A yarn quality prediction method based on a meta-learning method

Abstract

A yarn quality prediction method based on a meta-learning method relates to the technical field of yarn quality prediction and comprises the following steps: step 1, the weights of input parameters affecting yarn quality are analyzed by using a grey correlation degree method, and a mutual information algorithm is used to remove redundant parameters; step 2, a BiLSTM model with confidence is introduced to mine the correlation between input parameters and yarn quality; step 3, a meta-machine learning model is used to finally predict the yarn quality by using a small amount of yarn quality prediction data. The present application realizes feature selection by fusing the grey management method and the mutual information algorithm, ensures that the input data is closely related to the yarn quality, proposes the BiLSTM model with confidence, mines the relationship between the features and the yarn quality, solves the sample scarcity problem of spinning quality prediction by using the meta-learning method, and improves the prediction accuracy of the algorithm.

Description

Technical Field

[0001] This invention relates to the field of yarn quality prediction technology, and specifically to a yarn quality prediction method based on meta-learning. Background Technology

[0002] Yarn quality prediction technology forecasts the final yarn quality based on material selection, production processes, and key parameter configurations. Because the relationship between these key parameters and the final yarn quality prediction is not a simple linear one, traditional machine learning algorithms are not fully adaptable to quality prediction, resulting in low accuracy and efficiency. With the development of computer hardware and software technology, especially the emergence of deep learning methods, new opportunities and challenges have arisen for using artificial intelligence for classification and prediction. Therefore, whether deep learning technology can be used for yarn quality prediction, addressing the adaptability and accuracy issues of prediction techniques, and improving the practicality of yarn quality prediction is significant for enhancing factory intelligence and increasing production efficiency.

[0003] Currently, many institutions both domestically and internationally have conducted research on yarn quality prediction methods based on machine learning or deep learning. Donghua University, focusing on the impact of processing sequence on yarn quality and integrating convolutional and temporal networks in convolutional neural networks, designed three models: CNN-BP, CNN-GRNN, and CNN-LSTM for prediction. The input parameters include 13 cotton fiber parameters and 3 process parameters. Zhang Yutong et al. selected combed wool yarn samples and used a computational method comprised of grey relational analysis and a data optimization method based on the degree of input-output influence to select the four process parameters with the greatest impact on yarn unevenness and breaking strength from seven process parameters, establishing a yarn quality prediction model based on a BP neural network. Jiang et al. proposed a multilayer neural network (MNN) based on BLS to predict yarn unevenness, selecting 26 parameters for prediction.

[0004] While the above prediction methods represent a certain level of technological advancement, they still suffer from the following drawbacks:

[0005] 1. Low algorithm accuracy: Traditional machine learning or deep learning algorithms have failed to improve the yarn quality prediction problem, and direct application of existing methods will result in low accuracy.

[0006] 2. Existing methods typically input a large number of features directly into the model, resulting in a particularly large number of parameters, while failing to effectively analyze the relationships between different parameters.

[0007] 3. The amount of data currently used for yarn quality prediction is relatively small, resulting in a scarcity of samples. Existing models are unable to deeply explore the complex relationships within a limited amount of data. Summary of the Invention

[0008] This invention provides a yarn quality prediction method based on meta-learning. By integrating the grey management method and the mutual information algorithm, feature selection is achieved, thereby ensuring that the input data is closely related to yarn quality. At the same time, a BiLSTM model with confidence is proposed to explore the relationship between features and yarn quality, eliminate the synergistic influence of multicollinear (highly correlated) independent variables in different features on the results, and solve the problem of sample scarcity in spinning quality prediction through meta-learning, thereby improving the prediction accuracy of the algorithm.

[0009] To achieve the above objectives, the technical solution of the present invention is as follows:

[0010] A yarn quality prediction method based on meta-learning includes the following steps:

[0011] Step 1: Use the grey relational analysis method to analyze the weights of the input parameters affecting yarn quality, and use the mutual information algorithm to remove redundant parameters;

[0012] Step 2: By introducing a BiLSTM model with confidence, the correlation between input parameters and yarn quality is explored;

[0013] Step 3: Construct a meta-machine learning model using other quality prediction data from the textile industry and production quality prediction methods from other industries, and use a small amount of yarn quality prediction data to make the final yarn quality prediction.

[0014] Preferably, in step 1, the input parameters affecting yarn quality include cotton fiber diameter, fiber length, cotton fiber short fiber ratio, cotton fiber specific strength, traveler number, draw ratio, spindle speed, twist, and yarn fineness. The predicted yarn quality evaluation criteria are: yarn evenness and breaking strength.

[0015] Preferably, in step 1, the grey relational analysis method is first used to calculate the weights of all input parameters and sort them. After numerical normalization, the maximum value of all input parameters is used as a reference sequence, defined as y(k) = {1, 1, 1, ..., 1}, where k is the input parameter index. All different input data ξ are compared using formula (1). i (k), and then the weight for each input parameter is calculated using formula (2):

[0016]

[0017]

[0018] In formula (1), x i(k) is the i-th parameter of the k-th input sample, ρ is the resolution coefficient, which is usually taken as 0.5; in formula (2), n is the number of all samples.

[0019] Preferably, in step 1, the first m input parameter feature sets X are selected based on the weights of the input parameters, and then the set X is optimized using a mutual information algorithm, including the following specific steps:

[0020] (1A) The algorithm for calculating the mutual information of two vectors is defined as follows: Given two random variables x and y, the mutual information of x and y is obtained by formula (3) through the probability density functions p(x), p(y), p(x,y):

[0021]

[0022] In formula (3), p(x), p(y), and p(x,y) are all probability density functions;

[0023] (2A) Calculate the redundancy among input parameters in set S using formula (4):

[0024]

[0025] In formula (4), R represents the mutual information between features, S is the set of all input parameters, |S| is the number of input parameters in the set, and I(x) i x j ) represents the mutual information between the two parameters.

[0026] (3A) Select the input parameters in set S in order according to the gray-level correlation matrix, assuming that the obtained parameter subset is S. m-1 Then from the remaining parameter set XS m-1 Choose a parameter x j Ensure parameter x j The obtained R(S) is the smallest among all choices, as shown in formula (5):

[0027]

[0028] In formula (5), S m-1 Let m-1 be a subset of parameters, and S be a subset of parameters. m-1 Quantity;

[0029] Using formula (5), we finally obtain a subset that minimizes R(S).

[0030] Preferably, in step 2, the weights of different input parameters are allocated by introducing a BiLSTM model with confidence, and the outputs of the two directions of the input sequence are merged by using a bidirectional LSTM algorithm (BiLSTM).

[0031] Preferably, in the BiLSTM model that introduces confidence, the quality data of all input parameters are first clustered to identify noisy frames in the quality data. Input parameters that are far from the cluster center in the quality data are assigned a lower confidence level, while input parameters that are far from the cluster center in the quality data are assigned a higher confidence level. After obtaining the confidence level, the input parameters with confidence are introduced into the BiLSTM model.

[0032] Confidence score calculation method: For each cluster center, calculate the distance between all parameters within its cluster and the cluster center, and construct K sets. To ensure that outliers that are too far away do not affect the confidence score calculation, Gaussian normalize the distances of all data within each set; define j... t The confidence level for the current input parameter is expressed by formula (6), which is used to calculate its value:

[0033] j t =σ(1-d nomarliz (x t (6)

[0034] In formula (6), σ is the activation function, and d nomarlize (x t For parameter x t Calculate its normalized distance;

[0035] The input threshold of BiLSTM is i t The forgetting threshold is f t The state at the previous moment was C. t-1 The input to be updated is The current state is C. t Then we have:

[0036]

[0037] Formula (7) adds a new confidence level between the original input gate and the forget gate, based on the input parameter x. t The confidence weights are introduced into BiLSTM, and finally the input parameter features output by the BiLSTM model are fed into the classifier for quality prediction.

[0038] Preferably, step 3 includes the following specific steps:

[0039] (1B) Construct a convolutional neural network learning model, defined as BiLSTM(T) k ), k represents different industry datasets, i.e., task numbers. The input parameter size of all datasets is kept consistent and converted to an m×n size. The final result prediction is achieved through multiple convolutional layers.

[0040] (2B) Let the quality parameters for each industry be... Where i is the parameter index and k is the dataset, i.e., the task index, for each task, BiLSTM in specific step (1B) is used for learning, and the relationship between the input parameters and the final quality is constructed. After training, the model parameters for the task can be obtained as θ. Then, the loss function for each task obtained through BiLSTM is L. k ;

[0041] (3B) Summing the loss functions obtained from training the neural network for all industry tasks yields the following loss function for meta-learning:

[0042] L mmeta =∑ k∈Tasks L k (8)

[0043] In formula (8), L k This is represented as the loss for the k-th task;

[0044] (4B) The objective function of meta-learning is calculated according to formula (9). The sum of test errors obtained on different tasks is used as the optimization object of meta-learning. The input parameters are updated by gradient descent, and finally a set of initial input parameters is obtained and applied to model optimization:

[0045]

[0046] In formula (9), θ is the parameter to be learned, and L k This is represented as the loss for the k-th task;

[0047] (5B) For the learned parameter θ, an optimal common initial point is found by using other tasks. However, when it is applied to the data of spinning product quality prediction, the optimization point of the task is close to the common initial point. Therefore, the model uses a small amount of spinning product quality data to perform a few steps of gradient descent to approach the optimal point of the task, thereby giving the model the ability to learn and adapt quickly through simple fine-tuning. Therefore, the data and prediction results of spinning product quality prediction are applied to the BiLSTM in the specific step (1B) where the input parameters have been updated to fine-tune the input parameters.

[0048] The beneficial effects of the yarn quality prediction method based on meta-learning in this invention are as follows:

[0049] 1. This invention uses the grey relational analysis method to analyze the weights of input parameters that affect yarn quality, and employs a mutual information algorithm to remove redundant features of the input parameters. This allows for the optimization of a large number of input parameters, improving calculation speed and the accuracy of calculation results, and ensuring that the redundancy of input parameter features is minimized.

[0050] 2. This invention proposes a BiLSTM model that incorporates confidence levels, which can deeply explore the relationship between input parameter features and yarn quality, thereby improving the model's prediction accuracy and efficiency.

[0051] 3. This invention solves the problem of scarce samples for yarn quality prediction by using a meta-machine learning method, thereby improving the prediction accuracy of the algorithm. Attached Figure Description

[0052] Figure 1 Bi-LSTM model unit with confidence level introduced:

[0053] In the image: i t For the input threshold, f t To set the forgetting threshold, the state at the previous time step was C. t-1 The hidden feature of the previous state is h. t-1 The input to be updated is The current state is C. t The current hidden feature output is h. t j t σ represents the confidence level of the current input parameters, and tanh and σ are the activation functions. Detailed Implementation

[0054] The following description provides a detailed explanation of the embodiments of the present invention in a step-by-step manner. This description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0055] In the description of this invention, it should be noted that the terms "upper," "lower," "left," "right," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or a specific orientational structure and operation. Therefore, they should not be construed as limiting this invention.

[0056] In the initial embodiment, the present invention provides a yarn quality prediction method based on meta-learning, comprising the following steps:

[0057] Step 1: Use the grey relational analysis method to analyze the weights of the input parameters affecting yarn quality, and use the mutual information algorithm to remove redundant parameters;

[0058] Step 2: By introducing a BiLSTM model with confidence, the correlation between input parameters and yarn quality is explored;

[0059] Step 3: Construct a meta-machine learning model using other quality prediction data from the textile industry and production quality prediction methods from other industries, and use a small amount of yarn quality prediction data to make the final yarn quality prediction.

[0060] In this embodiment, the various input parameters play different roles in the final yarn quality assessment. Some input parameters have a significant impact on yarn quality, while others have a smaller impact. Furthermore, some input parameters contain redundant information, and duplicate information input into the model can lead to a decrease in algorithm accuracy. This invention optimizes the input parameter selection method for yarn quality prediction through step 1, dynamically adjusting the model input parameters and integrating the grey relational analysis method and the mutual information algorithm to ensure minimal redundancy of input parameter features. However, since the grey relational analysis algorithm only considers the correlation between input parameters and selects the top-ranked input parameter features in a certain order, it cannot guarantee that the final feature set obtained is the optimal feature set. Features extracted by the grey relational analysis method may have strong redundancy, leading to a decrease in computational speed without improving the algorithm's accuracy. Therefore, redundant features need to be removed. Based on the top k feature sets X selected by the grey relational analysis algorithm, this patent uses the mutual information algorithm to optimize set X.

[0061] Furthermore, since a single-layer LSTM model cannot fully express the relationship between input parameters, the output of the current node depends not only on previous nodes but also on subsequent nodes. Therefore, this invention employs a bidirectional LSTM algorithm (BiLSTM) to merge the outputs of the two streams. However, for all input parameters, different input parameters have varying degrees of importance, and some input parameters may contain noise. This invention assigns weights to different input parameters to eliminate noise and interference from the asymmetry of different data sources, thereby improving recognition accuracy. Based on this, this invention proposes a BiLSTM model incorporating confidence to mine the correlation between input parameter features and yarn quality.

[0062] Furthermore, since the existing yarn quality prediction algorithms based on the above input parameters have too little data, it is difficult to collect and apply them to general machine learning algorithms in a short period of time. Therefore, this invention adopts a yarn quality prediction method based on meta-learning, which makes full use of other quality prediction data in the textile industry and production quality prediction methods in other industries to construct a meta-machine learning model, and uses the existing small amount of yarn quality prediction data to make the final yarn quality prediction.

[0063] In a further embodiment, in step 1, the input parameters affecting yarn quality include cotton fiber diameter, fiber length, cotton fiber short fiber ratio, cotton fiber specific strength, traveler number, draw ratio, spindle speed, twist, and yarn fineness. The predicted yarn quality evaluation criteria are: yarn evenness and breaking strength.

[0064] In a further embodiment, in step 1, the grey relational analysis method is first used to calculate the weights of all input parameters and sort them. After numerical normalization, the maximum value of all input parameters is used as a reference sequence, defined as y(k) = {1, 1, 1, ..., 1}, where k is the input parameter index. All different input data ξ are compared using formula (1). i (k), and then the weight for each input parameter is calculated using formula (2):

[0065]

[0066]

[0067] In formula (1), x i (k) represents the i-th parameter of the k-th input sample, ρ is the resolution coefficient, which is generally taken as 0.5; in formula (2), n is the total number of samples.

[0068] In a further embodiment, step 1 involves selecting the top m input parameter feature sets X based on the weights of the input parameters, and then optimizing set X using a mutual information algorithm, including the following specific steps:

[0069] (1A) The algorithm for calculating the mutual information of two vectors is defined as follows: Given two random variables x and y, the mutual information of x and y is obtained by formula (3) through the probability density functions p(x), p(y), p(x,y):

[0070]

[0071] In formula (3), p(x), p(y), and p(x,y) are all probability density functions;

[0072] (2A) Calculate the redundancy among input parameters in set S using formula (4):

[0073]

[0074] In formula (4), R represents the mutual information between features, S is the set of all input parameters, |S| is the number of input parameters in the set, and I(x) i x j ) represents the mutual information between the two parameters.

[0075] (3A) Select the input parameters in set S in order according to the gray-level correlation matrix, assuming that the obtained parameter subset is S. m-1 Then from the remaining parameter set XS m-1 Choose a parameter x j Ensure parameter x j The obtained R(S) is the smallest among all choices, as shown in formula (5):

[0076]

[0077] In formula (5), S m-1 Let m-1 be a subset of parameters, and S be a subset of parameters. m-1 Quantity;

[0078] Using formula (5), we finally obtain a subset that minimizes R(S).

[0079] In a further embodiment, in step 2, the weights of different input parameters are assigned by introducing a BiLSTM model with confidence, and the outputs of the two directions of the input sequence are merged by using a bidirectional LSTM algorithm (BiLSTM).

[0080] In a further embodiment, in the BiLSTM model with confidence incorporation, the quality data of all input parameters are first clustered to identify noisy frames in the quality data. Input parameters that are far from the cluster centers in the quality data are assigned lower confidence scores, while input parameters that are far from the cluster centers in the quality data are assigned higher confidence scores. After obtaining the confidence scores, the input parameters with confidence scores are introduced into the BiLSTM model.

[0081] Confidence score calculation method: For each cluster center, calculate the distance between all parameters within its cluster and the cluster center, and construct K sets. To ensure that outliers that are too far away do not affect the confidence score calculation, Gaussian normalize the distances of all data within each set; define j... t The confidence level for the current input parameter is expressed by formula (6), which is used to calculate its value:

[0082] jt =σ(1-d nomarliz (x t (6)

[0083] In formula (6), σ is the activation function, and d nomarlize (xt) represents the parameter x. t Calculate its normalized distance;

[0084] The input threshold of BiLSTM is i t The forgetting threshold is f t The state at the previous moment was C. t-1 The input to be updated is The current state is C. t Then we have:

[0085]

[0086] Using formula (7), a new confidence level is added between the original input gate and the forget gate. The confidence level weight of the input parameter xt is introduced into BiLSTM. Finally, the input parameter features output by the BiLSTM model are sent to the classifier for quality prediction.

[0087] In a further embodiment, step 3 includes the following specific steps:

[0088] (1B) Construct a convolutional neural network learning model, defined as BiLSTM(T) k ), k represents different industry datasets, i.e., task numbers. The input parameter size of all datasets is kept consistent and converted to an m×n size. The final result prediction is achieved through multiple convolutional layers.

[0089] (2B) Let the quality parameters for each industry be... Where i is the parameter index and k is the dataset, i.e., the task index, for each task, BiLSTM in specific step (1B) is used for learning, and the relationship between the input parameters and the final quality is constructed. After training, the model parameters for the task can be obtained as θ. Then, the loss function for each task obtained through BiLSTM is L. k ;

[0090] (3B) Summing the loss functions obtained from training the neural network for all industry tasks yields the following loss function for meta-learning:

[0091] L meta =∑ k∈Tasks L k (8)

[0092] In formula (8), L kThis is represented as the loss for the k-th task;

[0093] (4B) The objective function of meta-learning is calculated according to formula (9). The sum of test errors obtained on different tasks is used as the optimization object of meta-learning. The input parameters are updated by gradient descent, and finally a set of initial input parameters is obtained and applied to model optimization:

[0094] min θ L meta =min θ ∑ k∈Tasks L k (9)

[0095] In formula (9), θ is the parameter to be learned, and L k This is represented as the loss for the k-th task;

[0096] (5B) For the learned parameter θ, an optimal common initial point is found by using other tasks. However, when it is applied to the data of spinning product quality prediction, the optimization point of the task is close to the common initial point. Therefore, the model uses a small amount of spinning product quality data to perform a few steps of gradient descent to approach the optimal point of the task, thereby giving the model the ability to learn and adapt quickly through simple fine-tuning. Therefore, the data and prediction results of spinning product quality prediction are applied to the BiLSTM in the specific step (1B) where the input parameters have been updated to fine-tune the input parameters.

Claims

1. A yarn quality prediction method based on a meta-learning method, characterized by: It comprises the following steps: Step 1, the weight of the input parameters affecting the quality of the yarn is analyzed by using the grey correlation degree method, and the redundant parameters are removed by using the mutual information algorithm; Step 2, the correlation between the input parameters and the quality of the yarn is mined by introducing the BiLSTM model with confidence; Step 3, the meta-machine learning model is constructed by using other quality prediction data in the textile industry and other industry production quality prediction methods, and a small amount of yarn quality prediction data is used for final yarn quality prediction; The step 3 comprises the following specific steps: (1B) Construct a convolutional neural network learning model, defined as , k is different industry data set, that is, task serial number, the input parameter size of all data sets is consistent, and is converted to Size, and the final result is predicted through multiple convolutional layers; (2B) Set the quality parameter of each industry as Where i is the parameter number, k is the data set, that is, the task number, for each task, the BiLSTM in the specific step (1B) is used for learning, and the relationship between the input parameters and the final quality is constructed, and after training, the model parameters of the task can be obtained as Then the loss function of each task obtained by BiLSTM is ; (3B) the loss function obtained by training all industry tasks through the neural network is summed to obtain the loss function and for meta-learning: （8）； In equation (8), denotes the loss for the kth task. (4B) the objective function of meta-learning is calculated according to formula (9), the sum of test errors obtained on different tasks is used as the optimization object of meta-learning, the input parameters are updated by gradient descent, and finally a set of initialization input parameters is obtained, which is applied to model optimization: （9）； In Equation (9), is a parameter to be learned, represents a loss for the kth task; (5B) the data and prediction results of the spinning product quality prediction are applied to the BiLSTM in the specific step (1B) with updated input parameters for input parameter fine-tuning.

2. A yarn quality prediction method based on meta-learning method as claimed in claim 1, wherein: In the step 1, the input parameters affecting the quality of the yarn include the fiber diameter, fiber length, cotton fiber short fiber rate, cotton fiber specific strength, steel ring number, draft ratio, spindle speed, fine twist, and yarn fineness, and the predicted yarn quality evaluation standard is: unevenness and breaking strength.

3. A yarn quality prediction method based on meta-learning method as claimed in claim 2, wherein: In step 1, firstly, the grey correlation method is used to calculate the weight of each input parameter and to sort, after numerical normalization, the maximum value of all input parameters is taken as a reference sequence, defined as , is the input parameter number, all different input data are compared by formula (1) Then, the weight of each input parameter is calculated by formula (2): （1）； （2）； In formula (1), where is the i-th parameter of the k-th input sample, is a resolution coefficient, taken as 0.5; in formula (2), n is the number of all samples.

4. A yarn quality prediction method based on meta-learning method as claimed in claim 3 characterized in that: In the step 1, the first m input parameter feature sets X are screened out according to the weight ordering of the input parameters, and then the set X is optimized by using the mutual information algorithm, comprising the following specific steps: (1A) The mutual information algorithm for two vectors is defined as follows: Given two random variables x and y, the mutual information of x and y is given by the probability density functions From equation (3), we get: （3）； In equation (3), are probability density functions; (2A) the redundancy degree between the input parameters in the set S is calculated by formula (4): （4）； In Equation (4), R is the mutual information between the features, S is the set of all input parameters, is the number of input parameters in the set, is the mutual information of the two parameters; (3A) Select the input parameters in the set S according to the order of the gray correlation matrix, assuming that the obtained parameter subset is , then select a parameter from the remaining parameter set , ensure that the parameter obtained by is the smallest in all selections, as shown in equation (5): （5）； In equation (5), is a parameter subset, m-1 is the number of By formula (5), finally obtainable make The smallest one.

5. A yarn quality prediction method based on meta-learning method as claimed in claim 4, characterized by: In the step 2, the BiLSTM model with confidence is introduced to allocate the weight of different input parameters, and the bidirectional LSTM algorithm BiLSTM is used to combine the outputs of the two directions of the input sequence.

6. A yarn quality prediction method based on meta-learning method as claimed in claim 5 characterized in that: In the BiLSTM model with confidence, first, the quality data of all input parameters is clustered to identify the noise frames in the quality data, and the input parameters with larger distance from the cluster center in the quality data are assigned with lower confidence, and the input parameters with smaller distance from the cluster center in the quality data are assigned with higher confidence; After obtaining the confidence, the input parameters with confidence are introduced into the BiLSTM model; The confidence calculation method: for each cluster center, the distance between all parameters in the cluster center and the cluster center is calculated, and K sets are constructed to ensure that the outlying points with too far distance do not affect the calculation of the confidence; the distance of all data in each set is Gaussian normalized; define The confidence for the current input parameter is represented, and the value is calculated by formula (6): （6）； In equation (6), is an activation function, is a parameter compute the distance after normalizing it; The input threshold of the BiLSTM is , the forgetting threshold is , the state of the previous moment is , the input to be updated is , and the current state is , then （7）； Through formula (7), a new confidence is added between the original input gate and the forget gate, and the confidence weight of the input parameter is introduced into the BiLSTM, and finally the input parameter features output by the BiLSTM model are sent to the classifier for quality prediction.

Images