Textile production quality closed-loop management method and system based on large model agent

Abstract

The application discloses a kind of textile production quality closed-loop control method and system based on large model intelligent agent.The method first acquires the multimodal data including image, time series parameter and design document in real time, maps it to unified feature space by space-time anchor point projection algorithm to realize semantic alignment;When monitoring quality anomaly, AI intelligent agent hub dynamically calls cross-system research and development specification and historical cases, uses the large model reasoning engine embedded in the physical mechanism of textile to construct heuristic reasoning chain, realizes the deep cause and effect tracing from phenomenon to research and development design source;Subsequently, the system automatically generates the multi-dimensional decision-making strategy of production end short-term correction and research and development end long-term optimization, and executes through industrial communication interface, finally according to the execution gain effect, the effective cause and effect chain is automatically deposited to knowledge base.The application breaks the data barrier between research and development and production, significantly improves the accuracy of root cause analysis and shortens the new product research and development iteration cycle.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of intelligent textile manufacturing and industrial artificial intelligence, specifically relating to a closed-loop control method and system for textile production quality based on a large-scale intelligent model. This invention utilizes multimodal large-scale model technology to deeply correlate R&D design specifications, production process parameters, historical quality problem records, and real-time detection data, enabling automatic tracing of the root causes of textile quality anomalies and intelligent generation of targeted improvement strategies. Background Technology

[0002] In actual production within the textile industry (such as weaving and dyeing), quality problems often exhibit characteristics of being "cross-stage, multi-factorial, and latent." Traditional quality management models suffer from the following pain points: (1) Disconnect between design and production: There is a disconnect between the process design specifications (such as yarn twist setting and dye formulation logic) in the R&D stage and the actual execution parameters in the production site. When quality abnormalities occur, the existing monitoring system can only report the parameter fluctuations at the production end, making it difficult to trace whether the problem is due to insufficient design margin or ambiguous specification definitions.

[0003] (2) Data silos from multiple sources: R&D documents (Word / PDF), process sheets (Excel), equipment operation logs (time-series data), and fabric defect images (visual data) are scattered across different systems such as PLM, MES, and QMS. It is difficult for humans to comprehensively analyze cross-system and cross-modal data in a short period of time.

[0004] (3) Attribution depends on personal experience: Traditional root cause analysis (RCA) relies on the personal experience of QE engineers and lacks systematic access to historical R&D problem records (Lesson Learned) and 8D reports, leading to the recurrence of similar design defects or process errors.

[0005] (4) Lack of deep logical reasoning: Existing AI quality inspection solutions focus on the classification of surface defects (such as identifying defect types) and lack the ability to make deep causal reasoning by combining “R&D design intent” with “textile physical / chemical mechanism”.

[0006] Therefore, there is an urgent need for an intelligent system that can integrate data from the entire R&D and production chain and possess deep logical reasoning and closed-loop management capabilities. Summary of the Invention

[0007] This invention primarily addresses the technical problems of design-production disconnect, attribution difficulties, and lack of closed-loop optimization in existing technologies, providing a closed-loop control method and system for textile production quality based on a large-scale intelligent agent model.

[0008] The present invention addresses the aforementioned technical problems primarily through the following technical solution: a closed-loop control method for textile production quality based on a large-scale intelligent agent, comprising the following steps: S1: Multimodal data acquisition and cross-modal feature alignment: Real-time acquisition of visual modal data V, temporal modal data T, and textual modal data S; semantic alignment is achieved by projecting the multimodal data onto a unified feature space Z using a spatiotemporal anchor projection algorithm. S2: Deep causal source analysis based on R&D logic: When a quality anomaly occurs, the AI ​​intelligent agent central system dynamically calls the R&D design specifications and historical problem database of the related system according to the anomaly characteristics. It uses a large model combined with textile physical mechanisms to construct a heuristic reasoning chain to locate the root cause at the R&D or production end. This step is based on the Milvus vector database to build a multimodal textile process knowledge base to achieve vectorized knowledge storage and millisecond-level retrieval. S3: Multidimensional intelligent decision-making strategy generation and execution: Based on the root cause, it automatically generates short-term strategies that include production-side parameter adjustment instructions and long-term strategies that include R&D-side specification revision suggestions, and sends them to the Manufacturing Execution System (MOM / MES) or the Data Acquisition and Monitoring Control System (SCADA) through the API interface to implement closed-loop control. S4: Closed-loop effect verification and knowledge base self-updating: Monitor the production quality trend after the execution of decision instructions. If the verification is effective, the causal chain consisting of "abnormal phenomenon - root cause - decision strategy - execution effect" will be vectorized and stored in the historical problem database.

[0009] This solution is based on large model reasoning capabilities and Retrieval Enhanced Generation (RAG) technology, deeply integrating standard ontology and physical mechanism engine in the textile field. It performs intent classification and multimodal information matching for abnormal features through multi-level logic trees, and dynamically retrieves similar scenario solutions using the dual hard constraints of "mandatory standard specifications" and "physical mechanism laws". Furthermore, it constructs a heuristic hybrid reasoning chain with quantitative verification mechanism, and implements boundary quantification judgment by combining counterfactual causal inference, realizing in-depth correlation analysis of quality anomalies from appearance to essence, and finally accurately locating the root cause of insufficient design margin in R&D or execution deviation in production.

[0010] The core of this solution lies in breaking away from the limitations of traditional methods that only look for causes from the production end, and introducing R&D design data as the core variable for attribution. Through multimodal alignment in S1, a unified foundation is provided for subsequent reasoning; through causal tracing in S2, a penetrating analysis from phenomenon to essence is achieved; and through S3 and S4, a closed-loop upgrade from problem discovery to problem solving and prevention is realized.

[0011] Preferably, the process of constructing the unified feature space Z in step S1 includes: S11: Utilize a deep learning encoder to extract visual feature vectors V based on visual modal data V, temporal modal data T, and text modal data S, respectively. i Temporal feature vector T j and text semantic vector S k ; S12: Using the production work order number and timestamp as spatiotemporal anchors, the visual feature vector V at a specific moment is... i The temporal feature vector T within a preset time period before and after this moment j To couple; S13: The coupled feature vector V is coupled through a unified projection layer. i T j and the corresponding text semantic vector S k Mapped to space Z, and using the contrastive loss function L=-log(exp(sim(z)) i ,z j ) / τ) / Σexp(sim(z i ,z n ) / τ)) constrain positive sample pairs of features z i ,z j The cosine distance in the space Z is calculated by minimizing the loss function L to shorten the distance between positive samples and features in the space, and to increase the distance between negative samples and features; where τ is a preset temperature scalar parameter used to control the concentration of the probability distribution, usually set between 0.07 and 0.2.

[0012] By minimizing the contrastive loss function L, the system can learn the inherent semantic relationships between cross-modal data. For example, it can shorten the spatial distance between color difference image features and abnormal features of dye vat heating curves, thereby automatically identifying feature pairs with co-occurrence patterns in massive amounts of data and improving alignment accuracy.

[0013] The dynamic construction rules for comparing positive and negative samples in the loss function: For the contrast loss function in step S13, the system defines sample sampling rules based on the physical properties of textiles: (1) Positive samples (z) i , z j The delayed coupling construction is as follows: Due to the physical delay in the transfer of fabric from the loom / dyeing vat processing point to the AOI visual inspection station, the "preset duration" set by the system is not a fixed value, but is dynamically calculated based on the formula Δt=L / v (where L is the physical distance from the processing point to the inspection point, and v is the current linear velocity of the fabric transfer). The visual defect feature at a specific time t only forms a positive sample pair with the temporal features within the time window [t-Δt-δ, t-Δt+δ], where δ is the allowable tolerance for fluctuations in operating conditions.

[0014] (2) Negative samples (z)n Hard Example Mining: Negative Sample Temporal Features z n Instead of random sampling, samples are forcibly extracted from normal production periods of the same production order that are judged as "first-class products" (without visual defects). By introducing such boundary-hard-to-bearing samples, the contrastive learning model is forced to learn more accurately the essential difference between "compliance parameter fluctuations" and "abnormal parameter drifts" in the feature space.

[0015] Preferably, in step S2, the AI ​​agent central unit performs recursive link tracing: sequentially retrieving real-time parameters from the production layer, retrieving product design intent and process window settings from the R&D layer, and retrieving 8D report cases with similar feature vectors from the historical layer.

[0016] This dynamic logic of on-demand data retrieval avoids the waste of computing resources caused by full data capture. Through priority settings, the system can quickly identify the most relevant system interfaces based on the defect morphology (visual modality). For example, for a "weft shrinkage" defect, the system will automatically increase the weight of the loom tension sensor data collection and simultaneously backtrack the yarn design specifications in the PLM.

[0017] Preferably, step S2 includes R&D design specification boundary verification: if the actual operating parameters are within the safe process window of the R&D design but quality abnormalities still occur, then the design margin detection is triggered to calculate the material sensitivity under the current environmental parameters and deduce whether the process window range set in the original R&D did not take into account the impact of raw material batch fluctuations.

[0018] This involves in-depth analysis of inherent quality risks. This step can identify hidden defects where parameters appear normal but problems still occur. For example, it might determine that the cause was that the temperature control window set during the R&D phase did not account for the impact of raw material batch fluctuations, thus driving optimization in the design phase.

[0019] Preferably, the specific implementation of step S2 includes: S21: Abnormal Feature Structuring and Multi-level Intent Recognition: Construct a standard ontology for the textile field that includes standard document entities, standard clause entities, and standard hierarchical entities. After the large model transforms multimodal data into a structured query language, it uses a multi-level logic tree of appearance-state-attribution to classify and determine intent. S22: Dynamic retrieval of knowledge base: Retrieve historical cases and R&D design specifications from the vector database, introduce a dual constraint filter of mandatory standard specifications and textile physical mechanism, calculate the error rate between feature values ​​and theoretical values ​​for quantitative elimination, and dynamically re-rank based on similarity, timeliness and physical fit. S23: Heuristic Hybrid Inference Chain and Double-Key Joint Backtracking: The large model generates potential root cause hypotheses containing physical formula IDs and parameters in parallel. After quantitative verification by an external physical calculation module, the comprehensive confidence level is calculated by combining physical consistency and evidence chain matching degree for pruning. For the verified hypotheses, the evidence chain is traced back using a double-key joint index of production work order number and timestamp. S24: Boundary Quantification Judgment and Counterfactual Causal Attribution: Based on multi-dimensional indicators such as equipment health index and operation compliance rate, define the responsibility boundaries between the R&D end and the production end, call the causal prediction model to perform counterfactual inference, and output a structured traceability report after being verified by the physical consistency checker.

[0020] The above-mentioned preferred scheme effectively overcomes the problem of illusions in industrial mechanism reasoning caused by the introduction of dual hard constraints of standard ontology and physical mechanism, and combines counterfactual causal inference. It realizes the quantitative and accurate determination of the responsibility boundary for tracing quality anomalies, and ensures the safe and executable nature of the final decision-making instructions in the industrial field.

[0021] Preferably, the textile industry standard ontology comprises a four-layer architecture: The first layer is the entity concept layer, which includes basic entities such as raw materials, equipment, process parameters, quality defects, and systems, as well as dedicated entities such as standard documents, standard clauses, and standard classifications. The second layer is the standard metadata layer, which is used to establish a two-way mapping relationship between the basic entities and the standard clause entities; The third layer is the semantic relationship layer, which, in addition to causal, subordinate and constraint relationships, defines standard-specific related predicates including business entities being bound by standard clauses, abnormal phenomena violating standard limits, and standard clauses defining product quality levels. The fourth layer is the constraint axiom layer, which includes the mandatory compliance axiom, the hierarchical threshold axiom, the process tolerance axiom, and the standard conflict arbitration axiom, serving as the dual constraint filter base for large model reasoning.

[0022] This preferred solution innovatively constructs a highly structured data foundation, enabling abstract textile quality defects and process parameters to be subject to quantifiable national or industry standard constraints. This fundamentally provides a deterministic knowledge basis for the logical reasoning of large models and avoids cognitive biases during cross-domain reasoning.

[0023] Preferably, step S3 includes instruction conflict arbitration: when the parameter instructions of the short-term strategy conflict with the hard threshold of the R&D design specification, the AI ​​agent arbitrates based on the principle of quality gain priority and generates a temporary deviation permission record to be synchronized to the quality management system (QMS).

[0024] In real-world industrial scenarios, adjustments required on the production floor may conflict with R&D standards. The AI ​​agent calculates the quality gain value of the arbitration result, allowing reasonable process deviations while ensuring product quality. It also ensures these deviations are recordable and traceable in the QMS, preventing unauthorized temporary changes from becoming long-term issues.

[0025] A closed-loop control system for textile production quality based on a large-scale intelligent agent model, used to implement the method described above, the system comprising: Multimodal perception module: used to collect and align visual modal data, temporal modal data, and text modal data in real time; AI intelligent agent central module: used to respond to quality anomalies, dynamically call cross-system data and generate source tracing conclusions using a large model inference engine embedded with textile physical mechanisms; Collaborative decision-making and distribution module: used to generate multi-dimensional decision strategy packages and distribute instructions to PLM, MOM / MES and SCADA systems through industrial communication interfaces; Closed-loop evaluation and evolution module: used to evaluate decision gains and automatically deposit effective causal chains into the knowledge base.

[0026] Through the organic combination of four functional modules, a digital brain with the capabilities of "perception-reasoning-decision-evolution" is formed.

[0027] Preferably, the AI ​​intelligent agent central module is connected to the Product Lifecycle Management System (PLM), Manufacturing Operations Management System (MOM), Quality Management System (QMS), Laboratory Information Management System (LIMS), and Enterprise Resource Planning System (ERP) via API interfaces to achieve heterogeneous integration of R&D, production, testing, and material data.

[0028] This architecture completely breaks down information silos within the factory. By integrating physical property test data from LIMS and supplier batch data from ERP, the system can extend the depth of attribution to the raw material end.

[0029] Preferably, the large model inference engine adopts a private deployment mode and is fine-tuned for knowledge in the textile industry, enabling it to have the logical deduction ability to understand the impact of yarn twist, weaving tension and dye physical properties on quality.

[0030] By deploying the model privately within the enterprise's local area network and fine-tuning it using domain-specific corpora (such as the "Dye Property Handbook" and "Common Defect Atlas"), the large model is equipped with expert-level textile logic reasoning capabilities.

[0031] The substantial effects of this invention are: (1) Accurately locate the root cause: It realizes the traceability from the simple equipment / operation level to the design / specification level, and completely solves the repetitive quality problems.

[0032] (2) Knowledge asset accumulation: Transform the implicit experience that was originally scattered in the minds of engineers into a digital knowledge graph, and continuously optimize design specifications through the system's memory function.

[0033] (3) Shorten the R&D cycle: Through automatic feedback of production data, the maturity and optimization of the process window in the New Product Introduction (NPI) stage are accelerated.

[0034] (4) Reduce trial and error costs: AI simulation and historical comparison can be used to warn of potential design defects in the early stages of mass production, thereby reducing scrap losses. Attached Figure Description

[0035] Figure 1 This is a flowchart of a closed-loop control method for textile production quality based on a large-scale intelligent agent, according to the present invention. Detailed Implementation

[0036] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0037] Example: In-depth source tracing and closed-loop management of textile production anomalies This embodiment demonstrates how the system handles complex quality anomalies in textile production, achieving efficient data collection from the sensing side to in-depth traceability and closed-loop management of the core AI agent.

[0038] A closed-loop control method for textile production quality based on a large-scale intelligent agent, such as Figure 1 As shown, it includes the following steps: S1: Multimodal data acquisition and cross-modal feature alignment The system acquires visual modal data V, temporal modal data T, and textual modal data S in real time. To ensure the effective acquisition of defect features in the visual modal data V, a high-speed, high-resolution imaging technology assisted by multi-spectral illumination is deployed at the visual acquisition end, solving the problem that key defects on the surface of textiles cannot be fully displayed by a single spectral light source. The specific implementation method is as follows: Multi-spectral adaptive light source illumination technology: The system is equipped with an automated control device including a light source, a rotary motor, and a slide rail. The system establishes a correlation between the product's surface texture and defect characteristics and the light source parameters, and automatically matches the light source spectrum and brightness according to the characteristics of the textile fabric, achieving optimal adaptive illumination adjustment.

[0039] Variable-angle time-division stroboscopic imaging technology: A variable-angle time-division stroboscopic imaging system integrating reflection and transmission imaging was designed specifically for the characteristics of textile fabric processes. This system employs a shaped roller rotation control device; the shaped roller structure ensures that the imaging position remains constant regardless of the rotation angle. The system develops a multi-source image fusion algorithm to simultaneously acquire and fuse reflected and transmitted light images with defects, effectively capturing special defects such as abrasions, paper scars, and bright lines that are difficult to photograph from a conventional vertical angle.

[0040] Cross-modal feature alignment: The system utilizes a deep learning encoder to extract visual feature vector Vi, temporal feature vector Tj, and textual semantic vector Sk. Using the production work order number and timestamp as spatiotemporal anchors, the visual feature vector Vi at a specific moment is coupled with the temporal feature vector Tj within a preset time period. Subsequently, it is mapped to the space Z through a unified projection layer and the contrastive loss function L=-log(exp(sim(z)) is applied. i ,z j ) / τ) / Σexp(sim(z i ,z n The cosine distance of positive sample features in space Z is constrained by (τ) to achieve semantic alignment. i With z j Let z represent positive sample pairs of features in a unified feature space Z. By minimizing the loss function, the algorithm reduces the distance between these positive sample pairs of features in the space, making them closer together. n The negative sample features are represented in the comparison, and τ represents the preset temperature scalar parameter.

[0041] Database Construction: Multimodal Textile Standard Ontology and Knowledge Database As the core data foundation for subsequent causal tracing, the system constructs a standard ontology and knowledge database with a four-layer architecture during the initialization phase: The first layer (entity concept layer) is divided into five basic entities: raw materials, equipment, process parameters, quality defects, and systems / documents. New standard-specific entities are added, including: standard document entities with standard numbers (such as GB 18401, FZ / T14004) and level attributes; standard clause entities with clause types and quantification threshold attributes; and standard graded entities (such as first-class products, qualified products) bound to corresponding defect allowable upper limits.

[0042] The second layer (standard metadata layer) establishes a two-way mapping relationship between business entities and standard clauses, enabling abstract defects to have quantifiable constraints. For example, the defect entity "color difference" is mapped to the standard clause "GB / T 250 ΔE≤3 (first-class product)".

[0043] The third layer (semantic relation layer): In addition to causal, subordinate, and constraint relationships, six types of standard-specific relational predicates are defined: hasStandardConstraint, standardClassify, quoteStandard, violateStandard, standardOptimize, and standardDetectionRule. For example, generating a triple: ΔE=4.2 (outlier) → violateStandard → GB / T250_First-class color difference limit.

[0044] The fourth layer (constraint axiom layer): solidifies industry constraints and physical laws as the underlying reasoning axioms, including mandatory compliance axioms (such as pH value exceeding the limit directly locking R&D defects), graded threshold axioms, process tolerance axioms, and standard conflict arbitration axioms.

[0045] S2: Deep causal tracing based on the integration of RAG, physical mechanisms, and standard specifications When a quality anomaly occurs, the AI ​​agent's central processing unit executes the following detailed sub-steps: S21: Structuring of Abnormal Features and Multi-level Intent Recognition The establishment of ontology mapping relationships includes explicit definition of expert rules (such as defining triples <fabric crossbar, caused_by, warp tension periodic fluctuation>, with additional attributes {period: 0.5-2s, amplitude: >5cN}), implicit association through data mining (using the Apriori algorithm to mine data from MES, with confidence >0.8 for inclusion in the database), and semi-automatic extraction of large models.

[0046] The system adopts a four-level logic tree routing of "appearance-state-attribution": L1 layer (signal validity) checks the sensor signal-to-noise ratio (SNR), and if it is less than the threshold, it is determined to be a fault in the sensing layer equipment; L2 layer (abnormal mode characteristics) if the tension fluctuation frequency is equal to the motor rotation frequency and the current harmonics are abnormal, it is determined to be a fault in the execution layer equipment; L3 layer (operating condition correlation) if the actual value deviates from the set value and the equipment health is greater than 0.9 and the humidity change rate exceeds the limit, it is determined to be a drift in process parameters; L4 layer (design boundary check) if the parameters are within the safety window and the equipment is normal but the quality is NG, it is determined to be insufficient design margin.

[0047] The system automatically extracts the context to construct a JSON-formatted summary and interacts with it using a Prompt template. Example Prompt template: "You are a textile quality diagnostic expert. Based on the following multimodal context [insert JSON data] and ontology hints, please: 1) Confirm the anomaly type; 2) List the 3 most likely root cause hypotheses; 3) Specify the physical verification formula ID to be invoked for each hypothesis. Output strictly follows the JSON Schema."

[0048] S22: Dynamic Retrieval of Multimodal Textile Technology Knowledge Base under Dual Constraints Anomaly features are mapped to a vector database, and a dual hard filter combining national and industry standards and physical mechanisms is introduced. Large-scale models extract textual numerical values ​​and substitute them into physical formulas for verification. Built-in formulas include: 1) Formula for yarn friction and tension transmission: T out =T in ·e μθ(1+k·v) , In the formula, μ is the coefficient of friction, θ is the encirclement angle, v is the velocity, and k is the velocity sensitivity factor. This is used to verify the hypothesis that "wear / contamination of the yarn guide causes abnormal tension."

[0049] 2) Loom shedding dynamics equation: F max =m·a peak / η+F friction +F elastic (ε), where m is the mass of the heald frame, a peak η represents peak acceleration, and η represents transmission efficiency. This is used to verify that "abnormal heald frame movement leads to warp yarn breakage / crossbar breakage."

[0050] 3) Fiber moisture regain-strength relationship model: σ break (MR)=σ0·(1+α·(MR-MR O The formula is MR, where MR is the measured moisture regain and α is the material coefficient. It is used to verify the hypothesis that "changes in environmental humidity lead to insufficient yarn strength".

[0051] Error% = |V actual -V theoretical | / V theoretical ×100%. An error rate of ≤10% indicates a high degree of physical fit, 10% to 25% indicates a basic fit (with tolerance allowed), and >25% is considered physically impossible and rejected.

[0052] Recall fragments are ranked using a multi-weighted formula: Score final =w s ·S sim +w t ·D(t)+w p ·P phy Where D(t) = e -λ·Δd This is the time-dependent decay function, where Δd is the number of days since the last occurrence, and λ is the decay constant (default 0.005, i.e., half-life ≈ 139 days). λ can be dynamically increased after equipment modification or process change. phy =max(0,100-β·Error%) is the physical fit percentage score.

[0053] S23: Heuristic Hybrid Reasoning Chain and Two-Key Joint Backtracking The large model serves only as a JSON generator, outputting something like {"hypothesis": "Tension increase due to yarn guide wear", "formula_id":"YARN_FRICTION_V2", "parameters": {"T_in": 150, "theta": 1.57, "v": 800, "mu_assumed": 0.35}}. The physical calculations are handled by an external engine. The number of hypothesis branches generated in parallel, N, is a dynamically calculated value: N = min(N... max ,max(N min ,⌈C entropy ⋅ρ graph ⌉)). C entropy It is the normalized cross-entropy (0-1) of the multimodal anomalous signal, reflecting the anomaly complexity, ρ graph N is the neighbor density of the current node in the normalized anomaly association graph. min =2, N max =8. Example: Single tension anomaly (C=0.3, ρ=0.5) → N=2; simultaneous tension + humidity + motor anomaly (C=0.8, ρ=0.9) → N=6.

[0054] Calculate the overall confidence score for each node: Conf = α·P phy +(1-α)·S evidence S evidence The evidence chain backtracking match degree is (0-100). The pruning trigger condition is: P phy <30 (error >17.5%), or Conf <40, or return the PHYSICALLY_IMPOSSIBLE flag, with a maximum of Top-3 branches retained per round of ToT.

[0055] A double-key composite index is used with "production work order number" as the primary key and "time stamp ± 30s sliding window" as the time key. This index allows for reverse tracing within the database partition, preventing cross-batch data contamination.

[0056] The source of quantization for dynamic parameters in heuristic inference chains: The formula for calculating the number of root cause hypothesis branches N in step S23 is N = min(N max ,max(N min ,⌈C entropy ⋅ρ graph ⌉))middle: (1) Multimodal anomaly complexity C entropyThe quantification is as follows: its value is obtained by weighted fusion of the information entropy of the visual modality and the anomaly score of the temporal modality. Specifically, the gray-level co-occurrence matrix of the defect region is extracted to calculate the visual information entropy, while the isolated forest algorithm is used to calculate the anomaly score of the temporal parameters deviating from the normal baseline. The two are normalized and then weighted and summed to represent the comprehensive complexity of the current quality anomaly.

[0057] (2) Neighbor density ρ graph Quantification: Characterizes the physical mechanism complexity of the current anomaly's association in the knowledge graph. Its value is the ratio of the actual number of connected edges within the subgraph activated by the current anomaly feature in the graph database to the maximum possible number of connected edges (the total number of edges in the complete graph). Higher complexity and graph density result in more dynamically allocated inference branches N, avoiding the omission of deep-seated root causes.

[0058] S24: Boundary Quantification Determination and Counterfactual Causal Attribution The R&D responsibility determination is triggered only if the equipment health index is ≥0.90, the operation compliance rate is ≥98%, the environmental control deviation is within the workshop standard and the parameters are within the design window, and the quality is still unqualified.

[0059] The system invokes a causal prediction model trained on historical time-series data for extrapolation. The input structure is {current working condition feature vector, intervention variable ΔX, intervention magnitude}, and the output structure is {predicted quality index value, improvement probability P_improve, confidence interval}. Double physical verification is implemented: the pre-constraint intervention value ΔX must be within the physically feasible region; the post-verification prediction result must pass a physical consistency checker. If the confidence level < 0.6, it is forcibly marked as "requires manual verification" as a fallback against illusion. The final output is a JSON object containing {root cause category, associated evidence ID, physical mechanism basis, confidence level, and suggested strategy type} sent to the S3 decision module.

[0060] S3: Multidimensional Intelligent Decision-Making Strategy Generation and Execution Based on the root cause of the S2 output, the system automatically generates a short-term strategy that includes production-side parameter adjustment instructions and a long-term strategy that includes R&D-side specification revision suggestions, and distributes them to the MOM / MES or SCADA system.

[0061] When on-site adjustment strategies conflict with R&D design specifications, the standard conflict arbitration axiom is triggered: if it is a mandatory national standard, it cannot be violated, and the system rejects it; if it is an industry recommended standard or enterprise SOP, arbitration is conducted based on the principle of prioritizing quality gain, generating a temporary deviation permission record and synchronizing it to QMS. If the attribution points to a standard defect, the system retrieves the `standardOptimize` predicate and generates a long-term change proposal in PLM.

[0062] S4: Closed-loop effect verification and knowledge base self-updating Monitor production quality trends after the execution of decision-making instructions. If verified to be effective, the complete causal chain consisting of "abnormal phenomenon - root cause - decision-making strategy - execution effect" will be vectorized and stored in the historical problem database to realize risk warning in the new product introduction (NPI) process.

[0063] Quantitative determination mechanism for closed-loop effect verification: In step S4, the system abandons subjective human evaluation and sets dual quantitative thresholds to determine "valid verification" and trigger self-update of the knowledge base: After the decision command is issued and takes effect, the system automatically extracts the subsequent T-length production window (such as the subsequent 1000 meters of fabric or the subsequent 2-hour operation cycle) as the verification and evaluation period.

[0064] Judgment Criterion 1 (Target Improvement): The frequency of the specific defect currently being traced has decreased by more than the first threshold (e.g., 80%) compared to the period when the anomaly occurred. Judgment Condition 2 (Guardrail Indicators): The degradation of the product's basic quality indicators (such as overall strength loss and average color difference) compared to the standard baseline must not exceed the second threshold (e.g., 2%).

[0065] The system will mark the verification status as True and package the full-link data into a successful experience, vectorizing it and writing it into the multimodal textile process knowledge base, only if both of the above conditions are met simultaneously.

[0066] Specific implementation methods of model training and fine-tuning mechanism: (1) Large Model Domain Fine-tuning: The system uses LoRA parameter efficient fine-tuning technology to conduct textile domain adaptive training on the base large model. The fine-tuning dataset is constructed using the instruction follow format. The system automatically parses historical 8D reports, standard specification documents and equipment maintenance logs, extracts the "abnormal phenomenon-mechanism analysis-root cause conclusion" triplet, and transforms it into a question-and-answer pair corpus in the form of {instruction: "Analyze the root cause of tension abnormality", input: "warp tension fluctuation +20%, workshop humidity decreases by 10%", output: "Based on fiber moisture regain model..."}, so that the large model can master the logical deduction paradigm of the textile domain.

[0067] (2) Counterfactual causal prediction model training: The causal prediction model in step S24 is constructed based on the DoWhy framework and Causal-LSTM. The training time series data uses process parameters within the historical production cycle as intervention variables (X) and the final fabric quality score as the outcome variable (Y). The system defines environmental temperature and humidity fluctuations and raw material batch attributes as confounders (Z). During the model training phase, the backdoor criterion is used to correct the confounders, blocking the spurious correlation path from X to Y, thereby ensuring that the "probability of improvement" output by the model has true causal inference significance, rather than simple statistical correlation.

[0068] This invention, a closed-loop control system for textile production quality based on a large-scale intelligent agent, adopts a three-layer physical architecture and includes four core modules: Multimodal sensing module: A variable angle time-division stroboscopic imaging device integrating the aforementioned irregularly shaped rollers, multi-spectral light sources and slide rails, to achieve high-resolution acquisition of visual features; deploying an edge gateway to connect with the underlying PLC device through the OPC UA protocol to acquire time-series data in real time.

[0069] AI Intelligent Agent Central Module: It adopts a private deployment mode to load a large model of the textile industry and a physical mechanism calculation engine; it can be heterogeneously integrated with PLM, MOM, QMS, LIMS and ERP systems through API interfaces.

[0070] Collaborative decision-making and distribution module: Receives structured reports and distributes parameter commands to the corresponding control systems.

[0071] Closed-loop evaluation and evolution module: assesses the actual gain of decisions and deposits the successful causal chain into a vector knowledge graph.

[0072] On a continuous dyeing and printing production line, AOI (Automated Optical Inspection) equipment captures a color deviation signal (visual modality V) on the fabric surface. At this time, the AI ​​agent's central processing unit performs the following operations: Feature extraction: The color moments, histogram, and texture feature vector V of the color difference image are extracted using a pre-trained visual encoder. i .

[0073] Spatiotemporal coupling: The system uses the current production work order number to retrieve the dyeing vat temperature, pressure, pH value, and auxiliary agent addition amount (time mode T) within 5 minutes before and after that moment from the SCADA system, and extracts the fluctuation feature vector T. j .

[0074] Spatial projection: V i T jThe standard process specification (text modality S) in the PLM system is projected onto a unified feature space Z. By calculating cosine similarity, it is found that the current color difference feature has a high semantic correlation with "temperature fluctuation in the high-temperature range" and "sensitivity of additive formulation".

[0075] AI agent initiates recursive link tracing: Standardization verification: Comparison revealed that the actual temperature control curve was completely within the safety window set by the R&D team (e.g., 100°C ± 2°C).

[0076] In-depth mechanism derivation: The large model calls the built-in "dye fixation rate and temperature evolution model" and combines it with the moisture content data of this batch of greige fabric fed back by the LIMS system.

[0077] Reasoning chain: Observe color difference → production parameters are normal → search historical 8D reports and find similar batch fluctuation records → combine physical mechanism to deduce the conclusion: due to excessive residue in the pretreatment of this batch of fabric, the dye's sensitivity to temperature is increased, and the ±2°C process window originally set by R&D is too wide under the current material condition, which is insufficient in the design specification margin.

[0078] In conducting causal attribution analysis, this invention utilizes Retrieval Enhanced Generation (RAG) technology to perform semantic extraction and logical alignment of multi-source unstructured documents: Risk guidance for quality planning documents: The system automatically parses the PFMEA (Process Failure Mode and Effects Analysis) and Control Plan to extract defined potential risk points and key control points. When abnormal fluctuations occur in production, the AI ​​agent prioritizes identifying high-risk processes based on the failure probability levels (S / O / D) in the PFMEA, guiding the search direction of the inference chain.

[0079] Mechanism support for knowledge manual documents: The system has built a professional knowledge base including the "Common Defect Graphs and Causes Manual", "Dye Property Manual", and "Equipment Fault Code Table". The large model uses the physical / chemical mechanisms in these documents (such as dye sublimation risk and fiber friction coefficient changes) to help understand the deep correlation between defect morphology and process parameters, and establish a chain of evidence for tracing the source.

[0080] Short-term emergency response: AI generates instructions to adjust the heating slope of the subsequent dyeing vat in real time through the MES system, reducing it from 2°C / min to 1.5°C / min, and increasing the amount of fixing agent by 1%.

[0081] Long-term strategy: The system automatically generates a "Design Specification Change Proposal" in PLM, which suggests tightening the dyeing temperature control tolerance from ±2°C to ±1°C for specific specifications of greige fabric, and adding a mandatory moisture content test before feeding.

[0082] 1. Dynamic invocation logic and conflict arbitration When the loom experiences batches of "weft shrinkage", the AI ​​agent detects a conflict between the short-term adjustment strategy (increasing weft tension) and the "yarn strength limit" in the R&D SOP.

[0083] Conflict arbitration: AI assessment suggests that if the original tension design is forcibly maintained, the defect rate will remain above 5%; if deviation from the permitted design is implemented, the risk of damage due to force increases by 0.5%, but the defect rate can be reduced to 0.1%.

[0084] Result: The AI ​​prioritizes quality gain, issues temporary deviation instructions, and simultaneously creates non-conforming product review records in the QMS system.

[0085] 2. Knowledge base self-updating (S4) After verifying that production has stabilized, the system performs the following operations: Vectorized storage: Transform the successful experience of "low twist yarn + 85% high humidity → increase weft tension by 10%" into a standard case.

[0086] End-to-end closed loop: This case is stored in the "R&D Problem Record Library". When the subsequent New Product Introduction (NPI) process involves yarn of similar specifications, the system will automatically pop up the risk warning during the design stage and recommend a better process window, realizing the transformation from "firefighting" to "fire prevention".

[0087] To achieve the above functions, the textile production quality closed-loop control system based on large-model intelligent agents adopts the following deployment architecture: Edge layer: Deploy edge gateways in the weaving workshop to connect with the loom PLC via the OPC UA protocol and collect tension and speed data in real time down to the second level.

[0088] Computation layer: A large model inference server (configured with 4×80GB GPU resources) is privately deployed in the enterprise data center to support the multimodal encoder and causal inference engine.

[0089] Application layer: Through API interfaces, it integrates with existing PLM, MES, QMS, and ERP systems to build a unified AI intelligent agent control center, enabling cross-system command scheduling and data flow.

[0090] The system described in this invention achieves bidirectional data flow and command linkage across systems through a central control module: High-frequency data synchronization: The system collects second-level operating data from underlying equipment (such as looms and dyeing vats) in real time via the SCADA / DCS interface. This high-frequency time-series data is used to analyze the subtle impact of instantaneous fluctuations in process parameters on product quality.

[0091] Cross-system collaborative execution: Collaborating with PLM: The AI ​​agent not only reads the product design intent, but also automatically writes back the verified and effective process optimization suggestions and draft specification revisions to the PLM system's change process after the traceability is completed.

[0092] Collaboration with QMS: The system can read quality inspection results and historical 8D reports, and trigger automated 8D processes during the decision-making stage, directly writing the root cause analysis conclusions generated by AI into the QMS report draft.

[0093] The core of the closed-loop management and control method based on large-model intelligent agents described in this invention lies in deeply embedding AI reasoning capabilities into the core business processes of textile enterprises: In the New Product Introduction (NPI / APQP) process: During the requirements analysis and solution design phase, the system uses an AI agent to conduct a Design for Manufacturing (DFM) review of the design draft. The system automatically retrieves the Product Development Task Book from the PLM and combines it with historical quality anomaly records of similar products to predict the stability of design parameters in a mass production environment, thereby avoiding inherent quality risks before trial production.

[0094] In the process design and change workflow: When a process change occurs, the AI ​​agent monitors the quality data of the first batch of products after the change in real time and automatically compares the fluctuations before and after the change. The system can assess the risks of the change and prevent non-standard temporary experience-based operations from evolving into long-term standard deviations.

[0095] In the Supplier Quality Management (SQM) process: the system connects to raw material batches, supplier performance, and incoming inspection records in the ERP system via API interfaces. By analyzing the contribution of specific supplier raw materials to the final fabric defect rate, the AI-powered intelligent agent proactively outputs suggestions for adjusting procurement strategies.

[0096] Through pilot application in a large textile enterprise, this system reduced the incidence of recurring quality problems by 40%, shortened the average time for quality root cause analysis from 48 hours to 15 minutes, and shortened the maturity cycle of new products from R&D to mass production by about 20%.

[0097] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0098] Although this paper makes frequent use of terms such as modal data, feature vectors, and agents, the possibility of using other terms is not excluded. These terms are used merely for the convenience of describing and explaining the essence of this invention; interpreting them as any additional limitation would contradict the spirit of this invention.

Claims

1. A closed-loop control method for textile production quality based on a large-scale intelligent agent model, characterized in that, Includes the following steps: S1: Multimodal data acquisition and cross-modal feature alignment: Real-time acquisition of visual modal data V, temporal modal data T, and textual modal data S; semantic alignment is achieved by projecting the multimodal data onto a unified feature space Z using a spatiotemporal anchor projection algorithm. S2: Deep causal analysis based on R&D logic: When a quality anomaly occurs, the AI ​​intelligent agent central system dynamically calls the R&D design specifications and historical problem database of the related system according to the anomaly characteristics, and uses a large model combined with textile physical mechanisms to construct a heuristic reasoning chain to locate the root cause at the R&D end or the production end. S3: Multidimensional intelligent decision-making strategy generation and execution: Based on the root cause, it automatically generates short-term strategies that include production-side parameter adjustment instructions and long-term strategies that include R&D-side specification revision suggestions, and sends them to the Manufacturing Execution System (MOM / MES) or the Data Acquisition and Monitoring Control System (SCADA) through the API interface to implement closed-loop control. S4: Closed-loop effect verification and knowledge base self-updating: Monitor the production quality trend after the execution of decision instructions. If the verification is effective, the causal chain consisting of "abnormal phenomenon - root cause - decision strategy - execution effect" will be vectorized and stored in the historical problem database.

2. The method according to claim 1, characterized in that, The process of constructing the unified feature space Z in step S1 includes: S11: Utilize a deep learning encoder to extract visual feature vectors V based on visual modal data V, temporal modal data T, and text modal data S, respectively. i Temporal feature vector T j and text semantic vector S k ; S12: Using the production work order number and timestamp as spatiotemporal anchors, the visual feature vector V at a specific moment is... i The temporal feature vector T within a preset time period before and after this moment j To couple; S13: The coupled feature vector V is coupled through a unified projection layer. i T j and the corresponding text semantic vector S k Mapped to space Z, and using the contrastive loss function L=-log(exp(sim(z)) i ,z j ) / τ) / Σexp(sim(z i ,z n ) / τ)) constrain positive sample pairs of features z i ,z j The cosine distance in the space Z is calculated by minimizing the loss function L to shorten the distance between positive samples and features in space, and to increase the distance between negative samples and features; where τ is a preset temperature scalar parameter used to control the concentration of the probability distribution.

3. The method according to claim 1, characterized in that, In step S2, the AI ​​agent central system performs recursive link tracing: it sequentially retrieves real-time parameters from the production layer, retrieves product design intent and process window settings from the R&D layer, and retrieves 8D report cases with similar feature vectors from the historical layer.

4. The method according to claim 1, characterized in that, Step S2 includes boundary verification of R&D design specifications: if the actual operating parameters are within the safe process window of the R&D design but quality abnormalities still occur, the design margin detection is triggered to calculate the material sensitivity under the current environmental parameters and deduce whether the process window range set in the original R&D did not take into account the impact of raw material batch fluctuations.

5. The method according to claim 1, characterized in that, The specific implementation plan for step S2 includes: S21: Abnormal Feature Structuring and Multi-level Intent Recognition: Construct a standard ontology for the textile field that includes standard document entities, standard clause entities, and standard hierarchical entities. After the large model transforms multimodal data into a structured query language, it uses a multi-level logic tree of appearance-state-attribution to classify and determine intent. S22: Dynamic retrieval of knowledge base: Retrieve historical cases and R&D design specifications from the vector database, introduce a dual constraint filter of mandatory standard specifications and textile physical mechanism, calculate the error rate between feature values ​​and theoretical values ​​for quantitative elimination, and dynamically re-rank based on similarity, timeliness and physical fit. S23: Heuristic Hybrid Inference Chain and Double-Key Joint Backtracking: The large model generates potential root cause hypotheses containing physical formula IDs and parameters in parallel. After quantitative verification by an external physical calculation module, the comprehensive confidence level is calculated by combining physical consistency and evidence chain matching degree for pruning. For the verified hypotheses, the evidence chain is traced back using a double-key joint index of production work order number and timestamp. S24: Boundary Quantification Judgment and Counterfactual Causal Attribution: Based on multi-dimensional indicators such as equipment health index and operation compliance rate, define the responsibility boundaries between the R&D end and the production end, call the causal prediction model to perform counterfactual inference, and output a structured traceability report after being verified by the physical consistency checker.

6. The method according to claim 5, characterized in that, The standard ontology for the textile industry comprises a four-layer architecture: The first layer is the entity concept layer, which includes basic entities such as raw materials, equipment, process parameters, quality defects, and systems, as well as dedicated entities such as standard documents, standard clauses, and standard classifications. The second layer is the standard metadata layer, which is used to establish a two-way mapping relationship between the basic entities and the standard clause entities; The third layer is the semantic relationship layer, which, in addition to causal, subordinate and constraint relationships, defines standard-specific related predicates including business entities being bound by standard clauses, abnormal phenomena violating standard limits, and standard clauses defining product quality levels. The fourth layer is the constraint axiom layer, which includes the mandatory compliance axiom, the hierarchical threshold axiom, the process tolerance axiom, and the standard conflict arbitration axiom, serving as the dual constraint filter base for large model reasoning.

7. The method according to claim 1, characterized in that, Step S3 includes instruction conflict arbitration: when the parameter instructions of the short-term strategy conflict with the hard threshold of the R&D design specification, the AI ​​agent arbitrates based on the principle of quality gain priority and generates a temporary deviation permission record to be synchronized to the quality management system (QMS).

8. A closed-loop control system for textile production quality based on a large-scale intelligent agent, used to implement the method described in any one of claims 1-7, characterized in that, Include: Multimodal perception module: used to collect and align visual modal data, temporal modal data, and text modal data in real time; AI intelligent agent central module: used to respond to quality anomalies, dynamically call cross-system data and generate source tracing conclusions using a large model inference engine embedded with textile physical mechanisms; Collaborative decision-making and distribution module: used to generate multi-dimensional decision strategy packages and distribute instructions to PLM, MOM / MES and SCADA systems through industrial communication interfaces; Closed-loop evaluation and evolution module: used to evaluate decision gains and automatically deposit effective causal chains into the knowledge base.

9. The system according to claim 8, characterized in that, The AI ​​intelligent agent central module connects to the Product Lifecycle Management System (PLM), Manufacturing Operations Management System (MOM), Quality Management System (QMS), Laboratory Information Management System (LIMS), and Enterprise Resource Planning System (ERP) via API interfaces, enabling heterogeneous integration of R&D, production, testing, and material data.

10. The system according to claim 8, characterized in that, The large model inference engine adopts a private deployment mode and is fine-tuned for knowledge in the textile industry, enabling it to have the logical deduction ability to understand the impact of yarn twist, weaving tension and dye physical properties on quality.

Images