An end-to-end yarn quality prediction model and method

Abstract

The present invention belongs to the technical field of yarn quality prediction, and provides an end-to-end yarn quality prediction model and method. The model includes the following modules: Module 1 is a data preprocessing module, which vectorizes the spinning technical parameters input by the model. Module 2 is a cotton blending information feature extraction module, which builds a neural network framework based on the self-attention mechanism to fully learn the variable-length cotton blending information sequence data to achieve comprehensive feature expression of mixed raw materials. Module 3 is a feature fusion and yarn quality prediction module, which realizes the prediction of yarn quality. Compared with traditional yarn quality prediction models, the model provided by the present invention is a learning-based, universal end-to-end prediction model that can flexibly adapt to different types of information involved in the spinning production process, effectively improves the prediction accuracy and prediction efficiency, and is expected to be expanded to other intelligent manufacturing fields.

Description

Technical Field

[0001] The present invention relates to the technical field of spinning quality prediction, and in particular to an end-to-end yarn quality prediction model and method. Background Art

[0002] From the perspective of a spinning production system, manufacturing process parameters and product specifications interact with different fiber ratios to ultimately determine product quality. Accurately predicting yarn quality plays a crucial role in controlling yarn quality, reducing spinning costs, improving labor efficiency, and optimizing supply chain management. Traditionally, yarn quality prediction relies on the experience of process technicians. Establishing a mapping model from spinning data to yarn quality and mining the inherent production knowledge within historical production data can effectively improve production management.

[0003] The modeling approach used in yarn quality prediction models is primarily supervised machine learning, which can be categorized into traditional machine learning algorithms and deep learning algorithms. Traditional learning algorithms primarily include multivariate linear regression, support vector machines, artificial neural networks, and their various improved models, such as support vector machines with modified kernel functions, random vector function link neural networks, and radial basis function neural networks. These algorithms typically require manually designed feature extraction and data preprocessing steps. Deep learning algorithms, such as convolutional neural networks, are a class of machine learning algorithms based on multi-layer neural network structures. They can automatically learn hierarchical feature representations from raw data and construct complex nonlinear models. In terms of model training methods, in addition to typical analytical solutions and gradient descent, heuristic algorithms such as particle swarm optimization, genetic algorithms, butterfly optimization algorithms, and gray wolf algorithms have also received widespread attention. Due to the limitations of their model structure, these methods can only accept fixed-length features as input and cannot adapt to varying lengths of cotton blending information. They also lack the flexibility to adapt the model structure to specific problems. Consequently, they rely heavily on manually extracted features, limiting the model's versatility. Summary of the Invention

[0004] The present invention provides an end-to-end yarn quality prediction model. Unlike the previous manual extraction of cotton blending information features, the model framework proposed in the present invention can adaptively extract the features of variable-length cotton blending information, and establish the interaction between control parameters and raw material characteristics through a learnable gating structure, thereby realizing the modeling of the yarn spinning process. Compared with existing yarn quality prediction models, the end-to-end yarn quality prediction model proposed in this application can self-learn the representation and fusion methods of various features, has reasonable industrial manufacturing structured priors, has superior global information perception capabilities, avoids the limitations and instability of manual intervention, makes prediction tasks easier to implement, and has better flexibility.

[0005] In order to achieve the above object, the technical solution of the present invention is:

[0006] An end-to-end yarn quality prediction model includes the following modules:

[0007] Module 1: Data preprocessing module, which vectorizes the spinning technical parameters input into the model.

[0008] Module 2, cotton blending information feature extraction module, is to build a neural network framework based on the self-attention mechanism to fully learn the variable-length cotton blending information sequence data to achieve the comprehensive feature expression of the mixed raw materials.

[0009] Module 3, feature fusion and yarn quality prediction module, first uses a gate-controlled regression prediction module to fuse the control features with the mixed cotton attribute features, guides the model to focus on relevant information in the comprehensive features of the mixed cotton fibers, and then maps the fused features to the yarn quality space, thereby modeling the information transfer process in spinning manufacturing and realizing the prediction of yarn quality.

[0010] Preferably, in the module 1, the spinning technical parameters input by the model include cotton blending information, process parameters, and one or more parameters of yarn specifications. The cotton blending information includes cotton fiber quality data and the cotton ratio used in the blending process. The quality attributes of cotton fibers include length index, strength and elongation index, micronaire index, color index, impurity index, and other indicators. The number of cotton fiber types and their blending ratios in the cotton blending scheme are not fixed. The process parameters include the number of steel wire rings, drafting multiple, spindle speed, spun yarn twist, carding method, spinning method, etc. Yarn specifications include yarn count, etc.

[0011] Preferably, in the module 1, data preprocessing is to vectorize the spinning technical parameters input into the model, which is conducive to the effective learning of the model. The specific implementation method includes the following:

[0012] 1.1 For each piece of cotton fiber attribute data First, a neural network layer E is used to map the cotton blending attributes to a new feature space to form the cotton fiber feature x c On the one hand, it ensures the ability of features to express complex cotton blending information, and on the other hand, it facilitates the use of cotton blending information feature extraction module. The specific calculation method of the neural network layer E is:

[0013] x c =E(x)=σ(wx+b) (1)

[0014] In formula (1), σ(·) is the activation function, and are the weights and biases of the fully connected layer respectively. After being processed by the neural network layer E, the original d-dimensional cotton fiber properties are mapped to d k Wei, recorded as

[0015] 1.2 For each process parameter and yarn specification data, the p process parameters are arranged into a numerical vector using vector coding, which is recorded as Arrange the yarn specification parameters of item c into the form of a numerical vector, denoted as Then concatenate the two into a new p+c dimensional feature vector, recorded as the control feature

[0016] Preferably, in Module 2, the model proposed in the present invention designs a learnable self-attention module to extract cotton blending information. Specifically, the physical characteristics of the raw materials are mapped to a weighted feature space, and then weighted fusion is performed based on the raw material ratio to achieve a comprehensive feature expression of the mixed raw materials. The specific calculation steps include the following:

[0017] 2.1 The cotton fiber features x obtained by processing the neural network layer E c Convert to get d m Dimension keys (K), values ​​(V), and queries (Q):

[0018] K j =x c W j k (2)

[0019] V j =x c W j v (3)

[0020] Q j =x c W j q (4)

[0021] In formulas (2), (3) and (4), W j k , W j v , is the trainable weight matrix.

[0022] 2.2 Q j , K j and V j Perform dot product-based attention calculation:

[0023]

[0024] 2.3 The extracted attention information is mapped into a new, weighted cotton fiber attribute feature vector through the neural network module and recorded as x a ,in Then use the cotton ratio as the weight to aggregate x a , to realize the feature modeling of cotton blending information. The cotton blending ratio is represented by the vector x p =(p1, p2, p3, ..., p l ) indicates that by x p with x a By weighted summation, we get the characteristic vector x that can characterize the quality of mixed cotton m .

[0025] Preferably, in the module 3, a gate-controlled regression prediction module includes the following specific implementation method:

[0026] 3.1 Control feature x o Mapping to d m The gating unit of dimension acts on the feature vector x of the mixed cotton quality in the form of element multiplication m , thereby modulating the cotton allocation information. The specific calculation method is as follows:

[0027] G n =sigmoid(w i ×x o )×x m (6)

[0028] 3.2 The modulated cotton blending information feature G n , is mapped to the yarn quality attribute space to be predicted through a fully connected layer to complete the prediction of yarn quality. The specific calculation method is as follows:

[0029] y=σ(w o g i +b o ) (7)

[0030] In formula (7), y is the predicted yarn quality, σ(·) is the activation function, and w o and b o are the weight and bias of the fully connected layer respectively.

[0031] Preferably, the end-to-end yarn quality prediction method is based on the above-mentioned end-to-end yarn quality prediction model and includes three steps: data preparation, model training, and model use:

[0032] (1) Data preparation: Collect the spinning technical parameters x from the historical production records of the spinning workshop, including cotton blending information, process parameters, and yarn specification parameters, and their corresponding yarn quality indicators y;

[0033] (2) Model training: Substitute x into the above end-to-end yarn quality prediction model to obtain the prediction index , using optimization technology to is the objective function, and the trainable weights and biases in modules 1, 2, and 3 in the model are optimized alternately and iteratively until the optimization converges or the objective function value is less than the set value e; where ||·|| is the vector distance function;

[0034] (3) Model usage: The spinning technical parameters x of the yarn quality to be predicted during use m As input, substitute it into the trained model and output the yarn quality prediction result y m .

[0035] Beneficial effects of the present invention:

[0036] (1) The present invention constructs a neural network module based on a self-attention mechanism to extract cotton blending information. This method has the ability to automatically learn the characteristics of mixed raw materials from training samples, avoiding information loss caused by human prior intervention, thereby improving the versatility of the model.

[0037] (2) The present invention proposes to use a learnable gating structure to transmit production process and product specification characteristics, thereby guiding the model to focus on relevant information in the comprehensive characteristics of the mixed raw materials, modeling the information transmission process in spinning manufacturing, and realizing reliable prediction of yarn quality.

[0038] (3) The yarn quality prediction framework constructed by the present invention can self-learn the representation and fusion methods of various features, possesses reasonable industrial manufacturing structured priors, and has excellent global information perception capabilities, avoiding the limitations and instability of manual intervention. This method can not only be used for yarn quality prediction in the textile manufacturing field, but can also be extended to any manufacturing quality prediction field involving processing technology, product specifications, and raw material ratio information. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a framework diagram of an end-to-end yarn quality prediction model provided by the example of this application;

[0040] Figure 2 Schematic diagram of the dot product-based self-attention mechanism provided in an embodiment of the present application;

[0041] Figure 3 Schematic diagram of the gate control structure provided in the embodiments of the present application. DETAILED DESCRIPTION

[0042] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0043] In the following, in conjunction with the accompanying drawings, an end-to-end yarn quality prediction model framework provided by the embodiment of the present application is described in detail through specific embodiments and application scenarios.

[0044] Reference Figure 1 , showing an end-to-end yarn quality prediction model framework diagram provided in an embodiment of the present application.

[0045] This application provides an end-to-end yarn quality prediction model, which includes the following modules:

[0046] Module 1: Data preprocessing module, which vectorizes the spinning technical parameters input into the model.

[0047] Module 2, cotton blending information feature extraction module, is to build a neural network framework based on the self-attention mechanism to fully learn the variable-length cotton blending information sequence data to achieve the comprehensive feature expression of the mixed raw materials.

[0048] Module 3, feature fusion and yarn quality prediction module, first uses a gate-controlled regression prediction module to fuse the control features with the mixed cotton attribute features, guides the model to focus on relevant information in the comprehensive features of the mixed cotton fibers, and then maps the fused features to the yarn quality space, thereby modeling the information transfer process in spinning manufacturing and realizing the prediction of yarn quality.

[0049] In this embodiment, unlike the previous manual extraction of cotton blending information features, the present invention designs a learnable attention module to automatically extract variable-length cotton blending information. In the attribute dimension and element dimension, the cotton blending information exhibits the following characteristics: (1) Different attributes are distributed in different feature spaces, and the relationship between attributes is unclear, which may bring difficulties to the model to learn the law of their influence on yarn quality; (2) The mutual influence law of the attributes of various types of fibers is unclear, resulting in possible information loss when using the proportion weight aggregation feature in the traditional method. Therefore, the present invention maps the physical characteristics of the raw materials to a weighted feature space, and then performs weighted fusion according to the raw material ratio to achieve a comprehensive feature expression of the mixed raw materials. This method has the ability to automatically learn the characteristics of mixed raw materials from training samples, avoiding information loss caused by human prior intervention, thereby improving the versatility of the model.

[0050] Furthermore, since there are great differences in the data distribution between the values ​​of the control features and the properties of cotton fibers, simply splicing them together and inputting them into the regression module will inevitably lead to an imbalance in the weights of certain features, making the training of the model more difficult. Considering the effect of control features on mixed raw materials in actual production, they have a certain selective effect on the production process, similar to the effect of certain equipment switches. Therefore, the present invention proposes to use a learnable gating structure to transmit production process and product specification features, and by simulating the information transmission in the spinning manufacturing process, guide the model to focus on relevant information in the comprehensive characteristics of the mixed raw materials, so as to achieve reliable prediction of yarn quality.

[0051] Furthermore, the embodiments of the present application were experimentally verified on actual spinning production data, proving that the end-to-end yarn quality prediction model proposed in the present application has a significant improvement in prediction accuracy.

[0052] In a further embodiment, in the module 1, the spinning technical parameters input into the model include cotton blending information, process parameters, and one or more parameters of yarn specifications. The cotton blending information includes cotton fiber quality data and the cotton ratio used in the blending process. The quality attributes of cotton fibers include length index, strength and elongation index, micronaire index, color index, impurity index, and other indicators. The number of cotton fiber types and their blending ratios in the cotton blending plan are not fixed. The process parameters include the number of steel wire rings, draft ratio, spindle speed, spun yarn twist, carding method, spinning method, etc. Yarn specifications include yarn count, etc.

[0053] In a further embodiment, in the module 1, data preprocessing is to vectorize the spinning technical parameters input into the model, which is conducive to the effective learning of the model. The specific implementation method includes the following:

[0054] (1) For each piece of cotton fiber attribute data First, a neural network layer E is used to map the cotton blending attributes to a new feature space to form the cotton fiber feature x c On the one hand, it ensures the ability of features to express complex cotton blending information, and on the other hand, it facilitates the use of cotton blending information feature extraction module. The specific calculation method of the neural network layer E is:

[0055] x c =E(x)=σ(wx+b) (1)

[0056] In formula (1), σ(·) is the activation function, and are the weights and biases of the fully connected layer respectively. After being processed by the neural network layer E, the original d-dimensional cotton fiber properties are mapped to d k Wei, recorded as

[0057] (2) For each process parameter and yarn specification data, the p process parameters are arranged into a numerical vector using vector coding, which is denoted as Arrange the yarn specification parameters of item c into the form of a numerical vector, denoted as Then concatenate the two into a new p+c dimensional feature vector, recorded as the control feature

[0058] In a further embodiment, in Module 2, the model proposed by the present invention incorporates a learnable self-attention module to extract cotton blending information. Specifically, the physical characteristics of the raw materials are mapped to a weighted feature space, which is then weighted and fused based on the raw material ratios to achieve a comprehensive feature representation of the blended raw materials. The specific calculation steps are as follows:

[0059] (1) The cotton fiber feature x obtained by processing the neural network layer E c Convert to get d m Dimension keys (K), values ​​(V), and queries (Q):

[0060] K j =x c W j K (2)

[0061] V j =x c W j v (3)

[0062] Q j =x c W j q (4)

[0063] In formulas (2), (3) and (4), W j k , W j v , is the trainable weight matrix.

[0064] (2)Reference Figure 2 , shows a schematic diagram of the self-attention mechanism based on dot product provided by the embodiment of the present application. j , K j and V j The method for attention calculation is as follows:

[0065]

[0066] (3) The extracted attention information is passed through the feedforward neural network module to obtain a new, weighted cotton fiber attribute feature vector, denoted as x c ,in The feedforward neural network module can be expressed as:

[0067] FFN(x)=σ(W1x c +b1) (6)

[0068] In formula (6), σ(·) is the activation function, W1 and b1 are the weight and bias of the fully connected layer, respectively.

[0069] (4) Aggregate x with cotton proportion as weight a , to realize the feature modeling of cotton blending information. The cotton blending ratio is represented by the vector x p =(p1, p2, p3, ..., p l ) indicates that by x p with x a By weighted summation, we get the characteristic vector x that can characterize the quality of mixed cotton m .

[0070] In a further embodiment, in the module 3, a gate-controlled regression prediction module includes the following specific implementation steps:

[0071] (1)Reference Figure 3 , shows a schematic diagram of the gate control structure provided by the embodiment of the present application. Specifically, the control feature x o Mapping to d m The gating unit of dimension acts on the feature vector x of the mixed cotton quality in the form of element multiplication m , thereby modulating the cotton allocation information. The calculation method is as follows:

[0072] G n =sigmoid(w i ×x o )×x m (7)

[0073] (2) The modulated cotton information feature G n , is mapped to the yarn quality attribute space to be predicted through a fully connected layer to complete the prediction of yarn quality. The specific calculation method is as follows:

[0074] y=σ(w o g i +b o ) (8)

[0075] In formula (8), y is the predicted yarn quality, σ(·) is the activation function, and w o and b o are the weight and bias of the fully connected layer respectively.

[0076] Furthermore, after establishing the end-to-end yarn quality prediction model, the spinning technical parameters input into the model and the corresponding yarn quality data are normalized to the maximum and minimum values:

[0077]

[0078] In formula (9), x ij represents the i-th data of the j-th feature, and represent the maximum and minimum values ​​of the j-th feature respectively. Refers to the normalized data x ij .

[0079] Furthermore, the loss function is used to evaluate the output of the model and the true value of yarn quality y i Perform loss calculation and select Adam optimizer to optimize model parameters. The loss function of the model uses the mean absolute error algorithm (MAE), and the specific calculation formula is as follows:

[0080]

[0081] In formula (10), is the training set, |·| is the number of elements in the set, and MAE is the degree of loss.

[0082] Furthermore, the spinning technical parameter x to be used to predict the yarn quality is m As input, substitute it into the trained model and output the yarn quality prediction result y m .

[0083] Furthermore, to verify the effectiveness of the prediction model, this application will select production data from a cotton textile enterprise for five consecutive years in 2018, 2019, 2020, 2021, and 2022 for experimentation. Cotton fiber properties are derived from HVI inspection data, mainly including length index, strength and elongation index, micronaire index, color index, impurity index, and other indicators, as shown in Table 1. In addition, detailed information on the other two sets of characteristic yarn process parameters and yarn structure parameters selected by the model is also shown in Table 1. The labels of this dataset are eight yarn quality indicators, as shown in Table 2, covering all important indicators for cotton yarn quality evaluation, such as yarn evenness, strength, thin places, thick places, neps, and hairiness.

[0084] Table 1 Spinning data input to the model

[0085]

[0086] Table 2 Yarn quality data output by the model

[0087]

[0088] The yarn quality prediction model was built using PyTorch 1.11.0 in Python. The neural network E layer was used to map the cotton fiber properties to the dimension d through feature transformation. k =128, and the zero-padding method is used to ensure that the length of the cotton fiber sequence data in each training batch is constant. Using N = 2 self-attention module layers, d m = 64 hidden units. Other model hyperparameters are: 3000 training rounds, 0.001 weight decay coefficient, and 32 batch size.

[0089] Two loss functions representing regression errors are selected for quantitative evaluation of the model: mean square error (MSE) and coefficient of determination (R 2 ), the formula is as follows:

[0090]

[0091]

[0092] In formula (12), and y i are the predicted value of yarn quality and the true value of yarn quality, respectively. is the average value of the true value of the yarn quality in the test set, |·| is the number of elements in the set, For the test set.

[0093] It has been verified that the end-to-end yarn quality prediction model provided by this application can achieve MSE=0.0021, R 2 =0.93, which can meet the yarn production needs of cotton spinning enterprises.

Claims

1. An end-to-end yarn quality prediction model, characterized by: Includes the following modules: Module 1, data preprocessing module, which is to vectorize the spinning technical parameters input by the model; The data preprocessing module is used to vectorize the spinning technical parameters input by the model, including the following specific implementation methods: 1.1 For each cotton fiber attribute data First, a neural network layer E is used to map the cotton blending attributes to a new feature space to form the cotton fiber feature x c On the one hand, it ensures the ability of features to express complex cotton blending information, and on the other hand, it facilitates the use of cotton blending information feature extraction module; 1.2 For each process parameter and yarn specification data, the p process parameters are arranged into a numerical vector using vector coding, which is recorded as Arrange the yarn specification parameters of item c into the form of a numerical vector, denoted as Then concatenate the two into a new p+c dimensional feature vector, recorded as the control feature Module 2, cotton blending information feature extraction module, which builds a neural network framework based on the self-attention mechanism to fully learn the variable-length cotton blending information sequence data to achieve the comprehensive feature expression of the mixed raw materials; In Module 2, the proposed model designs a learnable self-attention module to extract cotton blending information. Specifically, the physical characteristics of the raw materials are mapped to a weighted feature space, and then weighted fusion is performed according to the raw material ratio to achieve a comprehensive feature expression of the mixed raw materials. The specific calculation steps include the following: 2.1 The cotton fiber features x obtained by processing the neural network layer E c Convert to get d m Dimension keys (K), values ​​(V), and queries (Q): Q j =x c W j q (4) In formulas (2), (3) and (4), is a trainable weight matrix; 2.2 Q j , K j and V j Perform dot product-based attention calculation: 2.3 The extracted attention information is mapped into a new, weighted cotton fiber attribute feature vector through the neural network module and recorded as x a ,in Then use the cotton ratio as the weight to aggregate x a , to realize the feature modeling of cotton blending information; the cotton blending ratio is represented by the vector x p =(p1,p2,p3,…,p l ) indicates that by x p with x a By weighted summation, we get the characteristic vector x that can characterize the quality of mixed cotton m ; Module 3, feature fusion and yarn quality prediction module, first uses a gate-controlled regression prediction module to fuse the control features with the mixed cotton attribute features, guides the model to focus on relevant information in the comprehensive features of the mixed cotton fibers, and then maps the fused features to the yarn quality space, thereby modeling the information transfer process in spinning manufacturing and realizing the prediction of yarn quality.

2. The end-to-end yarn quality prediction model according to claim 1, characterized in that: In the module 3, a gate-controlled regression prediction module includes the following specific methods: (1) Control feature x o Mapping to d m The gating unit of dimension acts on the feature vector x of the mixed cotton quality in the form of element multiplication m , thereby modulating the cotton allocation information; the specific calculation method is as follows: G n =sigmoid(w i x o )×x m (6) (2) The modulated cotton information feature G n , mapped to the yarn quality attribute space to be predicted through a fully connected layer, completing the prediction of yarn quality; the specific calculation method is as follows: y=σ(w o G n +b o ) (7) In formula (7), y is the predicted yarn quality, σ(·) is the activation function, and w o and b o are the weight and bias of the fully connected layer respectively.

3. An end-to-end yarn quality prediction model according to claim 1 or 2, characterized in that: In step 1.1, the specific calculation method of the neural network layer E is: x c =E(x)=σ(wx+b) (1) In formula (1), σ(·) is the activation function, and are the weights and biases of the fully connected layer respectively; after being processed by the neural network layer E, the original d-dimensional cotton fiber properties are mapped to d k Wei, recorded as 4. The prediction method of any one of the end-to-end yarn quality prediction models according to claims 1 to 3, characterized in that: It includes three steps: data preparation, model training, and model use: Step 1: Data preparation: Collect spinning technical parameters x from the historical production records of the spinning workshop, including cotton blending information, process parameters, and yarn specification parameters; and their corresponding yarn quality indicators y; Step 2: Model training: Substitute x into the yarn quality prediction model to obtain the prediction index Using optimization technology to is the objective function, and the trainable weights and biases in modules 1, 2, and 3 in the model are optimized alternately and iteratively until the optimization converges or the objective function value is less than the set value e; where ‖·‖ is the vector distance function; Step 3: Model usage: Use the spinning technical parameters x of the yarn quality to be predicted during use m As input, substitute it into the trained model and output the yarn quality prediction result y m .

5. The prediction method of the end-to-end yarn quality prediction model according to claim 4, characterized in that: In the second step, the output of the model is analyzed using the loss function and the true value of yarn quality y i Perform loss calculation and select Adam optimizer to optimize model parameters.

6. The prediction method of the end-to-end yarn quality prediction model according to claim 5, characterized in that: In the second step, the loss function of the model adopts the mean absolute error algorithm (MAE), and the specific calculation formula is as follows:

Images