A green production optimization management method and system for jacquard fabrics

Abstract

This invention belongs to the field of production control technology and provides a green production optimization management method and system for jacquard fabrics. The method includes: based on the weaving parameters and the preset process parameters, using a pre-trained Transformer energy consumption prediction model to predict the energy consumption per unit fabric length; in response to the energy consumption per unit fabric length exceeding a preset threshold, using a reinforcement learning model to adjust the process parameters to obtain first optimized process parameters that meet the energy consumption threshold; obtaining the first correlation degree between the optimized parameters in the first optimized process parameters and surface defects of the jacquard fabric; conducting a preset number of trial productions based on the first optimized process parameters; obtaining second optimized process parameters by dynamically inputting surface defects during the trial production process into the reinforcement learning model; and producing the jacquard fabric based on the second optimized process parameters, thereby achieving a balance between energy consumption reduction and quality assurance in the jacquard fabric process parameters.

Description

Technical Field

[0001] This invention relates to the field of production control technology, and more specifically, to a green production optimization management method, system, electronic equipment, and computer program product for jacquard fabrics. Background Technology

[0002] In the process of green manufacturing transformation in the textile industry, jacquard fabrics are experiencing increased demand in the high-end sector due to their unique textures and rich patterns. However, production requires a balance between energy consumption and quality, and existing technologies present challenges. In traditional production, process parameter optimization focuses on a single energy consumption target, failing to address both energy consumption and quality in jacquard fabric production.

[0003] Therefore, how to achieve efficient green production optimization management methods for jacquard fabrics is an urgent problem to be solved in this field. Summary of the Invention

[0004] In response, the present invention provides a green production optimization management method, system, electronic device and computer program product for jacquard fabrics, so as to at least partially solve the above-mentioned technical problems.

[0005] As one aspect of the present invention, the present invention provides a green production optimization management method for jacquard fabrics, comprising the following method steps:

[0006] The process involves: acquiring the weaving parameters of the jacquard fabric, including at least one of warp and weft density, weave specifications, yarn characteristics, or equipment adaptation parameters; acquiring corresponding preset process parameters based on the weaving parameters; predicting the energy consumption per unit fabric length using a pre-trained Transformer energy consumption prediction model based on the weaving parameters and the preset process parameters; adjusting the process parameters using a reinforcement learning model to obtain first optimized process parameters that meet the energy consumption threshold when the energy consumption per unit fabric length exceeds a preset threshold; acquiring the first correlation degree between the optimized parameters in the first optimized process parameters and surface defects of the jacquard fabric; conducting a preset number of trial productions based on the first optimized process parameters; identifying surface defects of the jacquard fabric using a convolutional neural network; dynamically inputting the surface defects into the reinforcement learning model as state input factors to predict second optimized process parameters, wherein the preset number is positively correlated with the correlation degree; and controlling the production of the jacquard fabric based on the second optimized process parameters.

[0007] As another aspect of the present invention, the present invention also provides a green production optimization management system for jacquard fabrics, comprising:

[0008] A first acquisition module is used to acquire the weaving parameters of the jacquard fabric, the weaving parameters including at least one of warp and weft density, weave specifications, yarn properties, or equipment adaptation parameters; a first prediction module is used to acquire corresponding preset process parameters according to the weaving parameters, and predict the energy consumption per unit length of fabric using a pre-trained Transformer energy consumption prediction model based on the weaving parameters and the preset process parameters; a first adjustment module is used to adjust the process parameters using a reinforcement learning model in response to the energy consumption per unit length exceeding a preset threshold, to obtain first optimized process parameters that meet the energy consumption threshold; a second adjustment module is used to acquire the first correlation degree between the optimized parameters involved in the first optimized process parameters and the surface defects of the jacquard fabric, conduct a preset number of trial productions based on the first optimized process parameters, and identify the surface defects of the jacquard fabric based on a convolutional neural network, dynamically inputting the surface defects into the reinforcement learning model as state input factors, and predicting the second optimized process parameters, wherein the preset number is positively correlated with the correlation degree; a production control module is used to control the production of the jacquard fabric based on the second optimized process parameters.

[0009] As another aspect of the present invention, the present invention also provides an electronic device comprising: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program, when executed by the processor, implements the method as described in any of the foregoing embodiments.

[0010] As another aspect of the present invention, the present invention also provides a computer program product comprising a computer program that can be executed by a processor to implement the methods as described in any of the foregoing claims.

[0011] The beneficial effects of this invention are as follows:

[0012] By incorporating surface defect data from the trial production process into the reinforcement learning state space and reward function, and combining a dual-policy network, a hierarchical attention module, and a multimodal information fusion module, dynamic adjustment of process parameters is achieved. This reduces energy consumption while strictly controlling the number of defects such as skipped yarns, broken ends, and weft shrinkage.

[0013] In the dual-strategy network module, the main strategy network is responsible for routine energy consumption optimization, while the auxiliary strategy network adjusts high-risk process parameters. Both share the underlying feature extraction layer and output actions through weighted fusion. This design effectively avoids quality incidents caused by over-optimization of energy consumption when the model faces extreme parameter adjustment requirements.

[0014] The layered attention module calculates attention weights for weaving parameters, process parameters, and quality parameters respectively, analyzes the cross-layer correlations between parameters, and dynamically adjusts the combined weights of each parameter to achieve refined control of process parameters.

[0015] The number of trial productions is dynamically adjusted based on the correlation between parameters and defects. In scenarios with high correlation, the number of trial productions is automatically increased to ensure that the optimization strategy has sufficient data support and can effectively cope with complex and ever-changing production environments. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of a green production optimization management method for jacquard fabrics disclosed in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of a green production optimization management system for jacquard fabrics disclosed in an embodiment of the present invention. Detailed Implementation

[0019] The following specific embodiments illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Furthermore, the technical features involved in the different embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0021] like Figure 1 As shown in the figure, this invention discloses a green production optimization management method for jacquard fabrics, including the following steps:

[0022] Step S101: Obtain the weaving parameters of the jacquard fabric, wherein the weaving parameters include at least one of warp and weft density, weave specifications, yarn properties, or equipment adaptation parameters.

[0023] In some embodiments, weaving parameters are fundamental data for jacquard fabric production, generally including but not limited to the following parameters: Warp and weft density: refers to the number of warp and weft threads per unit length (usually per inch or per centimeter), which directly affects the fabric's thickness, tightness, breathability, and hand feel. Weave pattern: refers to the interlacing pattern of warp and weft yarns in the fabric, commonly including plain weave, twill weave, and satin weave. Different weave patterns give fabrics different appearance and performance characteristics. For example, plain weave fabrics have a stable structure, twill weave fabrics have a diagonal pattern on the surface, and better luster and hand feel, while satin weave fabrics have a smooth and lustrous surface. Yarn count: indicates the fineness of the yarn; the higher the count, the coarser the yarn. Yarn count affects the fabric's thickness, weight, and abrasion resistance. Coarser yarns produce thicker, more durable fabrics, suitable for making outerwear, etc.; finer yarns produce delicate and soft fabrics, often used for underwear, etc. Equipment compatibility parameters: Based on the fabric design requirements, select the appropriate weaving equipment type and model, as well as the relevant equipment parameter settings, such as the loom width, shedding mechanism type, and weft insertion method. Different equipment compatibility parameters will affect production efficiency, fabric quality, and production costs.

[0024] Optionally, parameters such as warp and weft density and weave specifications can be directly obtained from the design drawings of the jacquard fabric. The yarn count is usually determined by the specifications of the selected yarn raw materials, which can be obtained by checking the yarn label, product instructions, or conducting relevant tests. Based on the fabric's weaving requirements and the performance characteristics of the equipment, determine the equipment adaptation parameters. For example, select a loom model that can meet the requirements of weaving complex fabric patterns and set appropriate loom operating parameters.

[0025] Optionally, as an example, the specific parameter items and data formats corresponding to the above parameter types are shown in the table below:

[0026]

[0027] Step S102: Obtain the corresponding preset process parameters according to the weaving parameters, and predict the energy consumption per unit length of fabric based on the weaving parameters and the preset process parameters using a pre-trained Transformer energy consumption prediction model.

[0028] In some embodiments, the initial production parameters matched based on weaving parameters, such as loom speed, shedding time, yarn tension, workshop temperature and humidity, are typically matched based on a rule base to preset process parameters, or based on a database of historical data. This embodiment does not impose any limitations. For example, the rule base is shown in the table below.

[0029]

[0030] In some embodiments, the Transformer energy consumption prediction model is based on the classic Transformer architecture and is designed to meet the prediction requirements of weaving parameters and energy consumption. Its structure mainly consists of three parts: the input layer, the encoder layer, and the output layer.

[0031] The input layer includes:

[0032] Parameter Encoding: The acquired weaving parameters and process parameters, such as warp and weft density, weave specifications, yarn count, as well as equipment adaptation parameters, loom speed, shedding time, yarn tension, workshop temperature and humidity, are encoded. For numerical parameters (warp and weft density, yarn count, loom speed, shedding time, yarn tension, workshop temperature and humidity), normalization is performed to place them within a specific range (e.g., [0,1]) to facilitate model learning; for categorical parameters (equipment type, weave specifications in equipment adaptation parameters), one-hot encoding or embedded encoding is used to convert them into vector form.

[0033] Positional encoding: Since the Transformer architecture itself lacks the ability to capture sequence order information, and the order of weaving parameters has a potential impact on energy consumption prediction, positional encoding is introduced. By adding a positional information vector to each parameter, the model can distinguish parameters at different locations, thereby better learning the relationships between parameters. For example, although warp density and weft density have similar numerical properties, they play different roles in the fabric structure; positional encoding helps the model identify this difference.

[0034] The encoder layer includes:

[0035] Multi-head attention mechanism: This mechanism uses multiple "heads" to compute attention across different representation subspaces in parallel, capturing the correlations between weaving parameters from multiple perspectives. Each "head" independently calculates the attention weights between elements in the parameter sequence to determine which parameters are more critical for energy consumption prediction. For example, when predicting energy consumption of complex jacquard fabrics, the multi-head attention mechanism might find a higher correlation weight between the frequency of changes in fabric specifications and energy consumption, while in simple fabrics, the warp and weft density might have a greater weight. The outputs of the multi-head attention mechanism are then spliced ​​and linearly transformed to obtain a fused parameter representation.

[0036] Feedforward Neural Network: After the output of the multi-head attention mechanism, a fully connected feedforward neural network is connected to further extract features and perform nonlinear transformations on the fused parameter representation, enhancing the model's ability to learn complex relationships and uncovering deeper connections between parameters and energy consumption.

[0037] Layer normalization: Layer normalization is introduced in both multi-head attention mechanisms and feedforward neural networks to normalize the input data of each layer, thereby accelerating model training convergence and improving the stability and generalization ability of the model.

[0038] Preferably, the weights of each attention head in the multi-head attention mechanism of the pre-trained Transformer model are dynamically adjusted based on the direct impact weights of each weaving parameter and its corresponding preset process parameter on energy consumption, as well as the joint impact weights of the correlation between each weaving parameter and the correlation between each preset process parameter on energy consumption.

[0039] In some embodiments, the weight allocation of each attention point is adaptively adjusted based on the degree of influence of different weaving parameters on energy consumption and the correlation between the parameters. Optionally, the basic degree of influence of the parameters on energy consumption is determined by statistical analysis of historical data, typically represented by a normalized coefficient (range [0,1]). As an example, the direct influence of parameters on weights is shown in the following table:

[0040]

[0041] Optionally, the synergistic effect of parameter relationships on energy consumption can be determined through mutual information, correlation coefficients, or process knowledge, and represented by a joint weight matrix. As an example, the joint weights of parameters are shown in the table below:

[0042]

[0043] Specifically, as an example, dynamically adjusting the weights of each attention head in the multi-head attention mechanism of the pre-trained Transformer model can include:

[0044] Data input and weight initialization:

[0045] Input parameters: speed 480 r / min (direct weight 0.8), tension 16 cN (direct weight 0.7), warp density 280 threads / cm (direct weight 0.6); identification parameter pairs: "speed-tension" joint weight 0.9, "speed-warp density" joint weight 0.5.

[0046] Attention head weights are dynamically allocated:

[0047] First (responsible for independent speed influence): Weight = default 0.2 × direct weight 0.8 = 0.16;

[0048] First 2 (responsible for independent influence of tension): Weight = default 0.2 × direct weight 0.7 = 0.14;

[0049] Head 3 (responsible for the combined velocity-tension effect): Weight = default 0.1 × combined weight 0.9 = 0.09 (significantly higher than other combined heads);

[0050] Head 4 (responsible for speed-detailed joint influence): Weight = default 0.1 × joint weight 0.5 = 0.05.

[0051] Through the above embodiments, the weight adjustment mechanism of this embodiment allows the model to prioritize core parameters according to direct weights when the parameter combination belongs to a "normal scenario" (e.g., medium speed + medium tension); when the parameter combination belongs to a "high-risk scenario" (e.g., high speed + low tension, high density + coarse yarn), the model automatically activates specific attention points through joint weights to predict "energy consumption anomalies that may be caused by parameter conflicts". Dynamic weight adjustment can improve the model's adaptability to complex process scenarios.

[0052] Step S103: In response to the energy consumption per unit fabric length exceeding a preset threshold, the process parameters are adjusted using a reinforcement learning model to obtain the first optimized process parameters that meet the energy consumption threshold.

[0053] In some embodiments, reinforcement learning models optimize process parameters through a "state-action-reward" cyclical mechanism:

[0054] Status: The process parameters (such as loom speed, yarn tension, temperature and humidity) and energy consumption value during the current weaving process.

[0055] Action: Adjustment of process parameters, such as reducing the loom speed by 5% or increasing the yarn tension by 10%.

[0056] Rewards: Feedback is provided to the model based on the adjusted energy consumption results. A positive reward is given if the energy consumption per unit fabric length decreases and falls below a preset threshold; conversely, a negative reward is given. By continuously trying different actions, the model learns the parameter adjustment strategy that yields the maximum reward (i.e., energy consumption meets the target).

[0057] As an example, the specific execution process of the model may include:

[0058] Initialization state: Obtain the current production process parameters (e.g., loom speed 600r / min, yarn tension 150cN) and energy consumption per unit fabric length (e.g., 1.2kW·h / m). The preset energy consumption threshold is 1.0kW·h / m. If the energy consumption exceeds the standard, reinforcement learning optimization is triggered.

[0059] Execution of actions: The reinforcement learning model selects an action from the action space, such as reducing the loom speed by 10% (adjusting it to 540 r / min) while fine-tuning the yarn tension to 140 cN.

[0060] Observations and Rewards: After the action was performed, the new energy consumption per unit fabric length was monitored (e.g., reduced to 1.1 kW·h / m). Since the energy consumption was still above the threshold, a negative reward was given, but it was an improvement over the previous result, and the reward value was higher than when there was no adjustment at all.

[0061] Iterative optimization: Based on reward feedback, the model continues to try other action combinations, such as further reducing the loom speed to 500 r / min and adjusting the workshop temperature to reduce yarn friction energy consumption. After multiple iterations, a set of parameters (loom speed 480 r / min, yarn tension 130 cN, workshop temperature 23℃) was found, which reduced the energy consumption per unit fabric length to 0.95 kW·h / m, meeting the preset threshold, and thus obtaining the first optimized process parameters.

[0062] Step S104: Obtain the first correlation degree between the optimized parameters in the first optimized process parameters and the surface defects of the jacquard fabric; conduct a preset number of trial productions based on the first optimized process parameters; identify the surface defects of the jacquard fabric based on a convolutional neural network; dynamically input the surface defects into the reinforcement learning model as a state input factor; and predict the second optimized process parameters, wherein the preset number is positively correlated with the correlation degree.

[0063] In some embodiments, the impact of a first optimization parameter (e.g., speed, tension) on quality is evaluated; the higher the correlation, the more trial productions are conducted to ensure data sufficiency; CNN identifies defects in real time, which serves as new state input for reinforcement learning, forcing the algorithm to balance quality in energy consumption optimization.

[0064] The first correlation refers to the statistical correlation between the first optimized process parameter (e.g., speed, tension, etc.) and surface defects (skipped yarns, broken ends, etc.) of the jacquard fabric. As an example, historical production records are shown in the table below:

[0065]

[0066] Based on the aforementioned historical production records, the correlation can be quantified using the Pearson correlation coefficient, which simplifies to the following formula:

[0067]

[0068] Where, x i y i These are sample values ​​of the variable. It is the sample mean.

[0069] The correlation between "speed" and "skipping rate" in the above data is calculated to be (r = -0.82) (negative correlation, with a higher skipping rate at lower speeds). It is understood that other existing correlation quantification methods can be chosen to measure the degree of correlation depending on the data type; this embodiment will not elaborate further.

[0070] For example, the correlation degree is normalized, mapping the results of different methods to the [0,1] interval; the decomposition criteria can be shown in the table below:

[0071]

[0072] Examples of the grading results include: correlation between speed and yarn skipping rate: 0.85 (high); correlation between tension and yarn breakage rate: 0.58 (medium); and correlation between humidity and defects: 0.21 (low).

[0073] In some embodiments, the preset quantity trial production refers to a limited number of actual production verifications based on the first optimized process parameters (such as the adjusted loom speed, tension, etc.). Through trial production, the quality / energy consumption out of control caused by direct large-scale production is avoided, and the prediction bias of the reinforcement learning model is corrected by real production data. While optimizing energy consumption, it verifies whether the fabric defects exceed the acceptable range.

[0074] As an example, a high correlation parameter (e.g., speed) corresponds to more trial production times (correlation 0.85 → number of trial production times = 20 + 0.85 × 10 = 28 times); a low correlation parameter (e.g., humidity) can reduce the number of trial production times (correlation 0.21 → number of trial production times = 5 + 0.21 × 5 ≈ 6 times).

[0075] In some embodiments, during trial production, a certain length of fabric is produced each time, and surface images are captured using an industrial camera; the captured images are then identified using a convolutional neural network (CNN), and the identification result can be, for example, as follows:

[0076] First 10 times: Skip rate 0.5%-0.8%, yarn breakage rate 0.1%-0.2%.

[0077] Tests 11-20: Skip rate 0.7%-1.0%, yarn breakage rate 0.1%-0.3%.

[0078] Tests 21-27: Skip rate 0.9%-1.2%, yarn breakage rate 0.2%-0.4%.

[0079] In some embodiments, the state space of the reinforcement learning model is expanded based on the above identification results, and the surface defects are dynamically input into the reinforcement learning model as state input factors. For example, the original state space is: [speed, tension, humidity, energy consumption]; the expanded state is: [speed, tension, humidity, energy consumption, yarn skipping rate, yarn breakage rate].

[0080] Preferably, in response to the first correlation degree being higher than a first threshold, the surface defects are simultaneously used as the reinforcement learning reward constraint. Specifically, the type, quantity, and location coordinate information of the surface defects are encoded into a state vector, and together with the original process parameters and energy consumption indicators, a new state space is formed. The reward function of the reinforcement learning model is adjusted, and the reward function includes an energy consumption target achievement reward, an energy consumption reduction reward, a defect penalty, and a parameter stability reward. Based on the new state space and the adjusted reward function, the reinforcement learning model is iteratively optimized to output a second optimized process parameter that balances energy consumption and quality.

[0081] Specifically, when the first correlation (such as the correlation between speed and yarn skipping) is higher than the first threshold (usually set to 0.5-0.7), it indicates that the impact of process parameter adjustments on quality is no longer negligible. In this embodiment, surface defects are incorporated into the reward constraint system of reinforcement learning.

[0082] As an example, a specific process may include:

[0083] 1. State space reconstruction

[0084] Defect information coding: One-hot encoding is used to distinguish defect types, such as skipped yarn → [1,0,0], broken yarn → [0,1,0], weft shrinkage → [0,0,1]. Quantification of quantity and density is used, such as number of skipped yarns / 100m (e.g., 2.5 times → normalized to 0.5, maximum value 5 times); Defect density = number of defects / fabric area (e.g., 1.2 defects / m²). 2 → Normalization 0.6, maximum value 2 / m 2 ); position coordinate processing, such as extracting the relative coordinates of the defect on the fabric surface (e.g., horizontal position x∈[0,1], vertical position y∈[0,1]), and calculating the distribution entropy (the higher the entropy value, the more dispersed it is).

[0085] Example of the new state space composition: Original state: [Speed ​​500, Tension 180, Energy consumption 1.6, Warp density 300]; Expanded state: [Speed ​​500, Tension 180, Energy consumption 1.6, Warp density 300, Yarn skips [1,0,0], Quantity 0.5, Density 0.6, Distribution entropy 0.8].

[0086] 2. Reward Function Reconstruction

[0087] For example, the multi-dimensional reward items are broken down as shown in the table below:

[0088]

[0089]

[0090] The reward function dynamically adjusts its weights, automatically increasing the quality penalty weight when the number of defects exceeds the process standard. For example, the original penalty coefficient was 30 for skipped yarn, but became 40 when the standard was exceeded (e.g., if the standard for skipped yarn is ≤ 1 time / 100m, and the actual measurement is 2 times, the coefficient is 1.3).

[0091] The comprehensive reward calculation can be: R = Energy consumption meets the standard (+100) + Energy consumption decreases (+10) + Flaw penalty (-60) + Stability (+20) = +70.

[0092] Through the above embodiments, by deeply integrating visual defect data into the reinforcement learning framework, energy consumption optimization and quality assurance are automatically balanced in highly correlated scenarios, thereby improving the green production of jacquard fabrics while ensuring the quality of jacquard fabrics.

[0093] Preferably, the reinforcement learning model is an improved DDPG algorithm. The reinforcement learning model structure includes: a dual-policy network module, comprising a main policy network and an auxiliary policy network. The main policy network is used for routine energy consumption optimization, and the auxiliary policy network is used to handle high-risk process parameter adjustments. The two policy networks share a bottom-level feature extraction layer, and their output branches are independent and output actions through weighted fusion; a hierarchical attention module, set in the Critic network, comprising a parameter layer and a correlation layer. The parameter layer calculates attention weights for weaving parameters, process parameters, and quality parameters respectively. The correlation layer analyzes cross-layer correlations between parameters and dynamically adjusts the weights of each parameter combination; and a multimodal information fusion module, which fuses the defect features identified by the convolutional neural network with the process parameters through an encoder. The dual-policy network module, the hierarchical attention module, and the multimodal information fusion module work together to achieve a balance between energy consumption and quality optimization of the jacquard fabric process parameters.

[0094] Specifically, the improved DDPG algorithm handles the multimodal requirements of parameter adjustment through a dual-policy network, captures the complex relationships between parameters through hierarchical attention, and integrates visual and numerical data through multimodal fusion. The main policy handles routine optimization, while the auxiliary policy addresses high-risk parameter adjustments (e.g., a sudden drop in speed may lead to a surge in yarn skipping). The hierarchical attention mechanism not only focuses on the impact of single parameters but also captures the synergistic effects of parameter combinations such as speed, tension, and humidity. Cross-modal knowledge fusion transforms the visual defect features extracted by the CNN (e.g., yarn skipping location and density) into computable decision-making criteria.

[0095] Specifically, the dual-policy network module includes:

[0096] 1. Shared feature extraction layer:

[0097] Input: [process parameters, energy consumption, weaving parameters, defect features output by CNN];

[0098] Processing: Extract shared feature representations (e.g., 128-dimensional vectors) through a multi-layer fully connected network.

[0099] Independent output branch:

[0100]

[0101] 2. Weight Calculation: Dynamically adjust the weights of the two networks based on parameter correlation: When the correlation between "speed-skipping yarn" > 0.7, the weight of the auxiliary strategy is increased to 0.6. The specific calculation formula can be simplified as follows:

[0102] a = w1·π1(s) + w2·π2(s)

[0103] The formula is used to fuse the outputs of the main strategy network and the auxiliary strategy network to generate the final action for adjusting process parameters, where w1+w2=1.

[0104] Here, 'a' represents the final output action, such as the adjustment instruction for process parameters (speed, tension, etc.), which determines how the parameters are changed in actual production; w1 and w2 represent the weights of the main strategy network and the auxiliary strategy network, respectively, and their sum is 1, representing the "influence ratio" of their respective outputs in the final action. When the correlation is high (e.g., the correlation between "speed-skipped yarn" > 0.7), w2 (the weight of the auxiliary strategy) is increased, making the auxiliary strategy more influential on decision-making in high-risk scenarios; π1(s) is the output of the main strategy network, which, based on conventional energy consumption optimization logic, provides conservative and stable parameter adjustment suggestions (e.g., small-scale speed adjustments) when parameter fluctuations are small and quality risks are low; π2(s) is the output of the auxiliary strategy network, which deals with high-risk parameter adjustments. When conventional optimization is difficult to achieve the target and aggressive adjustments are required (e.g., energy consumption is seriously exceeded), it provides a parameter adjustment scheme with a large magnitude, while controlling quality risks as much as possible through network design; 's' represents the state input of reinforcement learning, which includes weaving parameters (speed, tension, etc.), energy consumption indicators, and quality parameters (number and type of defects, etc.).

[0105] Specifically, the hierarchical attention module includes:

[0106] 1-parameter layer attention

[0107] Calculate the importance weights of each parameter: Input: (x = [speed, tension, humidity, yarn skipping rate, yarn breakage rate]); Output: Attention weight vector (a p = [0.35, 0.25, 0.1, 0.2, 0.1] (where speed has the highest weight); indicating that in the current state, speed accounts for 35% of the impact on the decision.

[0108] 2. Relationship layer attention

[0109] Uncovering synergistic effects between parameters: Calculate the correlation weight matrix of parameter pairs, for example, as follows:

[0110]

[0111] As illustrated in the table above, when the speed decreases, the model automatically increases tension compensation by using the "speed-tension" correlation weight (0.8) to reduce the risk of yarn skipping.

[0112] 3. Dynamic weight adjustment

[0113] Adjust attention allocation based on the current state: When the yarn skipping rate is >1 time / 100m, the parameter layer automatically increases the weight of "yarn skipping rate" to 0.3; the association layer enhances the association weight of "speed-yarn skipping" and "tension-yarn skipping" to 0.9.

[0114] Specifically, the multimodal information fusion module includes:

[0115] 1. Defect Feature Coding

[0116] The output of the CNN is processed as follows:

[0117] Spatial features: The 128×128 defect feature map is compressed into a 1024-dimensional vector by global average pooling;

[0118] Semantic features: Information such as defect type and quantity is encoded into a 256-dimensional vector through a fully connected layer;

[0119] Final defect vector: concatenated into a 1280-dimensional feature representation.

[0120] 2. Cross-modal fusion process

[0121] Process parameter encoding: Numerical parameters such as [speed, tension, humidity] are mapped into 512-dimensional vectors through an embedding layer;

[0122] In some embodiments, a multilayer perceptron (MLP) is used to integrate process parameter vectors (numerical features) and defect vectors (visual features) into a unified feature representation. This can be represented as:

[0123] z = MLP([process parameter vector; defect vector])

[0124] The MLP structure is: 512+1280→1024→512→256

[0125] Among them, the process parameter vector (512-dimensional) contains numerical parameters such as speed, tension, and humidity, which are mapped to a high-dimensional vector through the embedding layer; the defect vector (1280-dimensional) is a visual feature extracted by CNN, containing information such as defect type, quantity, and location; the two vectors are concatenated according to their dimensions to form a joint feature vector of 1792 dimensions (512+1280).

[0126] For example, the number of layers in an MLP network is 3 fully connected layers. The dimensionality change can be represented as follows:

[0127] 1792 (concatenated input) → 1024 (first-layer dimensionality reduction) → 512 (feature abstraction) → 256 (final representation). Activation functions: ReLU is typically used in intermediate layers, and Tanh or linear activation is used in the output layer. The specific calculation process will not be detailed in this embodiment. Through the above feature compression, redundant information (such as collinearity in process parameters) is reduced; cross-modal semantic relationships are extracted (such as the implicit relationship between "speed reduction" and "increased yarn skipping"); the 256-dimensional z-vector is a "semantically aligned" feature representation, containing information on both process parameters and quality status, and can be directly used as the state input for reinforcement learning.

[0128] Specifically, an example of three-module collaborative optimization is as follows: For instance, in the production of high-density jacquard fabric, the initial parameters are [speed 520 r / min, tension 190 cN, humidity 55%], energy consumption 1.6 kW·h / m (threshold 1.5), and skip rate 1.2 times / 100m (standard ≤ 1). The specific optimization process may include:

[0129] 1. State Input and Feature Extraction:

[0130] CNN detection: The skipped threads are mainly concentrated in the central area of ​​the fabric, with a density of 1.5 threads / m. 2 ;

[0131] Multimodal fusion: Encode speed (520), yarn skipping density (1.5), etc. into a 256-dimensional fusion vector.

[0132] 2. Attention Allocation:

[0133] Parameter layer: speed weight 0.35, yarn skipping rate 0.3, tension 0.25;

[0134] Association layer: "Speed-skip yarn" association weight 0.85, "Tension-skip yarn" 0.7.

[0135] 3. Strategic Decision-Making:

[0136] Main strategy recommendation: reduce speed to 500 r / min and increase tension to 195 cN; Auxiliary strategy recommendation: reduce speed to 480 r / min and increase tension to 200 cN; Fusion output: speed 490 r / min (w1 = 0.6, w2 = 0.4), tension 198 cN.

[0137] 4. Implementation and Feedback:

[0138] New status: Energy consumption 1.48 kW·h / m, yarn skipping rate 0.9 times / 100m;

[0139] Rewards: Energy consumption meets target +100, skipping yarn penalty -27, total reward +73.

[0140] Step S105: Control the production of the jacquard fabric based on the second optimized process parameters.

[0141] In some embodiments, controlling the production of jacquard fabric based on the second optimized process parameters involves converting the optimal parameter combination output by the reinforcement learning model (e.g., speed 510 r / min, tension 185 cN, humidity 58%) into actual production instructions, thereby ensuring the continuous achievement of energy consumption indicators (e.g., ≤1.5 kW·h / m) while guaranteeing the surface quality of the fabric (skip rate ≤ 1 time / 100m).

[0142] A green production optimization management method for jacquard fabrics, corresponding to the above embodiment, Figure 2 This diagram illustrates a structural block diagram of a green production optimization management system for jacquard fabrics provided in an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown.

[0143] See Figure 2 As shown in the embodiment of this application, a green production optimization management system 200 for jacquard fabrics includes:

[0144] The first acquisition module 201 is used to acquire the weaving parameters of the jacquard fabric, wherein the weaving parameters include at least one of warp and weft density, structure specification, yarn properties, or equipment adaptation parameters.

[0145] The first prediction module 202 is used to obtain the corresponding preset process parameters according to the weaving parameters, and based on the weaving parameters and the preset process parameters, use a pre-trained Transformer energy consumption prediction model to predict the energy consumption per unit length of fabric.

[0146] The first adjustment module 203 is used to adjust the process parameters using a reinforcement learning model in response to the energy consumption per unit length exceeding a preset threshold, so as to obtain the first optimized process parameters that meet the energy consumption threshold.

[0147] The second adjustment module 204 is used to obtain the first correlation degree between the parameters involved in the optimization in the first optimized process parameters and the surface defects of the jacquard fabric, conduct a preset number of trial productions based on the first optimized process parameters, identify the surface defects of the jacquard fabric based on the convolutional neural network, dynamically input the surface defects into the reinforcement learning model as state input factors, and predict the second optimized process parameters, wherein the preset number is positively correlated with the correlation degree.

[0148] The production control module 205 is used to control the production of the jacquard fabric based on the second optimized process parameters.

[0149] This invention also provides an electronic device comprising: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program, when executed by the processor, implements the method as described in any of the foregoing embodiments.

[0150] This invention also provides a computer program product comprising a computer program that can be executed by a processor to implement the method as described in any of the foregoing embodiments.

[0151] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0152] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A green production optimization management method for jacquard fabrics, characterized in that, The methods and steps include the following: Obtain the weaving parameters of the jacquard fabric, wherein the weaving parameters include at least one of warp and weft density, weave specification, yarn count, or equipment adaptation parameters; According to the weaving parameters, the corresponding preset process parameters are obtained. Based on the weaving parameters and the preset process parameters, the energy consumption per unit length of fabric is predicted using a pre-trained Transformer energy consumption prediction model. Based on the direct impact weights of each weaving parameter and its corresponding preset process parameter on energy consumption, as well as the joint impact weights of the correlation between each weaving parameter and the correlation between each preset process parameter on energy consumption, the weights of each attention head in the multi-head attention mechanism in the pre-trained Transformer model are dynamically adjusted. In response to the energy consumption per unit fabric length exceeding a preset threshold, the process parameters are adjusted using a reinforcement learning model to obtain the first optimized process parameters that meet the energy consumption threshold. Obtain the first correlation degree between the optimized parameters in the first optimized process parameters and the surface defects of the jacquard fabric, conduct a preset number of trial productions based on the first optimized process parameters, identify the surface defects of the jacquard fabric based on the convolutional neural network, dynamically input the surface defects into the reinforcement learning model as state input factors, and predict the second optimized process parameters, wherein the preset number is positively correlated with the correlation degree. In response to the first correlation degree being higher than the first threshold, the surface defects are simultaneously used as the reinforcement learning reward constraint. Specifically, the type of the surface defects is encoded using one-hot encoding, the quantity and density are quantified, and the distribution entropy of the position coordinates is calculated and encoded into a state vector, which together with the original process parameters and energy consumption indicators constitutes a new state space. The reward function of the reinforcement learning model is adjusted, and the reward function includes energy consumption compliance reward, energy consumption reduction reward, defect penalty, and parameter stability reward. When the number of defects exceeds the process standard, the quality penalty weight is automatically increased. Based on the new state space and the adjusted reward function, the reinforcement learning model is used for iterative optimization to output a second optimized process parameter that balances energy consumption and quality. The production of the jacquard fabric is controlled based on the second optimized process parameters.

2. The green production optimization management method for jacquard fabrics according to claim 1, characterized in that: The reinforcement learning model is specifically an improved DDPG algorithm. The reinforcement learning model structure includes a dual-policy network module, which includes a main policy network and an auxiliary policy network. The main policy network is used for routine energy consumption optimization, and the auxiliary policy network is used to handle high-risk process parameter adjustments. The two policy networks share the underlying feature extraction layer, and their output branches are independent and output actions through weighted fusion. The hierarchical attention module, set within the Critic network, includes a parameter layer and a correlation layer. The parameter layer calculates attention weights for weaving parameters, process parameters, and quality parameters respectively. The correlation layer analyzes cross-layer correlations between parameters and dynamically adjusts the weights of each parameter combination. The multimodal information fusion module fuses the defect features identified by the convolutional neural network with the process parameters through an encoder. The dual-strategy network module, the hierarchical attention module, and the multimodal information fusion module work together to achieve a balance between energy consumption and quality optimization of the jacquard fabric process parameters.

3. A green production optimization management system for jacquard fabrics, characterized in that, include: The first acquisition module is used to acquire the weaving parameters of the jacquard fabric, wherein the weaving parameters include at least one of warp and weft density, weave specification, yarn count, or equipment adaptation parameters. The first prediction module is used to obtain the corresponding preset process parameters according to the weaving parameters, and based on the weaving parameters and the preset process parameters, use a pre-trained Transformer energy consumption prediction model to predict the energy consumption per unit length of fabric. The second adjustment module is used to dynamically adjust the weights of each attention head in the multi-head attention mechanism of the pre-trained Transformer model based on the direct impact weights of each weaving parameter and its corresponding preset process parameter on energy consumption, as well as the joint impact weights of the correlation between each weaving parameter and the correlation between each preset process parameter on energy consumption. The first adjustment module is used to adjust the process parameters using a reinforcement learning model in response to the energy consumption per unit fabric length exceeding a preset threshold, so as to obtain the first optimized process parameters that meet the energy consumption threshold. The second adjustment module is used to obtain the first correlation degree between the parameters involved in the optimization in the first optimized process parameters and the surface defects of the jacquard fabric, conduct a preset number of trial productions based on the first optimized process parameters, identify the surface defects of the jacquard fabric based on the convolutional neural network, dynamically input the surface defects into the reinforcement learning model as state input factors, and predict the second optimized process parameters, wherein the preset number is positively correlated with the correlation degree. The first setting module is used to respond to the first correlation degree being higher than the first threshold by simultaneously using the surface defects as a reward constraint for the reinforcement learning. Specifically, the type of the surface defects is encoded using one-hot encoding, the quantity and density are quantified, and the distribution entropy of the position coordinates is calculated and encoded into a state vector, which together with the original process parameters and energy consumption indicators constitutes a new state space; the reward function of the reinforcement learning model is adjusted, the reward function including energy consumption compliance reward, energy consumption reduction reward, defect penalty, and parameter stability reward; when the number of defects exceeds the process standard, the quality penalty weight is automatically increased; Based on the new state space and the adjusted reward function, the reinforcement learning model is used for iterative optimization to output a second optimized process parameter that balances energy consumption and quality. The production control module is used to control the production of the jacquard fabric based on the second optimized process parameters.

4. The green production optimization management system for jacquard fabrics according to claim 3, characterized in that, include: The reinforcement learning model is specifically an improved DDPG algorithm. The reinforcement learning model structure includes a dual-policy network module, which includes a main policy network and an auxiliary policy network. The main policy network is used for routine energy consumption optimization, and the auxiliary policy network is used to handle high-risk process parameter adjustments. The two policy networks share the underlying feature extraction layer, and their output branches are independent and output actions through weighted fusion. The hierarchical attention module, set within the Critic network, includes a parameter layer and a correlation layer. The parameter layer calculates attention weights for weaving parameters, process parameters, and quality parameters respectively. The correlation layer analyzes cross-layer correlations between parameters and dynamically adjusts the weights of each parameter combination. The multimodal information fusion module fuses the defect features identified by the convolutional neural network with the process parameters through an encoder. The dual-strategy network module, the hierarchical attention module, and the multimodal information fusion module work together to achieve a balance between energy consumption and quality optimization of the jacquard fabric process parameters.

5. An electronic device, characterized in that: The electronic device includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program, when executed by the processor, implements the method as described in any one of claims 1-2.

6. A computer program product, characterized in that: The computer program product includes a computer program that can be executed by a processor to implement the method as described in any one of claims 1-2.

Images