Intelligent textile printing and dyeing process optimization method and system

Abstract

The application provides a kind of intelligent textile printing and dyeing process optimization method and system, according to the parameter of membership function determined according to the characteristics of current batch material, fuzzy process characteristics are obtained by fuzzy processing of real-time process parameters using membership function;According to the correlation degree of the historical prediction accuracy of fuzzy rule and process benefit, the activation contribution of each fuzzy rule is calculated, and then the process coordination state is obtained;Sensor time series data and process coordination state are input into multi-modal time series main capsule layer to obtain main capsule output vector, calculate the confidence of the vector, and adjust the influence of the vector when it is routed to the upper capsule layer subsequently according to the confidence, the upper capsule layer generates a set of control parameters for the subsequent process;Obtain the quality detection data of subsequent process, obtain multi-objective deviation vector according to quality detection data and control parameter set, and use multi-objective deviation vector to optimize the upper capsule layer, to obtain the optimized control parameter set.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of textiles, and particularly relates to an intelligent textile printing and dyeing process optimization method and system. BACKGROUND

[0002] As a component of traditional manufacturing industry, the textile printing and dyeing industry has a complex and variable production process involving many physical and chemical processes, which has an impact on product quality, resource consumption and environmental protection. Traditional textile printing and dyeing process control relies on manual experience and fixed process parameter settings. This approach is difficult to adapt to differences in raw materials, fluctuations in equipment state, and changes in market demand for product diversification and high quality. On the one hand, the physical and chemical properties of raw materials such as dyes, auxiliaries and fabrics vary between batches. Even if the same material is provided by the same supplier, subtle changes in its characteristics can cause color differences, poor levelness and other quality problems in the final product. Operators often need to make exploratory adjustments based on experience, which is not only time-consuming and labor-intensive, but also difficult to ensure the stability and consistency of the optimization results. On the other hand, the performance of printing and dyeing equipment changes subtly over time, such as pipe fouling and sensor aging. These factors can also interfere with the accurate execution of process parameters and the stable maintenance of process state, causing the fixed process parameters to no longer be the optimal choice. Traditional control systems have poor adaptability to such dynamic changes, often lagging behind the occurrence of problems, resulting in unnecessary waste and waste. In order to improve the level of intelligence in textile printing and dyeing, the academic and industrial communities have introduced expert systems, fuzzy logic, neural networks and other artificial intelligence technologies to model and optimize the printing and dyeing process. Fuzzy logic has certain advantages in dealing with the fuzziness and uncertainty in the printing and dyeing process, and can convert the experience and knowledge of operators into executable fuzzy rules. Neural networks are good at extracting non-linear relationships from large amounts of historical data for predicting key process indicators or diagnosing faults. The determination of the membership function of the traditional fuzzy system often relies on expert experience, making it difficult to dynamically adapt to real-time changes in material characteristics and process disturbances, which limits the accuracy of fuzzy reasoning. The construction and optimization of the fuzzy rule base also takes a long time, and the effectiveness of the rules is not closely related to actual production efficiency, making it difficult to ensure the continued applicability of the rules. For the fusion processing of time series data and multi-source heterogeneous information, traditional methods often fail to fully exploit their deep features and dynamic associations. For example, how to effectively combine high-frequency sensor data with macroscopic process coordination state information to accurately predict and actively control future process trends remains a technical challenge that needs to be addressed. At the same time, existing optimization methods mostly focus on the optimization of local processes, lack systematic consideration of the associated impact of the entire printing and dyeing process and global coordination optimization capability, and the feedback correction mechanism for optimization results is usually not precise and efficient, making it difficult to achieve continuous self-evolution of the model and precise achievement of multi-dimensional quality targets. SUMMARY

[0003] To solve the above problems, the application provides an intelligent textile printing and dyeing process optimization method, comprising:

[0004] Obtain real-time and historical process parameters of the printing and dyeing process, determine the parameters of the membership function according to the characteristics of the current batch of materials, and obtain fuzzy process characteristics by fuzzy processing of the real-time process parameters using the membership function;

[0005] Calculate the activation contribution degree of each fuzzy rule according to the correlation degree of the historical prediction accuracy of the fuzzy rule and the process benefit, and obtain the process synergy state by using the activation contribution degree of the fuzzy rule and the fuzzy process characteristics;

[0006] Input the sensor time series data and the process synergy state into the multi-modal time series main capsule layer to obtain a main capsule output vector, calculate the confidence of the vector, and adjust the influence of the vector when it is routed to the upper capsule layer according to the confidence, and the upper capsule layer generates a set of control parameters for the subsequent process;

[0007] Obtain quality detection data of the subsequent process, obtain a multi-objective deviation vector according to the quality detection data and the set of control parameters, and optimize the upper capsule layer using the multi-objective deviation vector to obtain an adjusted and optimized set of control parameters.

[0008] Optionally, the parameters of the membership function are determined according to the characteristics of the current batch of materials, comprising:

[0009] Obtain the material characteristic vector of the current processing batch of materials, input the material characteristic vector into the radial basis function network, and obtain the center point value and standard deviation value of the Gaussian membership function of the preset fuzzy subset for each process parameter.

[0010] Optionally, the fuzzy process characteristics are obtained by fuzzy processing of the real-time process parameters using the membership function, comprising:

[0011] Obtain the Gaussian membership function of the fuzzy subset corresponding to each process parameter and the corresponding center point value and standard deviation value;

[0012] Substitute the numerical value of the process parameter into the Gaussian membership function of each fuzzy subset corresponding to the process parameter Obtain the membership degree value of the process parameter for each fuzzy subset, wherein x is the parameter value, c is the center point value, and σ is the standard deviation value;

[0013] All membership degree values of each process parameter constitute the fuzzy process characteristics of the process parameter.

[0014] Optionally, the activation contribution degree of each fuzzy rule is calculated according to the correlation degree of the historical prediction accuracy of the fuzzy rule and the process benefit, comprising:

[0015] For each fuzzy rule in the fuzzy rule base, all historical process batches whose membership products of the fuzzy variables in the condition part of the rule are greater than a preset value are filtered out from the database of historical process parameters;

[0016] For each historical process batch, the root mean square error (RMSE) between the intermediate process target predicted value defined by the result part of the fuzzy rule and the actually recorded intermediate process target value of the batch is calculated, and the historical prediction accuracy of the rule is ;

[0017] At least three final product quality indicators and at least two process benefit indicators corresponding to each historical process batch are obtained, and the Pearson correlation coefficients between the prediction accuracy of the rule and each quality indicator and benefit indicator are calculated, and the weighted average value of each Pearson correlation coefficient is taken as the process benefit correlation degree CB;

[0018] The activation contribution degree of the fuzzy rule is calculated according to the formula , wherein , is a preset coefficient, and Optionally, the process synergy state obtained by using the activation contribution degree of the fuzzy rule and the fuzzy process characteristics comprises:

[0019] For each fuzzy rule, the fuzzy process characteristics are substituted into the condition part of the rule, the total activation intensity of the condition part of the rule is calculated by using an algebraic product T-norm operator, the weighted activation intensity of the rule is obtained by multiplying the total activation intensity of the rule by the activation contribution degree ACD of the rule;

[0020] The weighted activation intensity of each rule is multiplied by the characteristic value of the process synergy state defined by the result part of the rule, the product results of all rules are summed, and then divided by the sum of the weighted activation intensities of all rules to obtain the process synergy state at the current time.

[0021] Optionally, the confidence of the vector is calculated by:

[0022] Within a current processing time window preset or dynamically adjusted according to the characteristics of the process stage, all main capsule output vectors output by the multi-modal time sequence main capsule layer are obtained, the L2 norm of each main capsule output vector is calculated, and the maximum value in the L2 norms of all main capsule output vectors within the time window is taken as a reference value;

[0023] The confidence of each main capsule output vector is obtained by dividing the L2 norm of the main capsule output vector by the reference value.

[0024] Optionally, the adjusting the influence of the vector when being routed to a subsequent capsule layer according to the confidence level, the subsequent capsule layer generating a set of regulation parameters for a subsequent process, comprises:

[0025] for each primary capsule vector output by the primary capsule layer and its corresponding confidence level :

[0026] for each high-level capsule j, the primary capsule vector is converted into a prediction vector by a transformation matrix ;

[0027] the input of the high-level capsule j is initialized as a zero vector;

[0028] in the routing iteration process, the dot product of the prediction vector of the current iteration round and the current state vector of the high-level capsule j is calculated, and the connection routing coupling coefficient between each primary capsule vector and the high-level capsule j is obtained by normalizing all high-level capsules by a softmax function; the input of the high-level capsule j is updated according to , and the state vector of the high-level capsule j is updated by a squash function;

[0029] after the iteration is completed, the state vectors of all high-level capsules are spliced to obtain a feature vector, and the feature vector is input into a multi-layer perception MLP to obtain a set of regulation parameters.

[0030] Optionally, the optimizing the high-level capsule layer by using the multi-objective deviation vector to obtain an adjusted and optimized set of regulation parameters, comprises:

[0031] a loss function value is calculated based on the multi-objective deviation vector; the gradient of the loss function value with respect to the weight parameters of each layer of the MLP is calculated by back propagation, and the gradient of the loss function value with respect to each element in the transformation matrix connecting the primary capsule layer and the high-level capsule layer is calculated by back propagation;

[0032] the weight parameters of the MLP and the transformation matrix are updated according to the gradient; the updated MLP and the transformation matrix are used to regenerate a set of regulation parameters.

[0033] The application further provides an intelligent textile printing and dyeing process optimization system, comprising:

[0034] ​a feature extraction unit configured to obtain real-time and historical process parameters of a printing and dyeing process, determine parameters of a membership function according to characteristics of a current batch of materials, and obtain fuzzy process characteristics by fuzzy processing of the real-time process parameters using the membership function;

[0035] a state acquisition unit configured to calculate an activation contribution degree of each fuzzy rule according to a correlation degree between a historical prediction accuracy of the fuzzy rule and a process benefit, and obtain a process coordination state by using the activation contribution degree of the fuzzy rule and the fuzzy process characteristics;

[0036] a parameter adjustment unit configured to input sensor time series data and the process coordination state into a multi-modal time series main capsule layer to obtain a main capsule output vector, calculate a confidence degree of the vector, and adjust an influence of the vector when the vector is routed to a higher capsule layer in a subsequent process according to the confidence degree, and the higher capsule layer generates a set of regulation and control parameters for a subsequent process;

[0037] an optimization unit configured to obtain quality detection data of the subsequent process, obtain a multi-objective deviation vector according to the quality detection data and the set of regulation and control parameters, and optimize the higher capsule layer using the multi-objective deviation vector to obtain an adjusted and optimized set of regulation and control parameters.

[0038] Optionally, the determining of the parameters of the membership function according to the characteristics of the current batch of materials comprises:

[0039] obtaining a material characteristic vector of a current batch of materials, inputting the material characteristic vector into a radial basis function network to obtain a center point value and a standard deviation value of a Gaussian membership function of a preset fuzzy subset for each process parameter.

[0040] Optionally, the fuzzy processing of the real-time process parameters using the membership function to obtain the fuzzy process characteristics comprises:

[0041] obtaining a Gaussian membership function of a fuzzy subset corresponding to each process parameter and a corresponding center point value and a standard deviation value;

[0042] substituting a numerical value of the process parameter into the Gaussian membership function of each fuzzy subset corresponding to the process parameter to obtain a membership degree value of the process parameter for each fuzzy subset, wherein x is the numerical value of the parameter, c is the center point value, and σ is the standard deviation value;

[0043] all membership degree values of each process parameter constitute a fuzzy process characteristic of the process parameter.

[0044] Optionally, the calculating of the activation contribution degree of each fuzzy rule according to the correlation degree between the historical prediction accuracy of the fuzzy rule and the process benefit comprises:

[0045] For each fuzzy rule in the fuzzy rule base, all historical process batches whose membership products of the fuzzy variables in the condition part of the rule are greater than a preset value are filtered out from the database of historical process parameters;

[0046] For each historical process batch, the root mean square error (RMSE) between the intermediate process target predicted value defined by the result part of the fuzzy rule and the actually recorded intermediate process target value of the batch is calculated, and the historical prediction accuracy of the rule is AP = 1 / (1+RMSE);

[0047] At least three final product quality indicators and at least two process benefit indicators corresponding to each historical process batch are obtained, the Pearson correlation coefficients between the prediction accuracy of the rule and each quality indicator and benefit indicator are calculated, and the weighted average value of each Pearson correlation coefficient is taken as the process benefit correlation degree CB;

[0048] According to the formula The activation contribution degree of the fuzzy rule is calculated, wherein 、 is a preset coefficient, and .

[0049] Optionally, the process synergy state is obtained by using the activation contribution degree of the fuzzy rule and the fuzzy process characteristics, comprising:

[0050] For each fuzzy rule, the fuzzy process characteristics are substituted into the condition part of the rule, the total activation intensity of the condition part of the rule is calculated by using the algebraic product T-norm operator, the weighted activation intensity of the rule is obtained by multiplying the total activation intensity of the rule by the activation contribution degree ACD of the rule;

[0051] The weighted activation intensity of each rule is multiplied by the characteristic value of the process synergy state defined by the result part of the rule, the product results of all rules are summed, and then divided by the sum of the weighted activation intensities of all rules to obtain the process synergy state at the current time.

[0052] Optionally, the confidence of the vector is calculated, comprising:

[0053] Within a current processing time window preset or dynamically adjusted according to the process stage characteristics, all main capsule output vectors output by the multi-modal time sequence main capsule layer are obtained; the L2 norm of each main capsule output vector is calculated; the maximum value in the L2 norms of all main capsule output vectors within the time window is taken as a reference value;

[0054] The L2 norm of each main capsule output vector is divided by the reference value to obtain the confidence of the main capsule output vector.

[0055] Optionally, the adjusting the influence of the vector when being routed to a subsequent high-level capsule layer according to the confidence level, the high-level capsule layer generating a set of regulation parameters for a subsequent process, comprises:

[0056] for each primary capsule vector output by the primary capsule layer and its corresponding confidence level :

[0057] for each high-level capsule j, the primary capsule vector is converted into a prediction vector by a transformation matrix ;

[0058] the input of the high-level capsule j is initialized as a zero vector ;

[0059] in the routing iteration process, the dot product of the prediction vector of the current iteration round and the current state vector of the high-level capsule j is calculated, and is normalized among all high-level capsules by a softmax function to obtain a connection routing coupling coefficient between each primary capsule vector U_i and the high-level capsule j; the input of the high-level capsule j is updated according to , and the state vector of the high-level capsule j is updated by a squash function

[0060] after the iteration ends, the state vectors of all high-level capsules are spliced to obtain a feature vector, and the feature vector is input into a multi-layer perception MLP to obtain a set of regulation parameters.

[0061] Optionally, the optimizing the high-level capsule layer by using the multi-objective deviation vector to obtain an adjusted and optimized set of regulation parameters, comprises:

[0062] a loss function value is calculated based on the multi-objective deviation vector; the gradient of the loss function value with respect to the weight parameters of each layer of the MLP is calculated by back propagation, and the gradient of the loss function value with respect to each element in the transformation matrix connecting the primary capsule layer and the high-level capsule layer is calculated by back propagation;

[0063] the weight parameters of the MLP and the transformation matrix are updated according to the gradient; using the updated MLP and transformation matrix, a set of regulation parameters is regenerated.

[0064] The application adjusts the parameters of the membership function, and calibrates the activation contribution of the fuzzy rule according to the historical prediction accuracy and the actual process benefit correlation, thereby improving the adaptability of the fuzzy system to material diversity and process; the multi-modal time sequence main capsule layer is introduced to process heterogeneous sensor data and process coordination state, and the dynamic routing mechanism based on confidence evaluation is combined to transmit information to the high-level capsule layer, thereby enhancing the acquisition ability of inter-modal interaction; the high-level capsule layer can not only generate predictive control parameters for subsequent processes, but also correct its state representation and routing protocol by using the multi-objective deviation vector generated by subsequent actual quality detection, thereby realizing deep and accurate regulation of the printing and dyeing process and optimization of the model, thereby improving the quality stability of the textile printing and dyeing product, reducing resource consumption, and improving the intelligentization and self-adaptation level of the production process. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 is a flowchart of example one;

[0066] Figure 2 is a schematic diagram of the membership degree when the temperature is 40 degrees;

[0067] Figure 3 is a schematic diagram of the capsule network;

[0068] Figure 4 is a schematic diagram of the optimization loop based on the multi-objective deviation vector. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0070] In specific embodiments, the present application proposes an intelligent textile printing and dyeing process optimization method, as shown in Figure 1 , comprising:

[0071] S1, obtaining real-time and historical process parameters of the printing and dyeing process, determining the parameters of the membership function according to the characteristics of the current batch of materials, and adopting the membership function to fuzz the real-time process parameters to obtain fuzzy process characteristics;

[0072] Real-time data and past accumulated data on the printing and dyeing production line are collected. According to the specific properties of the batch of cloth currently being processed, a membership function is determined to describe the degree of fuzziness of the process parameters. For example, for the fuzzy concept of high temperature, if it is heat-resistant cotton, the center point of the membership function is set at 95°C, and the standard deviation is large; if it is silk, the center point may be at 80°C, and the standard deviation is small. The membership function converts the real-time accurate process parameter value into a set of characteristics describing its fuzzy state, for example, if the pH value is 6.8, it is converted into the degree of moderate acidity of the pH value, which is 0.7, and the degree of weak acidity is 0.3.

[0073] S2, according to the correlation between the historical prediction accuracy of the fuzzy rule and the process benefit, the activation contribution degree of each fuzzy rule is calculated, and the process coordination state is obtained by using the activation contribution degree of the fuzzy rule and the fuzzy process feature;

[0074] The high prediction accuracy of the rule in the past does not mean that the rule is good, further, the correlation between the prediction accuracy and the process benefits such as the good and bad of the final product quality and the high and low of the production efficiency is obtained. For example, a rule can accurately predict the temperature of a certain intermediate step, and historical data shows that such accurate prediction has brought better color fastness and lower energy consumption in most cases, so the activation contribution degree of this rule is relatively high. Then, combined with the activation contribution degree and the obtained real-time fuzzy process feature, the process coordination state is comprehensively judged by combining the cooperation of each link in the whole printing and dyeing process to obtain the process coordination state.

[0075] S3, input the sensor time series data and the process coordination state into the multi-modal time series main capsule layer to obtain a main capsule output vector, calculate the confidence of the vector, and adjust the influence of the vector when it is routed to the higher level capsule layer in the subsequent, and the higher level capsule layer generates a set of control parameters for the subsequent process;

[0076] The data from different sensors and the process coordination state are input into the multi-modal time series main capsule layer to obtain the mode of the change of these different types of data over time, and a set of main capsule output vectors are output. If a certain main capsule vector is generated based on sensor data with a lot of noise, its confidence is low. The confidence determines the influence of the vector when it is passed to the higher level capsule layer. The higher level capsule layer uses this information to generate a set of specific operation instructions for the subsequent printing and dyeing process, i.e. a set of control parameters.

[0077] S4, obtain the quality detection data of the subsequent process, obtain a multi-objective deviation vector according to the quality detection data and the control parameter set, and optimize the higher level capsule layer using the multi-objective deviation vector to obtain an adjusted and optimized control parameter set.

[0078] The actual product quality detection result is obtained, and by comparing the actual detection data and the expected quality based on the control parameters, a multi-target deviation vector containing multiple aspect differences is obtained, for example, the color difference deviates by ΔE=0.5, and the intensity is 2N lower than expected. The deviation vector is used to feedback adjust and the internal parameters of the high-level capsule layer, so that more accurate prediction and control can be made in the future, thereby outputting the further adjusted and optimized control parameter set.

[0079] In an optional embodiment, the parameters of the membership function determined according to the characteristics of the current batch of materials include:

[0080] The material characteristic vector of the current batch of materials is obtained, and the material characteristic vector is input into a radial basis function network to obtain the center point value and the standard deviation value of the Gaussian membership function of each preset fuzzy subset of the process parameter.

[0081] Specifically, assuming that a special blended fabric needs to be processed at present, the material characteristic vector of the blended fabric contains information such as cotton content 70%, polyester content 30%, yarn denier 150D, fabric gram weight 180g / m², etc. The characteristic vector is input into a pre-trained radial basis function (RBF) network. According to the input material characteristic combination, the optimal membership function parameters are output for each fuzzy subset of the subsequent process parameters. For example, for this batch of blended fabric, the RBF network may output the center point value of the Gaussian membership function of the “high temperature” fuzzy subset of the dyeing temperature as 92°C, and the standard deviation value as 3°C; and for the long-time subset of the dyeing time, the center point value of the Gaussian membership function is determined as 55 minutes, and the standard deviation is 5 minutes. Further, if a thick cotton fabric with a large gram weight is replaced, the material characteristic vector is input into the same RBF network, and the RBF network outputs the center point value of the high-temperature subset of the dyeing temperature as 98°C, and the standard deviation as 2°C; and outputs the center point value of the long-time subset of the dyeing time as 70 minutes, and the standard deviation as 4 minutes. Further, it can be identified that the thick cotton fabric needs a higher dyeing temperature and a longer dyeing time. The determination of the material characteristic parameters makes the definition of the membership function flexible, avoids the subjectivity of manual parameter setting, and also enables the entire optimization method to be applied to various types of textile printing and dyeing.

[0082] In an optional embodiment, the fuzzy process characteristic obtained by fuzzy processing of the real-time process parameter by using the membership function includes:

[0083] The Gaussian membership function of each fuzzy subset corresponding to the process parameter and the corresponding center point value and standard deviation value are obtained;

[0084] The numerical value of the process parameter is substituted into the Gaussian membership function of each fuzzy subset corresponding to the process parameter The membership values of each fuzzy subset are obtained by substituting the parameter value x into the membership function of the fuzzy subset, where x is the parameter value, c is the center point value, and σ is the standard deviation value.

[0085] All the membership values of each process parameter constitute the fuzzy process feature of the process parameter.

[0086] Each real-time process parameter that needs to be fuzzified is pre-defined with a set of fuzzy subsets describing different state intervals. For example, for dyeing temperature, three fuzzy subsets are defined: low temperature, medium temperature, and high temperature. Each of these fuzzy subsets is described by a Gaussian membership function, which is determined by two key parameters: the center point value c and the standard deviation value σ. The center point value represents the most typical value of the fuzzy subset, for example, the center point of medium temperature can be 70°C; the standard deviation determines the range within which the parameter value is still considered to belong to this subset around the center point.

[0087] When the specific real-time process parameter value is obtained, for example, the current dyeing temperature is 72°C, the value x = 72 is substituted into the Gaussian membership function formula of each fuzzy subset corresponding to the parameter dyeing temperature to calculate the membership values of the current temperature 72°C for the three fuzzy subsets of low temperature, medium temperature, and high temperature. For example, the calculated membership value for low temperature is 0.05, for medium temperature is 0.8, and for high temperature is 0.2. The combination of all the membership values of the fuzzy subsets corresponding to a process parameter constitutes the fuzzy process feature of the process parameter [0.05, 0.8, 0.2], Figure 2 The membership value at a temperature of 40 degrees is shown.

[0088] In an optional embodiment, the activation contribution degree of each fuzzy rule is calculated according to the correlation degree of the historical prediction accuracy and the process benefit of the fuzzy rule, comprising:

[0089] For each fuzzy rule in the fuzzy rule library, all historical process batches whose membership product of each fuzzy variable in the condition part of the fuzzy rule is greater than a preset value are filtered out from the database of historical process parameters;

[0090] For each historical process batch, the root mean square error RMSE between the intermediate process target prediction value defined by the result part of the fuzzy rule and the actual recorded intermediate process target value of the batch is calculated, and the historical prediction accuracy of the rule is ;

[0091] At least three final product quality indicators and at least two process benefit indicators corresponding to each historical process batch are obtained; the Pearson correlation coefficients between the prediction accuracy of the rule and each quality indicator and benefit indicator are calculated, and the weighted average value of each Pearson correlation coefficient is taken as the process benefit correlation degree indicator CB;

[0092] According to the formula The activation contribution degree of the fuzzy rule is calculated, wherein 、 is a preset coefficient, and .

[0093] For each batch of production in history, it is determined to what extent the process parameters at that time satisfy the condition part of the rule. If the product of these membership degrees exceeds a preset value, then this historical batch is related to the rule. For all the filtered historical batches, the intermediate process target value predicted by the rule is calculated according to the prediction effect of the conclusion part of the rule at that time, and the corresponding value actually recorded in history. The smaller the error, the more accurate the rule prediction, and the higher the historical prediction accuracy.

[0094] Further analysis of the actual correlation degree of accuracy and final product quality and production efficiency, that is, the process benefit correlation degree. Obtain the final product quality data and process benefit data corresponding to the historical batches with accurate rule prediction. The final product quality indicators include but are not limited to color difference, level of uniformity, color fastness, and the process benefit indicators include but are not limited to unit energy consumption and dye utilization rate.

[0095] The Pearson correlation coefficient between the historical prediction accuracy AP of the rule and the quality and benefit indicators is calculated, and a weighted average is performed to obtain a comprehensive CB value. If the AP value of a rule is high, the historical data shows that the product color fastness is generally high, and the unit energy consumption is low, then the CB value of this rule will be higher. Through the weighted formula The historical prediction accuracy AP and the process benefit correlation degree CB are combined to obtain the final activation contribution degree ACD of the fuzzy rule. Ensure that the rules that are predicted accurately and can bring better quality and higher production benefit are given greater weight and influence in the current process optimization decision.

[0096] In an optional embodiment, the process synergy state obtained by using the activation contribution degree of the fuzzy rule and the fuzzy process characteristics comprises:

[0097] For each fuzzy rule, the fuzzy process characteristics are substituted into the condition part of the rule, the total activation intensity of the rule condition part is calculated by using the algebraic product T-norm operator, and the weighted activation intensity of the rule is obtained by multiplying the total activation intensity of the rule by the activation contribution degree ACD of the rule.

[0098] The weighted activation intensity of each rule is multiplied by the characteristic value of the process synergy state defined by the result part of the rule, the product results of all rules are summed, and then divided by the sum of the weighted activation intensities of all rules to obtain the process synergy state at the current time.

[0099] For example, one rule is: if temperature is high and pH is low, then dyeing rate tends to be fast, and the activation contribution degree ACD of this rule is calculated as 0.9 according to historical performance. The fuzzy features of current process parameters are substituted into the condition part of the rule, and the original activation strength of this rule triggered by current condition is obtained using algebraic product operator, which is 0.56 with temperature high 0.8, pH low 0.7. The value obtained by multiplying the original activation strength by the ACD of the rule is 0.504, which is the weighted activation strength of this rule for the current overall state judgment. Further, not only whether the rule is triggered by the current situation is considered, but also the effectiveness and actual contribution to process benefit of the rule in history are combined. The result part of each rule corresponds to a preset characteristic value of process synergy state, for example, fast corresponds to synergy state value 8, medium corresponds to 5, and slow corresponds to 2. The weighted activation strength of each rule is multiplied by the characteristic value of the process synergy state pointed by it. For example, if the weighted activation strength of rule A is 0.504, and the result points to the characteristic value of synergy state 8, then the product is 4.032; the weighted activation strength of rule B is 0.3, and the result points to the characteristic value of synergy state 5, then the product is 1.5. The products of all rules are added up, and then divided by the sum of the weighted activation strengths of all rules, to obtain the process synergy state value at the current time.

[0100] In an optional embodiment, the calculating the confidence of the vector comprises:

[0101] In a current processing time window preset or dynamically adjusted according to process stage characteristics, all main capsule output vectors output by the multi-modal time sequence main capsule layer are obtained; the L2 norm of each main capsule output vector is calculated; the maximum value in the L2 norms of all main capsule output vectors in the time window is taken as a reference value;

[0102] The confidence of each main capsule output vector is obtained by dividing the L2 norm of the main capsule output vector by the reference value.

[0103] In particular, assume that in the fixation stage of textile dyeing, the current processing time window, say 60 seconds, is determined according to the current fixation rate feedback. Within these 60 seconds, the multi-modal temporal main capsule layer outputs multiple main capsule vectors, each of which represents a short-term dynamic feature extracted from the time-series data of different sensors. For example, one main capsule vector can characterize the smoothness of the temperature profile, and another can characterize the trend of rapid dye concentration drop. The L2 norm of each such vector is calculated, say, the L2 norm of the temperature smoothness vector is 0.8, and the L2 norm of the dye concentration drop trend vector is 1.5, and the L2 norm of the other vectors can be between the two. Find the maximum value among the L2 norms of all main capsule output vectors within the 60 seconds, say 1.5, and take it as the reference value. Divide the L2 norm of each main capsule output vector by the reference value 1.5 to obtain the respective confidence. The confidence of the dye concentration drop trend vector is 1.5 / 1.5 = 1.0, while the confidence of the temperature smoothness vector is 0.53, Figure 3 The relationship between the main capsule layer and the high-level capsule is shown.

[0104] In an optional embodiment, the adjusting the influence of the vector according to the confidence when it is subsequently routed to the high-level capsule layer, and the high-level capsule layer generates a set of control parameters for the subsequent process, comprises:

[0105] For each main capsule vector output by the main capsule layer and its corresponding confidence :

[0106] For each high-level capsule j, the main capsule vector is converted into a prediction vector by a transformation matrix ;

[0107] The input of the high-level capsule j is initialized as a zero vector;

[0108] In the routing iteration process, the dot product of the prediction vector of the current iteration round and the current state vector of the high-level capsule j is calculated, and the connection routing coupling coefficient between each main capsule vector and the high-level capsule j is obtained by normalizing all high-level capsules by the softmax function; the input of the high-level capsule j is updated according to , and the state vector of the high-level capsule j is updated by the squash function;

[0109] After the iteration is completed, the state vectors of all high-level capsules are spliced to obtain a feature vector, and the feature vector is input into a multi-layer perception MLP to obtain a set of control parameters. ​

[0110] Specifically, the master capsule vectors representing low-level features and their respective confidences are obtained. Each master capsule vector produces a prediction for each high-level capsule through the corresponding transformation matrix. Then, through an iterative routing mechanism, the contribution weight of each master capsule to the high-level capsules is assigned according to the degree of match between the master capsule’s prediction and the current state of the high-level capsules and the master capsule’s own confidence. The weighted and iterative process enables the high-level capsules to effectively integrate information from multiple master capsules, with master capsules of high confidence having a greater impact. After the iteration is completed, the state vectors of all high-level capsules are concatenated into a feature vector. The feature vector is input into a multi-layer perceptron network, which maps the integrated features to a specific set of control parameters for the subsequent process.

[0111] In an optional embodiment, the optimization of the high-level capsule layer using the multi-objective deviation vector to obtain an adjusted and optimized set of control parameters comprises:

[0112] The loss function value is calculated based on the multi-objective deviation vector. The gradient of the loss function value with respect to the weight parameters of each layer of the MLP is calculated through backpropagation, and the gradient of the loss function value with respect to each element in the transformation matrix connecting the master capsule layer and the high-level capsule layer is calculated through backpropagation.

[0113] The weight parameters of the MLP and the transformation matrix are updated according to the gradient. The updated MLP and transformation matrix are used to regenerate the set of control parameters.

[0114] Specifically, the multi-objective deviation vector between the effect of the previously generated set of control parameters after actual application, i.e., the quality detection data, and the expected target is calculated. Based on the multi-objective deviation vector, the loss function value is calculated, and the internal weights of the multi-layer perceptron (MLP) and the transformation matrix connecting the master capsule layer and the high-level capsule layer are updated using the backpropagation algorithm. According to the gradient information, the weights of the MLP and the transformation matrix are fine-tuned to reduce the loss function. After the parameter update is completed, the optimized MLP and transformation matrix are used to regenerate a more optimal set of control parameters, as shown in Figure 4 .

[0115] The above examples are only used to illustrate the technical solutions of the present application, but not to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application. In addition, the various embodiments of the embodiments of the present application can be combined in any manner, as long as it does not deviate from the idea of the embodiments of the present application, and it should also be considered as the disclosed content of the embodiments of the present application.

Claims

1. An intelligent textile printing and dyeing process optimization method, characterized in that, The method comprises the following steps: obtain real-time and historical process parameters of the printing and dyeing process, determine the parameters of the membership function according to the characteristics of the current batch of materials, and obtain fuzzy process characteristics by fuzzy processing of real-time process parameters using the membership function; calculate the activation contribution degree of each fuzzy rule according to the historical prediction accuracy of the fuzzy rule and the correlation degree of process benefit, and obtain the process synergy state by using the activation contribution degree of the fuzzy rule and the fuzzy process characteristics; input the sensor time series data and the process synergy state into the multi-modal time series master capsule layer to obtain a master capsule output vector, calculate the confidence of the vector, and adjust the influence of the vector when it is routed to a higher layer capsule layer according to the confidence, and the higher layer capsule layer generates a set of control parameters for the subsequent process; obtain quality detection data of the subsequent process, and obtain a multi-target deviation vector according to the quality detection data and the control parameter set, and optimize the higher layer capsule layer using the multi-target deviation vector to obtain an adjusted and optimized control parameter set; the calculation of the activation contribution degree of each fuzzy rule according to the historical prediction accuracy of the fuzzy rule and the correlation degree of process benefit comprises: for each fuzzy rule in the fuzzy rule library, filter all historical process batches whose membership degree product of each fuzzy variable in the condition part is greater than a preset value from the database of historical process parameters; For each historical process batch, the root mean square error RMSE between the intermediate process target predicted value defined by the result part of the fuzzy rule and the actually recorded intermediate process target value of the batch is calculated, and the historical prediction accuracy of the rule is ; obtain at least three final product quality indicators and at least two process benefit indicators corresponding to each historical process batch; calculate the Pearson correlation coefficient between the prediction accuracy of the rule and each quality indicator and benefit indicator, and take the weighted average value of each Pearson correlation coefficient as the process benefit correlation degree index CB; According to the formula The activation contribution degree of the fuzzy rule is calculated, wherein , is a preset coefficient, and .

2. The method of claim 1, wherein, the determination of the parameters of the membership function according to the characteristics of the current batch of materials comprises: obtain the material characteristic vector of the current processing batch of materials, input the material characteristic vector into the radial basis function network, and obtain the center point value and standard deviation value of the Gaussian membership function of the preset fuzzy subset for each process parameter.

3. The method of claim 1, wherein, the fuzzy processing of real-time process parameters using the membership function to obtain fuzzy process characteristics comprises: obtain the Gaussian membership function of the fuzzy subset corresponding to each process parameter and the corresponding center point value and standard deviation value; substituting the value of the process parameter into the Gaussian membership function of the corresponding each fuzzy subset obtaining the membership value of the process parameter for each fuzzy subset, wherein x is the parameter value, c is the center point value, and σ is the standard deviation value all membership degree values of each process parameter constitute the fuzzy process characteristics of the process parameter.

4. The method of claim 1, wherein, the calculation of the activation contribution degree of each fuzzy rule according to the historical prediction accuracy of the fuzzy rule and the correlation degree of process benefit comprises: for each fuzzy rule, substitute the fuzzy process characteristics into the condition part of the rule, calculate the total activation intensity of the rule condition part using the algebraic product T-norm operator, multiply the total activation intensity of the rule by the activation contribution degree ACD of the rule to obtain the weighted activation intensity of the rule; multiply the weighted activation intensity of each rule by the characteristic value of the process synergy state defined by the result part of the rule, sum the product results of all rules, and then divide by the sum of the weighted activation intensities of all rules to obtain the process synergy state at the current time.

5. The method of claim 1, wherein, the calculation of the confidence of the vector comprises: Within a current processing time window preset or dynamically adjusted according to process stage characteristics, all main capsule output vectors output by the multi-modal time-series main capsule layer are obtained; L2 norms of each main capsule output vector are calculated; a maximum value in the L2 norms of all main capsule output vectors within the time window is taken as a reference value; A confidence degree of each main capsule output vector is obtained by dividing the L2 norm of the main capsule output vector by the reference value.

6. The method of claim 1, wherein, The influence of the vector on subsequent routing to a high-level capsule layer is adjusted according to the confidence degree, and the high-level capsule layer generates a set of regulation parameters for a subsequent process, which comprises: for each primary capsule layer output primary capsule vector and its corresponding confidence : For each high layer capsule j, the transform matrix The primary capsule vector is converted to a prediction vector ; Initialize the input to high-level capsule j is a zero vector; In the routing iteration process, the prediction vector of the current iteration round is calculated The dot product of the current state vector of the high-level capsule j , and the normalization of all high-level capsules by the softmax function, to obtain each main capsule vector The connection routing coupling coefficient between the main capsule vector and the high-level capsule j ; according to Update the input of the high-level capsule j, and update the state vector of the high-level capsule j through the squeeze function; After iteration, state vectors of all high-level capsules are spliced to obtain a feature vector, and the feature vector is input into a multi-layer perception (MLP) to obtain a set of regulation parameters.

7. The method of claim 1, wherein, The high-level capsule layer is optimized by using the multi-objective deviation vector to obtain an adjusted and optimized set of regulation parameters, which comprises: A loss function value is calculated based on the multi-objective deviation vector; gradients of the loss function value with respect to weight parameters of each layer of the MLP and with respect to each element in a transformation matrix connecting the main capsule layer and the high-level capsule layer are calculated through back propagation; and the weight parameters of the MLP and the transformation matrix are updated according to the gradients; the set of regulation parameters is regenerated using the updated MLP and transformation matrix. Comprise:

8. An intelligent textile printing and dyeing process optimization system, characterized in that, A feature extraction unit is configured to obtain real-time and historical process parameters of a printing and dyeing process, determine parameters of a membership function according to characteristics of a current batch of materials, and perform fuzzy processing on the real-time process parameters by using the membership function to obtain fuzzy process characteristics; A state acquisition unit is configured to calculate an activation contribution degree of each fuzzy rule according to a correlation degree between a historical prediction accuracy of the fuzzy rule and a process benefit, and obtain a process synergy state by using the activation contribution degree of the fuzzy rule and the fuzzy process characteristics; A parameter adjustment unit is configured to input sensor time-series data and the process synergy state into a multi-modal time-series main capsule layer to obtain a main capsule output vector, calculate a confidence degree of the vector, and adjust an influence of the vector on subsequent routing to a high-level capsule layer according to the confidence degree, wherein the high-level capsule layer generates a set of regulation parameters for a subsequent process; An optimization unit is configured to obtain quality detection data of the subsequent process, obtain a multi-objective deviation vector according to the quality detection data and the set of regulation parameters, and optimize the high-level capsule layer by using the multi-objective deviation vector to obtain an adjusted and optimized set of regulation parameters. The activation contribution degree of each fuzzy rule is calculated according to a correlation degree between a historical prediction accuracy of the fuzzy rule and a process benefit, which comprises: For each fuzzy rule in a fuzzy rule library, all historical process batches in which a membership degree product of each fuzzy variable in a condition part of the rule is greater than a preset value are selected from a database of historical process parameters; At least three final product quality indicators and at least two process benefit indicators corresponding to each historical process batch are obtained; a prediction accuracy of the rule and Pearson correlation coefficients between each quality indicator and each benefit indicator are calculated; and a weighted average value of the Pearson correlation coefficients is taken as a process benefit correlation degree indicator CB. For each historical process batch, the root mean square error RMSE between the intermediate process target predicted value defined by the result part of the fuzzy rule and the actually recorded intermediate process target value of the batch is calculated, and the historical prediction accuracy of the rule is ; ​ According to the formula The activation contribution degree of the fuzzy rule is calculated, wherein , is a preset coefficient, and .

9. The system of claim 8, wherein, The parameter of the membership function determined according to the material characteristics of the current batch includes: The material characteristic vector of the current processing batch material is acquired, the material characteristic vector is input into the radial basis function network, and the center point value and the standard deviation value of the Gaussian membership function of the preset fuzzy subset for each process parameter are obtained.

Images