Intelligent prediction method for rear weighing of tread extrusion production line

Abstract

The invention relates to an intelligent prediction method for the post-weighing weight of a tread extrusion production line. The method comprises the following steps: 1) obtaining historical data and real-time data of a production process of a rubber tire extruder; 2) constructing a preliminary similar data set through a weighted similarity analysis method based on Euclidean distance; 3) determining the size of a current similar data set by using a Pearson correlation coefficient; 4) performing abnormal value elimination on the current similar data set by using an isolated forest algorithm; and 5) on the basis of a multi-head attention mechanism, constructing a local model of instant learning, training the model on the basis of the training set to obtain a post-weighing weight prediction model, obtaining actual real-time data, inputting the data into the prediction model, and obtaining a post-weighing weight prediction result of the product of the rubber tire extrusion production line. Compared with the prior art, the method has the advantages that the post-weighing weight prediction efficiency is improved, the prediction hysteresis is reduced, and the like.

Description

Technical Field

[0001] The invention relates to the field of product quality index prediction in process industries, and in particular to an intelligent prediction method for the post-weighing weight of a tread extrusion production line. Background Art

[0002] In the industrial production process, in order to meet the requirements of product quality, simple offline monitoring often cannot complete the detection task well due to its lag. The post-weighing weight measurement of the rubber tire extrusion production line has strong lag and inaccuracy. By predicting the post-weighing weight in real time, the measurement lag can be greatly reduced, which is of great significance for the real-time adjustment of the parameters of the production process. However, since the global modeling is affected by the product model and the collected data, the error in the indicator prediction is large, and the model training and detection time is long. Therefore, the real-time learning model is used to adjust the local model in real time according to the input data and complete the weight prediction to meet the changes in the production process. It can greatly improve the detection accuracy and adjust the local model structure to improve the detection efficiency.

[0003] When building a real-time learning model, similarity analysis is usually used first and abnormal data points are removed to obtain training data for the local model. In order to solve the problem of over-stretching and over-compression when applying DTW, the paper "Adaptively constrained dynamic time warping for time series classification and clustering" proposes an adaptive constrained DTW (ACDTW) algorithm, which introduces a new adaptive penalty function to more accurately calculate the distance between trajectories, can adaptively adjust the correspondence between two trajectories, and obtain higher accuracy between different trajectories. In order to mine time series based on LCSS, the paper "Elastic Similarity and Distance Measures for Multivariate Time Series" proposes a metric method called Developing the Longest Common Subsequence (DLCSS), which eliminates the shortcomings of LCSS by defining two similarity thresholds to determine their values. Its effect in determining better cluster representatives is greater than that under LCSS and DTW, with a confidence level of at least 99.95%. The paper "A New Local Distance-Based Outlier Detection Approach for Scattered Real-World Data" proposed a local distance-based outlier factor (LDOF) to measure outliers in scattered data sets. LDOF uses the relative position of an object and its neighboring objects to determine the degree to which the object deviates from its neighboring objects. Experimental results show that its detection accuracy is better than traditional KNN and LOF, especially it is not sensitive to parameter values. The paper "Variational LSTM Enhanced Anomaly Detection for Industrial Big Data" proposed a variational long short-term memory learning model (VLSTM) for intelligent anomaly detection based on reconstructed feature representation, which significantly improved the accuracy and false alarm rate of anomaly detection in industrial big data.

[0004] In addition, many researchers have designed JITL similarity analysis and local model structures through an overall framework. The paper "An ensemble JITL method based on multi-weighted similarity measures for cold rolling force prediction" proposes a new JITL model based on similarity analysis method to improve the accuracy of rolling force prediction on product quality. First, relevant samples are selected through multi-weighted similarity measurement. Then, local models are built and the output values ​​of the data are queried. Finally, the output of each local model is integrated according to the ensemble learning strategy. In addition, the cumulative similarity factor is introduced to optimize the number of samples for local modeling, and the threshold is set to adaptively update the local model, which has achieved good results. Considering that traditional JITL modeling is not suitable for industrial processes that exhibit time-varying and non-Gaussian behavior, the paper "Adaptive Gaussian Mixture Model-Based Relevant Sample Selection for JITL Soft Sensor Development" proposes a GMM-based similarity measure to improve the prediction accuracy of JITL soft sensors. The feasibility and effectiveness of the proposed method are demonstrated by discussing case studies involving numerical examples and industrial processes. Considering that traditional JITL methods mainly focus on equal sample sizes between process input variables and quality output variables, the paper "Semisupervised JITL Framework for NonlinearIndustrial Soft Sensing Based on Locally Semisupervised Weighted PCR" proposes a new semi-supervised JITL framework based on semi-supervised weighted probabilistic principal component analysis (SWPPCR). By selecting the most relevant labeled and unlabeled samples and assigning weights to them, local SWPPCR can be constructed to estimate the output variables of the query sample. The effectiveness and flexibility of this method are demonstrated through case evaluation of predictive ability in industrial processes.

[0005] As mentioned above, many studies have demonstrated the effectiveness of just-in-time learning models in real-time prediction. However, when applied in the field of weighing, most existing just-in-time learning models do not consider the complexity of the algorithm. Due to the similarity analysis stacking or the complex structure of the local model, it takes a long time to build the data set and train the local model for prediction, and it is impossible to significantly reduce the prediction lag. Summary of the invention

[0006] The purpose of the present invention is to provide an intelligent prediction method for the post-weighing weight of a tread extrusion production line in order to improve the prediction efficiency of the post-weighing weight and reduce the prediction hysteresis.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] Compared with the prior art, the present invention has the following beneficial effects:

[0009] A method for intelligently predicting the weight of a tread after extrusion production line, the method comprising the following steps:

[0010] 1) Obtain historical data of the rubber tire extruder production process and real-time data for model prediction;

[0011] 2) Using the weighted Euclidean distance similarity analysis method, a certain amount of similar data is obtained based on historical data and real-time data to construct a preliminary similarity data set;

[0012] 3) using the Pearson correlation coefficient, determining the size of the similar data set with the highest correlation according to the correlation of the different size subsets of the preliminary similar data set, and taking the subset of the similar data set size with the highest correlation as the current similar data set;

[0013] 4) Use the isolation forest algorithm to remove outliers from the current similar data set to obtain a training set;

[0014] 5) Based on the multi-head attention mechanism, a local model of instant learning is constructed, and the model is trained based on the training set to obtain a post-weighing weight prediction model. The actual real-time data is input into the prediction model to obtain the post-weighing weight prediction results of the products of the rubber tire extrusion production line.

[0015] Further,

[0016] 2) The specific steps are:

[0017] 201) Obtaining the Euclidean distance calculation formula;

[0018] 202) Mean centering of historical data;

[0019] 203) Perform dimensionality reduction projection to calculate the variance of each data point of the historical data after mean centering projected to the low-dimensional coordinates;

[0020] 204) maximize the value of the variance and solve the weight factor using the Lagrange multiplier method;

[0021] 205) calculating weighted Euclidean distance similarity based on the weight factor and the Euclidean distance calculation formula;

[0022] 206) According to the real-time data, a certain number of data with the greatest similarity are selected from the historical data through weighted Euclidean distance similarity to construct a preliminary similarity data set.

[0023] Furthermore, in 202), the historical data is represented by X = [x1, x2, ..., x m ], where m represents the total number of historical data. The historical data is m n-dimensional data. The historical data after mean centering is:

[0024]

[0025] Where i represents the sequence number of historical data.

[0026] Furthermore, the variance is:

[0027]

[0028] Among them, ω is the weight factor and Var represents the variance.

[0029] Furthermore, the weighted Euclidean distance similarity is:

[0030]

[0031] Among them, x i Represents various historical data, y j represents each real-time data, k represents the serial number of the dimension, x ik For data x i The kth dimension of y jk For data y j The kth dimension of i Indicates the weight factor of the i-th dimension data.

[0032] Furthermore, the specific steps of 3) include:

[0033] 301) obtaining subsets of different sizes of preliminary similar data sets;

[0034] 302) Calculate the correlation between subsets of different sizes and real-time data through the Pearson correlation coefficient and select the subset with the largest correlation. The size of this subset is the most appropriate size of the similar data set;

[0035] 303) The subset corresponding to the most appropriate similarity data set size is used as the current similarity data set.

[0036] Furthermore, the specific steps of 4) are:

[0037] 401) Construct an isolation forest algorithm model;

[0038] 402) using the post-weighing weight of the current similar data set as input to train the isolation forest algorithm model to obtain a trained algorithm model;

[0039] 403) Use the trained algorithm model to remove abnormal data from the current similar data set to obtain a training set.

[0040] Furthermore, the specific steps of 5) are:

[0041] 501) Construct a multi-head attention mechanism network structure model, that is, a local model;

[0042] 502) Determine the dimensions of the input and use the post-weighing weight as output;

[0043] 503) Select the mean absolute error loss function as the loss function of the model;

[0044] 504) The model is trained based on the training set to obtain a weighing weight prediction model, and actual real-time data is input into the prediction model to obtain a weighing weight prediction result of the product of the rubber tire extrusion production line.

[0045] Furthermore, the loss function is:

[0046]

[0047] Among them, y i refers to the true value, f(x i ) refers to the predicted value.

[0048] Furthermore, the network structure of the local model only includes three linear layers.

[0049] The invention collects the historical data set of rubber tire extrusion production, derives the weight factor that can reduce the impact of the physical dimension of data of different dimensions, and proposes a weighted Euclidean distance similarity analysis method. The new similarity analysis method is used to select data to construct a similar data set, which not only improves the quality of the selected data, but also speeds up the efficiency of similarity analysis and improves the training and prediction speed of the prediction model. At the same time, the Pearson correlation coefficient is used to measure the correlation size of different size subsets of the initial similar data set to determine the most appropriate data set size, thereby improving the detection accuracy. The model of the present invention adopts a multi-head attention mechanism for post-weighing weight prediction, which can perform real-time monitoring of quality indicators, is conducive to reducing measurement hysteresis, and is convenient for production parameter adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the present invention;

[0051] Figure 2 It is a schematic diagram of the system operation architecture flow of the present invention;

[0052] Figure 3 Determine the result graph for the data set size of the present invention;

[0053] Figure 4 This is a graph of outlier removal results of the present invention;

[0054] Figure 5 It is a post-weighing weight prediction result diagram of the present invention;

[0055] Figure 6 It is a result comparison diagram of the present invention, wherein Figure 6 (a) shows the comparison result of the present invention with the global prediction model ANN and LSTM. Figure 6 (b) shows the comparison results of the present invention, JITL-ANN and WSJITL-ANN. Figure 6 (c) shows a box plot comparing the present invention with other models. DETAILED DESCRIPTION

[0056] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0057] The present invention proposes an intelligent prediction method for weighing weight after tread extrusion production line, the flow chart and block diagram are as follows Figure 1 and Figure 2 As shown, the method comprises the following steps:

[0058] 1) Obtain historical data of the rubber tire extruder production process and real-time data for model prediction;

[0059] 2) Using the weighted similarity analysis method based on Euclidean distance, a certain amount of similar data is obtained based on historical data and real-time data to construct a preliminary similarity data set;

[0060] 3) using the Pearson correlation coefficient to determine the most appropriate size of the similarity dataset according to the correlation of subsets of different sizes of the preliminary similarity dataset, and using the subset of the most appropriate size of the similarity dataset as the current similarity dataset;

[0061] 4) Use the isolation forest algorithm to remove outliers from the current similar data set to obtain a training set;

[0062] 5) Based on the multi-head attention mechanism, a local model of instant learning is constructed, and the model is trained based on the training set to obtain a weighing weight prediction model. The actual real-time data is input into the prediction model to obtain the weighing weight prediction results of the products of the rubber tire extrusion production line.

[0063] 1), the historical data of the production process of different models of rubber tire extrusion production lines are obtained. After removing some invalid physical quantities, the size is 30630×27, and each data contains 27 physical quantities such as screw speed and current; the real-time data used for model prediction is obtained, and the dimension is adjusted according to the size of the historical data to construct a real-time data set of 60×27 size.

[0064] 2), a certain amount of similar data is obtained through a weighted similarity analysis method based on Euclidean distance to construct a preliminary similarity data set. The specific implementation steps are as follows:

[0065] 201) Considering the efficiency of similarity analysis, Euclidean distance is selected as the benchmark similarity analysis method. The similarity of two data is judged by the size of the Euclidean distance. The Euclidean distance formula is:

[0066]

[0067] 202) Calculate the weight factor of weighted similarity analysis. The main idea is to project high-dimensional historical data onto one dimension through dimensionality reduction projection, retain the original information through variance maximization, calculate the projection matrix, and obtain the weight factor.

[0068] Assume there are m n-dimensional data, which can be expressed as X = [x1, x2, ..., x m ], and centering its mean gives:

[0069]

[0070] 203) Assume that we have a direction vector ω, and the coordinates of each data point projected onto this dimension are ω t x i , the variance can be expressed as follows:

[0071]

[0072] 204) in ω T Under the condition of ω=1, maximize the value of variance, let Solve for ω using the Lagrange multiplier method.

[0073] L(ω,λ)=ω T Cω+λ(1-ω T ω)

[0074] 205) Combining the weight factor with the Euclidean distance formula, we get the weighted Euclidean distance similarity analysis method. We get the weight factor ω=[ω1,ω2,…,ω n ], combining the weight and Euclidean distance to obtain the weighted Euclidean distance similarity calculation formula:

[0075]

[0076] Among them, Dis(x i ,y j ) represents data x i and j The weighted Euclidean distance between i Indicates the weight factor of the i-th dimension data, x ik For data x i The kth dimension of jk For data y j The kth dimension of .

[0077] 206) The weighted Euclidean distance similarity analysis method is used to calculate the similarity between the real-time data set and the historical data set, and a certain amount of data is selected to construct a similar data set. Considering that the amount of data collected for each type of rubber material during data collection is about 1500-2000, 2000 data are selected as the benchmark data set size.

[0078] 3), the Pearson correlation coefficient is used to measure the correlation of different size subsets of similar data sets to determine the most appropriate size of similar data sets. The specific implementation steps are as follows:

[0079] 301) Obtain subsets of different sizes of a preliminary similar data set of size 2000. If 100 data are used as an incremental basis to construct a data set subset, it can be expressed as:

[0080] D={D1,D2,…,D 20}

[0081] Here, the present invention selects 100 data points as increments to determine the most appropriate size of similar data sets.

[0082] 302) The correlation between the subsets of different sizes and the original real-time data is calculated by the Pearson correlation coefficient and the subset with the largest correlation is selected. The subset size is the most appropriate subset size. Figure 3 As shown, each subset corresponds to a value between 0 and 1, indicating the relevance of the subset of this size to the real-time data. Here, it is shown that when the size of the data set is 1800, its size is the most appropriate.

[0083] 303) The subset is treated as a similar data set for further processing.

[0084] In 4), the isolation forest algorithm is used to remove outliers from the similar data set in 3). The specific implementation steps are as follows:

[0085] 401) Build an isolation forest algorithm model.

[0086] 402) The post-weighing weight of the similar data set is used as input to train the isolation forest algorithm to obtain a trained algorithm model.

[0087] 403) Use the trained model to remove abnormal data from the data set.

[0088] The results of outlier removal are as follows Figure 4 As shown, the outlier ratio of the isolation forest rule is set to 10%, and the algorithm prediction results are 1 and -1. 1 indicates that the point is a normal value, represented by a yellow dot, and -1 indicates that the point is an outlier, represented by a purple dot, and it will be eliminated.

[0089] In step 5), based on the multi-head attention mechanism, a local model for instant learning is constructed, and the local model is trained using the similar data set in step 4, and the trained model is used to predict the subsequent weighing weight. The specific implementation steps are as follows:

[0090] 501) Construct a multi-head attention mechanism network structure model. The network structure only contains three linear layers, which meets the requirements of lightweight network and fast training and prediction.

[0091] 502) Select the appropriate dimension as input and determine the dimensions of the three input layers; then weigh the weight as output.

[0092] 503) Select a suitable loss function. Here we use the mean absolute error as the loss function:

[0093]

[0094] 504) completes model training and model prediction. Use the first 26 dimensions of the similar data set obtained in 4) as input, and the 27th dimension, i.e., the post-weighing weight, as output to train the local model. After completing the model training, use the first 26 dimensions of the real-time data as input to predict the value of the post-weighing weight.

[0095] like Figure 5 The figure shows the post-weighing weight prediction result of the instant learning model of the present invention, wherein the root mean square error RMSE between the predicted value and the true value is 6.33, the mean absolute error MAE is 5.01, and the mean absolute percentage error MAPE is 0.26%.

[0096] like Figure 6 As shown in the figure, it is a comparison result diagram of this model and other models. Figure 6 (a) is a comparison result diagram of the present invention and the global prediction models ANN and LSTM. It can be seen that the prediction results of the instant learning model have a greatly improved effect compared with the global modeling results of ANN and LSTM; Figure 6(b) is a comparison result diagram of the present invention, JITL-ANN, and WSJITL-ANN. It can be seen that the prediction result of the proposed network architecture has the best stability and is closest to the true value of the predicted data, which can judge the effectiveness of local modeling; Figure 6 (c) is a box plot comparing the present invention with other models. The yellow short line in the box plot represents the median of the prediction error, and the green dotted line represents the mean of the prediction error. It can be seen that the median error of the model proposed by the present invention is closest to 0, and the average error is the smallest, which is basically 0. The black rectangular box represents the data with a prediction error within 50% of the overall data. It can also be seen that the prediction error of this model is relatively small overall. Table 1 is a quantitative description table of the comparison results of the present invention with other models. From RMSE, R 2 , MAE, and Train Time, the prediction effects of different models are compared. The results show that the errors of global modeling are relatively large. The effectiveness of the weighted similarity analysis proposed by the present invention can be seen through JITL-ANN and WS-JITL-ANN, which is significantly improved compared to the unweighted method; by comparing WS-JITL-ANN and Proposed, the effectiveness of the network model used in the present invention can be compared, and not only the accuracy is improved, but also the detection efficiency is improved; KWS-JITL-FBLS is a new instant learning network structure model. By comparing with the latest instant learning model, the method of the present invention also has a significant improvement effect. Therefore, through comprehensive comparison, the effectiveness of the quality monitoring model of the present invention is verified.

[0097] Table 1 Quantitative description of the comparison results of the present invention and other models

[0098]

[0099] The present invention has the following advantages:

[0100] (1) The present invention collects the historical data set of rubber tire extrusion production, derives the weight factor that can reduce the influence of the physical dimension of data of different dimensions, and proposes a weighted Euclidean distance similarity analysis method. The new similarity analysis method is used to select data to construct a similar data set, which not only improves the quality of the selected data, but also speeds up the efficiency of similarity analysis.

[0101] (2) The present invention determines the most appropriate data set size by referencing the Pearson correlation coefficient to measure the correlation between different size subsets of the initial similar data set, thereby improving the detection accuracy.

[0102] (3) The present invention uses a multi-head attention mechanism to perform post-weighing weight prediction, which can monitor quality indicators in real time, help reduce measurement lag, and facilitate production parameter adjustment.

[0103] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. An intelligent prediction method for the post-weighing weight of a tread extrusion production line, characterized in that: The method comprises the following steps: 1) Obtain historical data of the rubber tire extruder production process and real-time data for model prediction; 2) Using the weighted Euclidean distance similarity analysis method, a certain amount of similar data is obtained based on historical data and real-time data to construct a preliminary similarity data set; 3) using the Pearson correlation coefficient, determining the size of the similar data set with the highest correlation according to the correlation of the different size subsets of the preliminary similar data set, and taking the subset of the similar data set size with the highest correlation as the current similar data set; 4) Use the isolation forest algorithm to remove outliers from the current similar data set to obtain a training set; 5) Based on the multi-head attention mechanism, a local model of instant learning is constructed, and the model is trained based on the training set to obtain a post-weighing weight prediction model. The actual real-time data is input into the prediction model to obtain the post-weighing weight prediction results of the products of the rubber tire extrusion production line.

2. The intelligent prediction method for the post-weighing weight of a tread extrusion production line according to claim 1, characterized in that: 2) The specific steps are: 201) Obtaining the Euclidean distance calculation formula; 202) Mean centering of historical data; 203) Perform dimensionality reduction projection to calculate the variance of each data point of the historical data after mean centering projected to the low-dimensional coordinates; 204) maximize the value of the variance and solve the weight factor using the Lagrange multiplier method; 205) calculating weighted Euclidean distance similarity based on the weight factor and the Euclidean distance calculation formula; 206) According to the real-time data, a certain number of data with the greatest similarity are selected from the historical data through weighted Euclidean distance similarity to construct a preliminary similarity data set.

3. The intelligent prediction method for the post-weighing weight of a tread extrusion production line according to claim 1, characterized in that: 202), the historical data is represented by X = [x1, x2, ..., x m ], where m represents the total number of historical data. The historical data is m n-dimensional data. The historical data after mean centering is: Where i represents the sequence number of historical data.

4. The intelligent prediction method for the post-weighing weight of a tread extrusion production line according to claim 3, characterized in that: The variance is: Among them, ω is the weight factor and Var represents the variance.

5. The intelligent prediction method for the post-weighing weight of a tread extrusion production line according to claim 4, characterized in that: The weighted Euclidean distance similarity is: Among them, x i Represents various historical data, y j represents each real-time data, k represents the serial number of the dimension, x ik For data x i The kth dimension of y jk For data y j The kth dimension of i Indicates the weight factor of the i-th dimension data.

6. The intelligent prediction method for the post-weighing weight of a tread extrusion production line according to claim 5, characterized in that: The specific steps of 3) include: 301) obtaining subsets of different sizes of preliminary similar data sets; 302) Calculate the correlation between subsets of different sizes and real-time data through the Pearson correlation coefficient and select the subset with the largest correlation. The size of this subset is the most appropriate size of the similar data set; 303) The subset corresponding to the most appropriate similarity data set size is used as the current similarity data set.

7. The intelligent prediction method for the post-weighing weight of a tread extrusion production line according to claim 1, characterized in that: The specific steps of 4) are: 401) Construct an isolation forest algorithm model; 402) using the post-weighing weight of the current similar data set as input to train the isolation forest algorithm model to obtain a trained algorithm model; 403) Use the trained algorithm model to remove abnormal data from the current similar data set to obtain a training set.

8. The intelligent prediction method for the post-weighing weight of a tread extrusion production line according to claim 1, characterized in that: The specific steps of 5) are: 501) Construct a multi-head attention mechanism network structure model, that is, a local model; 502) Determine the dimensions of the input and use the post-weighing weight as output; 503) Select the mean absolute error loss function as the loss function of the model; 504) The model is trained based on the training set to obtain a weighing weight prediction model, and actual real-time data is input into the prediction model to obtain a weighing weight prediction result of the product of the rubber tire extrusion production line.

9. The intelligent prediction method for the post-weighing weight of a tread extrusion production line according to claim 8, characterized in that: The loss function is: Among them, y i refers to the true value, f(x i ) refers to the predicted value.

10. The intelligent prediction method for the post-weighing weight of a tread extrusion production line according to claim 9, characterized in that: The network structure of the local model only contains three linear layers.

Images