A tire calendering data fusion method based on Kalman filtering

Abstract

The application discloses a kind of based on Kalman filtering's tire calendering data fusion method, it includes: S1, utilize multiple linear regression method to establish tire calendering thickness model;S2, to obtain data from different sensing measurement equipment, realize the data cleaning work of multi-source data set;S3, carry out time sequence arrangement and constitute sequence data, realize the feature extraction work of complex tire calendering multi-source data by LSTM algorithm;S4, realize the evaluation of tire calendering process precision, by random forest algorithm Decision tree is constructed, data is classified, obtains observation weight value and attribute decision result;S5, according to the attribute decision result of random forest respectively carries out attribute association and data fusion, obtains the feature and estimated value of tire calendering data fusion.The application according to the important factor of influencing calendering machine thickness, consider data missing data anomaly and other problems, constructs multiple heterogeneous tire calendering algorithm, improves the thickness quality of tire, reduces the device operating energy consumption.

Description

Technical Field

[0001] This invention relates to a tire calendering data fusion method based on Kalman filtering, belonging to the field of tire industry automation and complex industrial process data fusion technology. Background Technology

[0002] Tire calendering is a key process affecting the overall quality of tires. Steel cord calendering is an important step in the production of all-steel radial tires, directly impacting their intrinsic quality. Currently, this process in China has long suffered from batch complexity and difficulty in coordinating processes. The mainstream calendering methods in industry mainly include four types: hot pressing, cold pressing, extrusion, and coating. Hot pressing involves passing tire strip material through the rollers of a hot press, where the temperature and pressure are controlled to ensure tire consistency. Cold pressing refers to preparing the rubber material for tire manufacturing without heating it to high temperatures; it is directly prepared in sheet or strip form for use in various parts of the tire. Extrusion involves passing a rubber compound through an extruder and then forming it into a tire shape through an extrusion head. Coating involves coating rubber onto tire strip material and then calendering it using pressure rollers.

[0003] During the hot pressing process, due to frequent batches, temperature changes, and varying roller distances, the calendered tire rubber material has long suffered from uneven thickness control at different locations and low batch yield. Since the tire calendering process involves complex parameters and incomplete or occasionally missing data from different sensors, it is of great significance to develop a multi-modal data fusion algorithm system for tire calendering in order to better obtain comprehensive and multi-dimensional information and improve the accuracy and yield of the calendering process for all-steel radial tires.

[0004] Therefore, a new technology is needed to solve at least some of the above-mentioned problems. Summary of the Invention

[0005] The purpose of this invention is to provide a tire calendering data fusion method based on Kalman filtering to improve tire thickness quality and reduce the operating energy consumption of the device.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is: a tire calendering data fusion method based on Kalman filtering, which includes the following steps:

[0007] S1. Based on the factors affecting the calendering process, a tire calendering thickness model is established using the multiple linear regression method.

[0008] S2. Based on the tire calendering thickness model, the data from different sensing and measurement devices are first obtained. The missing data is estimated multiple times using the multiple interpolation method. Each estimation introduces uncertainty. The missing data is filled in and the redundant data is deleted to achieve data cleaning of the multi-source dataset.

[0009] S3. After the tire calendering data has been cleaned, it is arranged into a sequence data by time sequence. Then, by introducing a gating mechanism, an LSTM multi-source data feature extraction model is constructed to store and forget the features of the multi-source tire calendering data. The LSTM algorithm is used to realize the feature extraction of complex multi-source tire calendering data.

[0010] S4. Based on the feature extraction data using LSTM, the accuracy of tire calendering process is evaluated. A decision tree is constructed using the random forest algorithm to classify the data and obtain the observation weight values ​​and attribute judgment results.

[0011] S5. Based on the attribute decision results of the random forest, perform attribute association and data fusion respectively. Use the output of the random forest attribute decision as the observation weight. Update and adjust the system's state estimate by initializing the Kalman filter, setting the initial state and covariance matrix parameters. If the consistency test can be passed, perform the final tire calendering data fusion. If not, repeat the process iteratively to finally obtain the features and estimates of the tire calendering data fusion.

[0012] Furthermore, in step S1, establishing the tire calendering thickness model includes:

[0013] S1-1. Assume there are n sample data related to tire calendering, each sample has k independent variables, and the model is expressed as:

[0014] Y = β0 + β1X1 + β2X2 + ... + β k X k +ε,

[0015] Where Y is the tire operating side and drive side thickness matrix, X1, X2, ..., X k This is a matrix of independent variables affecting the calendering thickness of tires, where β0 is the intercept, and β1, β2, ..., β k ε is the coefficient of the independent variable, and ε is the error term;

[0016] S1-2. The model parameters are estimated using the proposed calendering model. The sum of squared residuals between the predicted values ​​and the predicted values ​​is minimized using the least squares method. The formula for the least squares method is as follows:

[0017]

[0018] Where y iThese are observations from the rolling mill model. It is the model's prediction for the i-th sample.

[0019] Furthermore, in step S1, the factors affecting the calendering process include: the calender roll gap, the calender outlet temperature, the speed of the extrusion and pick-up conveyor belt, performance inspection indicators, extruder temperature, the thickness of the upper and lower sheet drive side and operating side, and the drive side balance.

[0020] Furthermore, in step S2, the data interpolation processing method for data from multi-source tire calendering includes:

[0021] Based on nine datasets from different sensing devices, the missing data portions were analyzed using Lagrange multiplication. For a given n+1 (n=9) data points (x0,y0), (x1,y1),...,(x... n ,y n The corresponding data points are represented as follows: (calender roll distance, corresponding measured value), (calender operating side outlet temperature, corresponding measured value), (calender drive side outlet temperature, corresponding measured value), (calender conveyor belt speed, corresponding measured value), (performance test index, corresponding measured value), (extruder temperature, corresponding measured value), (operating side thickness, corresponding measured value), (drive side thickness, corresponding measured value), (drive side balance, corresponding measured value). At this time, the Lagrange interpolation polynomial is:

[0022]

[0023]

[0024] Where P(x) is a pair of y with respect to tire calendering data types x1, x2, ..., x n The estimated polynomial, It is a Lagrange basis function used to weight each attribute, x = (x1, x2, ..., x...). n The expression represents the multidimensional attribute values ​​to be estimated.

[0025] Furthermore, in step S3, the LSTM-based tire calendering feature extraction algorithm comprises three parts: an input gate, an output gate, and a forget gate. The forget gate filters various calendering data states through an update function, retaining high-weight states and removing low-weight states. The input gate is responsible for determining the weight value of the input information and inputting the data with high weight into the state update module. This update process is divided into two parts: the first part updates various calendering data states through a function, and the second part uses an inverse trigonometric function to generate new variables.

[0026] Furthermore, the update function for the forget gate is:

[0027] f t =δ(w f ·[h t-1 ,x t ]+b f )

[0028] Where w f b f h represents the weight and paranoia value of the forgetting gate. t-1 If δ is the output value of the output gate at the previous time step, then δ is the activation function.

[0029] Furthermore, the function of the input gate is:

[0030] f t =δ(w i ·[h t-1 ,x t ]+b i ),

[0031]

[0032] Where w i b i w represents the weights and biases of the input gate. c b c Representing the state weight value and the paranoia value; the cell state update formula is shown below:

[0033]

[0034] c t The cell represents the current cellular state, is responsible for remembering long-term information, and its updates depend on the forgetting gate f. t and the output i of the input gate t ;

[0035] The output formula of the output gate is:

[0036] O t =δ(w0·[h t-1 ,x t ]+b0)

[0037] Among them O t The output of the output gate, w0 is the weight matrix of the output gate, and b0 is the bias term of the output gate.

[0038] Furthermore, in step S4, the attribute determination of the random forest algorithm on the multi-source tire data includes:

[0039] S4-1. Resample the dataset after feature extraction using the LSTM algorithm, randomly select N subsets, generate m decision trees, and the unselected data are called idle data.

[0040] S4-2. Use decision trees to poll and train multiple datasets for tire calendering process, calculate the weight value of each subset, and sort them.

[0041] S4-3. Calculate the weight value of each decision tree, and combine the weights of the subset datasets in multiple rounds. The method is as follows:

[0042]

[0043] Where, is a j For the decision tree weights, A ij B represents the weight value of the subset. i This represents the variation in the random forest.

[0044] Furthermore, in step S5, data fusion using the Kalman filter algorithm includes:

[0045] S5-1. Using attribute decisions based on LTSM and random forest, the tire data classification results are considered as a preliminary estimate of the system state. First, the state vector is defined as x, including multiple variables such as the calender receiving machine line speed, the distance between each roller, and the roller temperature. Simultaneously, based on the system's dynamic characteristics, a state transition equation is established to describe the evolution of the given state vectors at different times, expressed as:

[0046]

[0047] P k =AP k-1 A T +Q

[0048] in Let be the posterior state estimate of the tire data at time k. Let A be the posterior state estimate of the tire data at time k-1, A be the desired state transition matrix, B be the input matrix, and u be the value of the posterior state. k-1 For external input, w k It is the process noise at time k-1, P k and P k-1 Let represent the posterior estimation covariance at time k and time k-1, respectively, and let Q represent the process excitation noise covariance, which is used to represent the error between the state transition matrix and the actual tire calendering process.

[0049] S5-2. Define the observation equation based on the state transition equation, and map the classification results output by the LSTM and Random Forest algorithms to the state space, as follows:

[0050] z k =H*x k +v k

[0051] Where H is the transformation matrix from state variables to measurements, representing the relationship connecting the state and the observation, v k It is observation noise;

[0052] Then, the Kalman gain of the system is calculated, expressed as follows:

[0053]

[0054] Where K k This represents the Kalman filter gain, and is an intermediate calculation result of the filter. R is the prior estimate covariance at time k, and R is the covariance of the observation noise.

[0055] S5-3. Based on the calculation results, the state of the tire calendering prediction data is updated to the calibrated state using Kalman gain. Prediction is achieved through a time update process, that is, the state at the current time (k) is estimated based on the posterior estimate of the previous time step (k-1), expressed as:

[0056]

[0057] Among them, z k Indicates the measured value;

[0058] The error covariance is then updated, and finally the required fusion weights and estimates are output based on the consistency test results.

[0059] Furthermore, the fusion algorithm predicts and outputs estimates by assigning different weights to each data module. These weights are determined in each individual Kalman filter process by calculating the proportion of each sensor error covariance to the sum of the nine sensor error covariances. Let the orthometric errors of the data from the nine different sensors be ρ1, ρ2, ρ3, ρ4, ρ5, ρ6, ρ7, ρ8, and ρ9. Then the allocation coefficients for the nine data types are as follows:

[0060]

[0061]

[0062]

[0063]

[0064] The beneficial effects of this invention are as follows: This invention proposes a tire calendering data fusion method based on Kalman filtering. It collects multi-source radial tire calendering process data through various sensing devices such as eddy current thickness measurement devices, point laser thickness measurement devices, and temperature sensors. Based on the important factors affecting the thickness of the calender, and considering issues such as missing data and data anomalies, a multi-dimensional heterogeneous tire calendering algorithm is constructed to improve the thickness quality of the tire and reduce the operating energy consumption of the device. Attached Figure Description

[0065] Figure 1 This is a data fusion framework diagram of the present invention.

[0066] Figure 2 This is a thickness modeling diagram of the radial tire calendering process of the present invention. At point 1, two sensors measure the thickness of the four-roll shaft intersection and the operating side (drive side) of the upper and lower film; at point 2, two sensors measure the thickness of the steel cord fabric on the left and right operating sides (drive side); at point 3, two sensors measure the error of the steel cord fabric on the left and right operating sides (drive side); at point 4, one sensor measures the temperature of the four rolls and the change in distance between the four rolls; at point 5, one sensor measures the calender pick-up speed; and at point 6, one sensor measures C... pk C p C a Performance indicators, etc.

[0067] Figure 3 This is a diagram of the improved LSTM feature extraction network structure of the present invention.

[0068] Figure 4 This is a flowchart illustrating the process of multivariate data fusion using Kalman filtering in this invention.

[0069] Figure 5 This is a flowchart of the overall technology of the present invention. Detailed Implementation

[0070] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0071] like Figure 1-5 A method for fusing tire calendering data based on Kalman filtering includes the following steps:

[0072] S1. Based on the calendering process in tire production, a tire calendering thickness model is established using multiple linear regression based on key factors affecting the calendering process, such as calender roll gap, calender rubber temperature, extrusion conveyor belt speed, extruder temperature, thickness of upper and lower rubber sheets on the drive side and operating side, and roll gap adjustment distance.

[0073] S1-1. Assume there are n sample data related to tire calendering, each sample has k independent variables, and the model is expressed as:

[0074] Y = β0 + β1X1 + β2X2 + ... + β k X k +ε

[0075] Where Y is the dependent variable, X1, X2, ..., X... k The independent variable is β0, the intercept is β1, β2, ..., β k ε is the coefficient of the independent variable, and ε is the error term.

[0076] S1-2. The model parameters are estimated using the proposed rolling mill model. The sum of squared residuals between the predicted values ​​is minimized using the least squares method. The formula for the least squares method is as follows:

[0077]

[0078] Where y i These are observed values. It is the model's prediction for the i-th sample.

[0079] S2. Based on the tire calendering thickness model, data from nine sensing and measuring devices were first acquired, including calender roll distance, calender operating side discharge temperature, calender drive side discharge temperature, calender conveyor belt speed, and performance inspection index C. pk C p C a Data from multiple sources, including extruder temperature, operating side thickness, drive side thickness, and drive side balance, were collected. The Multiple Imputation method was used to estimate the missing data multiple times. Each estimation introduced some uncertainty. The missing data was filled in and redundant data was deleted, thus completing the data cleaning of the multi-source dataset.

[0080] S2-1. Based on nine datasets from different sensing devices, Lagrange interpolation is performed on the missing data. For a given n+1 (n=9) data points, the corresponding data points can be represented as (calender roll distance, corresponding measured value), (calender operating side discharge temperature, corresponding measured value), (calender drive side discharge temperature, corresponding measured value), (calender conveyor belt speed, corresponding measured value), (performance test index, corresponding measured value), (extruder temperature, corresponding measured value), (operating side thickness, corresponding measured value), (drive side thickness, corresponding measured value), (drive side balance, corresponding measured value). The Lagrange interpolation polynomial is then:

[0081]

[0082]

[0083] P(x) is a pair of y values ​​relating to tire calendering data types x1, x2, ..., x n The estimated polynomial. These are Lagrange basis functions used to weight each attribute. x = (x1, x2, ..., x...) n Its expression is the multidimensional attribute value to be estimated. This multidimensional Lagrange interpolation polynomial formula extends the traditional single-variable case and is applicable to the interpolation problem of multi-source data of tire calendering.

[0084] S3. After cleaning the tire calendering data, arrange the data in time sequence to form a sequence that meets the requirements. Then, by introducing a gating mechanism, construct an LSTM multi-source data feature extraction model to store and forget the features of the tire calendering multi-source data. The LSTM algorithm is used to extract features from the complex tire calendering multi-source data.

[0085] S3-1, the LSTM-based feature extraction algorithm mainly consists of three parts: an input gate, an output gate, and a forget gate. The forget gate filters various rolling data states through an update function, retaining high-weight states and removing low-weight states. The input gate is responsible for determining the weight values ​​of the input information and inputting the high-weight data into the state update module. This update process is divided into two parts: the first part updates various rolling data states through a function; the second part uses an inverse trigonometric function to generate new variables. The input gate function includes weight values ​​and bias values, as well as state weight values ​​and bias values. Cell states are updated through specific formulas, and the output of the output gate is determined by a corresponding formula.

[0086] Tire calendering data often involves long time spans, requiring the modeling of long-term dependencies. LSTM introduces mechanisms such as forget gates, which can effectively handle long sequences, helping the model better understand and utilize the long-term dependencies in tire calendering data, thereby improving the accuracy of feature extraction.

[0087] The forget gate primarily selects from various rolled data states, retaining states with high weights and removing states with low weights. The forget gate update function is...

[0088] f t =δ(w f ·[h t-1 ,x t ]+b f )

[0089] Where w f b f h represents the weight and paranoia value of the forgetting gate. t-1 If δ is the output value of the output gate at the previous time step, then δ is the activation function.

[0090] S3-2, the input gate determines the weight value of the input information and inputs the data with the higher weight into the state update module. The update process is divided into two parts: the first part updates the various rolling data states using the δ function; the second part uses the inverse trigonometric function to generate new variables. The function of the input gate is...

[0091] f t =δ(w i ·[h t-1 ,x t ]+b i ),

[0092]

[0093] In representing w i b i The weight and bias values ​​represented by the input gate, w c b c This represents the state weight value and the bias value. This represents a temporary value for cell state updates. The cell state update formula is shown below.

[0094]

[0095] c t The cell represents the current cellular state, is responsible for remembering long-term information, and its updates depend on the forgetting gate f. t and the output i of the input gate t .

[0096] The output formula of the output gate is

[0097] O t =δ(w0·[h t-1 ,x t ]+b0)

[0098] Among them O t The output of the output gate, w0 is the weight matrix of the output gate, and b0 is the bias term of the output gate.

[0099] S4. Based on the feature extraction data using LSTM, the accuracy of tire calendering process is evaluated. A decision tree is constructed using the random forest algorithm to classify the data and obtain the observation weight values ​​and attribute judgment results.

[0100] S4-1. Resample the dataset after feature extraction using the LSTM algorithm, randomly select N subsets, generate m decision trees, and the unselected data are called idle data.

[0101] S4-2. Use decision trees to poll and train multiple datasets of tire calendering process, calculate the weight value of each subset, and sort them.

[0102] S4-3. Calculate the weight value of each decision tree, and combine the weights of the subset dataset several times. The method is as follows:

[0103]

[0104] Where, is a j For the decision tree weights, A ij B represents the weight value of the subset. i This represents the variation in the random forest.

[0105] S5. Based on the attribute decision results of the random forest, perform attribute association and data fusion respectively. Use the output of the random forest attribute decision as the observation weight. Update and adjust the system's state estimate by initializing the Kalman filter, setting the initial state and covariance matrix and other parameters. If the consistency test can be passed, the final tire calendering data fusion is performed. If not, the process is repeated iteratively to finally obtain the features and estimates of the tire calendering data fusion.

[0106] S5-1. Using attribute decisions based on LTSM and random forest, the tire data classification results are considered as a preliminary estimate of the system state. First, the state vector is defined as x, including multiple variables such as the calender receiving machine line speed, the distance between each roller, and the roller temperature. Simultaneously, based on the system's dynamic characteristics, a state transition equation is established to describe the evolution of the given state vectors at different times, expressed as:

[0107]

[0108]

[0109] in Let be the posterior state estimate of the tire data at time k. Let A be the posterior state estimate of the tire data at time k-1, A be the state transition matrix, B be the input matrix, and u be the value of the tire data. k-1 For external input, w k This is the process noise at time k-1. P k and P k-1 Let represent the posterior estimation covariance at time k and time k-1, respectively, and let Q represent the process excitation noise covariance, which is used to represent the error between the state transition matrix and the actual tire calendering process.

[0110] S5-2. Define the observation equation based on the state transition equation, and map the classification results output by the LSTM and Random Forest algorithms to the state space, as follows:

[0111] z k =H*x k +v k

[0112] Where H is the transformation matrix from state variables to measurements (observations), representing the relationship connecting the states and observations, v k This is the observation noise; then the Kalman gain of the system is calculated, expressed as:

[0113]

[0114] Where K k This represents the Kalman filter gain, and is an intermediate calculation result of the filter. R is the prior estimate covariance at time k, and R is the covariance of the observation noise.

[0115] S5-3. Based on the calculation results, the state of the tire calendering prediction data is updated to the calibrated state using Kalman gain. Prediction is achieved through a time update process, that is, the state at the current time (k) is estimated based on the posterior estimate of the previous time step (k-1). The expression is:

[0116]

[0117] Among them, z k Indicates the measured value. This represents the posterior estimate at the previous time step (k-1). This represents the posterior estimate at the current time (k).

[0118] The error covariance is then updated, and its constraint formula is as follows:

[0119]

[0120] Where I is the identity matrix. Let P be the estimation error covariance matrix at the previous time step (k-1). k Let be the estimation error covariance matrix at the current time (k).

[0121] It's important to note that the fusion algorithm predicts and outputs estimates by assigning different weights to each data module. These weights are determined in each individual Kalman filter process by calculating the proportion of each sensor's error covariance to the sum of the error covariances of the nine sensors. Assuming the orthometric errors of the data from the nine different sensors are ρ1, ρ2, ρ3, ρ4, ρ5, ρ6, ρ7, ρ8, and ρ9, then the allocation coefficients for the nine data types are as follows:

[0122]

[0123]

[0124]

[0125]

[0126] In summary, this invention acquires multi-source radial tire calendering process data through various sensing devices such as eddy current thickness measurement devices, point laser thickness measurement devices, and temperature sensors. Based on the important factors affecting the thickness of the calender and considering issues such as missing or abnormal data, a multi-dimensional heterogeneous tire calendering algorithm is constructed to improve tire thickness quality and reduce the operating energy consumption of the device.

[0127] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the scope of protection of the present invention in any way, and all technical solutions obtained by equivalent substitution or other means fall within the scope of protection of the present invention. Parts not covered in this invention are the same as or can be implemented using existing technology.

Claims

1. A method for fusing tire calendering data based on Kalman filtering, characterized in that... Includes the following steps: S1. Based on the factors affecting the calendering process, a tire calendering thickness model is established using the multiple linear regression method; the establishment of the tire calendering thickness model includes: S1-1. Assume there are n sample data related to tire calendering, each sample has k independent variables, and the model is expressed as: ， in It is the thickness matrix of the tire operating side and drive side. It is a matrix of independent variables that affect the calendering thickness of tires. It is the intercept. It is the coefficient of the independent variable. It is an error term; S1-2. The model parameters are estimated using the proposed calendering model. The sum of squared residuals between the predicted values ​​and the predicted values ​​is minimized using the least squares method. The formula for the least squares method is as follows: in These are observations from the rolling mill model. It is the model's prediction for the i-th sample; S2. Based on the tire calendering thickness model, the data from different sensing and measurement devices are first obtained. The missing data is estimated multiple times using the multiple interpolation method. Each estimation introduces uncertainty. The missing data is filled in and the redundant data is deleted to achieve data cleaning of the multi-source dataset. S3. After the tire calendering data has been cleaned, it is arranged into a sequence data by time sequence. Then, by introducing a gating mechanism, an LSTM multi-source data feature extraction model is constructed to store and forget the features of the multi-source tire calendering data. The LSTM algorithm is used to realize the feature extraction of complex multi-source tire calendering data. S4. Based on the feature extraction data using LSTM, the accuracy of tire calendering process is evaluated. A decision tree is constructed using the random forest algorithm to classify the data and obtain the observation weight values ​​and attribute judgment results. S5. Based on the attribute decision results of the random forest, perform attribute association and data fusion respectively. Use the output of the random forest attribute decision as the observation weight. Update and adjust the system's state estimate by initializing the Kalman filter, setting the initial state and covariance matrix parameters. If the consistency test can be passed, perform the final tire calendering data fusion. If not, repeat the process iteratively to finally obtain the features and estimates of the tire calendering data fusion.

2. The tire calendering data fusion method based on Kalman filtering according to claim 1, characterized in that, In step S1, the factors affecting the calendering process include: the distance between the calender rolls, the calender outlet temperature, the speed of the extrusion and pick-up conveyor belt, performance test indicators, extruder temperature, the thickness of the upper and lower sheet drive side and operating side, and the balance of the drive side.

3. The tire calendering data fusion method based on Kalman filtering according to claim 1, characterized in that, In step S2, the data interpolation processing method for data from multi-source tire calendering includes: Based on nine datasets from different sensing devices, the missing data portions were analyzed using Lagrange multiplication. For a given set of n+1 (n=9) data points... The corresponding data points are represented as follows: (calender roll distance, corresponding measured value), (calender operating side outlet temperature, corresponding measured value), (calender drive side outlet temperature, corresponding measured value), (calender conveyor belt speed, corresponding measured value), (performance test index, corresponding measured value), (extruder temperature, corresponding measured value), (operating side thickness, corresponding measured value), (drive side thickness, corresponding measured value), (drive side balance, corresponding measured value). At this time, the Lagrange interpolation polynomial is: in yes Regarding the types of tire calendering data The estimated polynomial, These are Lagrange basis functions, used to weight each attribute. The expression represents the multidimensional attribute values ​​to be estimated.

4. The tire calendering data fusion method based on Kalman filtering according to claim 1, characterized in that, In step S3, the LSTM-based tire calendering feature extraction algorithm comprises three parts: an input gate, an output gate, and a forget gate. The forget gate filters various calendering data states through an update function, retaining high-weight states and removing low-weight states. The input gate is responsible for determining the weight value of the input information and inputting the data with high weight into the state update module. This update process is divided into two parts: the first part updates various calendering data states through a function, and the second part uses an inverse trigonometric function to generate new variables.

5. The tire calendering data fusion method based on Kalman filtering according to claim 4, characterized in that, The update function for the forget gate is: in The weight and paranoia values ​​for the forgetting gate. If the output value of the output gate is the value at the previous moment, then This is the activation function.

6. The tire calendering data fusion method based on Kalman filtering according to claim 4, characterized in that, The function of the input gate is: ， in , This represents the weight and bias values ​​of the input gate. , Representing the state weight value and the paranoia value; the cell state update formula is shown below: This represents the current cell state, is responsible for remembering long-term information, and its updates depend on the forgetting gate. and the output of the input gate ; The output formula of the output gate is: in The output of the output gate It is the weight matrix of the output gate. The bias term of the output gate.

7. The tire calendering data fusion method based on Kalman filtering according to claim 1, characterized in that, In step S4, the attribute determination of the random forest algorithm for multi-source tire data includes: S4-1. Resample the dataset after feature extraction using the LSTM algorithm, randomly select N subsets, generate m decision trees, and the unselected data are called idle data. S4-2. Use decision trees to poll and train multiple datasets for tire calendering process, calculate the weight value of each subset, and sort them. S4-3. Calculate the weight value of each decision tree, and combine the weights of the subset datasets in multiple rounds. The method is as follows: Among them, For the weight values ​​of the decision tree, The weight value of the subset. This represents the variation in the random forest.

8. The tire calendering data fusion method based on Kalman filtering according to claim 1, characterized in that, In step S5, data fusion using the Kalman filter algorithm includes: S5-1. Using attribute decisions based on LTSM and random forest, the tire data classification results are considered as a preliminary estimate of the system state. First, the state vector is defined as follows: This includes information on multiple variables such as the line speed of the calender receiving machine, the distance between each roller, and the roller temperature. Based on the dynamic characteristics of the system, a state transition equation is established to describe the evolution of the given multiple state vectors at different times, expressed as: in Let be the posterior state estimate of the tire data at time k. Let A be the posterior state estimate of the tire data at time k-1, A be the desired state transition matrix, and B be the input matrix. For external input, It is the process noise at time k-1. and Let represent the posterior estimation covariance at time k and time k-1, respectively, and let Q represent the process excitation noise covariance, which is used to represent the error between the state transition matrix and the actual tire calendering process. S5-2. Define the observation equation based on the state transition equation, and map the classification results output by the LSTM and Random Forest algorithms to the state space, as follows: in It is the transformation matrix from state variables to measurements, representing the relationship between states and observations. It is observation noise; Then, the Kalman gain of the system is calculated, expressed as follows: in This represents the Kalman filter gain, and is an intermediate calculation result of the filter. It is the prior estimate of the covariance at time k. It is the covariance of the observation noise; S5-3. Based on the calculation results, the state of the tire calendering prediction data is updated to the calibrated state using Kalman gain. Prediction is achieved through a time update process, that is, the state at the current time (k) is estimated based on the posterior estimate of the previous time step (k-1), expressed as: in, Indicates the measured value; The error covariance is then updated, and finally the required fusion weights and estimates are output based on the consistency test results.

9. The tire calendering data fusion method based on Kalman filtering according to claim 8, characterized in that, The fusion algorithm predicts and outputs estimates by assigning different weights to each data module. The weights are determined in each individual Kalman filter process by calculating the proportion of each sensor error covariance to the sum of the nine sensor error covariances. Let the orthometric error of the data from the nine different sensors be denoted as . So the allocation coefficients for the nine types of data are as follows: 。

Images