the corresponding level of business risk type to form a standardized interactive element risk system.

[0051] Buttons, text links, ordinary input boxes, and text fields are collectively classified as atomic-level interactive elements. They belong to the basic and general interactive controls of the page and are only associated with general interactive risks. The corresponding design flaws can easily lead to minor user errors such as accidental clicks, blank inputs, and invalid clicks. Amount-specific input components, calendar date pickers, and file upload controls are collectively categorized as component-level interactive elements. These are dedicated controls for business logic and are associated with input risks. Design flaws in these components can easily lead to medium-level business logic problems such as data format validation failures, amount entry errors, file upload anomalies, and date matching deviations. The bill issuance button, business endorsement button, risk warning label, and review status indicator control are uniformly classified as business-sensitive interactive elements. These are core financial risk control controls that are directly related to the security risks of funds and bill assets. Corresponding design flaws can directly induce high-risk asset losses and compliance issues such as incorrect fund transfers, abnormal bill ownership, release of funds in violation of regulations, and lack of risk warnings. The three levels of risk increase in severity, providing a clear classification basis for the subsequent differentiated assignment of risk factors.

[0052] Reference Figure 5 In S140, the steps for calculating the risk score of design defects using multi-factor product logic include S510-S560: S510, Obtain the basic risk value from the defect type library; S520 extracts business impact factors from the structured data package of the selected area profile; S530: Determine the element level factor based on the risk level of the interactive element. The element level factor increases with the atomic level, component level, and business sensitivity level. S540: Extract the scores of the corresponding dimensions from the five-dimensional cognitive radar chart to determine the cognitive dimension deviation factor; S550: Obtain the output of the large visual model or local recognition stream, and determine the recognition confidence factor; S560 multiplies the basic risk value, business impact factor, element level factor, cognitive dimension deviation factor, and identification confidence factor to obtain the risk score of the design defect.

[0053] Specifically, the system has a built-in library of comprehensive financial UI design defect types, which includes design defects across all categories such as text display, interactive layout, operation guidance, status feedback, exception handling, risk control prompts, and compliance display.

[0054] All defect types are pre-set with a fixed, standardized base risk value (Base), with a uniform range of 20-70. This range is derived from a simplified model of the Financial Factors Analysis (FMEA) model, commonly used in the financial industry: design defects are divided into three levels based on their business impact: low, medium, and high. 20-35 corresponds to minor display defects (such as misplaced layouts or non-standard color schemes), which have an extremely low probability of causing financial losses and minimal impact. 36-55 corresponds to routine operational defects (such as unreasonable interactive layouts or unclear feedback in normal states), which have a moderate probability of causing operational disruptions and some local business impact. 56-70 corresponds to severe risk control defects (such as missing risk control warnings, lack of protection for high-risk operations, or incorrect display of core business information), which have an extremely high probability of directly inducing erroneous fund transfers or compliance violations and are potentially fatal. The upper limit of 70 is designed to leave sufficient room for the combined amplification of business impact factors and element-level factors in the subsequent multiplication model, preventing the risk score from prematurely saturating and overflowing in extreme scenarios.

[0055] After identifying a specific design defect, the system accurately matches the defect features to the type library and uniquely retrieves the corresponding basic risk value Base. This basic risk value is a preset initial risk score, which serves as the core benchmark value for subsequent risk score calculations.

[0056] The system extracts core business attribute data from the generated selected area profile structured data package, comprehensively determines and generates a standardized business impact factor Bimpact, with a value range uniformly locked at 1.0-2.0. The factor value fluctuates dynamically based entirely on the business risk level of the page. The system comprehensively calculates the business impact of page operations by analyzing multiple dimensions, including the sensitivity of key business actions to financial losses, the compliance level of core business fields, the risk weight of business scenarios, and the business impact of page operations. The business impact factor for ordinary information display and static content pages is 1.0-1.2, the business impact factor for routine business review and ordinary information entry pages is 1.3-1.6, and the business impact factor for highly sensitive and high-risk pages such as bill issuance, fund payment, and ownership change is 1.7-2.0. This accurately quantifies the potential business harm caused by design flaws in different business scenarios.

[0057] The system sets a gradient of fixed element-level factors, Ecrit, based on the three-tiered risk levels of interactive elements. This ensures precise matching between risk severity and weighting coefficients. The three levels correspond to fixed standard values ​​with no fluctuation deviation. Specifically, the atomic-level general interactive elements have a factor value of 1.0, indicating that defects in this level of element do not incur additional risk and only retain the basic risk level. The component-level logical interactive elements have a factor value of 1.2, assigning a moderate level of risk, highlighting the business harm of logical input defects. The business-sensitive core risk control elements have a factor value of 1.5, assigning the highest level of risk, emphasizing the risk weight of design defects in core controls related to funds and bills. Through differentiated factor settings, the defects of high-risk business elements are highlighted in risk quantification, aligning with the core needs of financial risk control.

[0058] The system will accurately match the identified design defects with the final generated five-dimensional cognitive radar chart, clarify the core review dimension corresponding to each defect, extract the standardized score Scorek of the dimension, and calculate the dimension deviation based on the preset benchmark threshold, and finally generate the cognitive dimension deviation factor Δdim, with a value range of 1.0-1.5.

[0059] The specific calculation logic is as follows: The system presets a baseline safety threshold of 0.6 for the five-dimensional cognitive dimensions. It uses the linear interpolation formula Δdim = 1.0 + (Scorek - 0.6) × (0.5 / 0.4) for calculation. The higher the radar chart score (Scorek) of the dimension containing the defect, the greater the attention this dimension receives in the current page's weight shift, indicating a significant deviation in the defect's compliance with the dimension's requirements. The calculated Δdim approaches the upper limit of 1.5. When Scorek is below the baseline threshold of 0.6, it indicates that this dimension is not a core review focus of the current page, and the defect's deviation is low. The calculated Δdim approaches the lower limit of 1.0. This quantitative interpolation formula achieves a deep binding between defect risk and the core review focus of the page, making the risk quantification results more closely aligned with actual scenario needs.

[0060] The system automatically identifies the recognition operation mode used in this review process and determines the standardized recognition confidence factor Cconf based on this, with a uniform value range of 0.6-1.0. If the current process is a large visual model recognition stream, the system directly retrieves the accurate confidence value output by the model in real time as the factor value, truly reflecting the reliability of the AI ​​recognition results. If the current process is a local node tree recognition stream triggered by low confidence fallback, the system uniformly sets the local recognition confidence value to a fixed value of 0.9, assuming that the local rule recognition results are stable and reliable, effectively avoiding the interference of recognition accuracy fluctuations on risk score calculation, and ensuring the stability and accuracy of risk quantification results.

[0061] The system uses a multi-factor product model specific to financial design risk control to calculate the defect risk score. It performs a step-by-step product operation on five core parameters: the base risk value (Base), the business impact factor (Bimpact), the element level factor (Ecrit), the cognitive dimension deviation factor (Δdim), and the identification confidence factor (Cconf). Base is the initial risk score, and the other factors are dimensionless weighted correction coefficients. Finally, a unique standardized design defect risk score, RiskScore, is calculated.

[0062] Meanwhile, the system incorporates a risk mapping link from design flaws to business losses, mapping physical defects at the design level to deviations in business actions: for cognitive layer risks, displaying amounts without a thousandths place or with an excessively small font size is mapped to direct losses due to user misinterpretation leading to incorrect payments or endorsements; for feedback layer risks, vague submission status feedback is mapped to operational losses due to duplicate submissions, duplicate acceptances, or duplicate reviews; for fault tolerance layer risks, the lack of withdrawal or secondary confirmation for high-risk operations is mapped to compliance losses due to the inability to intercept erroneous instructions in a timely manner; thus, a safety baseline for business losses is established at the design stage, achieving precise quantitative transformation and logical mapping from design flaws to business loss risks.

[0063] In S150, the specific logic for determining the defect level is as follows: Set a first preset threshold and a second preset threshold, wherein the first preset threshold is greater than the second preset threshold; When the risk score exceeds the first preset threshold, the design defect is marked as a P0 blocking level design problem; when the risk score exceeds the second preset threshold but does not exceed the first preset threshold, a manual review is required; monitor the cross-threshold changes of the risk score, and when the preset rule conditions are met, trigger the level jump mechanism to adjust the defect priority and indicate the jump conditions in the report.

[0064] Specifically, the implementation of step S150 is as follows: the system combines the risk control grading standards designed for the financial industry with the boundary extrapolation of the multi-factor product model, presets two levels of dynamically adjustable risk thresholds, and strictly follows the hierarchical logic that the first preset threshold is greater than the second preset threshold. The preferred range for the first preset threshold is 60-80, and the preferred range for the second preset threshold is 35-50.

[0065] The derivation logic of this threshold is as follows: Taking a severe risk control defect as an example, when Base is the highest value of 70, if the interactive element is not business sensitive (Ecrit=1.0), has a moderate business impact (Bimpact=1.0), and has a moderate confidence level (Cconf=1.0), the base score of 70 is already an absolute high risk that must be blocked. Therefore, the upper limit of the first preset threshold is set at around 70. Taking a medium operational defect as an example, when Base is the median value of 45, after adding business sensitive level (Ecrit=1.5) and high business impact (Bimpact=1.6), the score will jump to 108, which will definitely exceed the blocking line. However, if only slight impact (Bimpact=1.0) is added, the score remains at 45, which is a critical state that needs attention but may not necessarily be blocked. Therefore, the lower limit of the second preset threshold is set at around 45.

[0066] In the current preferred embodiment, based on the balance between recall and precision of the system's internal test data, the second preset threshold is set to 45 and the first preset threshold is set to 70. The two thresholds allow the backend to fine-tune the parameter thresholds according to the risk control rigor of different business lines, providing a unified, quantitative, and implementable basis for subsequent defect level classification, manual review and judgment, and risk level transition, ensuring that the risk classification standards for all review projects are unified and standardized.

[0067] The system compares the calculated RiskScore with two preset thresholds in real time to automate defect level assessment and process handling. When the risk score exceeds the first preset threshold of 70, the design defect is deemed to pose an extremely high risk of financial loss, compliance violations, or business out of control. It is directly marked as a P0 blocking-level design problem, requiring designers to complete rectification and repair. Only after passing verification can the design proceed to the development and deployment process. When the risk score is greater than the second preset threshold of 45 but less than the first preset threshold of 70, the defect is deemed to pose a moderate potential business risk. Given the uncertainty in model identification, the system automatically triggers a manual review mechanism, prompting risk control experts and senior design reviewers to conduct a second manual verification to confirm the defect's authenticity, risk level, and rectification requirements. When the risk score is less than or equal to the second preset threshold of 45, it is determined to be a low-risk, routine design defect, automatically marked as a P1 or P2 general defect level. No manual review is required; only routine optimization and rectification are needed.

[0068] The system dynamically monitors the multi-factor attributes and business scenario context of identified defects throughout the entire process, and achieves objective and quantitative priority transitions based on a pre-set priority transition rule system. When an identified ordinary-level defect meets any of the following pre-set transition conditions, the priority transition mechanism is automatically triggered to raise the defect priority to P0 blocking level: Condition 1, the interactive element containing the defect is judged to be business-sensitive (Ecrit=1.5) and involves a missing tolerance dimension; Condition 2, the business impact factor Bimpact of the current page is greater than 1.6 and the number of defects with the same cognitive dimension detected on the same page exceeds 2, which is judged as a systemic design deviation. For example, for a P1 level information display defect on a regular page, if the element it contains is a ticket issuance button (trigger condition 1), the system automatically transitions it to P0 blocking level, and clearly records the trigger rule code and attribute matching results of this priority transition in the final review report, avoiding the risk of missed judgment caused by fixed classification rules, and achieving dynamic and accurate calibration of defect levels.

[0069] Reference Figure 6 In S150, the steps for recording review feedback data to update the historical feedback matrix include: S610 records the selected area profile snapshot, explicit business data options, generated five-dimensional cognitive radar chart, identified defects and corresponding risk scores and levels, and user feedback on the review results for each review. The feedback includes the acceptance, ignoring or false alarm of defects. S620, based on the type of operational feedback, converts the operational feedback into a historical correction vector for the five-dimensional cognitive dimension. The historical correction vector reflects the weight offset direction and offset amount of the current five-dimensional cognitive radar chart. S630, aggregate the historical correction vector with the current historical feedback matrix to update the historical feedback matrix; S640, the updated historical feedback matrix is ​​used as an input for the linear combination in subsequent reviews to adaptively adjust the dynamic weight matrix.

[0070] Specifically, for each complete design risk control review process, the system performs a full-element, full-link data snapshot retention operation, and structurally stores all core review data. This includes complete snapshot data of the selected area profile for this review, user-entered and system-corrected explicit data of fixed dimensions and deep business motivations, the scores and coordinate parameters of each dimension of the final generated five-dimensional cognitive radar chart, a detailed list of all identified design defects, the risk score, defect level, and transition records for each defect, and the full feedback operation data of the operators after the review. The system accurately records the operation status of each defect, such as adoption and rectification, ignoring and exemption, and false alarm marking. All data is uniformly encrypted and stored in the system's backend database, forming a complete review data ledger. This provides real, comprehensive, and effective raw data support for subsequent iterative updates of the historical feedback matrix, model optimization, and rule iteration.

[0071] The system performs refined classification, analysis, and quantitative transformation of the stored user review operation feedback data, converting subjective operation feedback into calculable standardized historical correction vectors. Each set of correction vectors precisely corresponds to the five-dimensional cognitive dimension, clearly recording the weight offset direction and specific offset amount.

[0072] The system presets the basic correction step size. Based on the feedback type, a differentiated vector generation strategy is implemented: for defect data that users have confirmed have adopted and rectified, the weight ratio of the corresponding dimension of the current five-dimensional radar is determined to be reasonable, and the generation direction is positive with a step size of 0.1. The fine-tuning correction vector is used to solidify existing rules; for defect data marked as false alarms or invalid by users, if the corresponding cognitive dimension weight is too high or the review criteria are too strict, the generation direction is negative and the step size is 1.0. The offset correction vector is applied with a strong penalty of reduced weights. For model omissions in detecting real business defects discovered through manual review, the corresponding dimension weights are deemed too low or the review standards too lenient; the generation direction is positive and the step size is 2.0. The gain correction vector is used to perform maximum weight enhancement. Meanwhile, to prevent drastic weight oscillations caused by a single feedback, a limiting mechanism is introduced to ensure that the absolute value of a single generated offset does not exceed 10% of the existing weight values ​​in the current dimension, achieving a precise, quantifiable, and controllable transformation from manual feedback to model weight optimization.

[0073] The system employs a batch iterative aggregation algorithm based on exponential moving average (EMA) to aggregate the five-dimensional historical correction vector generated in a single review with the historical feedback matrix already stored in the database, performing dimension-by-dimensional aggregation calculations. Specifically, for the weight parameters of the five cognitive dimensions in the historical feedback matrix, the system does not directly overwrite or replace them, but instead introduces a momentum decay factor. (Default value is 0.9), based on the formula The matrix is ​​then smoothed and fine-tuned. Here, Vector_correction is the five-dimensional historical correction vector generated in the current single review. The specific value of this correction is determined according to the aforementioned feedback type quantification strategy (e.g., for false alarm feedback, the corresponding dimension component). Values For feedback on missed detections, the corresponding dimension components Values (And constrained by a limit mechanism of no more than 10%). This algorithm mechanism can effectively suppress the impact of random errors and subjective noise from single operation feedback on global weights, ensuring the stability of the matrix iteration direction. At the same time, with the continuous input of review data from multiple rounds and multiple scenarios, it smoothly accumulates effective optimization rules and continuously optimizes the weight ratio of historical feedback matrices, allowing the matrix to truly adapt to the risk control review needs of various financial design scenarios, and realize the dynamic iteration, continuous evolution and self-optimization of the matrix.

[0074] The system will reintegrate the updated historical feedback matrix into the linear combination calculation system of the dynamic weight matrix. In all subsequent design risk control review tasks, the new historical feedback matrix will participate in weighted fusion in collaboration with the basic matrix and expert matrix. Based on real historical review feedback data, defect rectification effects, and manual review experience, the system will automatically adjust the proportions of each dimension of the dynamic weight matrix. This will allow the system review rules to continuously adapt to business scenario iterations, updates to financial compliance policies, changes in user operating habits, and historical risk omissions and false detections. This will form a closed-loop optimization mechanism across the entire chain of "design review - result feedback - matrix iteration - rule optimization - accurate review," continuously improving the system's design defect detection accuracy, business loss risk identification capabilities, and scenario-adaptive review level.

[0075] Based on the above method embodiments, the second embodiment of this application discloses a design risk control review system. The design risk control review system of this application embodiment can implement any of the above-described design risk control review methods, and the specific working process of each module in the design risk control review system can be referred to the corresponding process in the above method embodiments.

[0076] For ease of understanding, an example is as follows: A risk control review system is designed, including: The data acquisition module is used to acquire the selection area data of the design draft to be reviewed and to receive explicit business data input by the user. The feature extraction and profile building module is used to perform multimodal implicit feature extraction on the design draft selection area data, generate a selection area profile structured data package, and pre-fill and correct the explicit business data based on the selection area profile structured data package. The dynamic evaluation standard generation module is used to generate a five-dimensional cognitive radar chart of the current page based on the corrected explicit business data and dynamic weight matrix, using a dynamic offset algorithm as the dynamic evaluation standard. The dynamic weight matrix is ​​constructed based on the mapping relationship between business intent and cognitive dimensions. The risk quantification calculation module is used to scan the design draft selection area data to identify interactive elements and their corresponding risk levels. Based on the five-dimensional cognitive radar chart, the risk level of interactive elements, the structured data package of selection area profiles, and the defect type library, it calculates the risk score of design defects through multi-factor product logic. The defect classification and feedback update module is used to determine the defect level based on the risk score and output it, while recording review feedback data to update the historical feedback matrix.

[0077] The third embodiment of this application provides a computer device, which may include a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement a design risk control review method.

[0078] The memory can communicate with the processor via a communication bus, which can be an address bus, a data bus, a control bus, etc.

[0079] Additionally, the memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device.

[0080] Furthermore, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0081] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A design risk control review method, characterized in that, include: Acquire the selection area data of the design draft to be reviewed, and receive explicit business data input by the user; Multimodal implicit feature extraction is performed on the selected area data of the design draft to generate a structured data package of selected area profiles, and the explicit business data is pre-filled and corrected based on the structured data package of selected area profiles; Based on the corrected explicit business data and dynamic weight matrix, a five-dimensional cognitive radar chart of the current page is generated through a dynamic offset algorithm as a dynamic evaluation standard. The dynamic weight matrix is ​​constructed based on the mapping relationship between business intent and cognitive dimension. The design draft selection area data is scanned to identify interactive elements and corresponding risk levels. Based on the five-dimensional cognitive radar chart, the risk levels of the interactive elements, the structured data package of the selection area profile, and the defect type library, the risk score of the design defect is calculated through multi-factor product logic. The defect level is determined and output based on the risk score, and review feedback data is recorded to update the historical feedback matrix.

2. The design risk control review method according to claim 1, characterized in that, The steps for performing multimodal implicit feature extraction on the selected area data of the design draft to generate a structured data package of selected area images include: Recursively traverse the node tree of the design draft to extract text characters, node types, and size and style metadata; Export the canvas preview image and call the visual big data model to identify page profiles, layout, and key business actions; Monitor the recognition confidence level of the output of the large visual model. When the recognition confidence level is lower than a preset threshold, automatically switch to the local recognition stream based on the design draft node tree for recognition. The extracted metadata and recognition results are cleaned and denoised to generate a structured data package containing page title, key action semantics, business field summary, canvas size and structural signals.

3. The design risk control review method according to claim 1, characterized in that, The explicit business data includes fixed-dimensional data and in-depth business motivation data; The steps to generate a five-dimensional cognitive radar map include: Based on the mapping rules defined in the dynamic weight matrix, the options in the corrected fixed-dimensional data and deep business motivation data are discretized into weight vectors on five cognitive dimensions: rigor, certainty, efficiency, clarity, and fault tolerance. The algorithm uses a dynamic offset algorithm to perform weighted summation and offset calculation on the weight vector using the dynamic weight matrix to obtain the comprehensive score of each cognitive dimension. The dynamic weight matrix assigns a base offset weight to the deep business motivation data that is greater than that to the fixed dimension data, so as to ensure that the business intent plays a dominant role in the radar chart offset. The fixed dimension weight summation lays the foundation for the audit, and the deep business motivation weight summation represents the business intent. The comprehensive score is normalized to generate the basic five-dimensional cognitive radar chart of the current page; Obtain the environment multiplier and page type features, apply additional weighting to the review rules corresponding to the mobile terminal based on the environment multiplier, adjust the weights of the corresponding cognitive dimensions based on the page type features, and output the final five-dimensional cognitive radar chart.

4. The design risk control review method according to claim 1, characterized in that, The specific process for identifying interactive elements and their corresponding risk levels is as follows: Scan the interactive components in the node tree of the design draft, and classify buttons, links, input boxes, and text fields into atomic levels. The atomic level is associated with common interactive risks. The amount input component, date picker, and upload control are categorized as components, and the associated logic inputs risk at the component level. The issuance button, endorsement button, risk warning label, and review status indicator are classified as business-sensitive, and the business-sensitive level is associated with asset security risks.

5. The design risk control review method according to claim 1, characterized in that, The steps for calculating the risk score of design defects using multi-factor product logic include: Obtain the basic risk value from the defect type library; Extract business impact factors from the structured data package of the selected area profile; The element level factor is determined based on the risk level of the interactive element, and the element level factor increases with the atomic level, component level, and business sensitivity level. The scores of the corresponding dimensions are extracted from the five-dimensional cognitive radar chart to determine the cognitive dimension deviation factor; Obtain the output of the large visual model or the local recognition stream, and determine the recognition confidence factor; The risk score of the design defect is obtained by multiplying the basic risk value, business impact factor, element level factor, cognitive dimension deviation factor, and identification confidence factor.

6. The design risk control review method according to claim 5, characterized in that, The specific logic for determining the defect level is as follows: Set a first preset threshold and a second preset threshold, wherein the first preset threshold is greater than the second preset threshold; When the risk score exceeds the first preset threshold, the design defect is marked as a P0 blocking level design problem; When the risk score exceeds the second preset threshold but does not exceed the first preset threshold, a manual review is required. Monitor the risk score changes across thresholds. When preset rule conditions are met, trigger the level jump mechanism to adjust the defect priority and indicate the jump conditions in the report.

7. The design risk control review method according to claim 1, characterized in that, Before generating the five-dimensional cognitive radar chart of the current page using a dynamic offset algorithm based on the corrected explicit business data and dynamic weight matrix, the process also includes constructing the dynamic weight matrix: Based on the page title, key actions and terminal type in the structured data package of the selected area profile, the corresponding expert skill package is automatically matched and loaded from the expert skill center. The expert skill package includes supply chain invoice skill package, payment risk control skill package and mobile terminal usability skill package. Obtain a pre-set basic matrix, an expert matrix corresponding to the expert skill package, and a historical feedback matrix constructed from past review feedback data. The basic matrix includes standard weights for general audit rules and questionnaires, and the expert matrix includes business rule weights for vertical fields. A dynamic weight matrix is ​​formed by linearly combining the base matrix, expert matrix, and historical feedback matrix using fusion coefficients. The process of generating a five-dimensional cognitive radar chart for the current page using a dynamic offset algorithm specifically involves generating the five-dimensional cognitive radar chart for the current page based on the corrected explicit business data and the dynamic weight matrix.

8. The design risk control review method according to claim 7, characterized in that, The steps for recording review feedback data to update the historical feedback matrix include: For each review, the following are recorded: a snapshot of the selected area profile, explicit business data options, a generated five-dimensional cognitive radar chart, identified defects and their corresponding risk scores and levels, and user feedback on the review results. The feedback includes the acceptance, ignoring, or false alarm flags for defects. Based on the type of the operation feedback, the operation feedback is converted into a historical correction vector for the five-dimensional cognitive dimension. The historical correction vector reflects the weight offset direction and offset amount of the current five-dimensional cognitive radar chart. The historical correction vector is aggregated with the current historical feedback matrix to update the historical feedback matrix; The updated historical feedback matrix is ​​used as an input for the linear combination in subsequent reviews to adaptively adjust the dynamic weight matrix.

9. A design risk control review system, characterized in that, Performing the design risk control review method as described in any one of claims 1 to 8 includes: The data acquisition module is used to acquire the selection area data of the design draft to be reviewed and to receive explicit business data input by the user. The feature extraction and profile building module is used to perform multimodal implicit feature extraction on the design draft selection area data, generate a selection area profile structured data package, and prefill and correct the explicit business data based on the selection area profile structured data package; The dynamic evaluation standard generation module is used to generate a five-dimensional cognitive radar chart of the current page based on the corrected explicit business data and dynamic weight matrix, using a dynamic offset algorithm as a dynamic evaluation standard. The dynamic weight matrix is ​​constructed based on the mapping relationship between business intent and cognitive dimension. The risk quantification calculation module is used to scan the design draft selection area data to identify interactive elements and corresponding risk levels, and calculate the risk score of design defects through multi-factor product logic based on the five-dimensional cognitive radar chart, the risk level of the interactive elements, the structured data package of the selection area profile, and the defect type library. The defect classification and feedback update module is used to determine the defect level based on the risk score and output it, while recording review feedback data to update the historical feedback matrix.

10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the design risk control review method as described in any one of claims 1 to 8.

Images