A Data-Driven Method for Constructing a Tread Quality Prediction Model

Abstract

This invention relates to the technical field of product quality prediction in process industries, specifically to a method for constructing a tread product quality prediction model in the tire manufacturing industry. The method includes the following steps: S1: Collecting and organizing multi-source heterogeneous tread production process data and product quality inspection data to establish a structured dataset with a unified format; S2: Using the isolated forest algorithm to detect and remove outliers in the data; S3: Completing sensor time calibration alignment based on the dynamic characteristics of the production process and a numerical sensor time calibration method using timestamps; S4: Extracting features from the tread time-series data; S5: Decomposing the feature set and feeding it into an improved LSINet model to establish a tread quality prediction model, obtaining the final tread quality prediction result. This invention provides a method for analyzing tread quality in the tire manufacturing industry, helping companies trace quality problems of non-conforming products and improve tread product quality and production efficiency.

Description

Technical Field

[0001] This invention relates to the technical field of product quality prediction in process industries, specifically to a method for constructing a tread product quality prediction model in the tire manufacturing industry. Background Technology

[0002] The manufacturing industry has consistently pursued the goals of improving product quality, reducing production costs, and increasing production efficiency. Product quality issues typically originate in the manufacturing process. However, with continuous advancements in production technology and increasingly refined market demands, the manufacturing process has become increasingly complex, leading to a multitude of factors influencing product quality. As information technology infrastructure develops in the manufacturing sector, a vast amount of product-related data, including component design, production process, and quality inspection data, is accumulating in enterprise databases. However, many companies limit themselves to basic statistical analysis and visualization of this data, failing to fully leverage its potential value.

[0003] In traditional quality control processes, data is abundant but information is scarce, and many companies are unable to extract the valuable knowledge behind the data. The rise of next-generation data analytics technologies has opened up new possibilities for quality prediction. These technologies include machine learning, data mining, and artificial intelligence, which can extract useful information from massive datasets, revealing correlations, trends, and hidden knowledge, thereby providing manufacturing companies with more accurate and real-time product quality predictions.

[0004] Innovative methods for product quality forecasting not only help companies identify potential problems promptly and reduce defect rates, but also improve production efficiency, reduce costs, and thus enhance their product quality management capabilities. This makes product quality forecasting an important tool in modern manufacturing, helping companies provide higher-quality products, meet customer needs, and maintain competitiveness in a highly competitive market.

[0005] Tires, as a crucial component of automobiles, are vital to vehicle performance and safety. The tread, a major component of the tire, directly impacts its performance and lifespan. Therefore, the product quality and production efficiency of the tread production line significantly influence the entire tire manufacturing process. However, in actual tread production, the formulation of control strategies for group equipment and the setting of actual process operating parameters still largely rely on operator experience. To compensate for insufficient operator experience and inaccurate instructions, there is an urgent need for a method that can collect, process, analyze, and mine historical production data to establish a tread quality prediction model. Summary of the Invention

[0006] The main purpose of this invention is to provide the tire manufacturing industry with an analytical method for tire tread quality, helping companies trace the quality problems of non-conforming products and identify possible production influencing factors.

[0007] Technical solution of the present invention:

[0008] A data-driven method for constructing a tire tread quality prediction model includes the following steps:

[0009] S1: Collect and organize multi-source heterogeneous tire tread production process data and product quality inspection data, and establish a structured dataset with a unified format.

[0010] Multi-source heterogeneous data includes manufacturing execution system data, sensor online detection data, and production raw material report data. These data from different sources and formats need to be integrated and cleaned to ensure consistency and accuracy.

[0011] S2: There are outliers with large fluctuations in the tire tread production process. The Isolation Forest algorithm is used to detect and remove outliers from the data.

[0012] S3: To address the asynchronous acquisition frequency of multiple sensors, a numerical sensor time calibration method based on the dynamic characteristics of the production process and timestamps is proposed to complete sensor time calibration and alignment.

[0013] S4: Explore the potential relationship between tread quality and production factors, and then extract features from tread time series data.

[0014] Tire tread groove data is a typical example of time series data. Time series data refers to the results of observing a process at a given sampling rate over equally spaced time periods. The core of time series analysis is to discover patterns in the data and predict future value based on historical observations, providing reference and basis for decision-making. Utilizing industry background knowledge and experience, this study analyzes the potential relationship between tread quality and production factors, and then extracts a set of features from the time series data.

[0015] S5: Decompose the feature set and feed it into the improved LSINet model to establish a tread quality prediction model and obtain the final tread quality prediction result.

[0016] LSTNet (Long-and-Short-term Time-series network) combines the features of Long Short-Term Memory (LSTM) networks and convolutional neural networks, making it a deep learning model suitable for time-series data. Using LSTNet as the basic network architecture, tire tread quality data as the output, and the feature set extracted in step S4 as the input, a tire tread quality prediction model is established.

[0017] Furthermore, the training process of the isolated forest algorithm network in step S2 includes the following steps:

[0018] S21: Randomly select a feature from the data and randomly determine a threshold between the minimum and maximum values ​​of that feature.

[0019] S22: Segment the data points into two parts based on the selected features and thresholds. One part contains data points smaller than the threshold, and the other part contains data points larger than the threshold.

[0020] S23: For each set of data points generated in S22, recursively continue steps S21 and S22, each time selecting new features and thresholds to split the data points into two subsets, until a stopping condition is met. The stopping condition can be that the depth of the tree reaches a predetermined maximum depth, or that the number of data points contained in a node is less than a certain threshold.

[0021] S24: Calculate the isolation score for each data point using the tree depth and the position of the leaf node where the data point is located. The isolation score calculation formula is as follows:

[0022]

[0023] In the formula, S(x,n) is the isolation score of data point x, n is the number of data points contained in the node, h(x) is the depth of data point x in the tree, E(h(x)) is the expected path length of data point x at depth h(x), and c(n) is the average path length of the tree.

[0024] S25: Determine which data points are considered outliers by setting a threshold. Outliers are isolated at shallower tree depths and receive lower scores; normal values ​​are isolated at deeper tree depths and receive higher scores. Data points with scores below the threshold are identified as outliers.

[0025] Furthermore, step S3 completes the sensor time calibration and alignment, which specifically includes the following steps:

[0026] S31: Due to the natural shrinkage of rubber during the tire tread production process, the tire tread extrusion line has multiple floating rollers, with the speed of each floating roller decreasing sequentially. Based on the distance between the sensors, the floating roller speed at historical moments, and the speed of the linkage line, the numerical acquisition time deviation range is calculated.

[0027] S32: Based on the characteristics of data fluctuations, fine-grained time calibration is performed through time series shift analysis. The Pearson correlation coefficient is introduced to convert the correlation between the two data sets into a numerical value, using the magnitude of the value to describe the strength of the relationship between the sequences. Within the aforementioned data acquisition time deviation, the time point with the highest correlation coefficient is identified, completing the time calibration.

[0028] Furthermore, the LSTNet model described in step S5 consists of four parts, as follows:

[0029] S51: Convolutional Layer. Convolutional layers can capture local features in time series data through convolution operations, which helps to identify patterns, trends and other important information in the data, transforming the raw time series data into a more informative representation.

[0030] S52: Recurrent layer. Implemented using LSTM functions, it is suitable for processing time series data and can effectively capture the temporal dependencies and sequence patterns of the data. Each LSTM is designed with three thresholds: input gate i... t Forgotten Gate t Output gate o t Cellular states of long-term memory c t Waiting for candidate states stored in long-term memory The three thresholds represent the input features x at the current time. t Short-term memory h from the previous moment t-1 The function is expressed as follows:

[0031]

[0032] S53: Cyclic Skip Layer. Unlike the cyclic layer, the cyclic skip layer has a time-skip component, which increases the time span of the information flow, allowing the model to learn features at different time scales. Due to gradient vanishing, LSTMs often fail to capture very long-term correlations. Cyclic skip layers are needed to utilize periodic patterns and improve the model's prediction accuracy.

[0033] S54: Autoregressive Layer. Since both the convolutional and recursive components are nonlinear, the output size of this model is insensitive to the input size. The input signals in the data are often aperiodic. To address this issue, a linear component that is more sensitive to local data is added to the original model. This linear component employs the classic autoregressive (AR) model. The AR model is described below:

[0034]

[0035] In the formula, This represents the model output, y t-k,i W represents the model input vector. ar It is the autoregressive coefficient, b ar As a distractor, q ar It is the window size of the input matrix.

[0036] The improved LSTNet model performs data transformation during data construction, decomposing the input data into two parts: (1) short-term time series (INPUT1), including the current time point and the previous few adjacent time points (t-3, t-2, t-1, t). (2) long-term skipped time series (INPUT2), including the current time point and the more distant time points (t-3×skip, t-2×skip, t-skip, t);

[0037] The two decomposed data parts are used as model inputs. Short-term features are extracted from the data using one-dimensional convolution kernels, and then LSTM is used to capture the time dependence of the data. For short-term time series, the input dimension is adjusted by Lambda to complete the short-period link jump of the data. Dense is used to simulate the AR autoregressive process, adding a linear component to the prediction, while making the output respond to the scale change of the input.

[0038] Compared with the prior art, the present invention has the following advantages:

[0039] This invention introduces the Isolation Forest algorithm to effectively detect and remove outliers in the data, reducing the adverse effects of imbalanced data on the product quality prediction model. Addressing the asynchronous nature of sensor acquisition frequencies, this invention performs time calibration to ensure the timeliness and accuracy of the data. Simultaneously, this invention uses the LSTNet network model as its basic architecture, combining industry background knowledge and experience to analyze product quality influencing factors from tire tread production time-series data, thereby constructing a data-driven tire tread quality prediction model. This helps to trace quality issues of non-conforming products and improve tire tread product quality and production efficiency. Attached Figure Description

[0040] Figure 1 This diagram illustrates the method flowchart of the present invention.

[0041] Figure 2 This diagram illustrates the LSTNet model framework constructed according to an embodiment of the present invention.

[0042] Figure 3 shows a schematic diagram of the prediction results of the model for product quality in an embodiment of the present invention. Detailed Implementation

[0043] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0044] By analyzing the time-series characteristics of tire tread production line data, a data-driven method for constructing a tire tread quality prediction model is designed and specifically applied to tire tread quality prediction in tire tread production lines, such as... Figure 1 This includes the following steps:

[0045] S1: Collect and organize multi-source, heterogeneous tire tread production process data and product quality inspection data to establish a structured dataset with a unified format. This multi-source, heterogeneous data includes manufacturing execution system data, online sensor monitoring data, and raw material production report data. These data from different sources and formats need to be integrated and cleaned to ensure consistency and accuracy.

[0046] S2: Due to the existence of outliers with large fluctuations in the tire tread production process, the Isolation Forest algorithm is used to detect and remove outliers in the data.

[0047] S3: To address the asynchronous nature of sensor acquisition frequencies, a numerical sensor time calibration method based on the dynamic characteristics of the production process and timestamps is proposed.

[0048] S4: Utilize industry background knowledge and experience to analyze the potential relationship between tire tread quality and production factors, and extract feature sets from time series data.

[0049] The production factors affecting tread quality include raw material quality and production process parameters. Rubber, as the raw material for tread production, is affected by many factors that influence its expansion and contraction, such as rubber composition, temperature, humidity, and external stress. These factors determine the quality of the tread. During tread production, various stages are interconnected, resulting in dynamic coupling of production process parameters. Correlation analysis is performed between these parameters and tread quality to identify variables with significant impact on product quality, and a feature set is established for subsequent model building.

[0050] S5: Feed the extracted feature set into the improved LSTNet network to establish a tread quality prediction model.

[0051] In this embodiment, the tread production line dataset contains 15,200 data points. The factors affecting tread quality include the three screw speeds, the three screw currents, the extrusion temperature, the extrusion pressure, the multi-segment linkage speed, the cooling ramp speed, and the cooling ramp speed. The data representing the final tread quality include continuous weighing, sorting scale weight, and front width, totaling 14 dimensions of data.

[0052] Specifically, the training process of the isolated forest network described in step S2 includes the following steps:

[0053] S21: Randomly select a feature from the tread dataset and randomly determine a threshold between the minimum and maximum values ​​of that feature.

[0054] S22: Segment the data points into two parts based on the selected features and thresholds. One part contains data points smaller than the threshold, and the other part contains data points larger than the threshold.

[0055] S23: Recursively continue steps S21 and S22, each time selecting new features and thresholds to split the data points into two subsets, until a certain stopping condition is reached. The stopping condition may be that the depth of the tree reaches a predetermined maximum depth, or that the number of data points contained in a node is less than a certain threshold.

[0056] S24: Calculate the isolation score for each data point using the tree depth and the position of the leaf node where the data point is located. The isolation score calculation formula is as follows:

[0057]

[0058] In the formula, S(x,n) is the isolation score of data point x, n is the number of data points contained in the node, h(x) is the depth of data point x in the tree, E(h(x)) is the expected path length of data point x at depth h(x), and c(n) is the average path length of the tree.

[0059] S25: Determine which data points are considered outliers by setting a threshold. Outliers are isolated at shallower tree depths and receive lower scores; normal values ​​are isolated at deeper tree depths and receive higher scores. Data points with scores below the threshold are identified as outliers.

[0060] Specifically, step S3, sensor time calibration alignment, includes the following two steps:

[0061] S31: Due to the natural shrinkage of the rubber compound during the tread production process, the tread extrusion line has multiple floating rollers, with the speed of each floating roller decreasing sequentially. Let the speed of the floating roller be V. s =[v s,1 ,v s,2 ,…,v s,n The linear velocity of the linkage is V. r =[v r,1 ,v r,2 ,…,v r,n Based on the distance L between the sensors (where the distance between the floating roller segments is M), the floating roller speed V at historical moments is calculated. s and the linear velocity V r The numerical acquisition time deviation T is calculated using the following formula:

[0062]

[0063] The tread extrusion process is subject to speed fluctuation errors and sample accumulation during production. The range of the acquisition time deviation T can be calculated using the above formula [T...]. min ,T max Preliminary time calibration is performed within this range.

[0064] S32: Based on the characteristics of data fluctuations, fine-grained time calibration is performed through time series shift analysis. The Pearson correlation coefficient is introduced to convert the correlation between the two data sets into a numerical value, using the magnitude of the value to describe the strength of the relationship between the series. The formula is as follows:

[0065]

[0066] In the formula, X and Y represent two sample data, and n represents the number of samples. and This represents the mean of two sample data.

[0067] Assume the sampling time point for the two data sets is T. a =[t a,1 ,t a,2 ,…,t a,n ] and T b =[t b,1 ,t b,2 ,…,t b,n The collected data sample X(T) a )=[x(t a,1 ),x(t a,2 ),…,x(t a,n )] and Y(T b )=[y(t b,1 ),y(t b,2 ),…,y(t b,n )], X(T a Given a fixed time, let ΔT = [Δt1, Δt2, ..., Δt] n ], ΔT∈[T min ,T max ]. Then Y(T) b +ΔT) is Y(T) b At time point T b The data obtained by shifting upwards by ΔT. After aligning the two data sets in time, the correlation should be maximized. Therefore, the problem is transformed into finding:

[0068]

[0069] By using the above sensor time calibration and alignment method, the time of the four samples (corresponding to Y in the above formula) of the screw speed acquisition point, front width measurement point, continuous weighing measurement point, and sorting scale measurement point of the tire tread production line are all aligned to the screw speed acquisition point (corresponding to X in the above formula), thus eliminating the time lag of the data samples.

[0070] Specifically, step S5, the LSTNet model, consists of four parts, as follows:

[0071] S51: Convolutional Layer. Convolutional layers can capture local features in time series data through convolution operations, which helps to identify patterns, trends and other important information in the data, transforming the raw time series data into a more informative representation.

[0072] S52: Recurrent layer. Implemented using LSTM functions, it is suitable for processing time series data and can effectively capture the temporal dependencies and sequence patterns of the data. Each LSTM is designed with three thresholds: input gate i... t Forgotten Gate t Output gate o t Cellular states of long-term memory c t Waiting for candidate states stored in long-term memory The three thresholds represent the input features x at the current time. t Short-term memory h from the previous moment t-1 The function is expressed as follows:

[0073]

[0074] S53: Cyclic Skip Layer. Unlike the cyclic layer, the cyclic skip layer has a time-skip component, which increases the time span of the information flow, allowing the model to learn features at different time scales. Due to gradient vanishing, LSTMs often fail to capture very long-term correlations. Cyclic skip layers are needed to utilize periodic patterns and improve the model's prediction accuracy.

[0075] S54: Autoregressive Layer. Since both the convolutional and recursive components are nonlinear, the output size of this model is insensitive to the input size. The input signals in the data are often aperiodic. To address this issue, a linear component that is more sensitive to local data is added to the original model. This linear component employs the classic autoregressive (AR) model. The AR model is described below:

[0076]

[0077] In the formula, This represents the model output, y t-k,i W represents the model input vector. ar It is the autoregressive coefficient, b ar As a distractor, q ar It is the window size of the input matrix.

[0078] In this embodiment, the tire tread dataset samples are divided into training, validation, and test sets in a 7:2:1 ratio. The variable data undergoes min-max normalization and inverse normalization before evaluating the prediction results. The model is trained on the training set, its hyperparameters are tuned on the validation set, and the prediction results are obtained using the test set.

[0079] In this embodiment, the time interval of the tire tread time series dataset is relatively short. For the LSTNet model, it needs to pass a large tensor as input to the model. The implementation of the loop skip layer requires a large number of data slices, which leads to a significant increase in computational complexity.

[0080] Specifically, the original LSTNet model uses only a single one-dimensional convolution to process the data before performing data transformation. The improved LSTNet proposed in this invention is as follows: Figure 2 :

[0081] When constructing the data, data transformation is performed, and the input data is decomposed into two parts: (1) short-term time series (INPUT1), including the current time point and the previous few adjacent time points (t-3, t-2, t-1, t). (2) long-term skipped time series (INPUT2), including the current time point and the more distant time points (t-3×skip, t-2×skip, t-skip, t).

[0082] The decomposed data is used as the model input. Short-term features are extracted from the data using one-dimensional convolutional kernels, and then LSTM is used to capture the temporal dependencies of the data. For short-term time series, the input dimension is adjusted using Lambda to complete short-cycle link jumps in the data. Dense is used to simulate the AR autoregressive process, adding a linear component to the prediction, while making the output responsive to scale changes in the input. The improved LSTNet avoids the large number of data slices in the loop-skipping layers. For the short-interval tread time series dataset in this example, it can better capture the correlation between time series data and improve the accuracy of the tread quality prediction model.

[0083] Using the tread product quality prediction model constructed in this embodiment, the continuous weighing weight, sorting scale weight, and front width of the tread were predicted. The experimental results are shown in Figures 3(a), (b), and (c). By adopting the method proposed in this invention, the model can effectively learn the potential variation patterns in time series data, thereby achieving high accuracy in predicting tread quality. This helps to trace the quality problems of non-conforming products and improve the quality of tire tread products and production efficiency.

[0084] The above description is merely a description of embodiments of the present invention and is not intended to limit the scope of the invention in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered equivalent and valid embodiments, and all fall within the scope of the present invention.

Claims

1. A data-driven method for constructing a tread quality prediction model, characterized in that, Includes the following steps: S1: Collect and organize multi-source heterogeneous tire tread production process data and product quality inspection data, and establish a structured dataset with a unified format; Multi-source heterogeneous data includes manufacturing execution system data, online sensor detection data, and production raw material report data. These data from different sources and formats are integrated and cleaned to ensure consistency and accuracy. S2: There are outliers with large fluctuations in the tire tread production process. The Isolation Forest algorithm is used to detect and remove outliers from the data. S3: To address the asynchronous acquisition frequency of multiple sensors, a numerical sensor time calibration method based on the dynamic characteristics of the production process and timestamps is proposed to complete sensor time calibration and alignment. S4: Explore the potential relationship between tread quality and production factors, and then extract features from tread time series data; Tread extrusion line data is a typical time series data. By using industry background knowledge and experience, we can analyze the potential relationship between tread quality and production factors, and then extract feature sets from the time series data. S5: Decompose the feature set and feed it into the improved LSTNet model to establish a tread quality prediction model and obtain the final tread quality prediction result. The improved LSTNet model performs data transformation during data construction, decomposing the input data into two parts: (1) short-term time series, including the current time point and the previous few adjacent time points (t-3, t-2, t-1, t); (2) long-term skipped time series, including the current time point and the more distant time points (t-3, t-2, t-1, t). skip, t-2 skip, t-skip, t); The two decomposed data parts are used as model inputs. Short-term features are extracted from the data using one-dimensional convolution kernels, and then LSTM is used to capture the time dependence of the data. For short-term time series, the input dimension is adjusted by Lambda to complete the short-period link jump of the data. Dense is used to simulate the AR autoregressive process, adding a linear component to the prediction, while making the output respond to the scale change of the input.

2. The method for constructing a data-driven tread quality prediction model as described in claim 1, characterized in that, The training process of the isolated forest algorithm network described in step S2 includes the following steps: S21: Randomly select a feature from the data, and randomly determine a threshold between the minimum and maximum values ​​of that feature; S22: Segment the data points according to the selected features and thresholds, dividing the data points into two parts: one part contains data points less than the threshold and the other part contains data points greater than the threshold. S23: For each set of data points generated in S22, recursively continue steps S21 and S22, each time selecting new features and thresholds to split the data points into two subsets until a stopping condition is reached; the stopping condition may be that the depth of the tree reaches a predetermined maximum depth, or that the number of data points contained in a node is less than a certain threshold. S24: Calculate the isolation score for each data point using the tree depth and the position of the leaf node where the data point is located; the isolation score calculation formula is as follows: In the formula, Data points The isolation score, It is the number of data points contained in a node. Data points At depth in the tree, Data points In depth The expected path length, It is the average path length of the tree; S25: A threshold is set to determine which data points are considered outliers; outliers are isolated at shallower tree depths and receive lower scores; normal values ​​are isolated at deeper tree depths and receive higher scores; data points with scores below the threshold are identified as outliers.

3. The method for constructing a data-driven tread quality prediction model as described in claim 1, characterized in that, Step S3 completes the sensor time calibration and alignment, which includes the following steps: S31: Based on the distance between the sensors, the floating roller speed and the linkage line speed at historical moments, calculate the numerical acquisition time deviation range; S32: Fine-grained time calibration is performed by shifting the time series data; the Pearson correlation coefficient is introduced to convert the correlation between the two sets of data into a numerical value, and the magnitude of the value describes the degree of connection between the sequences; within the above-mentioned time deviation, the time point with the largest correlation coefficient is selected to complete the time calibration.

4. The method for constructing a data-driven tread quality prediction model as described in claim 1, characterized in that, The LSTNet model described in step S5 consists of four parts, as follows: S51: Convolutional layer: Convolutional layers can capture local features in time series data through convolution operations, which helps to identify patterns, trends and other important information in the data, transforming the original time series data into a more informative representation. S52: Recurrent Layer: Implemented using LSTM functions, suitable for processing time series data, effectively capturing the time dependencies and sequence patterns of the data; each LSTM is designed with three thresholds: input gate... Forgotten Gate Output gate Cellular states of long-term memory Waiting for candidate states stored in long-term memory ; The three thresholds represent the input features at the current time. Short-term memory of the previous moment The function is expressed as follows: S53: Cyclic Skip Layer: The cyclic skip layer has a time skip component, which increases the time span of the information flow and allows the model to learn features at different time scales; due to gradient vanishing, LSTM usually cannot capture very long-term correlations, and cyclic skip layers are needed to utilize periodic patterns to improve the model's prediction accuracy. S54: Autoregressive Layer: A linear component that is more sensitive to local data is added to the original model. This linear component adopts the classic autoregressive model, as follows: In the formula, This indicates the model output result. Represents the model input vector. It is the autoregressive coefficient. As a distractor, It is the window size of the input matrix.

Images