A lyocell short fiber mass prediction method and system

Abstract

The application relates to the technical field of intelligent manufacturing of fiber production processes, in particular to a lyocell short fiber quality prediction method and system.The method comprises the following steps: acquiring a monitoring data set formed by process parameters, environmental condition data and finished product quality evaluation index data in a lyocell short fiber production process; performing cleaning and complementing on the monitoring data set to obtain an optimized data set; performing correction on the optimized data set to obtain a corrected data set; determining key features based on the corrected data set, and screening the corrected data set by using the key features to obtain a training sample; constructing a first quality index prediction model by using the training sample, and predicting the quality of the lyocell short fiber by using the first quality index prediction model.The data quality is improved through cleaning, complementing and key feature extraction processing, and then the quality prediction is realized by using a quality index prediction model, so that the efficiency and accuracy of lyocell short fiber quality prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing of fiber production process, and particularly relates to a lyocell short fiber quality prediction method and system. BACKGROUND

[0002] Lyocell fiber is recognized as the "green fiber" of the 21st century in the world, which uses natural plant fiber as raw material and has multiple excellent properties of natural fiber and synthetic fiber. The production process uses cellulose / N-methyl morpholine-N-oxide (NMMO) / water solution as the spinning dope, and is prepared by dry-jet wet spinning technology. In the whole production process of lyocell short fiber, the detection and judgment of product quality play a crucial role in production control, cost control and product quality guarantee.

[0003] At present, the quality detection of lyocell short fiber in the industry is still mainly based on traditional methods, which have the following significant limitations: first, the subjectivity is strong and the labor cost is high, which depends on manual visual inspection, manual defect picking and manual length pulling, etc. Not only a large amount of labor is consumed, but also the judgment error is caused by the fatigue of the detection personnel and the experience difference; second, the detection efficiency is low, which involves complicated chemical analysis steps such as Soxhlet extraction for oil content and viscosity method for polymerization degree, and the operation period is long, which seriously affects the production turnover efficiency; third, the hardware cost is high, and the single fiber strength instrument and muffle furnace are not only high in purchase cost, but also have strict requirements on environmental conditions such as temperature and humidity, and the daily maintenance cost is further increased; fourth, the quality control has a lag, and the traditional detection is in offline mode, which cannot feedback the quality fluctuation in the production process in real time, so that it is difficult to quickly adjust when the process is abnormal, which directly affects the overall production quality and efficiency.

[0004] At the same time, there is no systematic scheme for the quality prediction method of lyocell short fiber. Since the dry-jet wet spinning process involves complex procedures, and each procedure is affected by multiple factors such as spinning solution characteristics, equipment parameters and environmental conditions, the correlation between quality and production parameters presents strong nonlinearity and time sequence dependence, which further increases the difficulty of quality prediction.

[0005] The defects of the above traditional detection method and the blank of the quality prediction technology make it difficult for the existing quality control mode to meet the needs of modern textile industry for efficient, accurate and real-time monitoring. Therefore, it is urgent to build a quality prediction scheme based on production data to break through the technical bottleneck and improve the intelligent quality control level of lyocell short fiber production. SUMMARY

[0006] In view of the defects in the prior art, the present application provides a lyocell short fiber quality prediction method and system, which solves the problems of low efficiency, poor accuracy and difficulty in realizing real-time quality control in the lyocell short fiber quality detection in the prior art.

[0007] In order to achieve the above-mentioned purpose, one aspect of the present application provides a lyocell short fiber quality prediction method, which comprises: acquiring a monitoring data set formed by process parameters, environmental condition data and finished product quality evaluation index data in the lyocell short fiber production process; cleaning and completing the monitoring data set to obtain an optimized data set; correcting the optimized data set to obtain a corrected data set; determining key features based on the corrected data set, and screening the corrected data set using the key features to obtain a training sample; constructing a first quality index prediction model using the training sample, and predicting the quality of lyocell short fibers using the first quality index prediction model.

[0008] The present application breaks through the limitation of single dimension of traditional detection data by comprehensively collecting process parameters, environmental conditions and finished product quality evaluation indexes in the lyocell short fiber production to form a monitoring data set. After cleaning and completing, the missing and abnormal data are effectively removed, laying a high-quality data foundation for subsequent analysis; and then the core parameters are optimized through data correction, so that they are more consistent with the physical law of the process and the noise interference is reduced. The key feature screening focuses on the core influencing factors, improves the model pertinence and running efficiency, and the finally constructed prediction model realizes accurate quality prediction, which not only realizes real-time quality control of lyocell short fiber quality prediction, but also improves the efficiency and accuracy.

[0009] Optionally, the cleaning and completion of the monitoring data set to obtain the optimized data set comprises: calculating a data missing rate according to the monitoring data set, and screening the monitoring data set according to the data missing rate to obtain a first data set; constructing a data reconstruction model, and reconstructing the first data set using the data reconstruction model to obtain a second data set; calculating a reconstruction error of the second data set; deleting abnormal data in the second data set according to the reconstruction error to obtain the optimized data set.

[0010] The present application can accurately remove data with too much missing, difficult to complete and easy to be distorted by calculating the data missing rate and screening the first data set, reduces invalid information interference from the source, reconstructs the first data set by constructing a data reconstruction model, can effectively complete the reasonable missing values, so that the data more completely reflects the production process state, finally calculates the reconstruction error and deletes the abnormal data, which can further identify and remove the abnormal values deviating from the normal law, reduces the influence of data noise on subsequent analysis, and improves the accuracy of the monitoring data.

[0011] Optionally, the data reconstruction model satisfies the following formula:

[0012]

[0013] wherein, is a loss function of the data reconstruction model, is a total number of time steps, is a number of sensors, is a time point is a measurement value of the i-th sensor at the time point, is a predicted value of the i-th sensor at the time point, is a time point is a measurement value of the i-th sensor at the time point, is a predicted value of the i-th sensor at the time point, is an element in a mask matrix of the i-th sensor at the time point. The data reconstruction model of the application can accurately quantify the data reconstruction error through the loss function, constrain the difference between the measurement value and the predicted value by using the mask matrix, and complete the missing part in the monitoring data of the lyocell short fiber production in a targeted manner, thereby improving the performance of the data reconstruction model.

[0014] Optionally, the reconstruction error satisfies the following formula:

[0015]

[0016]

[0017] wherein, is a reconstruction error of a time series sample in the second data set, is a total number of time steps, is a number of sensors, is a true value of the i-th sensor at the time point, is a time point is a true value of the i-th sensor at the time point, is a reconstruction value of the i-th sensor at the time point. The reconstruction error formula of the application can accurately quantify the reconstruction quality of the second data set by calculating the average of the absolute difference between the true value and the reconstruction value in the time step and the sensor dimension, provide an objective index for judging the data reconstruction effect, and effectively identify and screen out abnormal data with poor reconstruction effect, thereby improving the data quality of the second data set. Optionally, the method further comprises: modifying the optimization data set to obtain a modified data set; and obtaining the modified data set based on the modified process parameters.

[0018]

[0019]

[0020] ​​​​​The application corrects the process parameters in the optimization data set, reduces the errors caused by parameter deviation, and further improves the accuracy of the corrected data set.

[0021] Optionally, the correction of the process parameters in the optimization data set to obtain the corrected process parameters comprises: respectively correcting the spinning solution viscosity, the spinning speed, the spinning temperature, the coagulation bath parameters and the ring blowing air speed to obtain the corrected process parameters.

[0022] The application corrects the spinning solution viscosity, the spinning speed, the spinning temperature, the coagulation bath parameters and the ring blowing air speed, which are the key process parameters for producing lyocell short fibers, respectively, can eliminate the deviation and error of each parameter itself, the correction of the spinning solution viscosity can ensure the stability of the solution state, which is beneficial to the uniform formation of fibers, the correction of the spinning speed can ensure the consistency of fiber stretching and structure, the correction of the spinning temperature makes the temperature more suitable for process requirements, optimizes molecular motion and spinning efficiency, the correction of the coagulation bath parameters improves the stability of the coagulation process, and the correction of the ring blowing air speed makes the fiber cooling and setting more uniform, which greatly improves the accuracy of the process parameters.

[0023] Optionally, the determination of the key features based on the corrected data set comprises: obtaining a second quality index prediction model based on a long short-term memory network with self-attention mechanism using the corrected data set; calculating feature weights using the second quality index prediction model; and determining key features according to the feature weights.

[0024] The application builds a second quality index prediction model by introducing a long short-term memory network with self-attention mechanism, which can fully exert the modeling capability of LSTM on time series data and the focusing capability of attention mechanism on key information, accurately capture the nonlinear and time series dependence relationship between features and quality indexes in the corrected data set, objectively quantify the contribution of each feature to quality by means of the feature weights calculated by the model, and further efficiently screen out key features, thereby improving the scientificity of the key features.

[0025] Optionally, the construction of the first quality index prediction model using the training samples comprises: setting a basic loss function based on mean square error; setting a time consistency constraint term, and constructing a total loss function using the time consistency constraint term and the basic loss function; and training the first quality index prediction model using the training samples based on the total loss function.

[0026] The application sets a basic loss function through mean square error, provides a core error measurement standard for the model, and ensures that the model can accurately fit the quality index law in the training sample. The time consistency constraint term is introduced and the total loss function is constructed, which can effectively constrain the change amplitude of the model in the continuous time step prediction, avoid abnormal jitter that does not conform to the production physical law, and give the prediction result stronger rationality and interpretability. The first quality index prediction model trained based on the total loss function can not only ensure the prediction accuracy, but also conform to the time sequence characteristics of the lyocell short fiber production, thereby improving the performance of the first quality index prediction model.

[0027] Optionally, the time consistency constraint term satisfies the following formula:

[0028]

[0029] wherein, is the time consistency constraint term, is the total number of time steps, is the quality index prediction value of the time step is the quality index prediction value of the time step is the quality index prediction value of the time step is the quality index prediction value of the time step is the quality index detection delay time step.

[0030] The formula of the time consistency constraint term of the application quantifies the difference between adjacent time step prediction values, effectively constrains the smoothness of the quality index prediction result in the time dimension, and since the lyocell short fiber production is a continuous process and the quality change has time sequence correlation, this constraint can avoid unreasonable mutations in the prediction value, make the output of the first quality index prediction model more conform to the production physical law, and further improve the performance of the first quality index prediction model.

[0031] Another aspect of the application also provides a lyocell short fiber quality prediction system, comprising a processor, an input device, an output device and a memory, the processor, the input device, the output device and the memory are connected with each other, wherein the memory is used for storing a computer program, the computer program comprises program instructions, and the processor is configured to call the program instructions to execute the lyocell short fiber quality prediction method of any one of the previous aspect.

[0032] The lyocell short fiber quality prediction system of the application has compact structure, stable performance, high integration and simple structure, and can stably execute the lyocell short fiber quality prediction method provided in the previous aspect, thereby further improving the overall applicability and practical application ability of the application. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1A lyocell short fiber quality prediction method flow chart for an embodiment of the present application;

[0034] Figure 2 A lyocell short fiber quality prediction system structure schematic diagram for an embodiment of the present application. DETAILED DESCRIPTION

[0035] The specific embodiments of the present application will be described in detail below, and it should be noted that the embodiments described herein are only used for illustration and do not limit the present application. In the following description, a large number of specific details are set forth in order to provide a thorough understanding of the present application. However, it is obvious to those skilled in the art that the specific details need not be used to practice the present application. In other instances, well-known circuits, software or methods have not been specifically described in order to avoid obscuring the present application.

[0036] Throughout the specification, the reference to "one embodiment", "an embodiment", "one example" or "an example" means that a particular feature, structure or characteristic described in connection with the embodiment or example is included in at least one embodiment of the present application. Therefore, the phrases "in one embodiment", "in an embodiment", "one example" or "an example" appearing throughout the specification do not necessarily all refer to the same embodiment or example. In addition, specific features, structures or characteristics can be combined in any appropriate combination and / or subcombination in one or more embodiments or examples. In addition, those skilled in the art should understand that the diagrams provided herein are for illustrative purposes only and the diagrams are not necessarily drawn to scale.

[0037] Please refer to Figure 1 , in order to solve the problems in the prior art, in an optional embodiment, as shown in Figure 1 A lyocell short fiber quality prediction method, comprising the following steps:

[0038] Step S1, obtaining the monitoring data set formed by the process parameters, environmental condition data and finished product quality evaluation index data in the lyocell short fiber production process.

[0039] In the present embodiment, for the whole lyocell short fiber production process, through the high-precision sensors, industrial instruments and the like deployed at each key process point of the lyocell short fiber production line, process parameters and environmental conditions and finished product quality indexes are collected, and a multi-dimensional time series monitoring data set is integrated:

[0040] The process parameters cover key control variables across the entire process, including cellulose concentration and spinning solution viscosity in the spinning dosing preparation stage, and spinneret temperature, pump flow rate, annular air pressure / velocity, coagulation bath temperature, and component concentration in the dry-jet wet spinning stage. Environmental condition data focuses on key influencing factors in the production scenario, including overall temperature and humidity in the spinning workshop and atmospheric pressure in the production area. Finished product quality evaluation indicators are derived from offline precision testing and online sampling testing, including core quality parameters such as fiber breaking strength, breaking elongation, whiteness, moisture regain, single fiber linear density, and apparent defect rate. The above data are collected synchronously in time steps and stored in the industrial distributed control system (DCS), ultimately forming a multivariate time-series monitoring dataset that includes the entire process, environment, and quality chain, with timestamp alignment. The samples in the monitoring dataset meet the following requirements. This represents a multivariate time series sample with d sensor variables at a time step of T.

[0041] Step S2: Clean and complete the monitoring dataset to obtain an optimized dataset.

[0042] The process of cleaning and completing the monitoring dataset to obtain an optimized dataset specifically includes the following sub-steps:

[0043] Step S201: Calculate the data missing rate based on the monitoring dataset, and filter the monitoring dataset according to the data missing rate to obtain the first dataset.

[0044] In this embodiment, given that the monitoring data originates from sensors and detection equipment in multiple processes of the production line, it is inevitable that there will be missing NaN values ​​due to factors such as equipment accuracy, communication interruption, and environmental interference. For data with too many missing values, the difficulty and error of filling the missing data are relatively large, making it difficult to reflect the true situation. For such data, a deletion operation should be performed. The key to deletion is the calculation of the missing data rate. After calculating the missing data rate, the data with a missing data rate greater than 50% or lower (determined according to the actual situation) is deleted, and the resulting data is the first dataset.

[0045] The data missing rate satisfies the following formula:

[0046]

[0047] in, For monitoring data The missing rate, The total number of time steps. For the number of sensors, For indicator functions, For a moment No. The measurement values ​​of each sensor, These are missing values ​​during the sampling process of the DCS system.

[0048] Step S202, a data reconstruction model is constructed, and the first data set is reconstructed to obtain a second data set by using the data reconstruction model.

[0049] In the embodiment, the data reconstruction model is constructed based on LSTM-AE, process parameters and environmental condition data are collected from high-precision sensors, industrial instruments and the like deployed at key process points of the lyocell short fiber production process, and data with no missing or extremely small missing is selected as training data of the LSTM-AE (long short-term memory network-autoencoder), the dimension of the training data is consistent with that of the first data set, the data reconstruction model is obtained by training the training data, and the first data set is input into the data reconstruction model to obtain the second data set.

[0050] The autoencoder is an unsupervised learning model commonly used in deep learning, and the autoencoder is composed of an encoder and a decoder, and the target is to learn a low-dimensional representation (latent space) and to be able to reconstruct the original input.

[0051] For input data x, the encoder compresses it into a low-dimensional latent representation :

[0052]

[0053] wherein represents a sensor observation vector at a certain time step, represents its hidden state feature, is an encoder parameter.

[0054] is an encoder mapping function, which can compress the sensor observation vector at a certain time step into a low-dimensional hidden state feature under the action of the encoder parameter .

[0055] Then the decoder tries to recover the original data from :

[0056]

[0057] wherein is a reconstructed observation value, is a decoder parameter, is a decoder mapping function.

[0058] is a decoder mapping function, which can recover the hidden state feature obtained in the encoding stage into the original observation value Step-wise mapping to reconstructed observation vectors to achieve a temporal reconstruction of the original input data.

[0059] The model can eventually be trained by minimizing the reconstruction error between and .

[0060] Missing data in time series data is a common problem, and the combination of an autoencoder with LSTM (Long Short-Term Memory) is an effective method for handling missing data completion tasks with temporal characteristics. By combining LSTM with an autoencoder, a framework can be constructed that can handle both missing values and outliers in time series data. The architecture combining the two usually includes the following steps:

[0061] Encoding stage: Use LSTM as an encoder to map the input time series data into a low-dimensional latent space. This latent representation captures the main features of the time series.

[0062] Decoding stage: Use another LSTM as a decoder to reconstruct the latent representation into an approximation of the original input. By minimizing the reconstruction error, the model can learn how to recover the original data from the latent representation.

[0063] Missing value filling: During training, the model learns how to predict the missing parts from the known data. In the inference stage, the data containing missing values can be input into the encoder and then through the decoder to generate a complete data representation.

[0064] Specifically, the encoder module combined with LSTM: Given an input sequence containing missing values to be completed , the encoder processes step by step through the LSTM unit to generate the final state:

[0065]

[0066] where is the output hidden state representing the overall features of the input sequence, is the cell state used to capture long-term temporal dependency information.

[0067] is the encoder function based on the Long Short-Term Memory Network (LSTM), which receives a time series input sequence containing missing values to be completed , and after iterative processing through the LSTM unit step by step, generates the final state tuple .

[0068] The decoder part takes as the initial state and reconstructs the complete sequence step by step through LSTM :

[0069]

[0070] define a mask matrix wherein indicates that the data is an observation.

[0071] indicates a long short-term memory (LSTM) based decoder function, taking the encoder output as initial state, to generate the complete sequence by step-by-step prediction, achieving denoising and completion of the original time series.

[0072] wherein the data reconstruction model satisfies the following formula:

[0073]

[0074] wherein is a loss function of the data reconstruction model, is the total number of time steps, is the number of sensors, is the measurement value of the th sensor at time is the predicted value of the th sensor at time is an element in the mask matrix of the th sensor at time

[0075] the second data set satisfies the following formula:

[0076]

[0077] wherein is the reconstructed value of the th sensor at time is the true value of the th sensor at time is the reconstructed value of the th sensor at time

[0078] In step S203, the reconstruction error of the second data set is calculated.

[0079] The reconstruction error satisfies the following formula:

[0080] ​​​​​​

[0081] in, Time series samples in the second dataset Reconstruction error, The total number of time steps. For the number of sensors, For a moment No. The true value of each sensor, For a moment No. The reconstructed values ​​of each sensor.

[0082] Step S204: Delete abnormal data in the second dataset according to the reconstruction error to obtain an optimized dataset.

[0083] Abnormal data satisfies the following formula:

[0084]

[0085] in, Time series samples in the second dataset The overall reconstruction error, The mean reconstruction error of the training set for the data reconstruction model. This is an adjustment coefficient, the value of which determines the model's sensitivity in anomaly detection: The larger the value, the more lenient the judgment; only samples that deviate significantly from the mean will be considered abnormal. The smaller the value, the more stringent the detection will be, and more minor deviations may be identified as anomalies. This is typically seen in industrial process monitoring. A score between 2 and 3 is common. Samples that meet the criteria will be considered outliers and will be removed or flagged. The standard deviation of the reconstruction error of the training set for the data reconstruction model.

[0086] Deleting the abnormal data yields the optimized dataset.

[0087] Step S3: Correct the optimized dataset to obtain the corrected dataset.

[0088] The process of correcting the optimized dataset to obtain the corrected dataset specifically includes the following sub-steps:

[0089] Step S301: Correct the process parameters in the optimized dataset to obtain the corrected process parameters.

[0090] The optimized dataset includes process parameters such as spinning solution viscosity, spinning speed, spinning temperature, coagulation bath parameters, and annular airflow velocity. The process parameters in the optimized dataset are then modified to obtain the corrected process parameters, specifically including:

[0091] The spinning solution viscosity, the spinning speed, the spinning temperature, the coagulation bath parameters and the ring blowing air speed are respectively corrected to obtain corrected process parameters.

[0092] In the mass prediction of Lyocell dry-jet wet-spun fibers, the process parameters are collected from the real-time signals of the distributed control system (DCS), which often have noise, delay, drift or local anomalies. If directly input into the prediction model, it is easy to lead to distorted results. Therefore, a special correction method for core process parameters in the spinning process is proposed to convert them into effective features that conform to the physical laws and process characteristics.

[0093] The corrected process parameters satisfy the following formula:

[0094]

[0095] wherein, is the time corrected spinning solution viscosity, is the time spinning solution viscosity, is a correction factor function obtained based on experimental calibration, is the time spinning temperature, is the time spinning solution NMMO concentration.

[0096] The correction factor function is obtained from the experimental calibration for a specific Lyocell fiber production system. Since the spinning solution viscosity is affected by the coupling of temperature and concentration, the DCS measurement value has a systematic deviation, which needs to be constructed by a multivariate control experiment. Different temperature and concentration gradients are set within the process range, and the DCS viscosity measurement value and the real viscosity value detected by high-precision offline are collected synchronously. The variation law of the deviation of the two with temperature and concentration is analyzed, and then the specific expression of is obtained by regression analysis (such as linear, polynomial or machine learning fitting). The parameters and forms are determined by the scene experimental data of production equipment characteristics, raw material batches, etc., which are used for accurate compensation of systematic deviation.

[0097]

[0098] wherein, is the time corrected draw roller speed, is the time draw roller speed, is a sliding median filter function with a window size of .

[0099]

[0100]

[0101] where, is the time corrected coagulation bath temperature, is the number of coagulation bath measurement points, is the weight of the coagulation bath temperature measurement point, is the time coagulation bath temperature of the coagulation bath measurement point, is the time corrected coagulation bath composition concentration, is the weight of the coagulation bath composition concentration measurement point, is the coagulation bath composition concentration at the time .

[0102]

[0103] where, and represent the mean, variance and skewness, respectively, represents the measured loop air blowing speed, represents the corrected loop air blowing speed.

[0104] In summary, all parameter corrections can be approximately uniformly represented as:

[0105]

[0106] where, represents the original process parameters extracted by the DCS system, represents the parameter set required for correction (including weight, correction function, delay, etc.), represents the correction function combined with the physical law, represents the corrected effective parameter, which is convenient for input into the subsequent prediction model.

[0107] In addition to the above key process parameters represented by the spinning solution viscosity, spinning temperature, spinning speed, coagulation bath parameters and ring blowing air speed, other monitoring parameters (such as pump supply, environmental pressure, solution NMMO ratio, environmental temperature and humidity, etc.) in the spinning process are also applicable. The correction and processing of these parameters can be in accordance with the same principle, that is, combined with their physical meaning and process action in the Lyocell dry jet wet spinning process, the methods such as time delay alignment based on mechanism constraint, outlier filtering, weighted fusion or statistical feature extraction are adopted, so as to ensure that the input data is stable, true and consistent with the process law before entering the prediction model. In this way, the present application can take into account the universal processing of other secondary parameters while keeping the key optimization of core process parameters, further improving the reliability and adaptability of the overall prediction model.

[0108] Step S302, obtaining a corrected data set based on the corrected process parameters.

[0109] In this embodiment, the sequence of the corrected process parameters is replaced with the corresponding uncorrected process parameter column in the original optimization data set based on the timestamp of the optimization data set, while the environmental condition data column and the product quality evaluation index data column are retained, forming a time series data set with the structure of corrected process parameters, original environmental data and original quality indicators.

[0110] Step S4, determining key features based on the corrected data set, and screening the corrected data set using the key features to obtain a training sample.

[0111] In this embodiment, the key features are determined based on the corrected data set, which includes:

[0112] Step S401, obtaining a second quality index prediction model based on the corrected data set by using a long short-term memory network with self-attention mechanism.

[0113] In this embodiment, sample quality analysis is performed on the cleaned data set. Considering that there is a coupling relationship between different process feature parameters, and the relationship between the quality index and the feature parameter is nonlinear, it is necessary to find the feature parameters that are strongly related to the fiber quality index, i.e. the key features.

[0114] Based on this, the present application proposes an LSTM-Attention algorithm model for screening feature parameters that are strongly related to different quality indicators and identifying the dependency relationship between these feature parameters and the quality indicators.

[0115] Let the input feature parameters be multi-dimensional time series features , and the output be a quality index sequence . The target is to construct , where k is the window length.

[0116] Attention, which originates from natural language processing tasks, gives different "attention" weights to different parts of the input sequence. In multivariate input, it can be used to identify the most influential feature variable at each time point.

[0117] Specifically, let be the input feature at time t, where represents the acquisition value of the th sensor variable at time t, represents the number of sensors. The mechanism calculates the weight of each feature as follows:

[0118]

[0119] where is a trainable weight vector, , represents the importance of feature at time t, and represents its relative importance to the quality indicator.

[0120] In order to better improve the prediction performance of the model and enhance its ability to identify feature importance, the key features are selected. The invention combines the modeling ability of LSTM for time series data and the dynamic focusing ability of Attention mechanism for key information, so that it can automatically identify the input features and time points that have the greatest impact on the output variable when dealing with complex data with multiple variables and multiple time steps. The LSTM-Attention model can effectively identify input variables that have a significant impact on the final quality indicator and dynamically adjust their weights at different time steps, thereby revealing the temporal dependence between variables.

[0121] Here, the Attention module is embedded as a pre-feature selector before the LSTM network, and the specific structure is as follows:

[0122] At each time step , the Attention mechanism is used to extract a weighted combination from the input feature parameters:

[0123]

[0124] where represents the effective input that takes into account the contribution of each process variable to the quality indicator at that time.

[0125] The weighted sequence is input into the LSTM model to obtain the final hidden state​ :

[0126]

[0127] in, This represents the temporal dependency features learned from historical process variables.

[0128] When performing output prediction, a fully connected layer is used to predict the quality metric:

[0129]

[0130] in, Indicates at time The predicted quality index vector, The output layer weight matrix maps the hidden states to the quality metric space. This is a bias term.

[0131] The mean squared error is used as the loss function, and the training is carried out by correcting the dataset to achieve the purpose of fitting the actual quality index. After training, the second quality index prediction model is obtained.

[0132] The second quality indicator prediction model satisfies the following formula:

[0133]

[0134] in, The loss function for the second quality index prediction model is... These are the quality indicators obtained through actual measurement. These are the model's predicted values.

[0135] Step S402: Calculate feature weights using the second quality index prediction model.

[0136] The feature weights satisfy the following formula:

[0137]

[0138] in, For feature weights, The total number of time steps. For a moment No. The weights of each sensor.

[0139] Step S403: Determine key features based on the feature weights.

[0140] In this embodiment, based on the feature weights output by the second quality index prediction model, all input features are sorted from largest to smallest according to their weight values. The higher the weight value, the stronger the contribution and correlation of the feature to the quality index (such as breaking strength and elongation) throughout the entire time series. Subsequently, screening criteria are set according to the needs of the industrial scenario: a weight threshold method can be used to retain features with weights higher than a preset threshold (the threshold is determined based on the prediction accuracy of the model validation set), or the top N features with the highest weights can be selected (N is adjusted according to the model input dimension requirements). The high-contribution features are then selected as key features.

[0141] Finally, the corrected dataset is filtered using key features, that is, only the data in the corrected dataset that can match the key features are retained to form training samples.

[0142] Step S5: Construct a first quality index prediction model using the training samples, and use the first quality index prediction model to predict the quality of lyocell staple fiber.

[0143] The construction of the first quality index prediction model using the training samples specifically includes the following sub-steps:

[0144] Step S501: Set the basic loss function based on the mean square error.

[0145] In this Lyocell staple fiber quality prediction, the loss function not only serves as the model training objective but also directly determines the optimization direction of the prediction results across different quality dimensions. Traditional mean squared error (MSE) assumes all output dimensions have equal importance, which is clearly not true in actual Lyocell dry-jet wet spinning processes. Different quality indicators, such as fiber breaking strength, fiber uniformity, moisture regain, and apparent defect rate, have varying sensitivities and control priorities in production practice. Simply pursuing overall error minimization often leads to under-prediction of key indicators while overfitting non-key indicators. Therefore, a weighted and structured design tailored to industrial application scenarios is introduced at the loss function level to improve the model's process adaptability and controllability.

[0146] To reflect the varying degrees of importance of quality metrics in practical applications, the loss function is defined as a multi-task weighted form:

[0147]

[0148] Where m represents the number of quality indicators. and The first The actual and predicted values ​​of each quality indicator The error measurement function (usually mean squared error) represents the index. These are weighted parameters related to process objectives. Dynamic setting allows for predictive optimization tailored to production needs. For example, when a production batch demands higher filament strength, the weight of breaking strength should be increased accordingly; if the current process prioritizes the stability of the spinning process, the weights of parameters related to the spinning solution or the stability of the airflow will be assigned greater values. Thus, the loss function can be linked to process priorities, rather than simply seeking a mathematically optimal solution.

[0149] Therefore, the basic loss function satisfies the following formula:

[0150]

[0151] in, Based on the loss function, For the quantity of quality indicators, For the first The weight of each quality indicator, The total number of time steps. For time steps The true values ​​of the quality indicators For time steps Predicted values ​​of quality indicators The time step is the delay step for quality indicator detection.

[0152] Step S502: Set time consistency constraints, and construct the total loss function using the time consistency constraints and the basic loss function.

[0153] The time consistency constraint term satisfies the following formula:

[0154]

[0155] in, For time consistency constraints, The total number of time steps. For time steps Predicted values ​​of quality indicators For time steps The predicted values ​​of quality indicators.

[0156] This constraint can prevent abnormal fluctuations that are inconsistent with physical laws by limiting the magnitude of changes in the model's predictions across consecutive time steps. This smoothing regularization can effectively improve the interpretability and physical plausibility of the prediction results.

[0157] Finally, the time consistency constraint is added to the basic loss function to obtain the total loss function. .

[0158]

[0159] in, For the sample size, For the time consistency constraint term coefficient, Indicates the first The basic loss function for each sample For the first Time consistency constraints for each sample.

[0160] Step S503: Based on the total loss function, a first quality index prediction model is trained using the training samples.

[0161] In this embodiment, a first quality index prediction model is constructed based on a BiLSTM-Attention network, and the total loss function is set into the loss function of the first quality index prediction model. Finally, the first quality index prediction model is trained using training samples, and after training, a first quality index prediction model that can perform quality prediction is obtained.

[0162] LSTM (Long Short-Term Memory) is a special type of RNN (Recurrent Neural Network). Compared to traditional RNNs, which are prone to gradient vanishing or exploding problems when processing long sequences and struggle to capture long-term dependencies, LSTM alleviates this problem by introducing a gating mechanism. However, its unidirectional structure can only utilize historical information, while BiLSTM enhances the model's ability to model time series by simultaneously considering both forward and backward information.

[0163] The bidirectional LSTM layer uses two LSTM networks, one forward and one backward, to process the input sequence, and outputs the forward sequence. and reverse Then connect them to a completely hidden state:

[0164]

[0165] Meanwhile, this model introduces an attention mechanism into BiLSTM, enabling the model to dynamically focus on the most influential subsequences in the input sequence during prediction. The attention mechanism calculates attention weights using learnable parameters and performs a weighted summation of the hidden states of BiLSTM to generate a context vector for subsequent prediction tasks. This approach not only improves the model's ability to identify key time steps but also enhances its interpretability.

[0166] The attention mechanism can hide the state at each time step of the BiLSTM output. Different attention weights are assigned to obtain a weighted representation of the entire sequence. The weights are calculated as follows:

[0167]

[0168] wherein, is the hidden state of BiLSTM at time step t, is the weight matrix, is the bias term, is the context vector for measuring the importance of the current time step, is the intermediate attention score.

[0169] Subsequently, the attention scores of all time steps are normalized by Softmax:

[0170]

[0171] Thus, the final context representation of the sequence is obtained:

[0172]

[0173] The vector is a weighted combination of the most critical moments in the entire sequence, extracting the information that contributes most to the prediction of future quality.

[0174] Finally, the context vector is mapped to a quality indicator prediction value, and output.

[0175]

[0176] wherein, is the output weight matrix, is the output bias term, is the predicted multi-quality indicator vector.

[0177] After obtaining the first quality indicator prediction model after training, when predicting quality, collect real-time process parameters and environmental condition data of the production line to form timestamp-aligned multivariate time series data, which is input into the first quality indicator prediction model after interpolation, data correction and key feature extraction (consistent with the data dimension of the first quality indicator prediction model) to obtain multiple quality indicators.

[0178] As shown in Figure 2 , in another aspect, the present application also provides a lyocell short fiber quality prediction system, comprising: a processor, an input device, an output device and a memory, the processor, the input device, the output device and the memory are connected with each other, wherein the memory is used for storing a computer program, the computer program comprises program instructions, the processor is configured to call the program instructions, and execute the related steps of the related embodiments of the lyocell short fiber quality prediction method of the present application.

[0179] The lyocell short fiber quality prediction system provided by the application can integrate each functional component in one processing component, or each component can exist physically alone, or two or more components can be integrated in one component. The integrated components can be realized in the form of hardware or in the form of software function.

[0180] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the application, and they should be covered in the scope of the claims and the description of the application.

Claims

1. A method for predicting the mass of lyocell staple fibers, characterized by, The method comprises: obtaining a monitoring data set formed by process parameters, environmental condition data and product quality evaluation index data in the production process of lyocell short fiber; cleaning and completing the monitoring data set to obtain an optimized data set, comprising: calculating a data missing rate according to the monitoring data set, and screening the monitoring data set according to the data missing rate to obtain a first data set; constructing a data reconstruction model and reconstructing the first data set using the data reconstruction model to obtain a second data set; the data reconstruction model satisfies the following formula: wherein, is a loss function for the data reconstruction model, is a total number of time steps, is a number of sensors, is a time instant is a measurement value of the th sensor, is a predicted value of the th sensor at a time instant is a predicted value of the th sensor at a time instant is an element in a mask matrix of the th sensor; calculating the reconstruction error of the second data set; the reconstruction error satisfies the following formula: in, Time series samples in the second dataset Reconstruction error, The total number of time steps. For the number of sensors, For a moment No. The true value of each sensor, For a moment No. Reconstructed values ​​from each sensor; deleting abnormal data in the second data set according to the reconstruction error to obtain an optimized data set; correcting the optimized data set to obtain a corrected data set; determining key features based on the corrected data set, and screening the corrected data set using the key features to obtain a training sample; the determination of key features based on the corrected data set comprises: obtaining a second quality index prediction model using the corrected data set based on a long short-term memory network with self-attention mechanism; calculating feature weights using the second quality index prediction model; determining key features according to the feature weights; constructing a first quality index prediction model using the training sample, and predicting the quality of lyocell short fiber using the first quality index prediction model.

2. The method for predicting the quality of lyocell staple fiber according to claim 1, characterized in that, the correction of the optimized data set to obtain a corrected data set comprises: correcting process parameters in the optimized data set to obtain corrected process parameters; obtaining a corrected data set based on the corrected process parameters.

3. The method of claim 2, wherein the step of determining the quality of the lyocell short fiber is performed by using a regression model. the process parameters in the optimized data set include spinning solution viscosity, spinning speed, spinning temperature, coagulation bath parameters and circular blow air speed, and the correction of the process parameters in the optimized data set to obtain corrected process parameters comprises: respectively correcting the spinning solution viscosity, the spinning speed, the spinning temperature, the coagulation bath parameters and the circular blow air speed to obtain the corrected process parameters.

4. The method of claim 1, wherein the lyocell staple fiber quality is predicted by the equation: ###00002### wherein the variables are defined as in claim 1. the construction of a first quality index prediction model using the training sample comprises: ​ setting a basic loss function based on mean square error; setting a time consistency constraint term, constructing a total loss function using the time consistency constraint term and the basic loss function; training a first quality index prediction model based on the total loss function using the training sample.

5. The method of claim 4, wherein the step of determining the quality of the lyocell short fiber is performed by using a regression equation. the time consistency constraint term satisfies the following formula: wherein, is a temporal consistency constraint term, is the total number of time steps, is the time step is the quality indicator prediction value for the time step is the quality indicator prediction value for the time step is the quality indicator prediction value for the time step is the quality indicator detection delay time step.

6. A lyocell staple fiber mass prediction system characterized by, comprising: a processor, an input device, an output device and a memory, the processor, the input device, the output device and the memory are connected to each other, wherein the memory is used to store a computer program, the computer program comprises program instructions, the processor is configured to call the program instructions, and execute a lyocell short fiber quality prediction method according to any one of claims 1 to 5.

Images