A method for on-line prediction of process quality in a cuttage plant

Abstract

The application discloses a kind of process quality online prediction methods of silk making workshop, comprising the following steps: (1): process parameters and quality indicators are obtained from MES database;(2): the regularized representation and data fusion of the obtained silk making process parameter historical data are carried out, and pre-processing is carried out;(3): based on sequence-to-sequence learning structure, adopt mixed neural network to predict the loose rehydration quality index of next time as the target to establish workshop offline quality prediction model, i.e.CTCN_A_B network model;(4): by establishing communication with the sensor of silk making workshop online monitoring equipment, obtain online silk making workshop process data, select the same loose rehydration process data as offline data and process;(5): using double thread method, the network structure and parameters of feature extraction module in offline prediction model are frozen using transfer learning, and are transferred to online process quality prediction model, realize dynamic adjustment prediction model, satisfy production actual change prediction demand.

Description

Technical Field

[0001] This invention belongs to the field of industrial intelligent online prediction technology, specifically relating to an online prediction method for process quality in a silk-making workshop. Background Technology

[0002] "Intelligent manufacturing" is the development direction of my country's manufacturing industry, and intelligent production technology is listed as an important research area. The scientific and rational process quality control and management technology level and development have become key.

[0003] Currently, tobacco processing workshops have only achieved preliminary automation, failing to utilize online forecasting for real-time control of production plans and process parameters, thus hindering the realization of the strategic goal of a smart factory. Therefore, it is essential to conduct in-depth research into the correlation and inherent characteristics between process parameters and quality indicators, and to establish predictive models to characterize key control indicators in tobacco processing technology.

[0004] However, in actual production, processing data from different processes and production lines are often mixed together, making effective organization difficult. This poses a challenge to achieving factory digitalization. Furthermore, unforeseen circumstances such as changes in equipment parameters or the production environment can render fixed feature mining methods ineffective in all scenarios. Fixed model parameters within machines cannot be learned, updated, or reset according to actual conditions. Therefore, the ability to continuously learn and adapt to feature changes is a necessary research direction for the current transformation of intelligent factories. Considering the importance of timely response in actual production, rapid and effective autonomous learning is required. Therefore, utilizing the feature learning capabilities and rapid adaptive learning of machine learning in the silk-making process is a key means to achieve online prediction of process quality. This invention is proposed to address this issue. Summary of the Invention

[0005] The purpose of this invention is to provide an online prediction method for process quality in silk-making workshops, which solves the problems existing in the prior art: relying on traditional mathematical statistics methods and requiring expert experience to preset thresholds, resulting in a certain degree of subjectivity in the acquired correlation features; currently, industrial quality prediction focuses on artificial intelligence fields such as machine learning and neural networks, but using a single network structure cannot fully extract the complex temporal correlation information in industrial production, losing local information expression, resulting in prediction results that cannot achieve satisfactory accuracy; at the same time, existing prediction networks are based on feature extraction capabilities trained on offline historical data, but in practical applications, due to unforeseen circumstances, there are differences between online and offline data, and these differences become more and more serious over time, leading to serious distortion of prediction results.

[0006] The technical solution adopted in this invention is as follows:

[0007] A method for online prediction of process quality in a yarn processing workshop includes the following steps:

[0008] Step (1): Obtain historical data of yarn-making process parameters from different production plants and production lines in the MES database, including process parameters and quality indicators for different processes;

[0009] Step (2): Regularize and fuse the historical data of the obtained yarn-making process parameters, and perform preprocessing;

[0010] Step (3): Select the loose moisture regain process data from the historical data of the silk making process after preprocessing in step (2), and establish an offline quality prediction model for the workshop, namely the CTCN_A_B network model, based on the sequence-to-sequence learning structure and using a hybrid neural network to predict the loose moisture regain quality index at the next moment.

[0011] Step (4): Establish communication with the sensors of the online monitoring equipment in the yarn making workshop to obtain the online yarn making workshop process data, select the loose re-moistening process data that is the same as the offline data and process it;

[0012] Step (5): Using a dual-thread method, the network structure and parameters of the feature extraction module in the offline prediction model are frozen by transfer learning and transferred to the online process quality prediction model to achieve dynamic adjustment of the prediction model and meet the actual production change prediction needs.

[0013] Preferably, step (2) specifically includes the following steps:

[0014] Step (2.1): By analyzing the process schemes, process combinations, data acquisition methods, and data naming similarities of each plant area, formulate the configuration content of the silk-making process to achieve the regularized representation and data fusion of different process data in different plants area;

[0015] Step (2.2): For various abnormal data, data preprocessing is achieved through multiple processes such as data cleaning, data type conversion, data batch number collection, outlier handling, and data normalization.

[0016] Preferably, step (3) specifically includes the following steps:

[0017] Step (3.1): Select the loose re-moistening process data from the complete yarn making process data, and analyze the process parameters that have an important impact on the quality indicators. Select these process parameter data as the input data for the offline prediction model and train the prediction model.

[0018] Step (3.2): Validate the effectiveness of the constructed model by comparing the fitted values, standard errors, and evaluation absolute percentage errors with traditional regression methods, shallow machine learning methods, and neural networks (DNN, LSTM, GRU, etc.).

[0019] Preferably, step (3.1) specifically involves the following steps:

[0020] Step (3.1.1): The key process parameters selected are: process flow rate, process hot air temperature, cumulative material amount, water flow rate, cumulative water amount, and steam automatic valve opening. The quality indicators are: discharge temperature and discharge moisture content. These data are divided into training set and test set according to a certain ratio. The offline prediction model is trained using the training set and tested using the test set.

[0021] Step (3.1.2): Concatenate the encoder components of CNN_TCN and train the preprocessed offline prediction model using the training set data; use the BiLSTM network to mine the bidirectional temporal features in the loose moisture quality index data, connect the temporal correlation features of the loose moisture process parameters and the temporal features of the quality index data, and then use the connection vector as the input of the multi-head attention network to redistribute the weights of the features to achieve the fusion of multi-source feature information;

[0022] Step (3.1.2): Construct an offline prediction model for the process quality of the silk-making workshop based on a sequence-to-sequence learning structure. This model consists of two parts: an encoder component using a concatenated CNN_TCN and a BiLSTM decoder component incorporating a multi-head attention mechanism. In the encoder component, a convolutional neural network (CNN) is first used to automatically learn deep latent features in the process parameter data, effectively eliminating reliance on expert experience to obtain correlation information. Subsequently, a concatenated temporal convolutional neural network (TCN) is used to effectively analyze the temporal information of the workshop process while maintaining the causal convolutional characteristics of the process parameters. Furthermore, causal convolution is applied in the TCN network to improve the processing efficiency of long-span memory units, enabling the extraction of long-distance temporal features of process parameters. Simultaneously, in the decoder component, a BiLSTM network is used to extract historical temporal features from the quality indicator data. Then, a multi-head attention network is used to fuse all the extracted features and redistribute the weights of the features to achieve multi-source feature information fusion. Finally, the predicted value of the quality indicator is output through a fully connected layer.

[0023] Preferably, step (5) specifically includes the following steps:

[0024] Step (5.1): To cope with various emergencies that occur in actual production, two model platforms are established using the dual-thread method: a training platform and a prediction platform. The training platform is used to continuously train the model to achieve the purpose of real-time model updates, and the prediction platform is used to predict the quality of the silk-making workshop at future moments.

[0025] Step (5.2): Use the constructed CTCN_A_B network on the training platform to train the model on offline historical data, set the model prediction accuracy threshold, and save the model structure and parameters that reach the threshold.

[0026] Step (5.3): On the training platform, offline and online data obtained from the silk-making workshop under the same production process are used. The production information contained in them is similar, which makes the feature information mining methods interchangeable. On the prediction platform, the feature extraction network in the pre-trained model is frozen and transferred to the online prediction model. The fully connected layer of the model is randomly initialized to adapt to the actual production change requirements, and the silk-making online prediction model is rebuilt.

[0027] The beneficial effects of this invention are:

[0028] 1. The method of this invention combines the ideas of deep learning and transfer learning. Based on a sequence-to-sequence learning structure, it simultaneously learns historical and feature information of industrial quality indicators. In the encoder, a convolutional neural network is used to extract deep features of process parameters, and a temporal convolutional network is connected in series to deeply mine their potential temporal information. In the decoder, a bidirectional long short-term memory network is used to learn the historical temporal information of the quality indicators at both the front and back ends, and a multi-head attention mechanism is embedded to fuse all the extracted feature information, thereby improving the model's expressive power. The network model is constructed with future industrial quality as the prediction target. The constructed model is pre-trained to obtain a better network structure, and transfer learning is used to achieve accurate short-term online industrial quality prediction, providing a foundation for the construction of intelligent workshops.

[0029] 2. The method of the present invention addresses the issue that the inconsistent data sources in the silk-making workshop, the differences in the structural properties and data volume of various process parameters, and the resulting significant multi-source heterogeneous characteristics of silk-making process data, which makes data fusion difficult. The present invention proposes a method for solving the problem of multi-source heterogeneous data acquisition, fusion, and data preprocessing in silk-making.

[0030] 3. The silk-making process quality prediction model of this invention breaks away from the experience-dependent limitations of traditional correlation algorithms, autonomously and fully extracts the complex correlation time-series information contained in the process data, and realizes the needs of workshop quality prediction.

[0031] 4. This invention utilizes transfer learning to achieve online quality prediction and applies a dual-thread method to simultaneously perform training and prediction, thereby updating model parameters in real time according to actual production conditions and achieving adaptive quality prediction capabilities. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the process of the present invention.

[0033] Figure 2 The results of the batch data collection and analysis in step (2) of this invention are as follows.

[0034] Figure 3 This is a structural diagram of the silk-making workshop quality prediction model of the present invention.

[0035] Figure 4 This is a diagram showing the relationship between the loose rehydration process parameters of the present invention.

[0036] Figure 5 This is a diagram of the temporal convolutional network structure of the present invention.

[0037] Figure 6 This is a schematic diagram of the LSTM structure of the present invention.

[0038] Figure 7 This invention provides a comparative study on the prediction of moisture content in loose, rehydrated yarn output.

[0039] Figure 8 This invention provides a comparative study on the prediction of loose, rehydrated, and discharged material temperature for filament processing. Detailed Implementation

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the scope of the present invention is not limited to the description.

[0041] Example: Figure 1 As shown, an online prediction method for process quality in a yarn-making workshop mainly includes the following steps:

[0042] Step (1): Analyze the specific process flow of the yarn making workshop, determine the relevant process parameters and process indicators, and obtain historical data of yarn making process parameters under different production plants and production lines from the database of the Manufacturing Execution System (MES), including process parameters and quality indicators of different processes;

[0043] Step (2): Since different yarn-making process data are collected by different sensors, there is a data fusion problem. Therefore, it is necessary to study a multi-source heterogeneous yarn-making process data acquisition and fusion method, and to preprocess the obtained historical process data. The specific steps of the yarn-making industrial workshop data acquisition and fusion method include:

[0044] Step (2.1): Research the data fusion method of multi-source heterogeneous fiber making process. By analyzing the process schemes, process combinations and data acquisition methods of each plant area, as well as the commonality of data naming, formulate the configuration content of the fiber making process, and realize the regularized representation and data fusion of different process data in different plants.

[0045] Step (2.2): Data preprocessing and analysis method for silk making process. In order to deal with various abnormal data that occur in the actual silk making process, data preprocessing is achieved through multiple processes such as data cleaning, data type conversion, data batch number collection, outlier handling and data normalization.

[0046] Step (2.2.1): Data cleaning mainly includes deleting useless data columns, replacing special characters (such as replacing "--" with 0), and deleting the row numbers containing null values.

[0047] Step (2.2.2): Data type conversion is required after data cleaning to convert the data data types for subsequent data preprocessing and analysis. Simultaneously, it is necessary to configure a dictionary based on the preceding process parameters and convert them by category: convert "Time column name" to timestamp, convert "Process model name", "Process name", "Batch number", and "Brand" to string type, and convert other process parameter data types to numeric type; data batch and brand number collection;

[0048] Step (2.2.3): Data batch and grade collection is based on batch number, combined with grade, model name, process name, etc. The start and end times of each batch are extracted, and the batch duration and sampling interval are calculated. The results are as follows: Figure 2 As shown;

[0049] Step (2.2.4): Outlier handling. First, the 3σ method is used to identify outliers in the data, that is, outliers exceeding 3 times the standard deviation. Then, the Savitzky-Golay smoothing method is used to smooth and filter outliers.

[0050] Step (2.2.5): Data normalization uses the following formula to handle the problem of inconsistent dimensions and orders of magnitude of process parameters, realize data normalization, and control all parameters within the range of [0,1].

[0051]

[0052] In the formula: X min X is the minimum value among the parameter index data. max This represents the maximum value among the parameter index data.

[0053] Step (3): It is necessary to establish an offline quality prediction model for specific processes in the silk-making process. Select a certain process data from the historical data of the silk-making process preprocessed in the above steps. This invention selects the loose moisture regaining process as an example to construct an offline quality prediction model. Specifically, based on the sequence-to-sequence learning structure, a hybrid neural network is used to construct a quality prediction model for the silk-making industrial workshop with the goal of predicting the loose moisture regaining quality index at the next moment, namely the CTCN_A_B network model. The specific steps for constructing the offline quality prediction model for the silk-making industrial workshop include:

[0054] Step (3.1): Select any process in the silk-making process and construct an offline quality prediction model for the silk-making workshop. The model structure is as follows: Figure 3 As shown. From the complete yarn-making process data obtained in the above steps, the processing data for this process is selected, and the process parameters that have a significant impact on quality indicators are analyzed. These process parameter data are then used as input data for the offline prediction model to train the prediction model.

[0055] Step (3.1.1): This invention performs predictive analysis on individual processes within the yarn-making process. Therefore, through steps (1) and (2), the overall processing data of the pre-treated yarn-making process is obtained, and the complete processing parameter data of the loose re-moistening process is selected. The loose re-moistening process is then analyzed. Figure 4 A diagram showing the relationship between specific process parameters for loose rehydration is provided. Figure 4 The key process parameters affecting the quality indicators of loose and rehydrated materials can be identified, specifically including six process parameters: process flow rate, process hot air temperature, cumulative material quantity, water flow rate, cumulative water quantity, and automatic steam valve opening. The quality indicators in loose and rehydrated materials are discharge temperature and discharge moisture content. Therefore, these parameters were ultimately selected for predictive analysis. A total of 39,000 historical data points were extracted from March to September 2020. These data were divided into training and testing sets in a 7:3 ratio. The offline prediction model was trained using the training set and tested using the testing set to determine its performance.

[0056] Step (3.1.2): The CTCN_A_B network model is built on a sequence-to-sequence learning structure, and it mainly consists of two parts: a cascaded CNN_TCN encoder component and a BiLSTM decoder component that integrates a multi-head attention mechanism. The following is an analysis of the composition of each component:

[0057] (3.1.2.1) The encoder component of CNN_TCN is concatenated, and the preprocessed offline prediction model is trained using the training set data. Specifically, the normalized process parameter time series data X=(x1,x2,…,x…) is used. T )=(x 1 ,x2 ,…,x N ) T The aforementioned process flow rate, process hot air temperature, cumulative material quantity, water flow rate, cumulative water quantity, and steam automatic valve opening degree are used as inputs to the encoder component. First, these are fed into a CNN network, which relies on the convolutional neural network's ability to automatically learn deep latent features in the data, effectively eliminating reliance on expert experience to obtain correlation parameters. Utilizing the unique convolution and pooling structures of the CNN, the sliding window operation of the convolutional kernel captures the static features of the time-series data. Then, the scale invariance of key features in the pooling layer is used to reduce the dimensionality of the extracted features, highlighting key features. Simultaneously, parameter sharing reduces the network's complexity. Finally, the output of the CNN network is... As input, using a temporal convolutional neural network, Figure 5 The diagram shows the structure of a temporal convolutional network. While preserving the causal convolutional properties of the process, it effectively analyzes the temporal information of the workshop process. Dilated causal convolution is applied to improve the processing efficiency of long-span memory units, enabling the extraction of temporal features of process parameters. The specific calculation formula is as follows.

[0058]

[0059] In the formula: (sd·i) represents the sequence corresponding to the elements in the convolution kernel; f represents the dilated convolution operation;

[0060] Each residual block in a temporal convolutional network consists of a direct mapping part and a residual part, expressed as follows:

[0061]

[0062] In the formula: represents the dependency information contained in the nth convolutional layer; F represents the dilated convolution operation, i.e. the direct mapping part; R represents the residual mapping operation across layers.

[0063] (3.1.2.2) The BiLSTM decoder component incorporating a multi-head attention mechanism, with the normalized quality index sequence as Y = (y1, y2, ..., y... T )∈R T Specifically, the output temperature and output moisture content quality indicators are used as inputs to the encoder component. A BiLSTM network is then used to mine the bidirectional temporal features in the loose, rehydrated quality indicator data. The BiLSTM structure is as follows: Figure 6 As shown, a gated output method is adopted, namely an input gate, a forget gate, an output gate, and two time states (CellState and Hidden State);

[0064] The output values ​​at time t are i t ft o t c t and h t The specific solution method is as follows:

[0065]

[0066] Suppose the given input sequence is x = {x1, x2, ..., x...} t ,…,x T Let} represent time t, where t represents time t and T represents the total number of time intervals. Finally, the output result is obtained through the following formula.

[0067]

[0068] In the above formula: h t This represents the output of the hidden layer at time t; This represents the LSTM output at time t; σ is the Sigmids activation function; b α Denotes the deviation, where α∈{i,f,c,o,h}, W={W xi W hi W ci W xf W hf W cf W xo W ho W co W xc W hc W xh W hh} represents the weighted parameters obtained through time-reverse replay, W xh For example, it represents the weight matrix between the input layer and the hidden layer;

[0069] To comprehensively process the parameter feature information obtained from different components, the time-series correlation features of the loose re-moistening process parameters and the time-series features of the quality index data are first linked, and then the linking vector is... As input to a multi-head attention network, it can be used to redistribute weights on features, effectively avoiding information loss and achieving the fusion of multi-source feature information.

[0070] h = concat(h) enc +h dec )

[0071] d k =dim / h

[0072]

[0073]

[0074] Where dim represents the feature vector dimension; h represents the number of attention heads set;

[0075] Let Q represent the trainable weight matrix. h K h V h The problem vector, key vector, and value vector are obtained by multiplying them by the learning parameters, respectively.

[0076] The CTCN_A_B network model was trained using loosely refluxing training data to obtain model parameters. The model was then tested using test set data, and the test results for both the test set and the model were inversely normalized. The formula can be used to calculate the goodness of fit between the actual quality value and the predicted data:

[0077] MAE 出料含水率 =3.19, RMSE 出料含水率 =3.19, R2 出料含水率 =97.78%;

[0078] MAE 出料温度 =4.02, RMSE 出料温度 =7.55, R2 出料温度 =96.36%;

[0079] Depend on Figure 7 and Figure 8 It can be seen that the moisture content and discharge temperature of the loose rehydrated material predicted by the CTCN_A_B network model largely coincide with the actual values.

[0080] Step (3.2): To determine the superiority of the prediction results of the constructed CTCN_A_B network model, it is compared with common current workshop quality prediction models. The effectiveness of the constructed model is verified by comparing the fitted values, standard errors, and absolute percentage errors with traditional regression methods, shallow machine learning methods, and neural networks (DNN, LSTM, GRU, etc.), as shown in Table 1. Through comparison, in the test set, the prediction error of the CTCN_A_B network model proposed in this invention is much lower than that of common current prediction models.

[0081] Table 1 Comparison of results from different prediction models

[0082]

[0083] Step (4): In order to realize the online prediction capability of filament making, it is necessary to obtain online filament making process data, that is, to establish communication with the sensors of the online monitoring equipment in the filament making workshop to obtain online filament making workshop process data, and select the same loose re-moistening process data as the offline data: process flow rate, process hot air temperature, material cumulative amount, water addition flow rate, water addition cumulative amount and steam automatic valve opening, as well as the discharge moisture content and discharge temperature online data, and perform data outlier processing and normalization processing in step (2.2) on these online data;

[0084] Step (5): Utilizing normalized online silk-making workshop process data, online quality prediction of the silk-making workshop is achieved. A dual-thread method is adopted, using transfer learning to freeze the network structure and parameters of the feature extraction module in the offline prediction model and transfer them to the online process quality prediction model, thereby dynamically adjusting the prediction model to meet the actual production change prediction needs. Specific steps include:

[0085] Step (5.1): To cope with various emergencies that occur in actual production, two model platforms are established using the dual-thread method: a training platform and a prediction platform. The training platform is used to continuously train the model to achieve the purpose of real-time model updates, and the prediction platform is used to predict the quality of the silk-making workshop at future moments.

[0086] Step (5.2): Use the constructed CTCN_A_B network on the training platform to train the model on offline historical data, set the model prediction accuracy threshold, and save the model structure and parameters that reach the threshold.

[0087] Step (5.3): The training platform utilizes offline and online data acquired from the silk-making workshop under the same production process. The production information contained within these data is similar, allowing for interoperability of feature mining methods. In the prediction platform, the feature extraction network in the pre-trained model is frozen and transferred to the online prediction model. The fully connected layers of the model are randomly initialized to adapt to actual production changes, and the online prediction model for silk-making is rebuilt. As shown in Table 2, the accuracy of online prediction using transfer learning is above 98%, and the training time is significantly reduced, enabling timely response to industrial production prediction needs.

[0088] Table 2. Online prediction results using transfer learning

[0089]

[0090] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for online prediction of process quality in a silk-making workshop, characterized in that, Includes the following steps: Step (1): Obtain historical data of yarn-making process parameters from different production plants and production lines in the MES database, including process parameters and quality indicators for different processes; Step (2): Regularize and fuse the historical data of the obtained yarn-making process parameters, and perform preprocessing; Step (3): Select the loose moisture regain process data from the historical data of the silk making process after preprocessing in step (2), and establish an offline quality prediction model for the workshop, namely the CTCN_A_B network model, based on the sequence-to-sequence learning structure and using a hybrid neural network to predict the loose moisture regain quality index at the next moment. Step (4): Establish communication with the sensors of the online monitoring equipment in the yarn making workshop to obtain online yarn making workshop process data, select the loose re-moistening process data that is the same as the offline data and process it; Step (5): Using the dual-thread method, the network structure and parameters of the feature extraction module in the offline prediction model are frozen by transfer learning and transferred to the online process quality prediction model to realize dynamic adjustment of the prediction model and meet the prediction needs of actual production changes. Step (3) specifically includes the following steps: Step (3.1): Select the loose re-moistening process data from the complete yarn making process data, and analyze the process parameters that have an important impact on the quality indicators. Select these process parameter data as the input data for the offline prediction model and train the prediction model. Step (3.2): Validate the effectiveness of the constructed model by comparing the fitted values, standard errors, and absolute percentage errors with traditional regression methods, shallow machine learning methods, and neural networks such as DNN, LSTM, or GRU. Step (3.1) is as follows: Step (3.1.1): The key process parameters selected are: process flow rate, process hot air temperature, cumulative material amount, water flow rate, cumulative water amount and steam automatic valve opening. The quality indicators are: discharge temperature and discharge moisture content. These data are divided into training set and test set according to a certain ratio. The offline prediction model is trained using the training set and tested using the test set. Step (3.1.2): Concatenate the encoder components of CNN_TCN and train the preprocessed offline prediction model using the training set data; use the BiLSTM network to mine the bidirectional temporal features in the loose moisture quality index data, connect the temporal correlation features of the loose moisture process parameters and the temporal features of the quality index data, and then use the connection vector as the input of the multi-head attention network to redistribute the weights of the features to achieve the fusion of multi-source feature information; Step (3.1.2): Construct an offline prediction model for the process quality of the silk-making workshop based on a sequence-to-sequence learning structure. This model consists of two parts: an encoder component using a concatenated CNN_TCN and a BiLSTM decoder component incorporating a multi-head attention mechanism. In the encoder component, a convolutional neural network (CNN) is first used to automatically learn deep latent features in the process parameter data, effectively eliminating reliance on expert experience to obtain correlation information. Subsequently, a concatenated temporal convolutional neural network (TCN) is used to effectively analyze the temporal information of the workshop process while maintaining the causal convolutional characteristics of the process parameters. Furthermore, causal convolution is applied in the TCN network to improve the processing efficiency of long-span memory units, enabling the extraction of long-distance temporal features of process parameters. Simultaneously, in the decoder component, a BiLSTM network is used to extract historical temporal features from the quality indicator data. Then, a multi-head attention network is used to fuse all the extracted features and redistribute the weights of the features to achieve multi-source feature information fusion. Finally, the predicted quality index is output through a fully connected layer.

2. The online prediction method for process quality in the yarn-making workshop according to claim 1, characterized in that, Step (2) specifically includes the following steps: Step (2.1): By analyzing the process schemes, process combinations, data acquisition methods, and data naming similarities of each plant area, formulate the configuration content of the silk-making process to achieve the regularized representation and data fusion of different process data in different plants area; Step (2.2): For various abnormal data, data preprocessing is achieved through multiple processes such as data cleaning, data type conversion, data batch number collection, outlier handling, and data normalization.

3. The online prediction method for process quality in the yarn-making workshop according to claim 1, characterized in that, Step (5) specifically includes the following steps: Step (5.1): To cope with various emergencies that occur in actual production, two model platforms are established using the dual-thread method: a training platform and a prediction platform. The training platform is used to continuously train the model to achieve the purpose of real-time model updates, and the prediction platform is used to predict the quality of the silk-making workshop at future moments. Step (5.2): Use the constructed CTCN_A_B network on the training platform to train the model on offline historical data, set the model prediction accuracy threshold, and save the model structure and parameters that reach the threshold. Step (5.3): On the training platform, offline and online data obtained from the silk-making workshop under the same production process are used. The production information contained in them is similar, which makes the feature information mining methods interchangeable. On the prediction platform, the feature extraction network in the pre-trained model is frozen and transferred to the online prediction model. The fully connected layer of the model is randomly initialized to adapt to the actual production change requirements, and the silk-making online prediction model is rebuilt.

Images