Adaptive process adjustment method and system for small batch production of carpets

Abstract

The present application relates to the technical field of industrial automation, and in particular to a carpet production adaptive process adjustment method and system for small-batch production. The method comprises: obtaining abstract requirement data input by a customer; converting the abstract requirement data into a carpet pattern based on a multi-modal generation model; generating initial process parameters corresponding to the carpet pattern according to a process knowledge graph; adaptively adjusting the initial process parameters based on small-batch production historical data through a meta-learning module to obtain optimized process parameters; and controlling carpet production equipment to produce based on the optimized process parameters. The present application realizes automatic and accurate conversion from requirements to patterns through multi-modal generation technology, provides intelligent initial recommendations using a process knowledge graph, and realizes rapid adaptive optimization of process parameters for new orders with the help of meta-learning, effectively solving the problems of low efficiency, high cost and poor quality consistency in small-batch carpet customized production, and realizing the intelligent upgrading of flexible manufacturing.

Description

Technical Field

[0001] This invention relates to the field of industrial automation technology, specifically to an adaptive process adjustment method and system for carpet production in small-batch production. Background Technology

[0002] The carpet manufacturing industry is undergoing a crucial transformation from large-scale standardized manufacturing to small-batch, personalized customization. In traditional large-scale production, process parameters are relatively fixed, and the production process is stable. However, facing increasing personalized demands, small-batch, multi-pattern production is becoming the trend, posing significant challenges to the flexibility, rapid adjustment capabilities, and first-time success rate of production processes. Currently, in small-batch custom carpet production, the process from abstract, non-professional textual descriptions or style references provided by customers to the final productionable pattern, and determining complex process parameters such as matching weave density, yarn type, and dyeing formula, still heavily relies on the manual judgment and repeated trial and error of experienced designers and process engineers. This not only leads to long order response cycles and low production efficiency but also generates substantial material waste and production costs during the trial-and-error process. Although some computer-aided design (CAD) systems are used for pattern design, and manufacturing execution systems (MES) are used to manage production data, these systems are fragmented and lack end-to-end intelligent decision-making capabilities from demand to production. In particular, existing technologies struggle to effectively model and utilize the complex relationships between pattern features, process parameters, and production results implicit in historical production data. Furthermore, they are unable to quickly and accurately recommend and adaptively optimize the best production process parameters when faced with a completely new, never-before-produced pattern. Therefore, the industry urgently needs a solution that can automatically understand customer needs, intelligently generate producible patterns, and dynamically optimize production processes based on historical experience and real-time data to achieve high-quality, high-efficiency, and highly flexible small-batch carpet production.

[0003] Therefore, existing technologies still need further development. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide an adaptive process adjustment method and system for carpet production in small-batch production, so as to solve the problems existing in the prior art.

[0005] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides an adaptive process adjustment method for carpet production oriented towards small-batch production, comprising: S1. Obtain abstract requirement data input by the customer; S2. Based on the multimodal generation model, the abstract requirement data is transformed into a carpet pattern; S3. Generate initial process parameters corresponding to the carpet pattern based on the process knowledge graph; S4. The initial process parameters are adaptively adjusted based on historical data from small-batch production using the meta-learning module to obtain optimized process parameters; S5. Control the carpet production equipment to produce based on the optimized process parameters.

[0006] Specifically, the process of transforming the abstract requirement data into a carpet pattern based on a multimodal generation model includes: A diffusion model is used to process natural language descriptions or reference images to generate high-resolution carpet patterns, and pattern feature vectors are output simultaneously. The pattern feature vectors are then used to query the process knowledge graph to obtain pattern complexity indicators, which include the number of colors, texture detail density, and pattern geometric complexity. Based on the pattern complexity indicators, the number of denoising steps or the latent spatial dimension of the diffusion model are dynamically adjusted to optimize pattern generation efficiency and ensure that the pattern meets the technical constraints of small-batch production equipment.

[0007] Specifically, generating the initial process parameters corresponding to the carpet pattern based on the process knowledge graph includes: The process knowledge graph is constructed to associate pattern features, equipment parameters, and yield rates in historical production data; the pattern feature vectors are mapped to an initial set of process parameters, including yarn type, weaving density, dyeing formula, and weaving speed, through graph query; the process knowledge graph is embedded and learned using a graph neural network to complete missing associations and recommend initial process parameters suitable for small-batch production, wherein the recommendation process takes into account real-time equipment status data.

[0008] Specifically, the adaptive adjustment of the initial process parameters based on historical data from small-batch production via the meta-learning module includes: A model-independent meta-learning framework is adopted, and a meta-model is trained using historical data from multiple small-batch production tasks. The meta-model learns a strategy for quickly adapting to process parameter adjustments for new patterns. For each new pattern, based on the initial process parameters and real-time production monitoring data, the process parameters are iteratively optimized within a preset range using the meta-model. The production monitoring data includes data from yarn tension, loom vibration, and humidity sensors. The optimization process is implemented through meta-reinforcement learning to maximize the yield rate of small-batch production tasks and minimize adjustment time.

[0009] Specifically, the process of controlling carpet production equipment based on optimized process parameters includes: The optimized process parameters are sent to the programmable logic controller to adjust the operating parameters of the weaving machine, dyeing machine, and cutting machine. During the production process, IoT sensors collect equipment status data and product quality data in real time and feed the data back to the meta-learning module. Based on the feedback data, the meta-learning module dynamically updates the process parameter adjustment strategy to achieve closed-loop adaptive control of the production equipment and ensure the process adaptability of different patterns in small-batch production.

[0010] Specifically, the method also includes an incremental update step for the process knowledge graph: After each small-batch production task is completed, actual production data, including final process parameters, equipment logs, and quality inspection results, are collected. The actual production data is then integrated into the process knowledge graph using an incremental learning algorithm to update the association weights between pattern features and process parameters. The impact of the updated graph on the prediction accuracy of subsequent tasks is evaluated through a meta-learning module, and the graph update frequency is adjusted to adapt to the data sparsity of small-batch production.

[0011] Specifically, obtaining the abstract requirement data input by the customer includes: The system receives natural language descriptions, sketch images, or style reference images, and converts different modal data into a unified feature representation through a multimodal encoder. The unified feature representation is semantically enhanced to extract deep requirement information related to the carpet pattern, including color preferences, texture types, and size specifications. The enhanced feature representation is then input into a multimodal generation model to initiate the pattern generation process, ensuring accurate conversion of abstract requirements into patterns.

[0012] Specifically, the method further includes a production anomaly adaptive handling step: During the production process, abnormal equipment signals and product quality deviations are monitored in real time. When an abnormality is detected, the meta-learning module is triggered to perform rapid reasoning based on historical abnormal cases and generate process parameter adjustment plans. The effectiveness of the adjustment plans is evaluated by a reinforcement learning agent, and correction instructions are sent to the production equipment in real time to restore production stability and reduce waste in small-batch production.

[0013] According to a second aspect of the present invention, an adaptive process adjustment system for carpet production oriented towards small-batch production is provided, comprising: The data acquisition module is used to acquire abstract requirement data input by the customer; The pattern generation module is used to transform the abstract requirement data into a carpet pattern based on a multimodal generation model. The parameter generation module is used to generate initial process parameters corresponding to the carpet pattern based on the process knowledge graph. The meta-learning adjustment module is used to adaptively adjust the initial process parameters based on historical data from small-batch production to obtain optimized process parameters. The control module is used to control the carpet production equipment to produce carpets based on the optimized process parameters.

[0014] Specifically, the system also includes a real-time monitoring module and a feedback learning module: The real-time monitoring module is used to collect production environment data, equipment operation data and product image data through a sensor network, and transmit the data stream to the meta-learning and adjustment module. The feedback learning module is used to calculate process parameter deviations based on real-time monitoring data and to optimize parameter generation strategies online using a meta-learning model to dynamically adapt to pattern changes and equipment wear in small-batch production, ensuring the continuous effectiveness of process adjustments.

[0015] Beneficial effects: Compared with existing technologies, the adaptive process adjustment method and system for carpet production in small-batch production provided by this invention has the following significant advantages: First, this invention, by introducing a multimodal generative model and semantic enhancement technology, achieves high-precision understanding and transformation of abstract and ambiguous customer requirements. The system can automatically transform unstructured inputs such as text, sketches, or reference drawings into high-quality design patterns that conform to production specifications, and simultaneously extract key pattern features. This fundamentally changes the traditional design mode that relies on manual communication and repeated modifications, significantly improving the accuracy and efficiency of front-end requirement transformation and shortening order confirmation time.

[0016] Secondly, this invention systematically and structurally consolidates the scattered and implicit experience of master craftsmen and historical production data by constructing and continuously updating a process knowledge graph. This enables the system to perform intelligent reasoning based on rich "memory" when faced with new patterns, quickly providing scientific and reasonable recommendations for initial process parameters. This effectively solves the "cold start" problem in small-batch production and reduces the risk of high scrap rates caused by improper initial parameter settings.

[0017] Third, this invention creatively introduces a meta-learning framework into the field of production process optimization. The meta-learning module, trained on massive amounts of historical small-batch production tasks, acquires meta-knowledge on "how to quickly learn and adapt to new tasks." When faced with each new order, this module can combine real-time production monitoring data to iteratively optimize initial parameters in a very short time, quickly converging to the optimal or near-optimal set of process parameters for the current specific production context. This significantly reduces the time and material costs required for traditional process debugging, truly achieving highly efficient and flexible production with "one piece, first-class" results.

[0018] Fourth, this invention achieves dynamic adaptive control of the production process by constructing a complete real-time closed loop of "perception-decision-execution-feedback". The real-time monitoring and feedback learning module can continuously perceive subtle changes in equipment status and product quality, and drive the meta-learning model to fine-tune its decision-making strategy online. This dynamic adaptability enables the system to effectively cope with uncertainties such as equipment wear, material batch differences, and environmental fluctuations, ensuring production stability and quality consistency throughout the entire small-batch production process, and even between different orders.

[0019] Fifth, this invention possesses continuous self-evolution capabilities. After each production task is completed, its data is incrementally integrated into the knowledge graph and used to optimize the meta-learning model. This makes the system's "experience" increasingly rich, its decisions increasingly accurate, forming a virtuous cycle of becoming smarter with use. Furthermore, the integrated anomaly adaptive handling mechanism can perform rapid diagnosis and intelligent handling based on a historical case database, minimizing downtime and material waste, further ensuring the economy and reliability of small-batch production. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the adaptive process adjustment method for carpet production aimed at small-batch production, provided in a specific embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0022] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0023] Please see Figure 1 This invention provides an adaptive process adjustment method for carpet production aimed at small-batch production, comprising: S1. Obtain abstract requirement data input by the customer.

[0024] It should be further explained that this method provides an end-to-end adaptive process adjustment framework. The entire process begins with abstract requirements input by the customer, such as a piece of natural language text describing "a light gray geometric pattern with a slight embossed feel in a modern minimalist style," or a reference image containing a similar style.

[0025] S2. Based on the multimodal generation model, the abstract requirement data is transformed into a carpet pattern.

[0026] It should be further explained that the system first parses and understands these requirements. Then, the core multimodal generative model, such as one based on a latent diffusion model, transforms these abstract requirements into a concrete, production-ready high-resolution carpet design pattern, such as a 1024x1024 pixel PNG image. Simultaneously with pattern generation, the model also extracts the deep feature vectors of the pattern.

[0027] S3. Generate initial process parameters corresponding to the carpet pattern based on the process knowledge graph.

[0028] It should be further explained that the system queries a pre-built process knowledge graph containing rich historical production experience. This graph stores, in the form of knowledge triples, the complex relationships between different pattern features (such as color complexity and line density), the process parameters used (such as yarn type and weaving density), and the final production results (such as yield rate and production time) from a large number of past production tasks (e.g., over 10,000). By querying this graph, the system can match and recommend a set of initial process parameters for newly generated patterns, such as recommending the use of polypropylene yarn, a weaving density of 8 stitches / cm, and a reactive dyeing scheme.

[0029] S4. The initial process parameters are adaptively adjusted based on historical data from small-batch production using the meta-learning module to obtain optimized process parameters.

[0030] It should be further explained that, due to the nature of small-batch production, the pattern for each order may be unique, and directly applying historical parameters may not be entirely suitable. Therefore, this method introduces a meta-learning module. This module is pre-trained on a large number of historical small-batch tasks, and its core capability is "learning how to quickly learn and adapt to new tasks." When faced with a new pattern and its initial process parameters, the meta-learning module can combine the real-time status of the current production equipment (such as the readings of the yarn tension sensor, the temperature and humidity of the workshop environment) to fine-tune and optimize the initial parameters in a very short time (e.g., a few minutes), generating a set of optimized process parameters that are more suitable for the current specific production situation.

[0031] S5. Control the carpet production equipment to produce based on the optimized process parameters.

[0032] It should be further explained that the optimized parameters will be converted into specific equipment control commands and sent to production equipment such as weaving machines and dyeing machines to drive them to complete the production of this small batch of carpets.

[0033] Understandably, the core advantage of this method lies in its adaptability and intelligence. It solves the problems of traditional carpet production, especially small-batch customized production, such as the difficulty in converting abstract customer ideas into concrete, executable production processes, reliance on experienced craftsmen, and high trial-and-error costs. By combining multimodal generation, knowledge graphs, and meta-learning, the system can not only automatically complete pattern design but also intelligently and dynamically optimize production processes. This allows the production of small-batch, personalized orders to be as efficient, stable, and yield-efficient as large-scale standardized production, greatly improving production flexibility and market responsiveness.

[0034] Specifically, the process of transforming the abstract requirement data into a carpet pattern based on a multimodal generation model includes: A diffusion model is used to process natural language descriptions or reference images to generate high-resolution carpet patterns, and pattern feature vectors are output simultaneously. The pattern feature vectors are then used to query the process knowledge graph to obtain pattern complexity indicators, which include the number of colors, texture detail density, and pattern geometric complexity. Based on the pattern complexity indicators, the number of denoising steps or the latent spatial dimension of the diffusion model are dynamically adjusted to optimize pattern generation efficiency and ensure that the pattern meets the technical constraints of small-batch production equipment.

[0035] It should be further explained that this step elaborates on the specific mechanism of combining pattern generation with process prediction.

[0036] (1) First, the multimodal generative model adopts a stable diffusion model architecture. Its training process involves collecting a dataset containing over 100,000 pairs of "(text description / reference image, carpet design image)". The text descriptions are converted into text embedding vectors using the CLIP text encoder. ,in Represents the semantic features of the text. The reference image is converted into an image latent vector through an image encoder (such as the encoder part of a VAE). ,in Represents the visual features of an image. or As conditional input, a U-Net-structured denoising network is guided to perform denoising iterations in the latent space, ultimately generating a high-resolution carpet pattern via a VAE decoder. Simultaneously, the model extracts a 1024-dimensional feature vector from the intermediate layers of the U-Net. As a pattern feature vector It contains semantic and stylistic information about the pattern.

[0037] (2) Calculation and acquisition of pattern complexity indicators. The system predefines three key pattern complexity indicators: number of colors. The number of primary colors in the pattern is determined by clustering the pixel colors of the generated pattern using the K-Means clustering algorithm. The optimal cluster number K is 8 to balance computational efficiency and color distinguishability. Texture detail density... The variance of the local binary patterns of the pattern is calculated by multiplying the local binary patterns at multiple scales (e.g., 1x1, 3x3, 5x5 pixel windows), with the unit being "units of detail / pixel". A threshold of 0.05 is preferred; values ​​below this are considered simple textures. The geometric complexity of the pattern is also considered. The ratio of the total length of the outline after pattern edge detection (such as with the Canny operator) to the perimeter of its smallest bounding rectangle is obtained. This ratio is dimensionless, and a value greater than 2.0 is generally considered to indicate geometric complexity. The calculation results of these metrics will be related to... They can be stored together as key-value pairs in the process knowledge graph or used for real-time queries.

[0038] (3) Dynamic model adjustment based on complexity. To improve generation efficiency and ensure pattern producibility, the system dynamically adjusts the diffusion model parameters based on the complexity index obtained from the query. For example, when Less than 5 and When the value is less than 1.5, it is judged as a simple pattern, and the number of denoising steps in the diffusion model is reduced from the default 50 steps to 20 steps, which can reduce the generation time by about 60%. When the value is greater than 0.1, it is judged as a high-detail pattern, and the latent space dimension is increased from the default 256 to 768 to ensure that the model has sufficient capacity to depict fine textures and avoid generating blurry patterns. These adjustment thresholds (such as 5, 1.5, 0.1, 20 steps, 768 dimensions) are empirically optimized values ​​based on extensive testing of historical production data and model performance, achieving a good balance between generation quality and speed.

[0039] Understandably, this step integrates pattern generation with subsequent production process considerations upfront. It not only generates the pattern but also simultaneously outputs key features and complexity indicators for process reasoning. Furthermore, it introduces production knowledge-based optimization into the generation process itself, avoiding the generation of patterns that appear aesthetically pleasing but are extremely complex, inefficient to produce, or have low yield rates. This fundamentally improves the feasibility and economy of small-batch production.

[0040] Specifically, generating the initial process parameters corresponding to the carpet pattern based on the process knowledge graph includes: The process knowledge graph is constructed to associate pattern features, equipment parameters, and yield rates in historical production data; the pattern feature vectors are mapped to an initial set of process parameters, including yarn type, weaving density, dyeing formula, and weaving speed, through graph query; the process knowledge graph is embedded and learned using a graph neural network to complete missing associations and recommend initial process parameters suitable for small-batch production, wherein the recommendation process takes into account real-time equipment status data.

[0041] It should be further explained that this step details the construction, querying, and reasoning process of the process knowledge graph, including: (1) Construction of the process knowledge graph. The graph is stored in the form of triples, for example, "(pattern feature vector) Applicable to a set of process parameters "(Process parameters)" Impact, yield rate "(Device Status)" Not applicable to yarn types ")". Among them, pattern feature vector It is a 1024-dimensional real vector. (Set of process parameters) It is structured data containing discrete parameters (such as yarn type: {nylon, polypropylene, wool}) and continuous parameters (such as knitting density: 6-12 stitches / cm). Equipment parameters and yield rates are also correlated in a similar way. The map is automatically constructed by parsing historical MES (Manufacturing Execution System) databases, equipment logs, and quality inspection reports, and calibrated by manually annotating some key relationships.

[0042] (2) Initial parameter mapping based on graph query. When a feature vector of a new pattern is input... At that time, the system first searches for the most similar one in the spectrum. Each historical pattern feature vector ( Similarity is calculated using cosine similarity. Calculation, where Represents the dot product. Represents the magnitude of a vector. The optimal value is 5 to balance the diversity and accuracy of recommendations. Then, this is aggregated. The set of process parameters corresponding to each neighboring node is used to obtain a preliminary set of process parameters through voting (for discrete parameters) or weighted averaging (for continuous parameters, with weights being similarity). .

[0043] (3) Reasoning and completion based on graph neural networks. To handle novel patterns (without nearest neighbors in the graph) commonly encountered in small-batch production, the system employs graph convolutional networks (GCNs) for embedding learning of the knowledge graph. The propagation formula for each layer of the GCN is as follows: ,in It is an adjacency matrix with self-connections. It is the original adjacency matrix. It is the identity matrix; yes The degree matrix; It is the first Layer node characteristics; It is the first The trainable weight matrix of the layer; It is a non-linear activation function, such as ReLU. Through training, GCN learns to map nodes (patterns, parameters, devices, etc.) in the graph to a low-dimensional vector space, making closely related nodes in the graph closer in the vector space. For a new pattern Even if it is not in the original image, an embedding representation can be calculated based on its feature vector. Then, through nearest neighbor search in the vector space, the best-matching process parameter node can be inferred, thereby filling in the missing associations and making recommendations. .

[0044] (4) Recommendation correction based on real-time equipment status. When making recommendations, the system will read the status data of currently available production equipment in real time. For example, "The current spindle vibration amplitude of loom A is 0.05mm". During GCN inference or similarity weighting, a state matching factor is introduced. If a certain historical process parameter... Related historical device status and The difference is too large (e.g., High-precision weaving is required, but the current loom vibrates too much. Therefore, this... The weight of certain items will be reduced or they will be filtered out directly to ensure the recommendation is accurate. It is feasible under the current equipment conditions.

[0045] Understandably, this step systematizes and structures discrete historical experiences through knowledge graphs and graph neural networks, endowing them with reasoning capabilities. It not only provides reliable references when similar historical cases exist, but also, through deep learning, "draws inferences from one instance to another," recommending reasonable initial processes for entirely new patterns. Furthermore, by closely integrating with real-time equipment conditions, the recommendations become more robust and reliable, effectively reducing the "cold start" risk of small-batch production.

[0046] Specifically, the adaptive adjustment of the initial process parameters based on historical data from small-batch production via the meta-learning module includes: A model-independent meta-learning framework is adopted, and a meta-model is trained using historical data from multiple small-batch production tasks. The meta-model learns a strategy for quickly adapting to process parameter adjustments for new patterns. For each new pattern, based on the initial process parameters and real-time production monitoring data, the process parameters are iteratively optimized within a preset range using the meta-model. The production monitoring data includes data from yarn tension, loom vibration, and humidity sensors. The optimization process is implemented through meta-reinforcement learning to maximize the yield rate of small-batch production tasks and minimize adjustment time.

[0047] It should be further explained that this step is the core of the adaptive adjustment achieved in this invention, and it elaborates on the working mechanism of the meta-learning module, including: (1) Training of the meta-learning model (meta-training phase). A model-independent meta-learning framework is adopted. Historical production data is organized into a large number of small-batch production tasks. Each task Each historical order corresponds to a specific pattern feature. Initial process parameters Monitoring data collected in time series during the production process (such as yarn tension and vibration at time t), and the final production results. (e.g., yield rate, total production time). The goal of the MAML algorithm is to train a process tuning model with well-initialized parameters. ,in These are the model parameters. The training objective function is: in, Indicating the model in the task The adapted parameters are obtained after performing one or more gradient descent steps on the support set (a small amount of data for this task). This is the internal learning rate, with an optimal value of 0.01. In the mission The loss function, calculated on the query set (the remaining data for the task), is defined as follows: ,in It's the yield rate. It's the time spent adjusting the parameters. and These are weighting coefficients, with optimal values ​​of 0.7 and 0.3, emphasizing yield rate as the primary optimization objective. Through training on a large number of tasks, the meta-model... We learned a general parameter adjustment strategy that can quickly adapt to new tasks.

[0048] (2) Application to new tasks (meta-testing phase). When a new pattern order is received, i.e., a new task... At that time, the system obtains its and After production starts, real-time data collection and monitoring begins. The meta-learning module loads the pre-trained meta-model. The adjustment process begins: With current process parameters (Initially) ) and recent monitoring data window As input, output a set of process parameter adjustment amounts. . Each component is limited to a preset physically feasible range, such as the adjustment amount of the weaving speed. Hours per second to prevent drastic changes. Then, The command is sent to the device for execution.

[0049] (3) Meta-reinforcement learning optimization loop. The above adjustment process is essentially a reinforcement learning problem. State It is a combination of current process parameters and monitoring data, and the action... It is the parameter adjustment amount. ,award The reward is dynamically calculated based on production results; for example, a positive reward is given for improved yield, and a negative reward is given for defective products. We combine MAML with reinforcement learning, i.e., meta-reinforcement learning. In the meta-training phase, the agent (adjusting the model) learns in multiple different production tasks (environments), with the goal of finding a set of initial policy parameters that allow it to quickly obtain high cumulative rewards (i.e., high yield and short adjustment time) with only a small amount of trial and error (i.e., minimal interaction with the environment) when facing new tasks. In the meta-testing phase, facing... The agent rapidly explores based on its learned initial policy and continuously fine-tunes its policy (i.e., adjusts it) according to real-time reward signals. (the generation method), thereby finding a target in a short time. The optimal or near-optimal process parameters .

[0050] Understandably, the meta-learning mechanism in this step is key to this method's adaptation to the "one-piece-per-quality" characteristic of small-batch production. Instead of learning and parameter tuning from scratch for each new pattern, it utilizes meta-knowledge of "how to learn quickly" learned from past tasks, significantly accelerating the process convergence for new tasks. This significantly reduces production preparation time and trial-and-error material waste for each small-batch order, truly achieving efficiency improvements in flexible production while ensuring quality.

[0051] Specifically, the process of controlling carpet production equipment based on optimized process parameters includes: The optimized process parameters are sent to the programmable logic controller to adjust the operating parameters of the weaving machine, dyeing machine, and cutting machine. During the production process, IoT sensors collect equipment status data and product quality data in real time and feed the data back to the meta-learning module. Based on the feedback data, the meta-learning module dynamically updates the process parameter adjustment strategy to achieve closed-loop adaptive control of the production equipment and ensure the process adaptability of different patterns in small-batch production.

[0052] It should be further explained that this step constitutes a complete closed loop of "perception-decision-execution-feedback", and the specific design includes: (1) Execution layer control. Optimize process parameters. It is a set of structured data. The system uses industrial communication protocols such as OPCUA or Modbus TCP to transmit... The parameters in the code are translated into instructions that can be executed by specific equipment. For example, "knitting density: 8.5 stitches / cm" is translated into a target speed pulse signal for the knitting machine's servo motor; "dyeing formula C" is translated into the injection ratio and duration of each dye pump in the dyeing machine. These instructions are sent to the programmable logic controllers of each device on the production line, driving the equipment to operate according to the new parameters.

[0053] (2) Real-time data sensing. During the production process, IoT sensors deployed at key parts of the equipment continuously collect data. Equipment status data includes: real-time vibration acceleration of the braiding machine spindle (unit: ), collected by an accelerometer, with a sampling frequency preferably of 1000Hz, to determine mechanical stability; yarn tension (unit: ), collected by a tension sensor; dyeing tank temperature (unit: pH value; workshop temperature and humidity. Product quality data is collected through an online visual inspection system: industrial cameras periodically photograph the carpet weave surface, and image processing algorithms calculate the number of defects and color differences in real time. Values, etc. All of this data is timestamped, forming a high-frequency data stream.

[0054] (3) Closed-loop feedback and strategy update. Real-time data streams are continuously fed into the meta-learning module. An online learner is maintained within the module. This learner continuously compares the current actual monitoring data with the expected data range based on the current process parameters. For example, if the expected vibration amplitude is less than 0.1... However, it actually remained at 0.15. This generates a negative feedback signal. The core model of the meta-learning module... The system will be fine-tuned online based on this continuous feedback stream. Specifically, the current production segment (including recent process parameters, monitoring data, and feedback rewards) is used to create a small online experience, which is then added to the model's replay buffer. The model periodically (e.g., every 5 meters of carpet produced) samples data from the buffer for online gradient descent updates, adjusting its network parameters accordingly. This allows the model to dynamically adapt to performance drift of the equipment itself (such as machine wear), material batch differences, and environmental changes, ensuring that its process adjustment strategy maintains optimal or near-optimal performance throughout the entire small-batch production process and even in subsequent orders.

[0055] Understandably, this step represents a leap from static optimization to dynamic adaptive control. It transforms a one-time parameter optimization into a continuous optimization process that spans the entire production process. Through real-time feedback closed-loop, the system can cope with various unforeseen disturbances during production, ensuring that even for a single small-batch order, the stability of the production process and the quality of output are guaranteed in a closed-loop manner, significantly improving the reliability and consistency of small-batch production.

[0056] Specifically, the method also includes an incremental update step for the process knowledge graph: After each small-batch production task is completed, actual production data, including final process parameters, equipment logs, and quality inspection results, are collected. The actual production data is then integrated into the process knowledge graph using an incremental learning algorithm to update the association weights between pattern features and process parameters. The impact of the updated graph on the prediction accuracy of subsequent tasks is evaluated through a meta-learning module, and the graph update frequency is adjusted to adapt to the data sparsity of small-batch production.

[0057] It should be further explained that this step enables the system to continuously evolve, and the specific design includes: (1) Data collection and structuring. When a production order is completed, the system automatically packages all the data of the task into a complete data package. .include: ① The final set of complete process parameters optimized through meta-learning. ; ② A detailed equipment log for the entire production process, recording the adjustment time and amount of all parameters; ③ The final quality inspection results, such as the yield rate and the specific distribution of various defects; ④ Statistical characteristics of key monitoring data sequences during the production process (such as mean and variance).

[0058] (2) Incremental fusion of knowledge graphs. Traditional graph updates require full retraining, which is computationally intensive. This method uses an incremental learning algorithm based on graph neural networks. When new data... Upon arrival, the algorithm first considers its pattern features. This is then inserted into the graph as a new node. Then, based on... process parameters Based on the production results, create new edges between this new node and existing process parameter nodes and result nodes in the graph, or update the weights of existing edges. Edge weights This can be set based on the frequency of the association's occurrence in history or the average yield rate. Then, an incremental graph learning algorithm, such as the incremental Personalized PageRank algorithm based on random walks, is used to locally update the embedding representations of relevant nodes in the graph, instead of recalculating the entire graph. This significantly reduces update overhead.

[0059] (3) Meta-learning evaluation and update strategy adjustment. After each knowledge map update, the system initiates an evaluation process. The meta-learning module uses the knowledge maps before and after the update to perform "simulated recommendations" on a batch of recently completed historical tasks that were not included in this update. That is, it uses the old and new knowledge maps to give predictions of initial process parameters and compares them with the optimized parameters actually used in the task to calculate the change in prediction accuracy. If the average prediction accuracy of the new knowledge map on multiple tasks improves by more than a threshold (e.g., 5%), the update is considered effective and the update frequency of the knowledge map may be accelerated (e.g., from once every 10 tasks to once every 5 tasks). If the accuracy decreases or the improvement is not significant, the update frequency may be slowed down and the new data may be checked for noise or anomalies. This evaluation mechanism enables the system to adapt to the slow and sparse characteristics of small-batch production data, intelligently manage the update rhythm of its own knowledge base, and ensure the stable accumulation and effective use of knowledge.

[0060] Understandably, this step endows the system with the vitality for self-learning and continuous improvement. Each production run not only completes an order but also contributes new experience to the system. Through incremental updates, the process knowledge graph, like a master craftsman with ever-growing experience, becomes increasingly rich and accurate in its "experience base." The meta-learning module, acting as the "decision-making brain," can also utilize higher-quality knowledge to make better initial judgments, forming a co-evolution between the algorithm and the knowledge base. This makes the entire system increasingly intelligent and reliable as production tasks progress.

[0061] Specifically, obtaining the abstract requirement data input by the customer includes: The system receives natural language descriptions, sketch images, or style reference images, and converts different modal data into a unified feature representation through a multimodal encoder. The unified feature representation is semantically enhanced to extract deep requirement information related to the carpet pattern, including color preferences, texture types, and size specifications. The enhanced feature representation is then input into a multimodal generation model to initiate the pattern generation process, ensuring accurate conversion of abstract requirements into patterns.

[0062] It should be further explained that this step is the starting point of the process, aiming to accurately understand and anchor the customer's vague needs, specifically including: (1) Multimodal data reception and encoding. The system provides a web interface or mobile application where customers can input natural language descriptions (e.g., "I want a blue carpet with an ocean theme and wave texture"), upload hand-drawn sketches (e.g., a simple line drawing created using drawing software), or upload style reference images (e.g., a photograph of the ocean or an image of a home decor style). In the system backend, the text descriptions are encoded into text feature vectors using a pre-trained BERT model. The sketch or reference image is encoded into an image feature vector using a ResNet-50 model pre-trained on ImageNet (with the last fully connected layer removed). To align features from different modalities into the same space, a cross-modal alignment network is used. Its training objective is to maximize the feature similarity of matched (text, image) pairs and minimize the similarity of mismatched pairs. Ultimately, all inputs are mapped to a unified 512-dimensional feature vector. .

[0063] (2) Semantic enhancement. Since customer input may be brief, vague, or contain irrelevant information, semantic enhancement is necessary. Semantic enhancement is performed. This is specifically achieved through a semantic understanding module based on an attention mechanism. This module has a built-in knowledge base related to carpet design, including RGB value mappings for color names, a texture vocabulary (such as "plain," "jacquard," and "embossed"), and a standard size library. The attention mechanism calculates... The relevance weights of the key vectors to each concept in the knowledge base are then weighted and summed with the concept value vectors in the knowledge base to obtain an enhanced demand feature vector. For example, when When the text contains "blue" and "wave," the attention mechanism assigns higher weights to "color: RGB(0,0,255)" and "texture: curves," thus increasing the likelihood of it appearing as a color. The text emphasizes these specific requirements and may supplement implicit information such as "material: soft".

[0064] (3) Initiate pattern generation. Use the semantically enhanced feature vectors... It is concatenated with a randomly generated noise vector and input together into the conditional input port of the diffusion model described in this invention. As a strong condition, it guides the diffusion model generation process, ensuring that the final output pattern closely matches the customer's clarified and enhanced true intention in terms of color, style, and composition, reducing repeated modifications caused by misunderstandings of requirements, and improving the first-time success rate of small-batch customization.

[0065] Understandably, this step effectively addresses the pain points of "unclear demand description and distorted communication" in personalized customization through advanced multimodal understanding and semantic enhancement technologies. It transforms the customer's subjective and vague intentions into standardized and structured features that computers can accurately understand and process, laying a solid foundation for subsequent high-quality pattern generation and process adaptation, thereby improving customer experience and order conversion efficiency.

[0066] Specifically, the method further includes a production anomaly adaptive handling step: During the production process, abnormal equipment signals and product quality deviations are monitored in real time. When an abnormality is detected, the meta-learning module is triggered to perform rapid reasoning based on historical abnormal cases and generate process parameter adjustment plans. The effectiveness of the adjustment plans is evaluated by a reinforcement learning agent, and correction instructions are sent to the production equipment in real time to restore production stability and reduce waste in small-batch production.

[0067] It should be further explained that this step is a rapid response mechanism for unexpected problems during the production process, and its specific design includes: (1) Anomaly Monitoring and Detection. The system has set a series of monitoring thresholds to automatically detect anomalies. Equipment anomaly signals include: spindle vibration amplitude exceeding the threshold for 3 seconds. The instantaneous fluctuation of yarn tension exceeds three times the standard deviation. Product quality deviations are detected by an online vision system: three or more consecutively detected defects of the same type (such as broken warp); color difference. Value exceeding tolerance for 5 consecutive meters These thresholds , It is an optimal value set based on equipment safety specifications and industry quality standards, which can achieve a balance between sensitivity and false alarm rate.

[0068] (2) Rapid anomaly reasoning based on meta-learning. Once an anomaly is triggered, the system immediately pauses or reduces the equipment speed (depending on the severity of the anomaly). Simultaneously, the meta-learning module is activated and enters "anomaly handling mode." This module internally stores an "anomaly case library," where each case records the anomaly type, the process parameter context at the time of occurrence, the adjustment measures taken, and their effects. The module quickly matches the feature vector of the current anomaly (such as vibration spectrum features and defect image features) with the case library to find the most similar case. A historical anomaly case. Then, utilizing the "rapid adaptation" capability learned in meta-training, combined with the current specific process parameters and equipment status, this... Based on experience in handling individual cases, one or more candidate emergency adjustment plans for process parameters can be generated within milliseconds. For example, the solution might be to "immediately reduce the weaving speed by 20% and fine-tune the X-axis servo gain."

[0069] (3) Reinforcement Learning Agent Evaluation and Execution. After generating the adjustment plan, it is not executed immediately because the same plan may have different effects in different situations. A lightweight reinforcement learning agent simulator is used within the system. This agent starts with the current production state and quickly simulates the execution of each candidate plan. It predicts the changing trends of key indicators (vibration, tension, defect rate) within a short future timeframe (e.g., 10 seconds). The agent calculates based on a predefined reward function. Evaluate each option, where It is to predict vibration. It predicts the defect rate. It's about weighting. Choose the option with the highest predicted reward. Immediately issue the order to the equipment for execution. After execution, continue to closely monitor changes in indicators. If the anomaly is not eliminated within a preset time (e.g., 30 seconds), trigger a new round of reasoning-evaluation-execution cycle until the anomaly is resolved.

[0070] Understandably, this step adds a powerful "immune system" to small-batch production. Faced with unexpected anomalies, the system no longer relies on manual experience for downtime troubleshooting and trial-and-error adjustments. Instead, it can automatically, quickly, and accurately diagnose problems and take corrective measures based on historical experience and online learning. This significantly reduces downtime due to anomalies, avoids the risk of scrapping an entire small-batch order due to a single anomaly, minimizes waste, and ensures production continuity and economy.

[0071] To more clearly illustrate the technical solution of this invention, the adaptive process adjustment method and system for carpet production in small-batch production, as described below, is illustrated with a specific order processing flow. This case demonstrates the entire process from customer order placement to product completion, with particular emphasis on meta-learning and the collaborative work of various modules.

[0072] Suppose a customer enters a natural language description through the system interface: "I want a dark blue rug with an irregular wavy texture, measuring 1.5 meters by 2 meters, with a modern art style." Further, the following process will be executed: 1. Requirements Acquisition and Analysis Specifically, the data acquisition module receives the text description input by the customer. The multimodal encoder (based on BERT) converts this text into a 512-dimensional text feature vector. Attention mechanism computation in the semantic enhancement module Similarity with concepts in the knowledge base significantly enhanced features such as "color: dark blue (RGB(0,0,139))", "texture: wavy, irregular", "style: modern", and "size: 1.5m×2m", and supplemented the implicit information of "expected process complexity: medium", outputting an enhanced requirement feature vector. .

[0073] 2. Pattern Generation and Feature Extraction Specifically, the pattern generation module will A random noise vector is input into the latent diffusion model. After 20 steps of denoising and sampling (dynamic adjustment was not enabled due to the initial assessment of moderate complexity), the model generates a 1024×1024 pixel dark blue wavy carpet pattern. Simultaneously, a 1024-dimensional pattern feature vector is extracted from the model's cross-attention layer. The number of colors in the pattern has been calculated. (Mainly dark blue, light blue, white, and gray), texture detail density Geometric complexity .

[0074] 3. Recommended initial process parameters Specifically, the parameter generation module receives... First, in the process knowledge graph, calculations are performed. With all historical pattern feature vectors in the atlas cosine similarity Find the five historical patterns with the highest similarity, and their corresponding sets of process parameters are as follows: arrive The initial set of process parameters is obtained by weighted averaging (with weights equal to normalized similarity). The yarn type is "polypropylene", the knitting density is "9.0 needles / cm", the dyeing scheme is "reactive dye - dark blue formula A", and the knitting speed is "450 needles / minute". Furthermore, the graph neural network inference did not find any parameters conflicting with the current equipment status (loom vibration is normal), therefore this setting is recommended. .

[0075] 4. Meta-learning adaptive optimization Specifically, the meta-learning adjustment module is launched. The module loads the pre-trained meta-model. Production begins, initial parameters. The data is sent to the equipment. The real-time monitoring module begins transmitting data: yarn tension. Loom spindle vibration Ambient humidity The meta-learning module combines these real-time monitoring data with... , Together with the input, further perform the following steps: (1) Initial adjustment: Output the first set of adjustment amounts Knitting speed Needles / minute (because the actual vibration is 0.12, slightly higher than the ideal value of 0.10), staining temperature (Based on the performance of similar historical patterns under current humidity). After adjustment, vibration was monitored to decrease. With more stable tension, the system awards a small positive reward.

[0076] (2) Iterative optimization: This is an online meta-reinforcement learning process. State It includes current parameters and monitoring data, and actions. yes Based on reward signals (yield prediction, adjustment cost), the model continues to fine-tune over the next five adjustment cycles. For example, in the third cycle, the model attempted to fine-tune the knitting density to "9.2 stitches / cm", finding that the texture was more refined, the reward increased, and this adjustment was retained. After five rounds of rapid iteration (approximately 15 minutes in total), the model converged to a set of optimized process parameters. Yarn type: "polypropylene", knitting density: "9.2 stitches / cm", dyeing scheme: "reactive dye - dark blue formula A (temperature adjustment)", knitting speed: "445 stitches / minute".

[0077] 5. Closed-loop production and anomaly handling Specifically, the control module is based on The control equipment is used for mass production. The feedback learning module continues to run. When production reaches approximately 8 meters, the real-time monitoring module detects spindle vibration. Suddenly rise to Exceeding the threshold This triggers exception handling, including: (1) Anomaly reasoning: The meta-learning module matched in the anomaly case library and found that the similarity with the case "slight knotting of yarn caused instantaneous tension unevenness" was 85%. The historical solution for this case was "briefly reduce speed by 10% and slightly relax the tension sensor threshold".

[0078] (2) Scheme Evaluation and Execution: The reinforcement learning agent simulates the execution of this scheme, predicting that the vibration will subside within 5 seconds. The system immediately issues an instruction: the knitting speed is temporarily reduced to 400 stitches / minute, and the tension alarm threshold is temporarily increased by 5%. After execution, the vibration subsides to normal within 4 seconds. Production has returned to normal. This event has been recorded as a new anomaly case to enrich the knowledge base.

[0079] 6. Results Analysis and Knowledge Update After the completion of this small-batch production run (totaling 20 meters), the final yield rate reached 96.5%. The entire process adjustment phase (from initial parameters to stable production) took only about 18 minutes. In contrast, in the traditional mode without this system, process debugging for similar new patterns typically takes 2-3 hours, and the yield rate usually fluctuates between 85% and 90%. Specifically, after the task is completed, the system collects the final parameters. (and (Basically consistent), equipment logs, a yield rate of 96.5%, and data packets of abnormal vibration events during production. The incremental learning algorithm adds the data from this task as a new node to the process knowledge graph and updates the "pattern features". "and process parameter set" The meta-learning module uses this new data to perform a minor online fine-tuning of its internal policy network, enabling it to make faster and better decisions when encountering similar scenarios such as "wave textures" and "abnormal vibrations" in the future.

[0080] Understandably, this case study fully demonstrates the entire process of processing a small batch of customized orders using the system of this invention. The case study clearly illustrates this through specific data (such as feature vector dimensions, similarity calculations, specific process parameter values, monitoring thresholds, adjustment steps, and time). ① How are abstract customer needs accurately analyzed and transformed? ② How can the process knowledge graph provide reliable initial recommendations? ③ How the meta-learning module plays a core role: through multiple rounds of rapid iteration, it optimizes the initial parameters into the optimal parameters that are adapted to the specific production environment, greatly shortening the parameter tuning time; ④ How real-time closed-loop control and adaptive anomaly handling ensure stable production; ⑤ How the system achieves self-evolution through incremental learning. This case powerfully demonstrates that the solution of this invention can effectively resolve the contradiction between personalization and high efficiency and high quality in small-batch carpet production, possessing significant practicality and progressiveness.

[0081] This invention provides another embodiment, which offers an adaptive process adjustment system for carpet production in small-batch production. The adaptive process adjustment system for carpet production in small-batch production includes: (1) Data acquisition module, used to acquire abstract requirement data input by customers.

[0082] It should be further clarified that this system refers to the hardware entity or software system architecture corresponding to the aforementioned methods. The data acquisition module is deployed on the front-end interactive interface of a cloud server or local server, responsible for receiving and preprocessing the diverse demand data input by the client through the terminal.

[0083] (2) Pattern generation module, used to convert the abstract requirement data into carpet patterns based on a multimodal generation model.

[0084] It should be further noted that the pattern generation module contains one or more GPU servers running pre-trained multimodal generative AI models (such as diffusion models) responsible for computationally demanding pattern generation tasks.

[0085] (3) Parameter generation module, used to generate initial process parameters corresponding to the carpet pattern based on the process knowledge graph.

[0086] It should be further explained that the parameter generation module is connected to a knowledge graph database, which can be built based on graph databases such as Neo4j, and is responsible for storing and querying process knowledge.

[0087] (4) Meta-learning adjustment module, used to adaptively adjust the initial process parameters based on historical data of small batch production to obtain optimized process parameters.

[0088] It should be further explained that the meta-learning adjustment module is the "intelligent hub" of the system. It can be a dedicated server equipped with a high-performance CPU and memory, on which a pre-trained meta-learning model and an online learning inference engine are deployed.

[0089] (5) Control module, used to control carpet production equipment to produce based on the optimized process parameters.

[0090] It should be further explained that the control module, as the interface with the physical world, is typically deployed on industrial gateways or edge computing devices in the production workshop, responsible for protocol conversion, command issuance, and data acquisition. All these modules interact with each other via networks (such as industrial Ethernet and 5G) to form a collaborative, distributed intelligent system.

[0091] Understandably, this system deeply integrates the perception, cognition, and decision-making capabilities of artificial intelligence with the execution and control capabilities of industrial production, forming a hardware-software integrated intelligent solution. Each module has a clear division of labor and works collaboratively, integrating the production process that originally relied on human experience and independent software into an automated, adaptive, and continuously learning intelligent production system, providing a solid technological infrastructure for small-batch flexible manufacturing in the carpet industry.

[0092] Specifically, the system also includes a real-time monitoring module and a feedback learning module: The real-time monitoring module is used to collect production environment data, equipment operation data and product image data through a sensor network, and transmit the data stream to the meta-learning and adjustment module. The feedback learning module is used to calculate process parameter deviations based on real-time monitoring data and to optimize parameter generation strategies online using a meta-learning model to dynamically adapt to pattern changes and equipment wear in small-batch production, ensuring the continuous effectiveness of process adjustments.

[0093] It should be further explained that this step is a modular implementation of the core closed-loop control function of the system, and the specific design includes: (1) Composition of the real-time monitoring module.

[0094] This module is a hardware and software integrated edge computing unit. On the hardware side, it includes vibration sensors, tension sensors, temperature and humidity sensors, high-resolution industrial cameras, and other components deployed at key points on the equipment, forming an IoT sensor network. On the software side, it includes data acquisition programs, preprocessing programs (such as filtering, noise reduction, and image feature extraction), and streaming data transmission programs running on the edge computing box. Data is continuously sent to the data buffer of the meta-learning and adjustment module via streaming data protocols such as MQTT or Kafka with millisecond-level latency.

[0095] (2) The working mechanism of the feedback learning module.

[0096] This module is a subprocess within the meta-learning tuning module. It continuously monitors the data stream from the real-time monitoring module. Internally, it maintains a "current desired state model," which is based on the optimization process parameters being executed. It predicts the theoretical values ​​or reasonable ranges for each monitoring indicator. The feedback learning module compares real-time data with theoretical values ​​and calculates the deviation. Then, it invokes the online learning component in the meta-learning model. This component uses the current process parameters, real-time monitoring data, calculated deviations, and the resulting reward / penalty signals as training samples to perform extremely small online gradient updates (learning rate as low as 0.0001) on the policy network of the meta-learning model. This update is gradual, designed to allow the model to slowly adapt to the slow time-varying characteristics of the equipment (such as tool wear, sensor drift) and subtle differences between different patterns without forgetting previously learned general knowledge. The updated model is immediately used to generate subsequent process parameter fine-tuning instructions, thus forming a real-time adaptive closed loop of "perception-deviation calculation-model fine-tuning-decision optimization-execution-re-perception".

[0097] Understandably, the real-time monitoring module acts as the system's "sensory nerves," providing a capillary network to perceive subtle changes in the production environment. The feedback learning module, on the other hand, is the system's "cerebellum," responsible for rapid and automated fine-tuning based on sensory feedback to maintain dynamic balance in production. The addition of these two modules transforms the system from an "open-loop optimization" system based on a fixed model into a "closed-loop adaptive" system capable of coping with uncertainty and possessing self-adjusting capabilities. This allows it to maintain a high level of process control even when facing the diversity and changes in small-batch production.

[0098] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the adaptive process adjustment method for carpet production in small-batch production. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0099] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0100] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0101] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0102] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for adaptive process adjustment in carpet production for small-batch production, characterized in that, include: S1. Obtain abstract requirement data input by the customer; S2. Based on the multimodal generation model, the abstract requirement data is transformed into a carpet pattern; S3. Generate initial process parameters corresponding to the carpet pattern based on the process knowledge graph; S4. The initial process parameters are adaptively adjusted based on historical data from small-batch production using the meta-learning module to obtain optimized process parameters; S5. Control the carpet production equipment to produce carpets based on the optimized process parameters.

2. The method according to claim 1, characterized in that, The process of transforming the abstract requirement data into a carpet pattern based on a multimodal generative model includes: A diffusion model is used to process natural language descriptions or reference images to generate high-resolution carpet patterns, and pattern feature vectors are output simultaneously. The pattern feature vectors are then used to query the process knowledge graph to obtain pattern complexity indicators, which include the number of colors, texture detail density, and pattern geometric complexity. Based on the pattern complexity indicators, the number of denoising steps or the latent spatial dimension of the diffusion model are dynamically adjusted to optimize pattern generation efficiency and ensure that the pattern meets the technical constraints of small-batch production equipment.

3. The method according to claim 2, characterized in that, The step of generating initial process parameters corresponding to the carpet pattern based on the process knowledge graph includes: The process knowledge graph is constructed to associate pattern features, equipment parameters, and yield rates in historical production data; the pattern feature vectors are mapped to an initial set of process parameters, including yarn type, weaving density, dyeing formula, and weaving speed, through graph query; the process knowledge graph is embedded and learned using a graph neural network to complete missing associations and recommend initial process parameters suitable for small-batch production, wherein the recommendation process takes into account real-time equipment status data.

4. The method according to claim 3, characterized in that, The adaptive adjustment of the initial process parameters based on historical data from small-batch production via the meta-learning module includes: A model-independent meta-learning framework is adopted, and a meta-model is trained using historical data from multiple small-batch production tasks. The meta-model learns a strategy for quickly adapting to process parameter adjustments for new patterns. For each new pattern, based on the initial process parameters and real-time production monitoring data, the process parameters are iteratively optimized within a preset range using the meta-model. The production monitoring data includes data from yarn tension, loom vibration, and humidity sensors. The optimization process is implemented through meta-reinforcement learning to maximize the yield rate of small-batch production tasks and minimize adjustment time.

5. The method according to claim 4, characterized in that, The process of controlling carpet production equipment based on optimized process parameters includes: The optimized process parameters are sent to the programmable logic controller to adjust the operating parameters of the weaving machine, dyeing machine, and cutting machine. During the production process, IoT sensors collect equipment status data and product quality data in real time and feed the data back to the meta-learning module. Based on the feedback data, the meta-learning module dynamically updates the process parameter adjustment strategy to achieve closed-loop adaptive control of the production equipment and ensure the process adaptability of different patterns in small-batch production.

6. The method according to claim 5, characterized in that, The method also includes an incremental update step for the process knowledge graph: After each small-batch production task is completed, actual production data, including final process parameters, equipment logs, and quality inspection results, are collected. Incremental learning algorithms are used to integrate the actual production data into the process knowledge graph and update the association weights between pattern features and process parameters. The impact of the updated graph on the prediction accuracy of subsequent tasks is evaluated through the meta-learning module, and the graph update frequency is adjusted to adapt to the data sparsity of small-batch production.

7. The method according to claim 1, characterized in that, The abstract requirement data obtained from customer input includes: The system receives natural language descriptions, sketch images, or style reference images, and converts different modal data into a unified feature representation through a multimodal encoder. The unified feature representation is semantically enhanced to extract deep requirement information related to the carpet pattern, including color preferences, texture types, and size specifications. The enhanced feature representation is then input into a multimodal generation model to initiate the pattern generation process, ensuring accurate conversion of abstract requirements into patterns.

8. The method according to claim 7, characterized in that, The method also includes an adaptive handling step for production anomalies: During the production process, abnormal equipment signals and product quality deviations are monitored in real time. When an abnormality is detected, the meta-learning module is triggered to perform rapid reasoning based on historical abnormal cases and generate process parameter adjustment plans. By using reinforcement learning agents to evaluate the effectiveness of adjustment plans and issuing correction instructions to production equipment in real time, production stability can be restored and waste in small-batch production can be reduced.

9. An adaptive process adjustment system for carpet production in small-batch production, characterized in that, include: The data acquisition module is used to acquire abstract requirement data input by the customer; The pattern generation module is used to transform the abstract requirement data into a carpet pattern based on a multimodal generation model. The parameter generation module is used to generate initial process parameters corresponding to the carpet pattern based on the process knowledge graph. The meta-learning adjustment module is used to adaptively adjust the initial process parameters based on historical data from small-batch production to obtain optimized process parameters. The control module is used to control the carpet production equipment to produce carpets based on the optimized process parameters.

10. The system according to claim 9, characterized in that, The system also includes a real-time monitoring module and a feedback learning module: The real-time monitoring module is used to collect production environment data, equipment operation data and product image data through a sensor network, and transmit the data stream to the meta-learning and adjustment module. The feedback learning module is used to calculate process parameter deviations based on real-time monitoring data and to optimize parameter generation strategies online using a meta-learning model to dynamically adapt to pattern changes and equipment wear in small-batch production, ensuring the continuous effectiveness of process adjustments.

Images