A multi-quality index output prediction method and system for a loose rehydration process of silk production

Abstract

This invention discloses a method and system for predicting multiple quality index outputs in a loosening and rehydration process for silk reeling. The method includes: acquiring process parameters and quality indices for the loosening and rehydration process; acquiring historical data of the process parameters and quality indices for the loosening and rehydration process; preprocessing the historical data to obtain the preprocessed historical data and dividing it into training and testing sets; training a multi-output prediction model CB_A_L for the loosening and rehydration process using the training set; and testing the test set using the trained prediction model CB_A_L. This invention avoids the experience-dependent limitations of traditional correlation algorithms, autonomously and fully extracts the complex correlation time-series information contained in the process data, and achieves high prediction accuracy, meeting the needs of workshop quality prediction and enabling timely early warning of abnormal quality data.

Description

Technical Field

[0001] This invention relates to a method and system for predicting the output of multiple quality indicators in a loose and rehydrated silk-making process, belonging to the field of industrial workshop quality prediction technology. Background Technology

[0002] Emerging information technologies such as the Internet of Things (IoT), process big data, and artificial intelligence (AI) have laid a solid technological foundation for the development of intelligent process manufacturing workshops. However, for process industries, production processes consist of numerous steps, and the coupling relationships and technological mechanisms between these steps are extremely complex. Furthermore, the material, energy, and data information involved in the production process are often isolated and uncommunicated, leading to a lack of temporal analysis and correlation integration of process parameters in the workshop production process. This reduces the predictability, linkage, and autonomy of all elements in the production process, severely impacting the improvement of the workshop's intelligent production system. Therefore, researching accurate prediction methods for process manufacturing quality based on high-dimensional, multi-scale process data is a core issue supporting intelligent decision-making and optimized control of workshop production lines. Summary of the Invention

[0003] This invention provides a method and system for predicting multiple quality indicators in the loose and moisture-retaining process of silk making, in order to solve the technical problem of predicting multiple quality targets of loose and moisture-retaining during the silk making process.

[0004] The technical solution of this invention is: a method for predicting the output of multiple quality indicators in a loose and moisture-retaining silk-making process, comprising:

[0005] S1. Obtain the process parameters and quality indicators of the loose and re-moistened yarn processing technology;

[0006] S2. Obtain historical data on process parameters and quality indicators for the loosening and rehydration process of yarn making;

[0007] S3. Preprocess the historical data, obtain the preprocessed historical data, and divide it into training set and test set;

[0008] S4. Train the CB_A_L multi-output prediction model for loose re-moistening process quality using the training set;

[0009] S5. Test the test set using the trained prediction model CB_A_L.

[0010] S1 specifically refers to: determining process parameters and quality indicators based on the specific process flow of loose re-moistening in the yarn-making workshop; wherein, there are multiple process parameters and quality indicators.

[0011] The process parameters are adjustable, including process flow rate, opening degree of automatic gas-water mixing valve, water addition ratio, water addition flow rate, cumulative water addition amount, cumulative material amount, opening degree of automatic steam valve, and process hot air temperature.

[0012] The quality indicators include: discharge temperature and discharge moisture content.

[0013] The acquisition of historical data on process parameters and quality indicators for the loose and rehydrated silk refining process includes: collecting different sensor data information by production batch number, connecting brand, model name, process name, process parameter name, indicator type, and data acquisition start and end time; and then fusing these sensor data according to batch number, process name, process parameter, quality indicator, and acquisition time to form historical data on process parameters and quality indicators for the loose and rehydrated silk refining process.

[0014] The preprocessing specifically involves: performing outlier processing and normalization on historical data of process parameters and quality indicators for the loosening and rehydration process of yarn.

[0015] The loose rewetting process quality multi-output prediction model CB_A_L includes: an encoder based on CNN-BiLSTM feature learning and a decoder based on a temporal attention network;

[0016] The encoder based on CNN-BiLSTM feature learning includes a convolutional neural network and a BiLSTM. A one-dimensional process parameter time series matrix constructed from historical process parameter data of loose rehumidification process is used as the input of the encoder. First, the convolutional neural network is used to construct a two-dimensional time series feature information matrix from the one-dimensional process parameter time series matrix. Then, the time features in the extracted two-dimensional time series feature information matrix are learned by BiLSTM and defined as a fixed-length vector to output the process parameter associated time series feature sequence.

[0017] The decoder based on the temporal attention network includes a temporal attention network, an LSTM network, and a fully connected network. Historical data of quality indicators from the loosely woven refluxing process are used as input to the LSTM network in the decoder to extract temporal information of the quality indicators, i.e., the hidden states of the LSTM network. The temporal feature sequence associated with process parameters obtained from the encoder learned by CNN-BiLSTM is used as input to the temporal attention network, and the attention weights for all time steps at time t are calculated iteratively. The softmax function is used to normalize all the calculated attention weights, refining the expression of the influence of the encoder's process parameter-associated temporal feature sequence on the prediction target at each time step. The obtained influence and the encoder's process parameter-associated temporal feature sequence are weighted and summed to synthesize the information from all encoder time steps, resulting in the intermediate semantic vector at time t. After T iterations of calculation in the decoder LSTM network, a single-layer fully connected network is used to integrate the hidden states of the LSTM network with the intermediate semantic vector output by the temporal attention network, and finally, the predicted process quality value is obtained after dimensionality transformation.

[0018] According to another aspect of the present invention, a multi-quality index output prediction system for a loose and re-moistening process of yarn making is also provided, comprising:

[0019] The first acquisition module is used to acquire the process parameters and quality indicators of the loosening and rehydration process of yarn making;

[0020] The second acquisition module is used to acquire historical data on process parameters and quality indicators of the loosening and rehydration process of yarn making;

[0021] The partitioning module is used to preprocess historical data, obtain the preprocessed historical data, and partition it into training and test sets.

[0022] The training module is used to train the CB_A_L multi-output prediction model for loose re-moistening process quality using the training set;

[0023] The testing module is used to test the test set using the trained prediction model CB_A_L.

[0024] According to another aspect of the present invention, a terminal device is also provided, including a memory, a processor, and a program stored in the memory and executable by the processor, wherein the processor executes the program to implement the multi-quality index output prediction method for the loose and re-moisturized yarn process described in any one of the above.

[0025] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium comprising a stored program, wherein, when the program is executed, the device on which the computer-readable storage medium is located executes the multi-quality index output prediction method for the loosening and moisture regaining process of the filament as described in any one of the above.

[0026] The beneficial effects of this invention are as follows: First, traditional mathematical statistics methods rely on expert-preset thresholds, leading to a 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 effectively solve the complex problems of multi-distributed parameters, high-dimensional multi-scale, and strong coupling in process manufacturing workshops. Consequently, it cannot fully extract the complex temporal correlation information in industrial production, losing local information representation, resulting in unsatisfactory prediction accuracy. This invention, however, first analyzes the loose moisture regaining process of yarn processing to determine adjustable parameters and quality indicators. Then, it fuses these parameters and summarizes them into a data table. The process parameter data and historical quality indicator data are used as inputs to the encoder and decoder in the model, respectively. Based on a sequence-to-sequence learning structure, and utilizing BiLSTM, CNN, and temporal attention mechanisms, a loose moisture regaining prediction model is constructed. This avoids the experience-dependent limitations of traditional correlation algorithms, autonomously and fully extracting the complex temporal correlation information contained in the process data, and achieving high prediction accuracy. This fulfills the workshop quality prediction needs and enables timely warnings for abnormal quality data. Attached Figure Description

[0027] Figure 1 A schematic diagram of the multi-output prediction model for the quality of the loose and re-moistening process of silk fabrication constructed in this invention;

[0028] Figure 2 This is a diagram of the BiLSTM network structure of the present invention;

[0029] Figure 3 This is a diagram showing the relationship between the loose rehydration process parameters of the present invention;

[0030] Figure 4 The experimental comparison and analysis of the discharge moisture content of the CNN_BiLSTM model and the CB_A_L model of this invention is shown in the figure.

[0031] Figure 5 The discharge temperature is compared and analyzed using experimental data from the CNN_BiLSTM model and the CB_A_L model of this invention. Detailed Implementation

[0032] The invention will be further described below with reference to the accompanying drawings and embodiments, but the scope of the invention is not limited to the description.

[0033] Example 1: As Figure 1-5 As shown, a method for predicting the output of multiple quality indicators in a loosening and rehydration process for silk production includes:

[0034] S1. Obtain the process parameters and quality indicators of the loose and re-moistened yarn processing technology;

[0035] S2. Obtain historical data on process parameters and quality indicators for the loosening and rehydration process of yarn making;

[0036] S3. Preprocess the historical data, obtain the preprocessed historical data, and divide it into training set and test set;

[0037] S4. Train the CB_A_L multi-output prediction model for loose re-moistening process quality using the training set;

[0038] S5. Test the test set using the trained prediction model CB_A_L; or you can use the trained model to predict the data to be predicted.

[0039] Optionally, S1 specifically involves: determining process parameters and quality indicators based on the specific process flow of loose re-moistening in the yarn-making workshop; wherein, the number of process parameters and quality indicators is multiple.

[0040] Optionally, the process parameters are adjustable parameters, including process flow rate, opening degree of automatic gas-water mixing valve, water addition ratio, water addition flow rate, cumulative water addition amount, cumulative material amount, opening degree of automatic steam valve, and process hot air temperature;

[0041] Optionally, the quality indicators include: discharge temperature and discharge moisture content.

[0042] Optionally, the acquisition of historical data on process parameters and quality indicators of the loose re-moistening process of silk production includes: collecting different sensor data information by production batch number, connecting brand, model name, process name, process parameter name, indicator type, and data acquisition start and end time; and then fusing these sensor data according to batch number, process name, process parameter, quality indicator, and acquisition time to form historical data on process parameters and quality indicators of the loose re-moistening process of silk production.

[0043] Optionally, the preprocessing specifically involves: performing outlier handling and normalization on historical data of process parameters and quality indicators for the loose rehydration process of yarn. Outlier handling involves filtering each historical data point for null or zero values; if a null or zero value exists, that historical data point is deleted. This processing can remove outliers caused by equipment problems, batch production stoppages leading to material shortages, or material head and tail issues.

[0044] Optionally, the loose rewetting process quality multi-output prediction model CB_A_L includes: an encoder based on CNN-BiLSTM feature learning and a decoder based on a temporal attention network;

[0045] The encoder based on CNN-BiLSTM feature learning includes a convolutional neural network and a BiLSTM. A one-dimensional process parameter time series matrix constructed from historical process parameter data of loose rehumidification process is used as the input of the encoder. First, the convolutional neural network is used to construct a two-dimensional time series feature information matrix from the one-dimensional process parameter time series matrix. Then, the time features in the extracted two-dimensional time series feature information matrix are learned by BiLSTM and defined as a fixed-length vector to output the process parameter associated time series feature sequence.

[0046] The decoder based on the temporal attention network includes a temporal attention network, an LSTM network, and a fully connected network. Historical data of quality indicators from the loosely woven refluxing process are used as input to the LSTM network in the decoder to extract temporal information of the quality indicators, i.e., the hidden states of the LSTM network. The temporal feature sequence associated with process parameters obtained from the encoder learned by CNN-BiLSTM is used as input to the temporal attention network, and the attention weights for all time steps at time t are calculated iteratively. The softmax function is used to normalize all the calculated attention weights, refining the expression of the influence of the encoder's process parameter-associated temporal feature sequence on the prediction target at each time step. The obtained influence and the encoder's process parameter-associated temporal feature sequence are weighted and summed to synthesize the information from all encoder time steps, resulting in the intermediate semantic vector at time t. After T iterations of calculation in the decoder LSTM network, a single-layer fully connected network is used to integrate the hidden states of the LSTM network with the intermediate semantic vector output by the temporal attention network, and finally, the predicted process quality value is obtained after dimensionality transformation.

[0047] The following is a detailed description of an optional embodiment of the present invention.

[0048] Step (1): Analyze the specific process flow of loose re-moistening in the yarn making workshop, determine the process parameters and quality indicators, and obtain the historical data of the above-mentioned process parameters from the workshop MES database;

[0049] Step (2): Due to equipment problems and batch production, there are abnormal values ​​caused by machine stoppages and material shortages, as well as inconsistent magnitudes of various process parameters. Therefore, it is necessary to preprocess the collected loose and rehydrated process historical data.

[0050] Step (3): Analyze the characteristics of loose rehydration parameters, and establish a multi-output prediction model CB_A_L for loose rehydration process quality based on the internal characteristic information;

[0051] Step (4): Train the constructed prediction model CB_A_L using the preprocessed historical process data, and compare the training results with the current process quality prediction algorithm to verify the practicality and effectiveness of the proposed algorithm.

[0052] Step (1) includes the following steps:

[0053] Step (1.1): Analyze the loose re-moistening process flow, which is completed in the re-moistening machine. It consists of a feeding device, roller, steam system, water supply system, compressed air system, circulating air system, dehumidification system, and discharge device. The internal processing parameters can be divided into three categories: static parameters, process parameters, and quality indicators.

[0054] Step (1.2): Static parameters are preset parameters before the entire equipment processing and remain fixed; process parameters are divided into adjustable parameters and non-adjustable parameters. The adjustable parameters are process flow rate, automatic gas-water mixing valve opening, water addition ratio, water addition flow rate, cumulative water addition, cumulative material addition, automatic steam valve opening, and process hot air temperature (8 parameters); the quality indicators are discharge temperature and discharge moisture content, as shown in Table 1. To achieve advance prediction of quality indicators, it is necessary to closely monitor the relationship between adjustable parameters and quality indicators; this invention analyzes the processing characteristics of each process parameter in the actual production process of loose rehydration, removes non-adjustable parameters, and selects adjustable parameters and quality indicator data; on the one hand, it directly selects adjustable parameters for subsequent optimization work without considering non-adjustable parameters; on the other hand, it reduces the calculation work without affecting the prediction effect.

[0055] Table 1 Performance Parameters of Loosening and Rehydration Equipment

[0056]

[0057] Step (1.3): Obtain the raw data of the loose re-moistening process from the MES database of the yarn processing workshop. Since different process data are collected by different sensors, it is necessary to fuse the process data to form a data table that can be used directly. Specifically, according to the loose re-moistening processing process, as shown in Table 2, different process data are collected from different sensors by batch number, connecting the brand, model name, process name, process parameter name, index type, and data collection start and end time. Then, these sensor data are fused according to batch number, process name, and collection time to form historical data of process parameters and quality indicators of the loose re-moistening process of yarn processing, as shown in Table 3.

[0058] Table 2

[0059] Batch number Brand Model Name Process Name Process parameter name Indicator Type Start time End time 70 Cloud and Smoke (Purple) Module Two Data Model 0 Loose and damp Loose re-moistening process flow steady-state indicators 2021-01-26 22:50:18 70 Cloud and Smoke (Purple) Module Two Data Model 0 Loose and damp Loose, rehydrated discharge moisture content steady-state indicators 2021-01-26 21:59:54 2021-01-26 22:54:30 70 Cloud and Smoke (Purple) Module Two Data Model 0 Loose and damp Loose, rehydrated discharge temperature steady-state indicators 2021-01-26 21:59:54 2021-01-26 22:54:30 70 Cloud and Smoke (Purple) Module Two Data Model 0 Loose and damp Loosening and rehydration process hot air temperature steady-state indicators 2021-01-26 21:59:54 2021-01-26 22:50:18 71 Cloud and Smoke (Purple) Module Two Data Model 0 Loose back to the lake Loose re-moistening process flow steady-state indicators 2021-01-27 17:19:48 2021-01-27 18:13:48 71 Cloud and Smoke (Purple) Module Two Data Model 0 Loose and damp Loose, rehydrated discharge moisture content steady-state indicators 2021-01-27 17:23:24 2021-01-27 18:18:00 71 Cloud and Smoke (Purple) Module Two Data Model 0 Loose and damp Loose, rehydrated discharge temperature steady-state indicators 2021-01-27 17:23:24 2021-01-27 18:18:00 71 Cloud and Smoke (Purple) Module Two Data Model 0 Loose and damp Loosening and rehydration process hot air temperature steady-state indicators 2021-01-27 17:23:24 2021-01-27 18:13:48 73 Cloud and Smoke (Purple) Module Two Data Model 0 Loose and damp Loose re-moistening process flow steady-state indicators 2021-01-27 20:48:36 73 Cloud and Smoke (Purple) Module Two Data Model 0 Loose and damp Loose, rehydrated discharge moisture content steady-state indicators 2021-01-27 20:52:48 73 Cloud and Smoke (Purple) Module Two Data Model 0 Loose and damp Loose, rehydrated discharge temperature steady-state indicators 2021-01-27 20:52:48 73 Cloud and Smoke (Purple) Module Two Data Model 0 Loose and damp Loosening and rehydration process hot air temperature steady-state indicators 2021-01-27 20:48:36 75 Cloud and Smoke (Purple) Module Two Data Model 0 Loose and damp Loose re-moistening process flow steady-state indicators 2021-01-28 16:54:18 2021-01-28 17:48:30 75 Cloud and Smoke (Purple) Module Two Data Model 0 Loose and damp Loose, rehydrated discharge moisture content steady-state indicators 2021-01-28 16:58:00 2021-01-28 17:52:42

[0060] Table 3

[0061]

[0062] Step (2) includes the following steps:

[0063] After data fusion and processing in step (1.3) above, loose and rehydrated data produced from January 2 to August 25, 2020 were obtained. Nearly 100,000 historical data were collected. These data contained outliers caused by equipment problems and batch production shutdowns and material shortages, and these outliers needed to be deleted. At the same time, the order of magnitude of the process data was not uniform, so normalized data was applied to process the data and control all data within the range of [0, 1], as shown in equation (1).

[0064]

[0065] Step (3) includes the following steps:

[0066] The loose re-humidification normalized data obtained in step (2) is divided into a 7:3 ratio as the training set and the test set, respectively. The training set data is used as the input data for the prediction model to construct a prediction model targeting the loose re-humidification quality index. This model is based on a sequence-to-sequence learning structure, simultaneously learning the temporal features associated with workshop process parameters and the temporal information of quality indicators. In the encoder, CNN is used to autonomously mine the correlation between key processes in loose re-humidification, and then BiLSTM is cascaded to extract bidirectional temporal information to improve the process feature information. On this basis, a temporal attention mechanism is used to focus on important historical moments of the process to reduce redundant information. In the decoder, LSTM is used to extract the temporal information of quality indicators. All extracted feature information is effectively fused and input into a fully connected network. Finally, a complete prediction model is constructed with the prediction of the process quality at the next moment as the modeling objective.

[0067] An encoder based on CNN-BiLSTM feature learning: This addresses the complex temporal correlations within the process flow of a workflow workshop, using loosely packed process parameters X = (x1, x2, ..., x...). T )=(x 1 ,x 2 ,…,x N ) T As input data to the encoder, a convolutional neural network is first used to construct a two-dimensional temporal feature information matrix from the one-dimensional process time-series matrix according to certain rules. Then, a BiLSTM is used to learn the extracted temporal features from the matrix and define them as fixed-length vectors. BiLSTM is an extended structure evolved from Long Short-Term Memory networks. Its core idea is to add an additional LSTM layer on top of the LSTM to process the data in reverse. The specific structure is as follows: Figure 2As shown, BiLSTM uses two independent LSTMs to mine process timing information from both the front and back ends, and can completely capture the bidirectional time structure information in multi-sequence data. The lower layer is a forward LSTM, and the upper layer is a backward LSTM, both using the same computation process.

[0068] LSTM employs a gated output method, namely an input gate, a forget gate, an output gate, and two time states (CellState and Hidden State). The output value at time t is i. t f t o t c t and h t The specific solution method of LSTM is as follows:

[0069]

[0070] Suppose the given input sequence is x = {x1, x2, x3, ..., 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.

[0071]

[0072] 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 backpropagation over time, such as the weight matrix between the input layer and the hidden layer.

[0073] BiLSTM obtains the final output value by concatenating the outputs of two LSTM layers, as follows:

[0074]

[0075] in: and These are the outputs of the forward and backward LSTMs, respectively, with the completely reverse sequence order as input; α and β are sequence control factors for the forward and backward LSTMs (α+β=1); This represents the hidden state output value of the BiLSTM at time t, i.e., the time-series feature sequence associated with process parameters; This represents the overall output value of the BiLSTM.

[0076] A decoder based on temporal attention mechanism addresses the issue that in encoder-decoder networks, as the length of the input process parameter sequence increases, the acquired deep features may be continuously covered by subsequent features, resulting in the loss of some key information. To address this problem, a temporal attention mechanism is introduced that automatically focuses on key region information and weakens the representation of other redundant regions. This mechanism analyzes and determines the hidden state sequence associated with each time step of the encoder output. For predicting target process quality The degree of influence. The process parameter-related time-series feature sequence obtained above is used as the input to the time-series attention network, and the quality index sequence Y = (y1, y2, ..., y) in loose reabsorption is used. T )∈R T As input to the decoder LSTM network, it is used to extract the temporal information d of the quality index. T-1 As shown in equation (5), the attention weights for all time steps at the current time are calculated iteratively. The calculated attention weights are normalized using the softmax function, as shown in Equation (6), to refine the hidden state sequence of the encoder association time sequence at each time step. The degree of impact on the prediction target

[0077]

[0078]

[0079] In the formula: Let d be the hidden state d of the decoder LSTM network at the previous time step. t-1 and gate information Fusion expression; V d ∈R q W d ∈R q×(2q) and u d ∈R p×p represents the weights and bias parameters of the attention mechanism network; p represents the number of hidden units in the BiLSTM network of the encoder.

[0080] The degree of influence obtained and all hidden state information of the encoder A weighted summation is performed to synthesize the information from all time points of the encoder, as shown in Equation (7), to obtain the intermediate semantic vector c. t ∈R 1×p This vector comprehensively expresses the time-series information of the correlation of process parameters and the time-series information of historical quality indicators.

[0081]

[0082] Equation (8) is used to integrate the hidden state d after the decoder LSTM network has undergone T iterations of computation. T-1 The intermediate semantic vector c output by the temporal attention mechanism T Finally, the predicted process quality value is obtained through dimensional transformation. As shown in equation (9).

[0083] d T =f LSTM (d T-1 ,y T )∈R 1×q (8)

[0084]

[0085] In the formula: W y ∈R q×(p+q) b w ∈R q V y ∈R q b y ∈R represent the weights and biases of the fully connected network and the network under dimensional transformation, respectively.

[0086] Step (4) includes the following steps:

[0087] The parameters of the CB_A_L model for predicting the quality of loose re-humidification processes can be specifically set as follows: time step s = 10, i.e., using historical loose re-humidification process data from the previous minute to predict the quality index data for the next moment; number of iterations = 100; learning rate = 0.001; in the model structure, the BiLSTM and LSTM networks have the same network layer L = 3; the number of hidden layers are p = 8 and q = 10, respectively. Experiments have verified that the optimal CNN convolutional kernel size in the CB_A_L model is 512, at which point the prediction accuracy reaches its best state. To further verify the model's predictive performance, as shown in Table 4, the CB_A_L model is compared with 11 other models: traditional regression methods (linear regression and multinomial regression), support vector regression (LINEAR, POLY, and RBF), DNN, recurrent neural networks (RNN, LSTM, and GRU), Seq2Seq, and CNN_BiLSTM. The deep learning model parameter configurations all refer to the model settings in this paper, with the default neural network layer number set to 1. Finally, the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are used as model evaluation and analysis metrics. Plots are generated using partial test set data from the optimal results of the CNN_BiLSTM and CB_A_L models compared above, as shown below. Figure 4 , Figure 5 As shown.

[0088] Table 4. Comparison of prediction results of different models

[0089]

[0090] Due to the complexity of the loose re-moistening process, numerous quality-influencing factors, and high degree of process coupling, it is difficult to characterize the correlation between parameters in detail. Simultaneously, production materials flow along different processes and production flows, and the internal material flow, information flow, and energy flow change in an orderly manner during this flow. The equipment states at different times influence each other, resulting in significant temporal characteristics in the acquired process data. Therefore, it is crucial to analyze the correlation and temporal information between process parameters, while also grasping the temporal information within the quality indicator data. The CB_A_L model constructed in this invention can solve the aforementioned difficulties. Comparison with test sets demonstrates that the CB_A_L model has significant superiority in predicting multiple quality outputs in the loose re-moistening process of yarn making. The actual values ​​of quality indicators and the predicted values ​​output by the model fit each other to the greatest extent, and the evaluation and analysis indicators of the CB_A_L model are also much smaller than those of other comparative models, proving the predictive stability of the CB_A_L model.

[0091] According to another aspect of the present invention, a multi-quality index output prediction system for a loose and re-moistening process in yarn making is also provided, comprising:

[0092] The first acquisition module is used to acquire the process parameters and quality indicators of the loosening and rehydration process of yarn making;

[0093] The second acquisition module is used to acquire historical data on process parameters and quality indicators of the loosening and rehydration process of yarn making;

[0094] The partitioning module is used to preprocess historical data, obtain the preprocessed historical data, and partition it into training and test sets.

[0095] The training module is used to train the CB_A_L multi-output prediction model for loose re-moistening process quality using the training set;

[0096] The testing module is used to test the test set using the trained prediction model CB_A_L.

[0097] It should be noted that the first acquisition module, the second acquisition module, the partitioning module, the training module, and the testing module mentioned above correspond to steps S1 to S5 in Embodiment 1. The examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1.

[0098] According to another aspect of the present invention, a terminal device is also provided, including a memory, a processor, and a program stored in the memory and executable by the processor, wherein the processor executes the program to implement the multi-quality index output prediction method for the loose and moistened yarn process described in any one of the above embodiments.

[0099] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the multi-quality index output prediction method for the loosening and moisture regaining process of the yarn as described in any one of the above.

[0100] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A multi-quality indicator output prediction method for a loose bale reconditioning process, characterized by, The application relates to a method for predicting the quality of a loose humidification process in a tobacco making process. S1, acquiring process parameters and quality indexes of a loose humidification process in a tobacco making process; S2, acquiring historical data of process parameters and quality indexes of the loose humidification process in the tobacco making process; S3, preprocessing the historical data, acquiring preprocessed historical data, and dividing a training set and a test set; S4, training a loose humidification process quality multi-output prediction model CB_A_L by using the training set; S5, testing the test set by using the trained prediction model CB_A_L; The S1 is specifically: according to the specific process flow of the loose humidification in a tobacco making workshop, determining process parameters and quality indexes; wherein the number of the process parameters and the quality indexes is multiple; The process parameters are adjustable parameters, including process flow, air-water mixed automatic valve opening, water adding ratio, water adding flow, water adding cumulative amount, material cumulative amount, steam automatic valve opening, and process hot air temperature; The historical data of the process parameters and the quality indexes of the loose humidification process in the tobacco making process are acquired by: according to production batch numbers, connecting brand names, model names, process names, process parameter names, index types, and data collection start and end times, collecting different sensor data information, and then fusing the sensor data according to batch numbers, process names, process parameters, quality indexes, and collection times to form the historical data of the process parameters and the quality indexes of the loose humidification process in the tobacco making process; The loose humidification process quality multi-output prediction model CB_A_L comprises an encoder based on CNN-BiLSTM feature learning and a decoder based on a time sequence attention network; The encoder based on CNN-BiLSTM feature learning comprises a convolutional neural network and a BiLSTM; a one-dimensional process parameter time sequence matrix constructed by the historical data of the process parameters of the loose humidification process is used as the input of the encoder; the one-dimensional process parameter time sequence matrix is first constructed into a two-dimensional time sequence feature information matrix by using the convolutional neural network, then time features in the two-dimensional time sequence feature information matrix are learned and extracted by the BiLSTM and defined into fixed-length vectors, and a process parameter correlation time sequence feature sequence is output. The decoder based on the time sequence attention network comprises a time sequence attention network, an LSTM network and a full connection network; historical data of quality indexes of the loose conditioning process is taken as input of the LSTM network in the decoder network, so as to extract time sequence information of the quality indexes, that is, hidden layer states of the LSTM network; and time sequence feature sequences of the process parameters obtained by the encoder of the CNN-BiLSTM feature learning are taken as input of the time sequence attention network, and attention weights of all time steps at t time are calculated through iteration; the calculated attention weights are normalized by using a softmax function, so as to refine the influence degree of the encoder process parameter time sequence feature sequence at each time on the prediction target at the time step; the influence degree and the encoder process parameter time sequence feature sequence are used for weighted summation to comprehensively encode all time information of the encoder, so as to obtain an intermediate semantic vector at t time; after the LSTM network in the decoder is calculated through T times of iteration, the hidden layer states of the LSTM network and the intermediate semantic vector output by the time sequence attention network are integrated by using a single-layer full connection network, and finally a process quality prediction value is obtained through dimension transformation.

2. The multi-quality indicator output prediction method of loose re-drying process of lap preparation according to claim 1, wherein, The quality indexes comprise a discharge temperature and a discharge moisture content.

3. The multi-quality indicator output prediction method of loose re-drying process of lap preparation according to claim 1, wherein, The preprocessing specifically comprises: performing outlier processing and normalization processing on historical data of process parameters and quality indexes of the loose conditioning process.

4. A multi-quality index output prediction system for a loose and re-moistening process of silk reeling, characterized in that, Comprise: The first acquisition module is configured to acquire process parameters and quality indexes of the loose conditioning process. The second acquisition module is configured to acquire historical data of process parameters and quality indexes of the loose conditioning process. The division module is configured to preprocess the historical data, acquire preprocessed historical data, and divide the preprocessed historical data into a training set and a test set. The training module is configured to train the loose conditioning process quality multi-output prediction model CB_A_L by using the training set. The test module is configured to test the test set by using the trained prediction model CB_A_L. The S1 specifically comprises: determining process parameters and quality indexes according to a specific process flow of the loose conditioning in the cut tobacco workshop; wherein the number of the process parameters and the quality indexes is multiple. The process parameters are adjustable parameters, and comprise process flow, air-water mixed automatic valve opening, water adding ratio, water adding flow, water adding cumulative amount, material cumulative amount, steam automatic valve opening and process hot air temperature. The historical data of the process parameters and the quality indexes of the loose conditioning process of cut tobacco are acquired by: collecting different sensor data information according to production batch numbers, model names, process names, process parameter names, index types and data collection start and end times, and then fusing the sensor data according to the batch numbers, the process names, the process parameters, the quality indexes and the collection times, so as to form the historical data of the process parameters and the quality indexes of the loose conditioning process of cut tobacco. The loose conditioning process quality multi-output prediction model CB_A_L comprises an encoder based on CNN-BiLSTM feature learning and a decoder based on a time sequence attention network. The encoder based on CNN-BiLSTM feature learning includes a convolutional neural network and a BiLSTM. A one-dimensional process parameter time series matrix constructed from historical process parameter data of loose rehumidification process is used as the input of the encoder. First, the convolutional neural network is used to construct a two-dimensional time series feature information matrix from the one-dimensional process parameter time series matrix. Then, the time features in the extracted two-dimensional time series feature information matrix are learned by BiLSTM and defined as a fixed-length vector to output the process parameter associated time series feature sequence. The decoder based on the temporal attention network includes a temporal attention network, an LSTM network, and a fully connected network. Historical data of quality indicators from the loosely woven refluxing process are used as input to the LSTM network in the decoder to extract temporal information of the quality indicators, i.e., the hidden states of the LSTM network. The temporal feature sequence associated with process parameters obtained from the encoder learned by CNN-BiLSTM is used as input to the temporal attention network, and the attention weights for all time steps at time t are calculated iteratively. The softmax function is used to normalize all the calculated attention weights, refining the expression of the influence of the encoder's process parameter-associated temporal feature sequence on the prediction target at each time step. The obtained influence and the encoder's process parameter-associated temporal feature sequence are weighted and summed to synthesize the information from all encoder time steps, resulting in the intermediate semantic vector at time t. After T iterations of calculation in the decoder LSTM network, a single-layer fully connected network is used to integrate the hidden states of the LSTM network with the intermediate semantic vector output by the temporal attention network, and finally, the predicted process quality value is obtained after dimensionality transformation.

5. A terminal device, characterized by: The method includes a memory, a processor, and a program stored in the memory and executable by the processor, wherein the processor executes the program to implement the multi-quality index output prediction method for the loose and re-moisturized yarn process as described in any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the multi-quality index output prediction method for the loosening and moisture regaining process of filament as described in any one of claims 1-3.

Images