A machine learning-based quality management digital transformation roadmap generation method

Abstract

The application discloses a kind of quality management digital transformation roadmap generation methods based on machine learning, it is related to product data management technical field.The method, by obtaining process time series data and quality result data, and calling pre-trained machine learning model self-adapting execution time lag correlation analysis and causal inference, the influence path of process parameter to quality management efficiency is quantified;According to the influence path quantization result, generate capability gap subgraph under different risk conditions, and carry out data labeling;Combined with the mapping relationship chain of pre-set phased target constraint, form the mapping relationship chain of phased task and verification scene, and carry out dynamic priority sorting;Summarize phased task, sorting result and resource planning, form digital transformation roadmap, and drive its advance;In the process of advancing, determine whether the final output executable, measurable and can be automatically rearranged with the change of working condition digital transformation roadmap, to realize the whole life cycle quality management closed-loop optimization for dynamic working condition.

Description

Technical Field

[0001] This invention relates to the field of product data management technology, and in particular to a method for generating a digital transformation roadmap for quality management based on machine learning. Background Technology

[0002] Against the backdrop of enterprise digital transformation and increasingly stringent quality supervision, quality management refers to an organization's systematic management activities of the quality of its deliverables. Manufacturing focuses on the quality of products and their components, materials, and manufacturing processes, while the service industry focuses on the quality of service processes and outcomes (such as consistency, compliance, and experience indicators). It is typically led by the quality department, with collaboration from R&D, procurement, production / delivery, and after-sales departments, forming a closed-loop system encompassing quality planning (quality objectives, key quality characteristics, control plans, and inspection schemes), quality assurance (system / process audits, document and change control, training and authorization), quality control (incoming material / process / finished product inspection and statistical process control), and quality improvement (non-conformity handling, root cause analysis, and effectiveness verification). Typical application scenarios include anomaly isolation and handling, batch and work order traceability, supplier quality control, process stability assessment, customer complaint closure, and reliability improvement, particularly prevalent in manufacturing sectors with high consistency and strong traceability requirements.

[0003] Existing technologies often employ maturity models or industry benchmarking lists for generation: Through questionnaires, interviews, audit records, and existing indicators, current status scores and gap identification are performed on capability domains such as R&D quality, supplier quality, process quality, inspection and testing, customer quality, and compliance traceability. This results in a phased construction task library (such as data governance and indicator system construction, quality analysis and prediction capability building, etc.), and priorities and investment sequences are determined through expert review, hierarchical analysis, Delphi, and other methods. Finally, phase goals, milestones, project portfolios, and resource plans are output. These roadmaps typically aim to cover the entire lifecycle, incorporating demand, design, procurement, manufacturing / delivery, and after-sales service into the planning scope.

[0004] Taking the production process of lithium-ion battery electrodes as an example, this type of product typically has strict requirements for key quality characteristics such as coating areal density, thickness uniformity, porosity, residual solvent / water content, adhesion, and final capacity consistency. Under conditions such as production line speed adjustment, batch switching, slurry viscosity drift, oven temperature fluctuations, and tension and roller pressure disturbances, electrode coatings are prone to uneven thickness, edge defects, micropores / pinholes, cracks, or compaction density fluctuations, which can lead to quality risks such as capacity dispersion, internal resistance shift, and decreased consistency. To address this scenario, it is usually necessary to incorporate process timing data such as line speed, tension, coating gap, slurry viscosity / solid content, temperature and humidity, oven multi-temperature zones, roller pressure, and roller gap, as well as quality result data such as online thickness / area density detection, random sampling electrical performance data, and records of scrap, rework, and abnormal handling into the evaluation input during the roadmap development stage. This forms a phased task library and implementation sequence, ultimately outputting a roadmap that can support stable process control and continuous achievement of key quality characteristics under multiple operating conditions.

[0005] However, existing roadmap generation processes primarily rely on questionnaires, interviews, and industry benchmarking as inputs, focusing on static scoring of whether quality processes are available, implemented, and covered. This makes it difficult to fully utilize process time-series data such as line speed, temperature zone, rolling pressure, and tension, along with outcome data such as thickness uniformity and capacity consistency, for correlation analysis. This results in the inability to objectively identify key risk conditions and critical influencing factors. Furthermore, roadmap phases are often simplified to process digitization and system integration, failing to simultaneously establish a phased verification path for closed-loop capabilities such as unified traceability primary keys for dynamic operating conditions, process capability assessment, anomaly identification threshold setting, predictive warnings, and parameter write-back. This makes it difficult for roadmaps to undergo rolling evaluation and automatic reordering as operating conditions change, ultimately hindering the continuous adaptation of generated roadmaps to the process stability control requirements of scenarios with frequent line speed adjustments and batch switching. Summary of the Invention

[0006] To address the technical problems in existing technologies, this invention provides a method for generating a roadmap for digital transformation of quality management based on machine learning. The technical solution is as follows: S1: Acquire real-time data including at least process time series data and quality result data; perform time-delay correlation analysis and causal inference within a preset quality analysis period; quantify the impact path of each process parameter on quality management effectiveness; the time-delay correlation analysis and causal inference are adaptively executed by calling a pre-trained machine learning model.

[0007] S2, based on the quantitative results of the influence path, generates a sub-graph representing the capability gap of each capability domain under different risk conditions, in which the quantitative gap between the current capability value and the target value calculated based on the data and the dominant influencing factors are marked.

[0008] S3, based on the capability gap subgraph and combined with the preset phased target constraints, forms a mapping relationship chain between each phase task and its corresponding verification scenario, and performs dynamic priority sorting.

[0009] S4 summarizes the resource planning results after summarizing the tasks of each stage and the dynamic priority ranking, and forms a digital transformation roadmap. At the same time, it drives the advancement of the digital transformation roadmap based on the verification results of the tasks of each stage.

[0010] During the process, S5 determines whether to trigger the roadmap update engine based on the results of status verification or abnormal operating condition identification, and then further determines whether to output the final version of the digital transformation roadmap.

[0011] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. This invention acquires real-time data, including at least process time-series data and quality result data. Within a preset quality analysis period, it invokes a pre-trained machine learning model to adaptively perform time-delay correlation analysis and causal inference, quantifying the impact path of each process parameter on quality management effectiveness. This enables the objective identification of key risk conditions and dominant influencing factors, providing a quantitative basis for roadmap planning. Based on the quantification results of the impact path, a capability gap sub-map is generated under different risk conditions, marking the current capability value, target capability value, and quantified gap calculated from the data, and associating them with dominant influencing factors, making the source of the capability gap explainable and traceable. Combined with preset phased targets... The process involves constructing a mapping chain between phased tasks and verification scenarios, and dynamically prioritizing them based on quantified gaps and causal strength. This results in a phased orchestration where tasks are executable and verification is reproducible, preventing tasks from merely remaining at the level of process digitization or system integration. The process then summarizes the tasks at each phase and the dynamically prioritized resource planning results to form a digital transformation roadmap. The roadmap's progress is driven by phased verification results, making milestones and acceptance criteria quantifiable and measurable. During the process, the roadmap update engine is triggered based on status verification or abnormal operating condition identification results, allowing for rolling adjustments to the task order and verification scenarios, ultimately determining the output of an iterative or final version of the roadmap. Through these steps, the roadmap transitions from static, experience-based planning to data-driven, closed-loop iterative optimization, improving its adaptability to dynamic operating conditions and the efficiency of continuous quality improvement.

[0012] 2. First, when establishing data preprocessing and feature engineering rules, the time alignment rules for process time-series data and the label definition rules for quality result data are written into the metadata table corresponding to the pre-trained machine learning model, making the analysis configuration reusable and version-manageable. Then, within a preset quality analysis period, the process time-series data and quality result data are timestamped according to the time alignment rules, and the quality result data undergoes deviation quantification processing according to the label definition rules, yielding performance response parameters used to uniformly characterize quality management effectiveness. Using the performance response parameters as the target variable, the machine learning model adaptively determines the candidate time lag range and candidate feature set, performs time lag causality verification on each process parameter, obtains the optimal lag time and its causal strength, and achieves self-adaptation to different operating conditions and response speeds. Furthermore, the influence path type is determined based on causal strength and causal relationship labeling. Through this mechanism, the direction, intensity, and lag characteristics of the impact of key process parameters on quality management effectiveness can be quantitatively identified without relying on subjective questionnaires, distinguishing between direct and indirect influence chains transmitted through mediating variables, and improving the stability and interpretability of the analysis results under changing operating conditions.

[0013] 3. When a direct impact path exists, firstly, based on the pre-set capability domain mapping rules in the pre-trained machine learning model, the direct impact edges are assigned to the corresponding capability domains. Within each capability domain, the path is sliced ​​according to risk conditions, constructing a set of nodes for the capability quantification gap subgraph. Then, the current capability value is calculated based on causal strength, optimal lag time, and effectiveness response parameters, and aligned with the preset target capability value to obtain the direct quantification gap and capability gap labels. Simultaneously, the causal strength, direction of action, and lag time of the dominant influencing factor are bound and displayed, generating corresponding task candidate nodes and labeling them with gaps and causal attributes, thus achieving visualization of the direct path and interpretable gap location. When an indirect impact path exists, mediator variables are identified according to the mediation mapping rules, and the indirect impact edges are split and assigned to at least two capability domains according to business semantics. A capability coupling gap subgraph containing mediator nodes and cross-domain coupling edges is constructed under each capability domain, and the mediator variables must pass the mediation validity verification. Furthermore, the cross-domain coupling effect coefficient is calculated using causal strength to form indirect quantification gap labels, which are then labeled to relevant capability item nodes, thereby generating cross-domain collaborative task candidate nodes and forming coupling association edges. Through the above mechanism, the data-driven impact path can be transformed into a computable gap subgraph and task candidates, which can not only explain the source of the gap and quantify the priority, but also reveal the shortcomings of cross-domain coupling and improve the roadmap's adaptability to complex working conditions and cross-link collaborative governance.

[0014] 4. First, obtain the candidate task nodes and their verification scenario instances after dynamic priority sorting, and generate task execution sequences according to stage affiliation. Under the premise of dependency constraints, tasks are split into stage work packages in parallel and bound to verification scenario instances to obtain compliance judgment items. Then, based on the compliance judgment items, the execution results of the verification scenario instances are quantified for deviation, and compared with the corresponding direct or indirect quantification gap to obtain the gap convergence value. The gap convergence value is written into the acceptance record table, and the stage completion status is marked on the direct related edge of the capability quantification gap subgraph or the coupled related edge of the coupled capability gap subgraph, thereby outputting a staged executable and measurable roadmap. During the roadmap advancement, the acceptance deviation is continuously extracted based on the compliance judgment items. When the gap convergence value is not less than the preset convergence threshold and the acceptance deviation is not greater than the preset deviation threshold, the stage work package is judged to be compliant and the final version roadmap is output. When the gap convergence value is lower than the preset convergence threshold or the acceptance deviation is greater than the preset deviation threshold, the update engine is triggered to output the iterative version. When both fail to meet the threshold, an anomaly is judged to exist and an anomaly block is marked to trigger task reordering and scenario reconfiguration. This mechanism transforms the roadmap from a planning document into a verifiable closed loop, using quantitative convergence and acceptance results to drive rolling optimization, thereby improving the roadmap's adaptability and continuous improvement efficiency in dynamic operating conditions and cross-process collaboration. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a method for generating a quality management digital transformation roadmap based on machine learning, provided in an embodiment of the present invention; Figure 2 The output determination flowchart of the final version of the digital transformation roadmap provided in the embodiments of the present invention; Figure 3 Causal strength comparison curves provided for embodiments of the present invention; Figure 4 This is an iterative acceptance monitoring curve during the implementation of the digital transformation roadmap provided in this embodiment of the invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0020] Taking the production process of lithium-ion battery electrode coating, drying, and rolling as an example, to solve the problems of existing roadmaps relying on static scoring, difficulty in identifying key risk conditions and dominant factors, and inability to be re-arranged as conditions change, this invention provides a machine learning-based method for generating a digital transformation roadmap for quality management. Figure 1 The flowchart shown is a method for generating a quality management digital transformation roadmap based on machine learning. The processing flow of this method may include the following steps: S1: When establishing data preprocessing and feature engineering rules, the time alignment rules for process time series data and the label definition rules for quality result data are written as analysis configuration fields into the metadata table corresponding to the pre-trained machine learning model. The metadata table is used to structurally record the data scope and processing strategies required for model calls, including at least the data source and field mapping (process points such as line speed, tension, coating gap, slurry viscosity / solid content, temperature and humidity, oven multi-temperature zone temperature, roller pressure and gap, equipment status, etc., are from the data acquisition gateway; online thickness / area density is from the online measurement system; sampling electrical performance is from the testing / laboratory system; rework and scrap and abnormal handling are from the non-conforming product handling order, rework and scrap order and abnormal event closed-loop record in the manufacturing execution system or quality management system), unified time base and sampling period, aligned primary keys (batch number / work order number / volume number / equipment number / time window number), missing and abnormal handling rules, window division and working condition slicing rules, feature engineering rules, candidate time delay range configuration, preset significance threshold, and weight fusion rules for performance response parameters.

[0021] Within a preset quality analysis period (e.g., selecting data from the most recent 30 days or the most recent N batches including risky operating conditions such as linear speed adjustment, batch switching, viscosity drift, and temperature fluctuation), data of different frequencies and granularities are mapped to a unified analysis window according to time alignment rules. Process time-series data are resampled and aggregated to obtain process characteristics such as mean, variance, rate of change, drift, and stability. Quality results such as online / offline detection, rework and scrap, and abnormal closure are associated with window backfilling by primary key. According to label definition rules, the quality result data is subjected to deviation quantification processing, and results such as thickness / area density deviation, uniformity fluctuation, and capacity consistency dispersion are converted into calculable labels. Management performance indicators such as non-conformance rate, rework and scrap rate, abnormal handling closure time, and closure achievement rate are statistically analyzed. Then, at least two or more indicators are normalized and aligned according to the weight rules in the metadata table and then weighted and fused to obtain performance response parameters used to uniformly characterize the effectiveness of quality management.

[0022] Using performance response parameters as target variables, the machine learning model first adaptively generates candidate time delay ranges for each process parameter based on multi-lag correlation / mutual information and process priors. Then, it constructs a candidate feature set according to feature engineering rules and removes redundant features through sparse filtering or attention weights. Subsequently, it constructs lag features for each process parameter within the candidate time delay set and performs time delay causality verification under confounding factors such as batch switching markers, equipment status, and other process variables, calculating the causal strength and significance of each lag point. Among the lag points that pass the significance threshold, the lag value with the largest causal strength is selected as the optimal lag time, and the causal strength corresponding to this lag value (which can simultaneously output the direction of action and confidence level) is used as the quantitative result of the impact path of the process parameter on quality management performance. This provides an interpretable and verifiable basis for subsequent capability gap subgraph construction and task priority generation.

[0023] The machine learning model can be a combination of a temporal feature encoder and a causal inference head (e.g., an encoder based on a temporal convolutional network + a lag selection module with an attention mechanism + a causal strength regression module). Its input consists of process parameter sequence features (including sliding window statistics / rates of change / stability indices for linear velocity, temperature zones, tension, and roller pressure, etc.) aligned to timestamps according to the metadata table, quality result deviation features, and label information corresponding to the performance response parameters. Its output includes the candidate time lag range, optimal lag time, corresponding causal strength, and direct / indirect causal relationship labels for each process parameter. The training method involves: first, self-supervised pre-training (mask reconstruction and short-term prediction) on historical multi-source time series to learn a general temporal representation; then, fine-tuning using the performance response parameters as the supervised target, and incorporating temporal constraints and sparsity regularization to suppress spurious correlations, thereby obtaining a pre-trained model capable of adaptive time lag and causal strength estimation.

[0024] If the causal strength is higher than the preset significance threshold, and the causal verification result between the process parameter and the performance response parameter is marked as a direct causal relationship within the optimal lag time, then it is determined that the corresponding process parameter has a direct impact path on quality management performance. The reason is that: the preset significance threshold is used to constrain the causal effect to be statistically stable and reproducible, excluding accidental correlations and noise disturbances; the optimal lag time is used to satisfy the time constraint that the cause precedes the effect, indicating that the change of process parameter has the greatest and most consistent impact on performance response within this lag window; the direct causal label indicates that after controlling for confounding factors such as batch switching, equipment status, environment, and other process variables, the direct effect of the process parameter on the performance response parameter still holds, and the performance change can be explained without introducing additional transmission variables, so it can be identified as a direct impact path.

[0025] If the causal strength is higher than the preset significance threshold, or if the causal verification result between the process parameter and the performance response parameter within the optimal lag time is not marked as a direct causal relationship and meets the preset mediator variable conditions, then it is determined that the corresponding process parameter has an indirect influence path on quality management effectiveness. The reason is that in this case, there are signs of stable influence between the process parameter and the performance response, but the direct side is unstable or weakened by the control variable, which is more in line with the structure transmitted through the business mechanism. When there is at least one mediator variable that makes the process parameter directly causal with the mediator variable, and the mediator variable directly causal with the performance response parameter, it indicates that the process parameter mainly affects the quality management effectiveness by influencing the mediator variable (such as process stability indicators, quality deviation indicators, or abnormal event indicators, etc.), so it is determined to be an indirect influence path.

[0026] Conversely, if the corresponding process parameters have no impact on quality management effectiveness during the current quality analysis period, they will not be included in the impact path analysis process. Including them would introduce spurious correlations and interfere with subsequent capability gap identification and task priority generation. Therefore, they will not be included in the impact path analysis process and will be re-evaluated in subsequent rolling analysis.

[0027] The pre-defined conditions for mediating variables include: the existence of at least one mediating variable such that the causal verification result between the process parameter and the mediating variable is marked as a direct causal relationship, and the causal verification result between the mediating variable and the performance response parameter is marked as a direct causal relationship.

[0028] S2: Scenario 1: The process of generating the subgraph of the capability gap corresponding to the path that directly affects the path.

[0029] First, based on the pre-defined capability domain mapping rules in the machine learning model, the direct impact edges of process parameters and performance response parameters in the direct impact path are assigned to the corresponding capability domains. A capability domain is a set of capabilities divided by lifecycle or business process in the digital transformation of quality management, including at least one or more of the following: data governance and traceability, inspection and testing, anomaly closure-loop handling, supplier quality, and R&D quality. The capability domain mapping rules include at least the mapping relationship between process parameters and business processes / procedures / system objects (equipment, work orders, batches, inspection points, abnormal events, etc.), and the attribution relationship between business objects and capability domains. For example, process parameters and event fields related to batch primary keys, work order traceability, and data consistency are mapped to the data governance and traceability capability domain; process parameters related to online / offline inspection points, sampling strategies, and judgment thresholds are mapped to the inspection and testing capability domain; and process parameters related to anomaly triggering, classification, handling time, and closure-loop achievement rate are mapped to the anomaly closure-loop handling capability domain, thus ensuring that the attribution of impact edges has business semantic consistency and interpretability.

[0030] Secondly, operating condition slices are performed on the directly affected paths under each capability domain. Operating condition slices are used to separate and characterize the differences in the effects of the same process parameter under different risk conditions. The basis for operating condition slices may include operating condition labels such as linear speed gear / adjustment events, batch switching events, temperature fluctuation levels, tension / roller pressure disturbance levels, and equipment alarm status. Specifically, the data within the preset quality analysis period can be divided into several sub-intervals according to the operating condition labels, and the mean, quantiles, fluctuation degree, and anomaly density of the performance response parameters can be statistically analyzed in each sub-interval to form a two-dimensional slice of capability domain and operating condition.

[0031] Then, after completing the attribution and slicing, a node set is constructed for the capability quantification gap subgraph. The node set includes at least capability item nodes, dominant influencing factor nodes, and task candidate nodes: Capability item nodes are used to express measurable capability items within the capability domain, such as traceability primary key coverage, data caliber consistency, inspection rule completeness, threshold strategy completeness, and anomaly closure timeliness; Dominant influencing factor nodes consist of a preset number of process parameters with the highest causal intensity in the direct impact path. The preset number can be set according to the capability domain size or risk level, and is used to focus on the few key process parameters that contribute the most to the effectiveness response; Task candidate nodes are used to express the digital construction tasks or governance tasks that can be implemented to narrow the capability gap, such as traceability primary key unification, data collection supplementation, threshold strategy configuration, online detection and sampling linkage, anomaly classification and handling arrangement, and parameter write-back interface construction.

[0032] Furthermore, based on the causal strength, optimal lag time, and corresponding performance response parameters of the directly affecting path, the current capability value of the corresponding capability domain is calculated and marked on the capability item node in the node set. The calculation of the current capability value can be completed by combining the capability measurement caliber in the metadata table: on the one hand, the performance response parameters are decomposed or mapped according to relevant indicators within the capability domain to obtain capability performance (for example, the ability to handle abnormal closed loops can be obtained by fusing the closed loop achievement rate and the normalized compliance of the closed loop duration; the inspection and testing capability can be constructed by indicators such as non-conformance detection rate, missed detection rate, and online / offline consistency; the data governance and traceability capability can be constructed by indicators such as primary key coverage, traceability chain integrity, and cross-system coding consistency rate). On the other hand, the causal strength and optimal lag time of the directly affecting path are introduced as influence weighting factors, that is, higher weights are given to influence edges with higher causal strength, shorter lag, and more consistent direction of action, so that the current capability value can reflect the actual support degree of the capability domain for performance response under the current working conditions.

[0033] Simultaneously, pre-defined target capability values ​​are marked on capability item nodes. These target capability values ​​can originate from corporate quality objectives, regulatory / customer requirements, benchmarks, or historical best levels, and their source, scope, and version number are also fixed in the metadata table. Subsequently, the absolute value of the difference between the current capability value and the target capability value is recorded as the direct quantitative gap, forming a capability gap label. The capability gap label is used to quantitatively characterize how far away from the target capability is, and is bound to the dominant influencing factor node to explain the source of the gap: the bound display includes at least the causal strength (indicating the magnitude of the impact), the direction of action (indicating the directionality of the improvement / deterioration of the performance response due to the increase / decrease of the process parameter), and the optimal lag time (indicating the speed of impact transmission and response window). This makes the capability gap not only measurable but also traceable to specific process variables and time scales, facilitating the subsequent development of verifiable governance measures.

[0034] Next, based on the direct quantification gap and the dominant influencing factors, the task candidate nodes corresponding to the capability item nodes are generated. The generation rules may include: when the direct quantification gap exceeds the preset gap threshold and the dominant influencing factors are concentrated in a certain type of missing data or rule gap, tasks such as data collection and supplementation, field standardization, primary key unification, and indicator caliber governance are generated; when the dominant influencing factors are manifested as amplified fluctuations and short lags under a certain type of working condition, tasks such as threshold strategy configuration, anomaly identification rule optimization, and real-time alarm are generated; when the direction of action and lag characteristics point to an excessively long handling chain or an unstable closed loop, tasks such as anomaly classification, handling arrangement, and closed loop duration reduction are generated; and the task inputs and outputs, dependencies, verifiable indicators, and suggested verification scenario types are recorded in the task candidate nodes.

[0035] Finally, the business semantic relationships between nodes in the node set are used as direct edges to construct a subgraph topology. For example, directed or bidirectional edges are established between dominant influencing factor nodes, capability item nodes, and task candidate nodes to express the correspondence between the source of influence, gap bearing capacity, and governance measures. At the same time, the direct quantitative gap, causal strength, direction of action, and optimal lag time are labeled on the direct edges to form a visualized capability quantitative gap subgraph. This allows for a clear presentation under different risk conditions: which process parameters affect the effectiveness response with what intensity and lag time, thus causing the capability item to deviate from the target; and which task candidate nodes should be prioritized to reduce this deviation.

[0036] Scenario 2: The process of generating the capability gap subgraph for indirect influence paths.

[0037] First, based on the pre-defined mediation mapping rules in the machine learning model, mediating variables in the indirect influence path are identified. The indirect influence edges of process parameters on performance response parameters are then decomposed according to business semantics and assigned to at least two capability domains, ensuring that cross-stage transmission relationships have interpretable business implications. The mediation mapping rules include at least: the definition of a candidate set of mediating variables (which may be process stability indicators, quality deviation indicators, abnormal event indicators, inspection consistency indicators, traceability chain integrity indicators, etc.), the mapping relationship between process parameters and mediating variables (e.g., the impact of line speed / tension / temperature fluctuations on thickness uniformity fluctuations, pinhole defect density, and the probability of triggering scrap and rework), and the mapping relationship between mediating variables and performance response parameters (e.g., the correlation between thickness uniformity fluctuations and non-conforming rate, rework and scrap rate, and abnormal closure time). This transforms influences that cannot be stably determined as direct causality into transmission chains that can be explained by mediation.

[0038] After identifying mediating variables, a mediation validity check needs to be performed on them. This check ensures that the mediating variables do indeed play a causal role rather than being merely coincidental. The check criteria may include: the causal relationship between the process parameter and the mediating variable is marked as a direct causal relationship with a causal strength exceeding a preset significance threshold; the causal relationship between the mediating variable and the performance response parameter is also marked as a direct causal relationship with a causal strength exceeding a preset significance threshold. Mediating variables that pass the mediation validity check are recorded as mediating variable nodes, forming an upstream edge with the corresponding process parameter and a downstream edge with the performance response parameter, structurally creating an indirect influence path between the process parameter, the mediating variable, and the performance response parameter.

[0039] Secondly, under each capability domain, the indirect influence path is sliced ​​into operating conditions, and a capability coupling gap subgraph containing intermediary nodes is constructed. The operating condition slices can be divided into sub-intervals based on operating condition labels such as linear speed gear / adjustment events, batch switching events, viscosity drift levels, temperature fluctuation levels, tension / roller pressure disturbance levels, and equipment alarm status, so that the transmission strength differences of the same intermediary chain under different operating conditions can be independently characterized. Within each sub-interval after slicing, the mean, fluctuation amplitude, drift trend, and anomaly density of the intermediary variables are statistically analyzed, and the changes in the performance response parameters are statistically analyzed simultaneously to form a cross-domain coupling operating condition profile. The node set of the coupling capability gap subgraph includes at least task candidate nodes, intermediary variable nodes, and cross-domain coupling edge nodes. The cross-domain coupling edge nodes are used to explicitly express the coupling relationship between different capability domains through intermediary variables, such as the transmission coupling between upstream capability domains (such as inspection and testing, process capability assessment, or data governance and traceability) and downstream capability domains (such as anomaly closed-loop handling or quality improvement). The task candidate nodes are used to express governance tasks that require cross-domain collaborative implementation.

[0040] like Figure 3 The causal strength comparison curves shown intuitively reveal the evolution of cross-domain transmission effects in quality management by comparing the dynamic trends of the causal strength of process parameters on mediating variables (①) and the causal strength of mediating variables on performance response parameters (②): Causal strength ① continuously increases as the analysis window advances (from 0.28 to 0.77), indicating that the upstream process control is constantly strengthening; while causal strength ② shows a downward trend (from 0.79 to 0.48), reflecting that the direct influence of mediating variables on final performance is gradually weakening. The two causal strength curves show significant dynamic decoupling and transmission bottleneck shift characteristics: the monotonous increase of strength ① reveals the continuous improvement of process parameter control capabilities, while the monotonous decrease of strength ② suggests that the direct dependence of quality management performance on a single mediating variable is decreasing. The curve formed by the combined effect of the two intuitively quantifies the concurrent evolution process of front-end control strengthening and back-end response passivation in the cross-domain transmission path.

[0041] Based on this dynamic relationship, by calculating the coupling effect coefficient (such as the product), the cross-domain transmission strength of upstream disturbances to the final performance can be quantified, and an indirect quantitative gap label can be generated. For example, in lithium battery manufacturing, when the intensity ① of oven temperature fluctuation on residual solvent is 0.71 and the intensity ② of residual solvent on overall performance is 0.52, the coupling coefficient is 0.3692. This coefficient not only quantifies the transmission gain, but its difference in growth rate with intensity ① also exposes the bottleneck of the transmission path. That is, although the existing technology has strong process control capabilities, the transmission efficiency of intermediate variables to performance has become the key constraint on overall performance. This enables managers to accurately identify that the focus of quality improvement should shift from strengthening upstream control to optimizing transmission link links such as detection feedback and closed-loop handling, thereby achieving a leap from local parameter control to global transmission link health management, and improving the targeting of resource investment and the ability to proactively control quality risks.

[0042] Based on this, the indirect quantitative gap is calculated by combining the degree to which the mediating variable deviates from the target interval and the degree to which the effectiveness response parameter deviates from the target, and an indirect quantitative gap label is formed. The indirect quantitative gap label includes at least the gap size, the main mediating variable, the strength of the two causal segments, the direction of action, and the key lag time, and these are respectively labeled on the relevant capability item nodes to express which capability items contribute to the gap in the cross-domain transmission chain. Finally, based on the cross-domain coupling effect coefficient and the indirect quantitative gap in the indirect quantitative gap label, cross-domain collaborative task candidate nodes are generated according to the preset collaborative generation rules. At the same time, mediating variable nodes, cross-domain coupling edge nodes, and task candidate nodes are connected by coupling association edges, and the cross-domain coupling effect coefficient, indirect quantitative gap, strength of the two causal segments, and direction of action are labeled on the coupling association edges, thereby forming a visualized coupling capability gap subgraph.

[0043] Based on the capability quantification gap subgraph generated in Scenario 1 or the coupled capability gap subgraph generated in Scenario 2, combined with preset phased target constraints and task orchestration rules, S3 is executed to form a mapping relationship chain between tasks at each stage and verification scenarios, and to complete dynamic priority ranking. The preset phased target constraints refer to the set of constraints used to limit the digital transformation roadmap's phased progression, including at least phase target type constraints, verification pass constraints, and resource boundary constraints. Phase target type constraints are used to divide candidate task nodes into the data foundation stage, capability enhancement stage, and closed-loop control stage according to construction goals. The data foundation stage is used for... The process involves several stages: 1) Unifying traceability primary keys, standardizing fields, ensuring data connectivity, and maintaining consistent indicator definitions; 2) Enhancing capabilities by configuring rules and threshold strategies, identifying and issuing warnings, assessing process capabilities, and improving stability; 3) Closed-loop control by orchestrating actions, writing back strategies, and achieving closed-loop operation, with gap convergence as the acceptance guide; 4) Verification constraints to limit the threshold for entering the next stage, such as ensuring the gap convergence value is not lower than a preset convergence threshold and the acceptance deviation is not higher than a preset deviation threshold; and 5) Resource boundary constraints to limit the upper limits of manpower, time, and computing power input for each stage, ensuring that task orchestration is executable, measurable, and capable of rolling updates.

[0044] S3: The direct quantification gap marked on the directly related edge in the capability quantification gap subgraph is recorded as the single-domain gap weight, or the indirect quantification gap marked on the coupled related edge in the coupled capability gap subgraph is recorded as the cross-domain gap weight; for each task candidate node, the gap weight on its related edge, the target constraint type of the corresponding capability item node, and the prerequisite dependency relationship are read, and the stage division is performed according to the preset staged target constraints.

[0045] When a candidate node for a task meets the basic data constraints (such as field standardization, data collection supplementation, and caliber alignment) and its difference weight reaches the threshold of the first stage, it is classified into the first stage and the verification scenario type is marked as data connectivity and traceability consistency verification.

[0046] When a task candidate node satisfies capability enhancement constraints (such as threshold strategy configuration, anomaly identification and early warning) and its gap weight reaches the second-stage threshold, and its prerequisite dependencies are satisfied, it is classified into the second stage and the verification scenario type is marked as rule validity and stability verification.

[0047] When a candidate node of a task satisfies closed-loop control constraints (such as disposal orchestration and parameter write-back) and its gap weight reaches the threshold of the third stage, and passes the verification scenario of the previous stage, it is classified into the third stage and the verification scenario type is marked as closed-loop achievement and gap convergence verification.

[0048] For collaborative tasks corresponding to cross-domain gap weights, if the basic data conditions of the capability domain are not met, they are first split into pre-tasks of each domain and the stage affiliation is moved forward. After the dependencies are met, they are then assigned to the corresponding stage for parallel implementation, thereby determining the stage affiliation and verification scenario type of each task candidate node.

[0049] The aforementioned first-stage threshold, second-stage threshold, and third-stage threshold can be set in layers based on the quantiles or risk levels of the gap between the historical baseline and the target: using the statistical distribution of the weights of the gaps in similar tasks as a benchmark, three threshold zones of low, medium, and high are selected respectively (e.g., based on the 25th, 50th, and 75th percentiles or based on low / medium / high risk mapping), and calibrated in conjunction with the upper limit of resources and the stage target constraints; the three satisfy a progressive relationship, usually that the third-stage threshold ≥ the second-stage threshold ≥ the first-stage threshold, which is used to reflect the phased advancement logic of first strengthening the foundation, then enhancing capabilities, and finally closed-loop control.

[0050] Based on the stage affiliation and verification scenario type, verification scenario instances are matched for it according to the preset scenario matching rules and task arrangement rules: the scenario matching rules include at least retrieving the corresponding template from the scenario library according to the verification scenario type, verifying the availability of data according to the task input and output fields, selecting the sample range covering risky working conditions such as line speed adjustment / batch switching according to the working condition label, and binding at least one executable verification scenario instance (including scenario triggering conditions, collection point list, judgment indicators, thresholds and pass conditions) to each task candidate node.

[0051] Task orchestration rules include at least dependency constraints, parallel conditions, and resource constraints: tasks that satisfy dependencies within the same stage are sorted from high to low according to their difference weights, and tasks with prerequisite dependencies are processed sequentially; tasks that satisfy parallel conditions and do not conflict with resources are split into parallel work packages and each is bound to its verification scenario instance.

[0052] Subsequently, the candidate nodes of each task are linked by the numerical order of single-domain difference weight or cross-domain difference weight, and the dependency constraints and verification loop identifiers are marked on the linked edges to form a mapping relationship chain that represents and visualizes the task orchestration order, verification execution and result writing.

[0053] Finally, the prerequisite dependencies between each candidate task node are recorded as dependency constraints. Prerequisite dependencies refer to the basic conditions or upstream task set that a candidate task node must complete before execution, which are used to ensure the feasibility of task implementation and the integrity of the verification loop. The acquisition methods include: automatic extraction based on the preset task dependency rules in the metadata table and task template library (for example, field standardization is a prerequisite for indicator alignment, data collection and supplementation is a prerequisite for threshold configuration, and anomaly classification and handling orchestration is a prerequisite for closed-loop duration pressure reduction / parameter writeback), and identification of cross-system dependencies of cross-domain collaborative tasks by combining the associated edge types in the mapping relationship chain. At the same time, the system status inventory results (whether the interface has been opened, whether the field has been launched, and whether the data has been covered) can be read to convert unmet basic conditions into prerequisite tasks.

[0054] Within the same stage, the candidate task nodes are first sorted in descending order according to the single-domain difference weight or the cross-domain difference weight to obtain the initial sequence; then, dependency verification is performed. If the dependency constraint relationship of any task in the sequence is not satisfied (i.e., its predecessor task is not located before it in the sequence and has not been marked as completed), then the delay processing is performed according to the minimum delay principle: the task is removed from its current position and inserted into the earliest feasible position after all its predecessor tasks, and the associated edges connected to it are updated synchronously; for tasks with multiple dependencies, the position of the last completed predecessor task is used as the insertion anchor point.

[0055] If candidate task nodes with the same gap weight exist in the descending sort results, the roadmap generation engine extracts the causal strength (maximum or average value) of the corresponding dominant influencing factor within the task set with the same weight and performs a secondary sort, rearranging them from high to low causal strength to prioritize tasks with a more direct and significant impact on performance response. Gap convergence is defined based on the execution results of the verification scenario: the deviation quantification calculation result corresponding to the compliance judgment item is normalized and compared with the direct or indirect quantification gap to obtain the gap convergence value. If the gap convergence value is not less than the preset convergence threshold and the acceptance deviation is not greater than the preset deviation threshold, it is considered converged; otherwise, it is considered unconverged. When the single-domain / cross-domain gap weight is updated or an unconverged event is triggered, the roadmap generation engine recalculates the sorting and outputs the updated task sequence. Both the preset convergence threshold and the preset deviation threshold are versioned by the preset personnel based on changes in product model, process version, and operating conditions.

[0056] S4: After obtaining the candidate nodes of each task and their verification scenario instances after dynamic priority sorting, the task execution sequence is first generated according to the stage affiliation: the roadmap generation engine forms an initial list in each stage based on the descending result of the difference weight, and constructs a directed dependency graph in combination with the dependency constraint relationship. Tasks that do not meet the dependency are processed by extension, and then the dependency graph is topologically sorted to obtain the executable sequence of that stage; for cross-domain collaborative tasks, if they are split into preceding sub-tasks, the preceding sub-tasks are inserted into their respective stages and a dependency edge pointing to the original collaborative task is established in the dependency graph, thereby ensuring that the sequence meets the executability requirements.

[0057] Secondly, under the premise of dependency constraints, candidate task nodes are split in parallel to form stage work packages: Tasks within the same stage that are independent, do not conflict with each other in resources, and have the same or parallel execution of verification scenarios are grouped, with each group forming a stage work package, which is then bound to a corresponding verification scenario instance. The criteria for achieving the desired outcome are automatically generated by scenario matching rules: the roadmap generation engine reads the criteria, thresholds, statistical windows, and pass conditions for a scenario from the verification scenario template library, and maps them one-to-one with the expected output fields of the candidate task nodes, forming a set of criteria for achieving the desired outcome for the task, scenario, and pass conditions.

[0058] Subsequently, based on the compliance criteria, the execution results of the verification scenario instance are subjected to deviation quantification calculation: for each criterion, the deviation between the measured value and the target value is calculated (which can be the difference, relative difference, or excess amount), and the acceptance deviation is obtained by summing them according to the index weight; then the acceptance deviation is normalized and compared with the corresponding direct or indirect quantitative gap to obtain the gap convergence value. For example, the gap improvement amount / original gap amount is used as the convergence ratio, or 1-normalized deviation is used as the convergence degree, and the larger the value, the more sufficient the gap reduction.

[0059] Finally, the gap convergence value is written into a pre-set acceptance record table as the acceptance result of the phase work package. Based on the gap convergence value, the completion status (meets standards / needs improvement / blocked) is marked on the direct or coupled edges of the capability quantification gap subgraph. Simultaneously, the actual completion time, resource consumption, and verification evidence links for this phase are written back, thus forming a phased, executable, and measurable digital transformation roadmap. Furthermore, the machine learning model continuously incorporates process time-series data and quality result data, continuously updating performance response parameters, causal strength, and gap weights. This drives the automatic reordering of the roadmap's progress and monitors gap convergence values, acceptance deviations, and the frequency of abnormal operating conditions to achieve closed-loop iteration and continuous adaptation of the roadmap.

[0060] S5: like Figure 2The output judgment flowchart of the final version of the digital transformation roadmap shown illustrates the process of advancing the digital transformation roadmap. For each candidate node of a task, the verification scenario instance is bound to it. The statistical window and weight rules in its pass condition set are read, and the measured indicators of the scenario execution output are compared with the target values ​​item by item. The individual deviation (which can be the difference, relative difference, or exceeding the limit) is calculated, and then the deviation score of the task is obtained by weighting and summarizing according to the item weight. The deviation scores of multiple tasks in the stage work package are summarized according to the maximum value, weighted average, or risk priority principle to obtain the acceptance deviation quantity used to characterize the degree of acceptance of each stage task. The larger the deviation quantity, the more obvious the deviation from the pass condition.

[0061] If the gap convergence value is not less than the preset gap convergence value and the acceptance deviation is not greater than the preset acceptance deviation, it means that both criteria are met simultaneously: on the one hand, the gap convergence value indicates that the direct / indirect quantitative gap has been reduced according to the target magnitude, and the capability gap has substantially converged; on the other hand, the acceptance deviation indicates that the key indicators in the verification scenario are all within the allowable deviation range, and the results are stable and reproducible. Thus, it can be determined that the corresponding stage work package has met the standards, and it is believed that the roadmap has completed the measurable acceptance loop in the current stage. Therefore, the formed digital transformation roadmap can be used as the final version output.

[0062] If the gap convergence value is less than the preset gap convergence value or the acceptance deviation is greater than the preset acceptance deviation, then at least one type of acceptance target is not met, which means that the gap reduction is insufficient or the verification result is unstable / exceeds the limit. Continuing to use the original task arrangement will reduce the input and output and the certainty of achieving the target. Therefore, the roadmap update engine is triggered to recalculate the gap weight, task priority and verification scenario, and output the iterative version roadmap to adapt to the latest working conditions and capability shortcomings.

[0063] The roadmap update engine refers to the automatic iterative orchestration mechanism set in the digital transformation roadmap of quality management. When the status verification fails to meet the standard or abnormal working conditions are identified, it rereads the latest gap weights, dependency constraints and verification results, triggers task reordering, stage affiliation adjustment and verification scenario reconfiguration, and generates an iterative version roadmap or determines whether the conditions for outputting the final version are met.

[0064] If the gap convergence value is less than the preset gap convergence value and the acceptance deviation is greater than the preset acceptance deviation, a situation occurs where the gap has not narrowed and the indicator exceeds the limit simultaneously. This indicates a significant non-convergence anomaly or a change in operating conditions causing the solution to fail. The corresponding stage work package needs to be marked as an abnormal blocked state to forcibly trigger task reordering (completing preceding tasks, splitting / moving collaborative tasks) and verification scenario reconfiguration (changing sample operating conditions, adjusting threshold caliber, or adding key indicators) to avoid erroneous progress. The preset gap convergence value and preset acceptance deviation are both set by the preset personnel based on the historical stage task baseline distribution and stage target intensity, and are calibrated on a rolling basis as the process version and operating conditions change.

[0065] It should be added that, such as Figure 4 The iterative acceptance monitoring curves shown in the digital transformation roadmap implementation process include the gap convergence value and acceptance deviation, illustrating the trends of these two data points with the number of iterations. The gap convergence value curve gradually rises from an initial 0.31 to approximately 0.90 before stabilizing, consistently remaining below the preset convergence threshold (0.52). This indicates that while the quantitative gap, whether direct or indirect, continuously narrows during iterations, it fails to reach the preset substantial convergence target, suggesting insufficient narrowing of the capability gap. The acceptance deviation curve gradually decreases from an initial 0.90 to approximately 0.33, remaining below the preset deviation threshold (0.10) throughout. This indicates that the deviation between the actual output of the stage tasks in the verification scenario and the target value gradually decreases, and the acceptance compliance level continuously improves. However, its value remains above the allowable deviation range, indicating that the stability or reproducibility of the results has not yet fully met the requirements.

[0066] Throughout the iterative process, despite continuous optimization of the execution (resulting in a decrease in acceptance deviation), the key capability gaps did not converge to within the expected threshold, and the acceptance results consistently failed to meet stable standards. This situation corresponds to the abnormal blocking scenario in the aforementioned mechanism where the gap convergence value is less than the preset convergence threshold and the acceptance deviation is greater than the preset deviation threshold. This indicates that the current digital transformation roadmap may have become ineffective under existing conditions, necessitating the triggering of the roadmap update engine to intervene with task rescheduling, verification scenario reconfiguration, and other measures. Otherwise, a measurable acceptance loop cannot be achieved.

[0067] In a specific embodiment, for example, when analyzing lithium-ion battery electrode production process data, a strong causal relationship is identified between a temperature parameter in a certain temperature range and the electrode thickness uniformity under specific line speed and temperature range fluctuations, and this is determined to be a direct influence path. The system maps this path to the process capability assessment capability domain, calculates the quantitative difference between the current control capability value and the target value, and automatically generates a task to deploy a feedforward feedback control algorithm, while binding it to a test scenario to verify the reduction of thickness fluctuations under simulated operating conditions. This task is listed as a priority task due to its high difference weight and strong causal support. As the production line conditions change, if subsequent data analysis finds that the difference convergence of this task is insufficient or the verification exceeds the limit, the system will automatically trigger a roadmap update, adjusting the order of subsequent tasks or reconfiguring the verification parameters.

[0068] This invention utilizes machine learning to perform time-delay causal inference on multi-source time-series data, achieving objective and quantitative identification of key risk conditions and dominant influencing factors. This allows for deeper problem localization from the process level to the technological parameter level. Based on a capability gap subgraph constructed using quantified impact paths, abstract capability gaps are transformed into visualized tasks bound to specific operating conditions and parameters, generating digital measures with clearly defined verification conditions. By embedding dependency constraints and dynamic priority ranking, an executable and measurable phased roadmap is formed. More importantly, the roadmap is continuously evaluated based on the acceptance deviations and gap convergence values ​​of phased tasks during its implementation, automatically triggering iterative updates and task reordering. This ensures that the roadmap dynamically evolves with changes in production conditions, ultimately outputting a digital transformation roadmap that continuously adapts to complex and dynamic manufacturing environments, possessing closed-loop verification and self-optimization capabilities. This fundamentally improves the accuracy, reliability, and success rate of transformation planning.

[0069] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0070] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0071] In various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0072] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0073] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for generating a digital transformation roadmap for quality management based on machine learning, characterized in that, The method includes: S1. Acquire real-time data including at least process time-series data and quality result data, and perform time-delay correlation analysis and causal inference within a preset quality analysis period to quantify the impact path of each process parameter on quality management effectiveness. The time-delay correlation analysis and causal inference are adaptively executed by calling a pre-trained machine learning model; S2, based on the quantification results of the influence path, generates a sub-graph representing the capability gap of each capability domain; S3. Based on the capability gap subgraph and combined with the preset phased target constraints, a mapping relationship chain between each phase task and its corresponding verification scenario is formed, and dynamic priority sorting is performed. S4. Summarize the resource planning results after the tasks of each stage and the dynamic priority ranking to form a digital transformation roadmap, and drive the digital transformation roadmap forward based on the verification results of the tasks of each stage. During the process, S5 determines whether to trigger the roadmap update engine based on the results of status verification or abnormal operating condition identification, and then further determines whether to output the final version of the digital transformation roadmap.

2. The method for generating a quality management digital transformation roadmap based on machine learning as described in claim 1, characterized in that, S1 includes: When establishing data preprocessing and feature engineering rules, the time alignment rules of the process time series data and the label definition rules of the quality result data are written as analysis configuration fields into the metadata table corresponding to the pre-trained machine learning model. Within a preset quality analysis period, the process time series data and the quality result data are timestamped according to the time alignment rules in the metadata table, and the quality result data are subjected to deviation quantification processing according to the label definition rules, thereby obtaining performance response parameters used to uniformly characterize the effectiveness of quality management. The performance response parameters are obtained by combining at least two or more of the following: non-conformance rate, rework and scrap rate, abnormal handling closed-loop time, and closed-loop achievement rate, according to the weighting rules in the metadata table. Using the performance response parameters as target variables, a machine learning model adaptively determines the candidate time delay range and candidate feature set, performs time delay causality verification on each process parameter, and obtains the optimal lag time and the causal strength under that lag time.

3. The method for generating a quality management digital transformation roadmap based on machine learning as described in claim 1, characterized in that, S1 further includes: If the causal strength is higher than the preset significance threshold, and the causal verification result between the process parameter and the performance response parameter is marked as a direct causal relationship within the optimal lag time, then it is determined that the corresponding process parameter has a direct impact path on the quality management performance. If the causal strength is higher than the preset significance threshold, and the causal verification result between the process parameter and the performance response parameter is not marked as a direct causal relationship within the optimal lag time, and meets the preset mediating variable condition, then it is determined that the corresponding process parameter has an indirect influence path on the quality management effectiveness. Except for the above situations, if the corresponding process parameters have no impact on quality management effectiveness during the current quality analysis period, they will not be included in the impact path analysis process. The preset mediator variable conditions include: the existence of at least one mediator variable such that the causal verification result between the process parameter and the mediator variable is marked as a direct causal relationship, and the causal verification result between the mediator variable and the performance response parameter is marked as a direct causal relationship.

4. The method for generating a quality management digital transformation roadmap based on machine learning as described in claim 1, characterized in that, S2 includes: Based on the preset capability domain mapping rules in the machine learning model, the direct influence edges of the corresponding process parameters and performance response parameters in the direct influence path are assigned to the corresponding capability domain. Under the capability domain, the direct influence path is sliced ​​according to working conditions to construct a node set of the capability quantification gap subgraph. The capability domain represents a set of capabilities in the digital transformation of quality management, divided according to the life cycle or business process. The node set includes at least capability item nodes, task candidate nodes, and dominant influencing factor nodes, wherein the dominant influencing factor nodes are composed of a preset number of process parameters with the highest causal intensity in the direct influence path. Based on the causal strength, optimal lag time, and corresponding performance response parameters of the directly affecting path, calculate the current capability value of the corresponding capability domain and mark it to each node in the node set. At the same time, mark the preset target capability value to each node in the node set. The difference between the current capability value and the target capability value is recorded as the direct quantification gap, forming a capability gap label, and is bound to the dominant influence factor in the dominant influence factor node for display. The dominant influence factor includes at least the causal strength, direction of action, and optimal lag time of the corresponding process parameter. Based on the direct quantification gap and the dominant influencing factor, a task candidate node corresponding to the capability item node is generated, and the correlation between each node in the node set is used as a direct correlation edge. Simultaneously, the direct quantitative gap, causal strength, direction of action, and optimal lag time are labeled on the directly related edges to form a capability quantitative gap subgraph that characterizes and visualizes the direct influence path and supports the location of capability gaps and the generation of task priorities under the direct influence path.

5. The method for generating a quality management digital transformation roadmap based on machine learning as described in claim 1, characterized in that, S2 further includes: Based on the pre-set mediation mapping rules in the machine learning model, the mediator variables in the indirect influence path are identified, and the indirect influence edges of the corresponding process parameters are split and assigned to at least two capability domains according to business semantics. Under each capability domain, the indirect influence path is sliced ​​into working conditions to construct a capability coupling gap subgraph containing mediator variable nodes. The mediator variable nodes are composed of mediator variables in the indirect influence path that have passed the mediator validity verification. Based on the causal strength of the process parameters on the mediating variables and the causal strength of the mediating variables on the effectiveness response parameters, the cross-domain coupling effect coefficient is calculated, and indirect quantitative gap labels are formed. At the same time, the indirect quantitative gap labels are marked to the relevant capability item nodes respectively. Based on the cross-domain coupling effect coefficient and the indirect quantization gap in the indirect quantization gap label, cross-domain collaborative task candidate nodes are generated. Based on the coupling association edges therein, a coupling capability gap subgraph is formed to characterize and visualize the indirect influence path and support the capability gap location and task priority generation under the indirect influence path.

6. The method for generating a quality management digital transformation roadmap based on machine learning as described in claim 4 or 5, characterized in that, The specific formation process of the mapping relationship chain includes: The direct quantization gap marked on the direct association edge in the capability quantization gap subgraph is recorded as the single-domain gap weight, and the indirect quantization gap marked on the coupling association edge in the coupled capability gap subgraph is recorded as the cross-domain gap weight. The single-domain difference weight or the cross-domain difference weight is used to divide each task candidate node into stages according to the preset phased target constraints, thereby determining the stage affiliation and verification scenario type of each task candidate node. Based on the stage affiliation and verification scenario type, at least one verification scenario instance is matched for each candidate task node according to the preset scenario matching rules and task arrangement rules. Establish association edges for each task candidate node according to the numerical order corresponding to the single-domain difference weight or the cross-domain difference weight, and obtain a mapping relationship chain that represents and visualizes the task arrangement order and its verification closed-loop relationship at each stage. Based on the mapping relationship chain, the candidate nodes of each task are dynamically prioritized and sorted, and a phased task list and corresponding verification scenario set are output.

7. The method for generating a quality management digital transformation roadmap based on machine learning as described in claim 6, characterized in that, The dynamic priority sorting includes: The prerequisite dependencies between each candidate task node are denoted as dependency constraints. Based on the obtained single-domain difference weight or cross-domain difference weight, the task candidate nodes in the same stage are sorted in descending order. During the descending sorting process, the task candidate nodes that do not meet the dependency constraint relationship are processed to be postponed so that their sorting position is after their predecessor task. Based on the descending order of the results, if there are task candidate nodes with the same single-domain difference weight or cross-domain difference weight, then they are sorted in a secondary order according to the causal strength from high to low. When the single-domain gap weight or cross-domain gap weight is updated, or when the execution result of the verification scenario shows that the capability gap has not converged, a dynamic priority rearrangement of the task candidate nodes is triggered.

8. The method for generating a quality management digital transformation roadmap based on machine learning as described in claim 1, characterized in that, The specific process of forming the digital transformation roadmap includes: Obtain each candidate node of the task and its corresponding verification scenario instance after dynamic priority sorting, and generate a task execution sequence according to stage affiliation; Under the premise of the aforementioned dependency constraints, each candidate task node is split in parallel to form a phase work package, which is then bound to the verification scenario instance to obtain the compliance judgment items. The pass criteria entry represents a set of pass conditions that correspond one-to-one between the task candidate node and the verification scenario instance; Based on the compliance criteria, the execution results of the verification scenario instance are subjected to deviation quantification calculation, and the calculation results are normalized and compared with the corresponding direct quantification gap or indirect quantification gap to obtain the gap convergence value, which is used to characterize the degree of convergence of the direct quantification gap or indirect quantification gap. The gap convergence value is written into the pre-set acceptance record table as the acceptance result of the phase work package. Based on the gap convergence value, the completion status of each task phase is marked on the directly associated edge of the capability quantification gap subgraph or on the coupled associated edge of the coupled capability gap subgraph, thereby forming a phased, executable, and measurable digital transformation roadmap.

9. The method for generating a quality management digital transformation roadmap based on machine learning as described in claim 1, characterized in that, S5 includes: During the implementation of the digital transformation roadmap, based on the execution results of the verification scenario instances according to the compliance judgment items, an acceptance deviation quantity is obtained to characterize the degree of acceptance of tasks at each stage. If the gap convergence value is not less than the preset convergence threshold and the acceptance deviation is not greater than the preset deviation threshold, then the corresponding stage work package is determined to meet the standard, and the resulting digital transformation roadmap is directly output as the final version. The preset convergence threshold and preset deviation threshold represent the dual threshold constraints corresponding to the acceptance and qualification determination of each stage of the work package.

10. The method for generating a quality management digital transformation roadmap based on machine learning as described in claim 9, characterized in that, The S5 also includes: If the gap convergence value is less than the preset convergence threshold, or the acceptance deviation is greater than the preset deviation threshold, the update engine of the digital transformation roadmap is triggered, and an iterative version of the digital transformation roadmap is output. If the gap convergence value is less than the preset convergence threshold and the acceptance deviation is greater than the preset deviation threshold, then it is determined that there is a non-convergence anomaly in the corresponding stage work package, and it is marked as an abnormal blocking state to trigger task reordering and verification scenario reconfiguration.

Images